This repository contains code and pre-trained weights for Transformer protein language models from Facebook AI Research, including our state-of-the-art ESM-2 and MSA Transformer, as well as ESM-1v for predicting variant effects and ESM-IF1 for inverse folding. 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-2 outperforms all tested single-sequence protein language models across a range of structure prediction tasks. The MSA Transformer (ESM-MSA-1) can improve performance on some proteins by leveraging MSA information.
Citation
@article{lin2022language,
title={Language models of protein sequences at the scale of evolution enable accurate structure prediction},
author={Lin, Zeming and Akin, Halil and Rao, Roshan and Hie, Brian and Zhu, Zhongkai and Lu, Wenting and dos Santos Costa, Allan and Fazel-Zarandi, Maryam and Sercu, Tom and Candido, Sal and others},
journal={bioRxiv},
year={2022},
publisher={Cold Spring Harbor Laboratory}
}
@article{rives2021biological,
title={Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences},
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 others},
journal={Proceedings of the National Academy of Sciences},
volume={118},
number={15},
pages={e2016239118},
year={2021},
publisher={National Acad Sciences}
}Table of contents
What's New
- August 2022: ESM-2 - new SOTA Language Models released (see Lin et al. 2022)
- April 2022: New inverse folding model ESM-IF1 released, trained on CATH and UniRef50 predicted structures.
- August 2021: Added flexibility to tokenizer to allow for spaces and special tokens (like
<mask>) in sequence. - July 2021: New pre-trained model ESM-1v released, trained on UniRef90 (see Meier et al. 2021).
- July 2021: New MSA Transformer released, with a minor fix in the row positional embeddings (
ESM-MSA-1b). - Feb 2021: MSA Transformer added (see Rao et al. 2021). Example usage in notebook.
- Dec 2020: Self-Attention Contacts for all pre-trained models (see Rao et al. 2020)
- Dec 2020: Added new pre-trained model ESM-1b (see Rives et al. 2019 Appendix B)
- Dec 2020: ESM Structural Split Dataset (see Rives et al. 2019 Appendix A.10)
| Shorthand | esm.pretrained. |
Dataset | Description |
|---|---|---|---|
| ESM-2 | esm2_t36_3B_UR50D() esm2_t48_15B_UR50D() |
UR50 (sample UR90) | SOTA general-purpose protein language model. Can be used to predict structure, function and other protein properties directly from individual sequences. Released with Lin et al. 2022 (Aug 2022 update). |
| ESM-1b | esm1b_t33_650M_UR50S() |
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 | esm_msa1b_t12_100M_UR50S() |
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 | esm1v_t33_650M_UR90S_1() ... esm1v_t33_650M_UR90S_5() |
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. |
| ESM-IF1 | esm_if1_gvp4_t16_142M_UR50() |
CATH + UR50 | Inverse folding model. Can be used to design sequences for given structures, or to predict functional effects of sequence variation for given structures. Enables SOTA fixed backbone sequence design. Released with Hsu et al. 2022. |
For a complete list of available models, with details and release notes, see Pre-trained Models.
| Task | Unsupervised contact prediction | Structure Prediction | |||
|---|---|---|---|---|---|
| Test set | Large valid | CASP14 | CAMEO (Apr-Jun 2022) | CASP14 | CAMEO (Apr-Jun 2022) |
| Gremlin (Potts) | 39.3 | ||||
| TAPE | 11.2 | ||||
| ProtBert-BFD | 34.1 | ||||
| Prot-T5-XL-BFD | 35.6 | 46.1 | 62.6 | ||
| Prot-T5-XL-Ur50 (3B) | 47.9 | 49.8 | 69.4 | ||
| ESM-1 | 33.7 | ||||
| ESM-1b | 41.1 | 24.4 | 39 | 41.6 | 64.5 |
| ESM-1v | 35.3 | ||||
| ESM-MSA-1b | 57.4 | ||||
| ESM-2 (8M) | 15.9 | 9.8 | 15.7 | 36.7 | 48.1 |
| ESM-2 (35M) | 28.8 | 16.4 | 28.4 | 41.4 | 56.4 |
| ESM-2 (150M) | 42.2 | 26.8 | 40.1 | 49.0 | 64.9 |
| ESM-2 (700M) | 50.1 | 32.5 | 47.6 | 51.3 | 70.1 |
| ESM-2 (3B) | 52.7 | 34.0 | 49.9 | 52.5 | 71.8 |
| ESM-2 (15B) | 54.5 | 37.0 | 51.7 | 55.4 | 72.1 |
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.
- For structure prediction, a AlphaFold2 structure module is trained directly from the frozen language model embeddings. For more details on the method, see Lin et al. 2022.
