Predicting drug-disease associations (DDAs) is a crucial task in drug discovery and repositioning, as it enables the identification of novel therapeutic indications for existing drugs. However, current technologies for predicting these associations face limitations, such as challenges in data integration, difficulty in capturing multi-relational biological interactions, and limited generalizability. To address these issues, we present a network-based framework that integrates heterogeneous biological data-including drugs, proteins, biological functions, and diseases-into a unified graph structure. To effectively model the complex and multi-relational nature of this network, we employ a Relational Graph Convolutional Network (RGCN), which explicitly captures various types of biological interactions. Our model named DDA-RGCN achieves superior predictive performance, with an AUC of 0.9745 and an AUPR of 0.9728, outperforming several baseline and state-of-the-art methods. Case studies further demonstrate the model's potential in uncovering novel drug indications, supported by literature evidence. These findings underscore the value of combining heterogeneous biological networks with graph neural network techniques and highlight the potential of our RGCN-based approach to advance computational drug repositioning and support personalized medicine initiatives.

Understanding the biological functions of molecules in specific human tissues or cell types is crucial for gaining insights into human physiology and disease. To address this issue, it is essential to systematically uncover associations among multilevel elements consisting of disease phenotypes, tissues, cell types and molecules, which could pose a challenge because of their heterogeneity and incompleteness. To address this challenge, we describe a new methodological framework, called Graph Local InfoMax (GLIM), based on a human multilevel network (HMLN) that we established by introducing multiple tissues and cell types on top of molecular networks. GLIM can systematically mine the potential relationships between multilevel elements by embedding the features of the HMLN through contrastive learning. Our simulation results demonstrated that GLIM consistently outperforms other state-of-the-art algorithms in disease gene prediction. Moreover, GLIM was also successfully used to infer cell markers and rewire intercellular and molecular interactions in the context of specific tissues or diseases. As a typical case, the tissue-cell-molecule network underlying gastritis and gastric cancer was first uncovered by GLIM, providing systematic insights into the mechanism underlying the occurrence and development of gastric cancer. Overall, our constructed methodological framework has the potential to systematically uncover complex disease mechanisms and mine high-quality relationships among phenotypical, tissue, cellular and molecular elements.

SUMMARY: Although many quantitative structure-activity relationship (QSAR) models are trained and evaluated for their predictive merits, understanding what models have been learning is of critical importance. However, the interpretation and visualization of QSAR model results remain challenging, especially for 'black box' models such as deep neural network (DNN). Here, we take a step forward to interpret the learned chemical features from DNN QSAR models, and present VISAR, an interactive tool for visualizing the structure-activity relationship. VISAR first provides functions to construct and train DNN models. Then VISAR builds the activity landscapes based on a series of compounds using the trained model, showing the correlation between the chemical feature space and the experimental activity space after model training, and allowing for knowledge mining from a global perspective. VISAR also maps the gradients of the chemical features to the corresponding compounds as contribution weights for each atom, and visualizes the positive and negative contributor substructures suggested by the models from a local perspective. Using the web application of VISAR, users could interactively explore the activity landscape and the color-coded atom contributions. We propose that VISAR could serve as a helpful tool for training and interactive analysis of the DNN QSAR model, providing insights for drug design, and an additional level of model validation. AVAILABILITY AND IMPLEMENTATION: The source code and usage instructions for VISAR are available on github https://github.com/qid12/visar. CONTACT: [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.