TL;DR
The COVID-19 pandemic generated a massive amount of research, leading to many proposed drug repurposing candidates but also a lack of consensus. This paper introduces the COVID-19 PHARMACOME, a comprehensive knowledge graph that harmonizes diverse data sources to rationally select and predict synergistic drug combinations for treating COVID-19. Experimental validation confirmed its utility in identifying effective drug pairings.
What Problem Does This Paper Address?
The rapid onset of the COVID-19 pandemic spurred an unprecedented global research effort, particularly in the area of drug repurposing. This resulted in a deluge of experimental data, hypotheses, and publications. However, a significant challenge emerged: the proposed therapeutic compounds from various SARS-CoV-2 specific assays often lacked agreement, creating a fragmented landscape of information. This fragmentation hinders the rational selection of promising drug repurposing candidates. The authors address this by aiming to create a singular, unified source of COVID-19 related information that can guide the identification of effective drug repurposing strategies, especially for combination therapies.
Methods
The core of this work is the development of the COVID-19 PHARMACOME, a comprehensive drug-target-mechanism graph. This was achieved through a multi-step process:
- Knowledge Graph Harmonization: The authors integrated data from 10 separate COVID-19 disease maps and various experimental sources. This involved converting disparate knowledge graphs and interactomes into OpenBEL (Biological Expression Language), a format designed to represent cause-and-effect relationships in biomedicine and ensure interoperability. Custom pipelines were developed to normalize molecular interactions and handle data that didn't initially conform to the OpenBEL grammar. When standard naming conventions failed, they utilized the recently generated COVID-19 ontology and other standard ontologies.
- Enrichment with Drug-Target Information: The integrated knowledge graph (termed the COVID-19 supergraph) was further enriched with drug-target information from curated databases like DrugBank, ChEMBL, and PubChem. This resulted in the final COVID-19 PHARMACOME, which includes drug targets and their ligands.
- Subgraph Annotation: The PHARMACOME was organized into subgraphs representing key aspects of COVID-19 biology: the viral life cycle, host response, and clinical pathophysiology.
- Gene Expression Data Mapping: Single-cell sequencing data from infected and non-infected human samples were mapped onto the PHARMACOME to identify areas responsive to SARS-CoV-2 infection at the gene expression level.
- Pathway Enrichment: Standard tools like Enrichr were used to identify associated pathways for subgraphs and significant targets.
- Drug Repurposing Screening and Combination Assessment: Phenotypic assays were performed to screen for repurposing drugs. Additionally, existing combinatorial therapy data was mapped to the PHARMACOME to evaluate its potential in guiding rational combination therapy approaches. The authors also experimentally tested novel drug combinations predicted by their model.
Datasets
The paper leverages a variety of data sources for building the COVID-19 PHARMACOME:
- 10 separate COVID-19 disease maps: These provided core knowledge about SARS-CoV-2 pathophysiology.
- Proprietary and public COVID-19 knowledge graphs and experimental data: Used for integration into the supergraph.
- Drug-target databases: Specifically mentioned are DrugBank, ChEMBL, and PubChem, used to annotate the supergraph with drug-target information.
- Six published drug repurposing screening experiments: Data from these were downloaded, mapped, and curated to identify "hits."
- Two single-cell sequencing datasets: One from human samples and another from cultured human bronchial epithelial cells (HBECs), used for mapping gene expression data.
- Existing combinatorial therapy data: From published work by Bobrowski et al. and Ellinger et al., mapped to the PHARMACOME.
Key Insights and Ideas
- The Power of Harmonization: The study highlights that no single disease map or knowledge graph captures the full complexity of COVID-19. Integrating multiple sources into a unified structure like the COVID-19 PHARMACOME is crucial for a comprehensive understanding.
- Network Structure Predicts Synergy: The authors propose that the "distance" (shortest path length) between drug nodes in the COVID-19 PHARMACOME can predict the effectiveness of drug combinations. Shorter distances suggest synergistic effects, while longer distances indicate antagonistic effects. This is a novel way to leverage network topology for drug combination prediction.
- Host vs. Viral Targets: The research suggests that combinations of drugs targeting different viral mechanisms are more likely to be synergistic than combinations where one or both drugs target host mechanisms.
- Nelfinavir and Raloxifene Synergy: An interesting finding was the synergistic effect observed between Nelfinavir (an HIV protease inhibitor) and Raloxifene, which aligns with the PHARMACOME's predictions. This highlights the potential of the platform to uncover unexpected synergistic pairings.
Key Results
- The COVID-19 PHARMACOME was constructed as a comprehensive drug-target-mechanism graph containing 150,662 nodes and 573,929 edges.
- Comparative analysis of drug repurposing screenings revealed no direct overlap in hits between experiments using different cell types, but over one hundred common protein targets were identified when comparing target spaces.
- The PHARMACOME successfully predicted synergistic drug combinations. For example, pairs known to have synergistic effects in SARS-CoV-2 treatment had an average shortest path length of 2.43 in the graph, while antagonistic combinations had an average shortest path length of 4.0.
- Experimental validation of predicted combinations, such as Remdesivir/Thioguanosine and Nelfinavir/Raloxifene, demonstrated synergistic effects on virus-induced cytopathic effect in Caco-2 cells.
Limitations and Future Work
- Assay Condition Differences: The authors acknowledge that the PHARMACOME integrates experimental results generated under different assay conditions. They note that assay responses from Vero E6 cells and Caco-2 cells may only partially overlap and plan to perform comparative analysis using transcriptome-wide gene expression analysis in future work.
- Model-Driven Approach: While the model-driven information fusion approach was shown to work well, the authors recognize the "putative differences between drug repurposing screening assays."
- Continuous Improvement: The authors emphasize that the COVID-19 PHARMACOME is a living resource that will be "continuously improved and expanded" to serve integrative approaches in anti-SARS-CoV-2 drug discovery and development. They encourage colleagues to contribute useful information and further develop the platform.
Summary updated