facebookresearch/esmArchived
Esm
This project is a collection of pretrained transformer protein language models designed for sequence embedding, structure prediction, variant effect estimation, and generative inverse folding. It provides a framework for transforming amino acid sequences into high-dimensional numerical vectors and predicting biological properties through these embeddings.
The system includes a generative model for inverse folding that designs amino acid sequences to fit a specific target protein backbone structure. It also features tools for protein variant analysis that use zero-shot prediction to estimate the functional impact of mutations on protein stability and activity.
Additional capabilities cover the prediction of 3D atomic coordinates from single sequences, the generation of protein contact maps via attention mechanisms, and the retrieval of structural data and embeddings from large metagenomic datasets. The project also includes utilities for protein sequence design, fitness scoring, and memory management for processing long sequences on limited hardware.
Features
- Protein Sequence Embeddings - Transforms amino acid sequences into high-dimensional numerical representations using pretrained transformer models.
- Transformer Language Models - Learns protein sequence patterns by processing amino acid strings through self-attention mechanisms to generate high-dimensional embeddings.
- Inverse Folding Sequence Sampling - Generates amino acid sequences conditioned on a target 3D structural backbone using generative sampling.
- Protein Inverse Folding Models - Designs new amino acid sequences that fit a specific target protein backbone structure using inverse folding.
- Per-Residue Representations - Generates high-dimensional numerical embeddings for individual amino acids to predict specific protein properties.
- Protein Structure Prediction - Predicts the three dimensional atomic coordinates of proteins from a single amino acid sequence using transformer models.
- Protein Variant Effect Predictors - Estimates the functional impact of mutations using zero-shot inference by comparing sequence likelihoods.
- Generative Protein Design - Implements generative models that create amino acid sequences based on learned evolutionary patterns.
- Generative Sequence Design - Uses generative models to optimize amino acid sequences based on specific structural constraints.
- Sequence Sampling - Samples candidate amino acid sequences from probability distributions conditioned on structural backbones.
- Protein Folding Coordinates - Predicts 3D atomic coordinates by mapping sequence embeddings to geometric representations through a structure-prediction transformer.
- Sequence Fitness Scoring - Calculates the conditional log-likelihood of sequences given a structure to evaluate biological fitness.
- Metagenomic Protein Data Retrieval - Provides bulk access to predicted structures, databases, and embeddings for millions of metagenomic proteins.
- Zero-Shot Mutation Likelihoods - Estimates the functional impact of mutations by comparing the model likelihood of wild-type and mutant protein sequences.
- Contact Map Predictions - Infers spatial proximity between residues by extracting interaction strengths from the internal attention heads of the transformer.
- Metagenomic Protein Analysis - Processes and retrieves structural data and embeddings for millions of proteins from large metagenomic datasets.
- Protein Contact Map Predictions - Identifies residues likely to be in physical contact by analyzing self-attention maps within a language model.
- Geometric Backbone Encoders - Transforms three-dimensional protein coordinates into numerical vectors using a geometric encoder for inverse folding tasks.
- Protein Variant Analysis - Estimates the functional impact of mutations on protein stability and activity using zero-shot prediction.
- Transformer and Language Models - Evolutionary-scale language models for structure prediction and design.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
