Large-Scale and Systematic Prediction of ADMET Properties through Heterogeneous Network and Kernel-Based Deep Learning

Abstract Accurately predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties remains a major challenge in accelerating drug discovery, especially for underexplored natural products (NPs). Herein, we present a unified multitask learning framework that integrated heterogeneous network (HN) embeddings and a modified kernel-based deep neural network (DNN) to simultaneously predict all ADMET properties. Our model named as hdADMET captures complex molecular interactions through HNs that incorporated compound–drug relationships, experimental validated ADMET features, and the interrelationships among ADMET properties to transform the embedding representations into a kernel-based DNN for boosting the learning ability of the prediction model. Hold-out cross-validation on benchmark data sets, including ChEMBL, BindingDB, CTRPandGDSC, and NPASS, demonstrated improved and robust accuracy through data integration. Compared with the state-of-the-art multitask models, the hdADMET model achieved competitive performance with a significantly lower computational cost. Its practical utility was further demonstrated by accurately predicting the ADMET profiles of newly identified NPs mined from the literature studies. The promising candidate with high anticancer activity, low toxicity, and favorable metabolic properties was identified for further experimental validation. The hdADMET model offered an alternative way to prioritize leading compounds for clinical translation through HN embeddings and deep learning.