Identification of disease treatment mechanisms through the multiscale interactome

Abstract

Most diseases disrupt multiple proteins, and drugs treat such diseases by restoring the functions of the disrupted proteins. How drugs restore these functions, however, is often unknown as a drug's therapeutic effects are not limited to the proteins that the drug directly targets. Here, we develop the multiscale interactome, a powerful approach to explain disease treatment. We integrate disease-perturbed proteins, drug targets, and biological functions into a multiscale interactome network. We then develop a random walk-based method that captures how drug effects propagate through a hierarchy of biological functions and physical protein-protein interactions. On three key pharmacological tasks, the multiscale interactome predicts drug-disease treatment, identifies proteins and biological functions related to treatment, and predicts genes that alter a treatment's efficacy and adverse reactions. Our results indicate that physical interactions between proteins alone cannot explain treatment since many drugs treat diseases by affecting the biological functions disrupted by the disease rather than directly targeting disease proteins or their regulators. We provide a general framework for explaining treatment, even when drugs seem unrelated to the diseases they are recommended for.

Fig. 1. The multiscale interactome models drug treatment through both proteins and biological functions.

a Existing systematic network approaches assume that drugs treat diseases by targeting proteins that are proximal to disease proteins in a network of physical interactions^–. However, drugs can also treat diseases by targeting distant proteins that affect the same biological functions (Supplementary Fig. 3)^–. b The multiscale interactome models drug-disease treatment by integrating both proteins and a hierarchy of biological functions (Supplementary Fig. 1). c The diffusion profile of a drug or disease captures its effect on every protein and biological function. The diffusion profile propagates the effect of the drug or disease via biased random walks which adaptively explore proteins and biological functions based on optimized edge weights. Ultimately, the visitation frequency of a node corresponds to the drug or disease’s propagated effect on that node (see the “Methods” section). d By comparing the diffusion profiles of a drug and disease, we compare their effects on both proteins and biological functions. Thereby, we predict whether the drug treats the disease (Fig. 2a–c), identify proteins and biological functions related to treatment (Fig. 2d–h), and identify which genes alter drug efficacy or cause dangerous adverse reactions (Fig. 3). For example, Hyperlipoproteinemia Type III’s diffusion profile reveals how defects in APOE affect cholesterol homeostasis, a hallmark of the excess blood cholesterol found in patients^–. The diffusion profile of Rovustatin, a treatment for Hyperlipoproteinemia Type III, reveals how binding of HMG-CoA reductase (HMGCR) reduces the production of excess cholesterol^,. By comparing these diffusion profiles, we thus predict that Rosuvastatin treats Hyperlipoproteinemia Type III, identify the HMGCR and APOE-driven cholesterol metabolic functions relevant to treatment, and predict that mutations in APOE and HMGCR may interfere with treatment and thus alter drug efficacy or cause dangerous adverse reactions.

Fig. 2. The multiscale interactome accurately predicts what drugs treat a disease and systematically identifies proteins and biological functions related to treatment.

a To predict whether a drug treats a disease, we compare the drug and disease diffusion profiles according to a correlation distance. b By incorporating both proteins and biological functions, the multiscale interactome improves predictions of what drug will treat a given disease by up to 40% over molecular-scale interactome approaches. Reported values are averaged across five-fold cross validation (see the “Methods” section); multiscale interactome values are in bold. c The multiscale interactome outperforms the molecular-scale interactome most greatly on drug classes known to harness biological functions that describe processes across the body (i.e., pituitary, hypothalamic hormones and analogs). d Existing interactome approaches are black boxes: they predict what drug treats a disease but do not explain how the drug treats the disease through specific biological functions^–. e By contrast, the drug and disease diffusion profiles (r^(c) and r^(d)) reveal the proteins and biological functions relevant to treatment. For each drug and disease pair, we induce a subgraph on the k most frequently visited nodes in the drug and disease diffusion profiles to explain treatment. f Drugs with more similar diffusion profiles have more similar gene expression signatures (Spearman ρ = 0.392, p = 5.8 × 10⁻⁷, n = 152, two-sided), suggesting that drug diffusion profiles capture their biological effects. g The multiscale interactome explains treatments that molecular-scale interactome approaches cannot faithfully represent. Rosuvastatin treats Hyperlipoproteinemia Type III by binding to HMG CoA reductase (HMGCR) which drives a series of cholesterol biosynthetic functions affected by Hyperlipoproteinemia Type III^–. h Anakinra treats cryopyrin-associated periodic syndromes (CAPS) by binding to IL1R1 which regulates immune-mediated inflammation through the Interleukin-1 beta signaling pathway^,. Inflammation is a hallmark of CAPS. Abbreviations: reg. regulation, path. pathway, proc. process, cell. cellular, + positive, − negative. Boxplots: median (line); 95% CI (notches); 1st, 3rd quartiles (boxes); data within 1.5 × the inter-quartile range from the 1st, 3rd quartiles (whiskers). Sample sizes in parentheses.

Fig. 3. Diffusion profiles identify which genes alter drug efficacy and cause serious adverse reactions and identify biological functions that help explain the alteration in treatment.

a Genes alter drug efficacy and cause serious adverse reactions in a range of treatments. A pressing need exists to systematically identify genes that alter drug efficacy and cause serious adverse reactions for a given treatment and explain how these genes interfere with treatment. b Genetic variants alter drug efficacy and cause serious adverse reactions by targeting genes of high network importance in treatment (median network importance of treatment altering genes = 0.912 vs. 0.513; p = 2.95 × 10⁻¹⁰⁷, Mood’s median test, two-sided; n = 1,223 vs. 1,223). We define the network treatment importance of a gene according to its visitation frequency in the drug and disease diffusion profiles (see the “Methods” section). c The treatment importance of a gene in the drug and disease diffusion profiles predicts whether that gene alters drug efficacy and causes serious adverse reactions for that particular treatment (AUROC = 0.79, average precision = 0.82). d Genes uniquely alter efficacy in one indicated drug but not another by primarily targeting the genes and biological functions used in treatment by the affected drug. In patients with Hypertensive Disease, a mutation in AGT alters the efficacy of Benazepril but not Diltiazem. Indeed, AGT exhibits a higher network importance in Benazepril treatment than in Diltiazem treatment, ranked as the 45th most important gene rather than the 418th most important gene. e Overall, 71.0% of genes known to alter efficacy in one indicated drug but not another exhibit higher network importance in treatment by the affected drug. f Diffusion profiles can identify biological functions that may help explain alterations in treatment. Shown are the proteins and biological functions identified as relevant to the treatment of Hypertensive Disease by Benazepril and Diltiazem. AGT, which uniquely alters the efficacy of Benazepril, is a key regulator of the renin–angiotensin system, a biological function harnessed by Benazepril in treatment but not by Diltiazem^–. Abbreviations: reg. regulation, proc. process, + positive, − negative. Boxplots: median (line); 95% CI (notches); 1st, 3rd quartiles (boxes); data within 1.5 × the inter-quartile range from the 1st, 3rd quartiles (whiskers).