Interpretable Multimodal Molecular Language Model for Drug-Target Interaction Prediction.

Accurate prediction of drug-target interactions (DTI) plays a vital role in accelerating drug discovery through multimodal data integration. While deep learning has shown significant potential for DTI prediction, its effectiveness is fundamentally limited by the scarcity of labeled training data, due to the expensive and time-consuming nature of experimental DTI validation. This constraint substantially hinders the full utilization of deep learning capabilities in computational drug discovery. Therefore, we propose an interpretable multimodal molecular language network to enhance DTI prediction, named M3LNet. Specifically, we first construct the multimodal network of the multiple perspectives of a drug, and then extract both the multimodal features of the drug and the features of the target sequence. Then, we employ a multikernel approach to learning different distributions of multimodal data. After that, we integrate large language models into the heterogeneous neural network, enabling knowledge-aware representation learning and contextual interaction prediction to label potential labeled data of DTI. Extensive experiments on publicly accessible datasets demonstrate that M3LNet significantly outperforms state-of-the-art methods on various tasks. Furthermore, we also visualize the attention weights of DTI to interpret DTI prediction, aiding researchers in understanding the mechanisms behind potential DTI. In summary, M3LNet provides an effective and interpretable approach for DTI prediction, facilitating both drug discovery and rational drug design.