MTAN-ADMET: A Multi-Task Adaptive Neural Network for Efficient and Accurate Prediction of ADMET Properties

Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties is essential in early-stage drug discovery but remains challenging due to data sparsity, imbalance, and the complexity of existing modeling approaches. We present MTAN-ADMET, a Multi-Task Adaptive Network designed to predict multiple ADMET endpoints directly from SMILES representations without the need for molecular graph preprocessing or extensive feature engineering. Leveraging pretrained continuous molecular embeddings and incorporating adaptive learning techniques—such as task-specific learning rates, gradient noise perturbation, and dynamic loss scheduling—MTAN-ADMET effectively balances regression and classification tasks within a unified framework. Benchmarking on 24 ADMET endpoints shows that MTAN-ADMET achieves performance on par with or exceeding state-of-the-art graph-based models, particularly in difficult toxicity predictions like cardiotoxicity and cytochrome P450 inhibition. The model demonstrates strong generalizability and robustness in handling sparse and imbalanced datasets, offering a scalable and efficient solution for comprehensive ADMET property prediction.