Genomic AI like Evo 2 and AlphaGenome can't explain the genome, Quanta argues

The piece opens by challenging a long-standing idea: that DNA, since its structure was deduced in the 1950s, is a blueprint, script or computer program holding the secret of life. The article says biologists have yet to find any such greater secret. The human genome is described as less a script than a puzzle that gets harder the closer researchers look. The Human Genome Project (1990 to 2003) nearly fully deduced the sequence of the roughly 3 billion chemical building blocks, but that did not help as much as hoped. It showed that barely 2 percent of the genome consists of actual genes, the information-coding sequences.

The article's central point is that understanding the genome is no longer about working out what each gene does. The harder question is how genes are regulated, which seems to involve some, and perhaps much, of the rest of the genome. Switching suites of genes on and off lets one set of DNA produce many cell types, and cells also adjust their genes moment to moment in response to signals from neighbours and surroundings. The processes are proving so complex that some biologists wonder whether a full understanding will ever be within reach.

Some are counting on AI. The article names genomic foundation models such as Evo 2, Genos and Google DeepMind's AlphaGenome. They are trained on vast quantities of genomic data, and biologists use them to predict how differences in DNA sequence affect biological processes and ultimately the traits, including disease risk, of a whole organism. These models do not worry about the regulatory detail; that is supposedly subsumed by training, through which the algorithm deduces correlations from known cases. The author's verdict: the approach is likely to be useful, but for anyone who craves real understanding, a computational black box will never suffice. The genome might also not submit to the straightforward input-output approach such models ultimately assume.

The rest of the visible text builds the case for why. The genome is the product of around 4 billion years of evolution. Karen Adelman, a biologist who studies gene regulation at Harvard Medical School, says: "Our genome is not what we might make it if we sat down at the drawing board." The traditional picture still holds roughly: about 20,000 genes, each holding the instruction for a protein, with sequences from a couple of dozen to almost 3 million DNA letters. A polymerase copies the gene into messenger RNA (transcription), and the ribosome builds the protein from it (translation). But Adelman says the genes are probably not the most interesting part of the genome.

Regulation has been known since the 1960s, but most early knowledge came from bacteria, where the principles are straightforward. In complex organisms there are overlapping layers of control. Transcription factors, proteins that recruit the polymerase, do not fit unique binding sites like keys in locks. They show weak preferences for DNA sites, tend to work in pairs or groups, and can activate a gene in one cell type and suppress it in another depending on which other factors are present. Adelman says bacterial regulation tends to follow an "OR" logic, while human regulation is more like "AND": many signals are integrated, and the knobs are tunable rather than simply on or off. "This is part of the beauty," she says.

Transcription factors bind to enhancers, which are thought to be the decisive influence on transcription. The genome has hundreds of thousands, perhaps millions, of them, far more than genes. Each gene might be influenced by many enhancers, and each enhancer might influence several genes. Wendy Bickmore, a genome biologist at the University of Edinburgh, calls it embarrassing that 25 years after the Human Genome Project, researchers do not know where all the enhancers are, what they do, or which genes they control. Most enhancers need a "cocktail" of factors, which gives them specificity to a particular cell at a particular time. Some enhancers sit millions of nucleotides from their gene, with other genes in between. Bickmore says such distal enhancers "seem bonkers" and that how the signal travels is a largely unanswered question.

One partial answer is looping. A protein motor called cohesin extrudes loops of chromatin (DNA plus its packaging proteins) that bring distal enhancers to their genes. The components do not assemble into a molecular machine. They form a loose but dense blob, sometimes called a condensate, in which they interact weakly, fleetingly and indiscriminately. These hubs are extremely fluid and differ from cell to cell, so even notionally identical cells, two skin cells say, never have quite the same regulatory activity at any moment.

