AI Foundation Models Still Can't Decode Gene Regulation

Artificial intelligence models trained to read the human genome can make useful predictions about DNA, but they cannot yet explain how the genome actually regulates itself, according to an account published by Quanta Magazine and reposted by Wired on Oct. 17, 2026. The piece, by science writer Philip Ball, reports that the deeper machinery of gene regulation remains too complex and too poorly measured for current AI systems to model in any mechanistic way.

Why hasn't sequencing the genome explained how it works?

Researchers have had the full order of the roughly 3 billion chemical building blocks in human DNA since the international Human Genome Project finished its work between 1990 and 2003, Quanta reported. That milestone did not resolve the central mystery of biology. The sequencing effort showed that barely 2 percent of the human genome consists of genes, the stretches of DNA that code for proteins, according to the report. The rest does not carry instructions in any straightforward sense, leaving biologists with a puzzle that Quanta describes as getting harder the closer scientists look.

What is the harder question biologists are now chasing?

The report frames the real challenge as regulation, not sequence: how genes are switched on and off so that the same genetic material can produce every different cell type in the body. Cells also adjust their gene activity moment to moment in response to signals from neighboring cells and their environment, per Quanta. That regulatory system, the report states, may involve much of the genome beyond the small fraction that codes for genes, and some biologists quoted in the piece question whether a complete understanding of it is within reach at all.

What are AI genomic models actually trying to do?

To tackle that complexity, researchers have built what the industry calls genomic "foundation models," including Evo 2, Genos, and Google DeepMind's AlphaGenome, Quanta reported. These systems train on large volumes of genomic data to predict how changes in DNA sequence affect biological processes and, ultimately, traits such as disease risk, according to the report. The appeal is that the models do not need to explicitly solve the regulatory puzzle. Instead, they are built to infer statistical correlations from cases already documented in their training data, treating the regulatory machinery as something the algorithm absorbs rather than something a scientist has to diagram.

Why do researchers say that's not enough?

Quanta's reporting is direct about the limits of that approach: a computational black box will not satisfy anyone who wants to actually understand how the genome, and by extension life, works. The deeper problem, according to the report, is that the genome may not behave like the kind of blueprint or algorithm that an input-output AI model assumes it to be. The genome is the product of roughly 4 billion years of evolution, the report notes, and its structure reflects that history rather than any engineered logic.

Harvard Medical School biologist Karen Adelman, who studies gene regulation, put it plainly in the piece: "Our genome is not what we might make it if we sat down at the drawing board." That remark captures the gap the report identifies between how engineers build predictive systems and how evolution actually assembled the genome over billions of years, without a design document.

The report also points to a largely unmapped layer of complexity: the three-dimensional folding and physical knotting of DNA inside the cell, which it describes as something researchers have "only scratched the surface" of understanding. That spatial structure appears to influence which genes get activated and when, adding a dimension that sequence-based AI training data does not fully capture, according to the report.

What should readers watch for next?

• Whether genomic AI tools move beyond one-dimensional sequence data to incorporate the three-dimensional folding structure that Quanta's report says shapes gene regulation.
• Whether AlphaGenome, Evo 2, and Genos are tested against known regulatory outcomes in ways that reveal where their predictions break down, rather than being treated as self-contained black boxes.
• Whether biologists manage to generate the kind of detailed regulatory data that current AI training sets lack, since the report frames data scarcity, not just model design, as a core obstacle.

The original reporting, published by Quanta Magazine and syndicated by Wired, does not offer a timeline for when AI models might close that gap. It frames the current moment as one in which both human researchers and machine learning systems are working with the same shortage: a genome that resists being read like a script.