Guiding generative AI

Abstract

Lab-based protein evolution is an approach that improves enzyme activity or development of protein-based binders, but it is very labor intensive. By contrast, computational approaches (potentially) offer rapid results without laborious lab work; however, the designs proposed by computational workflows have often not provided the required improvement in activity. More recently, generative machine learning methods have shown progress in proposing static protein scaffolds, but expanding these approaches to design in enzymatic reactivity or other protein functions has been problematic, and optimizing an initial model typically still requires lab-based directed evolution. Now, Jiang et al. have developed a computational method called EVOLVEpro to help guide experimental directed evolution efforts.

EVOLVEpro combines a protein language model with an active learning layer. A protein language model is a machine learning algorithm that analyzes and learns patterns from datasets of protein sequences. The active learning layer interprets the protein language model using an iterative process to decipher the relationship between sequence and experimentally determined activities. Based on a random forest design, the active learning layer can simultaneously optimize multiple protein properties during iterative rounds of experimental testing using as few as 10 data points per round.