Explore Small Molecule-induced Genome-wide Transcriptional Profiles for Novel Inflammatory Bowel Disease Drug.

Inflammatory Bowel Disease (IBD) is a chronic and relapsing disorder, which affects millions people worldwide. Current drug options cannot cure the disease and may cause severe side effects. We developed a systematic framework to identify novel IBD drugs exploiting millions of genomic signatures for chemical compounds. Specifically, we searched all FDA-approved drugs for candidates that share similar genomic profiles with IBD. In the evaluation experiments, our approach ranked approved IBD drugs averagely within top 26% among 858 candidates, significantly outperforming a state-of-art genomics-based drug repositioning method (p-value < e-8). Our approach also achieved significantly higher average precision than the state-of-art approach in predicting potential IBD drugs from clinical trials (0.072 vs. 0.043, p<0.1) and off-label IBD drugs (0.198 vs. 0.138, p<0.1). Furthermore, we found evidences supporting the therapeutic potential of the top-ranked drugs, such as Naloxone, in literature and through analyzing target genes and pathways.

Introduction

Inflammatory bowel disease (IBD) is a group of chronic and relapsing inflammatory conditions in the intestinal tract[1-4]. It affects 3.6 million people in the USA and Europe, and has an increasing incidence around world in recent years[5]. IBD is currently not medically curable. Available drug options, such as corticosteroids, aminosalicylates and targeted drugs, focus on remitting the inflammation and are only effective for a subset of the patients[1-4]. In addition, these drugs may cause severe side effects, such as an increased rate of malignancies or infectious diseases[6]. Traditional drug development experiments have identified novel targeted drugs, such as TNF-alpha blockers, which showed a promising efficacy in treating IBD. However, recent clinical researches show that a large number of patients do not respond to these drugs[2,6] Therefore, discovering new drug treatment for IBD is of compelling need.

Computational drug repositioning approaches predict new indications for existing drugs, which have passed a number of toxicity and other tests, thus may lead to rapid and cost-effective drug discovery. Existing drug repositioning strategies can be classified into three categories: drug-based, disease-based and profile-based (Figure 1). Drug-based approaches (Figure 1(a)) assume that similar drugs can be used to treat the same disease. They leverage chemical structure[7,8], side-effects[9], and other drug characteristics to determine the drug similarity. Disease- based approaches (Figure 1(b)) predict new drug indications exploiting disease-disease similarity based on disease phenotypes[10], genetic and genomic signatures[11]. Recently, profile-based drug repositioning approaches (Figure 1(c)) have been developed to identify drug candidates for small-cell lung cancer[12], Parkinson’ disease[13], as well as IBD[14]. Unlike the drug-based and disease-based strategy, profile-based approaches do not depend on existing drug- treatment knowledge and may have increased ability to discover new drug-disease associations. Early profile-based approaches match the drug target genes with the disease-associated genes obtained from GWAS[15,16], OMIM[17] and network-based disease-gene predictions [13,18-20]. Several recent studies also investigate the similarity in phenotype profiles between diseases and drugs[21,22]

Figure 1.

Computational drug repositioning strategies: (a) disease-based methods, (b) drug-based methods, and (c) profile-based methods.

Most current profile-based drug repositioning approaches exploit the well-studied drug targets to match genetic and phenotypic profiles of the drugs and diseases[15-17]. However, the interactions between chemical compounds and human proteins are complex, and the off-targets of existing drugs may provide a unique path to link diseases with new potential treatments[14,23]. For example, Connectivity Map (CMap) is the first database that provides the genomic signatures associated with perturbations of small molecular compounds[23], and has facilitated the drug discovery for IBD[14]. Recently, The Library of Integrated Network-based Cellular Signatures (LINCS) offers more comprehensive drug-gene interaction data through significantly extending the experiments in the first version of CMap on much more cell lines (from 4 to 18) with more perturbagens (from 164 to 20413 chemical perturbagens) using more signatures (from 564 to millions) [23,24] Only several recent studies have used LINCS for drug discovery[25,26] through calculating drug-drug similarities based on gene expression profiles. In this study, we developed methods to mine genomic profiles for approved drugs from vast amounts of data in LINCS and matched profiles between IBD and the drugs. We demonstrated the improved performance of IBD drug prediction comparing with the state-of-art approach based on CMap[14], and show several valuable new drug candidates. Our framework can be easily extended to predicting drugs for other diseases with the LINCS data.

