Designing simplified models for protein simulation has been a significant challenge for several decades. Using a diverse set of test proteins, and a deep-learning architecture, we have now developed a simple and chemically transferable force field for efficient simulation of protein sequences.