EPINTLM: enhancer–promoter prediction with pretrained k-mer embeddings and residual cross-attention

TL;DR

EPINTLM is a deep learning model that predicts enhancer-promoter interactions using DNA sequences and genomic features with improved accuracy.

Contribution

EPINTLM introduces a novel deep learning framework with cross-attention and residual aggregation for enhancer-promoter interaction prediction.

Findings

EPINTLM achieves competitive AUROC and AUPR performance on a benchmark across six human cell lines.

Ablation studies show cross-attention and residual aggregation are key to model performance.

A unified preprocessing pipeline improves training efficiency and reproducibility.

Abstract

Enhancer–promoter interactions (EPIs) play an important role in gene regulation, yet experimental mapping remains costly and limited in coverage. As a result, computational approaches are commonly evaluated under curated benchmark datasets, which pose challenges related to long-range sequence modeling, multimodal feature integration, and reproducible preprocessing. In this study, we present EPINTLM (Enhancer–Promoter Interaction Nucleotide Transformer Large Model), a deep learning framework designed to investigate architectural strategies for EPI prediction under standardized benchmark settings. EPINTLM integrates DNA sequence representations and genomic features by leveraging pretrained k-mer embeddings from the Nucleotide Transformer and explicitly modeling intra- and inter-sequence dependencies through residual self-attention and bidirectional cross-attention. We additionally introduce a unified preprocessing pipeline to improve training efficiency and reproducibility, and perform post hoc motif analysis to provide limited interpretability of learned sequence patterns. Evaluated on a widely used benchmark across six human cell lines, EPINTLM achieves competitive area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPR) performance relative to existing methods, with ablation studies highlighting the contributions of cross-attention and residual aggregation. These results demonstrate the utility of explicit cross-attention designs for paired regulatory sequence modeling within current benchmark constraints.

Figures