Predicting receptor-ligand pairs through kernel learning

Regulation of cellular events is initiated, often, via extracellular signaling. Extracellular signaling occurs when a circulating ligand interacts with one or more membrane-bound receptors. Identification of receptor-ligand pairs is thus an important and specific form of protein-protein interaction (PPI) prediction. While the problem of predicting PPI has been highly studied, little effort has been placed on the sub-problem of predicting receptor-ligand interactions.

A tremendous amount of research has been applied to the problem of predicting PPI. Foremost in the field has been prediction via phylogenetic profile analysis. Generally, this type of investigation studies the similarity of the phylogenetic history of a protein A and its putative protein partner B. Broadly speaking, the assortment of these types of phylogenetic studies examine the most accurate measure of similarity between the phylogenetic histories of A and B. Findings from these studies support the idea that proteins which interact have similar phylogenetic profiles, as these proteins should co-adapt as they are under the same evolutionary pressures [1, 2].

Bhardwaj et al. [3] make use of the phylogenetic information strategy while introducing expression data to predict PPI. Their findings support the idea that integrating gene expression profile and phylogenetic information increases the accuracy of predictions than phylogenetic analysis alone. The rational of using co-expression as an indicator of PPI originates from the observation that proteins which interact for the purpose of performing a similar function are likely to be co-expressed as they will need to be present at the same time to carry out their common biological activity [4, 5].

The notion of combining expression and phylogenetic information to predict PPI is clearly a step in a direction which leads us to consider a wider variety of data integration. Here we propose a framework in which other sources of data (such as domain content) can be applied to a kernel solution to the problem. The rational behind incorporation of domain content information is as follows: as certain domains are known to interact, it is self evident that this data would provide additional insight into the problem of determining receptor-ligand pairs.

One of the groups which have tackled the receptor-ligand prediction task is Gertz et al. [6]. In their work, they match members of the chemokines and tgfβ ligand families with their respective receptor families. They used distance matrices of the receptors and ligands families to measure similarity between the groups. Through a Metropolis Monte Carlo optimization algorithm, they explored and scored possible matches between the two matrices, until they reached optimal solutions. While their work was successful, they rely only on phylogenetic distance matrices, here we propose the integration of multiple data sources to help make more accurate matches.

We look into the use of creating a combined kernel classifier to carry out this learning task. While many kernel-based machine learning techniques have been applied to the PPI task [7, 8], it has hitherto never, to our knowledge, been used on the receptor-ligand problem. Kernel learning provides the means to utilize enigmatically related data (such as expression measures, domain content, etc.) and perform classification in higher dimensional space via kernel methods. In our work, we apply the least-squares support vector machines (LS-SVM) method based on the conclusions by Suykens et al. [9] which shows this implementation to be robust. As different data sources are used, separate LS-SVM kernel classifiers were built and the combined output used to provide a final result.

While the task addressed here is the predictions of successful protein ligand-receptor pairings, a related area of research is the protein-chemical interaction prediction task for which kernels have, sometimes, been applied. For example, Nagamine et al. [10] approach this task through the use of a SVM trained on vector representations of protein-chemical pairs. Building on this, Jacob et al. [11] demonstrate the utility of using hierarchical kernels to match proteins with chemical ligands in a similar learning task. This line of research was then further advanced by Bleakley et al. [12] who introduce the use of bipartite local models which use kernels to successfully predict several reported drug-target interactions.

We first describe our combined kernel classifier method to predict receptor-ligand pairings. We then present the bipartite-graph we derive from our findings and compare it to our "Golden Standard" and to results previously published by Gertz et al. [6]. Following this, we interpret the performance of our method with respect to this comparison. To conclude, we discuss the benefits and limitations of our method and possible future directions for this work.

We discover that our predictions for the tgfβ family accurately reconstruct over 76% of the supported edges (0.76 recall and 0.67 precision) of the receptor-ligand bipartite graph defined by the DLRP. In addition, the combined kernel classifier is able to relatively improved upon the Gertz et al. [6] work by a factor of approximately two as the Gertz et al. [6] work reconstructs 44% of the supported edges (0.44 recall and 0.53 precision) of the receptor-ligand bipartite graph defined by the DLRP. For this family of receptors and ligands, there exists an advantage in our approach to make predictions as we reconstruct more known edges and introduce less noise. Comparing F-measures, we see that our method improved upon that of Gertz et al. [6] significantly as our method has an F-measure of 0.71 while that of Gertz et al. [6] has a value of 0.48.

