Stowers Researchers Decode AI 'Black Box' to Reveal DNA Secrets

Deep-learning models are powerful, but they often function as black boxes, making it difficult for biologists to understand why a model predicts a specific DNA behavior. PISA, or pairwise influence by sequence attribution, changes this by providing a two-dimensional, base-pair-resolution map of the model's internal logic. This high-resolution view enables researchers to distinguish between true biological signals and technical artifacts, such as the enzyme preferences inherent in nucleosome mapping data.

In a study published in Nature Communications, the team demonstrated how PISA could mathematically isolate and remove experimental noise, revealing DNA sequences that organize the genome into 3D chromatin domains. Previously, mapping these boundaries required expensive, sequencing-intensive experiments. By applying this method, the researchers identified specific sequences that position nucleosomes, offering a clearer path from computational prediction to experimental design. This breakthrough, led by Julia Zeitlinger and Charles McAnany, turns predictive models into active discovery tools, helping scientists formulate hypotheses about genetic disease and gene regulation that can be tested directly in living cells.