Biological Background

Switch language: 生物学背景

Enhancer-promoter interaction (EPI) prediction is a sequence-based learning task centered on distal gene regulation. Enhancers are cis-regulatory DNA elements that can increase transcription of target genes, often across long genomic distances. Promoters are proximal regulatory regions near transcription start sites that recruit the transcriptional machinery. Their coordinated interaction is one of the main mechanisms through which cell identity, developmental programs, and disease-associated dysregulation are established.

Why this problem matters

  • Experimental assays such as Hi-C, ChIA-PET, Capture-C, and related 3D genome technologies are informative but expensive, slow, and not uniformly available across cell types.

  • A computational model that infers likely enhancer-promoter pairs from sequence can help prioritize regulatory hypotheses, guide follow-up experiments, and support interpretation of non-coding variants.

  • The project also serves as a benchmark bed for comparing classical CNN-style genomic models with newer sequence architectures such as Transformers, Mamba, RWKV, mLSTM, and DNA language model encoders.

Biological framing

In biological terms, the task is not simply motif classification. A practical EPI predictor must learn several levels of signal:

  • local sequence composition, including GC content and low-level nucleotide arrangement;

  • regulatory motifs associated with transcription factors and chromatin state;

  • combinational logic between enhancer and promoter regions;

  • long-range context patterns that may be represented better by modern sequence models than by shallow local filters alone.

Input and output

The project models each sample as a pair of DNA sequences:

  • enhancer sequence;

  • promoter sequence.

The output is a binary label:

  • 1 means the pair is considered interacting;

  • 0 means the pair is considered non-interacting.

In the current implementation, the core framework supports three input representations:

  • one-hot encoding;

  • k-mer token encoding;

  • precomputed DNA LLM embeddings.

Project role

The current repository is a modernization of the original DeepChrInteract code base. The old Keras/TensorFlow implementation is preserved for historical reference under DeepChrInteract-main(old)/. The new codebase reimplements the workflow in PyTorch, removes the PNG-based preprocessing path, expands the model family to fourteen encoder variants, and standardizes evaluation and experiment organization.

../_images/div.png