Unsupervised Machine Learning for Drug Repurposing and Medical Countermeasure (MCM) Identification
Abstract
We propose an integrated research program to develop a combination unsupervised/supervised ML platform for drug repurposing and the identification of biological factors contributing to viral pathogenesis. This computational platform will leverage multidimensional experimental data that our lab will generate, both from empirical screens of large compound libraries (39,218 structurally distinct molecules) for anti-viral activity as well as in vitro microscopy experiments and in silico protein binding simulations. Our platform will combine these data with graph neural network models to predict novel anti-viral compounds effective against Zika, flu, and dengue viruses. In order to do so, a baseline, supervised graph neural network model will be developed based on chemical screens for anti-cytopathogenic activity in host cells. This model will be augmented with microscopy feature data extracted using unsupervised ML approaches, in addition to data from computational protein binding simulations, in order to predict compound mechanism of action and the specific effects of compound treatment on viral pathogenicity. Compound hits predicted in silico from our combination unsupervised/supervised ML approach will be experimentally validated in vitro for anti-viral activity using the aforementioned chemical screens, microscopy assays, and protein binding simulations. For hits exhibiting promising anti-viral activity, the effects of compound treatment on viral pathogenicity will be further studied by applying nonnegative matrix factorization to transcriptomic data in order to identify gene expression networks and pathogenicity and toxicity factors of interest. We will iterate this approach and re-train our integrated ML platform in order to improve neural network prediction accuracy and better identify feature information from validation microscopy, protein binding, and transcriptomics experiments. Taken together, our combination unsupervised/supervised ML approach will discover novel anti-viral compounds, elucidate the effects of these compounds on viralpathogenicity, and propose lead compounds for development into MCMs.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 19, 2022
- Source ID
- HDTRA12210032
Entities
People
- James J. Collins
Organizations
- Defense Threat Reduction Agency
- Massachusetts Institute of Technology