This article provides a comprehensive review of significant advancements in the practical application of large language model (LLM) algorithms to contemporary problems in structural bioinformatics. The discussion focuses on several demonstrated successes of LLM implementations, including their use in predicting antigen surface epitopes, assessing the antigen-binding capabilities of specific CDRH3 fragments, and forecasting antibody cross-reactivity patterns. Particular attention is given to concrete examples where LLMs have been successfully employed for identifying hemagglutinin-binding antibodies against influenza virus, predicting the effects of point mutations, and improving the accuracy of protein sequence alignments. The analysis further examines critical limitations inherent in current LLM approaches, with specific emphasis on challenges related to model weight interpretability, constraints imposed by training dataset characteristics, and the substantial computational resources required for effective model training.
Gabibov et al. (Wed,) studied this question.