Persistent pain is a major healthcare issue, as defined by WHO, affecting about a fifth of the European population increasing to a third in the over 70-year old (Elliott et al., 1999; Breivik et al., 2006). It has a highly complex pathophysiology (Julius and Basbaum, 2001; Schaible, 2007) and it is triggered by several different causes, such as cancer (Portenoy, 1992) and surgery (Kehlet et al., 2006). Therefore, the search for novel analgesic strategies is an active research topic receiving public funding (Kringel and Lötsch, 2015). In addition to the identification of novel drug targets from molecular research as the main line of research, scanning the existing pharmacopeia for repurposing candidates becomes increasingly successful (Ashburn and Thor, 2004). This is facilitated by developments in computational data science (President's Information Technology Advisory Committee, 2005).

An accepted source of novel options for the pharmacological treatment of (persistent) pain is the study of genes causally involved in hereditary syndromes with insensitivity to pain (Goldberg et al., 2012; Table 1). This set of genes has provided the targets of novel analgesics (Table 2). However, this unique set of genotypes also enables the study of key biological processes of pain from the perspective of the functions of the involved genes. As shown recently, similarities between the profiles of biological processes in which the genes coding for the targets of available drugs with the profile of the biological processes in which a given set of genes is involved can be employed for drug classification or repurposing (Lötsch and Ultsch, 2016a,b). Hence, picturing the genetic background of human insensitivity to pain could be explored for drug repurposing based on functional similarity in addition to the development of novel substances targeting some of the respective gene products (Table 2). The present analysis made extensive use of computational biology, knowledge discovery methods, publicly available databases and data mining tools (Table 3) to merge results from pain, genetics and pharmacological research.

Data were analyzed using the R software package (version 3.3.3 for Linux; http://CRAN.R-project.org/; R Development Core Team, 2008). Following querying the relevant sets of genes by mining publicly available databases, the analyses aimed at (i) identifying the systems biology of hereditary insensitivity to pain as conferred by the functions of the causally involved genes and (ii) to explore whether this knowledge can be employed for drug repurposing approaches aimed at identifying potential candidates for the treatment of pain. The analytical steps are summarized in Figure 1 and described in full detail in the following paragraphs.

The present analysis used empirical evidence for human genes that when nonfunctional may cause the phenotype of insensitivity to pain. The set of n = 20 genes included those causally involved in hereditary sensory and autonomic neuropathies and for long in the center of pain research, and in addition further genes involved in several different neurological syndromes or, such as SCN9A, genes that when nonfunctional are mainly associated with insensitivity to pain. Indeed, the only additional phenotype in subjects carrying loss-of-function mutations in SCN9A was anosmia (Cox et al., 2010). In the present analysis, the common phenotype of insensitivity to pain associated with the complete gene set was used to address the functional background of pain insensitivity and to explore a computational attempt at using the information for drug repurposing. The computational approach allowed using data on the biological functions of genes acquired in any context, without restriction to pain research (Lötsch et al., 2013).

The computational analysis of genes associated with human insensitivity to pain included, in addition to basic functions as neuronal signaling or the perception of pain, as major components (i) nervous system development and structural component assembly and (ii) lipid-mediator based signaling including sphingosines and ceramides. Nervous system development, while presently probably also associated with the hereditary developmental disorders of the nervous system, appear to be a consistent part of the genetic background of pain since they also emerged as particularly prominent systemic functions of larger sets of genes related to all aspects of pain (Ultsch et al., 2016). As discussed there, as possible explanation of the consistent appearance of nervous system development among key biological processes exerted by pain-relevant sets of genes involves the concept of chronic pain as a dysregulation in biological processes that describe its systemic features of learning and neuronal plasticity (Mansour et al., 2014). Thus, the structural aspect of nervous system development would be compatible with both, hereditary developmental neuropathies and pain.

A role of lipid signaling is compatible with prior knowledge of the pathophysiology of pain. Indeed, the ceramide-to-sphingosine pathway has been proposed already as a therapeutic target in pain (Salvemini et al., 2013). Similarly, sphingosine-1-phosphate induced nociceptor excitation and ongoing pain behavior in mice and humans (Camprubi-Robles et al., 2013) and therefore, sphingosine-1-phosphate receptors have been proposed as novel targets for the treatment of pain (Welch et al., 2012). While these findings have emerged from basic research, the present prominent role of this pathway highlights the suitability of the computational approach analyzing available information about a particular subset of pain-related genes. Again, the biological plausibility applies to both, developmental neuropathies and pain.

