swhitfield/tape-proteinnet_contact_prediction

tape-proteinnet-contact-prediction

Sourced from TAPE ProteinNet LMDB splits.

Intended use

Protein residue-contact prediction from sequence. Contacts report C-alpha proximity in the source structure; they are structural labels, not sequence-only measurements.

Data files

Parquet files are stored under data/ with Hugging Face split names. TAPE valid is published as validation; train_unfiltered is omitted by default and included only when explicitly configured.

Dataset statistics are available in `stats.json` at the dataset root. The file includes row counts by split, columns, and SeqKit sequence-length metrics.

Contact definition

Positive contacts are upper-triangle residue pairs (including self-pairs) with valid coordinates and C-alpha Euclidean distance below 8 Angstroms. Rows store raw pairs in targets and valid residue indices in valid_positions; sequence-separation filtering (at least 6) is applied downstream by normalize_contact_record.

Processing and splits

• —Source type: lmdb
• —Local LMDB inputs: proteinnet_test.lmdb, proteinnet_train.lmdb, proteinnet_valid.lmdb
• —include_train_unfiltered: false. Omitted.
• —create_split_subsets: true. Canonical splits are retained and pooled _random, _stratified, and _cluster alternatives are generated with seed 1957723; cluster identity is 0.3, coverage is 0.8, and threads is 1.
• —Card generated (UTC): 2026-09-24

Output row counts:

• —test: 40 rows
• —test_cluster: 40 rows
• —test_random: 40 rows
• —test_stratified: 40 rows
• —train: 25299 rows
• —train_cluster: 25299 rows
• —train_random: 25299 rows
• —train_stratified: 25299 rows
• —validation: 224 rows
• —validation_cluster: 224 rows
• —validation_random: 224 rows
• —validation_stratified: 224 rows

Provenance

• —TAPE project and LMDB guidance: https://github.com/songlab-cal/tape#lmdb-data
• —Archive fallback: http://s3.amazonaws.com/songlabdata/proteindata/data_pytorch/proteinnet.tar.gz
• —Dataset family: TAPE ProteinNet contact prediction