Genomic AI Models Hit the Wall of Gene Regulation

Twenty-five years after the Human Genome Project, biologists still cannot say where all the gene-regulating enhancers in the human genome are, let alone what they do. That gap, University of Edinburgh genome biologist Wendy Bickmore said, is the core problem facing the AI genomic "foundation models" — Evo 2, Genos, and Google DeepMind's AlphaGenome — now being trained to predict how DNA differences shape traits and disease risk.

The stakes are high. If these models work, drug developers and geneticists gain a fast computational route from DNA variant to biological consequence. If they don't, the field will have spent enormous effort approximating a system that resists input-output prediction altogether.

What did the Human Genome Project actually reveal?

The international project, which ran from 1990 to 2003, deduced nearly the full sequence of our roughly 3 billion DNA building blocks. It also delivered a shock: barely 2 percent of the human genome consists of actual genes, the protein-coding sequences. Humans carry around 20,000 such genes, ranging from a couple of dozen to almost 3 million DNA "letters" (nucleotides) in length.

The remaining 98 percent is not filler. Much of it appears to govern how genes are switched on and off — the process biologists call gene regulation. "The genes are probably not the most interesting part of the genome," said Karen Adelman, who studies gene regulation at Harvard Medical School. The hard question is no longer what each gene does, but how cells use them, moment to moment, in response to signals from their surroundings.

Takeaway: reading the genome was the easy part; understanding how it is run is the unfinished project.

Why is gene regulation so hard to model?

In bacteria, regulation follows simple "OR" logic, Adelman said: one signal, one on-or-off decision. In humans, the logic resembles what computer scientists call "AND" — many signals integrated before a regulatory decision, with tunable rather than binary outcomes. "This is part of the beauty" of our regulatory complexity, Adelman said.

Several mechanisms deepen the difficulty:

• Enhancers: hundreds of thousands, perhaps millions, of DNA segments that gather transcription factors (the proteins that initiate gene reading) and deliver the "go" signal. Each gene may respond to many enhancers; each enhancer may affect multiple genes. Some sit millions of nucleotides from their targets.
• DNA loops: a protein motor called cohesin extrudes loops of chromatin (DNA plus its packaging proteins) to bring distant enhancers to their genes. The resulting interaction hubs, or condensates, are fluid, short-lived, and differ between cells — even notionally identical ones.
• Chromatin shape: DNA exists in densely packed heterochromatin, where genes are silenced, or open euchromatin, where genes are accessible. Special enzymes constantly repackage it.
• Epigenetic marks: small chemical molecules attached to histones (packaging proteins) or to DNA itself, altering chromatin packing and copied when cells divide. Whether a systematic "epigenetic code" exists, the source notes, is far from clear.

"We've stopped thinking about the genome as a linear piece of DNA code," Bickmore said. "Thinking about this incredibly dynamic three-dimensional folding as absolutely inherent to regulation is a very exciting change."

Takeaway: the genome operates as a dynamic physical structure in space, not a linear text.

What happens after transcription?

Regulation does not stop when a gene is read. Small noncoding RNAs called microRNAs can silence messenger RNA before it is translated into protein; a single microRNA may regulate many mRNAs, and vice versa. The spliceosome, a molecular assembly of proteins and noncoding RNAs, edits pre-mRNA — cutting out introns and splicing exons together — and can produce different protein isoforms in different cell types, sometimes with opposite regulatory effects.

Every player in this system, from transcription factors to noncoding RNAs, is itself produced by the same context-dependent process. The genome, in other words, acts on itself — a recursive system the computer scientist Douglas Hofstadter called "a strange loop." It is not, the source argues, a blueprint or a program.

"I wouldn't have designed it this way if I was God," Bickmore said. "But here we are!"

Takeaway: layered, self-referential decision-making makes the genome a moving target for any static model.

Why did evolution build it this way?

Bickmore suspects regulatory complexity was the only route to organismal complexity. "As you get more complexity, you need to add more and more enhancers," she said. "But where are you going to put them? You start to put them farther and farther away. Once they are [far enough], you start to need TADs and three-dimensional [chromatin] folding to allow those things to work." TADs (topologically associating domains) are chromatin compartments within which genes switch on or off in groups.

The genome also accumulated transposable elements — parasitic viral DNA that replicates across our chromosomes — forcing cells to evolve extra regulatory layers to suppress harmful translations. Evolution tinkers rather than designs, but eukaryotic regulation follows distinct principles, not just bacterial messiness.

Takeaway: the complexity may be functional, not accidental — which makes it harder to compress.

Can AlphaGenome absorb all of this into a black box?

The developers of genomic foundation models hope the full stack of regulation — transcription factors, splicing, epigenetic marks, loops, chromatin packing — will be implicitly captured in the correlations the algorithms learn between sequence and traits. They accept the black box so long as predictions are accurate.

Bickmore is not convinced. "I'm sure [AlphaGenome] is going to be useful, but with limitations," she said. "To me the big gap is in the complexity of the human body — in all the cell types and how they change over time in development. And all that data is missing." Whether algorithms can capture the genome's combinatorial, contextual logic, she added: "I don't know."

Adrian Woolfson, a biologist and cofounder of California-based biotech Genyro, pushes the critique further. He calls the full information cloud — genome plus diet, environment, microbiome, and culture — the "informiome." "While the human genome forms the foundation of the human informiome, other layers of extra-genetic information are equally important," Woolfson wrote in his book On the Future of Species, published in April 2026. Genomic foundation models cannot predict all consequences of mutations, he argued, because the relevant information is not in the genome sequence to begin with.

Takeaway: sequence-only models are missing data that never lived in the sequence.

What is the genome, then?

The best metaphors now come from biology itself. In 2020, the historian Evelyn Fox described the genome as "an exquisitely sensitive reactive system" — "a device for regulating the production of specific proteins in response to constantly changing signals it receives from its environment." That echoes geneticist Barbara McClintock's 1983 Nobel address, in which she called the genome "a highly sensitive organ of the cell, monitoring genomic activities and correcting common errors, sensing the unusual and unexpected events and responding to them, often by restructuring the genome."

Bickmore remains optimistic that humans can crack it. "We've got a handle on it now," she said. "We might not know the details, but I think the whole field is coalescing now into a framework where we're thinking along similar lines." Her caveat: AI can assist the sense-making, but human reasoning will be needed to discern the fundamental principles.

"McClintock was far more on point than people realized at the time," Adelman said. "What she said was that the genome isn't static — it's living."

The so-what for the AI sector is blunt: models like AlphaGenome may deliver useful predictions within limits, but until the missing developmental and cell-type data exists — and until model builders grapple with a genome that restructures itself in real time — full understanding stays outside the black box.