Home

EPInformer — pipeline

A clean, self-contained pipeline for EPInformer, built around EPInformer/models.py (EPInformer_v2 + the 256 bp enhancer_predictor_256bp). It trains two models, in order, end-to-end from raw ENCODE data across 6 cell lines (K562, GM12878, H1, HepG2, HUVEC, NHEK):

1. Enhancer-activity encoder — predicts 256 bp enhancer activity (H3K27ac·DNase) from sequence.
2. Gene-expression model (EPInformer_v2) — predicts RNA / CAGE from a gene's promoter plus its ABC-nominated enhancers, reusing the frozen encoder as the sequence backbone.

Do Part 1 (encoder) first, then Part 2 (expression) — the expression model loads the encoder.

Results (12-fold leave-chromosome-out, pooled out-of-fold Pearson R)

Part 1 — enhancer encoder (log2 activity):

	H1	HepG2	K562	HUVEC	NHEK	GM12878
R	0.820	0.743	0.740	0.742	0.677	0.617

Encoder headline values use forward-plus-reverse-complement averaging. For the exact pinned K562 narrowPeak, the corresponding single-reverse result is approximately 0.734; keep reverse-only and fwd+RC predictions in separate output directories.

Part 2 — gene expression (shipped f3: frozen encoder + 3 enhancer features + promoter signal):

	K562	GM12878	HepG2	HUVEC	NHEK	H1
RNA	0.856	0.860	0.845	0.839	0.828	0.781
CAGE	0.867	0.890	—	—	—	—

CAGE labels exist only for K562/GM12878 (the other four are RNA-only). The H1/HepG2/HUVEC/NHEK expression numbers are new — upstream EPInformer reports expression for K562/GM12878 only.

Contents

• Setup — environment + reference data
• Datasets and Accessions — ENCODE IDs, genome, expression labels
• Part 1: Enhancer Encoder — preprocessing → train → evaluate
• Part 2: Gene Expression — gene HDF5 → train f3 → evaluate
• Other Cell Lines — H1 / HepG2 / HUVEC / NHEK (RNA-only)
• Notebooks — interactive enhancer-activity + expression demos
• Troubleshooting and Notes — recipe details and common pitfalls

Pretrained checkpoints

Download the pretrained enhancer encoders from Hugging Face: JiecongLin/EPInformer-pipeline.