Evolutionary Profiles for Protein Fitness Prediction

The introduction of EvoIF marks a significant advancement in protein fitness prediction, a critical area in protein engineering. By leveraging evolutionary signals from both within-family and cross-family profiles, EvoIF provides a more efficient approach to predicting the impact of mutations. This model has demonstrated state-of-the-art performance on the ProteinGym dataset, which includes 217 mutational assays and over 2.5 million mutants. Notably, EvoIF achieves this competitive performance while utilizing only 0.15% of the training data, highlighting its efficiency compared to recent large models. The integration of masked language modeling and inverse reinforcement learning principles further enhances its predictive capabilities, making it a valuable tool for researchers in the field.