AI Foundation Models Show Promise But Cannot Explain How Genomes Actually Work
Genomic AI systems trained on massive datasets can make predictions about DNA, but lack the interpretability that biologists need to truly understand gene regulation.
What happened
Genomic foundation models such as Evo 2, Genos, and Google DeepMind's AlphaGenome have been developed and trained on vast quantities of genomic data to predict how DNA sequence differences affect biological processes and traits, including disease risk. These algorithms operate as computational black boxes, deducing correlations from known cases without explicitly modeling the regulatory mechanisms that control genes. The models sidestep the problem of understanding gene regulation—the process by which cells switch genes on and off—which remains poorly understood despite decades of research.
Context
Completing the human genome sequence in 2003 revealed that only 2 percent of the 3 billion DNA building blocks code for proteins; the rest appears to govern gene regulation through complex, overlapping systems of control that differ markedly from simpler organisms. In bacteria, gene regulation follows straightforward logic—particular signals turn genes on or off. In humans, regulation involves multiple transcription factors working in groups, often with context-dependent effects where the same factor may activate genes in one cell type and suppress them in another, following an 'AND' logic rather than simple 'OR' switches. This gap between predictive capability and mechanistic understanding matters because biologists seeking genuine comprehension of how the genome and ultimately life works cannot rely on algorithms that hide their reasoning, regardless of their predictive accuracy.