 The compositional patabation autoencoder, CPA, is a new method for predicting the effects of drug or genetic combinations on individual cells. It combines the interpretability of linear models with the flexibility of deep learning algorithms to accurately predict the responses of single cells to different doses, cell types, time points, and species. The CPA was validated against newly generated single cell drug combination data, demonstrating its ability to accurately predict unseen drug combinations. Additionally, it was shown to be able to impute missing combinations from a genetic screen, providing a powerful tool for hypothesis generation. This article was authored by Mohammed Lotfolahi, Anna Klimovskaya-Susmilj, Carlo Dodono, and others.