Need help with esm?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

503 Stars 111 Forks MIT License 57 Commits 2 Opened issues


Evolutionary Scale Modeling (esm): Pretrained language models for proteins

Services available


Need anything else?

Contributors list

Evolutionary Scale Modeling

This repository contains code and pre-trained weights for Transformer protein language models from Facebook AI Research, including our state-of-the-art ESM-1b and MSA Transformer. Transformer protein language models were introduced in our paper, "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences" (Rives et al., 2019).

ESM-1b outperforms all tested single-sequence protein language models across a range of structure prediction tasks. The MSA Transformer (ESM-MSA-1) can improve performance further by leveraging MSA information.

  author={Rives, Alexander and Meier, Joshua and Sercu, Tom and Goyal, Siddharth and Lin, Zeming and Liu, Jason and Guo, Demi and Ott, Myle and Zitnick, C. Lawrence and Ma, Jerry and Fergus, Rob},
  title={Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences},
Table of contents
What's New

Main models you should use

| Shorthand |

| Dataset | Description | |-----------|-----------------------------|---------|--------------| | ESM-1b |
| UR50 | SOTA general-purpose protein language model. Can be used to predict structure, function and other protein properties directly from individual sequences. Released with Rives et al. 2019 (Dec 2020 update). | | ESM-MSA-1b|
| UR50 + MSA | MSA Transformer language model. Can be used to extract embeddings from an MSA. Enables SOTA inference of structure. Released with Rao et al. 2021 (ICML'21 version, June 2021). | | ESM-1v |
| UR90 | Language model specialized for prediction of variant effects. Enables SOTA zero-shot prediction of the functional effects of sequence variations. Same architecture as ESM-1b, but trained on UniRef90. Released with Meier et al. 2021. |

For a complete list of available models, with details and release notes, see Pre-trained Models.

Comparison to related works

Task Unsupervised contact prediction Supervised contact prediction SSP
Test set Large valid CASP13-FM CAMEO CASP13-FM CAMEO CB513
Gremlin (Potts) 39.3 16.9 24.0 40.1 47.3
UniRep 11.2 17.8 58.4
SeqVec 13.8 22.5 62.1
TAPE 11.2 5.5 6.8 12.3 15.9 58.0
ProtBert-BFD 34.1 13.5 23.9 24.7 37.0 70.0
Prot-T5-XL-BFD 35.6 16.5 25.9 25.0 40.8 71.4 ± 0.3
ESM-1 33.7 13.6 21.4 (todo) (todo) 69.2
ESM-1b 41.1 17.0 30.9 28.2 44.4 71.6 ± 0.1
ESM-1v 35.3 14.2 24.4
ESM-MSA-1b 57.4 44.8 43.5 54.6 55.8 73.4 ± 0.3

Comparison to related protein language models on structure prediction tasks.

  • All contact numbers are the top-L,LR precision metric, where long range means sequence separation of at least 24 residues
  • For unsupervised contact prediction, a sparse linear combination of the attention heads is used to directly predict protein contacts, fitted with logistic regression on 20 structures. For more details on the method, see Rao et al. 2020.
  • Supervised contact prediction all uses the same resnet (32 layers) and trRosetta training data, cf Rao et al. 2021.
  • (SSP) Secondary structure Q8 accuracy on CB513, transformer finetuned with convolution + LSTM head.
  • Direct coupling analysis methods (Gremlin, mfDCA, Psicov) and ESM-MSA-1 use the trRosetta MSAs, while other methods predict from single sequence.


Quick Start

As a prerequisite, you must have PyTorch 1.5 or later installed to use this repository.

