[ad_1]
Researchers at the Stowers Institute for Medical Research have built an interpretation method that shows, base by base, what a deep-learning model has learned from DNA, and used it to strip a hidden experimental bias out of genomic data. The method, called PISA, is described in a paper published in Nature Communications in August 2026 and announced by the institute on August 25, 2026.
Sequence-to-function neural networks take raw DNA as input and predict the readout of genomics experiments, from transcription factor binding to nucleosome organization. What they cannot ordinarily tell you is why they make a given prediction. PISA, short for pairwise influence by sequence attribution, traces a model’s prediction at one exact genomic position back to every other base that influenced it, producing a two-dimensional map at single-base resolution of what the model learned rather than just what it predicted.
The study was led by Julia Zeitlinger at Stowers in collaboration with Anshul Kundaje at Stanford University, with Charles McAnany, a Stowers AI Fellow, as first author. PISA runs inside BPReveal, the lab’s latest extension of BPNet, a deep-learning framework the team first developed in 2021.
How PISA Separates Experimental Bias From Biology
The team applied PISA to MNase-seq, a widely used assay that maps nucleosomes, the structures formed when DNA wraps around histone proteins. The assay works by using an enzyme that cuts exposed DNA while leaving nucleosome-protected DNA intact, but the enzyme prefers some sequences over others. The data therefore contain two overlapping signals, and the model learned both.
Because earlier interpretation tools collapse each base’s influence into a single value, positive and negative effects can cancel out and disappear. PISA keeps the information at full resolution, and at that resolution the enzyme’s sequence preference showed up as a distinctive fingerprint on the maps. The team extracted that signature mathematically, trained a separate model on the bias alone, and subtracted it, leaving a second model that had learned only the biology.
“It’s a bit like super-resolution microscopy,” Zeitlinger said in the institute’s announcement. “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.”
The Surprise Hiding in Clean Data
Inside the bias-corrected model, PISA revealed DNA sequences that help position nucleosomes, with effects extending hundreds of base pairs in either direction. Many were asymmetric, influencing one side differently from the other. Following that asymmetry led the team to chromatin domain boundaries, the borders that determine which regulatory sequences can reach which genes. Those boundaries are normally mapped with 3D chromatin methods that demand enormous sequencing depth; the model surfaced thousands of them from nucleosome data alone, often more precisely than the 3D data allow, according to the paper.
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, evidence that the rules the model learned can generate testable hypotheses rather than merely describe existing data.
The work lands in a field that has been investing heavily in ever-larger sequence models. Google DeepMind’s AlphaGenome, published earlier in 2026 and cited in the PISA paper’s introduction, predicts the effects of regulatory variants across the genome. PISA addresses the complementary problem: once a model makes such predictions, understanding which sequence features it actually used. The method has already spread beyond the Zeitlinger lab, implemented in a separate software package by a collaborator and adopted by Stowers neuroscientist Neşet Özel for a different biological question.
PISA by the Numbers
- 2021 – the BPNet deep-learning framework, PISA’s foundation, first developed by the team
- April 8, 2025 – PISA preprint first posted to bioRxiv
- August 2026 – peer-reviewed publication in Nature Communications
- Hundreds of base pairs – the reach of individual nucleosome-positioning sequences the models uncovered
- Thousands – chromatin domain boundaries identified from nucleosome data alone
The Limits the Authors State
This is a methods and genomics paper, and its disease relevance is a direction, not a result. Most disease-associated genetic variation sits in regulatory DNA rather than genes, and Zeitlinger is explicit that placing a variant in a binding site or at a domain boundary proposes a mechanism without producing a drug. The work also required a lab fluent in both deep learning and experimental biology, and Zeitlinger names that dual expertise as the field’s persistent gap rather than computational power.
The designed DNA sequences were validated only for a subset tested experimentally, and the paper’s bias-correction demonstration is specific to MNase-seq, though the authors show PISA applied to several data types. The preprint’s revision history shows the chromatin-domain analysis was added during peer review, after a revised version posted on January 5, 2026.
What the work does establish now is a way to audit what genomic models absorb from their training data, correct the parts that came from the experiment rather than the biology, and extract sequence rules precise enough to design and test against living systems.
[ad_2]
Source_link
