Prediction of new drug indications based on clinical data and network modularity.

The human symptom-disease data is got from 57 . Based on 322 disease symptoms, a weighted disease network is constructed, which contains 133,106 interactions between 1,596 distinct diseases (see Supplementary Table S1 ). The disease network combines phenotypic relations with shared molecular mechanisms. The drug-side effect data is downloaded from SIDER (Side Effect Resource) version 2 58 . SIDER contains information on 996 marketed drugs, corresponding 4,192 recorded adverse drug reactions, and 99,423 drug-side effect pairs (see Supplementary Table S2 ). The information is extracted from public documents and package inserts. The available information include side effect frequency, drug and side effect classifications as well as links to further information, for example drug–target relations. We obtain the drug–disease associations from Comparative Toxicogenomics Database (CTD) in December 2015 22 . CTD contains curated and inferred chemical–disease associations. They can help researchers develop hypotheses about environmental diseases and their underlying mechanisms. We only preserve the curated associations with marker “therapeutic” or “marker/mechanism” as known associations. “Therapeutic” represents that a chemical has a known or potential therapeutic role in a disease (e.g., chemical X is used to treat leukemia). “Marker/mechanism” represents that a chemical correlates with a disease (e.g., increased abundance in the brain of chemical X correlates with Alzheimer disease) or may play a role in the etiology of a disease (e.g., exposure to chemical X causes lung cancer) 22 . Finally, we get 82,858 chemical–disease associations in all (see Supplementary Table S3 ). The relationship between drug j and side effects is treated as a feature vector d j : where w i,j quantifies the strength of the association between side effect i and drug j . The prevalence of the different side effects and diseases is very different. For example, there are highly abundant side effects like abdominal pain, and publication biases towards certain drugs. To account for this heterogeneity, we do not use the absolute co-occurrence W i,j to measure the strength of an association between side effect i and drug j , but the term frequency-inverse document frequency 59 w i,j : where N denotes the number of all drugs in the dataset and n i denotes the number of drugs in which side effect i appears. W i,j is equal to 1, if drug j displays side effect i , otherwise, it will be 0. Since all side effects in our data have at least one associated drug, the potential problem of dividing by zero does not arise. For the 996 drugs with recorded clinical side effect data, each will be assigned a 4192-dimension vector. We use the cosine similarity 60 to measure the similarity between the vectors d i and d j of two drugs i and j . The formula is shown as: where the cosine similarity ranges from 0 (no shared side effects) to 1 (identical side effects). After filtering drug and disease networks, we respectively cluster drug network and disease network using ClusterONE (Clustering with Overlapping Neighborhood Expansion) 61 and obtain two kinds of modules: drug module and disease module. ClusterONE is a graph clustering algorithm that is able to handle weighted graphs. Owing to these properties, ClusterONE is especially useful for detecting modules in networks with associated confidence values. In order to improve the reliability of the drug network, we filter the edges with lower similarity. To remain relatively more reliable edges, the cutoff is set to be 0.4 and we get a new drug network with 248 nodes and 379 edges, as shown in Supplementary Table S4 . We run ClusterONE with default parameter values in the filtered drug network and get 27 modules (See Supplementary Table S5 ). The p-values of 26 modules are lower than 0.05. The network topologies of 27 modules are also shown in Fig. 3 (nodes representing drugs). Rectangles and diamonds represent nodes in clusters and overlap between different clusters, respectively. When the scores of edges in disease network are not lower than 0.5, we can obtain more meaningful modules (p–value ≤ 0.05). Therefore, we discard the edges with scores lower than 0.5. After filtering, the disease network includes 1,367 nodes and 9,792 edges, as shown in Supplementary Table S6 . We also run ClusterONE with default parameter values in disease network and get 145 modules (see Supplementary Table S7 ). The p-values of 64 modules are lower than 0.05. Based on reliable chemical-disease associations got from CTD database, we construct connections between disease and drug modules. As an example shown in Fig. 4 , we assume drug module i includes four drugs (drug1 to drug4) and disease module j includes four diseases (disease1 to disease4). The strength of the association between drug module i and disease module j , w ( i , j ), is defined as: where M ij denotes the number of chemical-disease associations verified by CTD (marked as T or M/T) between drug module i and drug module j ; N ( i ) and N ( j ) denote the number of elements in drug module i and drug module j , respectively. In Fig. 4 , M ij  = 5 and N ( i ) =  N ( j ) = 4, so . In this way, we can get the correlation between any drug-disease module pair. Finally, we obtain 1,180 drug-disease module pairs whose score are not zero. According to the definition of the association between a drug-disease module pair (see formula (5)), the drug-disease module pairs with higher score will be preserved. There are 27 drug-disease module pairs whose scores are all not