DT-Web: a web-based application for drug-target interaction and drug combination prediction through domain-tuned network-based inference

In the last decades, pharmacology and therapeutic fields have encountered several development shortcomings due to prohibitive clinical costs related to novel drug discovery. The development of a new molecular entity is typically based on the process of discovering a new drug by modifying an existing one [1]. Recent trends in the pharmacogenomics area are going to exploit data mining and bioinformatics approaches, such as those based on drugs similarity, in connection with biological networks analysis. In Phatak et al. [2], a computational method for drug repositioning differs from previous similarity-based approaches since it combines chemical drug structures and drug target information computing similarity of drug target profiles via a bipartite-graph based approach. Bipartite graphs can also be used to provide a drug-target network for assessing the similarity between different disease inhibitors based on the connection to other compounds and targets. In this case, classical structure-based drug design and chemical-genomic similarity methods are combined with molecular graph theories. Another example of network based drug study is shown in Iorio et al. [3], where drug mode of action and drug repositioning are assessed using available gene expression profiles [4] to build a drug-drug network. The Connectivity Map databases [4] is a comprehensive reference catalog of genome-wide expression data from cultured human cells perturbed with many chemicals and genetic reagents, to connect human diseases with the genes that may cause them and drugs that can treat them. Literature-mining studies show that a large majority of new drugs bind to targets in some way related to a previously existing one [5–7]. A taxonomy and a comprehensive survey of new drugs discovery is contained in Csermely et al. [8].

An important role in the development of new drugs is given by the methods predicting drug-target interactions (DTI). Traditionally researchers have focused their attention on the development of drugs acting only on a specific protein family. On the other hand the more recent poly-pharmacology approach [9] combines actions of drugs and related multiple targets. The knowledge of these targets is fundamental to find out alternative applications (drug repositioning) as well as to identify side effects [10, 11]. Despite all such efforts, today many interactions are still unknown and in situ experiments are too costly and time-consuming to be used as the sole strategy.

Various techniques to solve such a problem have been proposed [6, 12–16]. In particular, the naive application [17] of the recommendation algorithm developed in Zhou et al. [18] has shown extremely promising results. In Alaimo et al. [19] an extension of the above method with the addition of domain-tuned knowledge led to the definition of the DT-Hybrid algorithm, combining bipartite networks projection and network resources transfer. The process is driven by 2D drug structural similarity, and target sequential similarity. The basic idea is that structurally similar drugs tend to have analogous behavior in similar proteins.

Another DTI prediction software is STITCH 4.0 [20–23]. It is a database of multi-species protein-chemical interactions, which, not only integrates many experimentally validated and manually curated data sources, but also provides a technique to compute predictions based on text-mining in research articles and analysis of chemical structures. The database combines this information to get an overall score that indicates the degree of confidence of the interaction. However, it can only be applied to the data in their internal database, and users can only search within such data or download the entire list of predictions, rather than providing their own compounds and/or proteins. Moreover, STITCH 4.0 does not compute drugs targeting combinations.

In addition, nowadays, research has been focused on new strategies exploiting metabolic and signaling networks since they show properties for the treatment of infection and cancer diseases. Graph searching algorithms help to exploit such data [24–29]. A metabolic pathway describes a series of chemical reactions occurring within a cell, in which enzymes catalyze the modification of an initial metabolite to form another product. A pathway may include biochemical reactions, complex assembly, transport events, catalysis events and physical interactions involving metabolites, namely proteins, DNA, RNA, small molecules and complexes. The set of metabolic pathways forms the metabolic network of a cell, which is part of its complex system, together with signaling pathways, representing the response to external signals, and other physical processes. Enzymes catalyzing a single reaction are usually essential [30], however related metabolites are frequently involved in several different pathways. Consequently, analyzing metabolic networks helps to identify interesting entities, such as: metabolites used as disease biomarkers [31, 32]; or choke points (reactions that consume or produce a certain metabolite) [33, 34], whose inhibition may cause lethal deficiencies or toxic accumulation of metabolites. Motivated by the success of manual pathway analysis, a pathway-based approach was proposed in Li et al. [35–37] where a causal drug-to-disease network is built by taking advantage of expert-curated biological knowledge including drug targets, pathways and disease downstream genes. Since, many human diseases cause metabolic deficiencies, the enzymes involved in such essential reactions can be considered promising drug targets [38–41]. Due to the complexity of diseases, the development of multi-target drugs [42, 43] or drug combinations may be considered crucial. The choke point analysis, the comparison of metabolic networks of pathogenic and non-pathogenic strains and the load point analysis, may improve effective combinations prediction. These commonly consist in the identification of nodes having a high ratio of incident k-shortest paths [44, 45]. On the other hand, it has been shown that co-targeting of crucial pathway points [46–49] is efficient against drug resistances both in anti-infective [50] and anti-cancer [51, 52] strategies. Two relevant examples are RAS and Survivin associated diseases.

