Subscribe now

Financials

Beam Therapeutics ($beam) is a clinical-stage gene editing company with active programs in sickle cell disease (SCD), alpha-1 antitrypsin deficiency (AATD), and glycogen storage disease type 1a. They are sitting on 1.25B in cash as of Dec 31, 2025 which includes the cash they received from BMS’ acquisition of Orbital Therapeutics (Beam owned ~17% of Orbital). Beam‘s cash is expected to be sufficient into 2029 through “risto-cel launch and BEAM-302 (AATD gene editor) pivotal trial development plan”. Beam’s stock price sits at $25.5/share which sits comfortably off the 2025 lows of $17/share. But, the share price has been flat over the past 2 years despite many updates from a clinical and regulatory standpoint. 117M shares fully-diluted is the count as of Sep 30, 2025. Therefore, the effective market cap is ~3B as of Mar 11, 2026.

On Feb 24, 2026, Beam announced a new financing agreement for a “senior secured term loan facility of up to $500 million” consisting of 1) initial draw of 100M on Feb 24, 2026, 2) potential 100M draw after risto-cel (sickle cell ex vivo editor) BLA submission, 3) potential 100M after risto-cel FDA approval, 4) potential 100M after risto-cel achieves revenue target, 5) a potential 100M subject to specific agreements among company and lenders. Matures on Feb 24, 2023 (7 years from start), with annual interest rate equal to 3-month SOFR + 6.5% or base rate plus margin. Further, “Among other permissions, the Company is permitted, on terms and conditions set forth in the Financing Agreement, to have outstanding convertible unsecured notes in an amount not to exceed $400 million”. The effective interest rate is 10%. So, Beam must pay 2.5M per quarter to service the initial 100M. If they draw more, they will pay more. Reasonably, the BLA-submission will enforce Jan 2027 draw down, BLA-acceptance maybe Q3 2027, and unclear about revenue target and other agreements. Maximum per quarter interest payment should be ~12.5M per quarter. Of course, it is tied to SOFR so if Fed lowers or raises rates, things may change. This loan implies two things: 1) cash reserves went up by 100M immediately, and 2) if beam is able to commercialize risto-cel, they have capital to use for that specific purpose with an interest rate of 10% and upside for their stock far above that.

Beam therapeutics is a platform company. A non-exhaustive set of relevant comparisons are KRYS (7.5B market cap), ARWR (8B market cap), WVE (2.6B market cap), ALNY (42B market cap), IONS (12B market cap). These companies have gone after different indications and market sizes, and some have executed way better than others. But, this is to say that where beam ends up comes down to continued execution.

More relevant comparisons for Beam are: 1) Verve Therapeutics (PCSK9 and ANGPTL3 editor for ASCVD) which was acquired by Lilly in 2025 for 1B in cash plus an extra 300M contingent on certain outcomes and 2) Orbital therapeutics (in vivo CAR T that used the same LNP technology as Beam was acquired by BMS in 2025 for 1.5B in cash. Note that Verve had 434M in cash when the Lilly deal was announced and its unclear how much cash Orbital had on hand at the time of acquisition. Orbital had non-human primate data but no human data and Verve had a phase 1b trial in progress. This means LLY paid ~600M in guaranteed cash for Verve’s assets. The Orbital premium over cash is likely to be over 1B given their last round was a 270M Series A in late 2023. Translated to Beam, an acquisition target of cash on hand plus 600M or 1B implies a 1.85B or 2.25B takeout value, respectively. Share price on fully diluted count tracks to $16 and $19, respectively. This is, of course, the lowest value possible given that BEAM has significantly more human data than Verve or Orbital and at least as high a probability of success in their intended indications. Further, for Verve-Lilly, Beam is eligible to receive up to 350M in future development-stage payments upon completion of certain clinical, regulatory and alliance events of which 25M has already been received. Before we cover BEAMs clinical programs, a word on their partnerships and management team. John Evans (Beam CEO) emphasized that Beam will continue to pursue drug development partnerships. Beam has received 900M in cash so far from partners, gets to learn from others’ endeavors into targeting different organs and optimize their own approach. Receiving cash to pressure test and iterate on their own technology stack is nice. An example of this is their deal with Orbital which included sharing lipid nanoparticle (LNP) technology. As John Evans (Beam CEO) pointed out, that means the same LNP stack can be used to target non-liver cells (i.e., T cells in Orbital’s case). Another such example is their partnership with Apellis, APLS, (75M up front in cash and 50-50 US co-development and co-commercialization agreement to one program). APLS plans to submit an IND in 2H 2026 for a one-time neonatal FcRn treatment using base editing technology from Beam and lipid nanoparticle technology licensed from Acuitas. I assume the APLS-BEAM LNP for FcRn is targeted to a cell type, and if so, It is likely that macrophages or endothelial cells are their intended target. We will know more as they disclose.

Once a manufacturer has demonstrated success with several consecutive patients with different bespoke therapies, the FDA will move toward granting marketing authorization for the product. Manufacturers will then be able to leverage platform data from such personalized products to gain marketing approval for similar products in additional conditions. Depending on the strength of evidence, either the accelerated or the regular pathway may be utilized.

But, what does this mean? What does it mean operationally for companies developing bespoke therapies for rare diseases? How many trials and what endpoints, surrogate markers, and controls will be sufficient? Recently, Makary and Prasad announced that FDA will default to one pivotal trial for approval not two. They cite the following list of factors that will influence their interpretation of a trial:

the magnitude of the effect, the use of a contemporary control group (versus a historical one), the nature of the control group (is it the best available therapy?), the prespecification of a hypothesis, the choice of a primary end point, the concordance with biologic correlates (including evidence of alteration of an in vivo target), alignment of intermediate end points, statistical power, blinding, concealment, independent review, whether post-protocol therapy is on par with the U.S. standard of care, the use of concomitant therapy, inclusion criteria, exclusion criteria, randomization, run-in periods, how missing data are handled, and many additional factors.

However, they do not offer a position on what evidence they would like to see for each of these factors with respect to rare diseases (where single pivotal trials with confirmatory evidence are more common). FDA, on February 23, 2026 released draft guidance for industry entitled “Considerations for the use of the Plausible Mechanism Framework to Develop Individualized Therapies that Target Specific Genetic Conditions with Known Biological Cause”. The entire 23 page document can be found here. The remainder of this post will be relevant highlights from the document mixed in with some commentary.

• A therapy that qualifies for plausible mechanism pathway (PMP) must target the underlying pathogenic alteration behind a disease, show that it improves clinical course. The company should have data characterizing the natural history of the disease in untreated people. Data to show that the therapy actually targets the proximate pathogenic alteration must be presented. There should be evidence that the targeted genetic variant plays a CAUSAL role in the disease.

• Natural history cohorts, if well put together, along with each treated patient acting as their own baseline can be an adequate trial design

• This is a direct read through for companies that offer gene replacement, gene editing, and gene silencing. Of note, successful clinical data from one vector (e.g., viral vectors and lipid nanoparticle formulation) can be re-packaged with different editors/replacement therapies/silencers. This guidance makes it MUCH easier and easier to commercialize treatment for ultra-rare and rare diseases, respectively. This lines up with Makary’s statement that, “In short: We want to see more Baby KJs” (STAT News).

• Cell line data showing appropriate on-target efficacy and lack of off-target effects will be sufficient for expanding applications. Technologies with known tail-risks are not as well positioned as relatively safe editing/silencing/replacement technologies. Specifically, I’m thinking of companies that introduce double stranded DNA breaks via Cas9.

• Arguably the most important section. If you can articulate why it is not feasible to run a RCT, FDA may not require one. Especially if you have a well-characterized natural history dataset and collect clinical and biomarker data before treatment starts. Off the top of my head, reasons for not needing an RCT would be that the disease is rapidly fatal, progresses too quickly, treatment is too invasive, or prevalence is so low that enrolling patients is impossible. Again, to make a compelling point for any of these reasons, a company will need solid natural history data, and data tying their specific alteration to the disease process.

• PRESPECIFIED protocols (relevant to those invested in uniQure). Anyhow, surrogate endpoints can be considered too, especially if that marker predicts clinical progression or outcomes reasonably well OR has been used in prior approvals (which we will see in the next highlight). Further, natural history can be used as external control if, for example, the outcome we are measuring takes a long time to manifest or slow. Cognitive outcomes are one such place.

• If using natural history, outcomes in the trial need to be HIGHLY concordant with those measured in the natural history data. It also goes without saying that the natural history patient population SHOULD REFLECT the trial population.

  Also, biomarkers can be used if they are established to predict clinical benefit or likely to predict clinical benefit. One such example in the neurology space is NfL. QALSODY (ASO for ALS due to pathogenic SOD1 alterations) did not meet statistical significant primary endpoint measured by ALS functional rating score. Secondary points of plasma NfL reduction and CSF SOD1 protein level were significant. And, it was approved via accelerated approval in 2023.

• “A benefit-risk assessment will be based on the data collected, considering

  demonstrated benefits and adverse events, with consideration of the severity of

  the disease and patient tolerance for risk”. How fast a disease progresses, other treatment options, and the demand by patients will be important.

• Postmarketing studies will be necessary for each therapy approved via PMP.

• It will be critical to ensure that the manufacturing process is locked and commercial ready early in the process, possibly even pre-IND. CMC reasons lead to FDA CRLs. One recent example is ScholarRock’s Apitegromab for SMA which received a CRL solely for CMC reasons.

FTD-GRN accounts for 10% of all frontotemporal dementia (FTD) cases and are caused by loss-of-function mutations in the progranulin gene. FTD-GRN is an autosomal dominant disease with 90%+ penetrance by age 70. Survival is six years from symptom onset. Cells in the CNS make progranulin which they then secrete into the extracellular space. From here, cells endocytose progranulin where they reside in lysosomes, are processed into granulin proteins, and help lysosomes function properly. In FTD-GRN, an extracellular progranulin deficiency leads to lysosomal dysfunction (reduced cathepsin D activity, impaired trafficking of prosaposin and decreased glucocerebrosidase activity). In microglial cells, progranulin deficiency prevents lysosomal clearance of myelin debris. In astrocytes, progranulin deficiency promotes profound synaptic degeneration. Progranulin-deficient neurons accumulate TDP-43, tau and phosphorylated alpha-synuclein. All together, these changes lead to the progressive loss of neurons in the frontal and temporal lobes.

FTD-GRN is similar in setup to Alzheimer’s disease (AD) and therefore, we can learn a lot from the trials and drugs in AD. EVOKE/EVOKE+ tried semaglutide in AD and there have been a number of trials of anti-amyloid beta (anti-AB) antibodies in AD. Semaglutide did not change the course of AD neurological functioning scores compared to placebo but a number of the anti-AB drugs did. They slowed the progression of AD by a small amount. The issue is these are IV drugs that require dosing at least every 4 weeks, and carry a decent risk of brain bleeds. The principle of the anti-AB drugs are that they attempt to intervene on the pathological process in AD (amyloid beta plaques) and seek to slow progression of disease. Very similar ideologically to the FTD-GRN goal. The issue is in execution with the anti-AB drugs: dosing frequency, route of administration, side effects, intervention too late in disease course, and limited ability to halt progression of disease.

However, if you offered a patient with AD the pure treatment benefit of anti-AB drugs (small help in cognitive scores) BUT it is now a one-time treatment and significantly safer, I think a lot of folks would take it. Anti-AB drugs were offered too late in the disease course in the AD trials (and carries many side effects). And, we can run into a similar problem with FTD-GRN. FTD-GRN may first start as a progranulin haploinsufficiency problem then, over time, progress into the buildup of other toxic byproducts in the brain which could continue damaging the brain independent of the original progranulin issue (credit to private conversations with medstudentinvst). It’s like a runaway train that’s off track and heading straight for a mountain. Nobody cares that the reason the train went off track was because of a malfunctioning screw. Replacing that screw at this point does nothing. From the AD space, we should also remember that treating patients as early as possible in the disease course is best. Therapies in FTD-GRN and AD cannot reverse damage or regrow neurons or their connections. We can only hope, at best, to preserve remaining function. So it stands to reason that the earlier in the disease course we intervene, the better.

Now that we have some guiding principles, let’s evaluate the FTD-GRN therapy space: 1) Prevail-Lilly, 2) Passage Bio, 3) Alector, 4) AviadioBio

1) Prevail-Lilly - AAV9-GRN gene therapy

Prevail Therapeutics is trialing an intracisterna magna injection1 of AAV9-GRN for treating FTD-GRN. They were bought for 880M + 160M CVR in 2020 based on early phase 1/2 trial data. In a 2024 paper, they reported low and mid dose data where mid dose was 2x low dose. CSF progranulin levels over time are shown below:

Normal CSF progranulin levels range from 3-8 ng/mL with variance in the range depending on who you ask and how you measure (we’re talking about NANOgrams). For 4 patients with 6 month mid-dose data, median CSF progranulin level was 8 ng/mL and low-dose was 6 ng/mL. Progranulin reached upper limit of normal (ULN) or 1.5x ULN with a clear trend of a peak in progranulin at 2 months followed by ~50% declines from the 2 month peak over the next 10 months.

Adverse events from the paper listed above. Three thrombotic events within 2 months noted. Risk factors included concurrent sirolimus and croticosteroids use and BMI high and immbolity per paper. Further, from looking at the table above, regardless of how they judge the relationship, we see 3 DVT/PE events, and at least 3 pure AAV immune-related conditions (fever, CSF inalmmatory response, sensorineural hearing loss). A good FTD-GRN therapy would not have these side effects.

CSF NfL is a sensitive but not specific marker of neurodegeneration. And, we can see in the above graphs a clear spike in CSF NfL at 2 months followed by a retreat over the following 10 months. At ~6 months, here are the rough CSF NfL changes: low dose, 6000 —> 3000; mid-dose 5500 —> 6000 (anti-aav9+); low dose, 6000 —> 6500, low dose 5000 —> 7000, mid-dose (anti-aav9+) 6000 —> 8000, low dose 9000 —> 9000, low dose 5000 —> 12000. Very rough estimates but directionally, we see that at 6 months NfL is 10%+ higher at least. Not good.

The authors propose that NfL transiently increased in CSF at both dose levels because of a secondary inflammatory response. NfL trends mirrored progranulin levels: 2 month peak followed by drop. Efficacy comes with adverse effects for Prevail.

But, what about their cognitive scores? CDR global was 1.42 (SD = 0.67) and 1.57 (0.53) in low (n = 6) and mid (n = 7) dose at baseline. CDR sum of boxes was 8.08 (5.07) and 9.42 (3.37), respectively. At month 6, the sum of boxes total scores increased compared to baseline in the low-dose and mid-dose cohorts by, respectively, 2.30 ± 2.61 and 3.90 ± 2.68 (mean ± s.d.). At month 12, in the low-dose cohort, the sum of boxes increased by 4.70 ± 3.23. The authors conclude that this is in line with the ranges of natural history FTD-GRN cohorts. Said another way, treatment does not confer materially different trajectories than that of untreated FTD-GRN patients. Unfortunate.

2) Passage Bio - AAV1-GRN gene therapy

Passage is trialing an AAV1-GRN gene therapy via intracisterna magna injection. The belief here, as it is with Prevail, is to replace progranulin in the CNS extracellular space such that neurons and glial cells can endocytose it, restore normal function, and prevent the progression of disease. As their CEO Will Chou said, “we do not replace neurons so we have to intervene early, we preserve”.

Normal range is 3-8 ng/mL of progranulin (PGRN). They are aiming to get mean CSF-PGRN at or above 8 per Will’s comments in early Nov 2025 on a Guggenheim webcast. Here is their most recent data from their ongoing phase 1/2 trial. Note that dose 2 is 50% dose 1 and the data for dose 2 is only for 1 patient. More data for dose 2 to come in the coming months.

Immediately, we notice the dose 1 PGRN levels are much higher than the Prevail ones and the PGRN levels go up consistently over time. For these 8 patients, CDR Sum of Boxes is mean 10.3 (5-17 range) with 50% CDR global 1 and 50% CDR global 2 with mean disease duration 2.9 years.

Passage’s NfL data are much better than Prevail’s. Jury is still out on whether this translates to cognitive scores declining at a slower rate than untreated patients.

Critically, no systemic inflammatory response or CNS inflammatory response observed. No complications with ICM administration. And, like Prevail, low dose anticoagulation will be given with treatment to prevent DVT. After the first patient in the first dose developed an asymptomatic venous sinus thrombosis and hepatotoxicity, the immunosupression regimen was updated to be 3d IV methylpred followed by methylpred PO.

3) Alector - anti-sortillin antibody

Alector trialed an anti-sortillin antibody for FTD-GRN given IV every 4 weeks. IV dosing every 4 weeks is much more of a burden for patients than a one time 1 hour procedure.

Sortillin is a cell surface receptor that allows for the uptake and subsequent lysosomal degradation of progranulin. Alector reasoned that inhibiting sortillin may be a viable strategy to preserve progranulin. Inhibit an inhibitor. This is an indirect way of approaching the primary problem of progranulin deficiency. But, there are two problems: 1) intralysosomal progranulin deficiency is the issue in FTD-GRN NOT extracellular, and 2) lysosomal processing of progranulin by sortillin, prosaposin, or other pathways is necessary for normal lysosome function.

Measuring extracellular progranulin is a proxy for the intracellular level ONLY when the pathways for endocytosis are not messed with. Otherwise, these the concentration in one compartment does not need to be a marker for the other. These were comments that were echoed by Passage CEO Will Chou too.

Let’s review the clinical data from their trials. Here is their INFRONT-2 open label data:

Notice that at maximum, CSF PGRN is not anywhere near the Passage CSF PGRN. Alector’s CSF PGRN data sits squarely in the 3.5-5 ng/mL range which is far lower than the AAV approaches. NfL CSF concentration in FTD-GRN cohort is stable over time which is certainly better than increases over time and maybe even in line with age-related changes.

Here are the cognitive outcomes trends:

We see that CDR-SB + NACC FTLD-SB trajectories are different between natural history data and open-label INFRONT-2 data. The 95% confidence intervals of the INFRONT-2 and natural history data overlap though.

INFRONT-2 finished end of 2021 (12 mo period), then Alector started INFRONT-3 that took until Q4 2025 to finish and read out. Here is the phase 3 trial structure which had a placebo group.

“In the 96-week, double-blind INFRONT-3 trial, latozinemab did not meet the clinical co-primary endpoint of slowing FTD-GRN progression. Although treatment resulted in a statistically significant effect on the biomarker co-primary endpoint of plasma progranulin (PGRN) concentrations, the secondary and exploratory endpoints, such as fluid biomarkers and volumetric magnetic resonance imaging (vMRI), demonstrated no treatment-related effects on FTD-GRN”.

We must await their presentation, but while we do there is one thing to note from the INFRONT-2 and INFRONT-3 inclusion/exclusion criteria2. The phase 2 allowed only for “good physical health” and “no clinically significant findings from medical history, physical exam, lab tests, ECGs or vital signs”. That enriches for people who are healthier and more likely to 1) be able to handle a drug with side effects, 2) come every 4 weeks for therapy, 3) more likely to respond because they have a lot of physiological reserve and clearly, their fatal condition has not progressed too far. This is a big change in my opinion.

