Data, weights, and code for running the TAPE benchmark on a trained protein embedding. We provide a pretraining corpus, five supervised downstream tasks, pretrained language model weights, and benchmarking code. This code has been updated to use pytorch - as such previous pretrained model weights and code will not work. The previous tensorflow TAPE repository is still available at https://github.com/songlab-cal/tape-neurips2019.
This repository is not an effort to maintain maximum compatibility and reproducability with the original paper, but is instead meant to facilitate ease of use and future development (both for us, and for the community). Although we provide much of the same functionality, we have not tested every aspect of training on all models/downstream tasks, and we have also made some deliberate changes. Therefore, if your goal is to reproduce the results from our paper, please use the original code.
Our paper is available at https://arxiv.org/abs/1906.08230.
Some documentation is incomplete. We will try to fill it in over time, but if there is something you would like an explanation for, please open an issue so we know where to focus our effort!
In this fork, we introduce a few extensions to the existing tasks and capabilities supported by TAPE
- Adding a new task for subcellular location, based on the DeepLoc dataset and publication.
- Adding Jupyter notebooks for experimenting with TAPE embeddings and the subcellular location task in particular. This is described in more detail in Notebooks.md.
We recommend that you install tape
into a python virtual environment using
$ pip install tape_proteins
We build on the excellent huggingface repository and use this as an API to define models, as well as to provide pretrained models. By using this API, pretrained models will be automatically downloaded when necessary and cached for future use.
import torch
from tape import ProteinBertModel, TAPETokenizer
model = ProteinBertModel.from_pretrained('bert-base')
tokenizer = TAPETokenizer(vocab='iupac') # iupac is the vocab for TAPE models, use unirep for the UniRep model
# Pfam Family: Hexapep, Clan: CL0536
sequence = 'GCTVEDRCLIGMGAILLNGCVIGSGSLVAAGALITQ'
token_ids = torch.tensor([tokenizer.encode(sequence)])
output = model(token_ids)
sequence_output = output[0]
pooled_output = output[1]
# NOTE: pooled_output is *not* trained for the transformer, do not use
# w/o fine-tuning. A better option for now is to simply take a mean of
# the sequence output
Currently available pretrained models are:
If there is a particular pretrained model that you would like to use, please open an issue and we will try to add it!
Given an input fasta file, you can generate a .npz
file containing embedding proteins via the tape-embed
command.
Suppose this is our input fasta file:
>seq1
GCTVEDRCLIGMGAILLNGCVIGSGSLVAAGALITQ
>seq2
RTIKVRILHAIGFEGGLMLLTIPMVAYAMDMTLFQAILLDLSMTTCILVYTFIFQWCYDILENR
Then we could embed it with the UniRep babbler-1900 model like so:
tape-embed unirep my_input.fasta output_filename.npz babbler-1900 --tokenizer unirep
There is no need to download the pretrained model manually - it will be automatically downloaded if needed. In addition, note the change of tokenizer to the unirep
tokenizer. UniRep uses a different vocabulary, and so requires this tokenzer. If you get a cublas runtime error, please double check that you changed tokenizer correctly.
The embed function is fully batched and will automatically distribute across as many GPUs as the machine has available. On a Titan Xp, it can process around 200 sequences / second.
Once we have the output file, we can load it into numpy like so:
arrays = np.load('output_filename.npz', allow_pickle=True)
list(arrays.keys()) # Will output the name of the keys in your fasta file (or if unnamed then '0', '1', ...)
arrays[<protein_id>] # Returns a dictionary with keys 'pooled' and 'avg', (or 'seq' if using the --full_sequence_embed flag)
By default to save memory TAPE returns the average of the sequence embedding along with the pooled embedding generated through the pooling function. For some models (like UniRep), the pooled embedding is trained, and so can be used out of the box. For other models (like the transformer), the pooled embedding is not trained, and so the average embedding should be used. We will be looking into methods of self-supervised training the pooled embedding for all models in the future.
If you would like the full embedding rather than the average embedding, this can be specified to tape-embed
by passing the --full_sequence_embed
flag.
Tape provides two commands for training, tape-train
and tape-train-distributed
. The first command uses standard pytorch data distribution to distributed across all available GPUs. The second one uses torch.distributed.launch
-style multiprocessing to distributed across the number of specified GPUs (and could also be used for distributing across multiple nodes). We generally recommend using the second command, as it can provide a 10-15% speedup, but both will work.
