KAIST

<(From Left) Ph.D candidate Hyojin Son, Professor Gwan-su Yi>

Proteins in our body function like switches. When a drug binds to a protein, the structure at the binding site changes, and this structural change propagates throughout the protein, turning its function on or off. Google DeepMind’s AlphaFold3 successfully predicted whether drugs bind to proteins and the three-dimensional structure of binding sites. However, it could not predict how signals propagate inside the protein after drug binding, how the entire structure changes, or whether the protein’s function is ultimately activated or inhibited. KAIST researchers have developed an AI that predicts not only whether a drug binds but whether it actually works.

KAIST (President Kwang Hyung Lee) announced on the 4th of March that a research team led by Professor Gwan-Su Yi of the Department of Bio and Brain Engineering has developed an artificial intelligence model called “GPCRact” that predicts whether candidate molecules not only bind to G-protein-coupled receptors (GPCRs)—a major drug target—but also actually activate the protein.

GPCRs act as “signal receivers” on the surface of cells. When hormones, neurotransmitters, or drugs send signals from outside the cell, GPCRs function as gates that receive these signals and transmit them into the cell. There are about 800 types of GPCRs in the human body, and roughly 30–40% of currently marketed drugs target them. They are key proteins involved in numerous physiological functions, including heart rate regulation, blood pressure control, pain sensing, immune responses, and emotional regulation.

However, a drug binding to a GPCR does not always trigger the desired biological function. Structural changes inside the protein and subsequent signal transmission determine whether the drug actually produces an effect. This process is known as allosteric signal propagation.

The research team designed the AI to learn the drug action process in two stages: ① the drug–target binding stage, and ② the intracellular signal propagation stage within the protein. The three-dimensional protein structure was represented as an atom-level graph, and an attention mechanism was applied to enable the model to learn important signaling pathways. Through this approach, the AI analyzes not only the drug binding signal but also the internal signaling pathways of the protein to predict whether the protein becomes activated.

As a result, the model significantly improved the prediction performance of drug activity even in proteins with complex structures that existing models struggled to analyze. Importantly, the model does not simply output “active” or “inactive.” It also presents the key internal signaling pathways that form the basis of its predictions, overcoming the limitations of so-called “black-box AI.”

<Schematic diagram of drug activity prediction and mechanism interpretation using the GPCRact artificial intelligence model>

This represents an important advance, as it allows researchers to interpret and verify predictions while simultaneously improving the reliability and efficiency of drug discovery. In the future, the model is expected to serve as a precision drug discovery AI platform capable of predicting not only whether drugs bind to GPCRs but also whether they truly activate them in various diseases targeting GPCRs.

<AI-generated image to help illustrate the research>

Professor Gwan-Su Yi explained, “Allosteric structural change refers to a phenomenon in which a drug binds to one part of a protein and its influence propagates internally, altering the function of other regions,” adding, “The key contribution of this research is incorporating this operational principle into deep learning.” He further noted, “We plan to expand the model to various proteins and ultimately develop technologies capable of predicting cellular and whole-body responses.”

Ph.D candidate Hyojin Son participated as the first author in this study. The paper was published on January 15 in the international journal Briefings in Bioinformatics, one of the leading journals in the field of bioinformatics.
※ Paper title: “GPCRact: a hierarchical framework for predicting ligand-induced GPCR activity via allosteric communication modeling”
DOI: https://doi.org/10.1093/bib/bbaf719
※ Author information: Hyojin Son (KAIST, first author), Gwan-Su Yi (KAIST, corresponding author)

This research was supported by the Basic Research Program for Individual Research funded by the Ministry of Science and ICT and the National Research Foundation of Korea (RS-2025-24533057).