AI Genomic Models Struggle With 3 Billion DNA Letters as Human Gene Control Defies Linear Prediction
Updated
Updated · WIRED · Oct 17
AI Genomic Models Struggle With 3 Billion DNA Letters as Human Gene Control Defies Linear Prediction
3 articles · Updated · WIRED · Oct 17
Summary
Current genomic AI systems such as Evo 2, Genos and DeepMind’s AlphaGenome can make useful predictions, but researchers say they still cannot fully capture how the human genome actually works.
Only about 2% of the roughly 3 billion DNA building blocks in the human genome are protein-coding genes; the harder problem is regulation across the rest of the genome, where genes switch on and off by cell type, timing and external signals.
Hundreds of thousands to millions of enhancers, long-range chromatin loops, 3D folding, TADs, epigenetic marks, noncoding RNAs and RNA splicing all shape gene activity, often differently from one seemingly identical cell to another.
That fluid, context-dependent behavior challenges AI models trained mainly on sequence-to-trait correlations, because key information may be missing from genome data alone and may also depend on development, environment, microbiome and culture.
Researchers remain optimistic that AI will aid discovery, but they argue black-box prediction will not replace human efforts to uncover the genome’s underlying principles.