The AI model interpretation method, PISA, shows what genomic AI models learn, allowing scientists to separate experimental bias from biology, train more focused models, and uncover unexpected insights related to gene regulation and genetic disease.

What’s new, and why it matters:

PISA (pairwise influence by sequence attribution) is a new method that traces a deep-learning model’s prediction at any single DNA base back to every other base that influenced it, producing a base-pair-resolution map of what the model learned, not just what it predicted.

Applying PISA to nucleosome mapping data, the team spotted and mathematically removed a technical bias baked into the data, revealing DNA sequences that position nucleosomes and, unexpectedly, mark the boundaries of larger 3D chromatin domains, normally identified only through expensive, sequencing-intensive methods.

Biologists have long been interested in better understanding how specific DNA sequences relate to gene regulation and genetic disease. PISA can point to potential sequence elements and mechanisms that warrant further investigation.

KANSAS CITY, Mo., Aug. 25, 2026 /PRNewswire/ — Artificial intelligence can predict how a stretch of DNA will behave inside a cell. What it usually can’t tell you is why it will behave that way. Researchers at the Stowers Institute for Medical Research have developed a new method that can.

The method gives scientists a high-resolution view of what a model learned and allows them to connect its predictions back to the DNA sequences that drive them. In the lab of Julia Zeitlinger, Ph.D., that new level of detail led to something scientists had not been able to see before: specific DNA sequences that help explain how the genome is organized in three dimensions.

PISA (pairwise influence by sequence attribution) traces a model’s prediction at one exact position in the genome back to every DNA base that influenced it. The result is a two-dimensional map at a single-base resolution of what the model learned, providing a new way to study the 3D organization of chromatin, which normally requires expensive, more challenging experiments.

The study, published in Nature Communications in August 2026, was led by Zeitlinger in collaboration with Anshul Kundaje, Ph.D., at Stanford University, and first author Charles McAnany, Ph.D., who is also an AI Fellow at the Institute. Together, they combined their expertise in computation and biology to build upon BPNet, a deep-learning framework developed by the team in 2021 and used today by researchers worldwide. PISA operates through BPReveal, the lab’s latest extension of BPNet.

“We’re pioneering how to bring deep learning to biologists to better understand biology in a systematic way,” Zeitlinger said. “We want to understand what the models learn and connect it to biological mechanisms.”

Zeitlinger, who leads the Institute’s AI Initiative, hopes PISA will better equip biologists to take advantage of already trained models to generate hypotheses that help inform their next experiments.

“With our DNA models, prediction usually isn’t the goal,” she added. “We already have the data. What we want to know is what the model had to learn to reproduce the data.”

Opening the black box, one base at a time

Deep learning models are powerful precisely because no one tells them what to look for. They find the patterns themselves.

“But that also makes them difficult to interrogate, and that is the black box problem,” Zeitlinger said.

Tools already existed that allowed scientists to peer inside these models and highlight which DNA bases mattered most overall. But because these tools collapse each base’s influence into a single value, positive and negative effects can cancel each other out and disappear entirely. PISA, however, does not collapse this information. The models predict experimental data at every base, allowing the interpretation to be just as fine-grained.

“It’s a bit like super-resolution microscopy,” Zeitlinger said. “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.”

How it works: separating biology from experiment

The team applied PISA to several kinds of genomic data, including MNase-seq, a widely used method for mapping nucleosomes, which are the structures formed when DNA wraps around proteins called histones. MNase-seq works by using an enzyme that cuts exposed DNA while leaving nucleosome-protected DNA intact. But the enzyme doesn’t cut all sequences equally; it prefers some sequences over others. As a result, the data contain two overlapping signals. The model learned both.

“The model got the answer right, but not only for the reason we expected,” McAnany said. “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.”

“Charles was really instrumental in seeing the opportunity where math was really required,” Zeitlinger said.

Seeing the difference between the two kinds of patterns meant the team could separate them. The enzyme’s sequence preference produced a distinctive signature on the PISA maps. The team could extract that signature mathematically, use it to train a separate model of the bias alone, and then subtract it from the original, leaving behind a second model that learned only the biology.

“Once you can see the fingerprint, you can tell the model to ignore it. You’re cleaning up the smudges on the picture,” Zeitlinger explained. “Then you ask, ‘okay, what do we see now?’ And now the model can specifically focus on what we wanted to learn.”

