UniversalEPI

Which enhancer contacts which promoter is a defining question in gene regulation, but the assays that answer it directly — Hi-C and Micro-C — are costly, so experimental contact maps exist for only a small number of well-studied cell lines. Computational predictors have filled part of the gap, yet most either demand inputs such as CTCF ChIP-seq that are unavailable for typical samples, or lose accuracy on cell types absent from training.

UniversalEPI, developed in Valentina Boeva's group at ETH Zurich with collaborators at the Swiss Data Science Center, Helmholtz Munich, and Dana-Farber Cancer Institute, predicts enhancer-promoter interactions up to 2 Mb apart from just two inputs: DNA sequence and chromatin accessibility measured by ATAC-seq. Instead of reconstructing a dense contact map over every genomic bin, it predicts the sparse set of quantitative interaction strengths between accessible regulatory elements. That reformulation keeps the model small enough to run on a standard GPU while retaining a multi-megabase receptive field.

The model is a deep ensemble that reports how confident each prediction is, which turns it from a point predictor into a tool for differential analysis: changes between two conditions can be tested against the model's own uncertainty rather than assumed real. The work was posted to bioRxiv in November 2024 and published in Nucleic Acids Research in 2026.

#Key Features

• Two widely available inputs: Only DNA sequence and ATAC-seq are needed — no CTCF ChIP-seq, no epigenomic panel — so the model applies to cell types for which little more than an accessibility profile exists.
• Generalization to unseen cell types: On held-out chromosomes of held-out cell lines, UniversalEPI was the only method benchmarked to exceed a Spearman correlation of 0.90 with experimental Hi-C, outperforming C.Origami, Akita, EPCOT, and ChromaFold.
• Calibrated uncertainty: Aleatoric uncertainty is learned through a β-NLL loss and epistemic uncertainty from the spread of a 10-model ensemble, enabling statistically grounded calls on differential interactions between conditions.
• Lightweight by design: 2.5M parameters and short inference times, because the model attends only to accessible regions rather than to every bin in a 4 Mb window.
• Single-cell compatible: Predictions from scATAC-seq pseudo-bulk profiles recover chromatin dynamics across cell states, extending 3D genome modeling to populations with no bulk assay.
• Precomputed public resource: Genome-wide interaction maps for 157 ENCODE ATAC-seq datasets — 116 cell lines and 41 primary cell samples — are distributed as a UCSC Genome Browser track hub.

#Technical Details

UniversalEPI runs in two stages. Stage 1 is a five-layer convolutional network over one-hot-encoded 1-kb sequences centered on ATAC-seq peak summits, plus the accessibility signal and a 36-bp mappability track; it predicts binding of three ubiquitously expressed transcription factors — CTCF (forward and reverse), YY1, and SP1 — reaching average Pearson correlations of 0.78, 0.62, and 0.47 on unseen chromosomes of unseen cell types. Its backbone is then frozen as a feature extractor. Stage 2 takes 401 consecutive accessible regions spanning roughly 4 Mb, applies stochastic gating to select convolutional features, and passes them through four multi-head attention blocks (4 heads, d_model 32) with a genomic-distance-aware positional encoding at 500-bp resolution, followed by a two-layer MLP decoder — 2.5M parameters in total.

Training used 5-kb ICE-normalized Hi-C from the 4D Nucleome Data Portal for GM12878, K562, IMR90, and HepG2 in a two-fold cross-validation: one model trained on GM12878 and K562 and tested on IMR90 and HepG2, the other reversed, with chromosomes 5, 12, 13, and 21 held out for validation and 2, 6, and 19 for testing. A separate variant was trained on 1-kb Micro-C data. Training takes about 12 hours on a fifth of an A100.

#Applications

The practical draw is coverage: any sample with an ATAC-seq profile — a rare primary cell type, a treatment time course, a tumor subpopulation resolved by scATAC-seq — can be given a predicted interaction map without new Hi-C. The authors track enhancer-promoter rewiring during human macrophage activation, where predictions agree with measured Hi-C at Spearman above 0.9, and contrast undifferentiated and differentiated cancer cell states in esophageal adenocarcinoma, identifying promoters of master transcriptional regulators that gain or lose connectivity.

#Impact

UniversalEPI argues that a deliberately narrow output space — sparse peak-to-peak interactions rather than a full contact matrix — buys both generalization and efficiency, a design choice relevant beyond this one task. Its limitations follow from that choice: it predicts nothing about regions outside ATAC-seq peaks, its sequence features are grounded in three ubiquitous transcription factors and so do not capture cell-type-specific regulators, and published benchmarking is confined to interactions of 1 Mb or less even though the model spans 2 Mb. The code is MIT-licensed with pretrained checkpoints on Zenodo; the paper is released under CC BY-NC. Adoption is still early, but the combination of minimal input requirements, released weights, and a browsable prediction resource lowers the barrier to using in silico 3D genome maps in regulatory genomics.