hazemessam/contact-prediction

TAPE Contact Prediction

Protein chains with 3D coordinates, for predicting which residues touch in the folded structure. This is the contact prediction task of TAPE (Rao et al., NeurIPS 2019); TAPE took the data from ProteinNet (AlQuraishi, 2019). The files are TAPE's JSON files, gzip-compressed; decompressed, they are byte-identical to TAPE's.

The task: given a chain's amino-acid sequence (primary), predict its residue–residue contact map.

Quick start

from datasets import load_dataset

default = load_dataset("hazemessam/contact-prediction", "default")
print(default["train"][0])

Files and splits

split	file	chains	sequence length (residues)
train	proteinnet_train.json.gz	25,115	20 to 999 (median 172)
validation	proteinnet_valid.json.gz	224	20 to 689 (median 163)
test	proteinnet_test.json.gz	40	75 to 670 (median 261)

Test chains are CASP12 targets (ids such as FM#T0859 and TBM#T0942).

Contacts

The labels are coordinates, not a contact map. TAPE's code (ProteinnetDataset in tape/datasets.py) builds the map as: two residues are in contact when their tertiary coordinates are less than 8.0 apart; pairs closer than 6 positions in the sequence, or involving a residue whose valid_mask is false, are ignored (set to -1).

import numpy as np
from scipy.spatial.distance import pdist, squareform

def contact_map(rec):
    contacts = np.less(squareform(pdist(rec["tertiary"])), 8.0).astype(np.int64)
    valid = np.asarray(rec["valid_mask"])
    i, j = np.indices(contacts.shape)
    contacts[~(valid[:, None] & valid[None, :]) | (np.abs(i - j) < 6)] = -1
    return contacts

Columns

field	type	role	meaning
id	string	identifier	ProteinNet record identifier
primary	string	input	amino-acid sequence, one letter per residue
tertiary	list of [3 floats] per residue	label	3D coordinates, one point per residue
valid_mask	list of bool	mask	false for residues without coordinates
evolutionary	list of [21 floats] per residue	feature	per-residue evolutionary profile from TAPE
secondary	list	unused	empty in every record
protein_length	int	metadata	length of primary

Verification

Checked on 2026-11-13 against the original source. SHA-256 of each file in this repo:

file	SHA-256	check against source
proteinnet_train.json.gz	e25c5b64b5a885816f842f9fc0f8ee648884893487f486414035156b0dc2880b	decompressed bytes identical to TAPE's .json
proteinnet_valid.json.gz	276d2beab254d4df65f61f071e290f95243c11d154cc7fb728c4a12b75ff35f4	decompressed bytes identical to TAPE's .json
proteinnet_test.json.gz	6d5d87407a137e8357d08a9ea43894ce7b378243096ff88c28b06afa4d7b89ff	decompressed bytes identical to TAPE's .json

Source, license and citation

Copied from TAPE's JSON download, http://s3.amazonaws.com/songlabdata/proteindata/data_raw_pytorch/proteinnet.tar.gz (TAPE's README: https://github.com/songlab-cal/tape#data). TAPE's repository is licensed BSD-3-Clause; the data come from the publications below. The files here are gzip-compressed to save space. Please cite TAPE and the original data source:

@inproceedings{rao2019tape,
  title     = {Evaluating Protein Transfer Learning with TAPE},
  author    = {Rao, Roshan and Bhattacharya, Nicholas and Thomas, Neil and Duan, Yan and Chen, Xi and Canny, John and Abbeel, Pieter and Song, Yun S.},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2019}
}

@article{alquraishi2019proteinnet,
  title   = {ProteinNet: a standardized data set for machine learning of protein structure},
  author  = {AlQuraishi, Mohammed},
  journal = {arXiv preprint arXiv:1902.00249},
  year    = {2019}
}

Changes

README rewritten and configs: added on 2026-11-13 (previous revision cf21bdd). Data files unchanged.