Abstract
The COVID-19 pandemic highlighted the urgent need for rapid therapeutic discovery. Traditional drug development timelines are incompatible with global health crises, making drug repurposing an essential strategy. This study presents a computational workflow combining Graph Convolutional Networks (GCNs) and deep-learning-based off-target drug predictions to identify existing medications with antiviral potential. By analyzing the multiscale host-virus interactome, we identified the MET inhibitor capmatinib as a potent, broad-spectrum antiviral candidate. Experimental validation confirmed that capmatinib’s efficacy stems from its interaction with the IRAK1/4 signaling pathway rather than its primary clinical target, demonstrating the power of in silico polypharmacology in identifying novel therapeutic mechanisms.
Introduction
Developing a new drug from scratch typically requires over a decade of research and clinical trials. During the SARS-CoV-2 pandemic, the scientific community turned to drug repurposing—finding new uses for existing, approved medications—to bypass early-stage safety testing. However, traditional repurposing is often limited to a drug's known primary mechanism or discovered through serendipity.
The human interactome, the vast network of interactions between viral proteins, human proteins, and biological pathways, contains hidden vulnerabilities that can be exploited by "off-target" drug effects (polypharmacology). This paper addresses the need for a systematic, computational approach to predict which small molecules can disrupt the host-virus relationship, even through non-obvious biological pathways.
Methodology
The research team employed a hybrid computational-experimental pipeline to screen and validate candidates:
1. Multiscale Interactome GCN
The researchers constructed a massive biological network integrating 1,661 drugs, 17,660 human proteins, 9,798 functional pathways, and 26 SARS-CoV-2 proteins. Using node2vec and Graph Convolutional Networks (GCNs), the system learned mathematical "embeddings" for each node. This allowed the model to rank drugs based on their proximity to the COVID-19 disease node within the multiscale network.
2. PolypharmDB and MatchMaker
To address the lack of known data for many drug-protein pairs, the team utilized MatchMaker, a deep-learning engine that predicts drug-target interactions (DTI) based on 3D structural features. This tool was critical for identifying "off-target" binders for the proteins highlighted by the GCN, effectively mapping the polypharmacology of the existing drug library.
3. Experimental Validation
From the computational predictions, 26 candidates were shortlisted for testing. The primary screen used a BSL2-compatible system (the 229E alphacoronavirus). Top-performing hits were further validated using plaque assays for other coronaviruses (OC43, NL63) and pseudovirus entry assays for SARS-CoV-1 and SARS-CoV-2.
4. Mechanistic Probing
To confirm the underlying biological mechanism, the authors used siRNA silencing and specific chemical inhibitors to determine if blocking the predicted targets (such as IRAK1/4) mimicked the antiviral effects of the drugs.
Key Insights and Results
The Discovery of Capmatinib
The most significant finding was the identification of capmatinib, an FDA-approved MET inhibitor used for lung cancer, as a potent antiviral. Crucially, the study revealed that capmatinib does not work through its primary target (MET). A different, highly specific MET inhibitor (AMG-337) showed no antiviral activity, proving that capmatinib’s efficacy is due to its off-target binding.
IRAK1/4 as a Viral Vulnerability
The computational model predicted that capmatinib binds to IRAK4. Experimental results confirmed that the IRAK1/4 signaling pathway is a critical host factor hijacked by coronaviruses. siRNA silencing of IRAK1 and IRAK4 significantly reduced viral infection, validating the model's prediction that this pathway is a viable therapeutic target.
Performance Metrics
- Hit Rate: 4 out of 26 predicted drugs (~15%) reduced viral abundance by at least 50% without significant cellular toxicity.
- Antiviral Potency: Capmatinib showed nearly 100% inhibition of 229E infection in initial screens. In plaque assays, it reduced viral production of NL63 by >70% and OC43 by ~50%.
- SARS-CoV-2 Entry: In pseudovirus assays, capmatinib inhibited the entry of both SARS-CoV-1 and SARS-CoV-2 with $IC_{50}$ values of 14.1 µM and 26.0 µM, respectively.
Discussion and Limitations
This "in silico first" approach demonstrates that host-centric drug targets can provide broad-spectrum protection against multiple coronavirus strains. Because these drugs target host proteins that the virus requires for replication—rather than the rapidly mutating viral proteins themselves—they may be more resilient against emerging variants.
However, limitations remain. The initial screens used "common cold" coronaviruses (229E) as proxies for SARS-CoV-2; while many mechanisms are conserved, some SARS-CoV-2 specific vulnerabilities might have been overlooked. Additionally, the $IC_{50}$ values for SARS-CoV-2 entry are in the micromolar range, and further research is required to determine if these concentrations can be safely achieved in human lung tissue. Finally, while functional evidence for the IRAK pathway is strong, direct structural biology (e.g., X-ray crystallography) of the drug-target complex would further solidify the MatchMaker predictions.
Conclusion
By coupling multiscale interactome analysis with deep-learning DTI predictions, this study successfully identified novel drug-target pairs that were previously hidden. The identification of capmatinib and the IRAK1/4 pathway provides a new roadmap for developing broad-spectrum antivirals and demonstrates the efficiency of AI-driven drug repurposing in responding to emerging viral threats.
Summary updated