Hugging Face's logo

Datasets:
adamyhe
/
procap-atlas-motifs

Tags:
genomics
transcriptomics
pro-cap
grch38
hg38
transcription-start-sites
License:
Dataset card Files
xet
 Community
1
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 80, in _split_generators
                  raise ValueError(
                  ...<2 lines>...
                  )
              ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

PRO-cap Atlas Motifs

This dataset contains TF-MoDISco motif discovery results and MotifCompendium atlas-wide motif clustering outputs for BPNet models trained on the ENCODE PRO-cap atlas (GRCh38/hg38). It includes per-experiment MoDISco .h5 files and HTML report archives for 224 experiments, plus the deduplicated atlas-wide motif compendium that collapses per-experiment motifs into shared clusters across biosamples.

This repository is intended to be used together with:

Dataset Details

Uses

Direct Use

Use this dataset to:

  • explore motifs discovered by TF-MoDISco from BPNet attribution scores across ENCODE PRO-cap experiments
  • load per-experiment contribution weight matrices (CWMs) and seqlet counts from .modisco.h5 files
  • browse per-experiment HTML motif reports (.modisco.tar archives)
  • access the atlas-wide MotifCompendium: deduplicated motif clusters with average CWMs, MEME-format motifs, cluster metadata, JASPAR annotations, and SVG logos
  • match per-experiment motifs to shared atlas-wide cluster identities via pattern-to-cluster mappings
  • use cluster-average CWMs (cluster_averages.h5) as a motif set for atlas-wide hit calling with Fi-NeMo

Out-of-Scope Use

This dataset does not contain trained model checkpoints, raw sequencing data, or processed signal tracks. Models are hosted in adamyhe/procap-atlas, and track assets are hosted in adamyhe/procap-atlas-tracks. CWMs are attribution-derived motif representations, not position frequency matrices (PFMs); they reflect model-learned sequence contributions and should not be treated as binding affinity measurements. These are research artifacts and should not be used for clinical or diagnostic decision-making.

Dataset Structure 

modisco/
├── {experiment}_{head}.modisco.h5         # TF-MoDISco results (CWMs, seqlets, metaclusters)
└── {experiment}_{head}.modisco.tar        # HTML motif report archive

motifcompendium/
├── motifcompendium_{head}_cluster_averages.h5        # Cluster-average CWMs (modisco-lite h5 format)
├── motifcompendium_{head}_cluster_averages.meme       # MEME-format cluster-average motifs
├── motifcompendium_{head}_pattern_to_cluster.tsv      # Per-experiment motif → atlas cluster mapping
├── motifcompendium_{head}_cluster_metadata.tsv        # Per-cluster stats: n_motifs, total_seqlets, experiments, JASPAR label
├── motifcompendium_{head}_cluster_report.html         # Logo-heavy cluster summary table
├── motifcompendium_{head}_cluster_summary.html        # Lightweight all-clusters table with SVG links
├── motifcompendium_{head}_cluster_logo_paths.tsv      # Cluster → logo SVG path mapping
└── motifcompendium_{head}_cluster_logos/
    ├── fwd/*.svg                                      # Forward-orientation cluster logos
    └── rev/*.svg                                      # Reverse-complement cluster logos

{experiment} is an ENCODE experiment accession (e.g., ENCSR882DWM). {head} is profile or count, corresponding to the BPNet prediction head whose DeepLIFT attributions were used for motif discovery.

Per-Experiment MoDISco Results

Each .modisco.h5 file contains the full TF-MoDISco output for one experiment and attribution head, including:

  • Discovered motif patterns organized into positive and negative metaclusters
  • Contribution weight matrices (CWMs) and hypothetical contribution scores for each pattern
  • Seqlet coordinates and per-seqlet attribution snippets
  • Pattern similarity and clustering metadata

The .modisco.tar archives contain the corresponding HTML motif reports with interactive logos, seqlet counts, and TOMTOM matches. Extract with tar xf {experiment}_{head}.modisco.tar.

Atlas-Wide Motif Compendium

The motif compendium, built with MotifCompendium, collapses per-experiment MoDISco motifs across all atlas experiments into one deduplicated set of atlas-wide clusters. Each cluster represents a motif family observed across one or more biosamples.