I should note that the phase 3 enrolled 16 of 119 patients with CDR + NACC FTLD of 0 which means these folks had no impairments as measured by the survey instrument. These patients, if untreated, will have better cognitive scores and NfL levels at 12 months compared to CDR+NACC > 0 patients.

We know the drug does not boost extracellular PGRN levels much and we know that this drug blocks one of the two main pathways of progranulin endocytosis. Therefore, this drug can’t boost intracellular levels much and these CDR+NACC FTLD = 0 patients on-treatment will experience a steady march of their disease and NfL patterns consistent with their off-treatment counterparts. Reasons the Alector Phase 3 failed: trial population changing to Alector’s detriment, drug producing a CSF PGRN of only 4 ng/ml, and issue that blocking sortillin impairs intracellular PGRN activity. Alector walked so others could run.

4) AviadioBio/Astellas - AAV9-GRN gene therapy

Intrathalamic AAV9-GRN gene therapy infusion currently in phase 1/2 trials (ASPIRE-FTD). They cite “biodistribution issues with approaches including systemic injection and cisternal magna injection”. And, I think the intrathalamic infusion risks outweigh the benefits.3 The only real differentiation between Aviado and Prevail is the route of administration and that Prevail has a multi-year head start. Maybe intrathalamic is the best but that’s not what I think.

They seem to be having a tougher time recruiting. Started trial 8/30/2023 and as of June 2025, only have data for 3 patients of which they have no NfL or CSF PGRN longitudinal data. This is a nothing burger.

Head to Head Passage and Prevail comparison:

• Prevail requires symptomatic FTD and sets CDR+NACC FTLD Sum of Boxes 0.5-15 in their inclusion criteria. Passage historically enrolled patients with similar CDR and NACC scores as Prevail did (in their 2024 paper), but are moving to now “enroll patients who are prodromal or have mild cognitive impairment and to exclude patients who are more severely progressed”. This is similar to the INFRONT-2 trial population requiring good health status and no abnormalities on medical exam. It is the at-risk population (n = 16) that INFRONT-3 enrolled. Intervention earlier in the process is better IF the drug can act successfully on the disease process.

• Prevail’s genetic elements are made to enable constant high-level expression of PGRN in CNS tissues. Prevail uses a CMV promoter which is susceptible to shutdown by endogenous cellular machinery. However, the CB7 promoter that Passage uses does not have the same shutdown risk. CB7 tends to have more sustained expression compared to CMV in multiple gene therapy contexts. 4

• AAV9 used by prevail nonspecifically transduce neurons and non-neuronal cells whereas AAV1 shows specificity for glial cells when given via intracisterna magna injection.

• Notably, 60-90% of AAV administered to CSF distributes systemically to major organs regardless of serotype. AAVs are going to circulate in the body and go to the liver and all other organs. A similar proportion of older adults are AAV9 and AAV1 seropositive (73% of adults over 65 have AAV9 antibodies. About 70% of adults have AAV1 antibodies). But, we know that AAV9 has greater systemic toxicity, in part because it is better at transducing cells. Watch out for liver toxicity, and broad inflammatory reactions.

Passage has the better drug, trial design, and focus. Prevail is split amongst a few programs and Lilly has hundreds. This is the least of Lilly’s worries or interests. Here is the only place in Lilly’s Q3 deck where Prevail’s program is mentioned.

Risks to Passage, at a high level, are a) Lilly beats them to market / pivotal trial readout, b) trial recruitment stalls, c) drug doesn’t work. Lilly and Passage are at very similar stages of the process but its been a year since we have heard anything from Lilly about this. Trial recruitment can always stall, but this is why we follow their updates and see how it is going. If Passage’s drug doesn’t work, Prevail’s won’t either. Passage has a market cap of 37M (as of Jan 4, 2026) and cash till Q1 2027 whereas Prevail was bought out for 880M. Passage is the clear play here with 10x upside pending solid clinical data and execution of their Phase 1/2 program.

Closing comments on UniQure and FDA evidence principles

The issue with UniQure was that they did not prespecify their statistical analysis plan. It was not their use of an external propensity matched control. Propensity matching is not randomization and is not a stand in for it. But, when the disease is severe, and we have no disease modifying agents, it is reasonable to use external controls from well-characterized natural history cohorts (e.g., GENFI2 and ALLFTD are the FTD cohorts). The recently released Real World Evidence Principle from FDA along with prior approvals based on natural history data (e.g., Zolengsma) is a clear signal that such external cohorts can be used for comparison. This is good for patients with rare diseases where running placebo-controlled trials is not always possible, ethical, or feasible. Lastly, from a regulation perspective, Will Chou (CEO of Passage) mentioned that the recent rare disease evidence principle from FDA suggests more openness to single arm studies under specific conditions: rare quickly progressing diseases where a drug that addresses genetic cause is available and no other treatments are available. This fits Passage and FTD-GRN well.

What can we look forward to for Passage and what to watch out for?

• They plan to have early discussions on experimental design of pivotal trial with FDA in 1H 2026

  • Does FDA want a placebo arm? Or will an open label Phase 1/2 and ongoing Phase 3 be sufficient?

• Data refresh from Phase 1/2 trial in 1H 2026 and will get guidance from FDA after data refresh

  • What happens to the NfL signal with more patients at the lower dose? Are we able to return NfL changes to that expected of age-related changes?

  • To what extent does NfL levels correlate with cognitive score levels (i.e., CDR-SB + NACC) over time? This is relevant to the whole field of FTD therapies.

  • How quickly are they able to enroll patients?

  • What happens to the cognitive scores over time?

  • What are the trends in CSF PGRN levels at the lower dose with more patients?

• All product made by Catalent, and no need to manufacture any more supply. Next supply will be for registrational cohort so nothing on the manufacturing front for now.

• Cash to 1Q 2027, 30M per year burn. Will need to fundraise in the next year, likely off updated Phase 1/2 data.

2

Inclusion criteria for INFRONT-3:

INFRONT-2 inclusion criteria:

The 539% number has been quoted a lot on social media. Cosmo has not found the cure for male pattern hair loss. The 539% only considers one of the two trials they ran. And, both trials are identical with respect to what treatments they gave, how the trial was designed and which types of patients they enrolled. Sample sizes are roughly the same, assuming equal dropout rate in each study1, so the true measure that should be quoted is the average of 539% and 168% or 353%.

These two numbers being so different from one another suggest that the trials actually enrolled markedly different patients. This was not on purpose as you can read from their clinicaltrials.gov page for both trials. The clinics enrolling were different, so perhaps patients with more severe male pattern hair loss were enrolled in the 168% study or there was uneven dropout between groups which juiced the numbers in one group. It’s hard to say what exactly was different, we just need to wait for them to present the results. Let’s just pretend 353% is the proper goalpost to compare to for now with the understanding that the pooled number is probably lower.

First, what exactly is the metric they are reporting? 539% relative improvement over placebo and 168% relative improvement over placebo means what? Their primary endpoint is change from baseline to month 6 in non-vellus target area hair count. So, each patient gets a value D = {hair count at month 6} - {hair count at baseline} for the target area on the scalp defined at baseline. Now, take the mean of all the D values for the patients in the placebo group. Then, take the mean of all the D value for the patients in the treatment group. So, maybe in the treatment group we had an average of 40 hairs extra grow. And, in the placebo group we had an average of 8 hairs extra grow. Then, we take the improvement over placebo. Specifically, we now do (40-8)/8 = 4. Now, express it as a percentage, 4 * 100 = 400%. Ta da, 400% relative improvement over placebo. In obscures the actual effect size of how many hairs grew back.

Note that if the placebo group had a mean D value < 0 meaning they lost hairs, then the relative improvement metric cannot be calculated. As far as I can tell, the relative improvement metric was not pre-specified and certainly not present in any materials of theirs I can find online. What was pre-specified was the change from baseline to month 6 in treatment group relative to placebo. In the numerical example above, that would be 40 - 8 = 32 hairs. There are two reasons management would not report this number: 1) the number is so small and lower than any of the other trials out there that explicitly measure this (e.g., finasteride trials), 2) it can confuse the lay audience of non-scientists / hair people. If you read the hair literature, reporting the change from baseline to month 6 in both groups is standard so anyone investing in this field would get it.

There is also an ERROR in their reporting. They quote a 539% improvement relative to placebo. Then, in their fine print on the same press release, they write “5.39x relative improvement in TAHC versus vehicle“. A 2x improvement is a 100% increase. So a 5.39x improvement is a 439% increase. A 539% improvement is a 6.39x increase. Which is it! Here is the screenshot:

Topical finasteride is the best comparison to topical clascoterone. So, how does finasteride perform on the same endpoint in a very similar trial population that COSMO measured for clascoterone?

Roughly a (20.2 - 6.7) / 6.7 = 200% improvement relative to placebo at 6 months with N = 105 in topical and N = 97 in placebo. Depending on how you interpret the error in their press release, Cosmo either has a 353% improvement OR a 253% improvement ( 5.39x improvement = 439% and 1.68x improvement = 68%). Fairly similar. Clascoterone (Cosmo’s drug) is clearly doing something, but not curing baldness levels of something. And, one more thing to note is that the finasteride trial was only once daily application whereas the clascoterone was twice daily. The dose makes the effect (and side effect) with these sorts of medications.

Clascoterone (brand name Winlevi) is an FDA approved drug for acne vulgaris. The acne trials Cosmo ran showed that it was generally safe. 7-12% report redness, itching, and scaling. Hyperkalemia and HPA axis suppression could occur but rare. However, the acne FDA label is for 1% whereas the SCALP1/2 hair loss trials tried 5%. Dose makes the poison and we should expect greater side effects in SCALP1/2. With sufficient topical usage, the drug will enter systemic circulation. We know this is true because cases of hyperkalemia and HPA axis suppression occurred which would be impossible with topical-only clascoterone penetration. At 5% applied twice daily indefinitely, patients will see systemic effects. This makes the following statement an overstep:

“clascoterone 5% topical solution is the first topical androgen receptor inhibitor designed to target the biological root cause of male-pattern hair loss without the risks associated with oral therapies.”

This builds into the next point of mechanism of action. Clascoterone is an androgen receptor inhibitor. In the hair follicles, it prevents dihydrotestosterone (DHT) from binding to the androgen receptor and triggering hair-destroying pathways. Finasteride is a 5-alpha-reductase inhibitor. It inhibits type II 5-alpha-reductase which converts testosterone to DHT. Lowers amount of DHT available nd therefore limits the ability of DHT to bind AR in the hair follicles and elsewhere in the body. Long story short, clascoterone does basically the same thing as finasteride. Oral finasteride like oral clascoterone will have more off-target effects than their topical counterparts. Topical finasteride is basically equivalent in terms of on-target effect to clascoterone. We don’t have as much long term safety data on clascoterone as we do finasteride so it’s hard to forecast what may be reported with respect to libido and other sexual side effects.

Brief chart of the effect of testosterone and its metabolites:

In the SCALP1/2 trials, the exclusion criteria reads:

Topical scalp treatments for hair growth including minoxidil, Aminexil, hormone therapy, anti-androgens, or other agents that are known to affect hair growth within 12 weeks of Visit 2/Baseline”, “Platelet rich plasma (PRP) procedure on the scalp within 6 months of Visit 2/Baseline”, micro-needling, light or energy treatment, or hair transplants. Also, “Any 5 alpha reductase medications (i.e.: Finasteride [Propecia®, etc.], Dutasteride or similar products within 6 months of Visit 2/Baseline

So basically, the person can’t be taking anything else for at least 3 months if minoxidil and 6 months if finasteride. That’s a long washout period and basically ensures that therapy if it has any on-target effect will work. We know DHT blocking works to improve hair growth. If you do it by blocking the receptor that DHT binds OR decreasing DHT, it’s all the same thing.

Novo’s US patent for semaglutide expires 2031 and as early in 2026 in China and Canada. Compounding pharmacies like Hims will continue to battle with Lilly and Novo. Novo market share continues to get eaten by Lilly both US and worldwide. Novo lost the Metsera acquisition to Pfizer. And their board got overhauled a few days ago. Novo appointed a new CEO in July of this year. This is why they trade at a 13.28 P/E ratio and sit near a 52 week low while Lilly rides high at P/E of 50 and near 52-week high. Novo management say this about their own trial in September: “We always presented it as a lottery ticket and it still is because it’s very uncertain” and their comments on their Q3 call echoed the sentiment. Morgan Stanley and other analysts “expects Novo Nordisk Alzheimer trial to fail”. Now that I have set the scene, we can talk science.

EVOKE and EVOKE+ are looking at semaglutide-placebo difference from baseline to week 104 in the clinical dementia rating sum of boxes score (CDR-SB). The CDR-SB survey captures cognitive and functional impairment across 6 areas: memory, orientation, judgement and problem solving, community affairs, home and hobbies, and personal care. Higher scores are worse. Improvements or slowing of decline on this axis is all that matters. It is irrelevant if the treatment treats the root cause, has a fancy mechanism of action, or improves some biomarker.

The trial population is 55-85 year old patients with mild cognitive impairment (MCI) or mild dementia due to AD with confirmed amyloid abnormalities on PET or CSF analysis. 104 week main treatment phase with 52 week extension. 8 week semaglutide dose escalation regimen (3 mg [0-4 wk], 7 mg[4-8 wk] and 14 mg [8-156 wk]). Plasma biomarkers and CSF will be analyzed too, these are exploratory. Up to 30% of overall trial population could have T2D, but no stratification by T2D status was conducted at randomization.2 Patients in either arm can be on standard of care (e.g., anti-amyloid antibody or memantine).

EVOKE+ allows patients with small vessel pathology (i.e., 1+ lacunar infarct plus/minus age-related white matter changes) to join. And we may find that small vessel associated dementia can be treated well by semaglutide, and that would be a broadening of their label beyond AD but for today, we will just think about early-stage AD patients. ECG, physical exam, neurological assessments, and markers (NfL, p-tau 181, GFAP, hs-CRP) at week 52, 104, 156. No futility or interim analyses are planned. In their exclusion criteria, they note patients cannot be on a GLP-1 in the 90 days prior to screening.

Why semaglutide will result in statistically significant improvements CDR-SB scores relative to placebo at week 104 compared to baseline?

1. Has there been an indication that semaglutide has NOT worked in? Diabetes, cardiovascular disease, obstructive sleep apnea, liver disease, and encouraging data in alcohol use disorder, opioid use disorder, and smoking cessation.

   A 9-week, placebo-controlled study of 48 adults with alcohol use disorder demonstrated that once-weekly semaglutide reduced alcohol self-administration in the laboratory (β, −0.48 [95% CI, −0.85 to −0.11];P = .01), craving (β, −0.39 [95%CI, −0.73 to −0.06];P = .01), and drinks per drinking day (β, −0.41 [95% CI, −0.73 to −0.09]; P = .04) despite a short treatment duration in a non–treatment seeking population.

   Primitive GLP1s have shown mixed results (3 positive, 2 negative) in Parkinson disease.3

2. Liraglutide, an older generation GLP1, led to successful results in AD. Phase 2b trial of liraglutide led to reduced shrinking in parts of the brain that control memory, learning, lanugage and decision making by nearly 50% compared to placebo and was presented at AAIC in 2024. However, a full paper was never released. But, Novo did fund the liraglutide study so you have to assume they understood what went on in the trial and made the informed decision to continue the semaglutide trials. Here is some more commentary from alzforum on the liraglutide trial:

   According to the published trial protocol, clinical secondary outcome measures included change on the ADAS-Cog-Exec z score—which contains the ADAS-Cog and the executive function portion of the neuropsychological test battery—as well as the Clinical Dementia Rating Scale Sum of Boxes and the AD Cooperative Study—Activities of Daily Living (Femminella et al., 2019). Edison reported that on the ADAS-Cog-Exec, participants on liraglutide declined less precipitously over the 12-month trial relative to those on placebo, and that, at p<0.01, the effect was statistically significant. However, cognitive trajectories of the treatment and placebo groups largely overlapped, as did the error bars. Any difference between the groups was small, at best.

   Edison reported that the drug curbed gray matter atrophy by 50 percent—a statistic then widely reported in the news. What was it based on? Participants in the placebo group lost a total of 13,500 voxels, or 1mm³ cubes, of cortical gray matter throughout the trial, while those taking liraglutide lost about half as many. To put those losses in perspective, both groups started the trial with some 555,500 voxels, meaning placebo and treatment groups lost roughly 2.4 and 1.2 percent of their cortical gray matter, respectively.

   I agree that the imaging based primary endpoint reduction reported on here means nothing. Functional status is all that matters. With 80 patients in each arm, it’d be great if you could see a difference. But with 1800 patients in each arm (pooled across EVOKE and EVOKE+), the trajectories won’t overlap - we’ll know the answer for sure. I do think the p < 0.01 on ADAS-Cog-Exec was nice though, but I will chalk that up to statistical artifact to try to raise the bear case as best I can.

   The liraglutide trial gave once-daily 1.8 mg subq injection of liraglutide for 48 weeks. EVOKE/EVOKE+ gives once daily 14 mg oral semaglutide for 104 weeks. We can look to PIONEER 4 to understand 15 mg PO daily semaglutide vs 1.8 mg SQ liraglutide. Weight loss at 52 weeks is grossly in favor of semaglutide, glucose control is in favor of semaglutide and both have similar discontiuation rate (8% semaglutide, 6% liraglutide) [Pratley et al 2019, Lancet]. 4 So in EVOKE/EVOKE+, we give a better drug for twice as long. Recipe for success.

1. Neuroinflammation is a plausible driver of AD.

   Semaglutide may have limited ability to cross the blood brain barrier. However, studies in animals have shown that semaglutide interacts with the circumventricular regions and specifically accesses GLP-1R–positive brain regions following peripheral administration, including the area postrema in the hindbrain, the arcuate nucleus of the hypothalamus, and the lateral septal nucleus [50]. These observations suggest that semaglutide may have at least limited direct access to the brain as well as exerting peripheral effects

   Semaglutide may be brain-penetrant at least to some degree. We do not fully understand how semaglutide works in the brain and perhaps its effects are neuroprotective through a mechanism we do not fully appreciate.

   We do know that semaglutide curbs peripheral inflammation (hs-CRP, a marker of inflammation, decreases on semaglutide treatment). For example, in the STEP 1, 2, and 3 trials, semaglutide 2.4 mg weekly reduced hsCRP by 39–48% versus placebo at 68 weeks, regardless of baseline BMI, glycemic status, or CRP level [Verma et al 2022].

   And peripheral inflammation can impact neuroinflammation. So, EVEN if you think semaglutide does nothing on the brain, it can exert anti-inflammatory effects on the brain by modulating inflammation in the rest of the body.

Although some benefits of GLP-1 medicines, such as improvements in obstructive sleep apnea and osteoarthritis, are largely driven by weight loss, other important effects appear independent of weight or glucose control. Large clinical trials such as SELECT (cardiovascular disease), ESSENCE (metabolic liver disease), and STRIDE (peripheral artery disease) have revealed weight-independent benefits, which may reflect reduced inflammation and direct effects of GLP-1 medicines on the heart, liver, and blood vessels.