- Direct coupling analysis methods (Gremlin, mfDCA, Psicov) and ESM-MSA-1 use the trRosetta MSAs, while other methods predict from single sequence.
As a prerequisite, you must have PyTorch installed to use this repository.
You can use this one-liner for installation, using the latest release of esm:
pip install fair-esm # latest release, OR:
pip install git+https://github.com/facebookresearch/esm.git # bleeding edge, current repo main branchWe 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:main", "esm2_t33_650M_UR50D")After pip install, you can load and use a pretrained model as follows:
import torch
import esm
# Load ESM-2 model
model, alphabet = esm.pretrained.esm2_t33_650M_UR50D()
batch_converter = alphabet.get_batch_converter()
model.eval() # disables dropout for deterministic results
# Prepare data (first 2 sequences from ESMStructuralSplitDataset superfamily / 4)
data = [
("protein1", "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"),
("protein2", "KALTARQQEVFDLIRDHISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
("protein2 with mask","KALTARQQEVFDLIRD<mask>ISQTGMPPTRAEIAQRLGFRSPNAAEEHLKALARKGVIEIVSGASRGIRLLQEE"),
("protein3", "K A <mask> I S Q"),
]
batch_labels, batch_strs, batch_tokens = batch_converter(data)
# 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)
plt.show()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-2 model:
python scripts/extract.py esm2_t33_650M_UR50D examples/data/some_proteins.fasta \
examples/data/some_proteins_emb_esm2 --repr_layers 0 32 33 --include mean per_tokDirectory some_proteins_emb_esm2/ now contains one .pt file per FASTA sequence; use torch.load() to load them.
scripts/extract.py has flags that determine what's included in the .pt file:
--repr-layers(default: final only) selects which layers to include embeddings from.--includespecifies what embeddings to save. You can use the following:per_tokincludes the full sequence, with an embedding per amino acid (seq_len x hidden_dim).meanincludes the embeddings averaged over the full sequence, per layer.bosincludes the embeddings from the beginning-of-sequence token. (NOTE: Don't use with the pre-trained models - we trained without bos-token supervision)
If you want to load very large models like 15B and/or do inference on long sequences on your machine, regular GPU inference may lead to OOM errors.
We show how to load the model with Fairscale's Fully Sharded Data Parallel (FSDP) and
use its CPU offloading feature.
This allows to do inference of large models on a single GPU.
Please check out examples/esm2_infer_fairscale_fsdp_cpu_offloading.py for more details.
See "examples/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).
Note that ESM-2 could be used for variant prediction as well, and is expected to have similar performance to ESM-1v.
See "examples/inverse_folding/" for detailed user guide. The ESM-IF1 model is described as GVPTransformer in Learning inverse folding from millions of predicted structures. (Hsu et al. 2022).
We also provide a colab notebook for the sequence design and sequence scoring functionalities.
The ESM-IF1 inverse folding model is built for predicting protein sequences from their backbone atom coordinates. We provide scripts here 1) to sample sequence designs for a given structure and 2) to score sequences for a given structure.
Trained with 12M protein structures predicted by AlphaFold2, the ESM-IF1 model consists of invariant geometric input processing layers followed by a sequence-to-sequence transformer, and achieves 51% native sequence recovery on structurally held-out backbones with 72% recovery for buried residues. The model is also trained with span masking to tolerate missing backbone coordinates and therefore can predict sequences for partially masked structures.
The environment setup is described in this subsection of examples/inverse_folding.
To sample sequences for a given structure in PDB or mmCIF format, use the
sample_sequences.py script. The input file can have either .pdb or
.cif as suffix.
For example, to sample 3 sequence designs for the golgi casein kinase structure (PDB 5YH2; PDB Molecule of the Month from January 2022), we can run the following command from the esm root directory:
python examples/inverse_folding/sample_sequences.py examples/inverse_folding/data/5YH2.pdb \
--chain C --temperature 1 --num-samples 3 --outpath examples/inverse_folding/output/sampled_sequences.fasta
The sampled sequences will be saved in a fasta format to the specified output file.
The temperature parameter controls the sharpness of the probability distribution for sequence sampling. Higher sampling temperatures yield more diverse sequences but likely with lower native sequence recovery. The default sampling temperature is 1. To optimize for native sequence recovery, we recommend sampling with low temperature such as 1e-6.
To score the conditional log-likelihoods for sequences conditioned on a given
structure, use the score_log_likelihoods.py script.
For example, to score the sequences in examples/inverse_folding/data/5YH2_mutated_seqs.fasta
according to the structure in examples/inverse_folding/data/5YH2.pdb, we can run
the following command from the esm root directory:
python examples/inverse_folding/score_log_likelihoods.py examples/inverse_folding/data/5YH2.pdb \
examples/inverse_folding/data/5YH2_mutated_seqs.fasta --chain C \
--outpath examples/inverse_folding/output/5YH2_mutated_seqs_scores.csv
The conditional log-likelihoods are saved in a csv format in the specified output path. The output values are the average log-likelihoods averaged over all amino acids in a sequence.
