TL;DR
This paper introduces a novel network medicine framework to accelerate drug repurposing for COVID-19. By combining multiple AI-driven prediction algorithms into a consensus approach, they significantly improve the accuracy of identifying effective drugs compared to individual methods. This multimodal approach successfully identified promising drug candidates, many of which act through indirect "network effects" rather than direct viral protein binding, highlighting the limitations of traditional drug discovery methods.
What Problem Does This Paper Address?
The COVID-19 pandemic underscored the urgent need for rapid drug development. Traditional de novo drug discovery is too slow, making drug repurposing a critical strategy. However, existing drug repurposing methods often lack reliability, leading to a "winner-takes-all" approach where resources are concentrated on a few candidates (like hydroxychloroquine and chloroquine), potentially missing other effective treatments. The authors aim to develop a more robust and accurate method for prioritizing existing drugs for COVID-19 treatment by leveraging network medicine principles.
Methods
The core of this work is a multimodal approach that fuses predictions from three distinct network-based drug repurposing algorithms:
- Artificial Intelligence (AI)-based algorithms (A1-A4): These use graph neural networks (GNNs) to learn embeddings of drugs, proteins, and diseases from a multimodal graph. They then predict potential drug-disease links.
- Network Diffusion algorithms (D1-D5): These leverage graph diffusion properties to measure similarity between drug targets and SARS-CoV-2 targets, considering how perturbations propagate through the network.
- Network Proximity algorithms (P1-P3): These rank drugs based on the shortest path distance between their targets and the SARS-CoV-2 host protein targets within the human interactome.
Crucially, instead of relying on a single best-performing algorithm, the authors developed a rank aggregation strategy, specifically focusing on the CRank algorithm. CRank iteratively refines importance weights for each pipeline's ranking, aiming to find a consensus ranking that best agrees with all individual predictions. This multimodal approach is designed to capture complementary information from different methods.
Datasets
- Human Interactome: A large collection of protein-protein interactions (PPIs) from 21 public databases, comprising 18,505 proteins and 327,924 interactions.
- SARS-CoV-2 Host Protein Targets: A list of 332 human proteins identified as targets of SARS-CoV-2 proteins.
- DrugBank Database: Used for drug target information and differentially expressed genes (DEGs) associated with drug exposure.
- Experimental Screening Data (E918 dataset): 918 drugs experimentally screened for their efficacy against SARS-CoV-2 in VeroE6 cells. This dataset was categorized into Strong (S), Weak (W), Cytotoxic (C), and No-Effect (N) outcomes.
- Experimental Screening Data (E74 dataset): An additional 74 compounds, selected by a medical doctor from top predictions, screened with a wider dose range.
- Clinical Trial Data (CT415 dataset): 67 drugs in ongoing clinical trials for COVID-19 as of April 2020, sourced from ClinicalTrials.gov.
- Clinical Trial Data (C615 dataset): Clinical trial data updated as of June 15, 2020.
Key Insights and Ideas
- No single algorithm is best: The study found that no single predictive algorithm consistently outperformed others across all validation datasets and metrics. This highlights the need for a more robust, ensemble approach.
- Consensus is key: A multimodal approach that fuses predictions from multiple algorithms, particularly using CRank for rank aggregation, significantly improves predictive accuracy and reliability.
- "Network Drugs" are important: A striking finding is that most effective drugs (76 out of 77) did not directly bind to SARS-CoV-2 viral proteins or their direct host targets. Instead, they act through indirect "network-based mechanisms" by perturbing the host's cellular network. This suggests that traditional docking-based strategies may miss many potential drug candidates.
- Prioritization matters: The CRank algorithm, combined with expert knowledge, dramatically increases the hit rate of effective drugs compared to brute-force screening, making experimental validation more efficient.
- COVID-19 disease module is distinct: The SARS-CoV-2 targets form a connected component in the human interactome that doesn't directly overlap with existing disease modules, arguing against repurposing based solely on disease similarity.
Key Results
- The multimodal approach, particularly using CRank, consistently outperformed individual pipelines across experimental and clinical trial datasets.
- The CRank algorithm achieved a 9% hit rate among the top 100 predicted drugs, a significant improvement over the 0.8% hit rate of unguided screening.
- Among the top-ranked drugs, they identified 77 drugs with positive experimental outcomes (S&W).
- Of these 77 drugs, 76 were identified as "network drugs," acting indirectly on the host network.
- The study proposed several promising drug candidates for COVID-19, including azelastine and digoxin, which showed strong outcomes and were not yet in clinical trials.
Limitations and Future Work
- Ground Truth Limitations: The authors acknowledge that ground-truth information (experimental outcomes and clinical trial status) is not available for all drugs, making direct comparison of all pipelines challenging, especially for pipeline P3.
- Conservative Estimates: Precision and recall values are presented as conservative estimates because the analysis considered all 6,340 drugs, not just the 918 experimentally screened ones. This means actual performance might be higher.
- Future Work: The paper suggests that this methodological pathway can be applied to identify repurposable drugs for future pathogens and neglected diseases, addressing the high costs and long timelines of de novo drug development. The authors also note that many potentially efficacious FDA-approved drugs remain to be tested, and some drugs that showed no effect in VeroE6 cells might still be effective in human cells.
Summary updated