Network Medicine Framework for Identifying Drug Repurposing Opportunities for COVID-19

Experimental and Clinical Validation of Drug Repurposing Pipelines

• For validation, the study identified 918 drugs ranked by all predictive pipelines that had been experimentally screened to inhibit viral infection and replication in cultured cells17.Source: Introduction · p. 2

Introduction

• Predictive power varied across datasets and metrics, making it impossible to choose a trusted method without prior ground truth; ensemble consensus improved prediction accuracy and reliability.The multimodal approach combines predictions from the different methods.Source: Introduction · pp. 2–3

Network-based Drug Repurposing

• Of 332 SARS-CoV-2 host targets, 208 formed a connected disease module, and that module did not directly overlap with proteins associated with any single disease.The authors argue that this supports finding repurposable drugs without relying on therapies approved for one specific disease.Source: Network-based Drug Repurposing · p. 3
• The AI, diffusion, and proximity algorithms showed low correlations, indicating that they extract complementary information from the network.The study implemented four AI pipelines, five diffusion pipelines, and three proximity pipelines.Source: Network-based Drug Repurposing · pp. 3–4

Experimental and Clinical Validation of Drug Repurposing Pipelines

• For E918, P1 had AUC 0.63, followed by D2 and P3 (0.58 each); for CT415, AI pipelines had AUC 0.73–0.76, followed by P1 (0.57) and P2 (0.56).The best-performing pipeline differed by dataset: proximity P1 led for experimental outcomes, while AI pipelines performed strongly for drugs selected for clinical trials.Source: Experimental and Clinical Validation… · p. 5
• Across the twelve pipelines, 22 positive drugs appeared among the top 100, but no pipeline offered consistently superior performance across outcomes, prompting a multimodal approach.The approach was intended to extract the joint predictive power of all pipelines.Source: Experimental and Clinical Validation… · p. 5

Multimodal Approach for Drug Repurposing

• Average Rank performed worst; Union and Dowdall improved on it but trailed the best individual pipelines, while Borda performed strongly for E918 but not CT415.CRank relied on Bayesian factors; the passages report its comparisons with Figure 4G-H.Source: Multimodal Approach for Drug Repurposing · p. 6
• CRank performed consistently across datasets and most K values, matching or exceeding individual pipelines across datasets and extracting their cumulative predictive power.It performed equally well on the prospective E74 list and the clinical-trial list updated 06/15/20 (C615).Source: Multimodal Approach for Drug Repurposing · p. 6

Network Effects

• 76 of the 77 drugs effective in the experimental screen were network drugs, perturbing the host subcellular network indirectly rather than targeting viral proteins or their host targets.Only amitriptyline was known to directly target a viral-protein binding target: SIGMAR1, targeted by SARS-CoV-2 NSP6; binding-based methods could not identify the other drugs, whereas network-based methods prioritized them.Source: Network Effects · pp. 6–7
• Targets of the 37 S drugs and 40 W drugs formed statistically significant connected components (Z-Score=2.05 and Z-Score=3.42), and their negative separation (S$% = −0.69) indicated a shared network neighborhood.The S&W drugs were diverse and lacked statistically significant shared pathway or mechanistic features.Source: Network Effects · p. 7
• Compared with random expectation, N-drug targets were far from the COVID-19 module, while S- and W-drug targets were closer; S targets were closer than W targets.This pattern suggests network proximity is a positive predictor of drug efficacy.Source: Network Effects · p. 7

Discussion

• All 918 drugs prioritized and screened here are FDA approved; by contrast, only 2 of the 12,000 compounds tested were FDA approved.Source: Discussion · p. 8
• 9% hit rate among CRank’s top 100 drugs versus 0.8% for brute-force screening, and its top 800 contained 58 of 77 S&W drugs.Source: Discussion · p. 8
• Combining CRank consensus with expert review enriched the screened set for S&W drugs by 28%.The review removed drugs with known significant in vivo toxicities and lower-ranked members of the same drug class.Source: Discussion · p. 8
• Azelastine (CRank #10) and digoxin (CRank #33) had strong experimental outcomes despite not being in clinical trials, supporting consideration for trials.Source: Discussion · p. 8
• Methotrexate (CRank #32) may attenuate hyperimmune responses, while omeprazole (CRank #50) alters lysosome acidification and has been linked to interference with viral formation.The passage also notes that omeprazole, along with other benzimidazoles, binds nsp3, a protein that enhances viral immune evasion.Source: Discussion · pp. 8–9
• Other highly ranked drugs with strong outcomes included fluvastatin (CRank #199), ivermectin (#235), and sildenafil (#493).The authors suggested these candidates for trial consideration based on their general pharmacology.Source: Discussion · p. 9
• Only 918 of the 6,340 prioritized drugs were screened because of compound availability, leaving many potentially efficacious FDA-approved drugs untested.Source: Discussion · p. 9
• Some VeroE6-negative drugs may work in human cells: loratadine inhibited viral activity in Caco-2 cells, while top-ranked ritonavir showed no effect in this screen despite over 42 clinical trials exploring it.Source: Discussion · p. 9