We also find that our predictions for the chemokine family accurately reconstruct over 65% of the supported edges (0.65 recall and 0.23 precision) of the receptor-ligand bipartite graph defined by the DLRP. In addition, the combined kernel classifier is able to relatively improved upon the Gertz et al. [6] work by a factor of approximately three as the Gertz et al. [6] work reconstructs 22% supported edges (0.22 recall and 0.37 precision) of the receptor-ligand bipartite graph defined by the DLRP. While the precision of the Gertz et al. [6] is higher, the recall of our method is about three fold higher. Comparing F-measures, we see that our method improved upon that of Gertz et al. [6], slightly as our method has an F-measure of 0.33 while that of Gertz et al. [6], has a value of 0.27.

Qualitatively, the performance of our method also seems to be matching the performance of Gertz et al. [6], as the novel interaction of CCR1 with SCY11 [18] reported in their work is also discovered using our method.

The overall results presented here support the notion that kernel learning presents a useful methodology for elucidating receptor-ligand pairing. Using disparate data sources, we propose a combined kernel classifier which is able to reconstruct the majority of known edges in the chemokine and tgfβ receptor-ligand bipartite graphs. In order to evaluate our pairings, we consider the bipartite graph which we construct from our results (see Figure 2). The success of the results are summarized by two performance measures; the recall and the precision of the edges predicted in the tgfβ and chemokine bipartite graphs. The relative performance of each method examined here is evaluated using the F-measure.

The combined classifier performs better using the tgfβ family of receptors and ligand than using the chemokine family of receptors and ligands. This can be attributed to two reasons. Firstly, the tgfβ has more positive examples than the chemokine family to train with. Secondly, the tgfβ family is more evolutionarily related while the chemokine family is related by function. Thus, it is more difficult to learn with data from the chemokine family as there is less evolutionarily related structure inherent to the data for the LS-SVM to learn with.

The benefits of the combined kernel classifier method are clear. Foremost in the advantages are the ability to predict multiple ligands for one receptor, which represents an imperative feature for receptor-ligand research. In addition, as the classifier output is continuous, the results can be considered to be prioritized, this presents a major convenience to researchers as often the set of candidate ligands are large and resources to validate few. The major limitation of the method rests in the need to have training examples for receptor-ligands which one is trying to predict. This is particularly true for predicting the pairing in the chemokine dataset as when we consider only ligand candidates with two or more receptor pairings, the precision performance of our method improves (0.79 recall and 0.31 precision) (see Table 1).

The advantage of using the three sub-classifiers instead of a global classifier which combines all features is two fold. The first reason is that the data sources used here are disparate and heterogeneous. A global classifier would require a mapping step which may introduce some noise. The second reason is that using separate sub-classifiers would allow for removal and addition of sub-classifiers. For example, if a better micro-array dataset becomes available in the future, it would be an advantage to be able remove the existing expression-based kernel with one derived from the new dataset without having to the retrain a global classifier. Also, if additional data sources become available, adding an additional sub-classifier based on the new data source would take less time to train than adding the data source and retraining the global classifier.

A practical advantage of using three sub-classifiers in our work became apparent when considering the performance of the individual classifiers versus that of the combined kernel classifier. More specifically, the combined kernel classifier performed equally as well or better than any of the individual classifiers. In the case of the chemokine family, the performance of all three individual classifiers was not nearly as good as the combined kernel classifier. In the case of the tgfβ family, the expression classifier performed nearly as well as the combined kernel classifier (see Additional File 1, Table S1).

The prediction of receptor-ligand pairings is a difficult and complex task. We have demonstrated that using multiple data sources provide an advantage over single data sources in solving this task. The use of multiple data sources allows us to extend our method as new data becomes available. Among our main contributions we count the ability of our method to prioritize candidate pairs, which represents an imperative feature for receptor-ligand research. As in-vivo validation is costly and time consuming, it's important that researchers have a ranking of a, potentially, large number of candidates. In addition, we provide a method which has high recall (0.76 and 0.67) and improved F-measures when compared to Gertz et al. [6] (0.71 for Iacucci et al. vs 0.48 for Gertz et al. [6] when evaluationg the tgfβ family and 0.33 for Iacucci et al. vs 0.27 for Gertz et al [6] when evaluating the chemokine family). Thus, the method is reliable in so far that it will retrieve a large portion of the true positives while not introducing too much noise. As more high throughput data becomes available, we expect to extend the current methodology to accommodate it.