4. It’s all glucose control. And weight loss. So why would it work in AD? This is the main point bears make. I agree with them that glucose control and weight loss do not stop AD from taking place or change the disease causing process itself. But, it could delay progression and offer quality of life improvements for patients with AD at least on par with the rather modest effects of anti-amyloid antibodies with significantly less cost, no risk of brain bleeds, and ease of use (oral pill versus IV infusion). I guess my point to the bears is, who cares if its all downstream of glucose control! The mean age was 72 years, ~15% had T2D, ADS-Cogs-13 was 27, mean MMSE was 24, and mean MOCA was 18.8 at baseline5. Common comorbidities in early-stage AD with these scores are diabetes, hypertension, and hyperlipidemia. About 50% of patients in this age group with around this level of cognitive scores have a BMI > 25. GLP1 is bound to help - we know it does. We also know semaglutide reduces MACE and offers improved glycemic control without the risk of hypoglycemia, so couldn’t everyone in this population benefit from the known effects of semaglutide?

   Having more normal glucose levels improves cognition and slows disease progression in MCI. Microvascular dysfunction contributes to disease progression in AD. [Morris et al 2014, Wang et al 2022, Lin et al 2023, Gonzalez-Rellan and Drucker 2025]

   Losing weight when BMI > 25 also improves energy and cognition [Verone et al 2017, Siervo et al 2011]

   GLP1s improve BP [Krumholz et al 2024] which in turn can improve cognition or decrease the risk of disease progression [Reboussin et al 2025, Hughes et al 2020, Dallarie-Theroux et al 2021]

5. Semaglutide may do a lot of other cool things like regulate blood brain barrier integrity, promote homeostasis of microglia and astrocytes, positively effect the adaptive immune systemic, manipulate synaptic viability and offer neuroprotection. Ok, maybe6. I have no idea. I am going to assume that for the sake of argument it has a zero percent chance of doing any of this but it STILL makes sense that the trial will hit.

6. We have lots of observational data on AD and GLP1s. Observational data results are hypothesis generating and not proof of anything. But still, its better than having observational data with results saying GLP1s speed up AD progression or increase AD incidence. [Schechter et al 2025, Zhang et al 2025, Wu et al 2025, Tang et al 2025]

Semaglutide will slow disease progression of early-symptomatic AD at least to the level offered by anti-amyloid antibodies if not more solely on the basis of its ability to decrease weight, control BP, decrease peripheral inflammation, and control blood glucose levels. Controlling these variables will lead to improved cognition, as is already established. If semaglutide has other magical neuroprotective properties that is great but totally unnecessary to read out positive at CTAD.

Semaglutide is a tolerable medication with a ~8% discontinuation rate and no risk of brain bleeds or necessity for IV infusion (compared to anti-amyloid drugs for AD patients). Semaglutide is cheaper than any other option out there and will continue to get cheaper as more competition arises, patents run out, and government sets new policies.

I am not saying the trial will hit. I am saying the probability of success is greater than what the market thinks. To be fair, a positive result here would move Novo stock like 5-10% in my opinion so not much.

4

Charts of weight loss and A1C on semaglutide and liraglutide treatment.

5

These characteristics are roughly what was observed in the anti-amyloid GRADUATE I and II trials (https://www.nejm.org/doi/full/10.1056/NEJMoa2304430) even down to the APOE e4 homozygous rate.

6

Please see references 17-26 here from the EVOKE and EVOKE+ paper

They ran AMPLIFY-201 and AMPLIFY-7P Phase 1 trials, which enrolled patients with G12D or G12R KRAS mutations who had minimal residual disease (MRD) positive pancreatic ductal adenocarcinoma (PDAC) or colorectal cancer (CRC) after locoregional therapy (e.g., surgical resection of tumor).

Let’s first look at AMPLIFY-7P results published November 2024. They enrolled KRAS G12D/R/V/C/S/A or G13D MRD+ PDAC patients and gave them CpG-DNA 10 mg with either 1.4 mg or 4.9 mg of Amph-Peptide 7P (7 mKRAS peptides linked to albumin binding moiety). 14 patients (13 PDAC, 1 CRC). No dose-limiting toxicities, or discontinuation due to AE. No cytokine release syndrome noted. They reported results for 12 patients, implying 2 patients’ data were thrown out. Maybe they left the study prior to treatment initiation or some other reason.

They ranked their 12 patients by T cell response and when subset to 3 worst responders versus the other 9, the worst 3 had a median disease-free survival (mDFS) of 3.1 months. This is an arbitrary splitting of the data. We should consider all 12 together. At 12 weeks, in the pooled data, there are 4 events meaning the DFS is 66%. (I know the third death in the Q1 group happens a little after 12 weeks, but we’ll just call it 12 weeks for ease of data analysis.) There are 8 patients at-risk after 12 weeks, all of whom get censored because the study is immature. There are nine ticks in the >Q1 group, suggesting nine patients were censored which doesn’t make sense since there are only 9 patients in the group and 1 had an event prior to any censoring. Anyhow, let’s be charitable and assume that was an error by an intern that didn’t get caught.

Going back to their data, higher dose of vaccine produces greater T cell response. But, counting dots in the below figure, I see 4 maroon triangles and 6 blue dots representing 10 patients. But, they report having peripheral blood samples for 12 evaluable dosed patients in their poster from which this figure was also taken. So, they threw out 2 more patients’ data in this T-cell analysis with no notice or explanation in their poster? Presumably these data points belong to 2 low dose patients to get to an even 6 and 6 patients in each dose group. The intern has made a second mistake. Evaluating the data as is, in spite of our concerns about the integrity of the data, we see that first quartile responders are made of 2 low dose and 1 high dose patients. So, that leaves 2 low dose and 5 higher dose patients in the not first quartile group. We have signal here that the 4.9mg dose has a better clinical response than 1.4 mg since non-first quartile patients only had one observed event that occurred at 8 weeks but first quartile patients all died within 12 weeks.

Now, let us look at AMPLIFY-201 published in Nature Medicine (first published in January 2024 and final results also presented in Nature Medicine in August 2025). Population enrolled is adjuvant PDAC (n = 20) patients (and a few CRC patients n = 5) who were MRD+ (ctDNA in 13/25 and CEA/CA199 in 5/25) with KRAS G12D or G12R mutation. This is basically the same as the 7P trial population except they no longer allowed patients with KRAS G12V/C/S/A or G13D mutants. Dose range was much more variable than 7P study, with 0.1mg, 0.5mg, 2.5mg, 5.0mg, and 10mg peptide dose levels. And, the peptide used was 2P - less potent than 7P per KOL call 2024 with Eileen M. O’Reilly (oncologist MSK). I’m assuming this means 2 mKRAS peptide fragments were given. Median of 19.7 months of follow up (2024 update had median follow up of 8.5 months), no new safety signals. Median OS was 28.94 months. Median RFS (recurrence-free survival) by imaging was 16.33 months, maintained from prior update in 2024. PDAC subset had mRFS of 15.31 months and mOS of 28.94 months (shown below).

You can’t look at the OS curve here and think the median value was truly 28.94 months. The curve skims the median hard. The real mOS probably lies between 16-20 months. We can debate where that may be, but it certainly is much less than 28 months. Median follow up was 19 months so the patients in the OS curve around 15-21 that get censored are probably censored because no more follow up available at the time of data cutoff. If any of the 7 censored prior to month 24 have an event, the mOS drops considerably. The Relapse-Free Survival (RFS) curve also skims the median here. The real mRFS is maybe somewhere from 6 to 12 months, not 15 months.

6 patients received subsequent chemotherapy following tumor biomarker increase yet remained free from progression through follow-up. 3 of six patients who remained free from radiographic progression received subsequent therapy, including ICI, following biomarker progression. So, subsequent therapy whether ICI or chemotherapy could be the key here. The argument may be that KRAS-vaccine improves response to ICI or chemotherapy but that is hard to prove with the current trial design.

They define a threshold for T cell response and there are 8 patients below the threshold and 17 patients above the threshold. A quick word on the threshold. In their first paper in 2024, they defined threshold as the median T cell response. Then, in 2025, presumably feeling that was not a robust marker, they did an ROC analysis of T cell response and radiographic progression or death (I think this was the endpoint for ROC, it hard to say what it is based off of) and for each T cell response value they computed the Youden index (sensitivity + specificity - 1) with larger values being better. This is an exploratory analysis.

All 8 below threshold progressed or died, 6 of whom got subsequent therapy. Of the 17 above threshold, 11 got subsequent therapy. Of the 11 who got subsequent therapy, 6 did not progress or die and 5 did not. Of the 6 who did not get subsequent therapy, 5 did not progress. First, this suggests that people with a really large response are going to not progress or die. We can look at individual T cell response and survival time data to confirm this. Second, subsequent therapy is offered if ctDNA levels increase which means the therapy is not working. Most people whether they were above or below threshold got subsequent therapy (17 of 25). Getting subsequent therapy is a 50-50 chance of non-progression/death if you had a good T cell response (above threshold) and a 0% chance if poor T cell response. We can read that in two ways: a) vaccines that create a strong T cell response enable the patient respond to subsequent therapy better, or b) those who respond well to vaccines are able to respond better to subsequent therapy because they have more intact immune systems at baseline, tumor microenvironment more friendly to incoming therapy (e.g., desmoplastic response is not so harsh). Option b is my most likely but I would certainly like to believe option a, I’m just not sure how to de-risk that hypothesis with the current trial design.

Subscribe now

All 12 patients who induced both CD4+ and CD8+ T cells had T cell responses above the 9.17 threshold. This means these 12 patients strongly activated their immune systems. Perhaps they were even naturally primed and already had mKRAS recognition abilities.

Let us study the below figure from the Nature Medicine 2024 paper, it’s filled with a bunch of great information on the trial.

First off, notice there are 25 patients here, the same as in the final analysis, we just have longer follow up in the final. Best overall biomarker response could have gotten more negative for some of these patients over time. But, if we look at the source data for Figure 1d in the 2025 paper with updated follow up and best biomarker response, we see that 6 patients had 100% clearance consistent with the 2024 waterfall plot above. So, 3 of the 6 clearance events were CRC. In the 3 PDAC clearance patients, their RFS censor date was around 24 weeks on study. And these patients shortly thereafter received subsequent therapy, per the chart. They are alive still but certainly had disease progression on therapy. And, we see that all the patients with increases in tumor markers (non-responders) were PDAC patients. They were censored for RFS quite early, around 6-10 weeks on study and received subsequent therapy. 2 of the 3 non-responders died at 16 and 48 weeks. Taken all together, the point is 1) basically everyone had recurrence and were given subsequent therapy, 2) non-responders are PDAC patients who have very poor outcomes and CRC patients enrich for better biomarker responders.

Notice that 3 of 5 biomarker clearers were CRC but only 5 of the 25 total cases were CRC. So, PDAC patients don’t respond as well as the total population. Look at the figure below, 5 patients died. 2 were CRC patients (both responders) that survived for 56+ weeks but the 3 PDAC who died, died quickly at 16 (nonresponder), ~24 (reduction), and 48 weeks (nonresponder).

I can’t really see a strong dose-response relationship in the below figure. Some patients respond really well to vaccines regardless of dose.

Let’s look at safety here from the 2024 paper (n = 25)

There is a dose-response relationship between adverse effect rate and dose. This dose seem like a pretty safe drug. Some anemia. Fatigue, myalgia, nasal congestion expected of something that interacts with the immune system as we so often see with basically any other vaccine.

Now let us look at company statements and what we might expect in standard of care treated patients. From Elicio Therapeutics press release:

IDMC’s positive recommendation to continue the AMPLIFY-7P study to final analysis without modification. Final DFS data [for phase 2 study] expected in Q4 this year, we are focused on advancing clinical preparation for a potential pivotal Phase 2 trial. Upon final DFS analysis, they will request and conduct End-of-Phase-2 meeting with the FDA to finalize regulatory strategy for the pivotal Phase 3 trial for ELI-002.

To refresh ourselves, AMPLIFY-7P is trial of (Amph-Peptide 7P) mKRAS-peptides and CpG-DNA conjugated to lipids as adjuvant treatment in PDAC patients with mKRAS.

Phase 2 of the study is PDAC patients randomized 2:1 to treatment or observation/SOC. Subjects randomized to observation will be able to elect to cross-over to treatment in the event of confirmed disease progression. Notably, patients can be MRD+ or MRD-. Previously, patients had to be MRD+. Dose is 10 mg adjuvant CpG DNA and 4.9 mg 7P peptide. Oligometastatic disease was eligible before, now it is not. This Phase 2 cohort is less sick than their previous patients.

I will note that the clinical data we have is from the Nature Medicine papers that describe adjuvant CpG DNA at the same dose as the phase 2. But, Nature Medicine results only gave 2P peptide fragment (i.e., 2 mKRAS fragments) versus the 7 peptide fragments that the phase 2 gives and the AMPLIFY-7P Phase 1 gave.

Observation is NOT the standard of care in MRD+ PDAC patients in the adjuvant setting. Chemotherapy, chemoradiation, or immunotherapy are the standard. Observation is not a good comparator. Even MRD- patients should be treated, at least a fraction of them per NCCN guidelines (see below).

If most patients are receiving observation then the control arm is doing a great disservice to these patients and is also not a good comparator. It is hard for their treatment to be worse than true observation. Even if the treatment is placebo, placebo against observation probably wins nine times out of ten because placebo can often times make patients feel like they are getting a benefit relative to nothing being done. In summary, if observation really is true observation and not investigator’s choice of adjuvant therapy (or lack thereof), then this trial does not offer as any useful clinical information to assess ELI-002 7P.

Subscribe now

Now, let us consider the scenario where observation means investigator’s choice in the Phase 2 ELTX trial. Eileen O’Reilly, the MSKCC GI oncologist, gave a KOL presentation webcast on ELTX’s IR site in June 2025. She presented the following slide:

Notice that PRODIGE24 established modified FOLFIRINOX (mFOLFIRINOX) as the standard of care in the adjuvant PDAC setting. The modified in mFOLFIRINOX refers to a dosing schedule lower than traditional therapy that seeks to decrease side effects of chemotherapy (and it does do that). The line to meet is 21.6 months for disease-free survival per ELTX’s own KOL. That is also to say that if the observation arm has a mRFS substantially lower than 21.6, then we need to ask a lot more questions. We have evidence that another type of chemotherapy, gemcitabine, tried in ESPAC-4 produces a DFS of 18 months in adjuvant PDAC (see below). So, the absolute lowest mDFS that the observation arm should have in ELTX Phase 2 should be 18 months.

ELTX therapy is less toxic than chemotherapy but in order for it to be approved, in my opinion, it would have to at least be close to gemcitabine mRFS of 18 months if not pushing 21 months. Current ELTX PDAC mRFS is 15 months. And, that was with skimming the median and most patients getting subsequent therapy.

They are telling us that MRD+ PDAC has a median RFS of 5-6.4 months mOS of 17 months. Look above, their own KOL tells a different story.

We can look at the above slide and why it does not make sense by trying to validate the mOS numbers using ESPAC-4. ESPAC-IV showed margin positive PDAC (R1) on gemcitabine had median OS of 26 months (see below figure). This is way higher than the 17 month number they quote in their slides. This leads me to believe that they are reporting the observation no treatment adjuvant PDAC estimates. And, this makes me think that they are using true observation for most patients in the Ph2 trial.

They say in their Phase 2, 80% have median T cell response above 9.5x which was basically the rate in the 7P Phase 1 data. Let’s investigate this claim. Figure 1 from 2024 Nature Medicine paper allows us to figure out which patient IDs had CRC and which had PDAC. CRC patients were ID 16, 19, 21, 13 and 7. Supplementary Source Data for Figure 1 from Nature Medicine 2025 paper gives us RFS in months and T cell max fold change for each patient ID. Granted this data is 2P (2 peptide) but the phase 2 is 7 peptides.

Here is the raw data of PDAC patients T cell response fold change (FC), whether they were censored or not, and RFS in months.

As you can see, there are a few patients with really high T cell responses so when we plot RFS by FC we can take log(FC) to make it interpretable. Here is a plot of non-censored RFS in months against T cell response.

Let us forget that this is an exploratory analysis and choose to use threshold as a predictor of outcomes. So, if a patient has a less than threshold response, they will almost surely have an RFS < 5 months. If a patient has above threshold response, they have a 50-50 shot of 5 month RFS or 18 months RFS. We know that in Phase 2 data, 80% have T cell response > threshold. That means 20% do not and they will have mRFS of ~5 months. The 80% are split between 5 months and 18 months. The median will be 12 months-ish. So 0.8*12 + 0.2*5 = 11 month mRFS is my guess. Of course, this is based only on MRD+ patients and on principle, MRD- patients should have longer survival but I can’t forecast that with the data at hand. This is higher than the 5-6.4m MRD+ PDAC comparison they put forth in their deck, but as mentioned above, this numebr is not the right comparison. 11 months is lower than the 21.6 months of PRODIGE trial shown above.

It is too hard to ascertain whether there treatment meaningfully shifts outcomes in the adjuvant setting for mKRAS PDAC. And I lean to there being no benefit here. There is much excitement in the PDAC space now with KRAS therapy (RVMD, VSTM, IMRX) and adjuvant vaccines are, from a biological perspective, incredibly cool.

Elicio has 22M in cash as of June 30, 2025 and expects to be able to fund operations into Q1 2026. This Q4 readout of Phase 2 disease free survival is a make-or-break event. They will raise cash if positive else I expect the stock to trade at cash after. This reminds me of the cardiff oncology readout earlier this year.

Thank you to Adu Subramanian for comments on the draft. View here only represent my thoughts.

• Be multidisciplinary, no silos (Munger-esque)

  • “As of 1481, the library at the Vatican, the largest in Europe, had around 3500 books and documents. The limited extent of humanity’s collective knowledge made possible and encouraged an interdisciplinary approach that would almost be certain to stall an academic career today. That cross-pollination as well as the absence of a rigid adherence to the boundaries between disciplines was vital to a willingness to experiment”

  • Charlie Munger stated that much of his success in investing came from learning principles from psychology, chemistry, biology, and other non-investment fields.

  • Medical knowledge doubles every 73 days. Nobody can keep up with that. But, the premise holds. People working on problems in medicine should have multidisciplinary backgrounds and interests. Study things like economics, art history, math, and physics AND learn medicine.

  • Looking further, medical students interested in cardiology often end up doing just cardiology research and work. I argue that to-be cardiologists should work hard on problems in seemingly disparate fields like oncology or psychiatry at least for a year or two in medical school.

Subscribe now

• Always ask yourself what the incentives are (Munger-esque view)

  • “At that salary, [Jerome] Powell is essentially volunteering his time to the country. His compensation as an employee of the federal government is negligible with respect to his net wort which has been reported to be in excess of 20 million, and he has said publicly that he essentially lives off his significant savings. But, why are we, as a country, the world’s wealthiest, asking for a volunteer to run the Federal Reserve? What incentives does that create and how dramatically does that winnow the pool of potential candidates who might be interested in the job. We complain about the influence of money in politics only to remain silent as wealthy individuals increasingly dominate political races. The unintended consequences of our approach to public sector compensation is that an increasingly disproportional number of the world’s wealthiest are running for and winning public office, both in the US and around the world.”

  • “In November 1994, Lee Kuan Yew, who served as the first prime minister of Singapore, was caught in a debate with other members of parliament regarding his proposed increases to government salaries. Lee had instituted a system under which the compensation of the island nation’s public officials was set based on comparable salaries in private sector professions, including banking and law. By 2007, for example, the average annual salary of the country’s ministers would rise to 1.26M per year. Lee’s critics argued that increasing salaries would attract the wrong type of candidate, those motivated to pursue government work for personal gain as opposed to public service. At a parliamentary debate on the mater, Lee responded that politicians ‘are real men and women, just like you and me, with real families who have real aspirations in life’. He continued ‘So when we talk of all these high-falutin noble lofty causes remember at the end of the day very few people become priests’.”

