ESM

Model Introduction

ESM (Evolutionary Scale Modeling) is a family of protein language models released by Meta AI / FAIR. It can be used for protein representation extraction, structure prediction, variant effect scoring, and fixed-backbone sequence design.

Paper: Evolutionary-scale prediction of atomic-level protein structure with a language model
https://www.science.org/doi/10.1126/science.ade2574

Model Description

This model package provides PyTorch inference support for ESM-1, ESM-2, MSA Transformer, ESMFold, ESM-1v, and ESM-IF1, together with adaptations for running on DCUs. Sample data is distributed with the Hugging Face model repository OneScience-Sugon/ESM.

Use Cases

Scenario	Description
Protein representation extraction	Takes a FASTA file as input and outputs per-token, mean-pooled, BOS, or contact representations
Protein structure prediction	Takes one or more amino acid sequences as input and outputs corresponding PDB structure files
Variant effect scoring	Takes a wild-type sequence and a DMS mutation table as input and outputs mutation effect scores
Fixed-backbone sequence design	Takes a PDB / CIF structure and chain ID as input and samples candidate sequences that satisfy the backbone constraints
Structure-conditioned sequence scoring	Takes a structure and candidate sequences as input and computes their conditional log-likelihoods
Hugging Face / OneCode execution	After downloading the model project, quickly verifies that the scripts run correctly in a life-sciences runtime environment

Usage Guide

1. OneCode Usage

Try one-click AI4S development in the OneCode online environment:

Try one-click AI4S development

2. Manual Installation and Usage

Hardware Requirements

GPU or DCU is recommended.
A CPU can be used for import checks and lightweight configuration tests; full training and inference will be slow.
DCU users must install DTK in advance. DTK 25.04.2 or later is recommended, or a OneScience-recommended version matching the current cluster.

Environment Check

NVIDIA GPU:

nvidia-smi

Hygon DCU:

hy-smi

Download the Model Package

hf download --model OneScience-Sugon/ESM --local-dir ./ESM
cd ESM

This model package includes a small set of sample data that can be used directly to validate the default workflow.

Install the Runtime Environment

DCU Environment

# Activate DTK and CONDA first
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# uv installation supported
pip install onescience[bio-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/  --trusted-host mirrors.onescience.ai

# If required libraries cannot be found, activate the CUDA compatibility environment as follows:
source ${ROCM_PATH}/cuda/env.sh
export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"
export LD_LIBRARY_PATH="$CONDA_PREFIX/lib/python3.11/site-packages/fastpt/torch/lib:$LD_LIBRARY_PATH"

After installation, return to the model package directory:

cd ./ESM

Training and Inference Data Overview

The FASTA, PDB / CIF, and DMS files used in the ESM examples are distributed with the Hugging Face model repository OneScience-Sugon/ESM. After downloading the complete model package, the files are available under data/. This model package does not include a training entry point; the data is intended for example inference and workflow validation. You can also download only the data directory:

hf download --model OneScience-Sugon/ESM ESM/data --local-dir ./data

Model Weights

The repository includes multiple ESM model checkpoints under weight/; select the appropriate checkpoint for inference.

Preparing Weights

Place the required ESM weights in the following directory:

weight/
  checkpoints/
    esm2_t6_8M_UR50D.pt
    esmfold_3B_v1.pt
    esm1v_t33_650M_UR90S_1.pt
    esm_if1_gvp4_t16_142M_UR50.pt
    ...

When using a shared runtime environment, you can specify the weights location through an environment variable:

export ESM_WEIGHT_DIR=/path/to/esm/weight

The default example uses:

weight/checkpoints/esm2_t6_8M_UR50D.pt

The ESMFold, ESM-1v, and ESM-IF1 examples require their respective weights to be available.

Default Example

bash scripts/infer.sh

The default example reads data/fasta/few_proteins.fasta, extracts protein representations using esm2_t6_8M_UR50D.pt, and saves the results to outputs/embeddings/.

Sequence Representation Extraction

python scripts/extract.py \
  weight/checkpoints/esm2_t6_8M_UR50D.pt \
  data/fasta/few_proteins.fasta \
  outputs/embeddings \
  --include mean per_tok \
  --repr_layers 6

ESMFold Structure Prediction

python scripts/fold.py \
  -i data/fasta/few_proteins.fasta \
  -o outputs/pdb \
  --model-dir weight \
  --cpu-only

The output directory will contain one or more .pdb files. For production GPU / DCU inference, remove --cpu-only and configure --chunk-size or --max-tokens-per-batch according to the available accelerator memory.

You can also explicitly enable ESMFold via the default script:

RUN_ESMFOLD=1 bash scripts/infer.sh

Inverse Folding — Sequence Sampling

python scripts/inverse_folding/sample_sequences.py \
  data/inverse_folding/5YH2.pdb \
  --chain A \
  --outpath outputs/sampled_seqs.fasta \
  --num-samples 1 \
  --nogpu

Inverse Folding — Sequence Scoring

python scripts/inverse_folding/score_log_likelihoods.py \
  data/inverse_folding/5YH2.pdb \
  data/inverse_folding/5YH2_mutated_seqs.fasta \
  --chain A \
  --outpath outputs/sequence_scores.csv \
  --nogpu

Variant Effect Prediction

Variant effect prediction requires a wild-type sequence that is consistent with the mutation annotations in the DMS table:

python scripts/variant_prediction/predict.py \
  --model-location esm1v_t33_650M_UR90S_1 \
  --sequence "${ESM_VARIANT_SEQUENCE}" \
  --dms-input data/variant_prediction/BLAT_ECOLX_Ranganathan2015.csv \
  --mutation-col mutant \
  --dms-output outputs/variant_prediction.csv \
  --offset-idx 24 \
  --scoring-strategy wt-marginals

Data Format

Sample data is stored under data/ by default:

data/
  fasta/
    few_proteins.fasta
    some_proteins.fasta
  inverse_folding/
    5YH2.pdb
    5YH2.cif
    5YH2_mutated_seqs.fasta
    example.json
  variant_prediction/
    BLAT_ECOLX_Ranganathan2015.csv
    rho_pp.csv
    aggregated_rho.csv
    aggregated_rho_round3.csv

In this structure:

FASTA files are used for sequence representation extraction and structure prediction.
PDB / CIF files are used for inverse folding sampling and structure-conditioned sequence scoring.
Variant effect prediction CSV files must include a mutation column. The default column name is mutant, and mutations use notation such as A123B.
For custom DMS data, the wild-type amino acid at each mutated position in the sequence provided via --sequence must match the corresponding mutation annotation.

Verification

Static import check:

python scripts/check_import_boundaries.py

Syntax check:

python -B -c "import ast, pathlib; [ast.parse(p.read_text(encoding='utf-8'), filename=str(p)) for root in ['model', 'scripts', 'tests'] for p in pathlib.Path(root).rglob('*.py')]"

OneScience Official Information

Platform	OneScience Main Repository	Skills Repository
Gitee	https://gitee.com/onescience-ai/onescience	https://gitee.com/onescience-ai/oneskills
GitHub	https://github.com/onescience-ai/OneScience	https://github.com/onescience-ai/oneskills

Citation & License

This repository is adapted from the open-source ESM model to support DCUs.
The ESM source code is licensed under the MIT License; see LICENSE. For the usage terms governing model weights and data, refer to the documentation provided by the respective publishers.
For scientific use, please cite the corresponding original ESM paper for each submodel used. For ESM-2 / ESMFold, cite: Evolutionary-scale prediction of atomic-level protein structure with a language model.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support