The idea behind using the knowledge about the biological processes characterizing insensitivity to pain for drug repurposing is to use the similarity measure in the high-dimensional vector space of the drug's interactions with biological processes for the identification of substances qualifying for the treatment of traits defined on the basis of biological processes (Lötsch and Ultsch, 2016b). A disease-relevant gene set is functionally analyzed and compared with the biological processes in which the targets of available drugs are involved. This approach presently resulted in the selection of a subset of n = 22 substances, which is the size of the cluster that surrounded the functional information of the present trait of interest in the high-dimensional data space (Figure 2). Results obtained with the present computational approach require basic research verification. However, the set included substances for which such evidence could be found. For example, myristic acid is contained at 18.64% in coconut oil, which when orally administered to rats produced moderate anti-inflammatory and anti-nociceptive effects (Intahphuak et al., 2010). Imatinib, a tyrosine kinase inhibitor, had no antinociceptive effects in a nerve injury model in rats when administered alone; however, when combined with a previously ineffective dose of morphine, complete pain relief was obtained (Donica et al., 2014), which was attributed to its platelet-derived growth factor inhibiting actions (Mcgary et al., 2004).

Similarly, flavopiridol or alvocidib, an inhibitor of cyclin-dependent kinases, was shown to facilitate the recovery from tactile allodynia when in a rat nerve injury model, which was attributed to its Janus kinase pathway inhibiting actions (Tsuda et al., 2011). Ellagic acid, a polyphenolic compound from plants such as raspberries, eucalyptus, and nuts (Clifford and Scalbert, 2000), showed antinociceptive effects and potentiated the effects of morphine in different rat models of pain (Mansouri et al., 2014). Furthermore, staurosporine, an alkaloid initially isolated from the bacterium Streptomyces staurosporeus (Omura et al., 1977) and found to inhibit protein kinases, inhibited Angiotensin II mediated sensitization of post mortem analyzed human nerves (Anand et al., 2015), which is in line with the targeting of angiotensin 2 type II receptors as a novel treatment for neuropathic pain (Rice et al., 2014).

However, the list of candidate drugs for repurposing also included the multitargeted receptor tyrosine kinase inhibitors sorafenib and sunitinib, for which evidence suggests hyperalgesic actions. This emphasizes that, as discussed previously (Lötsch and Ultsch, 2016b), a limitation of the present implementation of this computational approach to drug repurposing consist of the unsigned inclusion of drug vs. target interactions, i.e., without distinction of agonistic from antagonistic actions. Therefore, topical experts' knowledge is required to amend this limitation. Of further note, the present computational approach at drug repurposing based on computational analysis of gene functions and similarity measures in the high-dimensional data space still depends on the accuracy and completeness of the information in the queried databases. This makes it vulnerable to research bias and erroneous information in the databases, however, the increasing trend toward “big data” supports the expectation of a continuously broadening knowledge base.

The present computational analysis of the genes involved in human hereditary insensitivity to pain in a drug repurposing context extends previous applications of the recently proposed concept of “process pharmacology” (Lötsch and Ultsch, 2016b) by (i) the inclusion of a topically selected set of genes in addition to a previously used set of genes derived from microarray analysis, and (ii) the introduction of the concept of a “virtual drug,” i.e., a vector of biological processes representing the key functions of a gene set that can be entered as a further drug into the drug vs. biological process matrix, thereby facilitating the search for repurposing candidates by using similarity measures in the high dimensional vector space. As described previously (Lötsch and Ultsch, 2016a,b), in the present approach diseases are regarded as resulting from the activity of pathophysiological processes captured in the GO database via the category “biological processes,” i.e., series of events or molecular functions with a defined beginning and end (Ashburner et al., 2000). Such a focus on disease relevant biological processes (Lötsch and Ultsch, 2016b) has been shown to provide a robust basis for drug classification agreeing with classical approaches (Lötsch and Ultsch, 2016a) and may provide a phenotypic approach to drug discovery and repurposing.

