Core Insights - Dongyangguang Pharmaceutical's AI research team has developed multiple self-innovated models for optimizing drug molecular ADME/T properties, achieving an internal AUROC of 0.90 for their drug permeability/transport prediction model, significantly outperforming public open-source models [1] Group 1: Drug Permeability/Transport Prediction - The drug permeability and transport prediction model addresses high-cost, scarce data and small sample learning scenarios, crucial for understanding the impact of biological membranes and transporters on oral drug absorption [1] - The use of machine learning for proprietary data modeling allows for rapid and cost-effective predictions of drug interactions with biological membranes and transporters, facilitating early optimization of pharmacokinetic properties [1] Group 2: Multi-Task Learning Strategy - The HEC-Transporters model employs a multi-task learning strategy for joint modeling of permeability and transport tasks, with 80% of samples being shared across tasks to capture common structural features [2][4] - Internal benchmark tests show that the multi-task learning model achieves an average AUC of 0.90, improving by 0.33 over single-task models and 0.19 over baseline models, with the highest accuracy of 93% in membrane permeability tasks [4][6] Group 3: Technological Innovation and Application Value - HEC-Transporters is the first international predictive system for drug permeability/transport using multi-task learning, enhancing performance while addressing the limitations of small proprietary task data [7] - Since implementing the AI+ strategy in 2023, Dongyangguang Pharmaceutical has established a comprehensive AI research and development system covering target prediction, compound screening, lead optimization, and PK modeling, reducing new drug development costs and enhancing industry efficiency [7]
东阳光药(01558)AI研发团队发布HEC-Transporters模型 为早期药物研发提供全流程的药代动力学性质优化