DenseFS

This repository contains our (re)implementation of Protein family-specific models using deep neural networks and transfer learning improve virtual screening and highlight the need for more data (DenseFS).

If you found DenseFS useful, please cite our paper:

Imrie F, Bradley AR, van der Schaar M, Deane CM. Protein family-specific models using deep neural networks and transfer learning improve virtual screening and highlight the need for more data, Journal of Chemical Information and Modeling. 2018, 58(11), 2319-2330.

This implementation utilises libmolgrid for molecular gridding.

Requirements

This code was tested in Python 3.7 with PyTorch 1.4.

A yaml file containing all requirements is provided. This can be readily setup using conda.

conda env create -f DenseFS-env.yml
conda activate DenseFS-env

Models supported

We have implemented two CNN architectures, which can be found in models.py. These are specified when using the scripts by --model / -m.

Ragoza - This refers to the three-layer CNN architecture described in Ragoza et al., 2017.

Imrie - This refers to the DenseNet-based CNN architecture described in Imrie et al., 2018.

Example usage

Train from scratch (using random weights)

python CNN_train.py -m Imrie --train_file ./data/small.types -d ./data/structs/ -i 500 -b 32 -s 42 --display_iter 50 --save_iter 500 --anneal_iter 100 --rotate --translate 2.0

Train from pretrained model (using existing/trained weights)

python CNN_train.py -m Imrie --train_file ./data/small.types -d ./data/structs/ -i 250 -b 32 -s 42 --display_iter 50 --save_iter 250 --anneal_iter 100 --weights model.iter-500 --base_lr 0.005 --rotate --translate 2.0

Test

python CNN_test.py -m Imrie --weights model.iter-250 --test_file ./data/small.types -d ./data/structs/ -b 32 -s 42 --display_iter 50 --rotate --num_rotate 4

Contact (Questions/Bugs/Requests)

Please submit a Github issue or contact Fergus Imrie imrie@stats.ox.ac.uk.