ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants
Anusri PampariiD, Anna ShcherbinaiD, Evgeny Z. KvoniD, Michael KosickiiD, Surag NairiD, Soumya KunduiD, Arwa S. Kathiria, Viviana I. RiscaiD, Kristiina Kuningas, Kaur AlasooiD, William James GreenleafiD, Len A. PennacchioiD, Anshul Bharat KundajeiD
Abstract
Despite extensive mapping of cis-regulatory elements (cREs) across cellular contexts with chromatin accessibility assays, the sequence syntax and genetic variants that regulate transcription factor (TF) binding and chromatin accessibility at context-specific cREs remain elusive. We introduce ChromBPNet, a deep learning DNA sequence model of base-resolution accessibility profiles that detects, learns and deconvolves assay-specific enzyme biases from regulatory sequence determinants of accessibility, enabling robust discovery of compact TF motif lexicons, cooperative motif syntax and precision footprints across assays and sequencing depths. Extensive benchmarks show that ChromBPNet, despite its lightweight design, is competitive with much larger contemporary models at predicting variant effects on chromatin accessibility, pioneer TF binding and reporter activity across assays, cell contexts and ancestry, while providing interpretation of disrupted regulatory syntax. ChromBPNet also helps prioritize and interpret regulatory variants that influence complex traits and rare diseases, thereby providing a powerful lens to decode regulatory DNA and genetic variation.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 142
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 4 (23 Sept 2026) · CC BY 4.0
Learning and removing Tn5/DNase-I sequence bias lets a small base-resolution CNN (ChromBPNet) recover TF motifs and footprints from accessibility data and rival Enformer at predicting chromatin QTL effects.
Footprinting tools need deep coverage and predefined motifs, while nearly all contemporary sequence models (AlphaGenome aside) train on binarized peaks or 32-100 bp bins, and large multi-task models are costly to train on a single dataset. ChromBPNet models raw base-resolution ATAC-seq or DNase-seq profiles and separates the enzyme's own sequence preference from TF-driven signal within one model.
Key results
ShownEnzyme sequence preference distorts base-resolution profile shape but not total counts; a frozen bias model trained on chromatin background removes it, and does so better than TOBIAS, HINT-ATAC or naked-DNA bias models (including seq2PRINT).
A Tn5 bias model predicted peak profile shape (median JSD = 0.48) but not counts (r = -0.16). After factorization, Tn5 bias motifs vanished from TF-MoDISco motifs of 30,000 GM12878 peaks, as did a GATA1 footprint (GATA not expressed there). ATAC vs DNase profile JSD on 30K K562 peaks: 0.81 measured, 0.58 uncorrected, 0.26 bias-corrected. Alternatives left residual bias or G/C artefacts. · CNNs learn Tn5/DNase-I sequence preferences from…; A bias-factorized neural network (ChromBPNet)…; ChromBPNet outperforms alternative approaches…; ChromBPNet substantially improves concordance…
ShownModels trained on shallow data impute denoised profiles, footprints and motifs close to those of a deep-data model; motif recovery holds down to 25M reads, below which rarer motifs are lost.
GM12878 ATAC-seq (572M reads) subsampled: median JSD of measured profiles rose from 0.3 (250M) to 0.9 (5M), while ChromBPNet predictions stayed at 0.147 (250M) to 0.190 (5M) relative to the high-depth model. Frequent motifs (IRF, SPI, RUNX) were recalled better than rare ones (PAX, COE, HNF1). Pseudobulk scATAC-seq from 11,000 down to ∼100 cells followed the same trends. · ChromBPNet can impute base-resolution profiles…; 1.21 Subsampling; 1.28 Comparative Analysis of ChromBPNet…
ShownEach cell line's accessibility is explained by a compact, non-redundant motif lexicon that includes composite elements with strict spacing and orientation; motif contribution scores track TF occupancy far better than accessibility or footprint scores.
