pfnet-research / bmi219-2017-deepqsar

UCSF BMI219 Deep Learning (2017), Coding example (QSAR with Deep multitask learning)

Multitask learning with multi layer perceptron for QSAR

This is an example of the application of deep learning to Quantitative structure–activity relationship (QSAR) prediction.

The implementation is based on [1], but some minor modifications are applied. We use the PubChem dataset of chemical compounds and assay outcomes, and Chainer to build, train, and evaluate deep learning models.

See commentary.md for a detailed explanation.

Dependency

• Chainer
• NumPy
• pandas
• RDKit
• scikit-learn
• six
• nose (for testing only)

Usage

$ PYTHONPATH="." python tools/train.py

The training runs on CPU by default. If you want to run the program on GPU, add --gpu <GPU ID> option. Run python tools/train.py --help to see the complete list of options.

Q. What is PYTHONPATH? Why do we need to specify it?

Test

Run all tests including GPU ones.

PYTHONPATH="." nosetests tests

Without GPU tests

PYTHONPATH="." nosetests -a '!gpu' tests

Reference

[1] Dahl, G. E., Jaitly, N., & Salakhutdinov, R. (2014). Multi-task neural networks for QSAR predictions. arXiv preprint arXiv:1406.1231.