Methods

Our drug repositioning framework contains three steps (Figure 2): (i) generating genomic profiles for approved drugs from the LINCS data; (ii) constructing gene expression profile for IBD based on GEO; (iii) ranking all the candidate drugs based on the genomic similarity to IBD. The following sections describe the three steps as well as the evaluation method in detail.

Figure 2.

Workflow of our drug repositioning approach based on LINCS L1000 data.

A. Identify genomic profiles for approved drugs from the LINCS data

The Library of Integrated Network-based Cellular Signatures (LINCS) dataset is a catalog of gene-expression data collected from human cells treated with chemical compounds and genetic reagents. The LINCS L1000 dataset contains gene expression-profiling assays for 20413 small-molecular compound perturbagens and 22268 genes[24]. The expression profiles are generated for 1171 landmark transcripts while the rest of the transcriptome, approximately 22 000 genes, are estimated by a model built from computational processing of thousands of gene expression datasets from Gene Expression Omnibus (GEO)[27]. Several recent studies[25,26] computed transcriptional similarity based on landmark genes. Here, we explored the methods to integrate the complete data in LINCS L1000 and gain more prediction power in drug repositioning.

Gene expression data are available at different levels in the LINCS L1000 database (http://www.lincscloud.org/). We downloaded the differential expression level data, which consist of Z-scores representing the significance of the associations between genes and chemical perturbations. The data are stored in a 110GB GCTX file, which is a binary format file that allows for high-performance access. We extracted a large gene expression matrix from the file. The 22268 rows of the matrix represent probe sets (corresponding to genes) and the columns are millions of signatures, which represent replicated experiments for each chemical perturbation on all the genes. We obtained a list of 1761 FDA-approved drugs from DrugBank[28] and mapped them to 858 identical drugs in the L1000 database. Then we extracted the sub-matrix for the 858 approved drugs through the programmatic access to the dataset. The sub-matrix consists of 22268 rows and 100041 columns (each column represents one signature, and one drug corresponds to multiple signatures), and the Z-score stored in it was used to identify genomic profiles for the approved drugs.

In the sub-matrix of gene expression data for the 858 approved drugs, each gene may correspond to multiple probe sets (rows) and each drug is mapped to multiple signatures (columns). As a result, a drug-gene link is represented by a multiple-to-multiple relationship in the matrix. To represent the strength of a drug-gene association, we counted the number of genes that are over and under-expressed and constructed a sign score vector for each drug. Specifically, we determined the genes as over-expressed if the Z-score was greater than 2 and under-expressed if the Z-score was less than -2 following the previous study[25] (Figure 3(a)). For example, in figure 3, we marked the over- expressed genes as red and the under-expressed genes as blue in the Z-score matrix. Then we represented the gene- drug relationship by extracting corresponding probe sets (rows) and relative signatures (columns) into a small matrix (Figure 3(c)). Last, we computed a sign score (Ssign) for the drug-gene pair: Given a drug, a sign score was calculated for each of the 12716 genes, and vector consisting of 12716 sign scores was constructed to represent the genomic profile for the drug:

Figure 3.

The schematic representation of workflow to construct drug-gene profile: (a) Representation of determining over- and under-expression, (b) Z-score matrix of gene expression level profile in LINCS can be painted using our schema, (c) Relationship between gene and drug in LINCS.

