Released with End-to-End Learning on 3D Protein Structure for Interface Prediction (NeurIPS 2019) by Raphael J.L. Townshend, Rishi Bedi, Patricia A. Suriana, Ron O. Dror. The SASNet training and testing code, as well as a cleaned up version of Docking Benchmark 5 (DB5) can be downloaded here.

This repository contains processing methods for converting the raw pdb data into tfrecords containing the positive and negative interactions between all binary protein complexes in the DIPS dataset. A total of 42826 binary protein complexes will be generated. We also generate a couple intermediate representations that may be useful. Specifically, make_dataset outputs:

• parsed files: pickled pandas dataframes representing each pdb
• pair files: dill files containing interacting pairs of parsed pdbs

We also include tfrecord parsing functionality as described below.

To use the processing and parsing code, you can run make requirements to obtain most dependencies. To obtain the tensorflow dependency, if you do not have it already, you can use make tensorflow (CPU version) or make tensorflow-gpu (GPU version). We recommend you do so within a virtualenv or conda environment.

Download the raw PDB files:

rsync -rlpt -v -z --delete --port=33444 \
rsync.rcsb.org::biounit/coordinates/divided/ ./data/DIPS/raw/pdb

To process the raw pdb data into associated pair files:

python src/make_dataset.py ./data/DIPS/raw/pdb ./data/DIPS/interim

To apply the additional filtering criteria:

python src/prune_pairs.py ./data/DIPS/interim/pairs ./data/DIPS/filters/ ./data/DIPS/interim/pairs-pruned

To process the pair files into tfrecords:

python src/tfrecord.py ./data/DIPS/interim/pairs-pruned ./data/DIPS/processed/tfrecords-pruned -c 8

The DB5 dataset is provided as a fully processed set here. If, however, you wish to regenerate it, you can apply the above steps, with some additional flags and no pruning (as DB5 is already a gold-standard set):

python src/make_dataset.py ./data/DB5/raw/ ./data/DB5/interim --type=db5 --unbound
python src/tfrecord.py ./data/DB5/interim/pairs ./data/DB5/processed/tfrecords -c 8

You will want to use the parse_tf_example function in src/tfrecord.py:

import atom3.database as db
from src.tfrecord import parse_tf_example

filenames =  db.get_structures_filenames('./data/DIPS/processed', extension='.tfrecord')
tf_files = tf.data.Dataset.from_tensor_slices(
    tf.convert_to_tensor(filenames))
dataset = tf_files.interleave(tf.data.TFRecordDataset, 4)
dataset = dataset.map(parse_tf_example)