Authors:
Domonkos Pogány
and
Péter Antal
Affiliation:
Department of Measurement and Information Systems, Budapest University of Technology and Economics, Budapest, Hungary
Keyword(s):
Drug-Target Interaction Prediction, Drug Repositioning, Representation Learning, Metric Learning, Joint Embedding Models, Negative Sampling.
Abstract:
The challenges of modern drug discovery motivate the use of machine learning-based methods, such as predicting drug-target interactions or novel indications for already approved drugs to speed up the early discovery or repositioning process. Publication bias has resulted in a shortage of known negative data points in large-scale repositioning datasets. However, training a good predictor requires both positive and negative samples. The problem of negative sampling has also recently been addressed in subfields of machine learning with utmost importance, namely in representation and metric learning. Although these novel negative sampling approaches proved to be efficient solutions for learning from imbalanced data sets, they have not yet been used in repositioning in such a way that the learned similarities give the predicted interactions. In this paper, we adapt representation learning-inspired methods in pairwise drug-target/drug-disease predictors and propose a modification to one of
the loss functions to better manage the uncertainty of negative samples. We evaluate the methods using benchmark drug discovery and repositioning data sets. Results indicate that interaction prediction with metric learning is superior to previous approaches in highly imbalanced scenarios, such as drug repositioning.
(More)