B. Identify the genomic profile for IBD from GEO

We downloaded the study GSE6731 and identified the significant differentially expressed genes for IBD. A total of 36 samples assembled as gene expression patterns of active and inactive areas of IBD, infectious colitis and healthy controls from human colonoscopy biopsy. For a fair comparison, we selected the same study as the CMap-based method used. We used the GEO2R software[29] to detect the significant genes. Then, we converted P-values generated from GEO2R to Z-scores and set the threshold of 2/-2 (the same one in Method A) to identify significant genes for disease genomic profiles.

C. Calculate genomic similarity between IBD and drugs

We computed the cosine similarity between drug and disease genomic profiles to estimate the potential of each drug in treating IBD. The similarity is defined as: in which A and B are sign score vectors (represented as in equation (2)) of drug and disease, respectively. Then we ranked the candidate drugs based on this genomic similarity.

D. Compare with an existing baseline method

We compared our approach with a state-of-art drug repositioning method for IBD, which exploited the CMap data to identify the genomic profiles for drugs[14]. Our method and the CMap-based approach used the same GEO dataset to identify significant genes for IBD. They provided a ranked list of 79 drugs for IBD, and considered the rest of the 164 drugs irrelevant to IBD. We simulated the full rank of CMap-based approach by randomizing the irrelevant drugs for 100 times for the purpose of fair comparison. Then we constructed evaluation sets containing approved IBD drugs and novel IBD drugs respectively, and compared the two methods in prioritizing the evaluation drugs.

In our recent studies, a large-scale drug-disease treatment knowledge base was constructed based on multiple resources[30,31,32]. The databases include 9,216 drug-disease treatment pairs extracted from FDA drug labels, 34,306 pairs extracted from 22 million published biomedical literature abstracts, 69,724 disease-treatment pairs extracted from clinical trials and 111862 pairs from the FDA post marking surveillance system.

a. Evaluation on FDA-approved IBD drugs

We extracted 25 FDA-approved IBD drugs from FDA drug labels in this knowledge base, and found 10 of them appear in the LINCS dataset. The intersected approved drug set consists of 8 corticosteroids and 2 aminosalicylates. We plotted the number of approved IBD drugs that fell into 10 ranking ranges to estimate their distribution. Mean and median of the ranks for the approved IBD drugs were computed and compared with the CMap-based method. We also evaluated whether the two methods generated significantly different ranks for the approved IBD drugs.

b. Evaluation on novel IBD drugs

Based on this knowledge base, we constructed two evaluation sets as the proxy of novel IBD drugs. The first sets contain 58 potential IBD drugs that have been tested in clinical trials. The second one contains 238 off-label IBD drugs collected from the FDA post marking surveillance system. FDA-approved IBD drugs are removed from both sets. We used the histogram plot to estimate the distribution of drug ranks in two evaluation sets, respectively. We also calculated the precision-recall curve and the mean average precision to evaluate the ranking of the two methods.

Results

A. Our system prioritized the FDA-approved IBD drugs and outperformed the CMap-based method

Figure 4 show that our approach prioritized most approved IBD drugs. In particular, our method ranked 5 out of 10 approved IBD drugs within top 10% among 858 drugs. The average rank of 10 FDA-approved IBD drugs among 858 drugs is 26% (significantly higher than the random rankings, p-value < e-8, T test). The median rank is 19.5%, which is also significantly lower than random case (p-value < e-8, Wilcoxon signed sum test). Together, the results show that our ranking prioritized most of the approved IBD drugs and performed significantly better than random ranking. These results also support that our methods of processing the LINCS data and calculating the disease-drug profile similarity are reasonable. We noticed that two approved drugs, Methylprednisolone and Balsalazide, were ranked a bit lower than 50%. This might be due to small sample size for the two drugs: most of the chemical compounds in LINCS have several hundreds of replicated signatures, while the sample size of Methylprednisolone and Balsalazide is only 10% comparing to the average.

Figure 4.

Distribution of FDA-approved IBD drugs.

We also compared with the CMap-based method in predicting approved IBD drugs. Because of their smaller search scope (only three approved IBD drugs appear in the data), the CMap-based system only identified one approved IBD drug as the significant candidates (within top 79). The average rank of all the approved IBD drugs in the simulated ranks generated by the CMap-based method is 34%, which is significantly lower than ours average rank (p-value < e-8, T test).

