Massively parallel laboratory techniques, such as deep mutational scanning, have yielded large quantitative mutational effect datasets for disparate proteins from a myriad of organisms. These datasets deeply describe the effects of mutations on protein function, and stand in stark contrast to datasets used to train other variant effect predictors. To train Envision, a stochastic gradient boosting model, we employed 21,026 variant effect measurements from nine large-scale mutagenesis datasets. Each mutation was annotated with 27 features that aimed to describe evolutionary, structural or physicochemical characteristics of mutations. Envision was used to predict mutations from the proteomes of human, mouse, frog, zebrafish, fruit fly, worm and yeast, and all precomputed predictions can be queried using our web application. We offer predictions in three contexts: individual mutation, single protein or proteome-wide downloads. In all cases, we provide not only a prediction of quantitative mutational effect that ranges from ~0 (most damaging) to ~1 (most wild-type-like), but also the descriptive features used for prediction. We are most confident in predictions where all features are available.