--- license: mit library_name: pytorch tags: - genomics - enhancer - chromatin - dnase - h3k27ac - epinformer - chorus --- # chorus-epinformerseq-v2 Per-cell **EPInformer-seq** checkpoints for 11 Roadmap Epigenomics cell types. Drop-in artifacts for the [Chorus](https://github.com/pinellolab/chorus) `epinformerseq` oracle. ## Architecture `PerCellProfileNetWide` — a dilated-CNN, **2-channel** profile model. A **2114-bp** input is run through the body, then the central **1024 bp** is cropped for the heads (ChromBPNet-style "valid" geometry, so every output base has a full real-sequence receptive field). The two channels are: - **ch0 — DNase**: 5′ cut-sites (motif-sensitive, sharp). - **ch1 — H3K27ac**: read signal (the active-enhancer histone mark). Each cell line gets its **own** main checkpoint (no FiLM, no cell embedding), paired with a per-cell **frozen 2-channel `BiasNet`** (1024-bp, run on the central crop): ch0 subtracts the Tn5/DNase enzymatic cut bias, ch1 the H3K27ac background bias. H3K27ac has no enzymatic cut, so no separate bias model is needed. The retired joint `CellCondProfileNet` (FiLM over 11 cells) and the earlier 1024-bp SAME-padded `PerCellProfileNet` are not used by Chorus. ## Layout ``` per_cell_widewin/ K562/main.pt # PerCellProfileNetWide state_dict (~136K params) ... (11 cells) bias/ K562/bias.pt # 2-channel BiasNet state_dict (~37K params, frozen) ... (11 cells) ``` > **Shipped weights (2026-06-04): the `roadmap` retrain** — trained on Roadmap > DNase-summit peaks with the Roadmap-pipeline H3K27ac. Strongest per-cell DNase > test-r and a functional H3K27ac channel; supersedes an earlier ENCODE-IDR-peak / > full-coverage-H3K27ac variant. ## Training - **Window**: 2114-bp input → central 1024-bp profile crop, on roadmap DNase peak summits. - **Signals**: ch0 = per-bp 5′ DNase cut-site counts; ch1 = per-bp H3K27ac signal (Roadmap v2.1 pipeline). - **Data**: ENCODE DNase + H3K27ac BAMs (SE/PE), fold-10 leave-chromosomes-out split (chr11+chr21 held out). - **Loss**: multinomial NLL on the per-bp 2-channel profile + MSE on log10 count per channel. AdamW + OneCycleLR. - Per-cell held-out test-`r`: **DNase 0.59–0.88, H3K27ac 0.48–0.67**. ## Usage ```python from chorus.oracles import EPInformerSeqOracle oracle = EPInformerSeqOracle(cell_type="K562") oracle.load_pretrained_model() # downloads from this repo on first run result = oracle.predict( sequence="A" * 2114, assay_ids=["Enhancer_DNase:K562"], # or Enhancer_H3K27ac, Enhancer_H3K27ac_DNase ) ``` Available cells: K562, GM12878, HepG2, A549, H1, HeLa, HMEC, HSMM, HUVEC, NHEK, NHLF. Available assays (max over the central 256 bp of the 1024-bp output): `Enhancer_DNase` (default, max DNase), `Enhancer_H3K27ac` (max H3K27ac), `Enhancer_H3K27ac_DNase` (composite `sqrt(max DNase · max H3K27ac)`). ## Background CDFs Chorus pulls per-track CDFs from the companion dataset [`lucapinello/chorus-backgrounds`](https://huggingface.co/datasets/lucapinello/chorus-backgrounds) (`epinformerseq_pertrack.npz`, 33 tracks = 3 assays × 11 cells). ## Citation Cite the EPInformer paper and the Chorus pipeline (see the [Chorus README](https://github.com/pinellolab/chorus)).
