Abstract
Motivated by the critical need to identify new treatments for COVID-19, we present a genome-scale, systems-level computational approach to prioritize drug targets based on their potential to regulate host-virus interactions or their downstream signaling targets. We adapt and specialize network label propagation methods to this end. We demonstrate that these techniques can predict human-SARS-CoV-2 protein interactors with high accuracy. The top-ranked proteins that we identify are enriched in host biological processes that are potentially coopted by the virus. We present cases where our methodology generates promising insights such as the potential role of HSPA5 in viral entry. We highlight the connection between endoplasmic reticulum stress, HSPA5, and anti-clotting agents. We identify tubulin proteins involved in ciliary assembly that are targeted by anti-mitotic drugs. Drugs that we discuss are already undergoing clinical trials to test their efficacy against COVID-19. Our prioritized list of human proteins and drug targets is available as a general resource for biological and clinical researchers who are repositioning existing and approved drugs or developing novel therapeutics as anti-COVID-19 agents.
Introduction: Beyond the Black Box
Network propagation is a cornerstone of systems biology, functioning much like a social network where information about a disease flows from known "source" proteins through interaction networks to label unknown candidates. However, these algorithms often operate as "black boxes," providing lists of predictions without explaining the underlying experimental evidence. In the context of a global pandemic, researchers require transparency to justify the high costs of laboratory validation. This study addresses this gap by introducing an "interpretable" network propagation framework applied to the SARS-CoV-2 interactome.
Methodology: Interpretable Network Propagation
The researchers utilized the Regularized Laplacian (RL) algorithm to model the spread of biological signals across protein-protein interaction (PPI) networks. The core innovation of this approach is twofold:
- Provenance Tracing: By leveraging the linearity of the RL algorithm, the authors mathematically decomposed final scores into contributions from individual source proteins. This allows every prediction to be traced back to specific experimental origins.
- Automated Parameter Selection: To avoid manual trial-and-error, the authors derived a formula based on the expected path length of a random walk to set optimal diffusion parameters based on the network's specific topology.
The study integrated the STRING database (v11) for PPI networks with 332 human proteins known to physically interact with SARS-CoV-2 (Gordon et al., 2020) serving as the initial "seeds" for the algorithm.
Biological Insights: ER Stress and Ciliary Entry
The analysis yielded several high-priority hypotheses regarding the viral life cycle:
- The HSPA5-Clotting Connection: The model identified HSPA5 (GRP78), a protein involved in endoplasmic reticulum (ER) stress. Provenance tracing suggests a mechanism where viral-induced ER stress upregulates HSPA5, which then chaperones pro-coagulant proteins, potentially explaining the severe blood-clotting complications seen in COVID-19 patients.
- Cilia as a Viral Entryway: The results strongly implicated proteins involved in cilium assembly and tubulin modulation. Given that the ACE2 receptor is highly expressed on respiratory cilia, the virus may be utilizing these cellular structures as a primary entry route.
Validation and Key Results
The framework demonstrated exceptional predictive power when compared against independent experimental datasets. The top 332 and top 1,000 predictions showed massive, statistically significant overlaps with newer interactome and proximity-labeling (BioID) data.
Notably, approximately one-third of the top 1,000 RL predictions were confirmed by independent BioID datasets ($p$-value $\approx 3.9 \times 10^{-74}$). Furthermore, the study identified several actionable drug targets, including Aspirin, Colchicine, and Tenecteplase, which are already undergoing clinical trials for COVID-19 efficacy.
Conclusion
While the accuracy of these predictions is inherently tied to the quality of the underlying PPI networks, the introduction of provenance tracing makes network-based predictions more actionable for experimentalists. This framework is designed to be virus-agnostic, providing a tool that can be rapidly deployed for future emerging pathogens. The prioritized list of proteins and drug targets remains a public resource for the ongoing development of anti-COVID-19 therapeutics.
Summary updated