Researchers at the Stowers Institute for Medical Research, a nonprofit basic-research institute in Kansas City, published a new method on August 25 for figuring out what an AI model has actually learned when it makes predictions from DNA sequence, rather than just trusting that its predictions are correct. The method, called PISA, is described in a paper in Nature Communications, and it led the team to an unexpected finding about how the genome organizes itself in three dimensions. There is no company, ticker, or product behind this release; it is a piece of basic science from a nonprofit research institute, relevant to readers tracking the broader AI-and-genomics landscape rather than a specific investment.

The Black Box Problem

AI models trained on DNA sequence data have gotten good at predicting things like where a particular protein will bind, or how tightly DNA will be packaged in a given stretch of the genome. What those models typically cannot do is explain why they made a particular prediction, a limitation researchers call the black box problem. Existing interpretation tools could show which DNA bases mattered most to a prediction, but they did so by collapsing each base’s influence into a single number, which meant that positive and negative effects on nearby predictions could cancel each other out and disappear from view entirely.

What PISA Does Differently

PISA, short for pairwise influence by sequence attribution, avoids that collapsing step. Instead, it traces a model’s prediction at one specific position in the genome back to every other DNA base that influenced it, producing a detailed, base-by-base map of what the model learned rather than just what it predicted. The Stowers team, led by Julia Zeitlinger and first author Charles McAnany, built the tool on top of BPNet, a deep-learning framework the lab developed in 2021 that is now used by researchers elsewhere, working in collaboration with Anshul Kundaje at Stanford University.

Separating Biology From Lab-Technique Bias

The team’s key demonstration involved a common lab technique called MNase-seq, used to map nucleosomes, the structures formed when DNA wraps around proteins called histones to help package it inside a cell. The technique works by using an enzyme that cuts exposed DNA while leaving nucleosome-protected DNA intact, but the enzyme does not cut every DNA sequence equally; it has its own sequence preferences. That means data from this experiment actually contains two overlapping signals mixed together: real information about nucleosome positioning, and a technical artifact from the enzyme’s own chemical preferences. A model trained on this data learns both signals at once, without distinguishing between them.

Because PISA preserves fine-grained detail instead of collapsing it, the researchers could see the enzyme’s bias as a distinct, recognizable pattern on their maps. They used that pattern to build a separate model of the bias alone, then mathematically subtracted it from the original model, leaving a second model that had learned only the underlying nucleosome biology.

An Unexpected Discovery

Once isolated that way, the biology-focused model revealed DNA sequences that help position nucleosomes, with effects reaching hundreds of base pairs in either direction. Many of those sequences turned out to be asymmetric, meaning they influenced one side differently than the other, and following that asymmetry led the researchers to a bigger structural feature: the boundaries of larger 3D chromatin domains, the regions of DNA folding that help determine which regulatory sequences can physically reach which genes.

Those domain boundaries are normally identified only through expensive, sequencing-intensive 3D mapping experiments. The Stowers team found thousands of them directly from nucleosome-positioning data alone, in some cases more precisely than the standard 3D methods allow, and they did not stop at the computational finding: they used their bias-corrected model to design new DNA sequences predicted to arrange nucleosomes in specific patterns, then tested a subset of those designs in the lab. The designs worked as predicted, giving the team direct experimental evidence that the patterns the model learned reflect real, testable biology rather than just a statistical quirk of the training data.

Why This Might Matter Down the Line

The broader motivation, according to Zeitlinger, is that most genetic variation linked to disease sits not inside genes themselves but in the regulatory DNA that controls when and where genes turn on, and figuring out what a specific variant actually does remains extremely difficult. A tool like PISA does not identify a drug target or diagnose a disease on its own, but it can help researchers propose a specific mechanism for why a given DNA variant matters, which then guides which experiments to run next. That is a stated long-term aspiration from the lab rather than a near-term product or clinical claim, and Zeitlinger described it as a goal she hopes to make progress on rather than a problem already solved.

What’s Next

PISA has already spread beyond Zeitlinger’s own lab: a collaborator built a separate software implementation of the method, and Stowers neuroscientist Neşet Özel has adopted it for different biological questions, suggesting the approach generalizes beyond the specific chromatin problem described in this paper. The work was funded through Stowers’ own institutional support rather than outside grants or corporate sponsorship. This article covers a peer-reviewed basic-research publication and is not investment advice, since no publicly traded company or commercial product is involved.

Sources

Stowers Institute for Medical Research: Stowers scientists develop a new way to visualize what AI models learn from DNA, and discover how to control what the models learn next, PRNewswire, August 25, 2026.

Nature Communications: the peer-reviewed PISA study, published August 2026, cited as the primary scientific source for this article.

Editorial Disclosure

This article is based on a press release issued by the Stowers Institute for Medical Research on August 25, 2026, distributed via PRNewswire, describing a study published in Nature Communications, cited above. The Stowers Institute is a nonprofit research organization; no publicly traded company, product, or security is discussed in this article, and BioTech Stocks Daily was not compensated for this coverage. Statements regarding the long-term potential applications of this research are the researchers’ own stated aspirations rather than near-term clinical or commercial claims. This article is for informational and educational purposes only and does not constitute investment advice. See our full DISCLAIMER.