• A.A, D.M.G, M.Z, and X.G performed drug predictions; I.D.V analyzed disease comorbidities; R.D and J.J.P performed drug experiments and experimental screens.Source: Authors Contribution · p. 10
• A.A, D.M.G, I.D.V, M.Z, O.V, X.G, and J.L analyzed the data, while J.L manually curated drug candidates.Source: Authors Contribution · p. 10
• S.D.G guided A.A in designing diffusion-based similarity implementations, and O.V curated the list of drugs in COVID-19 clinical trials.Source: Authors Contribution · p. 10
• A.L.B, D.M.G, I.D.V, and X.G wrote the paper with input from all authors.Source: Authors Contribution · p. 10
• A.L.B designed the study, and all authors read and approved the manuscript.Source: Authors Contribution · p. 10

• J.L. and A.L.B. are co-scientific founders of Scipher Medicine, which applies network medicine strategies to biomarker development and personalized drug selection.Source: Declaration of interests · p. 10
• A.L.B. founded Nomix and Foodome, which apply data science to health.Source: Declaration of interests · p. 10
• O.V. and D.M.G. consult scientifically for Nomix, and I.D.V. consults scientifically for Foodome.Source: Declaration of interests · p. 10

• The supplementary information is titled “Network Medicine Framework for Identifying Drug Repurposing Opportunities for COVID-19.”Source: Supplementary Information: Network… · p. 21
• The listed authors are Deisy Morselli Gysi, Ítalo Do Valle, Marinka Zitnik, Asher Ameli, Xiao Gan, Onur Varol, and Susan Dina Ghiassian.Source: Supplementary Information: Network… · p. 21
• The listed authors also include JJ Patten, Robert Davey, Joseph Loscalzo, and Albert-László Barabási.Source: Supplementary Information: Network… · p. 21

1.1 Human Interactome and SARS-CoV-2 and Drug Targets

• The interactome drew on 21 public databases, including binary interactions from yeast-two-hybrid experiments, protein structures, and literature curation.Source: §1.1 · p. 22
• Additional interactome data covered affinity-purification/mass-spectrometry PPIs, kinase-substrate and signaling interactions, and regulatory interactions derived by ENCODE.Source: §1.1 · p. 22
• The final human interactome contained 18,505 proteins and 327,924 interactions after proteins that could not be mapped to Entrez IDs were removed.Source: §1.1 · p. 23
• The study retrieved interactions between 26 SARS-CoV-2 proteins and 332 human proteins reported by Gordon et al. (2020)19.Source: §1.1 · p. 23
• DrugBank data comprised 24,609 interactions between 6,228 drugs and 3,903 targets; the study also used literature-curated targets for 25 drugs and drug-exposure DEGs from different cell lines.Source: §1.1 · p. 23
• The largest connected component of human proteins binding SARS-CoV-2 proteins was calculated with a degree-preserving approach using 100 degree bins in 1,000 realizations.The approach was used to prevent repeated selection of the same high-degree nodes.Source: §1.1 · p. 23

1.2 Lung Gene Expression (Fig 2A)

• Lung gene expression was evaluated with GTEx, and genes with a median raw count below 5 transcripts were treated as not expressed.Source: §1.2 · p. 23

1.3 Disease Comorbidities

• Among 332 proteins targeted by SARS-CoV-2, 110 were implicated in other human diseases; the analysis retrieved 3,173 disease-associated genes across 299 diseases.Source: §1.3 · p. 23
• No major disease module directly overlapped the COVID-19 module, arguing against deriving treatment from drugs approved for a specific disease and for mapping drug targets irrespective of disease module.All disease network-overlap scores were positive.Source: §1.3 · p. 24, Figure S1 · p. 25
• The diseases closest to COVID-19 targets included cardiovascular diseases and cancers, whose comorbidity with COVID-19 is well documented [29–31].The network analysis also indicated proximity to pulmonary diseases.Source: §1.3 · p. 24, Figure S2 · p. 26
• Network evidence linked SARS-CoV-2 targets to neurological disease, consistent with reports of olfactory and taste loss [32] and neurological manifestations in 36% of hospitalized patients with severe infection.The host protein targets were also found to be expressed in the brain.Source: §1.3 · p. 24

