Gwangju Institute of Science and Technology (GIST) said on Aug. 6 that a research team led by Professor Lee Hyun-joo of its AI Department has developed an AI model called “Alpha DTA” that can predict the binding affinity of drug candidates to target proteins without experimentally determining protein-drug binding structures.
Alpha DTA focuses on “internal representations” generated by AlphaFold3, Google DeepMind’s AI model for predicting protein structures, as it predicts the binding structures of proteins and drugs.
Internal representations are numerical information that enables AI to learn the chemical and structural characteristics of proteins and drugs. They consist of “single representations,” which capture the characteristics of each protein and drug, and “pair representations,” which capture the relationships and interactions between the two molecules.
The research team developed Alpha DTA by combining a “multi-resolution encoder” that analyzes these single and pair representations with a “geometric encoder” that analyzes three-dimensional structural information from protein-drug complexes.
Alpha DTA takes protein sequences and drug information as inputs and analyzes both the three-dimensional structures and internal representations generated by AlphaFold3 to predict binding affinity.
When the researchers evaluated its predictive performance using new protein-drug combinations that were not included in the training data, Alpha DTA outperformed existing sequence-based methods. It also demonstrated predictive performance comparable to existing structure-based approaches, even without experimentally determined structural data.
The technology is expected to improve the efficiency of screening drug candidates and reduce the time and cost of drug development while maintaining high predictive performance for candidates lacking experimentally determined structural data. It could also have broad applications in drug repurposing research.
“This study is significant because it demonstrates that the binding affinity of drug candidates can be predicted with high accuracy without experimentally determining protein-drug binding structures,” Professor Lee said. “We expect the technology to help efficiently identify promising candidates for a wide range of target proteins with limited structural data and contribute to accelerating and improving the efficiency of drug development.”
The study, led by Professor Lee and conducted with master’s student Jung Min-jae as the first author, was published online July 21 in the international journal Journal of Cheminformatics, which specializes in cheminformatics.
Discussions regarding technology licensing can be conducted through GIST’s Technology Commercialization Center.