Stanford professor argues rare disease bottleneck is biology, not AI

Stanford professor Anshul Kundaje and Daniel McNinnon, a startup founder and parent of an affected child, debated whether AI can solve rare diseases. Kundaje's core view: AI and LLMs can only play a supporting role, easing the human-labor bottleneck in diagnostic interpretation, but the fraction of cases that benefit is relatively small; the real bottleneck in rare disease lies in the biology and engineering itself—completely unknown mechanisms cannot be reasoned into existence. McNinnon countered that the two were actually saying the same thing: the path to solving these problems at scale is precisely AI plus robotics, and he is confident discovery speed can be increased 100-fold.

Confirmed

• Kundaje believes AI can accelerate diagnostic interpretation for some N-of-1 rare disease cases, but this only addresses the human time problem, and only a small fraction of cases benefit.
• He noted the number-one bottleneck is identifying causal variants, genes, and mechanisms—the vast majority have never been characterized, and neither humans nor LLMs can reason their way to completely unknown problems.
• Many unsolved genetic cases are not classic protein-coding variants but highly context-specific effects in early developmental cell types and stages, requiring fetal and developmental sample data that is hard to obtain.
• After LLMs filter non-coding variants, experimental validation becomes a huge challenge: accurate experimental model systems for developmental cell types barely exist.
• Even if validation is solved, therapeutic delivery remains unsolved—there is no answer yet for how AAV, CRISPR, or small molecules reach most disease-relevant cell types.
• He concluded that rare disease requires: measuring the right entities in the right cellular context (fetal development and disease context), understanding basic developmental biology, reliably delivering therapeutics, and more reliable screening.
• McNinnon responded that discovery has been too slow and too expensive, and AI plus robotics is exactly the path to solving these problems at scale—he is quite confident of a 100x speedup.

Why it matters

This debate touches on the limits of AI's capabilities in the life sciences: when algorithms can screen candidate variants at scale, the real constraints shift to "wet lab" steps such as data accessibility (fetal samples), missing experimental models, and drug delivery. For AI healthcare founders and researchers, it suggests LLM-driven discovery must be paired with breakthroughs in experimental validation, sample access, and delivery technology—otherwise faster diagnosis will not translate into therapeutic outcomes.

2026-11-07 ~ 2026-11-07 · 10 related posts

• [source] Stanford's Kundaje: AI won't solve rare disease — the bottlenecks are measurement and delivery — anshulkundaje · 2026-11-07
• [source] Builder doubles down: AI plus robotics can 100x rare disease discovery pace — danielmckinn0n · 2026-11-07
• Geneticist: rare disease won't be solved by AI and robotics scaling alone — anshulkundaje · 2026-11-07
• [source] AI won't crack rare diseases' #1 bottleneck: uncharacterized causal mechanisms — anshulkundaje · 2026-11-07
• Stanford's Kundaje: LLMs fall short on noncoding developmental variants where data is scarce — anshulkundaje · 2026-11-07
• Stanford Prof: Once LLMs Can Act on Noncoding Variants, Validation Becomes the Bottleneck — anshulkundaje · 2026-11-07
• Kundaje: Even After Validation, Delivering AAV/CRISPR to Causal Cell Types Remains Unsolved — anshulkundaje · 2026-11-07

3 near-duplicate retellings: anshulkundaje · anshulkundaje · anshulkundaje