Stanford's Kundaje debates how far genomic prediction models still are from molQTLs
anshulkundaje · x · 2026-08-30
A substantive academic debate in AI genomics. Stanford's Anshul Kundaje argues that for chromatin phenotypes, current SOTA predictive models already outperform Prediexcan-like methods and are approaching heritability ceilings, and that fine-mapped caQTL maps built with lenient posterior thresholds contain many false positives.
His counterpart tuuliel concedes statistical limits but notes current predictive models capture only a small fraction of empirically observed molQTLs, and that QTL mapping or CRISPR validation across all cell states is impossible. Kundaje replies that expression prediction indeed has a long way to go — his papers explicitly temper expectations — and that his hybrid statistical-plus-predictive approach is precisely meant to reduce the need for massive sample sizes.
More from Research
- a16z addresses medical AI evaluation challenges; multimodal AI becomes biology's translation layer — agihouse_org · 2026-08-30
- Alibaba DAMO's LiON liver cancer model beats radiologists, catches missed metastases — jiqizhixin · 2026-08-30
- Anthropic's autonomous researcher beats humans at 1/37th the cost; OpenAI plans rival for September — daniel_mac8 · 2026-08-30
- Apple's Agent Seer Generates Agent Eval Suites Directly from MCP Specs — omarsar0 · 2026-08-30
- JIT-Agent: Weaker Models Beat Stronger Ones via Self-Written Scaffolding — krishnan · 2026-08-30
- FSU's Fall 2025 Machine Learning course by Christian Hubicki now free online — caglar_ee · 2026-08-30