Using attribution to decode binding mechanism in neural network models for chemistry

Source abstract

Significance Advances in machine learning have led to neural network-based methods for virtual screening, making it possible to sift through trillions of small molecules to find those that are pharmacologically important. Such methods have the potential to make chemical discoveries, but only if it is possible to untangle why models make the predictions that they do. Here we use attribution methods to investigate neural network models for small-molecule binding, and show that, while it is possible to identify pharmacophores with attribution, there is also the real possibility that a model which seems to perform perfectly instead learns spurious correlations in the underlying dataset that have nothing to do with binding chemistry. We propose an attribution-based test for determining whether a model can learn a hypothesized binding mechanism on a training set.