  • “It is skepticism of incentives in the domains that are are most important to our collective good that may be part of what is holding us back. Why should we, the public, cede the use of incentives to the finance and banking industries as well as the technology sector”

• Work on hard problems not instagram ads

  • “The market rewarded shallow engagement with the potential of technology, as startup after startup catered to the whims of late capitalist culture without any interest in constructing the technical infrastructure that would address our most significant challenges as a nation. The age of social media platforms and food delivery apps had arrived. Medical breakthroughs, education reform, and military advances would have to wait”

  • We can disagree on what other advances should have been prioritized but I think having some of the smartest scientific minds out there working on instagram ads is not a good thing. The problem is that the incentives have been set up to encourage this behavior, and therefore you cannot fault the individual engineering graduate who makes this choice.

• Eliminating hyper-structured corporate hierarchies preferentially select for employees who just want to get things done and not play status games. These ideas may be relevant when thinking about how to build life science/biotech/medicine organizations.

  • “One can also more readily identify pockets of talent and motivation within an organization once the veil of status, the constricting gaze through which everything is perceive in corporate life, is lifted”

  • “We undoubtedly have some form of ‘shadow hierarchy’ within the company, power structures that are no telegraphed explicitly but exist nonetheless. The lack of organizational legibility comes at a cost, increasing the price of navigation internally, for employees as well as for outside partners who often simply want to know who is in charge. But many discount the amount of open space that a de-emphasis on internal signs and signifier of status, for thousands of employees, can create. The benefit of it being somewhat unclear of ambiguous who is leading commercial sales in Scandinavia for example, is that maybe that someone should be you. Or what about outreach to state and local governments in the American Midwest”

  • “The point is only that voids or perceived voids within an organization in our experience have repeatedly har more benefits that costs, often being filled by ambitious and talented leaders who see gaps and want to play a role but might otherwise have been cowed into submission for fear of venturing onto somebody else’s turf.”

  • Meeting, committees, subcommittees make up the so called “meeting-industrial complex”. “Such meetings are mechanisms by which the ambitious self-promoter within an organization telegraph their status and power and many talented but less manipulative colleagues simply choose to relent at a significant cost to the institution.”

Three more points that I think are interesting. I would think about these three points as sections of the book that piqued my interest but I don’t know exactly how to think about it yet.

• “Rather than resist, we might see this next era as one of collaboration, between two species of intelligence, our own and the synthetic. The relinquishment of control over certain creative endeavors may even relieve us of the need to define our worth and sense of self in this world solely through production and output.” Then, how do we define our sense of self worth and what do we do with that time? Whatever you would like to for the sake of doing it. Spending time with your family, starting an ice cream shop, become a master tea maker. This reminds me of Daniel Ek’s conversation with David Senra where Daniel suggests that AI will make the good a pretty easy bar to achieve but will not replace the exceptional or master of their craft.

• Previous generations were more pragmatic

  • “Nothing of consequence is built in a straight line. A voracious pragmatism is needed as well as a willingness to bend one’s model of the world to the evidence at hand, not bend the evidence.

  • “A certain ravenous pragmatism and insensitivity to calculation had been lost on the current generation. After the end of World War II, U.S. defense and intelligence agencies launched a massive and secret effort to recruit Nazi scientists in order to retain an advantage in the coming years in developing rockets and jet engines. At least 1600 German scientists and their families were relocated to the United States. Some were skeptical about this late embrace of the former enemy. An officer in the US Air Force urged his commander to set aside any distaste for recruiting the German scientists to this new cause writing in a letter that there was an immense amount to be learned from the German-born information if only we are not too proud.”

• Endless possibility is not a superpower

  • “When Odysseys asked his crew to tie him to the mast of his ship as it sailed past the Sirens and their bewitching call, he warned his men, “if I should entreat you and bid you set me free then “with still more fetters bind me fast”. He was intentionally restraining his own range of motion, his ability to respond to the outside world and to the risk of being diverted by its enchanting and indeed deadly temptation. A freedom of motion, to maneuver at will, can masquerade as an imitation of power. A willingness to constrain choice, to cast oneself to the mast, is often the best, if not only, route to creative production for either a company or a culture”

AI models can now solve clinicopathologic conferences (CPCs). Basically, these are some of the most challenging cases to diagnose. The CPCs are presented as curated vignettes.

If AI can diagnose the hardest cases, then what is the role of the doctor? Specifically, what is the role of the mostly non-procedural physician (e.g., medical oncologist, internist)?

Please read Dhruv’s article in the New Yorker, this post builds off the ideas he makes in that article.

Subscribe now

• Bayesian update step

  What question do you ask next? What labs or imaging do you want to collect next? A patient sitting in front of you is not a CPC of neatly organized data points, they are relatively undifferentiated and you have to decide what data to collect.

  Each decision has an opportunity cost. It takes time to send them to get an MRI and their disease process may change in the interim, for example. There is a cost to the health system, insurer and patient too. And finally, there are other patients that also may need your time or access to a CT scanner so if you sit with a specific patient for three hours and order every test, other patients may not get timely access to what they need.

• Synthesis

  In a CPC, you do not need to sort through hundreds of medical notes, tens of imaging studies, and thousands of lab values and trends. In real life, you do. The clinician’s role is to create the clinical summary of the patient just as the staff of NEJM do for CPCs. The art of figuring out what to include in the summary and how to describe it is critical. “Real patients do not present as carefully curated case studies”.

  The natural counterargument to this is, “just throw most if not all the information at AI and have it summarize and sort”. Perhaps one day, but not today. In my experience, how you prompt the model makes a big difference. Here are two examples from Dhruv:

  Dhruv gave the same model that solved the CPC the broad strokes of what a real patient had experienced. The model made up lab values, exam findings, incorrectly interpreted imaging, and fabricated a CT scan read. And, it ended up with the wrong diagnosis. But, when Dhruv gave a formal summary of his visit in medical-ese, CaBot latched on to the pertinent data and did not make anything up. It reached an almost correct diagnosis. Importantly, the next steps CaBot suggested was identical to what should have been done under the exact right diagnosis. The underlying pathology and fix would have been done properly.

  Another example:

  

  These examples are representative of my experience with AI.

• Evaluating the output of medical AI

  A clinician has to know medicine to evaluate the outputs of the AI. If the clinician does not grok medicine, or at least the speciality they practice, then they will be forced to use AI / books / search to evaluate each statement the AI makes which is pointless.

• AI offers the clinician leverage

  With good AI, a clinician can scale their efforts. They can do what they could not have before. They can see more patients (and certainly PE firms will use this to squeeze productivity out of physicians), but they can also dedicate more effort to their patients. They can also think more deeply about medicine to contribute to meaningful clinical and research questions. AI allows the clinician to dedicate more time to working on hard problems.

• Considering the patient’s preferences, values, and goals

  Sometimes the ninth line of therapy for a 83 year old patient with metastatic colorectal cancer is not the right decision. It may be medically indicated but what does the patient want? What are their goals and priorities? What is their quality of life?

• There is often no right answer

  Reasoning under uncertainty is the name of the game especially when evidence is scant. A good example of this is the ICU setting. We don’t have RCTs delineating treatment decisions for each example. Clinicians have to reason under uncertainty.

  A related point is that it is hard to train LLMs in medicine because of the uncertainty. We can train really good math models because most math has a right answer that we can score model outputs against to reach an optimal set of weights. How do you develop a scoring function in medicine? Nobody knows the right answer.

• Underwriting risk

  This is a logistical question that will get figured out with time, and not a particularly strong reason. New insurance models and ways of underwriting risk will be introduced as AI gets better. I don’t find this to be a compelling argument.

There is no role for the non-procedural physician if we de-skill ourselves by relying on AI too much especially while training.

Please comment or reach out with thoughts or disagreements!

1. Engineering genome-scale drug manufacturing systems for biologics and crops

2. Developing biomarkers for disease detection, patient selection strategies for clinical trials, and monitoring response to treatment

3. Multiple in silico mutagenesis prediction

4. DNA FMs As-A-Service

1. Engineering genome-scale drug manufacturing systems for biologics and crops

Look no further than Brian Hie and team’s work fine tuning Evo2 (this is a gross oversimplification, please read their technical blog or preprint - it’s fantastic) to generate new bacteriophages that are functional when synthesized and have the ability to do things like overcome and kill E. Coli that are phage-resistant. A new era of genome design has started. Controllable generation of genomes with unique and specific functions is possible.

To flesh this idea out a little more, consider the problem of designing a 100 base pair sequence with a specific property defined by a known objective function. Suppose a company has collected data on their problem of interest and has an oracle prediction model. How do we search the 4¹⁰⁰ possible sequences? If each generate-predict loop takes 1 nanosecond, naively searching the space will take 10⁴³ years. Harnessing DNA FMs like Evo2 to generate sequences that reflect natural sequence variation and fitness allow us to arrive at optimal solutions faster. Further, we can take DNA FMs and fine-tune prediction heads using proprietary data on the optimization parameters of interest (e.g., antibody yield, drought resistance, etc.) to develop accurate and generalizable solutions.

Asimov has developed CHO cell line data and simulation models for antibody optimization. Ginkgo too. Agricultural companies have crop-related data and leverage predictive models to guide which edits may be most effective (Benson Hill, Ohalo, Inari, Corteva, Cropwise, Pairwise, Tropic). It is unclear if any of these companies are currently using DNA FMs.

2. Developing biomarkers for disease detection, patient selection strategies for clinical trials, and monitoring response to treatment

DNA FMs are a new tool for discovering genomic features that can aid in early detection of diseases like cancer and neurodegenerative diseases, monitoring response to treatment, and identifying patients who will respond to a therapy. DNA FMs can contribute to developing new low-coverage WGS features that capture transcription factor binding sites and other transcriptome- or epigenomic information from cell-free DNA. The devil is in the details here. Implementing this technology in this space and proving its value will be difficult. I think clinical trial patient selection or switching patients from one therapy to another are probably the first two places where DNA FMs could be leveraged.

Spotlight on Polygenic Risk Scores. Polygenic risk scores (PRS) are used by embryo selection companies (e.g., Orchid Health), clinical risk and prevention companies (e.g., Allelica). PRS are actively being developed by many academic groups. Pradeep Natarajan and Nilanjan Chaterjee are academic leaders in the PRS cardiology and oncology space, respectively. PRS are based on estimating effect sizes of each variant in the set and using that as a feature to predict risk or the desired outcome. The fraction of variants selected that are causal to the disease determines the performance of the PRS. We can enrich for biologically meaningful variants by leveraging DNA FMs.

3. Multiple in silico mutagenesis prediction

How can we predict the effect of multiple distant variants across tissues and genomic tracks? Suppose you want to understand the effects of 5 different mutations and a translocation on RNA expression, cell state, and methylation patterns. Currently, we can’t do this. One way to solve this problem is to scale context length.

Currently, we rely on large-scale CRISPRi screens of combinations of variants with multi-omic reporting. But, this method quickly becomes very costly and combinatorically infeasible. We may be able to adapt DNA FMs to do this in silico. There are a few technical modeling questions that need to be solve to achieve this milestone. Consider testing the effect of two mutations x and y. The DNA FM prediction can be represented by some arbitrary f(.). If the mutations act independently then f(x,y) = f(x) + f(y). However, mutations need not act independently and often do not if they act on the same pathway or network. We must model their joint distribution. Simple additive models perform pretty well and much better than current neural networks, so we have a long way to go.

One strategy could be to borrow from ProteomeLM’s method (see below figure) for modeling protein-protein interactions by introducing pair attention features and an interaction process (see schematic from ProteomeLM paper). We also need to capture pathway information given that two mutations in the same pathway (e.g., KRAS and MAPK) can cancel or amplify one another beyond what would be expected under additivity.

The follow-on effect of doing this is being able to propose specific edits to achieve a given cellular outcome. For example, we could predict lethal interactions of higher order combinations and greatly reduce preclinical costs of drug development.

4. DNA FMs As-A-Service

SaaS meets DNA FMs. Basically, serve different prediction (e.g., Borzoi, AlphaGenome, Yorzoi) and generation models (e.g., DFM, diffusion, autoregression-based models) with linkage between the two. On top of that, integrate a front-end LLM and optimization schema to generate guided sequences. Tamarind Bio offers a similar service for protein models. We can also offer fine-tuning of the models we serve for client-specific applications given a custom dataset. LatchBio offers a suite of bioinformatic solutions (AlphaFold, BLAST, Cell Ranger, RFDiffusion) including DNAChisel (simple search and optimizer) and could expand to offer DNA FMs too.

Before, we do that, let’s remember management’s paraphrased words:

Less than half of patients treated with standard of care GNP in this setting survive 9 months. 94% OS and 72% PFS observed at 6 months in 1L PDAC treated with atebimetinib (320 mg QD PO) + mGNP (n = 34). Cutoff for data was May 26, 2025.

The clinical trial data for GNP revealed 6-month OS of 67% and dropped to 50% in 8.5 months. $IMRX’s study does not have a placebo arm so we have to look externally for a GNP bar. The bar is set at 67% 6-month OS and basically 50% 9 month OS.

Two things to point out:

1. When $IMRX released their original 6 month data in mid-June, they guided for additional data update in Q4 2025. Sep 25, 2025 is ahead of schedule.

2. PFS is often a leading indicator of OS because you will progress before you die and if follow up appointments are regular enough, we should catch progression before death. PFS leading OS is pretty solid in pancreatic cancer. See more in the footnote including some caveats1

Part of the story is that Atebimetinib is an exceptionally well tolerated drug given previous high toxicity MEKi drugs. See below for their safety profile chart.

We need to compare apples to apples. Modified GNP is not the same as GNP (MPACT study). Modified GNP = 1000 mg/m2 of gem + 125 mg/m2 of nab-Pac days 1 & 15 q4 weeks whereas MPACT GNP = 125 mg/m2 of nab-Pac + 1000 mg/m2 of gem on days 1, 8, 15, 29, 36, and 43 after which all patients were administered treatment on days 1, 8, and 15 every 4 weeks. So GNP groups gets a lot more chemotherapy than mGNP. And chemotherapy AEs are drop blood counts (neutropenia, anemia, leukopenia), cause fatigue, and lead to GI side effects (diarrhea, vomiting, nausea). The safety profile of IMRX combination is likely due to less chemotherapy being given to patients. 2

Here are the IMRX survival analyses from their mid-June release.

Note that 5 patients were censored and off study early in follow up (grey ticks). And 10 more patients were censored before 6 months because they didn’t have follow up to that time yet (blue ticks). Last time, they presented data mid-June 2025 with data cutoff mid-May so it stands to reason that data until mid-August will be used in the mid-September update. Simply put, we will have 3 months more follow up for each patient in the study. The 10 patients with < 6 months follow up, will have 6 months follow up by Aug 25, 2025 (since last data was stopped May 2025), so we will know if they survived or not. Censoring early artificially props up short-term survival estimates until the data mature.

IMRX’s press release says N = 34 update with 9 month median follow up, so no new patients. They did mention expanding their cohort to 50, but we won’t be seeing that data here.

There were 5 patients that were censored early, within 3 months, who left the study. Why? Two reasons: 1) they sought curative or other treatment, 2) clinical progression necessitating a hospice admission or palliative non-treatment care. 2 is more likely than 1 for patients with metastatic unresectable pancreatic adenocarcinoma. It is my opinion that a conservative estimate would be that at least 3 of the 5 censored off-study patients died. If we factor that into our analysis, that drops the 6 month in the original analysis to around 85%. But, we won’t factor that in because that’s not fair to assume (even though I think its likely).

What might the 6 month OS be in the update? What might the 9 month OS be in the update?

1. 6 month OS updated with additional 3 month follow up. Anyone with at least 3 months of follow up will be included here. Previously, we had 17 at risk at 6 months. Now, we will have 27.

   1. What might the updated 6 month OS be? The current 6 months OS is defined as (1 - 1/34)*(1 - 1/29) = 0.937. If none of the 10 patients who previously had follow up ranging from 3-5 months did not die, the 6 month OS stays the same. If any one of them die, the OS estimate gets revised down.

   2. Of the ten, conservatively, 3 might die before 6 months. Let’s assume they die at 5 months. If that happens the 6 month OS is now (1 - 1/34)*(1 - 1/29)*(1 - 3/27) = 0.83. If 4 or more deaths occur, the 6 month OS rapidly approaches the GNP 6 month OS of 0.67.

2. 9 month OS updated with additional 3 month follow up. Anyone with at least 6 months follow up at the time of the first analysis will be included in the update and not censored. 17 patients had 6 months of follow up back then so we should expect 17 at risk for the 9 month OS presented Sep 25. The expression for 9 month OS is roughly 6-month OS * (1- K/17) where K is the number of deaths that occur in the 17 patients who will had follow up to at least 6 months in the old update and will have updated up to 9 months now.

   1. What might the 9 month OS be?

      1. Assume no deaths in the 10 patients with follow up from 3-6 months in the old update.

         If K = 3, 9 month OS = 77%; K = 5, 9 month OS = 66%; K = 9, 9 month OS = 44%

      2. Assume 3 deaths in the 10 patients with follow up from 3-6 months in the old update.

         If K = 3, 9 month OS = 68%; K = 5, 9 month OS = 58%; K = 9, 9 month OS = 39%

Let’s be as generous as we can and assume zero new deaths until 6 months and only 3 deaths between 6-9 months. The OS is 77%. That’s the most generous version. I think we hover around 9 month OS of 65%. To be clear, 65% 9 month OS is higher than the ~50% OS of GNP at 9 months with less toxicities. So, its still a better option than current standard of care. Mizhuo says “9 month OS of 75% would represent excellent outcomes”. I don’t see that happening.

Of course, confidence intervals will be wide given small sample sizes. We also need to keep our eye on the number of censored off study patients and how many more we get of them in the update. OS numbers will not hit the level people are expecting in part because this is not a miracle drug (though I think its a good drug) and metastatic unresectable pancreatic cancer is an incredibly aggressive disease.

What should the valuation of $IMRX be and where might it shake out in the competitive landscape are two excellent questions I will cover in the next post.3

The clinical benefit observed in this trial is backed up by biological data supporting the impact circadian rhythms have on response to immunotherapy. In this article, however, we will step through the data presented from this randomized trial and explore reasons it may not be a true benefit that we observe.

Subscribe now

Patients were randomized 1:1 to receive the initial four immunochemotherapy cycles before 3p in the early group or after 3p in the late group. 210 patients, 81% with metastatic NSCLC disease were randomized with similar baseline characteristics (per their word). Strikingly, median progression-free survival (PFS) was 13.2 months (95% CI 10.1-16.3) in the early group and 6.5 months (5.9-7.1) in the late group. After a median follow up of 18.9 months, the median overall survival (OS) not reached in the early group but was 17.8 months in the late group (see figure below). The ORR or objective response rate (complete response or partial response rate) was 75% vs. 56% in early versus late group.

