ProteinCLIP is a novel approach that enhances protein language models by applying contrastive learning between a protein's amino acid sequence and curated text describing its function. This method refines the sequence embeddings produced by pre-trained protein language models to create function-centric embeddings, which are more sensitive to functional changes and better capture the relationship between sequence and function. The authors demonstrate that ProteinCLIP significantly improves the performance of various tasks, including predicting protein mutations, identifying protein interactions, and detecting homologous proteins, even in cases with low sequence similarity. ProteinCLIP's effectiveness is attributed to its ability to harmonize sequence and function embeddings, providing a more comprehensive understanding of protein functions. The approach is flexible and can be applied to other protein language models, showcasing the potential of multi-modal learning in biological contexts. Despite its limitations, such as not improving performance in tasks unrelated to protein function, ProteinCLIP offers a valuable resource for advancing protein annotation and understanding biological systems.ProteinCLIP is a novel approach that enhances protein language models by applying contrastive learning between a protein's amino acid sequence and curated text describing its function. This method refines the sequence embeddings produced by pre-trained protein language models to create function-centric embeddings, which are more sensitive to functional changes and better capture the relationship between sequence and function. The authors demonstrate that ProteinCLIP significantly improves the performance of various tasks, including predicting protein mutations, identifying protein interactions, and detecting homologous proteins, even in cases with low sequence similarity. ProteinCLIP's effectiveness is attributed to its ability to harmonize sequence and function embeddings, providing a more comprehensive understanding of protein functions. The approach is flexible and can be applied to other protein language models, showcasing the potential of multi-modal learning in biological contexts. Despite its limitations, such as not improving performance in tasks unrelated to protein function, ProteinCLIP offers a valuable resource for advancing protein annotation and understanding biological systems.