README.md

metadata

license: mit
tags:
  - biology
  - genomics
  - gene-expression
  - enhancer
  - regulatory-genomics
  - dna
library_name: pytorch

EPInformer — reproducible pipeline (pretrained checkpoints)

Pretrained checkpoints for the EPInformer reproducible pipeline — a from-raw-ENCODE reproduction of EPInformer across 6 cell lines (K562, GM12878, H1, HepG2, HUVEC, NHEK).

Two models (defined in EPInformer/models.py on the reproducible_pipeline branch):

• enhancer_predictor_256bp — 256 bp enhancer-activity encoder; predicts log2(0.1 + sqrt(H3K27ac·DNase)) activity from sequence.
• EPInformer_v2 — gene-expression model (RNA / CAGE) that reuses the frozen encoder as its sequence backbone.

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

Enhancer encoder (log2 activity):

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

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).

Files

enhancer_encoders/{CELL}/fold_{i}.pt   # CELL ∈ {K562, GM12878, H1, HepG2, HUVEC, NHEK}, i ∈ 1..12

These are the reproduction's 256 bp enhancer-activity encoders — the best checkpoint for each of the 12 leave-chromosome-out folds, for all 6 cell lines (72 checkpoints). Any fold works for inference; the R values in the table above are pooled across all 12 held-out folds.

The KLF1 demo in the notebooks uses the original published EPInformer encoder shipped in the repo's trained_models/pretrained_enhancer_encoder/ — a separate checkpoint from these reproduction folds.

Usage

import torch
from huggingface_hub import hf_hub_download
# the model class lives on the reproducible_pipeline branch:
#   git clone -b reproducible_pipeline https://github.com/pinellolab/EPInformer
from EPInformer.models import enhancer_predictor_256bp

ckpt = hf_hub_download("JiecongLin/EPInformer-reproducible", "enhancer_encoders/K562/fold_8.pt")
net = enhancer_predictor_256bp()
net.load_state_dict(torch.load(ckpt, map_location="cpu", weights_only=False)["model_state_dict"])
net.eval()

See the project wiki for the full training and evaluation guide, and the reproducible_pipeline branch for code.