Key files:

  • cluster_averages.h5: Cluster-average CWMs in modisco-lite h5 format. Can be used directly as a motif set for atlas-wide hit calling (e.g., with Fi-NeMo via call_hits_bpnet.py --modisco-h5).
  • cluster_averages.meme: The same clusters in MEME format for compatibility with MEME Suite tools.
  • pattern_to_cluster.tsv: Maps each (experiment, local_motif_name) pair to its atlas-wide compendium_motif_name. Used by link_hits_to_compendium.py to relabel per-experiment hits with shared cluster identities.
  • cluster_metadata.tsv: Per-cluster summary statistics including number of contributing motifs, total seqlets, number of contributing experiments, and best JASPAR match.
  • cluster_logos/: SVG logos for each cluster in forward and reverse-complement orientations.

Dataset Creation

Source Data

MoDISco results are generated from DeepLIFT attribution scores computed on trained BPNet models from the companion model repository. Each experiment's attribution scores are computed over its processed peak regions and GC-matched negative regions across all seven chromosome folds.

Processing

  1. Attributions: BPNet DeepLIFT attributions are computed per experiment and prediction head using src/bpnet/attribute/attribute_bpnet.py.
  2. MoDISco: tfmodisco-lite discovers motifs from attribution scores using src/bpnet/modisco/launch.py. HTML reports are generated with src/bpnet/modisco/launch_report.py and tarred for upload.
  3. MotifCompendium: src/bpnet/motifcompendium/cluster_motifs.py loads all per-experiment MoDISco h5 files, builds a MotifCompendium, applies within-experiment and across-experiment similarity clustering, and exports the deduplicated cluster set.

The compendium excludes uncapped-library experiments and experiments below a minimum read-depth threshold. See src/bpnet/README.md (Motif Clustering section) for clustering parameters and variant compendium workflows.

Bias, Risks, and Limitations

  • CWMs are model-derived attribution summaries, not direct measurements of TF binding affinity. They reflect what the model learned and are subject to model limitations and training data biases.
  • ENCODE biosample coverage is uneven across cell types, tissues, and conditions. Motif recovery depends on sequencing depth, peak set size, and negative training set composition (see Supplementary Note 1 in the companion paper for an example with GATA motifs).
  • Compendium cluster identities (cluster_final IDs) are specific to the clustering run and its parameters. Different clustering thresholds, experiment subsets, or MotifCompendium versions produce different cluster IDs. The pattern_to_cluster.tsv mapping is the authoritative link between per-experiment motifs and atlas clusters.
  • The .modisco.tar report archives can be large. Extract individual archives as needed rather than unpacking all at once.

How to Use

Load a MoDISco h5 file 

import h5py

with h5py.File("modisco/ENCSR882DWM_count.modisco.h5", "r") as f:
    # List discovered patterns
    for metacluster in f["pos_patterns"]:
        pattern = f["pos_patterns"][metacluster]
        cwm = pattern["contrib_scores"][:]
        n_seqlets = pattern["seqlets"]["n_seqlets"][()]
        print(f"{metacluster}: {n_seqlets} seqlets, shape {cwm.shape}")

Load the atlas-wide compendium 

import pandas as pd

metadata = pd.read_csv(
    "motifcompendium/motifcompendium_count_cluster_metadata.tsv", sep="\t"
)
mapping = pd.read_csv(
    "motifcompendium/motifcompendium_count_pattern_to_cluster.tsv", sep="\t"
)

print(f"{len(metadata)} atlas-wide clusters")
print(metadata[["cluster_final", "n_motifs", "total_seqlets", "jaspar_label"]].head(10))

Use cluster averages for hit calling 

git clone https://github.com/kundajelab/procap-atlas.git
cd procap-atlas

python src/bpnet/hitcall/call_hits_bpnet.py \
    -e ENCSR882DWM \
    --modisco-h5 motifcompendium/bpnet/motifcompendium_profile_cluster_averages.h5

See src/bpnet/README.md in the companion repository for full hit calling, compendium relabeling, and motif analysis workflows.

Citation

If you use these motif results, please cite the PRO-cap atlas repository and the underlying ENCODE experiments (Shah et al., 2025). Please also cite the software dependencies used for motif discovery and clustering:

Contact

For questions, bug reports, or reuse notes, please use the GitHub repository issues: https://github.com/kundajelab/procap-atlas/issues
Downloads last month
34
Total file size:
148 GB

Models trained or fine-tuned on adamyhe/procap-atlas-motifs

Collection including adamyhe/procap-atlas-motifs