Unified lexicons: GM12878 51, HepG2 60, IMR90 47, H1-hESC 41, K562 47 motifs. In IMR90 the FOS-TEAD composite footprint exceeded the sum of its parts, and the cooperative gain disappeared when spacing or orientation changed. For CTCF in HepG2, ChromBPNet scores correlated with ChIP-seq BPNet scores at r > 0.89, versus r = -0.1 for total accessibility and r = 0.2 for TOBIAS footprints. · ChromBPNet reveals compact TF motif lexicons…; 1.22 Comparative Analysis of ChromBPNet and…
ShownTrained on one reference LCL using only the reference genome, ChromBPNet predicts chromatin QTL effects better than gkm-SVM and Enformer's published scores, competitively with locally recomputed Enformer scores, and consistently across ancestries.
Yoruba dsQTLs: average precision (AP) 0.54 (ATAC, 572M reads) and 0.43 (DNase, 68M) vs gkm-SVM 0.19, published Enformer 0.33, recomputed Enformer 0.53; effect-size r = 0.76 (ATAC) vs 0.56 for published Enformer. SPI1 bQTLs: ATAC-seq ChromBPNet AP 0.37 vs 0.34 for recomputed Enformer. Variant scores agreed at r = 0.96-0.97 between African-subgroup models. · ChromBPNet predicts effects of genetic variants…; ChromBPNet predicts effects of sequence…
ShownVariant scores learned from accessibility alone carry over to reporter assays, fine-mapped GWAS variants and a rare-disease variant whose predicted damaging effect was supported by transgenic mouse reporter assays.
CAGI5 MPRA: Pearson r within ±0.05 of Enformer at 9 loci and substantially higher at 5 of 14. High-scoring variants were enriched up to 70-fold among fine-mapped red-cell-trait variants (PIP threshold 0.4). A de novo variant predicted to cut accessibility by 40% abolished forebrain, hindbrain and eye enhancer activity at E11.5; a predicted-neutral control had no effect. · ChromBPNet predicts effects of sequence…; 1.26 Transgenic E11.5 embryo generation
Methods, data and reproducibility
A dilated CNN (BPNet-style; 2,114 bp input, 1 kb output) predicts total counts and base-resolution profile shape, while a small frozen bias CNN trained on GC-matched background regions absorbs Tn5/DNase-I preference and the main model learns the residual. DeepLIFT and TF-MoDISco extract motifs; allele-swap predictions score variants (logFC, JSD and integrative scores).
- ENCODE bulk ATAC-seq and DNase-seq from K562, HepG2, GM12878, IMR90 and H1-hESC (new H1-hESC ATAC-seq generated here) · human · n = 5 cell lines · GSE267154 (H1-hESC ATAC-seq); other ENCODE IDs are in Table 1, not in this text
- Deep GM12878 ATAC-seq subsampled to 250M-2M reads, plus GM12878 scATAC-seq pseudobulks · human · n = 572M reads; 19,000 cells · ENCSR637XSC; ENCSR371ZNU
- Naked-DNA ATAC-seq and DNase-seq used for bias-model comparisons · human · SRX030445; SRR1565781, SRR1565782
- Chromatin QTLs in LCLs: Yoruba dsQTLs (deltaSVM set), European and African caQTLs, each with matched controls · human · n = 560 dsQTLs; 7,900 European and 6,821 African caQTLs · GSE31388 (dsQTL effect sizes); QTD100018 (European caQTLs)
- SPI1 binding QTLs from pooled ChIP-seq across Yoruba LCLs · human · n = 999,799 variants; 60 LCLs
- CAGI5 MPRA measurements of single-nucleotide substitutions and deletions in disease-associated promoters and enhancers · human · n = over 30,000 variants in 20 elements
- Fine-mapped UK Biobank red blood cell trait variants (MCV, MCH, MCHC) plus background controls · human · n = 12,951 trait variants; 11,916,124 scored
- Primary-cell scATAC-seq (microglia, coronary smooth muscle cells) with matched caQTLs · human · n = 877 microglia and 386 SMC caQTLs · GSE147672 (microglia); GSE175621 (SMC)
- Methods & tools
- ChromBPNet, BPNet, DeepLIFT/DeepSHAP, TF-MoDISco / modisco-lite, TOMTOM (JASPAR), TOBIAS, HINT-ATAC, seq2PRINT, Enformer, gkm-SVM, LDSC, transgenic mouse lacZ enhancer assay
- Data & accessions
- https://enhancer.lbl.gov/
- Reagents & environment
- Versions stated for SAMtools 1.7, bedtools 2.29.2, caper 2.1.2, MACS2 v2.2.7.1, TensorFlow v1.6 (DeepLIFT), TF-MoDISco 0.5.16.0, modisco-lite 2.2.0, TOMTOM 5.5.5, BWA v0.7.17 and Nextflow v23.09.3; no container or requirements file described.
