Executive Overview
In the rapidly evolving landscape of computational biology, deep-learning models have emerged as revolutionary tools capable of predicting complex genomic functions straight from raw DNA sequences. From identifying transcription factor binding sites to mapping the intricate architecture of nucleosomes, these sequence-to-function neural networks can accurately forecast the outcomes of high-throughput genomics experiments. Yet, a fundamental limitation has long plagued the field: these models function as "black boxes." While they can tell researchers what is going to happen, they struggle to explain why they made a specific prediction.
This opacity has created a critical vulnerability in genomic artificial intelligence. When neural networks are trained on biological datasets, they do not just learn genuine biological rules—they also inadvertently ingest, memorize, and amplify technical noise and experimental biases inherent to the laboratory assays used to generate the data.
To solve this grand challenge, an interdisciplinary team of researchers at the Stowers Institute for Medical Research, in collaboration with Stanford University, has developed a groundbreaking interpretation method known as PISA (Pairwise Influence by Sequence Attribution). Published in Nature Communications in August 2026, PISA offers an unprecedented window into the inner workings of genomic AI. By tracing a model’s prediction at a single genomic position back to every individual base that influenced it, PISA constructs a two-dimensional, single-base resolution map of what the neural network has actually learned.
More than just an visualization tool, PISA has proven capable of performing a surgical operation on genomic data: isolating technical experimental bias, mathematically stripping it from the model, and leaving behind a pure representation of underlying biology. This capability allows scientists to audit AI models, uncover hidden structural elements in the human genome, design synthetic DNA sequences that manipulate cellular architecture with high precision, and bridge the stubborn gap between computational prediction and experimental biology.
Detailed Chronology: The Journey to PISA
The Genesis of Deep Genomic Modeling
The groundwork for PISA began years prior with the development of BPNet, a deep-learning framework engineered by the Zeitlinger Lab at the Stowers Institute in 2021. Designed to model transcription factor binding and chromatin organization from raw sequence data, BPNet represented a massive leap forward in predictive genomics. However, like other advanced sequence-to-function models—including Google DeepMind’s massive AlphaGenome architecture unveiled earlier in 2026—BPNet and its successors faced an interpretation bottleneck.
Traditional interpretation tools relied on collapsing the influence of each DNA base into a single, one-dimensional value. While this approach provided a crude summary of sequence importance, it suffered from a fatal flaw: positive and negative regulatory effects would often cancel each other out, completely disappearing from view. Furthermore, these methods lacked the resolution necessary to distinguish between true genomic signals and the artifacts introduced by experimental assays.
Building BPReveal and the PISA Breakthrough
Recognizing these limitations, the Stowers research team—led by principal investigator Julia Zeitlinger and spearheaded by Stowers AI Fellow Charles McAnany, in collaboration with Stanford’s Anshul Kundaje—set out to build a more powerful auditing system. Their efforts culminated in BPReveal, an advanced software extension of BPNet.
Within BPReveal, the team integrated PISA. Instead of compressing base-pair influences into a single metric, PISA calculates pairwise dependencies across the entire sequence. It maps out how base $A$ at position one influences base $B$ at position fifty, creating a comprehensive 2D matrix of single-base resolution interactions.
The team first deployed PISA on MNase-seq data, a widely utilized biochemical assay designed to map nucleosomes—the fundamental units of chromatin where DNA wraps tightly around histone protein spools. MNase-seq operates by applying an enzyme that cleaves unprotected, exposed linker DNA while leaving nucleosome-bound DNA intact. However, biochemists have long known that this enzyme possesses an intrinsic sequence preference: it cuts certain DNA motifs much more efficiently than others, regardless of whether a nucleosome is actually present.
Consequently, raw MNase-seq datasets contain a confounding mixture of true biological nucleosome positioning and the chemical bias of the cutting enzyme. Because previous interpretation models lacked fine-grained resolution, they inadvertently learned both signals, conflating technical artifacts with biological reality.
Unmasking and Eliminating the Artifact
When PISA was unleashed on the MNase-seq models, the enzyme’s sequence preference did not cancel out or vanish. Instead, freed from the constraints of 1D compression, the bias appeared as a distinct, highly structured fingerprint across PISA’s two-dimensional interaction maps.
Armed with this discovery, the researchers executed a clean computational maneuver:
- They extracted the mathematical signature of the enzyme’s sequence bias directly from the PISA maps.
- They trained a separate, dedicated neural network model focused exclusively on predicting and replicating this specific technical bias.
- They subtracted this bias model from the original network, effectively scrubbing the artificial noise away.
