data.md

RAPPPID Data

Szymborski & Emad Datasets

Szymborski & Emad describes a C-type dataset based on STRING H. sapiens protein pairs, and randomly sampled negative examples. For details, refer to the manuscript.

The data can be downloaded from the Internet Archive at this link, from this mirror hosted at OVH, or using the Academic Torrents BitTorrent file.

You can also download the datasets used for the BioLip transfer learning analysis at the Internet Archive, from this mirror hosted at OVH, or using this torrent.

Prepare Datasets for RAPPPID

RAPPPID expects four files when training a new dataset.

1. train.pkl.gz - Contains training pairs of protein IDs and a binary interaction label.
2. val.pkl.gz - Contains validation pairs of protein IDs and a binary interaction label.
3. test.pkl.gz - Contains testing pairs of protein IDs and a binary interaction label.
4. seqs.pkl.gz - Which is a map between protein IDs and their amino acid sequences.

Each of these files are expected to be a gzipped pickle file.

{train, val, test}.pkl.gz are pickled arrays of 3-tuples (tripels). The first two elements are two protein ids, and the last element is either a 1 (known interaction), or a 0 (no known interaction).

Here's an example of python code that creates a valid train.pkl.gz comprised of two pairs. val.pkl.gz and test.pkl.gz are created identically.

import gzip
import pickle

train_pairs = [
    #(protein_id_a, protein_id_b, label)
    ("PROTEIN_ID_ONE", "PROTEIN_ID_TWO", 1),
    ("PROTEIN_ID_THREE", "PROTEIN_ID_FOUR", 0)
]


with gzip.open('train.pkl.gz', mode='wb') as f:
    pickle.dump(train_pairs, f)

The seqs.pkl.gz file is created similarly. A dictionary of protein ID keys and encoded amino acid sequences. A function below is provided to encode the amino acid sequences.

Special amino acid codes X, Z, B are replaced with a randomly selected amino acid from the pool the represent. See the encode_seq function for more details.

def get_aa_code(aa):
    # Codes based on IUPAC-IUB
    # https://web.expasy.org/docs/userman.html#AA_codes

    aas = ['PAD', 'A', 'R', 'N', 'D', 'C', 'Q', 'E', 'G', 'H', 'I', 'L', 'K', 'M', 'F', 'P', 'S', 'T', 'W', 'Y', 'V', 'O', 'U']
    wobble_aas = {
        'B': ['D', 'N'],
        'Z': ['Q', 'E'],
        'X': ['A', 'R', 'N', 'D', 'C', 'Q', 'E', 'G', 'H', 'I', 'L', 'K', 'M', 'F', 'P', 'S', 'T', 'W', 'Y', 'V']
    }

    if aa in aas:
        return aas.index(aa)

    elif aa in ['B', 'Z', 'X']:
        # Wobble
        idx = randint(0, len(wobble_aas[aa])-1)
        return aas.index(wobble_aas[aa][idx])

def encode_seq(seq):
    return [get_aa_code(aa) for aa in seq]

seqs = {
    # protein_id: encoded amino acid seqs
    'PROTEIN_ID_ONE': encode_seq('MANL'),
    'PROTEIN_ID_TWO': encode_seq('MANG'),
    'PROTEIN_ID_THREE': encode_seq('MANW'),
    'PROTEIN_ID_FOUR': encode_seq('MANA'),
}

with gzip.open('seqs.pkl.gz', mode='wb') as f:
    pickle.dump(seqs, f)

With the above examples, you should be able to create your own datasets quite easily for RAPPPID.

Pre-Trained Weights

See the /data/pretrained_weights folder for pretrained weights for RAPPPID.