The present ORAs were performed with different p-value thresholds. The conservative p-value threshold of t_p = 1·10⁻¹⁵ used for the analysis of the drug targets was heuristically chosen to accommodate the intention to only include highly relevant terms and to obtain a set size of 500–1,000 terms that had proven suitable in previous similar analyses (Lötsch and Ultsch, 2016a,b). By contrast, the p-value threshold of 0.05 chosen for the ORA of the n = 20 genes associated with human insensitivity to pain was chosen for practical reasons. That is, it was the strictest criterion providing the generally accepted significance level of p = 0.05 and a correction for multiple testing, a stricter criterion or the use of more conservative α-correction according to Bonferroni (1936) led to a nearly empty ORA result for the present set of n = 20 genes involved inhuman insensitivity to pain and could therefore not applied. However, a systematic test of the optimum p-value threshold for the present computational drug repurposing approach remains an active research topic to be addressed in future work.

By placing the modulation of biological processes into the focus with molecular targets merely serving as intermediates, the present concept may exceed the particular molecular mechanism by which a drug acts making it particularly suitable for drug repurposing by eliminating the restriction to a specific molecular target. In this respect, the present concept complements non-redundantly alternative implementations of computational science in drug research and development. For example, the analysis of molecular interaction networks (e.g., protein-protein interaction networks or gene-gene co-expression networks) may be employed to study the possible consequences of target removals, considering a target as effective when its removal modifies the network in an essential way (Penrod et al., 2011). Other approaches proposed to use supervised machine learning, i.e., support vector machines, to analyze druggable protein-protein interaction on the basis of the number of shared GO terms indicating similarities in biological function between two interacting proteins (Sugaya et al., 2007; Sugaya and Ikeda, 2009). Further approaches include pathway based analysis where pathway annotations, such as provided by the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database are “translated into a two-dimensional statistical test problem that involves Wilcoxon's signed rank sum test in order to compute a Z-score for each pathway that quantifies the degree of alteration across the different experimental conditions” (Herwig and Lehrach, 2006). These approaches have also shown to provide clinically useful approaches to drug development such as the identification of anticancer drug combinations (Azmi et al., 2010), which complements a similar success of the present method (Lötsch and Ultsch, 2016b). This emphasizes the increasing utility of a variety of computational approaches to gene functions in drug research, for which the present analysis adds further support.

Most contemporary machine learning techniques for the analysis of genomic data are of the supervised learning type (for a recent review, see Libbrecht and Noble, 2015). These methods aim at the diagnosis (classification) of cases. By contrast, in the present work unsupervised methods were used, which mainly aim at structure detection in high dimensional (genetic) data. Typical methods for structure detection in high dimensional data are (i) nonlinear projections such as multidimensional scaling (MDS) (Tzeng et al., 2008), t-SNE (van der Maaten and Hinton, 2008; Bushati et al., 2011) or curvilinear component analysis (Alanis-Lobato et al., 2015) and (ii) clustering methods summarized in (Pirim et al., 2012). The ESOM/Umatrix method used here can be regarded as a combination of a disentangling and neighborhood preserving projection method combined with a clustering algorithm. This method has been demonstrated to be superior to other approaches in the sense, that complex and intertwined clusters can be identified, however, no spurious structures are artificially introduced by the clustering method itself (Ultsch and Lötsch, 2017) for which a superiority to many other projection methods has been shown (Tasdemir and Merényi, 2012).

Genes causally involved in human insensitivity to pain provide a unique source of studying the pathophysiology of pain and the development of novel analgesic drugs. In keeping with the contemporary trend toward “big data” analysis in biomedical research, an integrated computational analysis was performed to study the set of genes for emergent, principal pathophysiological processes that characterize insensitivity to pain. As a result, a particular importance for pain perception was observed for processes related to structural changes in the nervous system and to ceramide and sphingosine signaling pathways, which is compatible with suggestions of using these pathways as therapeutic target in pain. Using the biological processes characterizing hereditary insensitivity to pain for drug repurposing, a clear cluster of n = 22 substances emerged that comprised several drugs for which implications in pain have been shown occasionally in preclinical experiments. Thus, the present concept provides biologically plausible results and seems to be suitable for drug discovery by identifying a narrow choice of repurposing candidates, demonstrating that contemporary machine-learned methods offer innovative approaches to knowledge discovery from previous evidence.

Wrote the paper, conception or design of the work: JL. Analyzed the data: JL and CL. Data collection: DK and JL. Provided advice for data analysis: AU.