AI models trained on DNA face a basic problem: scientists still can't fully explain how human genes are regulated

What the report says

Wired published a long explainer arguing that artificial intelligence is not yet ready to make full sense of the human genome. The piece describes how genomic "foundation models" are trained on huge amounts of DNA data. Biologists use these models to predict how differences in DNA sequence affect biological processes and, in the end, traits such as disease risk.

According to Wired, the models do not try to account for the complicated machinery that controls genes. That complexity is supposed to be absorbed through training, in which the model finds correlations in cases scientists already understand. The article says this approach is likely to be useful. However, it also warns that the genome may not fit the straightforward input-output logic these tools assume.

Why the genome is harder than expected

Wired notes that the international Human Genome Project nearly fully worked out the order of the roughly 3 billion chemical building blocks in human DNA between 1990 and 2003. That work showed that barely 2 percent of the genome is made of actual genes. Humans have around 20,000 genes that hold instructions for proteins.

The bigger question now is how genes are turned on and off, the report says. Different cell types, such as muscle, brain and skin, need different proteins at different times. "The genes are probably not the most interesting part of the genome," Harvard Medical School biologist Karen Adelman told Wired.

Many signals, not one switch

In bacteria, Adelman explained, gene regulation tends to follow an "OR" logic, where one signal flips a gene on or off. In humans the logic is closer to "AND," meaning many signals are combined before a decision is made. Proteins called transcription factors work in pairs or groups and can have opposite effects in different cell types.

Enhancers nobody has fully mapped

Transcription factors gather on stretches of DNA called enhancers, and the genome may contain hundreds of thousands or even millions of them. Each gene can be influenced by many enhancers, and some sit millions of DNA letters away from the gene they control. University of Edinburgh genome biologist Wendy Bickmore told Wired, "It's embarrassing that 25 years after the Human Genome Project, we don't know where all the enhancers are in the genome."

Shape, folding and chemical tags

The report also describes how DNA's physical shape matters. A protein motor called cohesin pulls DNA into loops, and the activity inside cells changes so fast that no two cells are ever quite the same. Chemical tags called epigenetic marks can change how tightly DNA is packed. Wired says it is far from clear whether a systematic "epigenetic code" really exists.

More layers after a gene is read

Even after a gene is copied into messenger RNA, other controls decide whether a protein gets made. Small molecules called microRNAs can silence messenger RNA before it is translated, and they act in groups in overlapping ways. Wired describes this as an extra checkpoint that lets cells adjust to their immediate surroundings.

The article argues that this makes the genome work less like a fixed program and more like an adaptive, responsive process. For readers following AI and technology coverage, the takeaway from the report is that powerful prediction tools and real understanding of biology are not the same thing.

What it means for AI in biology

Wired names Evo 2, Genos and Google DeepMind's AlphaGenome as examples of genomic foundation models. The report suggests these tools may produce useful predictions. However, it adds that for scientists who want to know how the genome actually works, a computational black box will not be enough.

What to watch next

• Whether researchers can locate all the enhancers in the human genome and identify which genes they control, a question Wired describes as still open.
• How distant enhancers pass signals to the genes they activate, which Bickmore called a largely unanswered question.
• Whether a systematic "epigenetic code" exists, which the report says remains far from clear.
• Whether genomic AI models such as Evo 2, Genos and AlphaGenome can handle the genome's complexity using an input-output approach.