ENCODE Atlas — BPNet-family model file index

Tables

file	family	models	what the models predict
bpnet_model_files.tsv	BPNet	2,339	base-resolution TF binding (ChIP-seq/-nexus)
chrombpnet_model_files.tsv	ChromBPNet	1,512	base-resolution chromatin accessibility (ATAC/DNase)
procapnet_model_files.tsv	ProCapNet	6	transcription initiation (PRO-cap)
reporternet_model_files.tsv	ReporterNet	8	massively-parallel reporter activity

Each row is a trained model (ENCODE experiment accession), each column an output object type, and each cell the ENCODE file accession (ENCFF…) to download for that pair.

Client — encode_atlas.py

A single-file, dependency-free (stdlib-only) Python client any agent or user can use to resolve and download the models and their downstream products, and to reach the broader ENCODE portal. It fetches the tables from this repo automatically (cached), so you only need the one file.

import encode_atlas as ea

# what's released for a model, and where to download each object
ea.model_files("ENCSR032RGS")                    # family + metadata + all object types -> ENCFF + URL
ea.download("ENCSR032RGS", "model", "./models/") # the trained model tar; also: motifs, contrib_counts,
                                                 #   contrib_profile, signal_predicted, motif_hits, ...

# discovery -- you know the tissue, not the accession
ea.facets()                                      # what exists: counts by family / assay / tissue / target
ea.find_models("liver", family="chrombpnet")     # free-text over tissue/target/assay (ranked, QC first)
ea.find_models_by_organ("liver", family="chrombpnet", assay_title="ATAC-seq")
                                                 # ontology-aware: 'liver' -> HepG2 + hepatocyte + liver
                                                 #   tissue (via ENCODE organ_slims), which a string
                                                 #   match on the tissue column would miss

# search OUR model index (the four family tables)
ea.search(family="chrombpnet", biosample="K562") # -> [{accession, family, assay, target, tissue, qc}]

# reach the BROADER ENCODE portal (raw + processed data for ANY experiment, not just modelled ones)
ea.portal_files("ENCSR000EOT")                   # every file: reads, alignments, peaks, signal, ...
ea.portal_search(assay_title="ATAC-seq", biosample="liver")

Object types: model, motifs, motif_hits, contrib_counts, contrib_profile, signal_predicted, signal_observed, signal_biascorrected, regions, regions_train_test, metrics (see ea.object_types()). Download a file directly at https://www.encodeproject.org/files/<ENCFF>/@@download/.