To train the transformer on masked language modeling, for example, you could run this
tape-train-distributed transformer masked_language_modeling --batch_size BS --learning_rate LR --fp16 --warmup_steps WS --nproc_per_node NGPU --gradient_accumulation_steps NSTEPS
There are a number of features used in training:
* Distributed training via multiprocessing
* Half-precision training
* Gradient accumulation
* Gradient-allreduce post accumulation
* Automatic batch by sequence length
The first feature you are likely to need is the gradient_accumulation_steps
. TAPE specifies a relatively high batch size (1024) by default. This is the batch size that will be used per backwards pass. This number will be divided by the number of GPUs as well as the gradient accumulation steps. So with a batch size of 1024, 2 GPUs, and 1 gradient accumulation step, you will do 512 examples per GPU. If you run out of memory (and you likely will), TAPE provides a clear error message and will tell you to increase the gradient accumulation steps.
There are additional features as well that are not talked about here. See tape-train-distributed --help
for a list of all commands.
Once you've trained a language model, you'll have a pretrained weight file located in the results
folder. To evaluate this model, you can do one of two things. One option is to directly evaluate the language modeling accuracy / perplexity. tape-train
will report the perplexity over the training and validation set at the end of each epoch. However, we find empirically that language modeling accuracy and perplexity are poor measures of performance on downstream tasks. Therefore, to evaluate the language model we strongly recommend training your model on one or all of our provided tasks.
Training a model on a downstream task can also be done with the tape-train
command. Simply use the same syntax as with training a language model, adding the flag --from_pretrained <path_to_your_saved_results>
. To train a pretrained transformer on secondary structure prediction, for example, you would run
tape-train-distributed transformer secondary_structure \
--from_pretrained results/<path_to_folder> \
--batch_size BS \
--learning_rate LR \
--fp16 \
--warmup_steps WS \
--nproc_per_node NGPU \
--gradient_accumulation_steps NSTEPS \
--num_train_epochs NEPOCH \
--eval_freq EF \
--save_freq SF
For training a downstream model, you will likely need to experiment with hyperparameters to achieve the best results (optimal hyperparameters vary per-task and per-model). The set of parameters to consider are
* Batch size
* Learning rate
* Warmup steps
* Num train epochs
These can all have significant effects on performance, and by default are set to maximize performance on language modeling rather than downstream tasks. In addition the eval_freq
and save_freq
parameters can be useful, as they reduce the frequency of running validation passes and saving the model, respectively. Since downstream task epochs are much shorter (and you're likely to need more of them), it makes sense to increase these values so that training takes less time.
To evaluate your downstream task model, we provide the tape-eval
command. This command will output your model predictions along with a set of metrics that you specify. At the moment, we support mean squared error (mse
), mean absolute error (mae
), Spearman's rho (spearmanr
), and accuracy (accuracy
). Precision @ L/5 will be added shortly.
The syntax for the command is
tape-eval MODEL TASK TRAINED_MODEL_FOLDER --metrics METRIC1 METRIC2 ...
so to evaluate a transformer trained on trained secondary structure, we can run
tape-eval transformer secondary_structure results/<path_to_trained_model> --metrics accuracy
This will report the overall accuracy, and will also dump a results.pkl
file into the trained model directory for you to analyze however you like.
We have recently re-implemented the trRosetta model from Yang et. al. (2020). A link to the original repository, which was used as a basis for this re-implementation, can be found here. We provide a pytorch implementation and dataset to allow you to play around with the model. Data is available here. This is the same as the data in the original paper, however we've added train / val split files to allow you to train your own model reproducibly. To use this model
from tape import TRRosetta
from tape.datasets import TRRosettaDataset
# Download data and place it under `<data_path>/trrosetta`
train_data = TRRosettaDatset('<data_path>', 'train') # will subsample MSAs
valid_data = TRRosettaDatset('<data_path>', 'valid') # will not subsample MSAs
model = TRRosetta.from_pretrained('xaa') # valid choices are 'xaa', 'xab', 'xac', 'xad', 'xae'. Each corresponds to one of the ensemble models.
batch = train_data.collate_fn([train_data[0]])
loss, predictions = model(**batch)
The predictions can be saved as .npz
files and then fed into the structure modeling scripts provided by the Yang Lab.
