Your genome is the same in nearly every cell you have, yet a neuron and a liver cell behave nothing alike. Much of that difference comes down to which stretches of DNA are open for business in each cell type. ChromBPNet is a deep learning framework that learns to predict that openness, called chromatin accessibility, directly from the DNA sequence itself. Hand a trained model a region of sequence and it predicts the accessibility signal across that region, position by position, along with how strong the signal is overall. This article explains what ChromBPNet is, where it comes from, how its design works, and what researchers actually do with it day to day.
Chromatin accessibility, in plain terms
DNA inside a cell is not a loose thread floating in space. It is packaged, wound, and covered with proteins, and only some regions are physically exposed at any moment. An accessible region is one where the packaging is open enough for transcription factors, the proteins that switch genes on and off, to land and do their work. Accessibility is therefore one of the main ways a cell controls which genes are active, and it changes from one cell type to another even though the underlying sequence stays identical.
Laboratories measure accessibility with assays such as ATAC-seq and DNase-seq. These experiments scan the whole genome and return a signal: high where the DNA is open, low where it is closed. The results are usually summarized as peaks, the regions where signal piles up. That experimental signal is what ChromBPNet learns to predict from sequence alone, which means a trained model can estimate accessibility for DNA it has never seen measured, including sequence carrying a person’s genetic variants.
What a ChromBPNet model actually predicts
Many sequence models output a single number per region, a rough yes or no for activity. ChromBPNet does something more detailed. For each input region it predicts two things:
- A base-resolution profile, the predicted shape of the ATAC-seq or DNase-seq signal across the region, position by position, so you can see where inside the region the accessibility concentrates.
- A total read count, a prediction of how much signal the region carries overall, which captures how accessible it is in total.
The profile matters because biology lives in the details. Two regions can carry the same total signal while concentrating it in very different places, and those places often correspond to where transcription factors actually bind. Predicting the shape, rather than just the amount, is what makes the model useful for interpretation later, when you ask which short sequence patterns inside a region are driving the openness.
Where ChromBPNet comes from
ChromBPNet was built by the Kundaje lab at Stanford and published in Nature Genetics in 2025. It is open source, with the code public on GitHub under the kundajelab/chrombpnet repository, and it is written in Python on top of TensorFlow. If your group already works in Python-based genomics workflows, it slots into familiar territory.
The framework is a descendant of the lab’s earlier BPNet approach. BPNet showed that a neural network could predict base-resolution signal profiles for transcription factor binding experiments. ChromBPNet takes that core idea and adapts it to a harder problem: genome-wide accessibility data, which is far larger in scale and considerably noisier than a single binding experiment. That adaptation is not just a matter of training a bigger network. It involved rethinking how the model handles the systematic quirks of accessibility assays, which brings us to the design feature the framework is known for.
The bias problem, and the two-part design
Accessibility assays have a built-in distortion. The enzymes used to probe open DNA do not cut every sequence with equal enthusiasm. In ATAC-seq, the Tn5 enzyme prefers some local sequences over others, so part of the measured signal reflects the enzyme’s taste rather than genuine openness. A naive model trained on raw signal happily learns that enzyme preference and mistakes it for biology.
ChromBPNet is built to keep those two things apart. Training happens in two parts. First, a bias model learns the enzyme’s sequence bias from genomic regions that carry no real accessibility signal, so whatever pattern it finds there is bias by construction. Then the main accessibility model learns the remaining biological signal with the bias accounted for. The result is called a bias-factorized model. Because the main model never has to spend its capacity explaining enzyme preference, the patterns it learns are much more likely to be real regulatory biology, and that pays off directly when you interpret the model or use it to judge a variant.
How researchers use it in practice
A ChromBPNet model is always tied to a cell type or sample, because accessibility patterns differ so much between cell types. Training one needs three ingredients from an ATAC-seq or DNase-seq experiment in that cell type: the peak regions, the reference genome sequence, and the accessibility signal itself. Training is compute-heavy and normally runs on a GPU.
Once a model exists, three uses dominate:
- Prediction: estimate base-resolution accessibility profiles for new regions without running an experiment.
- Interpretation: use contribution scores and TF-MoDISco to find the sequence motifs the model relies on, which points at the transcription factors shaping accessibility in that cell type.
- Variant scoring: compare predictions for a reference sequence and the same sequence carrying a variant, such as a SNP, to estimate how that variant changes accessibility. This is a popular way to make sense of noncoding variants from genome-wide association studies.
If you do not want to train anything, the project also shares pretrained models for several cell types from the ENCODE project, so you can start with predictions and variant scoring straight away.
Who it is for, and what it is not
ChromBPNet is aimed at computational biologists and genomics researchers who work with accessibility data and want a sequence-based view of it. It rewards groups that already generate or download ATAC-seq or DNase-seq data and want to ask sharper questions of it: which motifs drive this peak, which variant in this locus is the likely functional one, what would happen to this region if the sequence changed.
It is not a general-purpose genome browser or a push-button answer machine. A model is only as good as the experimental data behind it, and its predictions describe one cell type at a time. Treat it as a rigorous instrument: powerful in the hands of someone who understands the assay, misleading if you feed it the wrong cell type or skip the bias step and then trust the motifs it reports.
Conclusion
ChromBPNet predicts chromatin accessibility from DNA sequence, at base resolution, with a design that deliberately separates enzyme bias from real biology. Built by the Kundaje lab, published in Nature Genetics in 2025, and open source on GitHub, it extends the BPNet idea from transcription factor binding to the genome-wide, noisy world of ATAC-seq and DNase-seq. Its daily value shows up in three places: predicting profiles for new sequence, interpreting models to recover transcription factor motifs, and scoring how genetic variants shift accessibility. If any of those problems sit on your desk, the setup guide below is the natural place to start.
Read next:
Pingback: Score Variant Effects with chrombpnet: A Guide
Pingback: chrombpnet vs BPNet: What Is the Difference?