AI Boosts Accuracy in Predicting Responses to Unseen Drug Candidates

According to AI TIMES, a GIST research team on October 19 unveiled an AI model called INFO-DRP that predicts responses not only to new drug candidates that were not used in training, but also to substances structurally different from existing drugs. In strict evaluations, the model posted a prediction correlation coefficient of 0.592, 35.8% higher than the best-performing comparison model. The team said the key to the performance gain was filtering out only the information useful for actual predictions from among a drug's many chemical features and examining it together with gene expression data from cancer cells. It explained that this approach aims to reduce the limitation of existing models, which tend to rely too heavily on drug features seen during training and weaken when applied to new substances. The team added that additional verification using separate data and non-cancer cell lines also showed strong results. However, it noted that such predictions are only an early-stage tool for narrowing candidates, and that final judgments on efficacy and safety must be confirmed through follow-up experiments and clinical trials.