1.4.1 Artificial Intelligence Based Algorithm (A1-A4)

• The GNN ranks candidates by treating drug–disease indication prediction as link prediction on a multimodal graph and scoring potential drug–disease edges.The graph contains drug, protein, and disease nodes; an encoder produces node embeddings and a decoder scores indication links.Source: §1.4.1 · pp. 27–28

1.4.2 Diffusion-Based Algorithms (D1-D5)

• Five DSD-inspired metrics were developed to measure drug-target impacts on SARS-CoV-2 targets, using diffusion-state distances and divergence measures.The described variants include L1-norm distance, Kullback-Leibler divergence, and symmetric Jensen-Shannon divergence.Source: §1.4.2 · pp. 31–32

1.4.3 Proximity Algorithm (P1-P3)

• The proximity method compares viral and drug target distances with degree-matched random protein sets and implements three target-set variants.The variants use all drug targets, exclude carriers, transporters, and metabolizing enzymes, or use drug-associated differentially expressed genes.Source: §1.4.3 · p. 32

1.5.1 Explanatory Subgraphs

• For each pipeline, explanatory subgraphs were identified to help understand predictions; proximity explanations exactly represent contributing nodes, while AI-based explanations were extracted with GNNExplainer.Source: §1.5.1 · p. 33
• Proximity and diffusion methods explore the PPI more broadly than AI methods, while selected genes and their neighborhoods differ across methods.Methods within the same pipeline tend to select similar genes; across methods, gene overlap and neighborhood similarity differ. AI subgraph sizes vary less than proximity subgraph sizes.Source: Figure S4 · p. 35, §1.5.1 · p. 35
• PCA shows AI methods and proximity methods P1 and P2 are consistent, while diffusion predictions vary more and P3 lies far from other proximity measures.Diffusion methods show greater variance in gene–drug pair predictions and a wider module spread.Source: §1.5.1 · pp. 35–36

1.5.2 Complementarity of Prediction Algorithms (Fig 2C)

• AI subgraphs overlap drug targets but are separated from the COVID-19 module, proximity subgraphs usually overlap the module, and diffusion subgraphs avoid both.These patterns indicate that the algorithms draw predictions from different regions of the protein–protein interaction network.Source: §1.5.2 · p. 36, Figure S5 · p. 37

2.1 Cell Cultivation and Viruses Used

• VeroE6 cells were maintained in serum-supplemented DMEM, and the USA-WA1/2020 virus stock was passaged twice on Vero cells.Source: §2.1 · p. 38

2.2 Virus Infection Inhibition Assay

• Compounds were added to Vero cells for at least 1 h before viral challenge (MOI <0.2), in two-fold dilutions from 25 µM to 0.2 µM; infection efficiency was infected cells divided by total nuclei, with nuclei loss flagging toxicity.Source: §2.2 · p. 38

2.3 Drug-Response Classification

• Drug-response classification tested toxicity before viral inhibition; significant reduction above 80% was strong, above 50% was weak, otherwise no-effect, while toxicity at half or more doses meant cytotoxic.Significance was assessed with goodness-of-fit tests using p < 0.01 with Bonferroni correction.Source: §2.3 · p. 39

2.4 Biological Interpretation of Effective Drugs in E918 Dataset

• 77 drugs showed strong or weak effects, but target and pathway analyses found no patterns explaining their efficacy.Source: §2.4 · pp. 39–40
• Hierarchical clustering of target profiles for 77 E918 S&W drugs found no binding pattern shared by all drugs, but identified four small groups, one comprising seven nervous-system-related drugs.Source: §2.4 · p. 40, Figure S6 · p. 41
• 42 of the 77 strong- or weak-effect drugs fell into three pathway groups, including 20 linked to transport and metabolism of different substrates.Source: §2.4 · p. 40, Figure S7 · p. 42
• 18 effective drugs were associated with membrane-receptor pathways, mostly nervous-system drugs targeting GPCRs such as ADRA1A, HTR2A, and HRH1.Source: §2.4 · p. 40, Figure S7 · p. 42

