This is a Python 3 version of Jinbo Xu's code for his recent paper "Distance-based protein folding powered by deep learning", PNAS August 20, 2019 116 (34). Sharing as per Prof. Xu's request.

Paper

Original code

Data (registration required)

Minor fixes:

• Indentation errors.
• Indexing error in the function utils.MidpointFeature().
• Replaced opt package by argparse for parsing arguments.
• Displaying inference progress.

Data preparation

After registration with an academic email, follow the link provided to access datasets and trained models.

Assume the working directory is 'DL4DistancePrediction2'.

Input data

Input for the 2D-Dilated Resnet model are pre-processed contact features for protein sequence, in .pkl format.

For example, '76CAMEO.2015.contactFeatures.pkl'.

Download the data to local folder, e.g., './data'

Trained models

There are 2 pretrained 2D-Dilated Resnet models provided: 'RXContact-DeepMode11410.pkl' and 'RXContact-DeepModel10820.pkl'.

Download them to local folder, e.g., './models'

Running inference

From the working folder, run the main script run_distance_predictor.py for inference:

python run_distance_predictor.py -m modelfiles -p predfiles [-d save_folder] [-g ground_truth_folder]

Parameters:

modelfiles: Specify one or multiple model files in PKL format, separated by semicolon.

predfiles: Specify one or multiple input feature files to be predicted. File(s) in PKL format, separated by semicolon in case of multiple input files.

save_folder (optional): Specify where to save the result files.

ground_truth_folder (optional): Specify the ground truth folder containing all the native atom-level distance matrix. When this option is provided, contact prediction accuracy will be calculated.

Example:

python run_distance_predictor.py -p data/76CAMEO.2015.contactFeatures.pkl -m models/RXContact-DeepMode11410.pkl -d result/76CAMEO.2015