They note PFS, OS, and ORR are consistently higher in the early group when adjusted for age, sex, performance status, tumor stage, histology, PD-L1 status, and immunotherapy agent. Hazard ratios (HRs) of PFS for each subgroup they analyzed largely favored the early group but the same was not true for OS.

Notably, females and never smokers have HRs for OS that considerably overlapped one indicating time of day did not matter much for these groups. Also, those with adenocarcinomas have HR for OS that significantly overlapped one or nearly touched one, respectively. Why might this be?

• Smoking: Individuals who do not smoke are associated with higher socioeconomic status [Garrett et al 2019; Leventhal et al 2019] . We know socioeconomic status matters for survival outcomes in cancer and can be quite pronounced. Socioeconomic status/financial stress/working jobs that have less flexibility and therefore having to come in later may influence dropout rates in the trial, hours of restorative sleep, baseline strength of their immune system, or even whether a patient accepted enrollment into the trial.

• Female vs Male: Immunoaging is more pronounced in males relative to females [Bartz et al 2020, Giefing-Kroll et al 2015, Taneja 2021] . Since females have generally more robust immune systems, they could be relatively insensitive to time of day of immunochemotherapy.

• Adeno vs squamous subtype: Chemotherapy regimen recommendations differ for adenocarcinomas and squamous NSCLCs per NCCN guidelines so we may be observing a chemotherapy efficacy difference. It is likely not the primary driver since trials investigating chemotherapy regimens in advanced NSCLC have a PFS/OS effect on the order of a few months, way shorter than observed here.

Put together, these interactions (subtype, socioeconomic status, smoking status, female/male immunoaging differences) if additive may account for much of the difference observed in the results.

Another post a few months ago profiled this conference presentation and had this to say:

But this paper was not a retrospective study of electronic health records, it was a randomized clinical trial, which is the gold standard. This means that we’ll be forced to immediately throw away our list of other obvious complaints against this paper. Yes, healthier patients may come in the morning more often, but randomization fixes that. Yes, patients with better support systems may come in the morning more often, but randomization fixes that. Yes, maybe morning nurses are fresher and more alert, but…well, randomization doesn’t fix evening nurse performance (which does dip during the night!), but I am inclined to believe the errors aren’t so high there as to cause this magnitude of a survival shift.

My counter points to this:

• There are a number of randomized trial whose interpretation shifted because of imbalances in baseline populations. ELITE postmenopausal women given estradiol had lower atherosclerotic events compared to placebo BUT they were found to have lower baseline coronary calcium scores. After adjusting for this, effect disappeared. Similar stories for EORTC22921 trial in rectal cancer, and BR.21 in NSCLC. With 210 patients, baseline imbalance could happen. We do not know if socioeconomic status, financial stress, or working fixed hour jobs with limited flexibility for time off are balanced between groups. A subgroup analysis should alleviate that concern.

• We see some important subgroups having no benefit in overall survival namely females, and never smokers.

• Adjustment for factors using a cox proportional hazards model assumes NON-INFORMATIVE censoring of patients in either arm. There is a lot more censoring especially in the first 12 months in the late group compared to the early group. So, if these patients had to drop out because they needed appointments that better fit their schedules, transporation needs, or other factors tied to socioeconomic status or financial reasons, we may be overestimating the benefit of time-of-day.

Lastly, I wanted to highlight a reason why all of these reasons I outlined before could be wrong. Grade 3+ adverse events were equal between groups including infections; mild leukopenia (grade 1-2) was higher in early time of day group. Despite higher rates of global reduction of white blood cells, after four cycles, CD8+ T cells comprised ~1.2% of peripheral white blood cells in the early group up from 1% at baseline whereas late treated patients had a decline to 0.9% from 1% after four cycles. The ratio of Activated:Exhausted circulating T cells were ~10x higher in the early group relative to late group after four cycles. This points to the possibility that there are more CD8+ T cells capable of killing tumor cells since the number of CD8+ T cells may have remained relatively stable (not completely sure because we do not have absolute count data) and a higher fraction of them were activated. It is unclear at present if these activated CD8+ T cells are able to recognize tumor antigen or penetrate into the tumor microenvironment.

Their inclusion criteria is shown below enrolls patients without detected driver genes in their cell-free DNA taken from venipuncture or tumor biopsy. Roughly 53% of patients with advanced NSCLC have no driver mutation (Ishida et al, 2024) so the treatment benefit to time-of-day would at maximum apply to 47% of advanced NSCLC patients. Thus, the narrowest reading of this conference presentation would be to try to schedule immunotherapy for advanced NSCLC driver negative male patients who smoke and have stable brain metastases before 3p.

If time-of-day matters a lot for immunotherapy, it should be incorporated into standard practice. The logistics and economics of infusion centers should adjust to this. But, before doing so, we need to wait for the randomized trial presented at ASCO to be published so we can look at key subgroups, and have at least one more follow up trial given the implications of cutting infusion center hours.

Personally, I think we do not study non-pharmacologic interventions enough. Examples of these studies include investigating the impact of exercise in colon cancer survival (exercise improves 8 year survival from 83% to 90%) or time-of-day of immunotherapy in lung cancer. In part, these studies are incentivized to not take place because pharmaceutical companies have no reason to do it. They want to fund things that lead to their drug being validated and approved. As Anirban Maitra points out, we need foundations or the government to fund such trials.

Appendix surveying some of the retrospective data on time-of-day impacting outcomes in immunotherapy:

Retrospective mRCC patients with early (before 4:30p) immunotherapy infusion had a median overall survival of 46.3 months compared to 41.7 months for the late infusion group (Dizman et al, 2023). Results were adjusted for age, gender, line of treatment, IMDC risk, and histologic subtype.

Another retrospective study considered time-of-day that patients stage III or IV non-small cell lung cancer were given immunochemotherapy. Patients who received the majority of their first four infusions before 11:30a had a median overall survival of 33 months compared to 19.5 months for the post-11:30a group. They adjusted for country, sex, age, WHO performance status, tumor stage, PD-L1 expression, number of metastatic sites, type of immunochemotherapy regimen administered. None of the studies I reviewed adjusted for socioeconomic status, financial stress, or lack of transportation.

Introduction

DNA Foundation Models (FMs) are large language models capable of 1) generating DNA sequences at scale that capture the grammar of DNA and 2) predicting diverse genomic properties (e.g., pathogenic variants) with minimal additional data or modeling. Before we jump in, let me tell you why you should care about DNA models.

Protein models like ESM3, AlphaFold3, Boltz2, and RFDiffusion generate, characterize, and design proteins thereby allowing us to make new drugs and discoveries. However, proteins compose just 2% of the human genome. What can we learn from the other 98% of the genome? If we can generate, predict, and characterize DNA sequences, what can we create?

Protein models primarily train on the protein data bank, a set of ~200,000 protein structures. In comparison, 490,000 whole genome sequences were deposited in the UK Biobank on August 6, 2025 alone. We have orders of magnitude more DNA sequencing data than protein sequence/structure data. The data advantage will grow even wider given that the cost of DNA sequencing and other molecular assays continues to drop. DNA FMs allow us to ingest this data and learn complex patterns.

This post is, in some ways, a response to Owl Posting’s ‘so-who-cares’ take on DNA models. DNA models have a lot of potential and it is early days both on the model building and application-side. Long story short, DNA models can be useful if coupled with high-quality domain-specific multi-modal data.

Subscribe now

Survey of DNA Models

Arc Institute’s Evo2 (Feb 2025) and DeepMind’s AlphaGenome (June 2025) are the frontier models. Both models are open source.

Evo2 is a 40B parameter 1M nucleotide context model that uses the StripedHyena2 (convolution and attention layers interwoven) architecture with nucleotide-level tokenization. Evo2 is trained on 9T DNA base pairs across all domains of life. Evo2 is able to generate sequences with specific epigenetic properties via inference-time search. Specifically, Evo2 was coupled with a guidance model (a prediction model that spits out a score given an input) to iterate and optimize the sequence to have certain properties (see figure below). Important caveat: we do not have experimental validation that the generated sequence actually creates the predicted chromatin accessibility.

AlphaGenome is a 450M parameter 1M nucleotide context model that uses an encoder-decoder architecture composed of convolution and transformer structures with single nucleotide-tokenization. It was trained on multimodal sequence data including RNA-seq, DNA sequences, and Hi-C genomic contact maps (see figure below). AlphaGenome predicts many genomic tracks across cell types in human and mice and is competitive or better than top task-specific models.

There are several other DNA FMs including JanusDNA (Fudan University), Caduceus (Cornell), NucleotideTransformer (InstaDeep), GENErator (Alibaba), DNA-BERT2 (Northwestern), and LucaOne (Alibaba). They make different modeling and data choices including k-mer tokenization, shorter context windows, different neural network architectures, and smaller training datasets with limited genetic diversity.

But, how do we generate DNA sequences from a DNA model? Three key ways. Autoregressive sampling (GPT-style next token prediction), diffusion sampling, or Dirichlet flow matching (DFM). DFM is a smooth diffusion process that allows a guidance model to steer all positions in the sequence. Autoregressive sampling is the standard (used by Evo2) though Dirichlet flow matching holds significant promise for generating sequences subject to constraints. DFM is faster and enforces global constraints but can be complex to train and less intuitive to grok than autoregressive sampling. We can work through these negatives to realize the advantages of DFMs. Dirichlet flow models are a natural choice for achieving guided sequence generation.

Technical areas for improvement

This is mostly a section of me asking questions and speculating. Feel free to skip, not super relevant to the rest of the post.

1. Data:

The quality of training data strongly influences the type of problems the model can solve, its performance on those tasks, and its ability to generalize. Open questions include: How much raw data can we acquire? What data should we collect to maximize the model’s ability to generalize to unseen states?

What other modalities of data should we acquire? What additional data should we collect to fine-tune a general model for our specific set of tasks? In short, it depends on the application. But in general, we can collect high-quality data for fine-tuning models by developing scalable assays that directly measure the molecular properties of interest. One example is the recently released Variant-EFFECTS assay developed by Eric Lander, Anshul Kundaje, and Jesse Engreitz, which quantifies how each bit of regulatory DNA influences gene expression. Other labs and companies have large amounts of data relevant to training DNA FMs including Octant Bio, Jay Shendure, and Pardis Sabeti.

2. Model building:

How do we capture long-range effects without forgetting small-scale patterns? One option is to scale context length. The context length of 1M base pairs is the longest range current DNA FMs can handle. If we can increase context length 1.7 orders of magnitude to 50M, the length of the smallest human chromosome, we can capture chromosome-scale interactions. They are rare but their effect sizes are often large (e.g., BCR-ABL1 fusions in leukemia). But, with increased context length comes memory, compute, training cost increases and the risk that the model will forget small-scale features. What new types of architectures allow for ultra-long-range pattern recognition while preserving short-range capabilities subject to cost, compute, and memory constraints? This is an open problem. For example, Evo2’s StripedHyena2 architecture offered a solution to this problem by using interwoven convolution and attention layers.

Caduceus, a DNA FM, implements reverse complement (RC) equivariance which models the symmetry of DNA’s structure and base pairing rules. It also implements bidirectional learning such that it ingests sequences left-to-right and right-to-left to extract patterns. Could RC-equivariance and bidirectional learning improve Evo2’s performance? What other biologically-informed architectures should we try to implement? These are open questions.

How can we map DNA FM embeddings to biological processes? Sparse autoencoders (SAEs) are one option. Interpretable biological signals using SAEs have been explored for protein models (ReticularAI), DNA FMs (Goodfire) and sc-RNAseq data (Markov Bio) with varying degrees of success. Some other interpretability methods for biological deep learning models include DeepLift and TF-MoDiSco.

3. Optimization environment:

A generative model coupled with a prediction model can iterate to create arbitrarily large DNA sequences with a specific set of functions if the set of functions can be translated into a score that can be optimized. Think Prime Intellect for reinforcement learning in biology.

4. Benchmarking:

Lastly, we need benchmarks to score DNA FMs without significant test-train leak. There are a few datasets people use including TraitGym, CAGI, Genomic Benchmarks, and BEND but each DNA FM paper uses a different set of datasets to measure performance. The field needs a consensus benchmark. Relatedly, there is no METR equivalent cataloging model performance in this space. There probably should be.

Applications

What can we do with DNA models?

A. Designing state-specific promoters for gene therapy and biological research

Regulatory DNA elements (ITR sequence, polyadenylation signal, promoter) and a therapeutic transgene packaged in Adeno-associated virus (AAV) capsids are a type of gene therapy approved to treat genetic diseases including retinal dystrophy (Spark Therapeutics), spinal muscular atrophy (Novartis), and Duchenne muscular dystrophy (Sarepta). Modifications of the each of the three regulatory elements in the AAV impair viability of the transgene in human tissue and alter transgene expression. Also, the DNA in the AAV capsid is limited to ~4700 bases due to space constraints in the capsid.

We currently design regulatory elements from rules (e.g., OnTarget), experimentally characterizing the search space of regulatory elements, or bioinformatically scanning databases hunting for a target tissue advantage. Can we engineer regulatory elements that allow the transgene to be expressed only in target tissue? Yes. We can leverage DNA FMs to create tissue-specific regulatory elements. AlphaGenome predicts cell-specific chromatin accessibility (ATAC-seq, DNAse-seq), promoter marks (histone modidifcations), and CAGE (transcription initiation). We can design an objective function based on a composite of these measures being activated in the target tissue and not activated in off-target tissue. Then, we can generate sequences perhaps through Evo2 or DFM and implement our guided generation to discover and rank candidate promoter sequences that are specific to a certain tissue type.

Efficiently designing tissue-specific promoter sequences enables AAV gene therapy, CAR-T therapy, DNA-nanoparticle gene therapy, ddRNAi drugs, and basic science work (lineage tracing, calcium physiology, cell ablation, tissue-specific disease models).

CAR-T promoter design matters a lot. It impacts anti-tumor effectiveness and side effects. Typically, always-on promoters are used to create predictable anti-tumor activity. But, always-on promoters can create significant side effects (i.e., cytokine release storm) when administered in the short term. In the long-term it can lead to T cell exhaustion which means that signaling gets shut down and cannot work to kill cancer cells anymore. We can design T cell-specific promoters with activity levels that are optimized to prevent exhaustion and severe cytokine release storm with DNA FMs. Typically, CAR-T cells are edited outside the patient’s body but recent efforts have shown that we can create CAR-T cells inside a patient’s body by giving a patient a lipid nanoparticle containing mRNA. The promoter design question becomes even more important for in vivo CAR-T therapies developed by companies like Capstan.

Looking forward, designing state-specific regulatory elements may be a new avenue for therapeutics. Instead of maximizing expression for liver tissue, we can maximize expression for hepatocellular carcinoma cells or hepatocytes with dysfunctional glucose regulation (i.e., Warburg effect).

There are several groups pursuing ML-based strategies for regulatory element design. Ryan Tewhey, Chris Reilly, and Pardis Sabeti published their deep learning model for designing tissue-specific regulatory elements across three cell lines in Nature (Oct 2024). Ginkgo Promoter-0 technology predicts promoter activity in cell types using Borzoi. Ginkgo gene therapy solutions offers AI-driven sequence optimization of cell-type specific promoters within 3 months (too long in my opinion) and licensing of their promoters. AskBio designs synthetic tissue-specific promoters using their own bioinformatics engine. Meira identifies tissue-specific mini-promoters using high-throughput experimental assays.

B. Discovering novel disease-causing targets through in silico mutagenesis

DNA FMs are a general platform that can identify disease causing variants. A clear example of the potential for DNA FMs to achieve this is demonstrated in a study published by DeepGenomics in 2020.

They used a deep learning DNA model to accurately describe the role of a missense mutation in Wilson’s disease. It was known that the mutation was associated with the disease but its mechanism was unclear given it had no effect on protein function in vitro. This study revealed that the mutation results in exon 6 skipping which enables the development of a new treatment for Wilson’s disease patients with this variant: steric blocking antisense oligonucleotides to restore splicing.

Also, DNA FMs can categorize missense mutations in proteins labelled as likely benign by AlphaMissense.

AlphaMissense classified 105139 mutations as likely benign. In ClinVar, 5108 of these were actually not benign. AUPRC is 0.18 for AlphaGenome on this very tough dataset. The prevalence of pathogenic variants is 4.8%. This is the expected performance if we randomly guess labels in this data. AlphaGenome does a lot better. I think the low AUPRC may not catch some people's eyes but I think its a great result. It shows an area (of many) where DNA models perform better than their protein counterpart

DNA FMs may have an advantage over protein models in predicting pathogenicity status of protein coding variants. With a multi-modal DNA FM, one can probe the effect a specific variant has across cell types and genomic tracks (e.g., expression, chromatin state, etc.).

Experimentally, scientists can make a specific change to a cell’s genome, observe its effect via molecular assays across cell types, compare their effect sizes and rank-order them by disease causing potential for further investigation in mice or other organisms. Issues with this approach include the cost and time of systematically probing the 3B base pair human genome and investigating the effect that cell-level changes like expression or chromatin state have on an organism’s phenotype. Consulting the scientific literature can aid in disentangling the cellular effect to phenotype linkage but is often sparse and hard to interpret. However, if we can prioritize the changes we wish to investigate (even loosely), we can then employ DNA FMs to quickly characterize their cell-level changes in silico and select a small set of variants for subsequent experimentation and de-risking. AstraZeneca fine-tuned a DNA FM to predict noncoding variant pathogenicity, so this topic is clearly top-of-mind for big pharma. The disruption of regulatory genomic elements (e.g., splice sites, promoters, enhancer, transcription factor binding sites) are understudied and DNA FMs allow us to probe into the effects of regulatory elements.

The process for choosing the changes to investigate by DNA FMs can be informed by clinicogenomic data. Two specific avenues for doing so are rare disease WGS data and human genetics (e.g., GWAS) studies. If a patient has a set of variants discovered by WGS and an observed phenotype (e.g., lab values, clinical syndrome, imaging), then we can investigate in silico, without assembling specific datasets, which one (or set of them) may lead to the phenotype. Companies that 1) sequence diverse populations or 2) assemble rare disease datasets are best positioned to leverage this technology.

(1) Sequencing diverse populations: We hope to emulate the process that yielded PCKS9 inhibitors for cholesterol management. Essentially, naturally occurring PCSK9 variants were observed to be associated with lower LDL and cardiovascular events in select populations. Then, mechanistic work was pursued in cell and animal studies. For example, VariantBio, a company working to sequence diverse individuals, could uncover future drug targets. Their downstream work could be aided by DNA FMs.

(2) Rare diseases: The undiagnosed diseases network (UDN) aggregates rare disease clinicogenomic patient data and seek to discover the genomic determinants of their conditions. They recently introduced a statistical method for prioritizing disease genes and discovering new diseases associations based on pooling WGS data across rare disease patients, finding recurrent mutations in genes, and grouping across pathways. UDN also works to deploy model organisms to characterize variant-phenotype relationships. Arcadia Science also works to develop new model organism systems and may be positioned to leverage DNA FMs to quickly test variant-phenotype relationships. Citizen Health is a rare disease company using AI to define trial endpoints and identify patient clusters to run trials. They have longitudinal clinical data (and potentially genomic data) that may enable variant to phenotype predictions. In silico assessment of variant to phenotype relationships also accelearte N-of-1 gene therapy treatments too.

