Part 2 Gene Expression

Part 2 — Gene-expression model

Do this second. EPInformer_v2 predicts RNA / CAGE from a gene's promoter + its ABC-nominated enhancers, reusing the frozen Part 1 encoder as the sequence backbone.

2a. Preprocessing — ABC links → factored gene HDF5

python run_pipeline.py --config config/config.yaml --samples K562,GM12878 --stages encoding
#   -> batch_output/{cell}/encoding/{cell}_samples.h5
#      (promoter + enhancer one-hot sequence, and activity / dhs / distance / contact features)

2b. Train EPInformer_v2 — the shipped f3 config

python train_EPInformer.py --model_type EPInformer-v2 --cell K562 --expr_type RNA \
    --n_enh_feats 3 --use_prm_signal \
    --h5_path   batch_output/K562/encoding/K562_samples.h5 \
    --expr_csv  data/GM12878_K562_18377_gene_expr_fromXpresso_with_sequence_strand.csv \
    --split_csv data/leave_chrom_out_crossvalidation_split_18377genes.csv \
    --use_pretrained_encoder --pretrained_encoder_dir results/seqencoder/K562/checkpoints \
    --output_dir EPInformer_models/K562 --epochs 50
# all 12 folds via SLURM (f3 is the default):  CELL=K562 sbatch slurm/train_epinformer_12fold.slurm
#   CAGE instead of RNA:   EXPR_TYPE=CAGE ...

What f3 means

The shipped config = --n_enh_feats 3 (distance + activity + Hi-C contact) + --use_prm_signal (promoter activity), with the encoder frozen. In the feature ablation:

• f1 = distance only
• f2 = + activity
• f3 = + Hi-C contact

Activity dominates; Hi-C is ~inert; the promoter signal adds up to +0.03 (cell-dependent), so it is bundled into f3 by default. --no_freeze_encoder fine-tunes the encoder end-to-end (optional, doesn't reliably help).

2c. Evaluate

python evaluate.py expression --pred_dir EPInformer_models/K562

Targets (pooled out-of-fold Pearson R):

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

Open visualize_results.ipynb for density scatters, per-fold R, and the cross-cell table.