KAIST researchers have developed an artificial intelligence model that can predict not only whether drugs bind to proteins but also whether they will actually work—a capability that even Google's AlphaFold3 drug discovery tool has yet to achieve.
KAIST announced on the 8th that a research team led by Professor Lee Kwan-soo of the Department of Bio and Brain Engineering developed "GPCRact," an AI model that predicts whether drug candidates actually activate G-protein-coupled receptors (GPCRs), a major drug target.
The human body contains approximately 800 types of GPCRs, which are key proteins involved in various physiological functions including heart rate regulation, blood pressure control, pain detection, immune response, and emotional regulation. About 30-40% of currently marketed drugs target GPCRs.
However, drug binding to a GPCR does not necessarily trigger the desired function. Structural changes and signal transduction processes occurring inside the protein after binding determine whether the drug actually works.
The research team divided the drug action process into two stages: drug-target binding and internal protein signal propagation. They designed the AI to learn each stage sequentially. The team represented three-dimensional protein structures as atomic-level graphs and applied an "attention mechanism" to enable learning of critical signal propagation pathways. This allows the AI to predict protein activation by identifying both drug binding signals and internal signal propagation pathways.
As a result, drug activity prediction performance improved significantly even in proteins with complex structures that challenged existing models. GPCRact is expected to serve as a precision drug development AI platform capable of predicting actual drug activity across various diseases targeting GPCRs.