EpiBinder

A CTCF motif reads the same in a lymphoblastoid cell as it does in a hepatocyte, yet the factor may occupy it in one and ignore it in the other. A model reading only DNA sees one string and can return only one answer, putting cell-type-specific occupancy out of reach. The field's usual response has been to widen the receptive field to hundreds of kilobases, on the theory that distal sequence supplies what the local motif cannot.

EpiBinder takes the other route. It keeps a compact 1,000 bp window and adds the measurements that genuinely differ between cell types: cytosine methylation in CpG, CHG and CHH contexts from whole-genome bisulfite sequencing, plus DNase I hypersensitivity. Each window becomes a 1,000 × 8 matrix — four one-hot nucleotide channels and four continuous epigenetic channels at single-base resolution — with separate models fit for GM12878, HepG2 and K562, the three most densely profiled ENCODE cell lines. The conclusion runs against the long-context trend: for TF binding, better characterization of local regulatory context beats a longer sequence span.

The model was developed at MBZUAI and the Weizmann Institute of Science, first presented at the ICML 2025 GenBio workshop and posted to bioRxiv in July 2026. It is benchmarked against DeepSEA, whose ENCODE and Roadmap corpus it reuses, plus HyenaDNA, DNABERT and BigBird.

#Key Features

• Eight-channel base-resolution input: Sequence and epigenome enter as parallel per-nucleotide streams rather than region-level covariates, so methylation is resolved at the same granularity as the bases it modifies.
• Cell-type-specific multitask heads: A single forward pass emits binding logits for every profiled target in that cell line — 94 classes for GM12878, 104 for HepG2 and 185 for K562, spanning TF ChIP-seq, histone marks and RNA polymerase.
• Zero-shot transfer to unseen cell lines: Models trained on one cell line are applied to A549 and H1-hESC, which contribute no training data, with the in-domain to out-of-domain gap quantified by Cohen's d.
• Methylation-sensitivity scanning: Zeroing methylation across a ±500 bp window genome-wide identifies CpG sites where demethylation raises or suppresses predicted binding, cross-referenced against the Illumina EPIC 850k array and EWAS Atlas mQTLs.
• In silico motif perturbation: Randomizing a neighbouring motif and re-scoring isolates cooperative and competitive TF pairs — in HepG2, disrupting a FOXA-family motif drops predicted TCF12 occupancy from 0.57 to 0.17.

#Technical Details

The architecture is a three-stage 1D convolutional backbone (channel widths 320, 480 and 960, kernel size 8, with ReLU, max-pooling and dropout) feeding a single 8-head transformer encoder layer at hidden dimension 960, whose flattened output passes through a 512-unit MLP classifier — roughly 46M parameters, between DeepSEA's 40M and DNABERT's 86M. Training used binary cross-entropy with Adam over roughly 2.2 million samples, holding out 227,512 on chromosomes 8 and 9. Labels come from ChIP-seq peak calls binarized over the central 200 bp of each window, with 400 bp of flanking context either side; the 318 TF profiles drawn from the DeepSEA corpus account for 45.9% of its ChIP-seq experiments. Mean auPRC for TF binding reaches 39.1 on GM12878, 43.1 on HepG2 and 41.1 on K562, against 26.1, 29.3 and 25.8 for DeepSEA — average gains of 14.0% for TFs, 11.6% for polymerase and 18.2% for K562 histone marks. Interpretation runs through SHAP attributions clustered by TF-MoDISco into seqlets and matched to JASPAR motifs, summarized by a Jaccard Overlap Score that separates factors recognized by one canonical motif from those predicted through broad contextual association.

#Applications

The practical draw is any question where a motif match is not enough: which enhancers a factor actually occupies in a given cell type, which CpG positions make a regulatory element methylation-sensitive, and which factor pairs act together at a locus. In-silico saturation mutagenesis of the SORT1 enhancer in HepG2 tracked measured MPRA variant effects at r ≈ 0.39 and recovered four of seven known regulators, including FOXA1, HNF4G, USF1 and SP5 — a route to prioritizing variants and cofactors before committing to a ChIP or reporter experiment. The zero-shot results matter most for cell types that will never receive dense TF profiling.

#Impact

EpiBinder is a preprint awaiting peer review, and its claims are bounded accordingly. Evaluation covers only cell lines with unusually deep ChIP-seq coverage; models are fit per cell line rather than jointly, so each learns from far fewer examples than a pooled model would; and SHAP attribution proved costly enough that interpretability analyses drew on roughly 1,000 binding sites. Its contribution is a controlled demonstration that base-resolution methylation and accessibility carry signal a longer sequence window does not recover — a counterweight in a subfield that has spent several years scaling context length. Training and inference code is public and archived on Zenodo under CC BY 4.0; the three per-cell-line checkpoints are not currently distributed.