--- title: Scientists Can Now See What AI Models Learn From Raw DNA description: A method called PISA lets scientists trace what deep-learning models learn from DNA, and tell real biology from experimental artifacts. author: Darie Nani (Editor-in-Chief) updated: 2026-08-26T03:17:19.176Z canonical: https://www.sovereignmagazine.com/article/pisa-ai-models-read-dna-genomics image: https://cdn.nanimediahouse.com/pisa-genomic-ai-dna-207684.webp categories: Science & Tech content_type: News region: Global publication: Sovereign Magazine schema_type: Article --- A team of researchers has developed a way to look inside the artificial intelligence models that read raw DNA, and to tell when one of those models has learned something real about biology rather than a quirk of how the data was made. The method is called PISA, short for Pairwise Influence by Sequence Attribution. It was published in Nature Communications in August 2026 by a group led by Charles McAnany, an AI Fellow at the Stowers Institute for Medical Research in Kansas City, Missouri, with senior author Julia Zeitlinger, an investigator at Stowers, and collaborator Anshul Kundaje of Stanford University. The same collaboration produced BPNet, a 2021 framework for base-resolution genomic deep learning, and PISA ships as part of a software toolkit the group maintains called BPReveal. ## A fuller map of what a model weighs A growing class of sequence-to-function models now predicts regulatory activity, gene expression and chromatin state directly from a stretch of DNA letters. BPNet reads transcription-factor binding at base resolution; DeepMind's Enformer, Calico's Borzoi and DeepMind's AlphaGenome, released in 2025, work on related problems. PISA does not make its own predictions. It interprets models of this kind after they are trained. Given such a model, PISA takes any single base in a prediction and traces it back to every other base that influenced it, producing a two-dimensional map at base-pair resolution. Earlier interpretation methods collapse each base to a single number, which can hide as much as it shows: a base that pushes a prediction up in one context and down in another averages out to nothing, and information about spacing and interaction between positions is lost. PISA keeps the full pairwise detail, so the influence of one base on another stays visible along with where each sits. The pairwise, position-preserving map is the genuinely new part, a real if incremental advance in an active field rather than a paradigm shift. ## A model can be right for the wrong reason A genomic model can be accurate and still be right for the wrong reason. It may have learned a batch effect, or, in data about how DNA is packaged, the sequence preferences of the enzyme used to cut the DNA rather than anything about the genome itself. Until now there was no clear way to tell the two apart from the inside. That is the gap PISA addresses. By mapping which bases a prediction actually rests on, it lets researchers separate experimental bias from real biology and flag when a model has memorized an artifact of the measurement. The payoff reaches beyond model diagnostics. Most of the genetic variants tied to human disease sit not in genes but in the vast non-coding, regulatory stretches of the genome, and [what most of them actually do remains unknown](https://pmc.ncbi.nlm.nih.gov/articles/PMC11044903/). Interpreting the models that read this regulatory DNA is one route to working out what those variants change. ## A nucleosome finding the team then tested in the lab The researchers applied PISA to MNase-seq data on nucleosome positioning. Nucleosomes are the protein spools that DNA wraps around, and how they are placed helps control which genes are switched on. The method identified DNA sequences that position nucleosomes with effects reaching hundreds of base pairs, along with asymmetric sequences that influence the boundaries of chromatin domains, features that normally require expensive three-dimensional mapping to study. The team did not stop at reading the model. Using its predictions, the researchers designed new DNA sequences intended to arrange nucleosomes in a particular pattern, then confirmed in the lab that the sequences produced the arrangement. PISA interprets an existing model and turns its predictions into DNA that behaves as designed when tested; it does not change how a model learns as it trains. Reading a model this way still leaves the confirming to the laboratory, and PISA is a tool for models that already exist rather than a new kind of predictor. Within those limits, it gives researchers a view into genomic deep learning that they did not have before, and a check on whether these systems have learned biology or the fingerprints of the experiments that trained them. ## FAQ **Q: What is PISA in genomics?** PISA, or Pairwise Influence by Sequence Attribution, is a method for interpreting AI models that read DNA. It traces a model's prediction at a single DNA base back to every other base that influenced it, producing a base-pair-resolution map of what the model is responding to. It was published in Nature Communications in August 2026 and is distributed in the BPReveal toolkit. **Q: What does it mean for an AI model to be a black box in biology?** It means the model produces accurate predictions without revealing the reasoning behind them. A genomic model can predict how a stretch of DNA behaves while giving no account of which bases drove the answer, which leaves open whether it learned real biology or an artifact of the data. Methods like PISA are attempts to open that black box. **Q: What is a non-coding genetic variant?** It is a change in a stretch of DNA that does not code for a protein but instead sits in the regulatory regions that control when and where genes are active. Most disease-associated variants fall in these non-coding regions, and the mechanism behind most of them is still unknown. **Q: Can AI predict what a DNA mutation does?** A class of genomic deep learning models predicts regulatory activity, gene expression and chromatin state from DNA sequence, which can suggest what a change might do. Those predictions still need laboratory validation, and interpretability tools like PISA help check that a model's prediction rests on real biology before anyone relies on it.