Deep Learning for Virtual Screening

Dependency

• deepchem 0.0.5.dev2704

  • https://github.com/deepchem/deepchem

• Tensorflow v1.3.0-rc2-20-g0787eee

• Anaconda 4.3

• Python 3.6.3

Method

• Multi-task DNN and Graph convolutional network

• 5-fold CV

Usage

$ make

...
 
PCBA dataset
Multi-task DNN
Train scores: 0.819380206847
Test scores: 0.789931466539
Train time: 608.6m
Graph-Convolution
Train scores: 0.889408119487
Test scores: 0.850056521039
Train time: 1958.0m

Tox21 dataset
Multi-task DNN
Train scores: 0.848936655301
Test scores: 0.789951449973
Train time: 2.2m
Graph-Convolution
Train scores: 0.90177625967
Test scores: 0.825843831165
Train time: 14.6m

Delaney dataset
Multi-task DNN
Train scores: {'mean-pearson_r2_score': 0.86165358015042837}
Validation scores: {'mean-pearson_r2_score': 0.61665541324854578}
Train time: 0.1m
Eval time: 0.0m
Graph-Convolution
Train scores: {'mean-pearson_r2_score': 0.97126488150811363}
Validation scores: {'mean-pearson_r2_score': 0.83998886621722046}
Train time: 0.2m
Eval time: 0.0m