MolP-PC: A Multi-View Fusion and Multi-Task Learning Framework for Drug ADMET Property Prediction

The accurate prediction of drug absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is a critical step in the early stages of drug development to reduce the risk of failure. The way deep learning is done now has problems with data sparsity and information loss because of the limitations of single-molecule representations and the isolation of predictive tasks. To solve these problems, this research suggests using MolP-PC, a multi-view fusion and multi-task deep learning framework that combines 1D molecular fingerprints, 2D molecular graphs, and 3D geometric representations. It also has an attention-gated fusion mechanism and a multi-task adaptive learning strategy to get accurate predictions of drug molecule ADMET properties. Experimental results show that MolP-PC achieves optimal performance in 27 of 54 tasks, with the multi-task learning mechanism notably improving predictive performance on small-scale datasets and outperforming single-task models in 41 of 54 tasks. Further ablation studies and interpretability analyses validate the importance of multi-view fusion in capturing multi-dimensional molecular information and improving the model's generalization ability. A case study on the anticancer compound Oroxylin A demonstrates MolP-PC’s good generalization in predicting key pharmacokinetic parameters such as half-life (T0.5) and clearance (CL), highlighting its practical potential in drug modeling. However, the model tends to underestimate volume of distribution (VD), suggesting room for improvement in handling compounds with high tissue distribution. In summary, this study provides an efficient and interpretable solution for predicting ADMET properties, offering a novel paradigm for molecular optimization and risk assessment in drug development.