Representational learning on biological data. A study on polypharmacy side-effects and graph embeddings

Summary

In recent years graphs, graph neural networks and graph embedding techniques are getting more attention in the area of machine learning in general, with biological applications being a major drive. Using Decagon, a graph neural network that predicts polypharmacy side-effects, as our starting point, we ιmplement a number of baseline models in order to identify the aspects that play the bigger part in predicting side-effects among pairs of drugs. Later, we focus on a subset of the initial dataset containing only the rarest side-effects and experiment with well known models from the graph embeddings area. We examine whether a normalization of the feature vectors in a tf-idf fashion helps a message passing network improve its performance. Finally, we use AnyBURL, a rule based model, to identify patterns in our data.