BPS2026 – A model for the sequence space of a protein family

The evolution of proteins into specialized molecular machines is complicated by the extreme vastness of protein sequence space. The improbability that a random protein sequence takes on a particular function combined with the stepwise nature of variation during evolution gives rise to the hypothesis that protein families live within locally dense, connected networks within sequence space. Generative, sequence-based probabilistic models for proteins enable, in principle, estimation of size, shape, and connectivity of sequence space for protein families. However, for these estimates to be informative, a candidate model must not be overfit, a condition that has been difficult to satisfy because the number of model parameters far exceeds the quantity of training data. Here, we employ a data-driven approach for noise reduction in inference, yielding a model capable of globally generating diverse functional protein sequences while improving consistency with local mutational data. With this noise-minimized model, we estimate the entropy of the functional sequence space for the AroQ family of chorismate mutases and investigate the accessibility of single-step paths between sequences. We observe that a sparse pattern of epistatic interactions between positions suffices to design functional chorismate mutases, acting collectively to reduce entropy and to shape paths across sequence space.