Multimodal machine learning and deep graph neural networks for the prediction of molecular inhibitory activity and disease associations.

Effectively integrating a variety of biological and chemical data sources is essential for the acceleration of drug discovery and precision diagnostics. Nevertheless, conventional computational methods frequently experience modality isolation, which results in their inability to simultaneously capture the intricate multi-omics regulatory landscape of diseases and the detailed electronic properties of potential drug molecules.. State-of-the-art models typically depend on one-dimensional SMILES representations or two-dimensional molecular signatures (e.g., ECFP4). Although these representations are computationally efficient, they substantially compress or disregard critical three-dimensional spatial arrangements and electronic characteristics that are essential for precise molecular interaction analysis. In addition, numerous existing drug-disease association models fail to account for the multi-scale topological regulatory signals that are present in heterogeneous biological networks. In complex maladies, such as calcific aortic valve disease (CAVD) and multidrug-resistant infections, this limitation results in diminished predictive performance. This study suggests a unified multimodal deep learning framework that incorporates molecular property prediction with genomic target identification to overcome these challenges. To be more precise, we present an Improved Convolutional Neural Network (ICNN) that has been optimised for omics-based target detection using Honey Bee Mating Optimisation (HBMO). Additionally, a Multi-scale Diffusion Graph Convolutional Network (MsDGCN) is implemented to significantly enhance the modelling of drug-disease relationships. Quantum-calculated three-dimensional electron density grids, which incorporate approximately 125 million data points, are a significant innovation of this work. By converting these grids into structured point cloud representations, it is possible to accurately characterise non-covalent interaction (NCI) regions with high resolution. Kernel Principal Component Analysis (KPCA) is executed to reduce the dimensionality of multimodal data. In addition, an attention-based fusion layer is intended to dynamically prioritise contributions from electronic, sequential, and structural data modalities.In an effort to enhance the robustness of the model, a hard negative mining strategy is implemented to more effectively differentiate between physiologically realistic confounding samples, thereby resolving class imbalance concerns. The proposed framework's successful scalability and robustness across numerous benchmark datasets are illustrated by experimental evaluations. Compared to conventional ECFP4-based models, the incorporation of three-dimensional electronic descriptors leads to a 9.1% increase in Area Under the Curve (AUC). Peak AUC of 0.98 is achieved by the fully fused framework. In addition, the biological relevance of the predictions is confirmed by molecular dynamics simulations that were conducted over a period of 50 ns. HLA-DRA is identified as a novel therapeutic target in CAVD by the framework, and stable natural product inhibitors, such as Chrysin, are predicted. Overall, the proposed system offers a scalable and high-precision platform for the real-world virtual drug screening and clinical biomarker discovery.