# AI Method Uncovers What Genomic Models Learn From DNA PISA technique removes hidden experimental bias from genomic data. - **Source:** [Unite.AI](https://www.unite.ai) - **Published:** 2026-08-25 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/d7f2cb8d-f065-4b4b-a249-31769a4a4035 ## Hidden Bias in Genomic Data Researchers at the Stowers Institute have developed PISA, a method that reveals, base by base, what deep learning models learn from DNA. This technique exposes hidden experimental biases in genomic data. PISA traces a model's prediction at a specific genomic position back to every other base that influenced it, creating a single-base resolution map of what the model learned, not just what it predicted. ## Technical Details of the Method PISA runs inside BPReveal, the latest extension of the BPNet deep learning framework. The team applied PISA to MNase-seq, a widely used assay that maps nucleosomes. However, the enzyme used in the assay prefers certain sequences over others, resulting in two overlapping signals in the data that the model learns. Earlier interpretation tools collapsed each base's influence into a single value, potentially canceling out effects. ## Mathematical Extraction of Bias PISA maintains information at full resolution, revealing the enzyme's sequence preference as a distinctive fingerprint. The team mathematically extracted this signature, trained a separate model on the bias alone, and subtracted it, leaving a second model that learned only the biology. Julia Zeitlinger likened PISA to super-resolution microscopy, where adding pixels reveals previously unseen details. ## Asymmetric Effects and Chromatin Boundaries PISA revealed asymmetric DNA sequences that position nucleosomes, with effects extending hundreds of base pairs. This asymmetry led the team to chromatin domain boundaries, which determine regulatory sequence access to genes. Normally mapped with 3D chromatin methods requiring deep sequencing, the model identified thousands of boundaries from nucleosome data alone, often more precisely than 3D data. ## Synthetic DNA and Experimental Validation Using biology-focused models, the team designed synthetic DNA sequences predicted to arrange nucleosomes in specific configurations. A subset was experimentally tested, confirming the predictions. This demonstrated that the model's learned rules can generate testable hypotheses, not just describe existing data. The work complements large sequence models like Google DeepMind's AlphaGenome. ## Current State of the Field This work enters a field heavily investing in larger sequence models. While AlphaGenome predicts regulatory variant effects across the genome, PISA addresses the complementary problem: understanding which sequence features a model actually used. The method has already spread beyond the Zeitlinger lab, implemented by collaborators and adopted by Stowers neuroscientist Neşet Özel for different biological questions. ## Clinical Applications and Limitations This is a methods paper; disease relevance is a direction, not a result. Most disease-associated variation sits in regulatory DNA. Zeitlinger notes placing a variant in a binding site or domain boundary proposes a mechanism without producing a drug. The work requires dual expertise in deep learning and experimental biology, which Zeitlinger identifies as the field's persistent gap, not computational power. ## Future Implications PISA establishes a way to audit what genomic models learn from training data, correct experimental artifacts, and extract precise sequence rules for designing and testing against living systems. The designed sequences were validated only for a tested subset, and bias-correction is specific to MNase-seq, though PISA applies to several data types. Peer review added the chromatin-domain analysis to the preprint in January 2026. ## Sources covering this story - https://www.unite.ai/ai-method-reveals-what-genomic-models-learn-from-dna-and-exposes-hidden-experimental-bias/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._