B. Our system identified novel IBD drugs and performed significantly better than the CMap-based method

In the two evaluation sets of novel IBD drugs, 24 drugs from the clinical trials and 141 from the post marketing surveillance system appear in our drug repositioning system. We ranked 7 drugs from the clinical trials within top 10%, and 13 within 30%. In addition, 23 off-label IBD drugs from the post marketing surveillance system were ranked within top 10% and 88 of them (58%) were ranked above 50%. We plot the distribution of the evaluation drugs among all ranked list in figure 5(a)-(b), and found an increasing trend for the number of positive drugs when the rank moves from the bottom to the top. In contrast, the off-label drugs spread evenly over different ranks in the result of the CMap-based approach (figure 5(d)). The evaluation drugs from the clinical trials are even likely to be ranked in the bottom by the CMap-based approach (figure 5(c)). The comparison of precision-recall curves between the two methods (figure 6) suggests the same result: our approach achieved a mean average precision of 0.072 and 0.198 for the two evaluation sets respectively, which are significantly better than 0.043 (p<0.1) and 0.138 (p<0.1) obtained by the CMap-based approach.

Figure 5.

Our approach: (a) Distribution of count of drugs from clinical trials. (b) Distribution of count of drugs from FDA post marking surveillance system. CMap-based approach: (c) Distribution of count of drugs from clinical trials. (d) Distribution of count of drugs from FDA post marking surveillance system.

Figure 6.

Precision-recall curves in ranking novel drugs from clinical trials (left); Precision-recall curve in ranking novel drugs from FDA post marking surveillance system (right).

We noticed that in predicting drugs from clinical trials, the precisions at different cutoffs of recall are relatively low for both two methods. This mainly dependents on the size of gold standard set, as CMap-based approach intersected only 7 drugs from clinical trials and 24 potential drugs from clinical trials appeared in our approach. This precision recall curve and corresponding mean average precision value do not represent the true predictive power of two methods. We evaluated precision recall here only for comparison of rankings of two methods.

C. Top-ranked predicted drugs show promise in treating IBD

Among the top-ranked drugs, we found Naloxone, which was ranked in top one, and Naltrexone in top seven have the potential in treating IBD. Both of them are opioid antagonists and have similar effects[33,34]. Low-dose naltrexone (LDN) has been demonstrated as a novel anti-inflammatory treatment for chronic pain conditions, such as fibromyalgia, Crohn’s disease, multiple sclerosis, and complex regional pain syndrome[33]. Naloxone is traditionally approved for complete or partial reversal of narcotic depression. We further extracted the significantly differentially expressed genes induced by Naloxone from LINCS, and found many common drug-interacting genes are closely associated with IBD pathways (Table 1). For example, the target gene CXCL1 is over-expressed in Naloxone and influences the pathway of Cytokine-cytokine receptor interaction, which regulates the innate as well as adaptive inflammatory host defenses. IL18 and STAT1 are over-expressed in Naloxone, and contained in the pathway directly associated with IBD. We ranked the pathways related to Naloxone and IBD respectively based on the gene-pathway associations and found that top six pathways are consistent between the drug and the disease (Table 2). The pathway analysis results support the therapeutic potential of Naloxone for IBD.

Table 1.

Common target genes that are significantly differentially expressed and related pathway of top-ranked drug Naloxone and IBD

Gene symbol	Pathway
HLA-DMA	Reactome immune system
RELB	Pid_IL12_2pathway
CXCL1	Cytokine-cytokine receptor interaction
IL18, STAT1	Inflammatory bowel disease
IL6	Jak-STAT signaling pathway

Table 2.

