Rewriting Protein Alphabets with Language Models 1. A groundbreaking study introduces TEA (The Embedded Alphabet), a novel 20-letter alphabet derived from protein language models, enabling highly efficient large-scale protein homology searches. This method achieves sensitivity comparable to structure-based approaches without requiring structural information, bridging the gap between deep learning representations and traditional sequence bioinformatics tools. 2. The core innovation lies in using contrastive learning to convert high-dimensional protein language model embeddings into a discrete alphabet. This allows for rapid sequence comparisons using optimized tools like MMseqs2, while retaining the ability to detect remote homologs with structural similarity, offering a powerful alternative to existing methods. 3. TEA demonstrates remarkable performance in benchmarks, achieving high sensitivity in detecting homologs within the SCOPe40 database and multi-domain proteins from the AlphaFold Database. It outperforms traditional sequence alignment methods and matches the accuracy of structure-based searches, all while maintaining low computational costs. 4. An intriguing aspect is the use of entropy as a confidence metric, providing a measure of prediction reliability. This allows researchers to filter results based on confidence levels, enhancing the accuracy of functional annotations and structural predictions, especially in challenging cases like disordered regions or poorly modeled structures. 5. The study highlights TEA’s potential to improve functional annotation by connecting structural singletons in the AlphaFold Database to cluster representatives. This approach identifies over 14 million new connections with high accuracy, suggesting TEA could revolutionize clustering efforts and reveal novel functional relationships. 6. Looking ahead, TEA offers a versatile framework for creating specialized alphabets tailored to specific biological goals, such as function prediction or interface description. This could extend its applications to RNA sequences and other areas, further integrating deep learning into bioinformatics workflows. 7. The TEA model and training scripts are available on GitHub, along with converted TEA sequences for popular databases like AFDB Clusters and UniRef50, making it accessible for researchers to integrate into their workflows. 📜Paper: biorxiv.org/content/10.1101/… #ProteinLanguageModels #Bioinformatics #ContrastiveLearning #HomologySearch #StructuralBiology #ComputationalBiology