C. Resolving variants of uncertain significance (VUS)

FDA-approved targeted therapies require patients to be ruled-in for eligibility based on if the patient has a pathogenic mutation in a predefined set of genes. Companion diagnostics offer tests to determine molecular eligibility and rule-in patients if they have a known pathogenic or likely pathogenic mutation. However, many patients who receive testing have only VUS and are generally not eligible for therapy. Also, as more patients receive WGS, more VUS will be reported. VUS reflect our scientific uncertainty about whether the variant causes the gene product to become dysfunctional and contribute to disease phenotype. DNA FMs can investigate VUS in silico to characterize their deleterious effects. What evidence do we have that this might be possible? We can look at the Evo2 and AlphaGenome papers for a base case.

Evo2 was able to accurately classify BRCA1 coding and noncoding SNVs as pathogenic/likely pathogenic or benign/likely benign but VUS were removed. PhyloP, a much simpler algorithm, performs nearly as well, which has contributed to skepticism about DNA FMs for variant classification.

AlphaGenome examined the effect of clinically observed non-coding variants near TAL1 and accurately predicted the mechanism of their functional impact in acute myeloid leukemia.

DNA FMs offer a computational way to bridge the variant-to-function gap.

Reclassifying VUS could materially impact sales of targeted therapies and health outcomes for cancer patients. In the US, HER2 negative breast cancer patients with germline pathogenic BRCA mutants are eligible for PARP inhibitors (PARPi). About 7% of patients receiving BRCA1/2 testing for breast cancer have VUS only. So, ~13,000 HER2- breast cancer patients have BRCA VUS each year whereas ~50,000 patients have BRCA pathogenic variants. Conservatively, if 5% of BRCA VUS patients are reclassified, 650 more patients are eligible. Assuming 180,000 new HER2- breast cancers in the US each year, $15k per month, 12 months of use on average, and 70% take-home after rebates, this translates to 81M per year in additional revenue if 5% are reclassified. PARPi is also indicated for prostate and ovarian cancers too. Reclassifying VUS at-scale could be a viable business model if value can be captured by the company doing the reclassification.

In this space, Velsera and QIAGEN QCI Intercept are software engines that serve labs and health systems to interpret variants. Myriad Genetics and Ambry Genetics develop clinical genetic tests and actively reclassify variants through RNA follow up, family studies, and surveillance. DNA FMs can reclassify at scale and therefore develop clinical tests that rule-in more patients for targeted therapy.

Myriad reported 25% of VUS were reclassified of which 9% were upgraded. Ambry reclassified a considerable number of VUS to likely pathogenic / pathogenic (see figure below, small sample size limits generalizability). These reclassifications are reported to ClinVar and then get used by other groups.

D. Biosecurity of generated DNA sequences

Given that bad faith actors can now generate novel DNA sequences with toxic or pathogenic features, we must develop technology to identify such sequences. In fact, Le Cong and Mengdi Wang wrote a paper showing how DNA language models can be used to generate problematic sequences. We need compliance software that can detect novel pathogenic sequences. DNA FMs are a solution to this problem.

Shelby Newsad and CompoundVC have written about the problem of DNA biosecurity:

I agree with their premise.

DNA synthesis companies (e.g., TWIST, IDT, Thermo Fisher, Ginkgo, Tierra, benchtop DNA printers) typically follow government guidance currently composed of rules-based sequence checkers and customer watchlists. This is not a good solution. We miss a lot of pathogenic sequences. A fine-tuned DNA FM can flag problematic sequences. Or DNA FMs can have enhanced safety alignments and tracing mechanisms. This is preferable over rules-based engines like UltraSEQ (TWIST is a customer) given that novel sequences with specific properties can now be generated at scale. Also, the current administration has expressed interest in using AI for biosecurity and we may be able to make progress on this front quickly at a national level.

I have a few more applications of DNA FMs that I will defer to part 2 of this series. If I could summarize this post in one brief sentence, it would be: DNA models can be useful if coupled with high-quality domain-specific multi-modal data.

Key catalysts:

1. ESSENTIAL3 Ph3 trial of Ulixacaltamide in Essential Tremor (Q3 2025)

2. New ASO drug candidates 2025. (PCDH19, SYNGAP1, SCN2A LoF)

3. POWER1 complete 2H 2025, readout 2-3 months later

4. EMBOLD2 (1H 2026)

5. EMBRAVE (ASO for SCN2A) Topline 1H 2026

6. EMERALD broad DEE (enrollment done 2026)

The three bolded events are the ones that matter most.

We’ll walk through each of these events.

1. Essential3 (Ph3) trial of ulixacaltamide (UX) in Essential Tremor

   If Essential3 results are good, UX becomes a new treatment for essential tremor. There are 5-10 million patients with essential tremor of which 30-50% are typically treatment-refractory. Current standard of care is primadone and propanalol both priced on the order of $10/month. That’s why it makes sense for UX to be priced a little higher but also as a second-line or refractory treatment option for patients whose conditions are not improved significantly on the cheaper first line therapies. So, conservatively, that’s 1.5M (0.3 x 5M) total patients ulixacaltamide can be useful for. At $100/month at 50% of 1.5M patients, peak yearly sales is 900M.

   The story with UX has been one of not statistically significant endpoints, discussion with FDA on endpoint measurement, and delays in reporting results. In March 2023, Essential1 results were presented.

   

   Enrolled patients could continue propanolol at stable dose during trial, and could not have had prior surgical treatment or focused ultrasound treatment for ET. Per ITT, 78 patients on UX (13 d/c - AEs for 9, withdrew consent 2, lack of efficacy 2) and 38 on placebo (4 d/c).

Primary endpoint was baseline to day 56 mADL score measured as such:

mADL had p=0.126 and per Praxis, no dose related difference in efficacy between 60 and 100 mg. Roughly 38 patients in UX 60 mg and 38 in UX 100 mg so if efficacy difference not seen its probably real given sample size is decent. The interesting part is that UX was uptitrated in dose from 5 to 60/100 depending on treatment arm protocol which led to patients receiving only 1 month of treatment before mADL measurement.

After these results, Praxis discussed with FDA on endpoints and the following was recommended: TETRAS-ADL can be used as an acceptable endpoint. But, exclude item 12. And rescore accordingly:

Basically, FDA is saying use mADL but get rid of PS6 and 7. Here are the results:

Looks better. And, there is some deepening of response over time but interestingly the response does not appear dose-dependent. Note that Day 28-56 is when patients got at least 60 mg of UX in Regimen 1 (up to 100 mg) and Regimen 2 (max 60 mg).

The dose-clinical efficacy relationship of UX appears to be sigmoidal with the point of plateau being at least 60 mg if not earlier. Praxis learned that they need to control for the presence of ET patients with intention tremor given that patients with intention tremor may not respond well to UX. Below is the mADL changes in patients without intention tremor. Note that UX effect size is larger than the -2.69 observed when including all patients.

So that was March 2023. They initiated Essential3 that has 2 studies. Originally Praxis messaging was that the UX 60mg arm would be treated for 6 weeks but as we see here from their Aug 2025 slide deck, they extended it to 12 weeks.

In February 2025, Praxis reported “Based on the predefined decision framework for Study 1, the [data monitoring committee] has recommended that the study be stopped for futility, due to the results being unlikely to meet the primary efficacy endpoint under the parameters set by the statistical model. The committee also indicated that some underlying assumptions of the statistical model might have influenced this outcome and encouraged Praxis to explore alternative analysis methods”. The primary endpoint was the mADL11 that they reported statistically significant results for in Essential1. Further, “Given the advanced state of enrollment for both Study 1 and Study 2 in the Essential3 program, and in the context of the advice received by the IDMC, Praxis has decided to continue both studies to completion, with topline results expected in the third quarter of 2025. The decision about whether the data supports the submission of an NDA will be made after analyzing the final results for Study 1 and Study 2”. Their CEO said “We are disappointed with and surprised by the outcome of the interim analysis for Study 1. Following the advice of the committee, we will explore different analysis methods for the final dataset, which is expected in the third quarter of 2025”.

Almost surely, the results of this trial will NOT be good. The stock dropped ~55% on the day these comments were released. Since then the stock has hovered around this. I think the market has priced in this failure.

RARE scenario. I don’t see this happening but just wanted to lay it out. If UX has minor AEs for the most part (as we saw in Essential1) and given the current FDA’s wantingness to allow for patient choice in diseases with high unmet need, there is a world in which UX can be approved for treatment-refractory essential tremor. Patients with severe treatment-refractory essential tremor are affected by their tremor in all parts of their lives and would likely be willing to try a relatively safe new drug if it could help them do some of their acts of daily living (ADL) more independently. It says something that Praxis has 200k+ people with Essential Tremor that expressed interest in joining the study.

2. Some new drug candidates to be announced in 2025 year end for PCDH19, SYNGAP1, and SCN2A LoF disease. All ultra-rare diseases. No other comments here. Will need to initiate trials after announcement. Will take 2-3 years to get to approval probably.

3. POWER1 (n=250) expected to complete 11/2025 and study completion 12/2025 per clinicaltrials.gov and company guidance of 2H 2025. Readout to be expected 1H 2026 by March. POWER1 is the placebo-controlled version of the open-label no-placebo RADIANT (n=37) trial that reported vormatrigine in focal onset seizures.

56.3% overall median seizure reduction at 8 weeks, and 60% of patients with over 50% seizure frequency reduction. 22% of patients seizure free in past 28 days. 56% with >12 seizures / month achieved 50+% reduction.

Safety profile fairly low, although with a decent discontinuation rate.

Investigators had the option to reduce the dose of the background medication to manage AEs and when this was done instead of d/c vormatrigine, 6/6 patients stayed on study. They know to advise investigators to do this and to counsel on what AEs to expect. Serious AE rate was 4.9% compared to 2.6% of XEN1101. What were these AEs?

The main comparison for vormatrigine is XEN1101 (aka azetukalner). XEN1101 Ph2b results were reported in JAMA neurology in Nov 2023. Since then, 2 Phase3 trials (X-TOLE2, X-TOLE3) are underway for focal onset seizures (enrolling 360 patients each). After completion of one of these trials, Xenon plans to submit NDA to FDA. In Oct 2024, Xenon announced it anticipated completion of patient enrollment in X-TOLE2 by end of 2024. The treatment period is only 12 weeks so results should have been reported by Q2 2025 at the latest. Now on clinicaltrials.gov, they are expected to complete 2026 and are still recruiting. Company guidance says to expect topline results early 2026.

Key differences: Xenon1101 was given with food, Vormatrigine is not. 25 mg XEN1101 had ~55% of patients have >50% seizure reduction. Vormat had 60%. Both were treated for 8 weeks.

Notably: XEN1101 is primarily metabolized by cytochrome P450 3A4 isozyme (CYP3A4) without substantial involvement of other CYPs, and plasma levels of XEN1101 may decrease in the presence of CYP3A4 inducers. Vormatrigine does not have this problem. CYP3A4 is inhibited by antifungals, some antibiotics, non-DHP CCBs (verapamil, dilt) and induced by a number of other drugs. More importantly, most patients that would receive XEN1101 or vormat are women with ~50% below 40 and potentially taking hormonal contraception. Taking hormonal contraception with XEN1101 may lead to reduced efficacy - a big concern.

So, Xene wants to submit NDA after topline X-TOLE2 expected in early 2026 (delayed from early 2025). Praxis could submit NDA after POWER1 in early 2026 (if they execute on time, we will know more after their presentation at epilepsy congress Aug 30, 2025). Issue is 250 patients in POWER1 but 260 in X-TOLE2. Praxis may need POWER2 which will finish Q4 2026 and readout likely early 2027. At worse, Praxis has a 1 year delay to market.

The Cenobamate question:

Above are the ASMs that patients on XEN1101 took per a Xenon poster. No cenobamate in XEN1101 trial. Per Aug call with Praxis management, >30% of n = 37 patients in RADIANT vormatrigine trial were on cenobamate and 55% had a >= 50% reduction in seizures from baseline. So vormatrigine works ON TOP of cenobamate. Does xenon? I’m not sure, I don’t have data on it. Cenobamate enhances GABA-A and inhibits persistent sodium current via blocking of voltage-gated Na channels. Perhaps cenobamate + Xen1101 = AEs that are too high or Xen1101 does not add much utility on top of cenobamate. We’ll see with X-TOLE2. Let’s look at mouse data from Xenon in 2021.

Cenobamate 5 mpk + XEN1101 1 mpk has a similar effect to cenobamate 10 mpk (1 mpk is roughly 5 mg for humans). There is some additivity but does that additivity get removed when actual doses are used? Is there a saturating clinical effect of this combination? 200 mg/day of cenobamate is the standard dose in humans after titration which is roughly the 30 mpk value in mice. The 30 mpk cenobamate is a near perfect eliminator of seizing in mice. So, adding XEN1101 to 30mpk won’t reveal anything useful since we’ve already squashed the problem in mice. This is a limitation of using mice models. We can cure lots of things in mice but it doesn’t translate to humans.

Praxis plans to initiate POWER3 in 1H 2026 to look at vormatrigine as monotherapy which was announced in Aug 2025 RADIANT call. It signals confidence that the management team wants to spend money to evaluate vormatrigine as monotherapy. POWER2 (n=400) with three different dose levels of vormatrigine and placebo for 12 weeks. Confirmatory trial of POWER1. POWER2 expected to complete enrollment in 2H 2026. Finally, it should be considered that patients with FOS / generalized seizures who are eligible to receive vormat or XEN1101 will need multiple treatment options. The same combination regimen will not work for everyone, and patients cycle through treatments. XEN1101 could work WITH vormat given different mechanisms of action to enhance their effect. Both could meaningfully help patients. I do not see it as a winner take all.

4. EMBOLD2 (n=80) is a continuation of EMBOLD1 (n-16) trial to confirm relutrigine’s efficacy in treating children with SCN2A GoF or SCN8A DEE. Topline results of EMBOLD2 expected 1H 2026. Study (trial design, endpoints, target patient population) is the same except larger patient size and relutrigine dose upped from 0.5 mg/kg/d to 1 mg/kg/d.

   

   EMBOLD1 results summarized above were great. For more, see here. EMBOLD2 should be a success.

5. EMBRAVE Part A (Ph 1/2) n=8 of ASO elsunersen in SCN2A GoF patients. Topline results expected 1H 2026.
   

   Relutrigine controls symptoms, but ASO can modify disease cause pathology up-front by preventing translation of some of SCN2A mRNA which normalizes ion flow and reduces seizure activity. They are complementary treatments for SNC2A GoF patients. A patient who received elsunersen under emergency circumstances went from 15+ seizures per day to 5 seizures per day (Nature Medicine).

   n = 4 EMBRAVE Part 1 looks good but small sample size and kind of hard to interpret reported summary metrics.

   

   EMBRAVE Part A has a placebo arm and topline results expected 1H 2026.

   

   EMBRAVE3 with a similar structure (notably 1:1 randomization) has 3 cohorts that are expected to initiate mid-2025. Cohort 1 (n=40) same as Part A patient population and structure. Cohort 2 (n = 5) and 3 (n = 5) will not have a placebo and will enroll 1-2 years old and 0-1 years old patients, respectively. Cohort 1 will support Part A in registration. Cohort 2 and 3 will allow for expansion to patients at birth.

6. EMERALD trial investigates relutrigine in broad DEE patients (n = 160) 1:1 randomized to relutrigine or placebo. Primary endpoint is change in baseline in monthly motor seizure frequency. No titration, once daily dosing. Completion of enrollment expected in 2026.

Praxis has a number of presentations and posters at the epilepsy congress Aug 30-Sep5, 2025 on Power1 (interim results or just study design?), updates of elsunersen in emergency use, updates from Ph1 trial in vormatrigine (more safety and tolerability data to help understand AE profile better), RADIANT trial results, EMERALD (study design), EMBOLD results (possibly reporting EMBOLD2 results)

So, where does this leave us? I expect relutrigine to be approved for SCN2A and SCN8A + receive extension for broad DEE treatment. If we believe management’s slides, this is a 3B US market opportunity. I expect vormatrigine to be approved for focal onset seizures too. This is a 2.5B US market opportunity per management. It’s a little early to tell if elsunersen will work but it mechanistically makes sense and early results + case study seem to really help patients. Assign that 500M US market opportunity per management slides. Let’s assume UX is worth nothing. I have not done the math to size the markets myself but this is something to do in the future. Praxis is currently valued at 1B dollars. I see 2-3x upside from here in the next 12-18 months based on execution of POWER1, EMBOLD2, and EMERALD. Formal valuation to follow.

Cash runway to 2028 per management. 446M in cash as of June 30, 2025. They burned 26M between March 31 and June 30, 2025 (1 quarter). So that’s 4 years of runway making their end of 2028 runway a good estimate. Hedge funds have been net buyers in the last 6 quarters and insiders have not been selling.

This is a living document of my thoughts on Praxis, subject to change, and formal valuations to come too.

Cardiff has ~113 patients across 6 arms. This is ~19 patients per arm. But, we can collapse the chemotherapy groups (FOLFOX and FOLFIRI) into one. They are relatively interchangeable. This means 38 patients per group (chemo + VEGFi + onvansertib 20 mg, chemo + VEGFi + onvansertib + 30 mg, chemo + VEGFi).

They released some interim data from the current study which we will cover later. First, let us consider their 2024 JCO paper describing onvansertib + chemo + VEGFi as 2L treatment. They treated 53 KRAS mutant metastatic CRC (mCRC) patients with onvansertib (PLK1i 15 mg/m2 on days 1-5 and 15-19) + FOLFIRI + VEGFi. These 53 patients had previously received at least 6 weeks of oxaliplatin and fluoropyrimidine, with or without bevacizumab, and either progressed within 6 months of treatment or shown intolerance to oxaliplatin. Functionally, these patients were resistant to their initial therapy. The overall response rate (ORR) was 26.4% for the 53 patients. Not great. Then, Cardiff looked post-hoc at the subset of patients who were VEGFi-naive. In the study, 13 patients received only prior chemo and 40 patients received VEGFi and chemo. Why were these 13 RAS-mutant mCRC patients not given VEGFi? Perhaps the RAS mutation was not detected at that time, uncontrolled hypertension, recent major surgery, or their disease was not metastatic. Something changed to rule them back into VEGFi because clearly these patients were eligible for VEGFi when Cardiff was recruiting. Within prior chemo-only patients, 10 of 13 had > 30% reduction in target lesion size compared to baseline (an objective response by RECIST) when treated with VEGFi+PLKi+chemo. Only 4 of 40 prior VEGFi exposed patients responded to the new treatment. These patients were already resistant to VEGFi so additional VEGFi was not going to do anything. The addition of PLK1i did not add much value here. Whatever mechanism makes a patient in Cardiff’s target population resistant to VEGFi also renders PLK1i ineffective. The hypothesis is that VEGFi and PLK1i given at the same time acts synergistically to produce deep and significant responses. This idea is similar to other dual blockade ideas like BRAFi and MEKi.

Cardiff then surmised that since 10 of 13 responded when given VEGFi for the first time and PLK1i, their treatment should be used in combination with VEGFi as a first line treatment. But, there was a group of patients who never qualified for the trial - those that responded to VEGFi and chemo. What is their ORR? This is equivalent to us asking what the ORR for first line VEGFi and chemo is for RAS-mutant mCRC patients? This informs the standard of comparison onvansertib will have to beat.

