AI models trained on genomes can predict how some DNA changes affect biology, but they still cannot provide a complete account of how cells regulate genes. WIRED’s report explains that the obstacle is not simply model size: scientists lack sufficiently rich data about the dynamic processes the models are supposed to learn.
Only about 2 percent of the human genome consists of protein-coding genes. Much of the harder problem lies in how cells switch genes on and off in response to their type, neighbors and environment. Genomic foundation models such as Evo 2, Genos and Google DeepMind’s AlphaGenome learn correlations from known sequences and measurements, which can help prioritize experiments or estimate the impact of a mutation.
Those predictions do not automatically reveal the underlying biological mechanism. Training records capture only a fraction of the relevant cell types, conditions and molecular interactions. A model may therefore perform well on familiar data while failing when a sequence appears in a different cellular context.
The systems can still be useful as scientific tools, especially for narrowing a large search space. But researchers need more diverse experiments and measurements before AI can reliably decode the genome’s regulatory machinery. Prediction and biological understanding remain separate goals.