The available models are:
transformer
(pretrained available)resnet
lstm
unirep
(pretrained available)onehot
(no pretraining required)trrosetta
(pretrained available)
The available standard tasks are:
language_modeling
masked_language_modeling
secondary_structure
contact_prediction
remote_homology
fluorescence
stability
subcellular_location
(WIP)trrosetta
(can only be used withtrrosetta
model)
The available models and tasks can be found in tape/datasets.py
and tape/models/modeling*.py
.
We have made some efforts to make the new repository easier to understand and extend. See the examples
folder for an example on how to add a new model and a new task to TAPE. If there are other examples you would like or if there is something missing in the current examples, please open an issue.
Data should be placed in the ./data
folder, although you may also specify a different data directory if you wish.
The supervised data is around 120MB compressed and 2GB uncompressed.
The unsupervised Pfam dataset is around 7GB compressed and 19GB uncompressed. The data for training is hosted on AWS. By default we provide data as LMDB - see tape/datasets.py
for examples on loading the data. If you wish to download all of TAPE, run download_data.sh
to do so. We also provide links to each individual dataset below in both LMDB format and JSON format.
Pretraining Corpus (Pfam) | Secondary Structure | Contact (ProteinNet) | Remote Homology | Fluorescence | Stability
Raw data files are stored in JSON format for maximum portability. This data is JSON-ified, which removes certain constructs (in particular numpy arrays). As a result they cannot be directly loaded into the provided pytorch datasets (although the conversion should be quite easy by simply adding calls to np.array
).
Pretraining Corpus (Pfam) | Secondary Structure | Contact (ProteinNet) | Remote Homology | Fluorescence | Stability
We will soon have a leaderboard available for tracking progress on the core five TAPE tasks, so check back for a link here. See the main tables in our paper for a sense of where performance stands at this point. Publication on the leaderboard will be contingent on meeting the following citation guidelines.
In the meantime, here's a temporary leaderboard for each task. All reported models on this leaderboard use unsupervised pretraining.
Ranking | Model | Accuracy (3-class) |
---|---|---|
1. | One Hot + Alignment | 0.80 |
2. | LSTM | 0.75 |
2. | ResNet | 0.75 |
4. | Transformer | 0.73 |
4. | Bepler | 0.73 |
4. | Unirep | 0.73 |
7. | One Hot | 0.69 |
Ranking | Model | L/5 Medium + Long Range |
---|---|---|
1. | One Hot + Alignment | 0.64 |
2. | Bepler | 0.40 |
3. | LSTM | 0.39 |
4. | Transformer | 0.36 |
5. | Unirep | 0.34 |
6. | ResNet | 0.29 |
6. | One Hot | 0.29 |
Ranking | Model | Top 1 Accuracy |
---|---|---|
1. | LSTM | 0.26 |
2. | Unirep | 0.23 |
3. | Transformer | 0.21 |
4. | Bepler | 0.17 |
4. | ResNet | 0.17 |
6. | One Hot + Alignment | 0.09 |
6. | One Hot | 0.09 |
Ranking | Model | Spearman's rho |
---|---|---|
1. | Transformer | 0.68 |
2. | LSTM | 0.67 |
2. | Unirep | 0.67 |
4. | Bepler | 0.33 |
5. | ResNet | 0.21 |
6. | One Hot | 0.14 |
Ranking | Model | Spearman's rho |
---|---|---|
1. | Transformer | 0.73 |
1. | Unirep | 0.73 |
1. | ResNet | 0.73 |
4. | LSTM | 0.69 |
5. | Bepler | 0.64 |
6. | One Hot | 0.19 |
If you find TAPE useful, please cite our corresponding paper. Additionally, anyone using the datasets provided in TAPE must describe and cite all dataset components they use. Producing these data is time and resource intensive, and we insist this be recognized by all TAPE users. For convenience,data_refs.bib
contains all necessary citations. We also provide each individual citation below.