You can use this one-liner for installation:

$ pip install fair-esm

We also support PyTorch Hub, which removes the need to clone and/or install this repository yourself:

import torch
model, alphabet = torch.hub.load("facebookresearch/esm", "esm1b_t33_650M_UR50S")

Then, you can load and use a pretrained model as follows:

import torch
import esm

Load ESM-1b model

model, alphabet = esm.pretrained.esm1b_t33_650M_UR50S() batch_converter = alphabet.get_batch_converter()

Prepare data (first 2 sequences from ESMStructuralSplitDataset superfamily / 4)


Extract per-residue representations (on CPU)

with torch.no_grad(): results = model(batch_tokens, repr_layers=[33], return_contacts=True) token_representations = results["representations"][33]

Generate per-sequence representations via averaging

NOTE: token 0 is always a beginning-of-sequence token, so the first residue is token 1.

sequence_representations = [] for i, (_, seq) in enumerate(data): sequence_representations.append(token_representations[i, 1 : len(seq) + 1].mean(0))

Look at the unsupervised self-attention map contact predictions

import matplotlib.pyplot as plt for (_, seq), attention_contacts in zip(data, results["contacts"]): plt.matshow(attention_contacts[: len(seq), : len(seq)]) plt.title(seq)

Compute embeddings in bulk from FASTA

We provide a script that efficiently extracts embeddings in bulk from a FASTA file. A cuda device is optional and will be auto-detected. The following command extracts the final-layer embedding for a FASTA file from the ESM-1b model:

$ python esm1b_t33_650M_UR50S examples/some_proteins.fasta examples/some_proteins_emb_esm1b/ \
    --repr_layers 0 32 33 --include mean per_tok


now contains one
file per FASTA sequence; use
to load them.
has flags that determine what's included in the
file: *
(default: final only) selects which layers to include embeddings from. *
specifies what embeddings to save. You can use the following: *
includes the full sequence, with an embedding per amino acid (seqlen x hiddendim). *
includes the embeddings averaged over the full sequence, per layer. *
includes the embeddings from the beginning-of-sequence token. (NOTE: Don't use with the pre-trained models - we trained without bos-token supervision)

Zero-shot variant prediction

See "./variant-prediction/" for code and pre-trained weights for the ESM-1v models described in Language models enable zero-shot prediction of the effects of mutations on protein function. (Meier et al. 2021).


Supervised variant prediction - training a classifier on the embeddings

To help you get started with using the embeddings, this jupyter notebook tutorial shows how to train a supervised variant predictor using embeddings from ESM-1. You can adopt a similar protocol to train a model for any downstream task, even with limited data. First you can obtain the embeddings for

either by downloading the precomputed embeddings as instructed in the notebook or by running the following:
# Obtain the embeddings
$ python esm1_t34_670M_UR50S examples/P62593.fasta examples/P62593_reprs/ \
    --repr_layers 34 --include mean

Then, follow the remaining instructions in the tutorial. You can also run the tutorial in a colab notebook.

Note this is somewhat outdated: use

instead, and see the newer instructions for zero-shot variant prediction, that is without any supervised training.

Unsupervised contact prediction

This jupyter notebook tutorial demonstrates contact prediction with both the ESM-1b and MSA Transformer (ESM-MSA-1) models. Contact prediction is based on a logistic regression over the model's attention maps. This methodology is based on our ICLR 2021 paper, Transformer protein language models are unsupervised structure learners. (Rao et al. 2020) The MSA Transformer (ESM-MSA-1) takes a multiple sequence alignment (MSA) as input, and uses the tied row self-attention maps in the same way. See MSA Transformer. (Rao et al. 2021).

To get unsupervised attention-based contacts, call

model(tokens, return_contacts=True)

ESMStructuralSplitDataset and self-attention contact prediction

And this jupyter notebook tutorial shows how to load and index the

, and computes the self-attention map unsupervised contact predictions using ESM-1b.

Available Models and Datasets

Pre-trained Models

| Shorthand |

| #layers | #params | Dataset | Embedding Dim | Model URL (automatically downloaded to
) | |-----------|---------------------|---------|---------|---------|---------------|-----------------------------------------------------------------------| | ESM-1v |
| 33 | 650M | UR90/S 202003 | 1280 | | | ESM-MSA-1b| `esmmsa1bt12100MUR50S` | 12 | 100M | UR50/S + MSA 201803 | 768 | | | ESM-MSA-1 |
| 12 | 100M | UR50/S + MSA 201803 | 768 | | | ESM-1b | `esm1bt33650MUR50S
| 33     | 650M    | UR50/S 2018_03 | 1280          |   |
| ESM-1     |
esm1t34670MUR50S` | 34 | 670M | UR50/S 201803 | 1280 | | | | `esm1t34670MUR50D
| 34      | 670M    | UR50/D 2018_03 | 1280          | |
|           |
esm1t34670MUR100` | 34 | 670M | UR100 201803 | 1280 | | | | `esm1t1285MUR50S
| 12      | 85M     | UR50/S 2018_03 | 768           |  |
|           |
esm1t643MUR50S` | 6 | 43M | UR50/S 201803 | 768 | |

