Μuch progress has been noted in the prediction of drug-drug interactions with machine learning. Usually, this is performed in the context of closed world assumption, whereby all training and test data stem from the same distribution. In reality, the test data can be a mixture of in and and out of distribution (OoD) instances. However, the predictor would misclassify the out of OoD instances, leading to erroneous predictions and thus misleading clinicians.

In the current work we apply supervised contrastive learning (SCL) to detect OoD instances, and distinguish them from in- distribution instances. In particular, the OoD instances are un- known interactions, i.e. interactions that have not been observed in the training set. The method applies SCL on embeddings of drug pairs, but also on the penultimate layer, and the on the output of an neural network that was trained to predict in-distribution instances. We compare the there aforementioned approaches to a baseline k-NN based method that does not use SCL. The role of SCL is to enforce a strong separation of the in-distribution and OoD instances. We also, study the role of negative samples representing lack of any interaction among drug pairs. Moreover, there are experiments with different mixtures of in-distribution (observed) and OoD (unobserved) interactions. Results are presented on a public data extracted from Twosides data set, and by and large SCL performs better than the baseline method.