SPOKE-related paper summaries

Repurposing Drugs for COVID-19


Network Medicine Framework for Identifying Drug Repurposing Opportunities for COVID-19. Gysi DM, Do Valle I, Zitnik M, ... Loscalzo J, Barabási A. ArXiv. Preprint. 2020 Apr 15: arXiv:2004.07229v1. PMCID: PMC7280907

...network analysis of the human proteins identified in:

A SARS-CoV-2 protein interaction map reveals targets for drug repurposing. Gordon DE, Jang GM, ... Shoichet BK, Krogan NJ. Nature. 2020 Jul 16;583(7816):459–468. PMID: 32353859
[back to paper list]

COVID-19 “Disease Module”

  • the human interactome (18,505 proteins with 327,924 interactions) was assembled from 21 public databases
    (...compare to 192,375 human proteins with 1,648,233 interactions in SPOKE, albeit fewer with interaction score at least 0.7, our default filter)
  • Gordon et al. identified 332 human proteins as interacting with SARS-CoV2 proteins
  • 208 of the 332 form a large connected component (LCC) within the human interactome
  • this LCC is treated as the disease module in subsequent analyses

Tissue Specificity?

Disease module enriched relative to all proteins (GTEx RNA-seq):


...how would ranking by ACE2 alone compare?

Comorbidities? (Risk Factors?)


 
  • of 3173 genes associated with 299 diseases, 110 are for proteins identified as SARS-CoV2 targets, but this overlap was not statistically significant
     
  • network-based overlap metric Svb calculated per disease gave values > 0, indicating lack of direct overlap
     
  • however, the closest diseases (those with smallest Svb) include the known comorbidities cardiovascular disease and cancer
     
  • other predicted comorbidities include neurologic, immune, endocrine

Identifying Drug Repurposing Candidates

  • Network: human interactome and (from DrugBank) 26,167 interactions between 7,591 drugs and 4,187 targets
  • Three general approaches: proximity, diffusion, artificial intelligence
  • network proximity:
    • P1: Z-score of drug targets vs. SARS-CoV2 targets
    • P2: same except discarding drug targets that are enzymes or transporters
    • P3: differentially expressed genes (also from DrugBank) rather than direct targets
  • diffusion state distance, a measure of similarity of drug targets vs. SARS-CoV2 targets
    • D1: L1 norm, sum over absolute differences between vector elements
    • D2: average minimum Kullback-Leibler divergence
    • D3: as in D2, except average median
    • D4: average minimum Jensen-Shannon divergence (symmetrized and smoothed compared to Kullback-Leibler)
    • D5: as in D4, except average median
  • drug and disease embedding by artificial intelligence
    • A1-A4
  • the 12 strategies give 12 lists of drugs
  • rank aggregation → a single list
  • manual removal of drugs with significant toxicities, “those not appropriate” (?), and some lower-ranked members of the same drug class

Several of these drugs make sense in that they are already being investigated. Some seem unlikely (e.g., doxorubicin causes heart damage).

From NYT article July 16 2020:

  • strong evidence: remdesivir, dexamethasone (in the list at left)
  • mixed: favipiravir, EIDD-2801 (experimental antiviral), recombinant ACE2, interferons, cytokine inhibitors, antibodies, anticoagulants, ...
  • not promising: lopinavir/ritonavir, hydroxychloroquine, chloroquine