- Reproducibility gaps
- Links to the code repository, figure-reproduction code and variant-scoring software are missing from this text. Models, outputs and benchmarks sit on ENCODE and Synapse (e.g. syn59449898, syn64126763); plasmids on request; CAGI5 and Enformer scores obtained by correspondence.
Limitations (6)
- Total-accessibility predictions trail pseudo-replicate concordance and recent long-context models, so predictions and derived interpretations are less reliable outside peaks. · stated by the authors
- Rare-motif sensitivity and variant classification drop notably below 25M reads; at matched depth, recomputed Enformer scores classify QTLs better, and ChromBPNet consistently trails AlphaGenome on these benchmarks. · stated by the authors
- Captures only TFs that change accessibility, not 'settler' TFs that bind already-open chromatin, and needs a separate model for each cell state or biosample. · stated by the authors
- Chromatin-QTL benchmarks test local QTLs only (no distal chromatin QTLs or eQTLs) with control variants restricted to peak regions; the authors note that variant classes beyond SNVs, such as indels and STRs, remain to be tested. · stated by the authors
- Bias-correction comparisons with TOBIAS, HINT-ATAC and naked-DNA models are judged mainly by motif and footprint inspection within the authors' own framework, and the HINT-ATAC and TOBIAS baselines were trained on fold 0 only. · noted in this summary
- In vivo support for rare-variant prioritization rests on one de novo variant plus one control in transgenic mouse embryos; embryo counts are not reported in this text. · noted in this summary
Who should read:Read this if you analyse bulk or pseudobulk ATAC-seq or DNase-seq and want bias-corrected footprints, motif syntax or variant-effect scores from a model you can train on a single dataset.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- ChAI-Net learns a regulatory code for chromatin co-accessibility
Jodi Lee, Kaite Zhang, … Katherine S. Pollard·bioRxiv·Oct 31, 2026
- Transfer learning identifies sequence determinants of regulatory element accessibility
Marco Salvatore, Marc Horlacher, … Robin Andersson·NAR Genomics and BioinformaticsPeer-reviewed·Jun 1, 2022
- Incorporating Gene Expression in Genome-wide Prediction of Chromatin Accessibility via Deep Learning
Qiao Liu, Kui Hua, … Rui Jiang·Genomics Proteomics & BioinformaticsPeer-reviewed·Mar 12, 2022
- circuitSNPs: Predicting genetic effects using a Neural Network to model regulatory modules of DNase-seq footprints
Alexander G Shanku, Anthony S Findley, … Roger Piqué-Regi·bioRxiv·Jun 3, 2018
- Transfer learning reveals sequence determinants of the quantitative response to transcription factor dosage
Sahin Naqvi, Seungsoo Kim, … Joanna Wysocka·Cell GenomicsPeer-reviewed·Feb 27, 2025
- Base-resolution models of transcription factor binding reveal soft motif syntax
Žiga Avsec, Melanie Weilert, … Julia Zeitlinger·Nature GeneticsPeer-reviewed·Feb 18, 2021