ProteinCLIP is a method that enhances protein language models (pLMs) by integrating natural language descriptions of protein functions. It uses contrastive learning to align amino acid sequences with their corresponding text descriptions, enabling the model to generate function-centric embeddings. These embeddings allow for state-of-the-art performance in various protein-related tasks, including predicting protein-protein interactions, detecting homologous proteins, and identifying functional changes due to mutations. ProteinCLIP demonstrates the effectiveness of multi-modal learning in biological contexts, helping to isolate key signals from large models and improve their utility. The method leverages pre-trained models for both proteins and natural language, without additional fine-tuning, and uses simple MLPs to project embeddings into a shared space. Training data comes from the UniProt database, and the model is evaluated on multiple tasks, showing improved performance in capturing functional changes and predicting interactions. ProteinCLIP also performs well in homology detection, outperforming other methods in retrieving similar proteins. Despite its success, ProteinCLIP has limitations, such as not improving tasks unrelated to protein function. The method is implemented using PyTorch and PyTorch Lightning, with all training data and models publicly available for reproducibility. Overall, ProteinCLIP provides a flexible and effective approach to enhance biological language models through multi-modal learning.ProteinCLIP is a method that enhances protein language models (pLMs) by integrating natural language descriptions of protein functions. It uses contrastive learning to align amino acid sequences with their corresponding text descriptions, enabling the model to generate function-centric embeddings. These embeddings allow for state-of-the-art performance in various protein-related tasks, including predicting protein-protein interactions, detecting homologous proteins, and identifying functional changes due to mutations. ProteinCLIP demonstrates the effectiveness of multi-modal learning in biological contexts, helping to isolate key signals from large models and improve their utility. The method leverages pre-trained models for both proteins and natural language, without additional fine-tuning, and uses simple MLPs to project embeddings into a shared space. Training data comes from the UniProt database, and the model is evaluated on multiple tasks, showing improved performance in capturing functional changes and predicting interactions. ProteinCLIP also performs well in homology detection, outperforming other methods in retrieving similar proteins. Despite its success, ProteinCLIP has limitations, such as not improving tasks unrelated to protein function. The method is implemented using PyTorch and PyTorch Lightning, with all training data and models publicly available for reproducibility. Overall, ProteinCLIP provides a flexible and effective approach to enhance biological language models through multi-modal learning.