Human disease-drug network based on genomic expression profiles

Abstract

Background: Drug repositioning offers the possibility of faster development times and reduced risks in drug discovery. With the rapid development of high-throughput technologies and ever-increasing accumulation of whole genome-level datasets, an increasing number of diseases and drugs can be comprehensively characterized by the changes they induce in gene expression, protein, metabolites and phenotypes.

Methodology/principal findings: We performed a systematic, large-scale analysis of genomic expression profiles of human diseases and drugs to create a disease-drug network. A network of 170,027 significant interactions was extracted from the approximately 24.5 million comparisons between approximately 7,000 publicly available transcriptomic profiles. The network includes 645 disease-disease, 5,008 disease-drug, and 164,374 drug-drug relationships. At least 60% of the disease-disease pairs were in the same disease area as determined by the Medical Subject Headings (MeSH) disease classification tree. The remaining can drive a molecular level nosology by discovering relationships between seemingly unrelated diseases, such as a connection between bipolar disorder and hereditary spastic paraplegia, and a connection between actinic keratosis and cancer. Among the 5,008 disease-drug links, connections with negative scores suggest new indications for existing drugs, such as the use of some antimalaria drugs for Crohn's disease, and a variety of existing drugs for Huntington's disease; while the positive scoring connections can aid in drug side effect identification, such as tamoxifen's undesired carcinogenic property. From the approximately 37K drug-drug relationships, we discover relationships that aid in target and pathway deconvolution, such as 1) KCNMA1 as a potential molecular target of lobeline, and 2) both apoptotic DNA fragmentation and G2/M DNA damage checkpoint regulation as potential pathway targets of daunorubicin.

Conclusions/significance: We have automatically generated thousands of disease and drug expression profiles using GEO datasets, and constructed a large scale disease-drug network for effective and efficient drug repositioning as well as drug target/pathway identification.

Figure 1. Human disease network.

In the disease network, each node corresponds to a disease colored according to their MeSH disease category as denoted by the MeSH Tree Number as shown on the right panel. The size of each node is proportional to the number of diseases connecting to it. A solid line links two diseases from same MeSH disease category while the dot line links two diseases from different MeSH disease category. Multiple nodes may represent the same disease, but they are from different studies or conditions, e.g. there are two bipolar disorder nodes, whose profiles are derived from studies using two different tissues (dorsolateral prefrontal cortex tissue and orbitofrontal cortex tissue). Disease abbreviations used in this figure include: C02, virus disease; C03, parasitic diseases; C04, neoplasms; C05, musculoskeletal diseases; C06, digestive system diseases; C07, stomatognathic diseases; C08, respiratory tract diseases; C10, nervous system diseases; C12, male urogenital diseases; C13, female urogenital diseases and pregnancy complications; C14, cardiovascular diseases; C15, hemic and lymphatic diseases; C16, congenital hereditary neonatal diseases and abnormalities; C17, skin and connective tissue diseases; C18, nutritional and metabolic diseases; C19, endocrine system diseases; C20, immune system diseases; C23, pathological conditions signs and symptoms; F03, mental disorders; AC, adenocarcinoma; AD, Alzheimer disease; AF, atrial fibrillation; ALS, amyotrophic lateral sclerosis; Anemia_R, refractory anemia; Anemia_RE, refractory anemia with excess of blasts; Anemia_S, sideroblastic anemia; BD, bipolar disorder; BE, Barrett esophagus; Carcinoma_B, basal cell carcinoma; Carcinoma_D, ductal carcinoma; Carcinoma_DB, breast ductal carcinoma; Carcinoma_L, lobular carcinoma; Carcinoma_PD, pancreatic ductal carcinoma; Carcinoma_SC, squamous cell carcinoma; CCS, clear cell sarcoma; CF cystic fibrosis; CIN, cervical intraepithelial neoplasia; CLL, chromic lymphocytic leukemia; CN, colorectal neoplasms; COPD, chronic obstructive pulmonary disease; DC, dilated cardiomyopathy; GF, gingival fibromatosis; HSC, hemoglobin sickle cell disease; HSP, hereditary spastic paraplegia; HT, hemorrhagic thrombocythemia; LM, lymphatic metastasis; M_myeloma, multiple myoloma; MD, muscular dystrophy; MD_D, Duchenne muscular dystrophy; MD_ED, Emery-Dreifuss muscular dystrophy; MD_F, Facioscapulohumeral muscular dystrophy; MDs, muscular diseases; MS, myelodysplastic syndromes; NM, neoplasm metastasis; Obesity_M, morbid obesity; PH, prostatic hyperplasia; POS polycystic ovary syndrome; RC, restrictive cardiomyopathy; TN, thyroid neoplasms; UC, ulcerative colitis; WM, waldenstrom macroglobulinemia; WT, Wilms tumor.

Figure 2. Disease-drug network.

This disease-drug network contains a total of 49 diseases in dark cyan nodes, 213 drugs in gold, and 906 connections. The size of the nodes is proportional to the number of links. Positive matches are shown by solid lines and negative relationships by dotted lines. Multiple nodes with the same descriptive name exist because the corresponding profiles were generated under different conditions or studies (refer to Supplementary Table S7 online for details). In addition to the abbreviations listed in the Figure 1 legend, other abbreviations used in this figure include: AK, actinic keratosis; BC, breast cancer; BL_cancer, basal_like cancer; BRCA1_cancer, BRAC1-associated cancer; HPV, human papillomavirus; IM, idiopathic myelofibrosis; LNM, lymph node metastasis; MCF, mild cystic fibrosis; NBL_cancer, non-basal-like cancer; NT_cancer, non-tumorigenic cancer cell; P_cancer, metastatic prostate cancer; RA_M, rheumatoid arthritis on methotrexate; SCC, squamous cell carcinoma; SCC_M, squamous cell carcinomas (lymph node metastasis); SCF, severe cystic fibrosis; T_cancer, tumorigenic cancer cell; VGP_melanoma, vertical growth phase melanoma; VIN, vulvar intraepithelial neoplasia.

Figure 3. The precision and recall of target and pathway deconvolution.

Precision is the fraction of the identified targets or pathways that are correct, calculated as “true positive”/(“true positive” + “false positive”). Recall is the fraction of all true targets or pathways that are successfully identified, calculated as “true positive”/(“true positive” + “false negative”).