Here is a chronological list of the released models and the paper they were introduced in:

| Shorthand | Release Notes | |-----------|---------------| | ESM-1 | Released with Rives et al. 2019 (Aug 2020 update). | | ESM-1b | Released with Rives et al. 2019 (Dec 2020 update). See Appendix B. | | ESM-MSA-1 | Released with Rao et al. 2021 (Preprint v1). | | ESM-MSA-1b | Released with Rao et al. 2021 (ICML'21 version, June 2021). | | ESM-1v | Released with Meier et al. 2021. |

ESM Structural Split Dataset

This is a five-fold cross validation dataset of protein domain structures that can be used to measure generalization of representations across different levels of structural dissimilarity. The dataset implements structural holdouts at the family, superfamily, and fold level. The SCOPe database is used to classify domains. Independently for each level of structural hold-out, the domains are split into 5 equal sets, i.e. five sets of folds, superfamilies, or families. This ensures that for each of the five partitions, structures having the same classification do not appear in both the train and test sets. For a given classification level each structure appears in a test set once, so that in the cross validation experiment each of the structures will be evaluated exactly once.

The dataset provides 3d coordinates, distance maps, and secondary structure labels. For further details on the construction of the dataset see Rives et al. 2019 Appendix A.10.

This jupyter notebook tutorial shows how to load and index the


, upon initializing, will download
. We also provide
for each of the domains. The data can be directly downloaded below.

| Name | Description | URL | |--------|-------------------------------------------------------------------------------|-----------------------------------------------------------------------| | splits | train/valid splits | | | pkl | pkl objects containing sequence, SSP labels, distance map, and 3d coordinates | | | msas | a3m files containing MSA for each domain | |

Pre-training Dataset Split

The split files establishing which UniRef50 clusters were used as held-out evaluation set for pre-training in Rives et al. 2019 and Rao et al. 2021 can be found here: * UniRef50 IDs of evaluation set: 3.016 M clusters * UniRef100 IDs of evaluation set: 13.745 M proteins, expanding the same UniRef50 clusters.

These files only contain only the UniRef50 IDs and UniRef100 IDs corresponding to the UniRef database, 2018-03 release which is released by the UniProt Consortium under a Creative Commons Attribution (CC BY 4.0) License.


If you find the models useful in your research, we ask that you cite the relevant paper:

  author={Rives, Alexander and Meier, Joshua and Sercu, Tom and Goyal, Siddharth and Lin, Zeming and Liu, Jason and Guo, Demi and Ott, Myle and Zitnick, C. Lawrence and Ma, Jerry and Fergus, Rob},
  title={Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences},

For the self-attention contact prediction:

  author = {Rao, Roshan M and Meier, Joshua and Sercu, Tom and Ovchinnikov, Sergey and Rives, Alexander},
  title={Transformer protein language models are unsupervised structure learners},

For the MSA Transformer:

  author = {Rao, Roshan and Liu, Jason and Verkuil, Robert and Meier, Joshua and Canny, John F. and Abbeel, Pieter and Sercu, Tom and Rives, Alexander},
  title={MSA Transformer},

For variant prediction using ESM-1v:

  author = {Meier, Joshua and Rao, Roshan and Verkuil, Robert and Liu, Jason and Sercu, Tom and Rives, Alexander},
  title = {Language models enable zero-shot prediction of the effects of mutations on protein function},

Much of this code builds on the fairseq sequence modeling framework. We use fairseq internally for our protein language modeling research. We highly recommend trying it out if you'd like to pre-train protein language models from scratch.


This source code is licensed under the MIT license found in the

file in the root directory of this source tree.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.