lower than 0.2, which are shown in Supplementary Table S8 . In order to analyze our results more targeted and find more valuable associations, we focused on the top-3 drug-disease module pairs for further analysis. The top-3 drug-disease module pairs whose correlations are all not lower than 0.36. Their details are shown in Table 1 . In Results section, we will analyze the top-3 drug-disease module pairs to predict new drug indications. In this section, we predict novel drug-disease associations based on the neighborhood partnerships of drugs and diseases in their own modules. Table 1 shows the information of top-3 drug-disease module pairs. We can find these modules are dense, that is to say, the nodes in a same module interact frequently. We know that the stronger the interactions between nodes within a same module, the more similar their functions. Therefore, based on the strong correlations between dense drug and disease modules, we can predict potential associations between drugs and diseases. The details are as following: Step 1: Recalculate the similarity of each pair of nodes in a module based on subgraph topology of the module. Given a and b is a pair of nodes in a drug module or a disease module, there may be many different paths between a and b . Suppose path i represents a path between a and b with i edges; sum ( path i ) represents the weight of path i , which equals to the sum of weight on each edge in it; avg ( path i ) represents the average weight of path i : avg ( path i ) =  sum ( path i )/ i . Then, the similarity between a and b is defined as: where n represents the number of nodes in the module. If there does not exist a path with i edges, we set avg ( path i ) = 0. In this way, we score all the correlations between each pair of nodes in a module. From the above steps, we know drugs or diseases have similar properties are clustered in a same group. The indications of a drug may be inferred by the enriched FDA-approved functions of its neighboring drugs in a same drug subgraph. Relationships in strong similar edges and connected components provide potential candidates for the previously unknown therapeutic effects of drugs. In the same way, the treatments of a disease may be inferred by its neighboring diseases in a same disease subgraph. Therefore, for a drug-disease module pair, we predict new drug-disease associations based on the subgraph topology of modules. Step 2: Recalculate the association of each pair of nodes belong to two different modules based on Step 1. Based on known drug-disease associations in CTD database, and correlations of drug-drug and disease-disease obtained from Step 1, we get a drug-disease heterogeneous network, which includes two types of nodes: drug and disease, and two types of edges: internal connections and external connections. Then, based on the heterogeneous network, we score the correlations between each pair of drug and disease nodes. Given v represents a node in disease module DI and w represents a node in drug module DR , that is v ∈ DI and w ∈ DR . Their relation c orr ( v , w ) is defined as: The direct neighbors of v in DI , which have connections with nodes in DR , are stored in N DI ( v ). The direct neighbors of w in DR , which have connections with nodes in DI , are stored in N DR ( w ). sim ( v , p ) and sim ( w , q ), defined as Equation (6) , denote the similarities between v and p , w and q , respectively. After the process of Step 1 and Step 2, we will score all the drug-disease associations between a drug-disease module pair and rank them in descending order. We do the same process for the top-3 drug-disease module pairs and get three ranking lists of drug-disease associations.

Based on the top-3 drug-disease module pairs, shown in Table 1 , we get three ranking lists of drug-disease association and predict new drug indications. The top associations in each list are reliable. Therefore, we evaluate the top-20 predicted associations in each list by their overlap with drug indications that are reported in published literature and Comparative Toxicogenomics Database (CTD; http://ctd.mdibl.org ) benchmark, and we also make Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis on them. Table 2 , 3 , 4 respectively show three lists of top-20 drug-disease associations corresponding to the three drug-disease module pairs in Table 1 (Rank = 1 to 3). In these tables, the predicted results are divided into two categories: known and potential associations. If there is direct evidence for a chemical-disease association marked as “therapeutic” or “marker/mechanism” in CTD database, we take it as a known association. Otherwise, it is considered as a potential drug-disease association (marked as bold italic items), such as prochlorperazine and Tremor (ID = 9) in Table 2 . In the following section, we abbreviate “therapeutic” to “T”, and “marker/mechanism” to “M”. Figure 2 shows the network topology of the first drug-disease module pair (Rank = 1 in Table 1 ). In the figure, internal connections within a module are labelled by blue, and external connections between two modules are labelled by purple. Green circle and red diamond nodes represent drugs and diseases, respectively. The purple edges represent the reliably curated drug-disease associations in CTD, which are also shown in Table 2 (CTD mark are “M” or “M&T”). In Table 2 , we find 18 of 20 are known associations. The percentage reaches up to 90%. Two results are new predictions and labelled by bold italic. The association between fluphenazine and Parkinsonian Disorders (ID = 19) is inferred by CTD (Inference Score = 7.38). The inference score 23 reflects the