ChromBPNet did not appear from nowhere. It descends from BPNet, the Kundaje lab’s earlier framework, and the family resemblance is obvious: both are neural networks that read DNA sequence and predict experimental signal at base resolution. Yet they were built for different assays, different scales, and different kinds of noise, and those differences decide which one belongs in your project. This article compares the two honestly, covering their shared roots, the data each was designed for, how ChromBPNet’s bias handling extends the original idea, and a simple rule for choosing between them based on the experiment sitting in front of you.
Shared roots: profiles instead of scores
The philosophical inheritance is the first thing to understand. Most sequence models of their eras predicted a single number per region, a summary score of activity. BPNet broke with that habit by predicting the full signal profile of an experiment, base by base, alongside a total count for the region. The shape of the signal, where exactly within a region the reads pile up, turned out to carry the interesting biology, because that shape reflects where proteins actually contact DNA.
ChromBPNet keeps that commitment wholesale. Its outputs are the same pair, a base-resolution profile and a total count, now applied to accessibility signal. If you have worked with one framework, the outputs of the other will feel immediately familiar, and the interpretation tools built around per-base signal, contribution scores in particular, carry across as well. The lineage is real: ChromBPNet is BPNet’s profile-prediction idea, re-engineered for a different class of experiment.
The data difference: one factor versus the whole genome
The assays they serve are the root of every other difference. BPNet was designed for transcription factor binding data: experiments that map where one specific protein binds across the genome. The data is comparatively focused, the signal concentrated at that factor’s binding sites, and the modeling question is sharp: what sequence determines this factor’s binding, and where exactly does it sit?
ChromBPNet serves chromatin accessibility assays, ATAC-seq and DNase-seq, which report something far broader: every open region in the genome of a cell type, shaped by many factors acting together. The data is genome-wide, the signal noisier, and the territory vastly larger. A model for accessibility must make sense of regions opened by any of dozens of regulators, in data with a heavier background. Same style of prediction, very different assignment, and the younger framework’s design choices all trace back to that difference in assignment.
Scale and noise: why adaptation was needed
Scaling the BPNet idea to accessibility was not a matter of feeding it more data. Genome-wide accessibility brings a noise structure that focused binding experiments largely avoid. The enzyme bias problem is the clearest example. In ATAC-seq, the Tn5 enzyme’s sequence preferences stamp a strong, systematic pattern across the entire dataset, and in genome-wide data that pattern is everywhere, entangled with real signal in region after region.
ChromBPNet’s answer is its bias-factorized architecture: a separate bias model, trained on regions without real signal, learns the enzyme’s preferences, and the accessibility model learns the remaining biology with that bias accounted for. This two-component design is the younger framework’s signature addition, and it exists precisely because accessibility data demands it. It is the difference between adapting an idea and merely enlarging it.
What each is typically used for
In practice the two frameworks occupy neighboring desks in the same lab:
- BPNet is the tool when your experiment profiles a specific transcription factor’s binding and you want base-resolution insight into that factor’s sequence rules and binding positions.
- ChromBPNet is the tool when your experiment profiles accessibility itself, and you want to predict open chromatin from sequence, discover the mix of motifs that create it, or score how variants shift it in a given cell type.
The interpretation stories rhyme. Both frameworks are valued because their models can be opened up: contribution scores highlight the bases driving predictions, and pattern-discovery tools gather those highlights into motifs. In a BPNet model the motifs describe one factor’s binding grammar in fine detail. In a ChromBPNet model the motifs describe the cast of factors whose binding keeps a cell type’s chromatin open, a broader census taken from a single model.
Choosing between them
The choice is made by your assay, not by fashion or by which paper is newer. Ask what your experiment measured:
- Binding of a specific transcription factor: reach for the BPNet approach, designed and validated for exactly that data.
- Chromatin accessibility from ATAC-seq or DNase-seq: reach for ChromBPNet, whose bias factorization and genome-wide design match the data you have.
- Both assays in the same project: the frameworks are complements, and groups running accessibility work alongside factor-specific experiments have sensible reasons to use each where it fits.
What you should not do is force one framework onto the other’s data because it is the one you already have installed. An accessibility model without bias handling will teach you enzyme preferences dressed as motifs, and a binding-focused setup pointed at genome-wide accessibility will struggle with a scale and noise profile it was never shaped for. The tools are specialists; let them specialize.
What carries over if you know one already
Experience with either framework transfers well. The input logic is shared: regions, reference sequence, and measured signal, assembled with care about genome builds and consistency. The output logic is shared: profiles and counts, checked against held-out signal before being trusted. The interpretation logic is shared: contribution scores first, motif discovery second, skepticism throughout. Even the failure modes feel familiar, since environment problems, input mismatches, and undertrained models announce themselves similarly in both. Learning ChromBPNet after BPNet is mostly learning its bias machinery and its genome-wide habits; learning BPNet after ChromBPNet is mostly learning to think about one factor at a time. Neither is a restart.
Outputs you can compare, region by region
One underappreciated consequence of the shared design is how comparable the two frameworks feel at the level of a single locus. Both hand you a predicted profile across a region and a total for that region, so in either framework you can lay the prediction beside the measured signal and judge the model with your eyes: does the predicted shape rise where the experiment rises, does it peak where binding or accessibility concentrates, does it stay quiet in the flanks? That visual, quantitative honesty is a large part of why the profile-prediction family displaced single-score models in the labs that adopted them.
ChromBPNet extends the same honesty to an entire cell type’s accessibility landscape. The regions change, the assay changes, the scale jumps, yet the working experience is continuous with BPNet’s: train carefully, predict profiles, compare them against held-out measurements, then interpret the bases the model depends on. Researchers moving between the two frameworks are not learning two tools so much as learning one tool’s two dialects, and the fluency compounds.
Conclusion
ChromBPNet and BPNet are two expressions of one idea from the same lab: predict the experimental signal itself, at base resolution, rather than a summary of it. BPNet applies that idea to transcription factor binding data, focused and factor-specific. ChromBPNet applies it to chromatin accessibility, genome-wide and noisy, and adds the bias-factorized design that accessibility assays, with their enzyme sequence preferences, make necessary. Choose by the experiment you ran, and the frameworks stop looking like competitors. They look like what they are: a matched pair of instruments, each tuned to its assay.
Read next:
