AI Still Can't Read the Genome: Wired Explores the Data Gap Holding Back DNA Models

Artificial intelligence has made real progress on protein folding and gene prediction, but the genome itself remains a harder problem than the field's marketing suggests. Wired reports that the DNA packed inside our cells is not a static list of instructions but a constantly shifting, context-dependent system, and the models built to interpret it are working from data that captures only part of that behavior. The gap is not one of computing power or model architecture. It is a data problem, and it is keeping genomic AI from delivering on its promise.

The core issue is that most training data comes from measurements taken after cells have been broken apart and processed, which flattens the dynamic three-dimensional structure DNA takes on inside a living cell. A model trained on that material learns a simplified picture, one that may not predict how a real genome folds, loops or switches genes on and off in response to its environment. Researchers quoted in the piece describe a field still assembling the raw material it needs before the models can be trusted on clinically meaningful questions.

That matters because the stakes attached to genomic AI keep rising. Drug developers and academic labs are pouring money into foundation models for DNA, betting that better predictions will speed up target discovery and reduce the cost of experiments. If the underlying data misses how the genome actually operates, those bets rest on shaky ground, and models can appear accurate on benchmarks while failing on the biology that matters. The result is a widening gap between published performance numbers and real-world usefulness.

The comparison to protein AI is instructive. Protein models advanced quickly because high-quality structural data became available at scale through shared databases and experimental efforts. Genomic data has no equivalent backbone: it is fragmented across labs, generated with inconsistent methods, and often collected without the cellular context that determines what the sequence actually does. Fixing that requires coordinated data collection and standards, work that moves slower than model releases and attracts less attention.

Wired's reporting lands as several companies and research groups race to ship genome-scale models, sometimes with limited disclosure of what their training sets contain. For now, the practical takeaway is caution: genomic predictions are useful for narrowing hypotheses, not for replacing experiments. The next thing to watch is whether new efforts to capture live-cell DNA structure at scale produce datasets that models can actually learn from, and whether the field will agree on ways to measure whether they do.

The open question is who pays for that data. Public consortia built much of the foundation for protein AI, and a similar push for dynamic genomic measurements would be expensive and slow. Without it, the models will keep getting bigger while learning from a thinner slice of biology than their benchmarks imply.