ProtGPS - A protein language model developed by MIT and other institutions.

What is ProtGPS?

ProtGPS (Protein Localization Prediction Model), developed by MIT and the Whitehead Institute for Biomedical Research, is a deep learning-based protein language model used to predict the subcellular localization of proteins. ProtGPS analyzes the amino acid sequences of proteins, using an evolutionary-scale protein transformer architecture to learn complex patterns and interrelationships within the sequence. ProtGPS can predict the distribution probability of proteins in 12 different subcellular regions (such as the nucleolus and nuclear spots), successfully guiding the generation of novel protein sequences that can specifically assemble into specific subcellular regions. ProtGPS can identify pathogenic mutations that alter protein subcellular localization, providing new tools and perspectives for understanding cellular function and disease mechanisms.

Main functions of ProtGPS

• Predicting the distribution of proteins within cellsPredict the probability of protein localization in 12 different subcellular regions (such as nucleolus, nuclear spots, stress granules, etc.).
• Designing proteins with specific subcellular localization: Generate new protein sequences and specifically assemble them into target subcellular regions (such as nucleoli or nuclear spots).
• Identifying the impact of pathogenic mutations on protein localizationTo analyze the impact of mutations on protein subcellular localization and predict whether pathogenic mutations will lead to abnormal protein distribution.

The technical principle of ProtGPS

• Transformer-based sequence learningBased on the ESM2 (Evolutionary Scale Model 2) architecture, it is a Transformer-based protein language model. It simultaneously learns the relationships between all amino acids in the input sequence, capturing complex patterns and interactions in protein sequences.
• Joint training of neural network classifiers:
  • The ESM2 algorithm is jointly trained with a neural network classifier. The classifier's task is to predict the probability of protein localization in different subcellular regions based on the features extracted by the ESM2 algorithm.
  • The training dataset includes 5,480 human protein sequences, which are annotated to belong to 12 different subcellular regions, and the protein sequence characteristics of different subcellular regions are learned.
• Algorithms for generating protein sequencesTo design proteins with specific subcellular localization, the Markov Chain Monte Carlo (MCMC) algorithm is used. When generating protein sequences, the chemical spatial properties and intrinsic disorder of the protein are considered to ensure that the generated sequences conform to the distribution of natural proteins and can specifically locate to the target subcellular region.
• Pathogenic mutation analysisThis study analyzes the impact of pathogenic mutations on protein subcellular localization. It compares the localization prediction results of wild-type and mutant proteins to identify mutations that alter protein distribution. Shannon entropy and Wasserstein distance from information theory are used to quantify the impact of mutations on the uncertainty of protein localization prediction.

ProtGPS project address

• GitHub repository:https://github.com/pgmikhael/protgps
• arXiv technical paper:https://www.biorxiv.org/content

Application scenarios of ProtGPS

• Disease mechanism researchTo identify the impact of pathogenic mutations on the subcellular localization of proteins and help understand the pathogenesis of diseases.
• Protein engineering and drug designDesign proteins with specific subcellular localization for use in the development of novel protein drugs or biosensors.
• Cell biology researchPredicting the localization of proteins in different subcellular regions can aid in the study of intracellular protein function and interactions.
• Gene therapy and gene editingDesign gene editing tools that specifically target subcellular regions to improve the efficiency and specificity of gene editing.
• Protein functional annotation and database constructionThis research provides clues for protein function studies and helps build a more comprehensive protein function database.