The RAS oncogenic mutations occur in human cancer quite frequently [53]. Therapeutic approaches based on the inhibition of RAS-mediated signalling have resulted ineffectively mainly due to the drug's toxicity. On the other hand, mutations of RAS, cause high levels of drug resistance. To overcome such shortcomings, Nussinov et al. [54] suggests an approach, inspired by Holzapfel et al. [55], based on drug combinations on parallel pathways.

The Survivin protein is linked to multiple pathways of cellular homeostasis [56]. These seem to confer to cancer cells a great plasticity, proliferation capability and resistance to death. Moreover, a large number of molecules, regulators, transcriptional networks and modifiers are directly or indirectly involved in Survivin related networks. Therefore, approaches considering Survivin in isolation and based on single-protein inhibitors [57] appear to be unsatisfactory. The development of drugs based on pathways inhibitors [58] exploiting the connectivity maps of Survivin to multiple signaling circuits [4] may result in a more effective approach.

An important application of biological network analysis concerns drug repositioning, a central task in new treatments discovery [13, 14, 59, 60]. In Jahchan et al. [61] and Smith et al. [62], a pathway enrichment analysis in combination with gene expression profiles to explore new applications of existing drugs has been exploited. In Pan et al. [63], sixteen FDA-approved drugs were studied in order to understand their clinical functions through pathway analysis. Targets interacting or affected by the investigated drugs were extracted by mining public databases. Pathways having a co-occurrence of such targets were ranked through a p-value. Although directly and indirectly correlated drug targets were identified, important limitations arose due to relevant pathway retrieval and unknown targets finding.

In this paper, we propose DT-Web a software resource accessible via web at [64], using known sources, such as DrugBank [65–67] and PathwayCommons [68], in connection with DT-Hybrid [19] recommendation algorithm. It implements two analysis tools relying on networks: the first one uses a bipartite drug-target interaction network to predict novel high confidence DTIs, while the second one uses a multi-drug, multi-target, multi-pathway approach to guide the early stages of experimental analysis in drug combinations studies. In fact, even if researches can relay on public-domain databases of drug combination references, such as Liu et al. [69], there is still a lack of systematic computational approaches, and often several possible drug combinations are disclaimed by expert knowledge and verified via clinical trials. Motivated by the success of network-based approaches, our multi-purpose pathway analysis aims to provide a limited set of candidate drugs both for drug repositioning and combination, that can be directly evaluated by the experts or combined with other methods.

The core of DT-Web has been developed in R and Java, while its front-end consists of a web interface developed in PHP with the backing of a MySQL database to store persistent information. We collected only drugs, targets and pathways for the Homo Sapiens species, and designed our methodology accordingly. DT-Web offers two major functionalities: (i) Drug-Target Interaction Prediction: the integration of Drug-Bank with DT-Hybrid to provide a comprehensive database of drugs, and their interactions with proteins (targets, enzymes, transporters, or carriers), either experimentally validated or predicted by DTI network inference; (ii) Multi-Purpose Pathway Analysis: the integration of DrugBank, DT-Hybrid, and Pathway-Commons to aid the experimental phase in drug combination studies by searching for drugs simultaneously acting on multiple targets.

