esm

ESM — Python SDK for Protein Language Models

Overview

ESM (Evolutionary Scale Modeling) is a Python SDK for protein language models provided by EvolutionaryScale. It supports protein sequence generation, structure prediction, inverse folding design, and embedding extraction, making it suitable for protein engineering, drug discovery, and computational biology research.

Use Cases

Protein Function Design and Optimization

Design protein sequences with specific functional properties. ESM3’s multimodal generation capabilities can be used to complete partial sequences, generate variants, or create novel proteins with desired structural features. It supports function-conditioned generation, enabling targeted design based on functional labels such as fluorescent proteins.

Protein Structure Prediction and Inverse Folding

Use ESM3’s structure track or ESMFold2 for all-atom structure prediction, predicting three-dimensional structures from protein sequences; or perform inverse folding to design sequences that are likely to fold into a given structure. This is suitable for research in structural biology and protein stability optimization.

Protein Embeddings and Machine Learning

Extract high-quality protein embedding vectors using ESMC models for sequence similarity calculation, function prediction, classification, and clustering analysis. Batch processing is supported, enabling protein representations to be used in downstream machine learning tasks such as functional annotation and protein property prediction.

Core Features

ESM3 Multimodal Protein Generation

Supports generation and prediction across the sequence, structure, and function tracks. ESM3-open (1.4B) can be deployed locally, while esm3-medium (7B) and esm3-large (98B) can be accessed through the Forge/Biohub API for higher quality. It provides chain-of-thought iterative optimization, temperature control, and asynchronous batch API calls.

ESMC Protein Embedding Extraction

Generates high-dimensional embedding representations of protein sequences for representation learning and downstream machine learning. It supports local use of esmc_300m and esmc_600m, as well as the esmc-6b cloud API. Multiple sequences can be encoded in batches, and normalized embeddings for similarity calculations can be returned through LogitsConfig.

ESMFold2 All-Atom Structure Prediction

Provides ESMFold2 inference through the Biohub platform for high-accuracy three-dimensional protein structure modeling. It can be integrated with ESM3 generation workflows to design sequences and then predict their structures, or to screen sequences before experiments. PDB-format output and coordinate extraction are supported.

Frequently Asked Questions

How do I install and use the ESM Python SDK?

ESM requires a Python environment with >=3.12,<3.13. Install the base version with uv pip install "esm==3.2.3". To accelerate inference on NVIDIA GPUs, it is recommended to additionally install Flash Attention: uv pip install flash-attn --no-build-isolation. To use ESM3-open locally, first accept the license agreement on Hugging Face; ESMC models can be loaded directly. For cloud API usage, simply install the esm package, then pass the model ID and ESM_API_KEY to esm.sdk.client().

What is the difference between ESM3 and ESMC, and which one should I choose?

ESM3 is a generative model that supports sequence generation, structure prediction, function annotation, and chain-of-thought optimization. It is suitable for protein design and creative tasks. ESMC is an encoder model focused on generating high-quality embeddings for representation learning, making it suitable for similarity calculation, classification, and feature extraction. Choose based on the task: use ESM3 when you need to generate or design proteins, and ESMC when you need embeddings or representations. In terms of model size, use 1.4B/300m for prototyping, 7B/600m for production, and 98B/6b for the highest accuracy.

Is ESM free to use, and what hardware is required for local deployment?

Open-source ESM models (ESM3-open 1.4B and ESMC 300m/600m) can be used locally for free, but you must accept the license on Hugging Face. Forge/Biohub APIs charge based on inference usage and require registration and an API key. For local deployment, ESMC 300m can run on a single GPU, while ESM3-open 1.4B is recommended to have at least 8 GB of VRAM; larger models require more resources. Inference speed can be optimized with Flash Attention and batch processing. Cloud APIs are suitable when you need the latest models, do not have a local GPU, or require large-scale batch processing.