Chromatin structure matters too. Outside cell division it is unwound into what looks like a tangled mess with order inside it: tightly packed heterochromatin, whose genes are typically silenced, and looser, accessible euchromatin. Enzymes repackage it, so what matters is not just the encoded information but how DNA exists physically in space. Bickmore says researchers have stopped thinking of the genome as a linear piece of DNA code, and she calls the dynamic 3D folding inherent to regulation a very exciting change. Chromatin also clusters into topologically associating domains (TADs), within which genes seem to be coregulated and switched on or off in groups; cohesin helps build them. Epigenetic marks, small molecules attached to histones or directly to DNA, can change chromatin packing and act like annotations that alter the DNA's meaning in context; they are copied when cells divide. How and when the marks are added, and what each means, has no simple answers, and whether a systematic "epigenetic code" exists is far from clear.

The article then turns to a further layer, post-transcriptional regulation, which determines whether mRNA is translated into protein and which protein results. The available text ends there, partway through that section.

Key facts

• The article argues the human genome is a regulatory puzzle rather than a blueprint or algorithm: barely 2 percent of it consists of actual genes, and the harder question is how roughly 20,000 genes are regulated.
• Genomic foundation models such as Evo 2, Genos and Google DeepMind's AlphaGenome are trained on vast genomic data to predict how sequence differences affect traits, including disease risk; their training is supposed to subsume the regulatory detail.
• The author's conclusion is that this approach is likely to be useful, but a computational black box will never suffice for real understanding, and the genome might not suit a straightforward input-output approach.
• Regulation involves hundreds of thousands, perhaps millions, of enhancers (more than genes), some millions of nucleotides from their targets, plus cohesin-made chromatin loops, 3D folding, TADs and epigenetic marks.
• Wendy Bickmore of the University of Edinburgh calls it embarrassing that 25 years after the Human Genome Project, researchers still do not know where all the enhancers are or what they control.

Why it matters

The piece reframes what the hard problem in genomics is. After the Human Genome Project delivered the sequence, the open question is not what each gene does but how genes are switched on and off, and the article says that may involve much of the non-gene portion of the genome. That bears directly on genomic foundation models, which skip the mechanisms and learn correlations from known cases. The article's position is that they will probably help with prediction but cannot deliver the kind of understanding some biologists want. No new model or result is announced; this is an explanatory feature.

Who it affects

Mainly biologists who study gene regulation and those counting on AI to analyse the genome. The article also touches the use of such models to predict how DNA sequence differences affect traits, including disease risk. Developers of models such as Evo 2, Genos and AlphaGenome appear only as named examples; none is quoted.

How to use it

There is nothing to install or run. It is a long read that works as a primer on gene regulation: transcription factors, enhancers, cohesin loops, chromatin states, TADs and epigenetic marks. For anyone judging claims about genomic AI, it supplies the author's argument that predictive accuracy and mechanistic understanding are different things.

How solid is it

The original appeared in Quanta Magazine and is republished by Wired. The biology is explained through two named experts: Karen Adelman, who studies gene regulation at Harvard Medical School, and Wendy Bickmore, a genome biologist at the University of Edinburgh. Their quotes concern gene regulation and enhancers, not the AI models. The verdict on AI is the article's own voice, not attributed to either of them. No benchmarks or test results for Evo 2, Genos or AlphaGenome are given. Several claims are hedged in the text itself: enhancers are "thought" to be decisive, genes "seem" coregulated within TADs, and the genome "might not" suit input-output modelling. The copy available here stops partway through the section on post-transcriptional regulation, so the article's closing argument is not covered.

Risks and caveats

The AI argument takes up only a few paragraphs of a much longer biology explainer, and it is a viewpoint rather than an experiment. The article says the AI approach is likely to be useful, so it should not be read as calling the models worthless. It sets no timescale for when, or whether, regulation will be fully understood. Because the text is cut off, anything the article says later about what the genome is instead of a blueprint is missing here.

“Our genome is not what we might make it if we sat down at the drawing board”

— Karen Adelman, biologist who studies gene regulation at Harvard Medical School