Gwangju Institute of Science and Technology (GIST) announced on the 19th that a research team led by Professor Lee Hyun-joo of its Department of AI has developed InFo-DRP, an artificial intelligence (AI) model that can accurately predict cancer cell responses to new drugs not encountered during training.
The model is expected to help screen drug candidates and determine which compounds should be prioritized for testing.
Even the same anticancer drug can produce different responses depending on the genetic characteristics of cancer cells. Researchers have therefore been developing AI technologies that analyze a drug’s chemical structure and a cancer cell’s genetic information to predict how the cell will respond to the drug.
However, existing AI models have limitations. Although they can predict responses to drugs used in training with reasonable accuracy, their performance tends to decline when they encounter unfamiliar drugs or compounds with substantially different chemical structures.
To overcome this limitation, the research team focused on methods that analyze the chemical structures and properties of drugs in multiple ways and select the information needed for prediction, reducing the AI model’s excessive reliance on characteristics of drugs used during training.
The team also applied a technique known as invariant learning to reduce the influence of features specific to the training drugs and extract information that remains useful for predicting responses to new compounds. InFo-DRP was designed to predict drug responses by combining the resulting drug information with gene expression data from cancer cells.
In evaluations, InFo-DRP demonstrated higher predictive correlations than six existing AI models for both cancer cells and noncancerous cells.
“This study is significant because it improved predictive performance for new drugs that the AI had not encountered during training, as well as for drugs with chemical structures different from those of previously learned compounds,” Professor Lee said. “We expect the technology to improve the efficiency of drug candidate screening by predicting responses to potential compounds in the early stages of drug development and helping prioritize which ones should be tested.”
The study was supervised by Professor Lee, with integrated master’s and doctoral student Lee Ji-sang serving as first author. Integrated master’s and doctoral student Yeom Tae-rim and Professor Kim Jun-woo of Wonkwang University’s Department of Electronic Engineering participated as co-authors. The findings were published online on Sept. 30 in the IEEE Journal of Biomedical and Health Informatics, an international journal in the field of biomedical informatics.
Technology commercialization of the research findings can be pursued through GIST’s Technology Commercialization Center.
