Symmetry-Embedded Siamese Neural Networks for Regression Tasks

Abstract

Machine learning (ML) has gained momentum in early drug discovery by reducing the number of expensive and time consuming real-world experiments. One class of ML that has recently shown promise are siamese neural networks (SNN) which take a pair of inputs and predicts the difference in a property, rather than the absolute property. It follows that such a model should be anti-symmetric with regards to the order of the inputs. We introduce a new design for an SNN that has this symmetry embedded directly into the architecture to increase the reliability of the model. By pairing a test compound with a series of reference compounds the model can produce an ensemble of predictions where the mean can be treated as the final prediction and the variance can be used as an uncertainty measure. The symmetry embedded SNN (SE-SNN) shows comparable performance to baseline models on five chemical datasets.