What is the response of chemo + VEGFi in RAS mt mCRC 1L? The BECOME trial, which enrolled only patients with RAS mutant, unresectable, liver-limited metastatic colorectal cancer, reported an objective response rate of 54.5% for bevacizumab plus mFOLFOX6, compared to 36.7% for mFOLFOX6 alone. Similarly, a randomized trial by Avallone et al. found an objective response rate of ~53% in RAS mutant subgroups treated with oxaliplatin-based chemotherapy plus bevacizumab. The FIRE-3 trial when subset to RAS-mutant mCRC had an ORR of 48%. So, we can expect a low-50s response rate for the control group in the upcoming readout. Will onvansertib beat is such that its confidence interval largely does not overlap this standard of care?

I do not think prior chemo in RAS mutant mCRC takes away the opportunity for a patient to have a objective response to VEGFi + PLK1i + chemo. So, the experimental arm in the upcoming readout should function much like the prior chemo only treatment group in their JCO reported results. ORR was 77% (CI 46% - 95%) and PFS was 15 months. If 8 of the 13 (so two less than observed) had an objective response then the ORR is about equal to that of 1L chemo + VEGFi. The issue with objective responses is that a patient can have a 30% target lesion decrease but recur shortly thereafter and it still counts. Progression-free survival or better yet, overall survival is a better basis for comparison. More importantly, in the KM curve of PFS, 6 patients in the bev-naive group were censored < 12 months from start date. 8 were censored in total. The effective sample size is 5 which makes it near impossible to tell what is going on with the actual PFS. In comparison only 8 of the 40 were censored in the prior VEGFi group. This is a more stable estimate.

Now, let us consider their Nov 26, 2024 update of the trial we will be getting results from on July 29. Results shown below.

3 of 9 responded in the standard of care. Consider that there is a patient with 29% change in the standard of care arm which will hit 30% in the next 6 months - the real ORR is 4 of 9 in this readout. This is well below our expected mid-50s so expect that to change when more patients are enrolled. Now, consider the patient in the 30 mg group with 100% response. This smells like the patient in the 2024 JCO study who enrolled in the expanded access program to continue receiving treatment and likely is an outlier relative to the other data observed in the paper. So perhaps the true ORR in the 30 mg group in the interim data presented is 6 of 11 (ORR 54%). This is in line with expected standard of comparison. And consider that the patient with ~5 months of follow up at just barely over 30% decrease could have easily flipped either way making the ORR 45%.

Let us look at this a different way. Assume onvansertib 30 mg treated group ORR is 64% like reported in the data above. Assume standard of care is on the lower end of what is shown in previous RCTs at 52%. We have 38 patients in the 30 mg group and 38 patients in the standard of care. We can use an inverse beta with alpha of 0.05 to compute confidence intervals for the ORR. So, the ORR CI for 30 mg is 48-80% and control is 33-67%. The confidence intervals have lots of overlap.

Dosage in 2024 JCO study was 15 mg/m2. At 20 mg, the assumed body surface area (BSA) is 1.3 and at 30 mg, the assumed body surface area is 2. Average BSA for men is 1.9-2 and women is 1.6-1.7. So, the JCO dose is closer to 30 mg than 20 mg. We can expect the 30 mg to perform like the 15 mg/m2 would. The pharmacokinetics have also been called into question. The pulse dosing and lower dose are probably there to avoid myelosuppression safety concerns noted with other PLK1i drugs (which ended up killing them). So, Cardiff is sacrificing efficacy here to avoid bad myelosuppression.

Cardiff postponed the release of the data we are about to see. Why did they prolong releasing their data and switch their CMO in between? Were they looking for more data to come in to see if they can show some signal? Some people have claimed that the new CMO is a good thing. I interpret it as having no bearing on the underlying trial. Sure, he has experience in a positive phase 3 trial for mCRC but the previous CMO was no slouch. I could even argue that the new CMO is a negative signal. Long story short, two companies had a new CMO appointed 2-3 months before a phase 2 or 3 readout (TARA and UBX). TARA fell 19% after readout and UBX fell 33% after readout. Opthea (OPT) hired a new CEO 6 months before readout and their Ph3 trial failed and the company went to zero.

Conclusion

Onvansertib will be at best a very thin additional benefit to 1L chemo + VEGFi. With 38 patients, the confidence interval for ORR will surely overlap considerably with standard of care. The JCO paper that led to the new VEGFi + PLK1i synergy hypothesis in 1L RAS mutant mCRC is based on minimal clinical evidence (e.g., 5 patients constructed the PFS curve and one patient had a grade 4 colon perf) and a few wet lab/bioinformatic experiments. There are questions on the pharmacokinetics too. Net-net in the long term I do not see onvansertib in the oral pulse-dose 1L setting for RAS mutant mCRC making meaningful improvements to patient outcomes. As always, open to comments, suggestions, and where I might have reasoned incorrectly. I am happy to be wrong. If onvansertib provides benefit to patients and improves median PFS and median OS, that will be great for patients and I am all for it. Better treatment options are always a good thing.

Share

We are going to dive into the landmark NEJM paper by Kiran Musunuru and team. From this, we will understand the science behind base editing, what goes into making a successful gene editing product, think about platforming gene editing treatments, and consider open clinical and scientific questions in this space.

Baby KJ was born with carbamoyl phosphate synthetase 1 (CPS1) deficiency, a rare genetic disease. 1 in 1.3 million are affected with this disease and 50% of those affected die in early infancy. CPS1 deficiency (CPS1D) prevents the urea cycle from making urea and offloading ammonia, a waste product generated from breaking down proteins (see figure below). This process primarily occurs in the liver and causes hyperammonemic crises (temperature instability, respiratory distress, seizures, coma) and irreversible neurologic injury.

Patients with CPS1D often end up with irreversible organ damage or die before they grow large enough for liver transplantation. So, how did this team of scientists and clinicians treat KJ?

To treat something, you first have to recognize that it exists. They noticed KJ had an ammonia level > 1000 μmol/L (normal in newborn is 37-118 μmol) which suggested that KJ had a urea cycle problem. Blood profiling of amino acids suggested a proximal urea cycle defect. Subsequent targeted sequencing identified two variants C10003T and G2140T in the CPS1 gene that caused a Q335X and E714X mutation, respectively. X signifies a stop signal that prevents further translation of the RNA sequence. The C to T mutation occurred on the paternal allele and the G to T occurred on the maternal allele, knocking out both copies of the CPS1 enzyme. The key to detection of KJ’s condition was the pediatricians flagging a high ammonia and then performing targeted sequencing. Only a few states (Florida, Maine, Massachusetts, Mississippi, New Hampshire, Rhode Island and Vermont) include CPS1 in the newborn screening panels. Even still, these screens cannot reliably identify all cases of this disorder. If we sequence (with enough resolution) each child and can turn around results in less than a day, we can catch many such rare genetic diseases. In a 2023 JAMA pediatrics study, of the 297 variants found on either WGS or targeted sequenced in hospitalized infants with a suspected genetic disease, 201 of them were due to unique single nucleotide substitutions. Assessing clinical significance between mutation and phenotype can be challenging when the connection is not as straight forward and validated like CPS1D. This remains an open problem2. However, this establishes the precedent that the majority of genetic diseases, if identified, can be treated with the exact technology set forth in this paper.

Subscribe now

From sequencing and biochemical labs, the team realized that KJ had effectively zero CPS1 activity. Based on mouse models, we know that if CPS1 activity is back to at least 10% of normal we can restore plasma ammonia and amino acid levels to within normal limits. Clinical observations in humans confirms that partial activity around 10% is sufficient. But, where does urea cycle activity take place? Which region of the body do we need to target? Almost all urea cycle activity happens in hepatocytes, primarily in periportal hepatocytes (region 1 in below figure) but also in perivenous hepatocytes (region 3 in below figure).

Blood from the branches of the portal vein and hepatic artery flows as a mixture through the sinusoids to the central vein. During this process, metabolism and re-composition of blood occurs (see figure below for high-level breakdown of processes including urea cycle activity). Our edit system ideally must selectively reach hepatocytes, preferably periportal hepatocytes.

Now, we know the exact genetic alteration that caused the urea cycle defect and where we need to edit to fix it. What do we fix? There are two alterations (E714X or Q335X) and fixing either one would suffice given we only need 15% restoration of function. We have adenine base editors (ABEs) and cytosine base editors (CBEs) which turn A —> G and C —> T, respectively. Consider that the E714X change occurred because a G flipped to a T. So, to fix it we need to switch a T to a G, which we cannot do with current base editors (prime editing could make this change, but that was not the chosen technology3). The opposite strand of the E714X mutation experienced a C —> A which we need to restore to a C, which we also cannot do. That leaves us with fixing the Q335X change (C —> T, opposite strand G —> A). With an ABE, we can flip the A —> G and restore the maternal allele.

Asimov Press wrote up a good explainer of genome editors and below is their image describing base editing. The NEJM paper used an ABE so mentally swap the cytosine deaminase for an adenine deaminase.

Long story short, the process can be boiled down to:

1) Two mRNA molecules enter the cell. One molecule encode cas9 and adenine deaminase and the other is the synthetic guide RNA

2) Cas9 enzyme binds the protospacer adjacent motif (PAM) found on the DNA strand

3) guide RNA binds complimentary region

The specificity of edits are tied to steps 2 and 3. The 3D structure of the region around the PAM matters in cas9 binding. For example, if the genomic region is tightly condensed or winding of the genome creates sterically unfavorable binding chemistry the cas9 will not bind well.

4) Adenine deaminase acts on exposed strand (created by binding described in 2-3 unwinding DNA at target site) turning adenine to inosine (which acts like a guanine).

5) Cas9 nickase breaks the opposite strand to what was just transformed into an inosine. This single-strand break forces the cell’s own mismatch repair process to replace this broken base with cytosine, the complimentary base to inosine.

The authors used the HuH7 cell line, a liver cancer cell line, and introduced into these cells a cassette with Q335X, E714X, and two positive controls via a lentiviral vector. A 100 base pair region centered around each variant was transduced. (The authors did not want to extract hepatocytes from KJ given the excess risk it posed to the patient, though primary hepatocytes from KJ would have been the ideal sample to use.) Now that the model system was set up, they designed seven different guide RNA’s tiling the Q335X site such that the variant base to be correct occupied different spots in the gRNA and unique distances from the PAM region.

Then used different Cas9 enzymes (SpG, SpRY) that differ primarily in their PAM preferences. SpG recognizes NGN PAM structures where N is any base, and SpRY is PAM-less. They also tested different ABEs - ABE8e [broadest edit window], ABE8.20 [intermediate] and ABE8.8 [narrowest]. The edit window is the distance from the distal end of the PAM that the ABE can chemically modify. All together, they tested different gRNA, ABE, and Cas9 combinations for the percentage of aligned sequencing reads with the edit.

SpG-AbE8e-gRNA8 had highest corrective editing efficiency for Q335X. They noticed that PAM associated with gRNA8 was NGC (see above figure of tiled gRNA’s schematic) so they performed a secondary screen with versions of Cas9 that prefer NGC PAM’s to boost specificity since SpG prefers a very generic NGN PAM. The secondary screen consisted of SpG, A1, A2, and A3 enzymes where A1-A3 each had their own set of mutations introduced (results shown below).

The left-most bar is the editing level for the reference control condition. Notice that A1-A3 offer gains in edit percentage over the base system (second from left bar). Because ABE8e has been reported to have gRNA-independent off-target editing, the ABE8e-V106W variant, which largely eliminates this off-target editing of ABE8e was used in a final transfection screen.

The A1 system had the best edit percentages when the V106W variant was introduced. Therefore, the A1-ABE8e-V106W/gRNA8 system was selected. The final drug product k-abe is a gRNA (20 nt spacer corresponding to protospacer DNA sequence matching CPS1 gene spanning Q335X + 80 nt tracrRNA domain that creates binding handle for nickase) + mRNA (encoding ABE8e-V106W-SpCas9D10A with a nuclear localization signal [NLS]) packaged in a lipid nanoparticle (LNP). The LNP design was taken from an earlier study that performed base editing of PCSK9 to lower cholesterol in primates. Acuitas therapeutics made this LNP from a mixture of the lipid components ALC-0307, PEG-lipid ALC-0159, 1,2-distearoyl-sn-glycero-3-phosphocholine, and cholesterol.

We previously established that the LNP must deliver the payload to periportal hepatocytes. The LNP is being infused intravenously and quickly binds ApoE in the bloodstream (Acuitas Therapeutics). ApoE-LNP can now bind LDL-R, CLDL-R, and ApoE-R2. ApoE binding affinity to its receptors is highest when ApoE is bound to lipid. LDL-R is expressed in different tissues including high expression in Leydig cells of the testes, adrenal gland, lung, stomach, duodenum, colon, liver, and kidney. Presumably the LNP is delivered to these tissues where it is endocytosed and subsequently protonated (intracellular vesicles are acidic) allowing for release of RNA payload into the cytosol. There is no periportal hepatocyte specific targeting agent yet the ApoE mechanism allows for adequate tissue specific targeting. Ideally, there would be less or no distribution to other organs. A well-designed LNP can ensure that the RNA product only touches the cells of interest. Recently, SORT LNPs and DUAL SORT LNPs and targeted LNPs have been developed allowing us finer control of drug distribution within the body. These are untested in trials but hold promise. We can predict protein structure well but cannot do the same for LNPs. Can we model LNP structure, tissue selectively, and intracellular protonation/payload release chemistry? Then, we can tweak lipid concentrations and component mixtures and understand their effects computationally. This could rapidly accelerate our ability to design better LNPs. End of tangent.

Before testing our drug product in animals, we need to understand k-abe’s ability to edit in an endogenous genomic context. The lentiviral cassette was randomly inserted into the HuH7 genome and thus may not reflect the editing activity in the endogenous gene locus. Reasons for edit activity differences include chromatin state differences and the cassette integrating into high transcription activity regions. To address this, they tested gRNAs that were complimentary to a region in the cassette and a region outside the cassette in the endogenous environment.

Lentiviral editing was 1.25 times greater than endogenous editing. Importantly, the rank order of gRNA performance was preserved between lentiviral and endogenous environments. Lentiviral edit activity is expected to overestimate actual edit activity by ~25% but it is a good proxy.

HuH-7 lentivirus-transduced Q335X cells tested with a clinical batch of k-abe demonstrated a maximum proportion of desired editing outcome of 67% with an EC50 of 17 ng/mL.

They tested a toxicology batch of k-abe on Rosa26-Q335X mice4 with two mice per dose and eight liver samples per mouse. At 3 mg/kg, there was 42% corrective editing of Q335X, 31% at 1 mg/kg, and 4% at 0.1 mg/kg (shown below). All observed bystander editing resulted in synonymous variants. The mice were treated at 1 month of age and were evaluated on necropsy several days later. In F1 CPS1-Q335X knock-in mice (endogenous edit of KJ variant then crossed with normal wild-type mouse to ensure 1 wild type and 1 mutant allele at CPS1 Q335), modestly higher editing rates were observed (52% at 3 mg/kg and 35% at 1 mg/kg) in line with the difference between lentiviral and endogenous cell line edits.

We know that k-abe makes the edit we want, but what others edits does it make? The BE-Hive machine learning model predicted only synonymous bystander edits that occurred within 5 base pairs of the target site. These predictions held true for the edit patterns seen in Huh-7 cells (only two non-synonymous edits totally 2.7% of reads sequenced).

Sequenced results from Q335X CPS1 deficient mice treated with k-abe revealed the same pattern of edits. In fact, the distribution of reads for each variant edit was exactly the same as the cell line. I am surprised that the distribution is exactly the same. See line by line exact number of reads for each variant in cell line work in Fig. S8 (above) and mouse work in Fig. S10 (below). Could the figure have been accidentally duplicated or perhaps I’m interpreting it wrong? Or maybe the mouse read distribution of variants is exactly that of cell lines. Should we not expect some variability on the number of reads assigned to each variant?

Let us pause here for a second to comment on an appropriate dose for baby KJ. The mouse studies showed more than 10% whole liver editing for 1 and 3 mg/kg implying that both could clinically resolve our CPS1 deficiency issue given the prior literature. Also, the authors note that 1) the LNP system was not designed for mice or cells but for primates, and 2) CPS1 is expressed in periportal hepatocytes preferentially so whole liver metrics may under-represent true edit efficacy. So, the authors chose to let safety guide the dose. More on this in a bit, but first we will consider cellular safety or off-target effects of our drug.

k-abe can, of course, make edits anywhere in the genome where the gRNA and cas9-abe complex find it chemically favorable. And, we know the LNP is going to deliver k-abe to many different tissue types. How do we assess off-target effects of k-abe? Assessing off-target effects is a regulatory requirement but the extent to which one must do it is not established. One CAR T product assessed only 20 off-target sites whereas others have evaluated thousands of sites. A summary of the k-abe off-target workflow is shown below.

What are each of these methods?5

• CHANGE-Seq: Take purified genomic DNA (in this case from KJ’s father because they could not draw enough genomic DNA from patient), use base editor, do some chemistry to tag any edit, sequence and align to identify mutated sites. Highly sensitive method for identifying mutated sites.

• ONE-Seq: Synthesize short DNA strands (~300 bp) on beads including common SNPs and incubated with base editor and sequence.

• UNCOVER-Seq: Add selected cas9 (without selected ABE) with gRNA and dsODN tags to HuH7 cells (primary hepatocyte preferred but did not want to biopsy patient due to risks). The tag integrates into each break caused by the cas9. We extract and purify genomic DNA and sequence tagged reads. After aligning to genome, we have identified off-target candidate sites.

• Guidescan2 (pure computational): Collect each 20-mer that matches gRNA with some limit flexibility and heuristics.

• IDT in silico assessment: some proprietary method that “leverages computational methods to nominate potential off-target regions for CRISPR edits”

Each method has its tradeoffs. Both CHANGE-Seq and ONE-Seq ignores chromatin context and native repair mechanisms. UNCOVER-Seq assesses the changes within cells and therefore is the most “natural”. Candidate off-target sites from each method were stacked together. For sites with at least 1-bp overlap, sites were merged and the candidate with lowest Levenshtein distance (number of edits necessary to turn candidate sequence into gRNA) was kept. Sites were prioritized as priority 1, 2, or 3 based on high frequency score from their original method, reproducibility across methods, exonic annotations. High priority sites (found in exon, higher score than on-target site) were evaluated by targeted PCR following next-generation sequencing after a clinical batch of k-abe was tested on Q335X lentivirus-transduced HuH-7 cells and primary human hepatocytes from three healthy patients (6, 7, and 13 months old). From HuH-7 cells, the estimated EC90 was 52 ng/mL and a 20x greater dose was used in the HuH-7 cells to stress test off-target effects. Of the sequenced sites, only an intronic site 369 bases upstream of closest coding sequence (a splice acceptor) at ATP7B displayed off-target editing.

ATP7B loss-of-function (LoF) causes Wilson disease but it is unlikely this finding could lead to LoF for two reasons: 1) low level editing so if edit leads to LoF it affects few cells, not enough to globally affect enough hepatocytes to mess with copper transport (see below figure), 2) unlikely that the edit itself leads to LoF given 369 bases from a splice acceptor which is the nearest coding sequence.

