In the contemporary era, the expeditious development of de novo new drugs to address infectious diseases such as COVID-19 is typically impractical due to prolonged cycles and exorbitant costs¹. With the primary goal of identifying novel indications for existing drugs, drug repositioning proves valuable in abbreviating development timelines and mitigating the risk of toxicity-related clinical effects². By narrowing down potential drug candidates, this approach is positioned as a promising alternative to the intricate process of de novo drug design³.

In the past decade, deep learning has emerged as a pivotal tool in drug repositioning, divided into two main approaches⁴: chemical structure-based and biomedical network-based methods. The former captures sequence information from drugs and viruses⁵ but struggles with oversimplified molecular representations that miss spatial atomic details and higher-order interactions. Meanwhile, network-based methods extract complex interactions from biomedical network structures⁶ but fail to quantify the interaction relationship between the micro-view drug spatial network and the macro-view biomedical network, limiting the understanding of drug substructures critical for activity. Furthermore, existing network-based methods commonly suffer from three limitations: homogeneous information sharing among heterogeneous nodes, short-range dependence, and feature over-smoothing with deeper layers⁷.

Recent research demonstrates that a multi-granularity hierarchical structure yields more accurate representations⁸. As a critical study object in the field of drug design, molecular 3D spatial structure analysis plays a critical role in bioactivity exploring⁹ and drug development¹⁰. Therefore, to precisely predict VDA based on the effective quantification and integration of 3D spatial network information and biomedical network information, we have devised a unified “spatial hierarchical network” structure, embedding the spatial network of the single drug as a subnetwork of the VDA network to describe the hierarchical structure, distinct from the existed community hierarchical network structure (Details in Methods Section 4.1 of the article). Specifically, we substitute the drug entity within the VDA network with the 3D structure network of molecules, thereby describing a micro-network hierarchy at the atom level. Alternatively, we conceptualize the complete 3D network structure of the molecule as a drug node within the VDA network, thereby forming a macro-network hierarchy at the entity level. This structure models the interaction processes between atom-level drug spatial information and entity-level biomedical association information. Unlike prior approaches that concatenate diverse features, modeling the interaction of multi-hierarchy information can effectively mitigate knowledge gaps and biases^{11, 12}.

We introduce a Spatial Hierarchical Network for Virus-Drug Associations (SpHN-VDA), a pipeline tailored to our spatial hierarchical network structure for VDA prediction (Fig. 1). Built on spatial and meta-path graph neural networks, SpHN-VDA embeds atom-level hierarchies into an entity-level framework, effectively modeling interactions between drug spatial and biomedical networks. By employing triple attention mechanisms, it captures implicit representations within and across hierarchical layers, providing a complete reasoning path from 3D molecular structure to biological associations. In addition, our model further alleviates three common issues inherent in network-based methods. Overall, SpHN-VDA utilizes its architectural superiority to establish a robust machine understanding of VDAs, guiding and promoting the screening of candidate compounds for antiviral drugs by confident predictions. 

In this study, we conduct experiments on three benchmark datasets to evaluate SpHN-VDA’s predictive capabilities in virus-drug associations (VDAs) across three key perspectives: (1) overall predictive performance on random split data with various positive-to-negative sample ratios; (2) generalization capacity to infer associations between drugs and previously unseen viruses; and (3) model robustness under random perturbations of VDA pairs. SpHN-VDA seamlessly integrates molecular spatial structure information with biological functional interaction data, optimizing them cohesively to extract advanced entity features. Experiments prove this methodology ensures not only minimized embedding distances for drugs sharing analogous biological functions at a macro scale but also effectively captures feature dependencies among long-range atoms at a micro level. SpHN-VDA employs triple attention mechanisms to understand VDAs thoroughly. It identifies crucial biological network neighbors and their key spatial structures, guiding drug screening. To assess whether both hierarchical information can regulate each other to promote reasonable explanations, we analyzed the prediction results of the model using two independent hierarchical information sources and examined the flow process of information through these hierarchies (Fig. 2). The analysis indicates a significant difference in more than half of the predictions between atom-level module and entity-level module, which can be avoided by our unified framework. 

The design of SpHN-VDA is rooted in the understanding that complex real-world phenomena demand a hierarchical approach to knowledge representation and problem-solving. We recognized a significant limitation in traditional biomedical modeling using heterogeneous networks. Hence, we developed the Spatial Hierarchical Network, embedding the spatial network of a single drug as a subnetwork within the VDA network to represent hierarchical complexity. Then, we further proposed SpHN-VDA to enhance the capability of capturing spatial structures at the atom level, preserving pharmacologically relevant properties while effectively representing complex heterogeneous interactions at the entity level. Furthermore, we proposed atom-level, node-level, and mode-level attention modules to establish a robust machine understanding of VDAs. It offers a complete reasoning process from macro to micro levels by identifying the most contributing biological network neighbors and further analyzing their critical spatial structures, which provides reliable guidance for drug design and subsequent wet experimental validations. Given its comprehensive reasoning from the molecular to entity level, we believe SpHN-VDA represents a bridge between computational models and wet lab validation, aiding drug discovery for emerging viral threats and guiding researchers and clinicians in a more targeted approach to antiviral therapy. Ultimately, SpHN-VDA’s predictive power, coupled with its interpretability, can enhance clinical decision support systems, bringing precision medicine closer to clinical reality and addressing the urgent need for adaptable antiviral treatments in a rapidly evolving healthcare landscape.

References:

1.    Dickson M, Gagnon JP. Key factors in the rising cost of new drug discovery and development. Nature reviews Drug discovery 3, 417-429 (2004).

2.    Fernández-Torras A, Duran-Frigola M, Bertoni M, Locatelli M, Aloy P. Integrating and formatting biomedical data as pre-calculated knowledge graph embeddings in the Bioteque. Nature Communications 13, 5304 (2022).

3.    Wang R-S, Loscalzo J. Repurposing drugs for the treatment of COVID-19 and its cardiovascular manifestations. Circul Res 132, 1374-1386 (2023).

4.    Chen Z-H, Zhao B-W, Li J-Q, Guo Z-H, You Z-H. GraphCPIs: A novel graph-based computational model for potential compound-protein interactions. Molecular Therapy-Nucleic Acids 32, 721-728 (2023).

5.    Ren Z-H, et al. DeepMPF: deep learning framework for predicting drug-target interactions based on multi-modal representation with meta-path semantic analysis. J Transl Med 21, 1-18 (2023).

6.    Ruiz C, Zitnik M, Leskovec J. Identification of disease treatment mechanisms through the multiscale interactome. Nature Communications 12, 1796 (2021).

7.    Song Y, Zhou C, Wang X, Lin Z. Ordered GNN: Ordering Message Passing to Deal with Heterophily and Over-smoothing. In The Eleventh International Conference on Learning Representations (Ithaca, 2023).

8.    Frolichs KM, Rosenblau G, Korn CW. Incorporating social knowledge structures into computational models. Nature Communications 13, 6205 (2022).

9.    Verma J, Khedkar VM, Coutinho EC. 3D-QSAR in drug design-a review. Curr Top Med Chem 10, 95-115 (2010).

10.    Yang S-Y. Pharmacophore modeling and applications in drug discovery: challenges and recent advances. Drug Discov Today 15, 444-450 (2010).

11.    Zeng Z, Yao Y, Liu Z, Sun M. A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals. Nature Communications 13, 862 (2022).

12.    Sun Y, et al. A graph neural network-based interpretable framework reveals a novel DNA fragility–associated chromatin structural unit. Genome Biol 24, 90 (2023).