A domestic research team has developed an artificial intelligence (AI) model that not only predicts whether a drug binds to a protein but also whether it actually turns the protein's function on or off. It is expected to help more accurately identify whether a drug candidate works.
KAIST said on the 8th that a research team led by Professor Lee Gwan-su of the Department of Bio and Brain Engineering developed an AI model, "GPCRact," that predicts whether drugs targeting G protein-coupled receptors (GPCRs) are active. The findings were published in January in the international journal in bioinformatics, "Briefings in Bioinformatics."
GPCRs are proteins on the cell surface that transmit external signals into the cell, responding to hormones, neurotransmitters and drugs. There are about 800 types in the human body, and 30% to 40% of marketed medicines target these proteins.
The problem is that even if a drug binds to a protein, it does not always produce the desired effect. After the drug binds, structural changes occur within the protein, and those changes must be transmitted to other regions for the actual function to be activated or suppressed. This process is called "allosteric signal propagation." Simply put, pressing one part of a protein triggers a switch-like action whose effects spread across the whole.
The researchers divided drug action into a "binding step" and a "protein internal signal transmission step," and designed the AI to learn them sequentially. They represented the protein structure as an atom-level graph and applied an AI training technique, the "attention mechanism," so it could find the key pathways through which signals pass. The attention mechanism is a Deep Learning technique in which an AI model prioritizes the most relevant parts, rather than all information, when processing data.
As a result, the model improved the prediction of drug activity even in complex protein structures that were difficult to predict using conventional methods. It did more than simply present active or inactive; it also showed which internal signaling pathways underpinned its judgment. In effect, it partially addressed the limitations of so-called "black box AI," whose results are hard to interpret.
Lee said, "We will expand to various proteins and advance the technology to predict cellular and human responses."
References
Briefings in Bioinformatics (2026), DOI: https://doi.org/10.1093/bib/bbaf719