FM-GWAS: From Personalized Genomes to Disease Gene Discovery | Eran Segal | ADIA Lab Symposium 2026

Value of the Information & Strength of the Argument

The talk presents a high-value integration of comprehensive multi-omics data with advanced AI models, demonstrating substantial progress in personalized health prediction and gene discovery. The argumentation is strong for the microbiome and glucose models, where clear examples and external validation (e.g., Spain cohort) bolster claims. For FM-GWAS, the comparison with MAGMA and replication in UK Biobank adds credibility, though the details of the model architecture and statistical procedures are not fully disclosed. The simulation of clinical trials is compelling, with 70% concordance across 192 trials, but the selection of biomarkers and potential biases are not discussed. Overall, the value is high, but the argumentation would benefit from more methodological depth and caution regarding causal claims.

Scientific Rigor, Source Quality, Title Accuracy

The talk reflects rigorous scientific practice, drawing on large-scale, longitudinal data and referencing multiple peer-reviewed publications implicitly. The Human Phenotype Project is a well-established resource, and the research is presented with appropriate caveats about causal inference (e.g., ongoing clinical trials for microbiome). However, the presentation does not cite specific papers or URLs, relying on the audience’s familiarity with the field. The description provides only a link to the symposium page, which serves as the sole explicit source. The title precisely matches the content, which focuses on the FM-GWAS framework and its application. The lack of detailed methodology and independent sources in the presentation limits the ability to fully assess robustness, but the overall impression is of high-quality, data-driven research.

The title accurately describes the core topic of the talk, which introduces the FM-GWAS framework and its application to disease gene discovery.

8/10

Presentation by a leading researcher in computational biology, based on large-scale longitudinal datasets and multiple published studies. The talk is informative but lacks detailed methodological transparency and independent verification in this format.

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

• 01:00Introduction to the Human Phenotype Project and its comprehensive data modalities (molecular, cellular, organ, whole-body).
• 04:00Discovery of bacterial SNPs in the microbiome associated with BMI, including an example where a single mutation explains 2 points of BMI.
• 07:00Development of a transformer-based glucose monitoring model that predicts prediabetes-to-diabetes conversion more accurately than HbA1c.
• 09:30Introduction of GeneUNet and FM-GWAS for gene-level association testing, leading to more significant gene-trait associations than MAGMA.
• 12:00Presentation of HealthFormer, a world model of human biology, predicting future biomarker levels and clinical transitions with improved accuracy.
• 14:30Demonstration of simulating clinical trials, with model predictions matching over 70% of randomized trial outcomes for various biomarkers.

• ADIA Lab Symposium 2026 — Official symposium page providing context for the presentation and related research updates.