Researchers say AI still struggles to predict how human DNA works

In short: Scientists say AI can help with genetics, but it is not ready to fully explain or predict how the human genome works.

What's going on

Researchers have long talked about DNA as if it were a blueprint or a computer program. A Wired report, reprinted from Quanta Magazine, argues that this metaphor breaks down for humans. Even after the Human Genome Project mapped the order of about 3 billion DNA “letters,” it became clear that only about 2 percent is made of genes, the parts that directly contain instructions for making proteins.

The harder problem is regulation, meaning how cells decide which genes to use, when, and how strongly. This depends on many moving parts. For example, proteins called transcription factors often work in groups, and DNA regions called enhancers can control genes from far away. The genome also folds in 3D inside the cell (like a long yarn ball that is constantly being reshaped), which affects which genes are reachable and active.

AI systems called genomic “foundation models” are being trained on large DNA datasets to predict how genetic differences might relate to traits or disease risk. Examples mentioned include Evo 2, Genos, and Google DeepMind’s AlphaGenome. The article says these models may be useful, but they often treat DNA like a straight line of text, while real gene control depends on cell type, timing, and the genome’s shifting physical shape.

What to watch

A key question is whether future AI models can combine DNA sequence data with missing context, such as detailed measurements from many cell types over time, plus outside influences like diet, environment, and the microbiome (the microbes living in and on our bodies). Without that broader picture, predictions may work in some cases but fail in others, especially when researchers want explanations and not just answers.

Source: Wired