For more information, see "./examples/inverse_folding/" for detailed user guide.
The ESM-IF1 inverse folding model predicts protein sequences from their backbone atom coordinates, trained with 12M protein structures predicted by AlphaFold2. This notetook guide you through examples of sampling sequences, calculating conditional log-likelihoods, and extracting encoder output as structure representation.
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 examples/data/P62593.fasta either by downloading the precomputed embeddings
as instructed in the notebook or by running the following:
# Obtain the embeddings
python scripts/extract.py esm1v_t33_650M_UR90S_1 examples/data/P62593.fasta \
examples/data/P62593_emb_esm1v --repr_layers 33 --include meanThen, follow the remaining instructions in the tutorial. You can also run the tutorial in a colab notebook.
Note, alternatively use the newer instructions for zero-shot variant prediction, which predicts mutational effects without any supervised training.
This jupyter notebook tutorial demonstrates contact prediction with both the ESM-2 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.predict_contacts(tokens) or model(tokens, return_contacts=True).
And this jupyter notebook tutorial shows how to load and index the ESMStructuralSplitDataset,
and computes the self-attention map unsupervised contact predictions using ESM-2.
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-IF1 | Released with Hsu et al. 2022. |
| ESM-2 | Released with Lin et al. 2022. |
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 ESMStructuralSplitDataset.
ESMStructuralSplitDataset, upon initializing, will download splits and pkl.
We also provide msas for each of the domains. The data can be directly downloaded below.
| Name | Description | URL |
|---|---|---|
| splits | train/valid splits | https://dl.fbaipublicfiles.com/fair-esm/structural-data/splits.tar.gz |
| pkl | pkl objects containing sequence, SSP labels, distance map, and 3d coordinates | https://dl.fbaipublicfiles.com/fair-esm/structural-data/pkl.tar.gz |
| msas | a3m files containing MSA for each domain | https://dl.fbaipublicfiles.com/fair-esm/structural-data/msas.tar.gz |
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:
@article{rives2019biological,
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},
year={2019},
doi={10.1101/622803},
url={https://www.biorxiv.org/content/10.1101/622803v4},
journal={bioRxiv}
}For the self-attention contact prediction:
@article{rao2020transformer,
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},
year={2020},
doi={10.1101/2020.12.15.422761},
url={https://www.biorxiv.org/content/10.1101/2020.12.15.422761v1},
journal={bioRxiv}
}For the MSA Transformer:
@article{rao2021msa,
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},
year={2021},
doi={10.1101/2021.02.12.430858},
url={https://www.biorxiv.org/content/10.1101/2021.02.12.430858v1},
journal={bioRxiv}
}For variant prediction using ESM-1v:
@article{meier2021language,
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},
year={2021},
doi={10.1101/2021.07.09.450648},
url={https://www.biorxiv.org/content/10.1101/2021.07.09.450648v1},
journal={bioRxiv}
}For inverse folding using ESM-IF1:
@article{hsu2022learning,
author = {Hsu, Chloe and Verkuil, Robert and Liu, Jason and Lin, Zeming and Hie, Brian and Sercu, Tom and Lerer, Adam and Rives, Alexander},
title = {Learning inverse folding from millions of predicted structures},
year = {2022},
doi = {10.1101/2022.04.10.487779},
url = {https://www.biorxiv.org/content/early/2022/04/10/2022.04.10.487779},
journal = {bioRxiv}
}
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.
Additionally, if you would like to use the variant prediction benchmark from Meier et al. (2021), we provide a bibtex file with citations for all data in ./examples/variant-prediction/mutation_data.bib. You can cite each paper individually, or add all citations in bulk using the LaTeX command:
\nocite{wrenbeck2017deep,klesmith2015comprehensive,haddox2018mapping,romero2015dissecting,firnberg2014comprehensive,deng2012deep,stiffler2015evolvability,jacquier2013capturing,findlay2018comprehensive,mclaughlin2012spatial,kitzman2015massively,doud2016accurate,pokusaeva2019experimental,mishra2016systematic,kelsic2016rna,melnikov2014comprehensive,brenan2016phenotypic,rockah2015systematic,wu2015functional,aakre2015evolving,qi2014quantitative,matreyek2018multiplex,bandaru2017deconstruction,roscoe2013analyses,roscoe2014systematic,mavor2016determination,chan2017correlation,melamed2013deep,starita2013activity,araya2012fundamental}This source code is licensed under the MIT license found in the LICENSE file
in the root directory of this source tree.