Enhancing ADMET property predictions using cross-aligned multimodal attention mechanisms

Accurate prediction of drug metabolism and pharmacokinetics (ADMET) properties is crucial in drug discovery. Here, we present a novel approach to enhance ADMET property predictions using Cross-Aligned Multimodal Attention (CMA) mechanisms, pretrained models, and multimodal techniques. ADMET data is collected and processed using image processing, graph neural networks, and chemical fingerprinting. Pretrained models like GROVER and ResNet generate a multi-channel data format, and the CMA mechanism aligns and correlates the data modalities. Grad-CAM technology interprets the model's predictions, visually demonstrating the relationship between compound properties and fragments. Our ADMET property prediction server ( http://guolab.mpu.edu.mo/CMA ) implements the CMA-based model and a substantial language model for ADMET property prediction. The innovation lies in the integration of multimodal data, the application of pretrained models, and the development of cross-modal alignment. This approach improves the efficiency and accuracy of ADMET property predictions and opens new avenues for research in molecular science, particularly in drug design and evaluation.