Understanding the structural determinants of a protein’s biochemical properties, such as activity and stability, is a major challenge in biology and medicine. Comparing computer simulations of protein variants with different biochemical properties is an increasingly powerful means to drive progress. However, success often hinges on dimensionality reduction algorithms for simplifying the complex ensemble of structures each variant adopts. Unfortunately, common algorithms rely on potentially misleading assumptions about what structural features are important, such as emphasizing larger geometric changes over smaller ones. Here we present DiffNets, self-supervised autoencoders that avoid such assumptions, and automatically identify the relevant features, by requiring that the low-dimensional representations they learn are sufficient to predict the biochemical differences between protein variants. For example, DiffNets automatically identify subtle structural signatures that predict the relative stabilities of β-lactamase variants and duty ratios of myosin isoforms. DiffNets should also be applicable to understanding other perturbations, such as ligand binding.