Matt Strang
Research Mentor: Furyal Ahmed
Mentor Department: Biophysics, LSA
Author(s): Matt Strang, Furyal Ahmed
Session: Session 6 (3:00 PM – 3:50 PM)
Presentation Type: Poster 45
Abstract
Computational modeling and molecular dynamics have become essential tools in accelerating the drug discovery process. Traditional experimental screening is often slow, expensive, and limited in scope, motivating the development of physics-based docking methods to make virtual screening faster, more accurate, and more accessible across a wide range of diseases. CDOCKER is a well-established docking tool that leverages molecular dynamics and grid-based energy calculations to evaluate ligand–protein interactions, achieving competitive performance across many target classes. However, benchmarking on the DUD-E diverse set reveals that CDOCKER achieves near-random performance in virtual screening, with an average AUC of 0.49, motivating the need for an optimized scoring function with improved discriminatory power. This project focuses on improving molecular docking workflows using CDOCKER by optimizing its scoring function through the identification of an improved set of parameter weights to enhance predictive accuracy. CDOCKER is used to evaluate ligand–protein interactions using experimentally determined structures from the Protein Data Bank (PDB). Parameter weights are tuned on the DUD-E benchmark to enable comparison with prior computational work, and then validated on the LIT-PCBA dataset, which contains experimentally confirmed high-throughput screening results, providing a real-world measure of predictive performance. We hypothesize that reweighting the CDOCKER scoring function, particularly correcting for systematic overestimation of van der Waals contributions, will improve ligand rank-ordering and yield higher ROC-AUC scores relative to the default, with direct implications for virtual screening in drug discovery. Preliminary docking results are presented in support of this hypothesis.
