Why AI Struggles With Genome Regulation — SkimNews
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- The Human Genome Project (1990–2003) found that only about 2% of the human genome consists of actual genes, leaving roughly 98% involved in regulatory processes that remain poorly understood more than two decades later.
- Genomic foundation models including Evo 2, Genos, and Google DeepMind's AlphaGenome are trained on vast quantities of DNA sequence data to predict how genetic differences affect biological traits and disease risk.
- These AI systems function as 'computational black boxes' that derive correlations from training data but cannot explain the underlying regulatory mechanisms, a limitation biologist Karen Adelman of Harvard Medical School says blocks any real understanding of how the genome works.
- The human genome holds roughly 20,000 genes but possibly millions of enhancers — regulatory DNA elements that can each influence multiple genes, and which scientists still cannot fully map or assign to their targets a quarter century after the Human Genome Project.
- Human gene regulation operates on an 'AND' logic, integrating many signals for context-dependent decisions, compared with the simpler 'OR' logic of bacteria, making the system tunable in ways that resist straightforward input-output modeling.
- Distal enhancers sit millions of nucleotides away from the genes they regulate and reach them through chromatin loops extruded by the motor protein cohesin, forming fluid 'condensates' that genome biologist Wendy Bickmore of the University of Edinburgh says vary unpredictably from cell to cell.
- Adelman and Bickmore argue that 4 billion years of evolution have produced a system whose nuances may exceed current analytical frameworks, questioning whether even large genomic foundation models can capture the full picture.
Why it matters: Researchers pursuing mechanistic explanations of gene regulation — central to understanding complex diseases — will not get them from AI black boxes that only return correlations. With hundreds of thousands to millions of enhancers still unmapped 25 years after the Human Genome Project, the gap between AI's predictive output and causal understanding of noncoding DNA could mislead drug-target discovery and disease-mechanism research.
Ask SkimNews
