Salah Mahmoudi at ARDD2026: AlphaForge predicts clinical outcomes

Value of the Information & Strength of the Argument

The information value is high in terms of practical applicability: AlphaForge could enable earlier go/no-go decisions, potentially saving years and substantial costs. The speaker provides concrete numbers (AUC, performance gap vs. LLMs) and illustrative case studies, which strengthens the claim. The argumentation is largely based on internal validation rather than external, and the lack of detailed methodological transparency and peer review limits the strength of the conclusions. Still, the prospective and ’timestamped’ predictions are a notable improvement over simple retrospective analyses.

Scientific Rigor, Source Quality, Title Accuracy

Scientific rigor is moderate: the data are unpublished and no full technical report is provided. The speaker discloses key design choices (exclusion of commercial failures, prevention of data leakage) but not the underlying algorithm or training details. Sources are not explicitly cited within the talk; however, the organizer’s website and LinkedIn page are listed in the description. The title matches the content exactly and does not overclaim beyond what is shown. The presentation is consistent with the conference’s focus on longevity, though the clinical-outcome scope is broader.

The title accurately reflects the content: the speaker presents AlphaForge, a model for predicting clinical outcomes.

7/10

The presentation shows promising quantitative results (AUC 0.90 retrospective, 0.886 prospective) but the data are unpublished and the methodological details are only briefly described. Validation is partly based on a single anecdotal fund case, and external confirmation is still pending.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 01:10Introduction: the problem of late drug development failures and cost.
• 03:00Presentation of AlphaForge: virtual drug review board using structure, target, disease.
• 05:20Retrospective cohort of 391 small-molecule programs, balanced, with AUC 0.90.
• 07:10Comparison with frontier LLMs: AlphaForge outperforms by ≥29 points; many LLMs near random.
• 09:00Prospective validation on 83 programs post-model freeze (Jan 2025), AUC 0.886.
• 10:50On-going tracking of 721 clinical programs; 32 revealed outcomes show AUC holding.
• 12:00Risk distributions across top pharma and AI-native companies: no improvement.
• 13:50Predicting primary failure reason (efficacy, toxicity) in ~80% of cases; examples (sitaxentan, intepirdine).
• 15:20Expanded dataset to 1,341 drugs, AUC 0.87; performance also holds in trials for patients over 60.
• 16:30Closing invitation: try AlphaForge for free assessments; external validation example.

• Aging Research and Drug Discovery (ARDD) official website — Organizer of the conference and organizer of the event, linked as general reference.
• ARDD LinkedIn page — Official channel for conference updates and future meetings.

• Wikipedia: Artificial intelligence in healthcare — General background supporting the growing use of AI in clinical prediction and drug discovery.