Network-based approach to prediction and population-based validation of in silico drug repurposing

Abstract

Here we identify hundreds of new drug-disease associations for over 900 FDA-approved drugs by quantifying the network proximity of disease genes and drug targets in the human (protein-protein) interactome. We select four network-predicted associations to test their causal relationship using large healthcare databases with over 220 million patients and state-of-the-art pharmacoepidemiologic analyses. Using propensity score matching, two of four network-based predictions are validated in patient-level data: carbamazepine is associated with an increased risk of coronary artery disease (CAD) [hazard ratio (HR) 1.56, 95% confidence interval (CI) 1.12-2.18], and hydroxychloroquine is associated with a decreased risk of CAD (HR 0.76, 95% CI 0.59-0.97). In vitro experiments show that hydroxychloroquine attenuates pro-inflammatory cytokine-mediated activation in human aortic endothelial cells, supporting mechanistically its potential beneficial effect in CAD. In summary, we demonstrate that a unique integration of protein-protein interaction network proximity and large-scale patient-level longitudinal data complemented by mechanistic in vitro studies can facilitate drug repurposing.

Fig. 1

The predicted drug-disease network. The high-confidence predicted drug-disease association network connects 22 types of cardiovascular disease (outcomes) (red circles) and 431 FDA-approved non-cardiac drugs. The edges between drugs and diseases are weighted and highlighted by different color representing the calculated z-score (Supplementary Data 2 and Methods section). Four selected drug-disease pairs, including carbamazepine-coronary artery disease (CAD) with z = −2.36, hydroxychloroquine-CAD (z = −3.85), mesalamine-CAD (z = −6.10), and lithium-stroke (z = −5.97), tested in patient data (Figs. 2 and 3), are highlighted. Drugs are colored by the first-level anatomical therapeutic chemical (ATC) classification system codes. The node size scales indicate the degree (connectivity) of nodes in the network

Fig. 2

Flow-chart of the pharmacoepidemiologic investigations using Truven MarketScan and Optum Clinformatics patient databases. AZA azathioprine, 6-MP 6-mercaptopurine, CBZ carbamazepine, HCQ hydroxychloroquine, IBD inflammatory bowel disease, LAM lamotrigine, LEF leflunomide, LEV levetiracetam, LIT lithium, MES mesalamine, RA rheumatoid arthritis. The balance achieved in patient characteristics and outcome risk factors between the two treatment groups compared through 1:1 PS-matching are provided in Supplementary Tables 2–5

Fig. 3

Hazard ratios and 95% confidence intervals for four cohort studies. Four cohort studies were performed using the pooled data from four drug pairs (a-d) Truven MarketScan and Optum Clinformatics databases (Methods section). In the primary analysis approach (as-treated), follow-up was stopped upon discontinuation of the index medication. Follow-up assumptions were varied in three sensitivity analyses to: (1) exclude the first 60 days of follow-up to reduce unmeasured baseline confounding, (2) truncate the follow-up to 1-year to minimize time-varying confounding, and (3) continue the follow-up for 1-year regardless of treatment discontinuation under an intent-to-treat (ITT) principle. Propensity score (PS) matching accounted for >50 relevant patient characteristics; all analyses were conducted separately in two databases and results were pooled using the DerSimonian and Laird random effects model with inverse variance weights. * In the as-treated approach, the follow-up was stopped if patients either filled a prescription for a drug in the other exposure group or discontinued the index exposure. ** In the ITT analysis, patients were followed in their index exposure group regardless of treatment change or discontinuation for up to 365 days

Fig. 4

Experimental validation of hydroxychloroquine’s likely mechanism-of-action in coronary artery disease (CAD). a A highlighted subnetwork shows the inferred mechanism-of-action for hydroxychloroquine’s protective effect in CAD by network analysis. A network analysis was designed to meet four criteria: (1) the shortest paths from the known drug targets (TLR7 and TLR9) in the human protein–protein interaction network; (2) the blood vessel-specific gene expression level based on RNA-seq data from Genotype-Tissue Expression database; (3) known CAD or cardiovascular disease (CVD) gene products (proteins); and (4) literature-reported in vitro and in vivo evidence. There are three proposed mechanisms: (i) ERK5 (encoded by MAPK7) activation prevents endothelial inflammation via inhibition of cell adhesion molecule expression (VCAM-1 and ICAM-1), (ii) suppression of pro-inflammatory cytokines (TNF-α and IL-1β), and (iii) improvement in endothelial dysfunction via enhanced nitric oxide production by endothelial nitric oxide synthase (NOS3). The node size scales show the blood vessel-specific expression level based on RNA-seq data from Genotype-Tissue Expression database (Methods section). b, d Endothelial cells were pretreated with various concentrations of hydroxychloroquine (HCQ, 10–50 µM) for 1 h prior to 24 h incubation with 5 ng/ml TNF-α. qRT-PCR was used to monitor gene expression of inflammatory genes (b) VCAM1 and IL1B; and (d) NOS3. Expression of the β-actin gene was used as an internal standard. VCAM1: no HCQ, no TNF, n = 8; TNF; n = 8; TNF+ 10 μM HCQ, n = 5; TNF+ 20 μM HCQ, n = 4; TNF+ 30 μM HCQ, n = 3; TNF+ 50 μM HCQ, n = 6. IL-1β and NOS3: no HCQ, no TNF, n = 9; TNF; n = 9; TNF+ 10 μM HCQ, n = 5; TNF+ 20 μM HCQ, n = 5; TNF+ 30 μM HCQ, n = 4; TNF+ 50 μM HCQ, n = 6. Error bars are standard deviations. *significantly different from TNF-α with no HCQ, p < 0.05 as determined by post-hoc testing using the Student's Newman–Keuls test. c Western blot of VCAM-1, IL-1β, and actin. Endothelial cells were pretreated with 10 µM hydroxychloroquine for 1 h prior to 24 h incubation with 5, 10 or 20 ng/ml TNF-α. Each condition was tested six times; shown are representative blots