3.1 Performance Evaluation using ROC Curves, Precision, and Recall

• AUC evaluates separation across the full ranked drug list, while precision@K and recall@K measure positive outcomes among the top-K predictions.Precision@K is the fraction of top-K drugs that are positive; recall@K is the fraction of all positive outcomes in the top-K.Source: §3.1 · p. 43
• Experimental screening covered 918 compounds (E918), identifying 806 no-effect, 40 weak-effect, and 37 strong-effect drugs, plus 74 additional compounds tested at 0.625–20μM and 0.2 MOI (E74).E74 comprised 39 no-effect, 10 weak-effect, and 11 strong-effect drugs.Source: §3.1 · p. 43
• Precision and recall estimates are conservative because evaluation ranked all 6,340 drugs, including untested drugs that could be positive.The passage states these estimates are lower than values from analysis limited to experimentally screened drugs.Source: §3.1 · p. 44
• An exact one-tailed binomial test with p-value < 0.05 assessed whether top-K lists contained more positive drugs than expected by chance.Expected counts accounted for the number of drugs with ground-truth information and the number receiving pipeline predictions.Source: §3.1 · p. 44
• In most cases, C615 has higher hit rates, precision, and recall than E74; Figure S8 also shows CRank’s top precision and recall against individual pipelines.Figure S8 compares twelve pipelines on E74 and CT615 and presents rank-aggregation results, describing CRank’s predictive performance as strong.Source: Figure S8 · p. 45, §3.1 · p. 45

4 Rank Aggregation Algorithms (RAAs)

• The Kemeny consensus maximizes pairwise agreements with input rankings, but computing it is NP-hard [48,49].This computational difficulty motivated heuristic and approximation methods.Source: §4 · p. 47

4.1 Average Rank Method

• Average Rank calculates each drug’s mean position across the 12 pipeline rankings, but the authors found average ranks can be a poor aggregation approach.It is described as a popular ad-hoc strategy.Source: §4.1 · p. 47

4.2 Borda Method

• Borda sums the number of drugs each candidate outranks across rankings and is a 5-approximation algorithm of the Kemeny optimal ranking [51,56].Source: §4.2 · p. 47

4.3 Dowdall Method

• Dowdall assigns each drug the reciprocal of its rank in each pipeline’s list and sorts candidates by their total score [57].Source: §4.3 · p. 48

4.4 CRank

• CRank iteratively reweights ranked lists using uncertainty across lists, allowing each pipeline’s importance to vary by drug; it converged in fewer than 20 iterations.The method updates weights using Bayes factors and re-aggregates the lists until the aggregate ranking stops changing.Source: §4.4 · p. 48

4.5 Comparison of RAAs

• For E918 at K=100, Average Rank’s Kemeny score is infinite, Borda’s is KS = 0.7131, Dowdall’s is KS = 0.4545, and CRank’s is KS = 0.2679.Across both E918 and CT415, CRank offers the best agreement with outcomes for K<250.Source: §4.5 · p. 49

Network resources

• Table S3 lists 332,749 pairwise binding interactions among 18,508 human proteins, and Table S4 covers interactions between 29 SARS-CoV-2 proteins and 332 human proteins detected by affinity purification and mass spectrometry.The SARS-CoV-2–human interaction dataset was retrieved from Gordon et al (2020).Source: §5 · p. 50, Table S3, Table S4

Drug data

• Tables S5 and S6 provide DrugBank drug-target lists and 17,222 differentially expressed genes identified after exposure to 793 drugs in different cell lines.The passage notes that the data were obtained from DrugBank.Source: §5 · p. 50, Table S5, Table S6

Network overlap and embeddings

• Table S7 reports network overlap between 299 diseases and SARS-CoV-2 targets; Tables S8 and S9 provide GNN-learned embedding vectors for diseases and drugs.The S7 measure captures network-based overlap between SARS-CoV-2 targets and each disease's associated gene pool.Source: §5 · pp. 50–51, Table S7, Table S8

Experimental datasets

• Tables S10 and S11 list experimental outcomes for 918 drugs screened in VeroE6 cells and 74 compounds selected from the top 10% of purchasable drug predictions.The 918-drug dataset records experimental outcomes for inhibition of SARS-CoV-2; the 74-compound dataset was selected by a medical doctor.Source: §5 · p. 51, Table S10, Table S11

Clinical-trial datasets

• Tables S12 and S13 list drugs under evaluation in COVID-19 clinical trials as of April 2020 (C415) and June 2020 (C615), respectively.Source: Table S12, Table S13

Drug rankings

• Table S14 provides drug rankings from the 12 pipelines and their aggregation.Source: Table S14