Stanford's Kundaje: LLMs fall short on noncoding developmental variants where data is scarce

anshulkundaje · x · 2026-11-07

Stanford professor Anshul Kundaje argues that many unsolved genetic cases don't act through canonical protein-coding variants, but through highly context-specific effects in early developmental cell types and stages — requiring hard-to-obtain fetal and developmental samples.

Both humans and LLMs need computational and experimental tools to crack these; both remain limited today.
For coding variants, strong ML models and variant scores exist, and LLMs are useful to reconcile evidence and build causal narratives.
For noncoding developmental effects, models need access to precious developmental-context data — the real bottleneck — with organoids and iPSC-derived cells only partially helpful.

Original post →

Stanford's Kundaje: LLMs fall short on noncoding developmental variants where data is scarce

More from AGI Musings