Loading…
Base-resolution chromatin accessibility model that factors out enzyme sequence bias to score regulatory variants and transcription factor footprints.
No providers recorded yet. Browse all providers
ChromBPNet is a convolutional DNA sequence model of chromatin accessibility from Anshul Kundaje's lab at Stanford University, with collaborators at Lawrence Berkeley National Laboratory, UC Irvine, and the University of Tartu. It predicts base-resolution ATAC-seq or DNase-seq coverage profiles from local sequence, and is built around one idea the field had largely ignored: the enzymes used to measure accessibility have their own sequence preferences, and a naively trained model learns those preferences alongside real transcription factor syntax.
The Tn5 transposase and DNase-I nuclease both cut DNA preferentially at particular sequences. ChromBPNet handles this in two stages. A small bias model is trained only on background regions with no meaningful accessibility, where the signal is essentially pure enzyme preference; it is then frozen and used to regress the enzyme's contribution out of the profiles, so the remaining "TF model" learns sequence determinants of accessibility rather than assay artifacts. Motifs recovered from a naive model's profile contribution scores include spurious patterns matching Tn5's own preference, which disappear after factorization.
The second theme is that a small, local model competes with much larger ones on the task people most want a regulatory model for — predicting the effect of a non-coding variant. ChromBPNet sees 2,114 bp of sequence and predicts a 1 kb profile, against the 200 kb context of Enformer, yet outperforms it across QTL, pioneer-factor binding, and reporter-assay benchmarks.
The architecture adapts BPNet — a fully convolutional network with dilated convolutions and residual connections — widening the first convolutional layer from 64 to 512 filters for the broader lexicon that genome-wide accessibility requires. Training uses reproducible peaks and GC-matched background regions in five-fold chromosome-partitioned cross-validation. On K562, predicted and measured total counts correlate at r = 0.70 for ATAC-seq and r = 0.71 for DNase-seq on held-out chromosomes, with auROC 0.98 separating peaks from background.
Variant effect benchmarks are the centrepiece. On DNase-seq QTLs from 70 lymphoblastoid cell lines, the composite prioritization score reaches average precision 0.54 for ATAC-seq models and 0.43 for DNase-seq, against 0.19 for a gkm-SVM on the same data and 0.33 for Enformer's published scores; recomputing Enformer's scores over local 2 kb windows raises it to 0.53, still below the deeply sequenced ATAC-seq model at 0.54 with effect-size correlation r = 0.76. On SPI1 binding QTLs, ChromBPNet beats Enformer on signed and unsigned effect sizes (r = 0.59 versus 0.53). Against CAGI5 reporter assay data across 14 loci, it matches Enformer within ±0.05 Pearson correlation at 9 and exceeds it at 5. On red blood cell trait GWAS loci, high-scoring variants reach up to 70-fold enrichment for fine-mapped causal variants.
The model suits regulatory genomics groups interpreting non-coding variation: prioritizing causal variants inside GWAS credible sets, evaluating rare de novo variants, mapping TF footprints and cooperative motif grammar in a cell type of interest, and reading out which motif a variant disrupts rather than only that it scores highly. It runs on bulk ATAC-seq and DNase-seq and on pseudobulk single-cell ATAC, reaching primary cell types such as microglia and coronary smooth muscle. One characteristic to plan around: applying ChromBPNet to a new cellular context requires training a model on that context's own reads. The chrombpnet pipeline command is a training entry point, and the artifacts that ship ready to use are the enzyme bias models on Zenodo plus the per-biosample models from the ENCODE model zoo. That zoo is mirrored on Hugging Face as an MIT-licensed collection covering 1,512 DNase-seq and ATAC-seq experiments across 408 biosamples, so a group working in a covered cell type can skip training entirely.
ChromBPNet is the reference implementation for bias-corrected, base-resolution accessibility modelling, distributed through pip, Docker, and an ENCODE-registered pipeline that has generated trained models across many biosamples. Its wider significance is evidence that context length is not the binding constraint for local regulatory prediction: a lightweight model on 2 kb of sequence, trained on deeply sequenced data, matched or beat a long-context multi-task transformer on the variant benchmarks both were built to serve. The counterpoint is scope — it models accessibility one context at a time and says nothing about long-range enhancer–promoter contacts. The work remains a preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.