Forcefield-guided Machine Learning Methods for Flexible Ligand Docking
Abstract
The goal of this proposal is to both develop and evaluate a pipeline for combining uHTS data of small molecule binding to targets of interest with machine learning methodology, to enable rapid determination of high-affinity binders to protein targets of interest. Figure 1 presents an overview of the proposed pipeline, and representation of where each task applies. Novel to our approach is the use of a biomolecular energy model to help provide regularization for a trained deep neural network model, in which a modest amount of uHTS data is augmented with in silico data from a traditional force-field based molecular docking approach. The optimal manner in which to combine both experimental and in silico data will be discovered, and a GPU-enabled version of molecular docking will serve to enable rapid training of such a network. The ultimate goal of this project is to develop a rapid system in which uHTS data can be rapidly mined to identify properties of small- molecule binders, yielding both confidence measures on the uHTS data as well as the ability to predict the most likely inhibitors.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 19, 2022
- Source ID
- HDTRA12210012
Entities
People
- Frank DiMaio
Organizations
- Defense Threat Reduction Agency
- University of Washington