Since the Human Genome Project completed a near-complete map of roughly three billion DNA bases between 1990 and 2003, researchers have learned that only about two percent of that sequence encodes proteins. The remaining stretch contains roughly twenty thousand genes, each ranging from a few dozen to several million nucleotides, yet the raw sequence alone has not revealed how life operates. The article originally appeared in Quanta Magazine and was later republished by WIRED.
Understanding how those genes are turned on or off proves far more difficult. In eukaryotes, transcription factors bind to DNA regions called enhancers, which can number in the hundreds of thousands or even millions. An enhancer may sit near its target gene or be located megabases away, requiring the protein cohesin to extrude chromatin loops that juxtapose distant elements. Each gene can be influenced by many enhancers, and each enhancer may affect several genes.
Karen Adelman, a gene-regulation researcher at Harvard Medical School, emphasized that the genome was not engineered like a blueprint. She noted that “the genes are probably not the most interesting part of the genome,” suggesting that the regulatory circuitry surrounding those genes holds the key to cellular diversity.
AI-driven foundation models such as Evo 2, Genos, and DeepMind’s AlphaGenome are trained on vast collections of genomic data. Researchers employ them to infer how specific DNA variations may influence biological processes and disease risk. The models rely on statistical correlations learned during training rather than explicit mechanistic rules, allowing rapid predictions but obscuring the underlying regulatory logic.
Wendy Bickmore of the University of Edinburgh highlighted that even after a quarter-century since the Human Genome Project, scientists still lack a comprehensive map of enhancer locations and functions. She described the regulatory system as employing an “AND” logic, where multiple signals must converge to activate a gene, contrasting with the simpler “OR” logic seen in bacteria. This combinatorial complexity challenges models that assume a straightforward input-output relationship.
While AI can accelerate hypothesis generation and flag potentially pathogenic variants, its black-box nature means it cannot replace mechanistic insight into gene regulation. The intertwined layers of transcription factors, enhancers, and chromatin architecture demand experimental dissection alongside computational tools. Until the full regulatory code is deciphered, predictions will remain limited, and the genome will continue to resist simple computational shortcuts.