Cross-protein transfer learning substantially improves disease variant prediction.

Our results demonstrate the utility of mutational scanning data for learning properties of variants that transfer to unseen proteins.