TAPE (Our paper):
@inproceedings{tape2019,
author = {Rao, Roshan and Bhattacharya, Nicholas and Thomas, Neil and Duan, Yan and Chen, Xi and Canny, John and Abbeel, Pieter and Song, Yun S},
title = {Evaluating Protein Transfer Learning with TAPE},
booktitle = {Advances in Neural Information Processing Systems}
year = {2019}
}
Pfam (Pretraining):
@article{pfam,
author = {El-Gebali, Sara and Mistry, Jaina and Bateman, Alex and Eddy, Sean R and Luciani, Aur{\'{e}}lien and Potter, Simon C and Qureshi, Matloob and Richardson, Lorna J and Salazar, Gustavo A and Smart, Alfredo and Sonnhammer, Erik L L and Hirsh, Layla and Paladin, Lisanna and Piovesan, Damiano and Tosatto, Silvio C E and Finn, Robert D},
doi = {10.1093/nar/gky995},
file = {::},
issn = {0305-1048},
journal = {Nucleic Acids Research},
keywords = {community,protein domains,tandem repeat sequences},
number = {D1},
pages = {D427--D432},
publisher = {Narnia},
title = {{The Pfam protein families database in 2019}},
url = {https://academic.oup.com/nar/article/47/D1/D427/5144153},
volume = {47},
year = {2019}
}
SCOPe: (Remote Homology and Contact)-
@article{scop,
title={SCOPe: Structural Classification of Proteins—extended, integrating SCOP and ASTRAL data and classification of new structures},
author={Fox, Naomi K and Brenner, Steven E and Chandonia, John-Marc},
journal={Nucleic acids research},
volume={42},
number={D1},
pages={D304--D309},
year={2013},
publisher={Oxford University Press}
}
PDB: (Secondary Structure and Contact)
@article{pdb,
title={The protein data bank},
author={Berman, Helen M and Westbrook, John and Feng, Zukang and Gilliland, Gary and Bhat, Talapady N and Weissig, Helge and Shindyalov, Ilya N and Bourne, Philip E},
journal={Nucleic acids research},
volume={28},
number={1},
pages={235--242},
year={2000},
publisher={Oxford University Press}
}
CASP12: (Secondary Structure and Contact)
@article{casp,
author = {Moult, John and Fidelis, Krzysztof and Kryshtafovych, Andriy and Schwede, Torsten and Tramontano, Anna},
doi = {10.1002/prot.25415},
issn = {08873585},
journal = {Proteins: Structure, Function, and Bioinformatics},
keywords = {CASP,community wide experiment,protein structure prediction},
pages = {7--15},
publisher = {John Wiley {\&} Sons, Ltd},
title = {{Critical assessment of methods of protein structure prediction (CASP)-Round XII}},
url = {http://doi.wiley.com/10.1002/prot.25415},
volume = {86},
year = {2018}
}
NetSurfP2.0: (Secondary Structure)
@article{netsurfp,
title={NetSurfP-2.0: Improved prediction of protein structural features by integrated deep learning},
author={Klausen, Michael Schantz and Jespersen, Martin Closter and Nielsen, Henrik and Jensen, Kamilla Kjaergaard and Jurtz, Vanessa Isabell and Soenderby, Casper Kaae and Sommer, Morten Otto Alexander and Winther, Ole and Nielsen, Morten and Petersen, Bent and others},
journal={Proteins: Structure, Function, and Bioinformatics},
year={2019},
publisher={Wiley Online Library}
}
ProteinNet: (Contact)
@article{proteinnet,
title={ProteinNet: a standardized data set for machine learning of protein structure},
author={AlQuraishi, Mohammed},
journal={arXiv preprint arXiv:1902.00249},
year={2019}
}
Fluorescence:
@article{sarkisyan2016,
title={Local fitness landscape of the green fluorescent protein},
author={Sarkisyan, Karen S and Bolotin, Dmitry A and Meer, Margarita V and Usmanova, Dinara R and Mishin, Alexander S and Sharonov, George V and Ivankov, Dmitry N and Bozhanova, Nina G and Baranov, Mikhail S and Soylemez, Onuralp and others},
journal={Nature},
volume={533},
number={7603},
pages={397},
year={2016},
publisher={Nature Publishing Group}
}
Stability:
@article{rocklin2017,
title={Global analysis of protein folding using massively parallel design, synthesis, and testing},
author={Rocklin, Gabriel J and Chidyausiku, Tamuka M and Goreshnik, Inna and Ford, Alex and Houliston, Scott and Lemak, Alexander and Carter, Lauren and Ravichandran, Rashmi and Mulligan, Vikram K and Chevalier, Aaron and others},
journal={Science},
volume={357},
number={6347},
pages={168--175},
year={2017},
publisher={American Association for the Advancement of Science}
}
Subcellular Location:
@article{10.1093/bioinformatics/btx431,
title = {DeepLoc: prediction of protein subcellular localization using deep learning},
author = {Almagro Armenteros, José Juan and Sønderby, Casper Kaae and Sønderby, Søren Kaae and Nielsen, Henrik and Winther, Ole},
journal = {Bioinformatics},
volume = {33},
number = {21},
pages = {3387-3395},
year = {2017}
}