DMS/NIGMS 1: Data-driven Ricci curvatures and spectral graph for machine learning and adaptive virtual screening

Abstract Computer-aided drug design is an integrated part of modern drug discovery and is of tremendous importance. The success of computer-aided drug discovery requires a thorough understanding of molecular biophysics and intensive methodological development. Modern mathematical methods, such as those based on differential geometry, algebraic topology and graph theory, are able to provide high-level abstractions of biomolecular systems. However, these methods were rarely properly applied to select the potential drug candidates and assist the experimental studies. The main challenge is to develop the scoring functions to accurately rank poses and simultaneously predict binding affinities and other small molecular properties, such as toxicity, solubility, partition coefficient, and blood-brain barrier penetration index. The current project aims to develop novel low-dimensional representations for biomolecular data analysis from mathematics-based approaches and robustness training data to revolutionize the current practice in structure-based virtual screening. First, the PIs will introduce for the first time molecular-shape- guided persistent Ricci curvature, at the same time, will provide local geometry and spectral information, to reduce the structural complexity but still maintain an essential and adequate description of biomolecular interactions. The PIs will develop a target-ligand adaptive deep learning protocol for post-docking pose selection, binding affinity prediction, ranking, and other molecular properties estimation. Finally, the PIs will extensively validate the proposed methods on a variety of datasets to optimize the mathematical representations and learning networks. Specifically, this project will focus on the development of the proposed model for the virtual screening of Phosphodiesterase-2 (PDE2) inhibitors, providing a promising therapeutic strategy for the treatment of various human diseases. A close loop integrating computational-experimental models will further strengthen the robustness and accuracy of the proposed models. The user-friendly software packages and online servers will be developed using parallel and GPU architectures for researchers who are not formally trained in advanced mathematics or sophisticated machine learning.