Primary hepatocytes from the three donors were treated at 20x the EC90 dose and sequenced three days later6. Only one site in primary hepatocytes, the ATP7B intronic site observed in HuH7 cells, displayed “even minimal off-target editing albeit at negligible levels that would not have been interpreted a priori as being indicative of off-target editing”.

Now that cellular safety has been characterized, let us move to clinical measurements of safety. AmplifyBio performed a safety study using the toxicology batch of k-abe in nonhuman primates. Nine juvenile cynomolgus monkeys (2.6-3.7 years old) with weights of 2.4 kg to 3.45 kg were split randomly while preserving mean body weights into three groups (control necropsy on day 14, 1.5 mg/kg necropsy 4 hours later, 1.5mg/kg necropsy on day 15). It’s a little unclear why 1.5 mg/kg was chosen. k-abe was infused over 30 minutes7. Lots of measurements and labs were taken including clinical observations, body weights, opthalmic observations, respiratory rates, anatomic pathology, labs (CBC, CMP, coagulation factors), organ weights, and microscopic evaluation of selected tissues.

“Administration of the drug product resulted in no adverse findings. The observed changes in the leukogram were minimal to mild and were consistent with a nonspecific physiological stress response and deemed non-adverse. The described serum enzyme elevations were minimal to mild and resolved by the conclusion of the study; therefore, these changes were considered non-adverse. The changes observed 4 hours following dose administration in the liver were transient and non-adverse as they were resolved by day 15.” From Musunuru et al 2025 NEJM Supplemental Text

The one mini-finding was increased liver weights corresponding to minimal diffuse cytoplasmic vacuolization of hepatocytes 4 hours after administration of k-abe. This was transient and resolved by day 15 hypothesized to have resulted from the LNP carrier. Toxicity from the carrier has been cited by other gene editing companies. Notably, VERVE-101 attributed the observed grade 3 transaminitis and thrombocytopenia event in their gene editing trial for lipid lowering to the LNP (note that VERVE-101’s target was LDL-R based uptake of hepatocytes). After this event, Verve shut down their trial on this asset. Subsequent trials from VERVE for lipid lowering used a different carrier and have been free of serious adverse effects. This is another reason why studying lipid chemistry, modeling lipid structure and tissue distribution is of paramount importance as we pursue more gene editing therapies.

No observable adverse events were detected at 1.5 mg/kg of k-abe in the monkeys. So, we know this to be the lower bound of a safe dose. Converting to a human equivalent dose requires scaling by body surface area. We arrive at an equivalent dose of 1.25 mg/kg. There was > 99.5% clearance of lipid excipients in the blood at 2 weeks after treatment compared to peak which informed their 3 week spacing requirement between doses to limit aggregate k-abe exposure. To be safe, the starting dose was dropped to 0.1 mg/kg.

Here is a summary of the timeline. The team was able to give KJ the first dose in ~7 months after recieving IND approval from FDA8. Incredible.

Can we shorten this timeline? What steps can we expedite? Folks like David Liu and Kiran Musunuru are the domain experts and can offer better opinions. However, here are mine. The cell line work to identify the right base editing system and mouse work to follow up whole liver editing activity data took 4 months. Can we streamline experimentation of base editing systems? I would argue the mouse work becomes less important for assessing editing activity is unnecessary if we can collect primary hepatocytes from a healthy donor (as they did for off-target work) and edit the normal genome to have the variant of interest. What about toxicology? The off target validation in mice, cell lines, and primary hepatocytes was time consuming and did not reveal anything discordant (which is good, but raises the question of what the minimal set of experiments are). Perhaps efficiency can be gained by establishing a company responsible for logistics and coordinating sub-teams executing each step of the process - a Palantir for gene editing.

Finally, lets cover KJ’s clinical course. KJ was placed on continuous renal replacement therapy (CRRT) due to his elevated ammonia shortly after birth and quickly transitioned off CRRT and onto glycerol phenylbutyrate (nitrogen scavenger), citrulline supplementation (200 mg/kg), and a protein-restricted diet. After an “expected honeymoon period from day 50-100”, KJ’s biochemical status worsened leading to an increased in the nitrogen scavenger dose to manage the elevated ammonia and glutamine and decrease in protein intake. KJ had one rotavirus-positive gastroenteritis around day 120. The first dose of the rotavirus vaccine is normally given to kids at 2 months of age. It is not stated whether KJ received his rotavirus vaccine but urea cycle disorders are not a classical contraindication to rotavirus vaccines. There were at least five distinct hyperammonemic crises (> 50 micromoles/L), and each event placed KJ at risk for permanent neurological damage and death. He was placed on the liver transplant list due to the severity of his CPS1 deficiency at 5 months.

Out of concern for KJ developing an immune response to full-length CPS1 protein (since he natively had none and his immune system might recognize it as foreign if presented on MHC class 1 proteins) prophylactic immunosupression with tacrolimus and sirolimus (exact doses not reported) was initiated on day 205. On day 208, dose 1 of k-abe (0.1 mg/kg) was given. On day 209, tacrolimus and sirolimus were re-administered. On day 230, dose 2 of k-abe (0.3 mg/kg) was given. These doses are way lower than the 1.25 mg/kg derived from the monkey experiments and the mice experiments (1 mg/kg and 3 mg/kg with max edit activity ~40%). The glycerol phenylbutyrate at 10.1 ml/m2/day kept the ammonia below 50 from day 150 onwards. And after the second dose, KJ was able to tolerate a 50% reduction in his glycerol phenylbutyrate without rising glutamate levels or other biochemical parameters. Side effects of glycerol phenylbutyrate in pediatric populations include neutropenia, fever, diarrhea, hypophagia, and cough. His first glycerol phenylbutyrate wean attempt after the first k-abe dose was unsuccessful due to rising glutamine levels.

After his first dose, KJ contracted a rhinovirus positive URI. After his second dose, he contracted gastroenteritis and a new rhinovirus or enterovirus infection. These infections may be due to long-term hospital stays, glycerol phenylbutyrate adverse effects, immunosuppression, normal (babies get sick), k-abe, or relative nutrient deficiencies. Importantly, KJ’s ammonia levels remained normal during these infections. When KJ had gastroenteritis pre-treatment, he experienced a hyperammonemic crisis. This is extraordinary. Further, KJ was able to tolerate a larger protein intake which will help him grow and develop. He remains on citrulline supplementation. k-abe allowed KJ to drop his glycerol phenylbutyrate dose by 50%, tolerate infections without hyperammonemic crises, normalize his biochemical profile, and handle a normal protein intake.

These charts suggest that the ammonia-related damage occurred from days 0-150. It remains to be seen the effect that these episodes had on KJ. Further, observing KJ as he grows will be important to understand the effects of k-abe. I am interested in following KJ’s neurocognitive testing and other parameters of brain function. Outstanding questions include KJ’s clinical and developmental history and whether he may need repeat doses of k-abe as he ages.

6

Primary hepatocytes had at least 4 times higher k-abe mRNA levels compared to HuH7 cells. Edit rates vary by host even amongst healthy primary hepatocytes (see below). Edit rates (both on and off-target) in primary hepatocytes should be larger than those observed in HuH7 cells. More so, it suggests that liver specific chromatin states and transcription factors favor greater edit rates. This makes sense since CPS1 and other urea cycle enzymes should be in an open chromatin state for most of the cell cycle to ensure urea can be made in the liver.

• Cell-free RNA to diagnose colorectal cancer

• In vivo CAR T cell generation to treat cancer and autoimmune disease

• PanEcho: AI-enabled echocardiography interpretation

• Time-varying Bayesian models for predicting disease risk

• NOD2 loss-of-function mutations lead to great immunotherapy response

• Risk-standardized mortality measures with EHR extracted features, CMS methodology report

• Protein language models are biased by unequal sequence sampling across the tree of life

At present we can detect deformities in bones and the bony pelvis; we can note the growth of bones and the malacosteon conditions important to the obstetrician; we can clearly see the shoulder-joint, the ribs, the vertebrae, the union of the pubes, the hip-joint and the curvatures of the spine. We can describe fractures of bones and readily distinguish them from dislocations, determine how to set them, and even watch their reunion without removing the splints and bandages. We can watch the absorption of a bone and readily find imbedded foreign substances. We can detect the presence of puss, watch the heart beating, see the form of the liver, the lungs, the kidneys and the spleen. I have seen with the fluoroscope the fetus in utero at the seventh month.

What is required are Crookes' tubes, a Ruhmkorff coil or a static machine, a fluoroscope, plate-holder, sensitized plates and other photographic apparatus. ... The expense of the machine costs only about $400 while a good coil is valued at from $200 to $600. But, of course, this need not be considered by physicians, since all are wealthy

These comments represent the instant click that many physicians felt when working with and look at XRs. It was a tool to answer their own personal curiosities about the human body and get answers to questions they spent much time thinking about and treating. And, for physicians like Stubenbord, they did not think about XRs in terms of return on investment - they had the money and this was a cool tool to try out at the very least.

During the initial wave of XR interest, there were physicians and scientists thinking about the risks associated with XRs too. Exceedingly painful dermatitis (more like a severe burn) caused by XR use was described in 1896. Codman (MGH surgeon) recognized the potential for XR to diagnose and study injuries and disease though he also paid close attention to reported injuries. In 1896, 55 were reported but by 1901 only 1 was reported. He felt the decline in reported occurred because people chose not to publish unless they exhibited unusual features, a reasonable conclusion. William Rollins (scientist and dentist) published over 200 papers from 1896-1903 urging minimum exposure, lead enclosures, and goggles to minimize risks. He became the father of radiation protection. Radiologist Alan Hart and others criticized XRs as “large loud sparking smelly devices prone to mishap and injury even when fully under the control of the physicians who in droves invested money and prestige in them”. Some surgeons doubted the image quality. Others felt that the bullet could be left safely in the body and worried about patient privacy (seeing through clothes).. The public was fascinated by this technology and studios offered the public “views of their bones” and “shoe fitting” images. These developments are expected of any new technology and necessary for its rapid adoption. People need to think about safety and drumming up interest with the public creates demand (although in this case they were likely causing harm to people who visited these studios because of the radiation dose and were condemned by medical societies). Carefree use led to calls for regulation at 1905 german radiology congress and American Ray Society protection committee in 1920. Wolfram Fuchs (1896) advised cutting exposures. Pfahler “unlikely to reveal much that palpation cannot” because long exposures blurred images and fractures could already be set empirically and bullets often harmless if left in situ.

Physicians started buying and assembling XR machines in 1896. Shortly thereafter, XRs were used during war time. In the Greco-Turkish war of 1897, the Ghe German Red Cross set up an XR machine in their hospital at Constantinople to aid wounded Turks. The British who supported the Greeks ad their machine in a villa outside Athens. The apparatus was bulky, and the power source unstable, so it was kept outside the front despite “British and German surgeons conclud[ing] that radiographs were of great use”. XRs were used in hospitals in the 1898 Spanish War with conclusions that XRs made probing for bullets obsolete. Medical journals published article after article in the following period describing XRs as indispensable for fractures and foreign bodies. XRs were used in the 1898 Spanish War allowing medics to triage injuries, reduce false incisions, and decide which limbs to amputate in the field. In 1908, at the American Roentgen Ray Society’s 9th meeting, it was stated that there was no excuse whatever for injury because radiographs can be taken in a fraction of a second. From 1896-1908, XRs were used for diagnosing conditions beyond fractures and bullet wounds like gastric cancers and ulcers. XR usage for diagnosing fractures at The Pennsylvania Hospital (TPH) was not common in 1909, despite the military and academic interest. At TPH, the chief resident was in charge of the XR along with other medical equipment. In 1910, the chief at the time stayed on for two years and insisted on purchasing new equipment. He was in charge of only the XR machine and limited his practice to the XR machine (radiology had yet to be created as a formal field) and he was given a percentage of the fees generated from patients who paid for the service. Creating dedicated positions and aligned incentives helped accelerate XR use at TPH. Other centers followed in the coming years and the first radiology residency program was established at MGH in 1915. Then, WW1 happened. This was a large catalyst for adoption. Marie Curie built mobile XR units and persuaded the Military Medical Service and Red Cross to use them. She was named the director of military radiology and built 20 mobile vans and 200 fixed stations training operators at each site. Robert Jones and John Hall-Edwards, who were both doctors and war office advisors, ordered XR equipment for each medical station and create a Roentgen service group within the Royal Army Medical Corps in England. Working with Siemens-Reiniger, each German field hospital recieved a wagon-mounted XR unit. US Surgeon-General William Gorgas commissioned a Division of Roentgenology where mobile XR units became increasingly available towards the end of the war. WW1 accelerated XR usage and proved lifesaving for shrapnel triage. Usage curves rose steeply in the 1920s and onwards. In 1933, 55% of all patients admitted to a university hospital had at least 1 XR. At Mayo, it was around 80% of all patients.

XR use cases continued to expand including Charles Leonard’s work (1900-1930) identifying renal stones missed on physical exam. Like many other XR superusers at the time, Leonard suffered many radiation-induced injuries including amputation of his right arm and died of radiation-induced cancer at 51. Alan Hart built statewide clinics in the 1920-30s for TB screening and in the early 1930’s positive contrast was shown to visualize ulcers, tumors, and strictures. The use cases had quickly expanded from fractures and bullets to most every organ system. By the early 1930s, radiology residency existed, professional societies had been established, hospital workflows were designed to rely on the XR for diagnosis and monitoring of disease.

XRs would not have taken off if patients and the public did not want to pay for it. In fact, hospitals charged $1-3 per image that were billed directly to the patient in a fee-for-service model. Self-pay made up about 70-80% of hospital payments, charity represented 5-10%, and the remaining was from state or municipal lump sum payments for patients who could not afford to pay. This is vastly different from today’s environment where 42% of MGH’s payment comes from Medicare, 14% from Medicaid, 37% from commercial insurance, 2% from self-pay and 5% from workers comp/liability/health safety net. Hospitals in the early 1900s did not answer to large insurers and controlled their pay-mix better so they could do more of whatever they wanted.

In the early 1900s, new services like XRs were desirable if patients would pay for the service or donors would fund the equipment and labor. So, to a first-degree, all that mattered was whether there existed demand. Today, new services are adopted if insurance covers it or it creates obvious increases in margin. Hospitals were fee-for-service largely operating off self-pay. Today, we have fixed payments, readmission penalties, value-based care that avoids excess procedures and emphasizes shortning length-of-stay, and regulatory oversight. These principles are good, but may be preventing other XR-like advancements from positively impacting patients such as cell-free DNA. For example, when Roentgen discovered XRs there was no federal device regulation in place. The first law to touch devices was in 1938. Today, we have heavy pre-market regulation. This regulatory oversight itself has many pros (protecting patients, exceeding standard of care) but also its faults (limiting pace of new technology). The new FDA under Makary has a bold vision for the future, that could protect patients while accelerating adoption of new technologies in the clinic. Beyond regulatory oversight, most technology requires CMS coverage (traditionally, commercial payers follow CMS) before mass adoption so hospitals know they will be reimbursed. An example is Galleri, the multi-cancer early detection test, which does not have CMS approval and low uptake.

How did the X-Ray gain mass adoption?

• Military use in WW1 accelerated its adoption (led by individuals like Marie Curie who dedicated time to making this a reality)

• The bar for comparison for fracture detection or bullet retrieval was palpation or surgical exploration - two crude technologies. Today, the standard of comparison is globally higher.

• Establishing incentives that aligned with physician interest that allowed them to run with the technology like the financial structure established at The Pennsylvania Hospital

• Lack of regulatory guardrails. There was none.

• Physicians felt strongly about the XR and what it could do - it captivated their curiosity. They tried to image beating hearts and see what a fetus looks like in utero. On the whole, physicians threw caution to the wind exposing themselves to burns, dermatitis, crazy injuries (amputations of limbs), and death (mostly cancer)

• The wins for use cases piled on across domains quickly within the first 10 years. It started for fracture detection and bullet localization but quickly spread to TB diagnosis, renal stones, strictures, ulcers, and tumors.

• Establishing specialist societies to turn an informal thing into a job; establishing a residency program

• Clear billing model (fee-for-service) that required only ~300 XRs to recoup the cost of capital.

• Anyone could interpret the end result. How well they could is a different question. But, at the least, they could look at it and go “there’s my hand bones”. It was a picture. And humans are visual in that “a picture is worth a thousand words”.

Consider the assumption that cell-free DNA for monitoring response to treatment and diagnosing disease will become a reality in the future. In 2100, I doubt we will be imaging every tumor to judge response to treatment. Why has it not already gained mass adoption? XRs introduced in 1896 gained “mass-adoption” by 1925. Cell-free DNA gained popularity in the 2005s so we have till 2035 to match XR adoption. Nevertheless, compared to XRs, what makes cell-free DNA harder to adopt:

• FDA review or CLIA LDT needed. You can’t just roll it out and see what happens like we did with XRs.

• Cell-free DNA reports are text-based, statistical, and require knowledge of genomics and statistics along with medicine in order to interpret. That is a lot of background knowledge. When we simplify the reports too much, they become unusable. But, at the most granular level, they become uninterpretable to basically every clinician. We need a middle ground where these reports not only fit into clinical workflows but can be interpreted by each physician.

• We do a bad job of making cell-free DNA reports accessible in Epic. The Genomics tab with scanned pdfs is a horrible interface.

• Medicare only covers ctDNA in narrow situations, and hospitals will not embed a test that generates denied claims. With XRs, insurers were not relevant, the hospitals could do more of what they wished and less of what the insurer wanted.

• A single broken bone convinced surgeons. No RCTs, no modeling. Now, we want trials that show survival benefits or cost savings relative to standard of care (when the standard of care (imaging) is deeply embedded in the hospital workflow)

• XRs are produced in minutes, visible to physician, and immediately alters the next cut of scalpel. cfDNA returns in days, the interpretation is statistical and triggers further workflow (scans, biopsies, genetic counseling). If the MAF of KRAS G12C is 5% down from 7%, and your patient is on a G12Ci, should you continue? What if it was 4%? What if it was 8%? These are non-trivial decisions that clinicians are not trained to make. It gets harder when they don’t have time to sit and deliberate because they are booked with patients or meetings all day.

• Slower clinician demand for cf-DNA. Maybe 30% of oncologists I have met are cf-DNA enthusiasts. This number is probably skewed upwards because I love cf-DNA. And, when talking about non-oncologists? Forget it. Most of them could care less.

• A positive factor driving cfDNA uptake in the US: COVID helped pushed liquid biopsies more into the mainstream. No procedure rooms are needed to make clinical decisions and keep revenue flowing. Now, every major insurer covers liquid biopsy therapy selection tests (like those for PARPi in prostate cancer), and prenatal cfDNA is basically universally covered. Guardant360 tests rose from 70k in 2019 to 200k in 2023. The number of Natera tests increased from 800k in 2019 to 2.5M in 2023.

Thanks to Simon Barnett for the discussion that led to this post, and comments on the work.

I plan to post 3-4 times a month and write about biotech companies I find interesting (or not interesting), commentary on clinical trials (e.g., covering the CAPE COD trial and why the signal could be artificially enhanced due to imbalance between the two groups), thoughts on new life sciences research (e.g., now that protein design and binding affinity prediction tools are getting really good, where do we go from here?), my current reading list, and other thoughts on the future of medicine based on what I see, hear, and read. Please consider subscribing if you find that what I have to say is interesting.

Subscribe now