degree of similarity between CTD chemical-gene-disease networks and a similar scale-free random network, which is computed as shown below: where Y represents the inference score; P represents the probability that a vertex in a large network interacts with another vertex decays according to a power law 24 ; G , C , and D represent a gene, chemical, and disease respectively; k represents the number of connection between G , C , or D ; n G represents a gene set. The higher the Y score, the more likely the inference network has atypical connectivity 25 . We also find the inferred association (ID = 19) is based on three genes: NGF 26 , DRD2 27 , and PRL 28 . Moreover, the two results (in bold italic) are also supported by external literature. Studies 29 have shown that prochlorperazine can cause Tremor (ID = 9). People possibly have tremor while taking prochlorperazine from FDA (Food and Drug Administration) and social media. Fluphenazine (ID = 19) is a drug used to treat psychotic disorders, agitation, and dementia 30 . And its use may lead to the development of symptoms that resemble Parkinson’s disease 31 . Supplementary Fig. S1 shows the relationships of the second drug-disease module pair (Rank = 2 in Table 1 ). Table 3 gives the details of top-20 drug and disease associations. In this table, sixteen associations are known and the remaining four ones marked as bold italic are new predicted results. The association between thorazine and Lewy Body Disease (ID = 16) is inferred associations in CTD (Inference Score = 3.77). The relationship is based on the gene “MAG”. Evidences show that thorazine affects the expression of MAG mRNA 32 and that MAG expression significantly relates to Lewy Body Disease 33 . Thorazine is an antipsychotic medication, which is primarily used to treat psychotic disorders such as schizophrenia. Lewy body dementia (LBD) is a type of progressive dementia that leads to a decline in thinking, reasoning and independent function because of abnormal microscopic deposits that damage brain cells over time 34 . LBD is found among people who take thorazine, especially for people who are male, more than 60 years old, also take medication protonix, and have insomnia 35 . The researchers found thorazine had shown better capacity to control hyperkinetie manifestations, and had not shown any undesirable side effects 36 37 . For the association (ID = 10), thioridazine is a typical antipsychotic drug used in the treatment of Psychoses (ID = 10) 38 . Prochlorperazine (ID = 18) belongs to a group of medicines called “phenothiazines”. It is prescribed for a variety of unrelated conditions, including problems with balance and dizziness, sickness, agitation and restlessness, and schizophrenia 39 . Bipolar Disorder (ID = 18) is found among people who take prochlorperazine, especially for people who are female, more than 60 years old, also take medication klonopin, and have nausea 40 . Symptoms typical of alcohol withdrawal (ID = 20) include agitation, seizures, and delirium tremens 41 . Laties and colleagues 42 determined that promazine and chlorpromazine (ID = 20) were equally efficacious in the treatment of delirium tremens. Chlorpromazine, promazine, and other low-potency typical antipsychotic agents have been reported 43 to have the greatest effect on lowering seizure threshold. The network topology of the third drug-disease module pair (Rank = 3 in Table 1 ) is shown in Supplementary Fig. S2 . From Table 4 , we find 2 associations (in bold italic) are new predictions by our method. The other 18 associations are all known in CTD. The percentage reaches up to 90%. Though, two other associations (ID = 15 and ID = 20) are not existed in CTD at present, their associations are supported by features of their ancestors and literatures. Circadian rhythms (“body clocks”) are controlled by a biological clock and work on a daily time scale. Circadian rhythm sleep disorders (CRSD) (ID = 15) are a family of sleep disorders. People with circadian rhythm sleep disorders are unable to sleep and wake at the times required for normal work, school, and social needs. We find amitriptyline (ID = 15) has strong connections with sleep disorders in CTD (marked as “M&T”). They are related by the gene “CHRNB2”: amitriptyline results in the decreased expression of CHRNB2 mRNA 44 and CHRNB2 protein results in the increased susceptibility to amitriptyline 45 . Hence, it is reasonable to infer that circadian rhythm sleep disorders and amitriptyline probably have close connection. Moreover, amitriptyline is found useful as a sleep aid continues 46 . Postpartum depression (ID = 20) is moderate to severe depression in a woman after she has given birth, which is a descendant of depressive disorder. The relationship between imipramine (ID = 20) and depressive disorder is marked as “therapeutic” and “marker/mechanism”, which is based on five genes: ALB 47 , BDNF 48 , CRH 49 , POMC 50 , SLC6A4 51 . Therefore, it is likely that postpartum depression has connection with imipramine. Furthermore, Cohen and Rosenbaum 52 pointed out the use of tricyclic antidepressants should not pose a risk when used in pregnancy or in the postpartum period. They maintained that the safest medications to use at this point in time are nortriptyline, imipramine, and fluoxetine. In this section, for the predicted drug-disease associations shown in bold italic in above tables, we perform KEGG pathway enrichment analysis on their related gene sets with the functional annotation tool of DAVID 53 . For DAVID, EASE Score, a modified Fisher Exact P-Value, is used as a threshold for gene-enrichment analysis 54 . It ranges from 0 to 1. When