A Generalized Partial Canonical Correlation Model to Measure Contribution of Individual Drug Features Toward Side Effects Prediction

Abstract

Identification of potential drug-side effects is an open problem of importance for drug development. Side effects are related to a variety of interlinked aspects such as chemical properties of drugs, drug–target interactions, pathways involved, and many more. Existing statistical methods and machine learning models toward creating models that incorporate such features to predict adverse drug reactions. One of the challenges in these efforts is to disentangle the interdependence of features to identify the contribution of individual features toward specifying side effects. We present a partial canonical correlation analysis (PCCA) model that facilitates enumeration of contribution from individual drug features toward the prediction of a class of side effects, irrespective of interdependence on other features. The model is a combination of analytical and numerical strategies, and can be used to arrive at the most effective set of drug features starting from a range of available descriptors. For eye and nose related side effects, we demonstrate the implementation of our model for identification of best 2D chemical features that are closely linked with organ-specific adverse reactions. Despite the presence of a large number of drugs that are simultaneously associated with both the organs, the model could discern distinct drug features specifically linked to each class. With the availability of large amounts of data with an array of interdependent drug descriptors, such a model is of value in the drug discovery process as it enables in dealing with multidimensional drug features space.