Comparing and Validating Machine Learning Models for Mycobacterium tuberculosis Drug Discovery

  • Thomas Lane, Daniel P. Russo, Kimberley M. Zorn, Alex M. Clark, Alexandru Korotcov, Valery Tkachenko, Robert C. Reynolds, Alexander L. Perryman, Joel S. Freundlich, Sean Ekins
  • Molecular Pharmaceutics, April 2018, American Chemical Society (ACS)
  • DOI: 10.1021/acs.molpharmaceut.8b00083

Comparing Machine learning methods with Tuberculosis in vitro data

What is it about?

We curated a dataset of compounds tested against TB and used this to compare different classic machine learning algorithms with DL. We used a wide range of metrics, different fingerprints as well as internal and external test sets.

Why is it important?

TB kills around 1.7M people every year. There are issues with drug resistance and we need to discover new drugs. There is plenty of in vitro data and our rationale has been that we can learn from it and find new compounds to fill the pipeline. In this study we show that Bayesian was as good or better than deep learning on external test sets that were diverse.

Perspectives

Dr Sean Ekins
Collaborations in Chemistry

The visibility of deep learning makes this a must try technology but we are finding that in many cases other classic methods may do just as well for making predictions on new molecules. We may be able to use the models we have to find new compounds. To be continued....

Read Publication

http://dx.doi.org/10.1021/acs.molpharmaceut.8b00083

The following have contributed to this page: Dr Sean Ekins