Dr. Thore Buergel is CEO at Pheiron, a company dedicated to predicting clinical trials using AI and big data.
We spoke about the data needed to do this, who the customer is, and about this year’s big trials (Novo Nordisk’s ZEUS trial, Novartis’ Lp(a)HORIZON trial, Abivax’s ABTECT trial).
The full episode is on Youtube and Spotify!
This episode is brought to you by my partners at Consensus. Here’s what’s new:
The Consensus Research Agent: Consensus can now plan and run multi-step research tasks with tools. 400M+ scholarly sources, up from 250M: new full-text partnerships with 7 of the 10 largest academic publishers, including APA, AAAS, Wiley, Sage, OUP, and Taylor & Francis. Citation Grounding: every citation in every AI summary is mapped to an exact quote from the paper. A Library you curate: upload grants, manuscripts, and book chapters alongside papers — point the Research Agent at everything you’ve saved.
You can get a free trial with my link!
My key takeaways
1. A quantitative approach for drug development
Pheiron’s idea is to use quantitative assessments (rather than human intuition) to help drug development - a bit like Moneyball, for pharma. They ask how much each evidence source increases the probability that a drug program succeeds, then backtest this against historical clinical outcomes.
Their analysis suggests that evidence quality varies enormously: rare human genetic evidence is particularly predictive, while some high-throughput expression or perturbation data adds surprisingly little. Crucially, they separate biological support from trial-specific probability of success - good biology can still fail in a badly designed trial.
2. The data that underpin predictions
Pheiron are human evidence maximalists. Their core inputs are individual-level human data: genetics, electronic health records, biobanks, hospital cohorts, multi-omics, tissue and cell datasets, plus clinical precedent and trial data. Their stack currently covers roughly 4.8 million individuals, hundreds of cell types and 50+ tissues, supplemented by essentially every relevant public database. AI agents retrieve and analyse these data rather than merely summarising literature. The key principle is that more data is not necessarily better: each source is weighted according to demonstrated predictive validity.
3. The prediction for Novo Nordisk’s ZEUS trial
Novo Nordisk's trial for an inflammation-blocking IL6 inhibitor just failed to prevent heart attacks in a major phase 3 trial. Pheiron predicted this at a 66% probability of success.
Why it should have worked:
- Precedent for the axis (from IL1β inhibition).
- Strong genetic evidence for the IL6 pathway.
- Convincing target engagement (reduction of inflammation - hsCRP) in previous trials.
Why it may have failed:
- Limited genetic evidence for the IL6 ligand.
- Questionable translatability of lifelong genetic loss to transient pharmacological inhibition.
- The previous clinical evidence (with IL1β) is strongest on non-fatal myocardial infarction, but the composite endpoint in this trial (which includes stroke) could dilute the effect.
4. The prediction for Abivax’s ABTECT trial
Pheiron gave obefazimod a 54% probability of success in its phase 3 trial for ulcerative colitis. It ultimately hit the efficacy endpoint decisively, outperforming what Pheiron expected. Their conservatism came from expecting efficacy to reduce during longer-term treatment - and from predicting bigger tolerability problems than the drug actually had.
Pheiron modelled obefazimod primarily as an miR-124 inducer, whereas the precise mechanism, it has emerged, is probably AhR modulation. It would be interesting to see if the probability of success would change if this altered mode-of-action was modelled instead.
5. The prediction for Novartis’ Lp(a) HORIZON trial
Pheiron gave Novartis’ Lp(a) HORIZON trial of pelacarsen a 49% probability of success, despite viewing Lp(a) itself as unusually strong biology. In the trial, pelacarsen did lower Lp(a) but did not lower cardiovascular events.
Pheiron’s concern was mainly the trial and population, not whether Lp(a) causally contributes to cardiovascular disease. Patients were already receiving substantial background cardiovascular therapy (statins, ezetimibe, PCSK9 inhibitors, GLP1s, SGLT2 inhibitors) leaving only a limited fraction of residual risk that Lp(a) lowering could modify. In Pheiron’s framework, HORIZON is a great example of strong target biology but much less certain clinical-trial success.
Follow Thore on X, here, and here is Pheiron’s Substack:
Thank you for reading/watching! The full transcript is here:
An interesting approach. I'm aware of (at least) one another company who are trying to predict clinical trial success (and claim an 80%+ success rate). That company is Quanthealth. I suspect these types of approaches can run into challenges when the MOA is truly novel and thus the actual clinical data is quite sparse.
One other thought: For large population diseases (such as RA, certain epilepsies and depression), a significant proportion of patients are non-responders to approved treatments. This figure invariably comes out at around 30% (although I can't confirm the actual data). It would be extremely valuable to identify potential non-responders in advance.
Samuel, any chance you could share a reminder with your substack subscribers. Here's text copy of the Reminder “The Aging Research and Drug Discovery Meeting October 1–3, 2026, at the David Rubenstein Treehouse at Harvard University in Boston, Massachusetts”. https://agingpharma.org/
Disclaimer: I have zero affiliation with the meeting though I am an Advocate for Longevity.