Software for Drug-Protein Interaction with Machine Learning

Executive Summary

Developing drugs for therapeutic applications is a time consuming and costly process. Selecting potential drug candidates and understanding the potential interactions in the body is a bewildering process due to the number of potential combinations. Thus, a number of efforts have been undertaking to use high throughput testing and artificial intelligence (AI) to speed the process, but currently, these approaches still take considerable computation time. MSU researchers have recently developed a new software platform called BINDSMART that accelerates the process for identifying drug-protein interactions. When used for single drug candidate analysis, BINDSMART significantly reduces computation time, thus speeding the process for drug development.

Description of Technology

The technology encodes structure-based protein-ligand interaction data (generated by the Protein Ligand Interaction Profiler, PLIP1) into a heterogeneous graph representation that encompasses the topology and physiochemical characteristics of both the ligand (small molecule, drug) and protein residues at the protein-ligand interface. The BINDSMART encoding process enables 3D structural data to be used in neural net models. For a single candidate drug, preprocessing PLIP data and performing neural net calculations takes only minutes on an NVIDIA GeForce GTX 1650 Ti GPU. The output of the software is a probability of the compound inhibiting a protein from 0 - 100%. Software is available as python source code.

Benefits

• Graphical representation allows molecular structures to be amiable for machine learning
• Reduces computational time compared to existing processes
• Potential to supplement in vitro investigations currently required by the FDA for investigational new drug candidates
• Outperforms traditional DDI prediction models which rely solely on ligand molecular features

Applications

• Drug discovery

IP Status

Copyright protection for software product

Publications

“Predicting Inhibitors of OATP1B1 via Heterogeneous OATP-Ligand Interaction Graph Neural Network (HOLIgraph)”, Journal of Cheminformatics, 2025

Licensing Rights

Full licensing rights of software available

Inventors

Dr. Daniel Woldring, Dr. Mehrsa Mardikoraem, Joelle Eaves, Theodore Belecciu

TECH ID

TEC2025-0046