Scientists at the Stowers Institute for Medical Research have created a breakthrough method called PISA that shows exactly what artificial intelligence models learn when analyzing DNA sequences, revealing patterns that were previously invisible and hidden experimental bias. The technique, published in Nature Communications in August 2026, transforms AI from a prediction tool into a discovery engine by allowing researchers to trace how models make decisions at the genetic level.

What Is PISA and How Does It Work?

PISA stands for "pairwise influence by sequence attribution." Unlike earlier interpretation methods that collapse information into single values, PISA traces a model's prediction at any single DNA base back to every other base that influenced it, producing a two-dimensional map at single-base resolution. Think of it like upgrading from a blurry photograph to a high-resolution image where you can suddenly see details that were always there but invisible before.

The team, led by Julia Zeitlinger at Stowers in collaboration with Anshul Kundaje at Stanford University, built PISA as an extension of BPNet, a deep-learning framework the team first developed in 2021. Charles McAnany, a Stowers AI Fellow, served as first author on the study.

"It's a bit like super-resolution microscopy. Even the earlier interpretation methods opened the black box. Then you realize you can see even more. You're adding pixels and suddenly you're seeing things you couldn't see before," explained Julia Zeitlinger.

Julia Zeitlinger, Director of AI Initiative at Stowers Institute for Medical Research

How Did Researchers Separate Bias From Biology?

The researchers applied PISA to MNase-seq, a widely used technique for mapping nucleosomes, the structures formed when DNA wraps around histone proteins. The assay uses an enzyme that cuts exposed DNA while leaving nucleosome-protected DNA intact, but here's the catch: the enzyme has preferences for certain DNA sequences over others.

This meant the data contained two overlapping signals, and the AI model learned both. Using PISA's high-resolution maps, the team spotted the enzyme's sequence preference as a distinctive fingerprint. They then extracted that signature mathematically, trained a separate model on the bias alone, and subtracted it from the original model, leaving behind a second model that learned only the actual biology.

"The model got the answer right, but not only for the reason we expected. It learned every pattern that helped predict the data. Some reflected nucleosome biology, while others reflected how the experiment was performed. Because PISA visualizes those patterns at high resolution, we were able to disentangle the two," noted Charles McAnany.

Charles McAnany, AI Fellow at Stowers Institute for Medical Research

What Unexpected Discoveries Did PISA Reveal?

Inside the bias-corrected model, PISA revealed DNA sequences that help position nucleosomes, with effects extending hundreds of base pairs in either direction. Many of these sequences were asymmetric, meaning they influenced one side differently from the other. Following that asymmetry led the team to an unexpected discovery: chromatin domain boundaries, the borders that determine which regulatory sequences can reach which genes.

These boundaries are normally mapped using expensive 3D chromatin methods that demand enormous sequencing depth. The PISA-enhanced model surfaced thousands of them from nucleosome data alone, often more precisely than the 3D data allow. The team then used the biology-focused models to design synthetic DNA sequences predicted to arrange nucleosomes in specific configurations, and tested a subset of those designs experimentally. The predictions held, providing evidence that the rules the model learned can generate testable hypotheses rather than merely describe existing data.

How to Apply PISA to Genomic Research

• Audit Model Behavior: Use PISA to examine what patterns AI models actually learned from training data, separating genuine biological signals from experimental artifacts that could mislead downstream research.
• Correct Systematic Bias: Identify distinctive fingerprints of technical bias in high-resolution maps, extract them mathematically, and train separate models to subtract bias from original predictions.
• Design and Test Hypotheses: Use the cleaned models to design synthetic DNA sequences with predicted properties, then validate predictions experimentally to confirm that learned rules apply to living systems.
• Discover Hidden Regulatory Elements: Apply PISA across multiple genomic data types to uncover sequence features and regulatory boundaries that earlier interpretation methods missed due to information collapse.

Why Does This Matter for Medicine and Biology?

The work addresses a fundamental challenge in AI: the "black box" problem. Deep learning models are powerful precisely because no one tells them what to look for; they find patterns themselves. But that also makes them difficult to interrogate. PISA solves this by allowing researchers to understand which sequence features a model actually used to make its predictions.

This has immediate implications for understanding gene regulation and genetic disease. Most disease-associated genetic variation sits in regulatory DNA rather than genes themselves. By understanding exactly which DNA sequences influence gene regulation, researchers can propose mechanisms for how genetic variants might cause disease. While this doesn't immediately produce drugs, it provides the foundation for targeted therapeutic development.

"Interpretation turns a model from a prediction machine into a discovery tool. Once we understand which sequence features are driving the output, we can generate biological hypotheses and design focused experiments to test whether those rules operate in living cells," stated Zeitlinger.

Julia Zeitlinger, Director of AI Initiative at Stowers Institute for Medical Research

The method has already spread beyond the Zeitlinger lab. It has been implemented in a separate software package by a collaborator and adopted by Stowers neuroscientist Neşet Özel for a different biological question, suggesting that PISA could become a standard tool for genomic AI research.

Zeitlinger emphasizes that the field's persistent gap is not computational power but rather the need for labs fluent in both deep learning and experimental biology. PISA represents a step toward bridging that gap by making AI models more interpretable and their insights more actionable for biologists.