Fisher Exact P-Value is 0, it represents perfect enrichment. We set it as 0.01. For each drug, we combine the genes obtained from Drugbank database 55 , which combines detailed drug data with comprehensive drug target information, and the top interacting genes got from CTD database. Similarly, the genes related to each disease are got from OMIM database 56 , which is a comprehensive, authoritative compendium of human genes and genetic phenotypes. For each predicted drug and disease pair, we respectively put their corresponding genes into DAVID and examine whether the drug and the disease have overlapped KEGG pathways. If they have more overlapped pathways, they are more relevant, that is the drug has strong correlation with the disease. For example, prochlorperazine and tremor (Rank = 9 in Table 2 ) has one overlapped pathway: “Neuroactive ligand-receptor interaction” (p-value = 1.2E-3). We find thioridazine (Rank = 10 in Table 3 ) has two pathways overlapping with psychoses, substance-induced: “neuroactive ligand-receptor interaction” and “calcium signaling pathway”. Their corresponding p-values are 2.14E-05 and 2.81E-04 respectively. Lower p-values indicate the predicted associations are reliable. Especially, the pathway overlapped between amitriptyline and dyssomnias (Rank = 16 in Table 4 ) is “neuroactive ligand-receptor interaction”, which has a very low p-value: 8.66E-29. These good results show their high reliability. Because of the incompleteness of data, the numbers of genes related to some drugs and diseases are all small, such as prochlorperazine and bipolar disorder (Rank = 18) in Table 3 . Therefore, it is hard to find pathways related to prochlorperazine and bipolar disorder using DAVID tool at present. With the improvement of data, the performance of our method will be more effectively. To evaluate the performance of our method, we compare it with PREDICT 21 . PREDICT integrates multiple data sources, including chemical structures, drug side effects, drug target protein sequences and target protein interactions and phenotype data. It applies network analysis for target protein distance calculation, applies text mining to identify disease phenotypes and use machine learning algorithms to classify true and false drug-disease associations 2 . Based on CTD benchmark, we make a comparison between our method and PREDICT 21 . We choose the top-20 drug-disease associations of PREDICT for analysis. They are shown in Table 5 . In the table, the column “Disease name” sometimes includes more than one disease name. As long as one disease is found related to a drug in CTD, the association is marked as “T”, “M”, “M&T” or “inferred”. Taking the fourth association (Rank = 4) as an example, its “Disease name” = “Pyogenic Sterile Arthritis, Pyoderma Gangrenosum, And Acne” (Rank = 4). We find “Arthritis” is related to “Gonadorelin” in CTD and it is an inferred association. Thus the association is marked as “inferred”, i.e. CTD mark = “inferred”. Finally, there are 14 drug-disease connections are found in CTD benchmark, but only 5 of them are known associations (CTD mark = “T” or “M&T” and marked as bold italic in Table 5 ) and the other 9 associations are inferred in CTD (CTD mark = “inferred”). Its precision is 5/20 = 0.25, which is lower than that of our method. The precision of our top-20 associations in each list is more than 0.8. From Table 2 , 3 , 4 , they are 0.9, 0.8, and 0.9 respectively.

How to cite this article : Yu, L. et al . Prediction of new drug indications based on clinical data and network modularity. Sci. Rep. 6 , 32530; doi: 10.1038/srep32530 (2016).

Based on side effects of drugs and symptoms of diseases, we construct drug and disease networks firstly. Then we cluster them to get two types of modules: drug module and disease module. According to the known drug-disease associations in CTD database, we score each pair of drug-disease module and reserve the top-3 drug-disease module pairs. For each selected drug-disease module pairs, we construct a drug-disease bipartite graph. We calculate all the connections to predict potential drug-disease associations. The significant enrichments of our predictions in the biomedical literature, clinical trials and KEGG pathways demonstrate that our approach can effectively identify new indications as an indicator for the mode of action. The success of our methods PDCIM can be attributed as follows: First, we integrate clinical data into our model, such as disease symptoms and drug side effects. Second, our approach is based on clustering and known drug-disease associations. Finally, we combine the neighborhood information of nodes in drug modules and disease modules. We believe that the combination of clinical data, network clustering and subnetwork connectivity could help us to predict new hypotheses to infer the drug-disease relationship and even improve the drug development. Moreover, our predicted drug-disease associations are not generated using the genes associated with diseases or drugs, so we can find some drug-disease associations that these drugs or diseases have less or no related genes. However, the limitation is the difficulty in distinguishing the positive and negative associations between drugs and diseases. In the future, we can choose different methods to calculate the similarity between drugs and diseases to enhance the reliability of our constructed networks. On the other hand, we can integrate various data sources such as pharmacological data, therapeutic/toxicological expression profiles and DNA methylation data to try to distinguish the positive and negative associations between drug and disease.