Like a set of Russian dolls

“This paper was a little bit like disassembling Russian dolls, where the more you look, the more information you can reveal,” Zeitlinger said. “We hadn’t even planned on developing PISA. Then we realized we could apply it to all sorts of datasets and see things we couldn’t see before. Then we could separate bias from biology and train a new model. And then we could see something else again.”

Inside the biology-focused model, PISA revealed DNA sequences that help position nucleosomes, the structures that package DNA, 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 chromatin domain boundaries, which determine which regulatory sequences can reach which genes. The boundaries are normally mapped with 3D methods requiring enormous sequencing depth. Earlier studies had suggested a link to nucleosomes; what the model added was the exact sequence information, with thousands of boundaries found in nucleosome data alone, often more precisely than the 3D data allow.

“We discovered that what was organizing these beads on a string is also organizing these larger-scale interactions,” Zeitlinger said. “That was a really nice surprise. What’s unique here is that we have the exact sequences driving it.”

The team then used the biology-focused models to design DNA sequences predicted to arrange nucleosomes in specific ways. They tested a subset of those designs experimentally. The predictions worked, providing evidence that the rules learned by the model can do more than describe existing data. They can help scientists generate and test new hypotheses and ultimately discover new biology.

“Interpretation turns a model from a prediction machine into a discovery tool,” Zeitlinger said. “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.”

“A model may predict experimental data very well, but that does not automatically tell us what it has learned,” McAnany added. “PISA lets us trace a prediction at one precise genomic position back to the DNA sequences influencing it.”

Putting PISA to use: decoding the genome

Most genetic variation associated with disease sits not in genes themselves but in the regulatory DNA that determines when and where genes are turned on. Researchers can identify those variants, but figuring out what they do is extremely difficult.

“We can’t just say that a piece of DNA is doing something,” Zeitlinger said. “We need to understand every base. Because if that base changes, it may have an effect, while another base may have none. To get there, we need predictions that are very precise, and then we need to understand why the model thinks a base matters.”

Knowing that a variant sits in a transcription factor binding site or at a domain boundary, and which cell type it acts in, does not produce a drug. It does, however, propose a mechanism, and that can be very important for deciding what to do next.

“Ultimately, there is a solution to the problem, and it’s a real goal that I hope to help accomplish in my lifetime,” Zeitlinger said.

Bridging a gap

PISA has already been used beyond the Zeitlinger Lab. The method was recently implemented in a separate software package by one of Zeitlinger’s collaborators and adopted by Stowers Institute Investigator and neuroscientist Neşet Özel, Ph.D., demonstrating that the approach can be applied to different biological questions and modeling frameworks.

“But the persistent gap in the field is not computational power,” Zeitlinger explained. “It’s bridging the gap between the researchers who build models and those who study biological mechanisms. Training a deep learning model still demands expertise most biologists don’t have.”

Zeitlinger believes that gap will narrow as AI makes coding and model-building more accessible. If that happens, tools like PISA could give more experimental biologists a way to examine what their models learned and use those insights to guide the next experiment.

“Once people who don’t routinely code can train these models and see these patterns, that’s a game changer,” Zeitlinger said. “Once you see your data at higher resolution, you wouldn’t want to go back.”

Additional authors include Melanie Weilert; Grishma Mehta; Fahad Kamulegeya; Jennifer M. Gardner; Jacob Schreiber, Ph.D.; and Anshul Kundaje, Ph.D. 

This work was funded by institutional support from the Stowers Institute for Medical Research.

About the Stowers Institute for Medical Research

Founded in 1994 through the generosity of Jim Stowers, founder of American Century Investments, and his wife, Virginia, the Stowers Institute for Medical Research is a non-profit, biomedical research organization with a focus on foundational research. Its mission is to expand our understanding of the secrets of life and improve life’s quality through innovative approaches to the causes, treatment, and prevention of diseases.

The Institute consists of 24 independent research programs. Of the approximately 500 members, over 370 are scientific staff that include principal investigators, fellows, technology center directors, postdoctoral scientists, graduate students, and technical support staff. Learn more about the Institute at www.stowers.org and about its graduate program at www.stowers.org/gradschool.

Media Contact:
Joe Chiodo, Director of Communications
724.462.8529
chiodo.joe@stowers.org 

View original content to download multimedia:https://www.prnewswire.com/news-releases/stowers-scientists-develop-a-new-way-to-visualize-what-ai-models-learn-from-dna–and-discover-how-to-control-what-the-models-learn-next-302859867.html

SOURCE Stowers Institute for Medical Research