DT-Web is an effective and user-friendly system that provides a web interface to visualize information on drugs and predicted targets, and to simplify the earlier experimental phases of drug combination studies by applying a multi-drug, multi-target, multi-pathway approach. All this is provided along with a constantly updated database containing the main and most reliable information from DrugBank. In addition, with the ability to upload their own data, our software does not limit the user to use only pre-computed information.

We, also, compared the prediction scores generated by STITCH 4.0 and DT-Web using a set of validated interactions taken from DrugBank. The analysis showed that the knowledge (i.e. experiments, structures, text-mining) given as input to STITCH 4.0 generate a distribution of scores which resulted bimodal and piled either on a low or a up range. On the other hand, DT-Web, which predictions scores are based on structure similarities and network inferences showed a Gaussian behavior. This difference is justifiable by both the kind of input knowledge and inference methods the two systems use.

To use the integrated database, user provides one or more keywords (see our website [64] for a more detailed list of searchable information). The results will be returned on a page, which, as described in Figure 1, contains a list of the main information extracted from DrugBank, experimentally verified targets, enzymes, transporters, and carriers (along with papers providing such experimental evidence), and the predictions calculated by DT-Hybrid.

To take advantage of the DTI prediction pipeline on his own data, the user can choose the "Drug-Target Interaction" job in the "Prediction Tools" menu of our web interface. He will be able to upload his own data (format information and examples are available on our website [64]), which, after checking to ensure that the format is correct, will be analyzed with our pipeline. When data is submitted, DT-Web automatically generates a temporary page (see Figure 2) that allows to check the job status and, when ready, view results. To avoid long waits, an e-mail address, where a notification will be sent as soon as the results are available, can be provided. When viewing predictions, the user can apply filters based either on p-value or correlation by choosing an appropriate threshold. He will also be able to download a file containing all the results in text format, or visualize a graphical representation of the resulting DTI network (see Figure 3).

To take advantage of our multi-purpose pipeline, the user can choose the "Drug Combination" job in the "Prediction Tools" menu. There, he can run the enriched analysis by indicating a set of indirectly influenced genes. Once the analysis starts, DT-Web automatically generates a temporary page (see Figure 4) that allows to check the job status and, if ready, view results. Once the analysis is performed, the results page will show the list of reachable targets and their interacting drugs. For each drug-target pair the DT-Hybrid prediction score is shown if the interaction does not come from a validated source. A list of filters can be applied to the results. The list of found drugs can be filtered by selecting only some of them as well as the drug-target interaction type, validated or predicted. Moreover, it is also possible to apply a filter on the set of found targets. For each target and drug, the user can visualize an interactive sub-network centered on the given entity. Additional information is available by clicking on the visualized elements, for example, the list of residing pathways can be visualized by clicking on each edge.

The prediction of novel drug-target interactions is a fundamental process in order to reduce the costly and time-consuming phases of drug discovery and design. As a matter of fact, knowing the possible unknown effects on the proteome of a drug can be extremely useful in understanding its true potential or predicting side effects.

Other studies, such as drug repositioning, drug combinations or substitutions, help to eliminate the need to develop new drugs. Drug repositioning studies exploit existing drug for new purposes that go beyond the original ones, while drug combination studies try to modify or intensify the overall effect of two or more drugs by administering them together. Similarly, drug substitutions studies try to replace drugs having significant side effects with others aiming to reduce them. For this purpose, a tool that guides the early stages of the experimental process can significantly reduce the time and costs associated therewith.

This is the context of DT-Web. Its main goal is to provide a simple system that allows a user to quickly browse predictions of probable novel DTI, to produce new ones from their own data, or to simplify the experimental studies described above. All this by using a database obtained combining a new valuable resource DT-Hybrid with data extracted from Drug-Bank and PathwayCommons. Finally, all the results can be downloaded in text format or viewed on-line through the help of interactive graphical interfaces that simplify their understanding.