Why AI Isn’t Ready to Tackle Our Convoluted Genomes
Researchers at companies and labs behind Evo 2, Genos and Google DeepMind’s AlphaGenome are training massive neural networks on the roughly 3 billion base‑pair human reference sequence and associated variant catalogs. Their goal is to predict how a single‑nucleotide change will alter disease risk or cellular phenotype without explicitly modeling transcription‑factor binding, enhancer‑gene wiring, or chromatin folding. The piece points out that even after the Human Genome Project delivered a near‑complete map of the sequence, only about 2 % of those bases encode proteins, leaving 98 % “non‑coding” and largely responsible for the dynamic regulation that defines cell type, developmental stage, and environmental response. Because transcription factors in mammals bind loosely, cooperate in combinatorial groups, and act through millions of enhancers whose target genes are often ambiguous, the genome behaves more like an AND‑gate network than a simple ON/OFF switch. AI models that treat the genome as a flat string of symbols therefore risk conflating correlation with causation, offering predictions that lack mechanistic grounding.
This debate mirrors a broader wave of AI foundation models spilling into life sciences—from AlphaFold’s structural breakthroughs to language‑model‑driven drug design. Companies are racing to claim the “next big AI win” in genomics, betting that scale and data volume will substitute for biological insight. Yet the article reminds us that unlike protein folding, where physics imposes relatively clear constraints, gene regulation is a product of billions of years of evolutionary tinkering, producing overlapping, context‑dependent control layers. The hype cycle therefore risks overlooking the need for multimodal data (Hi‑C, ATAC‑seq, single‑cell epigenomics) and explicit mechanistic frameworks that current large‑scale models typically ignore.
If the field proceeds without integrating regulatory logic, several risks emerge: false confidence in variant interpretation could misguide clinical decisions; funding may be diverted from experimental mapping of enhancer‑promoter contacts; and the black‑box nature of these models could stall regulatory acceptance. Watch for next‑generation hybrids that embed transcription‑factor motifs, 3D chromatin maps, or causal inference layers into deep networks, as well as any benchmarking studies that compare pure‑data‑driven predictions against experimentally validated regulatory perturbations.
Key Takeaways
AI foundation models are being applied to genomics, but they currently ignore the complex, non‑coding regulatory architecture that drives gene expression.
Only about 2 % of the human genome codes for proteins, while the remaining 98 % houses millions of enhancers and context‑dependent transcription‑factor networks.
The combinatorial “AND” logic of mammalian transcription factors makes simple input‑output modeling insufficient for accurate phenotype prediction.
Future progress will likely require hybrid approaches that fuse large‑scale AI with explicit regulatory and three‑dimensional chromatin data.
About the Source
This analysis is based on reporting by Wired. Here is a short excerpt for context:
The DNA inside our cells isn’t just a series of instructions, but something far more dynamic. Like scientists, AI models don't yet have the data to properly decode it.Read the original at Wired
