Mutational effect transfer learning (METL) is a protein language model framework that unites machine learning and biophysical modeling. Transformer-based neural networks are pretrained on biophysical simulation data to capture fundamental relationships between protein sequence, structure and energetics.