Introduction
Identifying missing associations between drugs and targets provides insights into polypharmacology and off-target mediated effects of chemical compounds in biological systems. Traditional machine learning algorithms like Naive Bayes, SVM and Random Forest have been successfully applied to predict drug target relations \cite{bib1,bib2,bib3,bib4}. However, using supervised machine learning methods requires training sets, and they can suffer from accuracy problems through insufficient sampling or scope of training sets. During the last years, the field of semi-supervised learning has been applied to methods based on graphs or networks. The data points are represented as vertices of a network, while the links between the vertices depend upon the labeled information. Thus, it is desirable to develop a predictive model based on using both labeled and unlabeled information. Network based models can avoid these kind of issues. The main advantages of network-based methods are:
- They use label information and as well as unlabeled data as input in the form of vectors.
- Once can use multiple classes inside the network structure.
- It uses multitude of paths to compute associations.
- Network based methods mostly use transductive learning strategy,in which the test set is unlabelled but while computation it uses the information from neighbourhood.
With the advent of the R open source statistical programming language \cite{bib5} and the gaining popularity of the RShiny package for interface development around R \cite{bib6} it has become straightforward for programmers to create and deploy web applications on windows and Linux servers. R and RShiny have already used in several biomedical applications. Table 1 shows some of these. We used R and R shiny to create Netpredictor standalone and web application respectively, which is freely available and open source. The web application framework in R allows creation of a simple intuitive user interface with dynamic filters and real-time exploratory analysis. Shiny also allows integration of additional R packages, Javascript libraries and CSS for customization. Web applications are accessible via browser or can be run locally on the user’s computer. The R package described in this paper provides utilities to compute recommendations in a bipartite network and well as unipartite network based on HeatS\cite{bib7}, Random walk with Restart (RWR), Network based inference (NBI)\cite{bib8} and combination of RWR and NBI(netcombo). In order to understand the topology of the network, the package also provides ways to compute bipartite network properties such as degree centrality, density of the network, betweenness centrality, number of sets of nodes and total number of interactions for given bipartite network. The package also performs graph partitioning such as bipartite community detection using the lp\-brim algorithm \cite{lpbrim1,lpbrim2} and visualization of communities, network permutations to compute the significance of predictions and performance of the algorithms based on user given data.