The resulting "bias-corrected" model was left with something immensely valuable: a pure, unadulterated map of true biological nucleosome positioning rules. As Dr. Julia Zeitlinger colorfully described the breakthrough in the institute’s official announcement, "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."
Supporting Context & Metrics: Uncovering Hidden Biology
Decoding Chromatin Domain Boundaries
Once the experimental noise was successfully purged from the models, PISA revealed an astonishing layer of hidden biology. Operating on the cleaned models, PISA exposed complex DNA sequence patterns responsible for positioning nucleosomes. These patterns extended hundreds of base pairs in both directions and frequently exhibited pronounced asymmetry—influencing genomic architecture differently on one side than the other.
Intrigued by this structural asymmetry, the researchers followed the signals outward and made a major discovery: the asymmetric patterns mapped directly to chromatin domain boundaries. These specialized genomic borders dictate which regulatory elements can physically interact with specific genes, forming the structural neighborhoods of the human genome.
Traditionally, mapping chromatin domain boundaries has required massive 3D chromatin conformation capture assays (such as Hi-C) that demand extraordinary sequencing depth, high costs, and significant experimental time. In contrast, PISA was able to surface thousands of these boundaries using only 1D nucleosome assay data, often achieving significantly higher spatial precision than traditional 3D methods could provide.
From Observation to Synthetic Design
In science, true understanding is often validated not just by observation, but by creation. To test whether the rules uncovered by PISA were genuinely predictive rather than merely descriptive, the Stowers team used their biology-focused models to design entirely synthetic DNA sequences.
These artificially generated sequences were engineered with the specific intent of forcing nucleosomes to arrange themselves into precise, predetermined spatial configurations. When the researchers synthesized these DNA strands and tested them in living systems, the computational predictions held true. This validation confirmed that PISA does more than just reverse-engineer existing data—it extracts foundational biological grammar robust enough to generate testable, real-world hypotheses.
Official Statements and Cross-Disciplinary Impact
The release of the PISA methodology arrives at a critical juncture for computational biology. As laboratories worldwide race to train larger foundational models of the genome—such as Google DeepMind’s AlphaGenome—the demand for robust interpretability tools has skyrocketed. While models like AlphaGenome excel at predicting the downstream functional consequences of regulatory mutations across the genome, PISA tackles the inverse, equally vital problem: auditing how and why the model arrived at its conclusions.
The impact of PISA is already transcending its origins in the Zeitlinger Lab. The software framework has been rapidly adopted by collaborators, integrated into external software repositories, and deployed by Stowers neuroscientist Neşet Özel to tackle entirely different biological questions regarding gene regulation in the developing nervous system.
Despite these successes, the authors maintain a rigorous and realistic perspective on the current boundaries of the technology:
- Association vs. Causation in Disease: The vast majority of disease-associated genetic variations reside not within protein-coding genes, but within non-coding regulatory DNA. While PISA can successfully place a genetic variant within a transcription factor binding site or a chromatin domain boundary, Zeitlinger emphasizes that this structural placement proposes a mechanistic hypothesis—it does not instantly produce a therapeutic drug.
- The Interdisciplinary Talent Gap: According to the Stowers team, the most persistent bottleneck in modern computational biology is not a lack of compute power or advanced hardware, but rather the relative scarcity of researchers who are natively fluent in both deep-learning engineering and hands-on experimental molecular biology.
- Assay-Specific Validation: While PISA’s foundational architecture applies broadly across genomic modalities, the specific bias-correction protocols demonstrated in the study were tailored to MNase-seq assays. Future applications will require similar rigorous auditing across alternative sequencing platforms.
Future Outlook: The Dawn of Transparent Genomic AI
The publication of PISA in Nature Communications marks a philosophical and technical maturation in computational genomics. For years, the field risked careening toward an over-reliance on opaque "black box" neural networks that could predict biological phenomena without granting researchers true mechanistic insight.
By providing a method to peer inside these complex architectures at single-base resolution, PISA bridges the gap between machine learning and empirical biology. It proves that artificial intelligence can be audited, that technical artifacts can be mathematically separated from genuine living signals, and that computational models can be used to write synthetic biological instructions that function reliably in the physical world.
As PISA continues to propagate through academic and industrial research pipelines, it paves the way for a new generation of transparent, bias-free genomic AI. By illuminating the hidden grammar inscribed within our DNA, tools like PISA are transforming genomics from a science of empirical observation into an exact science of programmable biology—bringing humanity one step closer to fully decoding the instructions of life.