Comparison of pathway ranks between top-ranked drug Naloxone and the disease IBD

Rank	Naloxone	IBD
1	REACTOME_GPCR_LIGAND_BINDING	REACTOME_GPCR_LIGAND_BINDING
2	REACTOME_CYTOKINE_SIGNALING_IN_IMMUNE_SYSTEM	REACTOME_CYTOKINE_SIGNALING_IN_IMMUNE_SYSTEM
3	REACTOME_PLATELET_ACTIVATION_SIGNALING_AND_AGGREGATION	REACTOME_PLATELET_ACTIVATION_SIGNALING_AND_AGGREGATION
4	REACTOME_METABOLISM_OF_LIPIDS_AND_LIPOPROTEINS	REACTOME_METABOLISM_OF_LIPIDS_AND_LIPOPROTEINS
5	REACTOME_ADAPTIVE_IMMUNE_SYSTEM	REACTOME_ADAPTIVE_IMMUNE_SYSTEM
6	REACTOME_HEMOSTASIS	REACTOME_HEMOSTASIS
7	REACTOME_METABOLISM_OF_CARBOHYDRATES	REACTOME_ANTIGEN_PROCESSING_UBIQUITINATION_PROTEASOME_DEGRADATION
8	REACTOME_NONSENSE_MEDIATED_DECAY_ENHANCED_BY_THE_EXON_JUNCTION_COMPLEX	REACTOME_PI3K_CASCADE
9	REACTOME_ANTIGEN_PROCESSING_UBIQUITINATION_PROTEASOME_DEGRADATION	REACTOME_SIGNALING_BY_FGFR
10	REACTOME_METABOLISM_OF_PROTEINS	REACTOME METABOLISM OF CARBOHYD RATES

In addition, previous researches show that severity of IBD varies during pregnancy, menopause, or oral contraceptives use, and gonadal sex hormones may be involved in the IBD pathogenesis[35]. Interestingly, Estradiol and Estriol, which are ranked in the second and fourth by our approach, are two female sex hormones. An earlier Estrogen study shows that female sex hormones may result in less severed IBD and this correlates with decreased pro-inflammatory mediators levels[35]. Scopolamine, a secondary metabolite from Solanaceae family, is atropine alkaloid drug with muscarinic antagonist effects[36] and has been used to control gastric peristalsis in bowel diseases[37]. The 6th in the rank list is Doxepin, which is quite interesting as it’s depression drug while shown Anti-ulcer effect[38]. Anti-ulcer effect shows relevance for treating IBD[39]. In the 10th of the list is Nafcillin, which is a narrowspectrum antibiotics of the penicillin class, and it is interestingly found to enhance innate immune-mediated killing of Staphylococcus aureus[40].

Discussion

In this study, we developed a drug repositioning framework based on genomic profile matching between drugs and diseases. We leveraged the data of differentially expressed genes induced by drugs from the most recent pharmacological database called LINCS. Our work has several limitations that can be improved in the future studies. First, the evaluation of FDA-approved drugs shows that the prediction power is affected by the sample size of drug expression signature. When the sample size of signatures of a drug is much smaller than the average size of replicates of signatures, the rank of it is relatively low as a result. In the future, we plan to experiment on adding weights to normalize the sample size in the sign score for drug-gene pairs and balance the low rank of minority caused by small sample size.

In addition, we currently focus on investigating the genomic profile similarity between drugs and diseases, and treat the over- and under-expression equally in calculating the similarity. In the future, we will consider the direction of the gene expression level and only match the significant genes with opposite directions between drugs and diseases. We also plan to incorporate higher-level phenotype profiles in evaluating the relationships between drugs and diseases[13].

Finally, we currently detected significantly expressed genes for IBD from one GEO study, primarily in order to make a fair comparison with the CMap-based approach. Since the gene expression studies in GEO are highly noisy, we will integrate the results based on multiple studies later when applying our system in predicting new drugs for IBD and other disorders.

Conclusion

We have developed a novel drug prediction framework leveraging drug-induced gene expression profiles in LINCS. We demonstrated that our approach identifies approved and novel IBD drugs and achieved significantly better performance comparing with a state-of-art drug repositioning approach based on an early pharmacological database called CMap. Literature-review and pathway analysis show that the top-ranked predicted drugs have the potential in treating IBD.