Awesome

LM-GVP:

Source code for "LM-GVP: A Generalizable Deep Learning Framework for Protein Property Prediction from Sequence and Structure" <img src="figs/lmgvp.png">

Dependencies

LM-GVP experiments were performed in Python 3.6 with Pytorch (v1.8).

To install all dependencies run:

$ pip install -r requirements.txt

Training

Training of LM-GVP and competitive models for protein property predictions can be done in train.py, which utilizes the PyTorch Lightning Trainer. All of the trainer flags in PyTorch Lightning are supported. To see the usage, run:

$ python train.py -h

usage: train.py [-h] [--logger [LOGGER]] [--checkpoint_callback [CHECKPOINT_CALLBACK]] [--default_root_dir DEFAULT_ROOT_DIR] [--gradient_clip_val GRADIENT_CLIP_VAL] [--gradient_clip_algorithm GRADIENT_CLIP_ALGORITHM]
                [--process_position PROCESS_POSITION] [--num_nodes NUM_NODES] [--num_processes NUM_PROCESSES] [--gpus GPUS] [--auto_select_gpus [AUTO_SELECT_GPUS]] [--tpu_cores TPU_CORES] [--log_gpu_memory LOG_GPU_MEMORY]
                [--progress_bar_refresh_rate PROGRESS_BAR_REFRESH_RATE] [--overfit_batches OVERFIT_BATCHES] [--track_grad_norm TRACK_GRAD_NORM] [--check_val_every_n_epoch CHECK_VAL_EVERY_N_EPOCH] [--fast_dev_run [FAST_DEV_RUN]]
                [--accumulate_grad_batches ACCUMULATE_GRAD_BATCHES] [--max_epochs MAX_EPOCHS] [--min_epochs MIN_EPOCHS] [--max_steps MAX_STEPS] [--min_steps MIN_STEPS] [--max_time MAX_TIME] [--limit_train_batches LIMIT_TRAIN_BATCHES]
                [--limit_val_batches LIMIT_VAL_BATCHES] [--limit_test_batches LIMIT_TEST_BATCHES] [--limit_predict_batches LIMIT_PREDICT_BATCHES] [--val_check_interval VAL_CHECK_INTERVAL]
                [--flush_logs_every_n_steps FLUSH_LOGS_EVERY_N_STEPS] [--log_every_n_steps LOG_EVERY_N_STEPS] [--accelerator ACCELERATOR] [--sync_batchnorm [SYNC_BATCHNORM]] [--precision PRECISION] [--weights_summary WEIGHTS_SUMMARY]
                [--weights_save_path WEIGHTS_SAVE_PATH] [--num_sanity_val_steps NUM_SANITY_VAL_STEPS] [--truncated_bptt_steps TRUNCATED_BPTT_STEPS] [--resume_from_checkpoint RESUME_FROM_CHECKPOINT] [--profiler PROFILER]
                [--benchmark [BENCHMARK]] [--deterministic [DETERMINISTIC]] [--reload_dataloaders_every_epoch [RELOAD_DATALOADERS_EVERY_EPOCH]] [--auto_lr_find [AUTO_LR_FIND]] [--replace_sampler_ddp [REPLACE_SAMPLER_DDP]]
                [--terminate_on_nan [TERMINATE_ON_NAN]] [--auto_scale_batch_size [AUTO_SCALE_BATCH_SIZE]] [--prepare_data_per_node [PREPARE_DATA_PER_NODE]] [--plugins PLUGINS] [--amp_backend AMP_BACKEND] [--amp_level AMP_LEVEL]
                [--distributed_backend DISTRIBUTED_BACKEND] [--move_metrics_to_cpu [MOVE_METRICS_TO_CPU]] [--multiple_trainloader_mode MULTIPLE_TRAINLOADER_MODE] [--stochastic_weight_avg [STOCHASTIC_WEIGHT_AVG]]
                [--model_name MODEL_NAME] --task TASK [--pretrained_weights PRETRAINED_WEIGHTS] [--lr LR] [--bs BS] [--early_stopping_patience EARLY_STOPPING_PATIENCE] [--num_workers NUM_WORKERS]

optional arguments:
  -h, --help            show this help message and exit
  --model_name MODEL_NAME
                        Choose from gvp, bert, bert_gvp, gat, bert_gat
  --task TASK           Task to perform: ['flu', 'protease', 'cc', 'bp', 'mf']
  --pretrained_weights PRETRAINED_WEIGHTS
                        path to pretrained weights (such as GAE) to initialize model
  --lr LR               learning rate
  --bs BS               batch size
  --early_stopping_patience EARLY_STOPPING_PATIENCE
  --num_workers NUM_WORKERS
                        num_workers used in DataLoader

pl.Trainer:
  --logger [LOGGER]     Logger (or iterable collection of loggers) for experiment tracking. A ``True`` value uses the default ``TensorBoardLogger``. ``False`` will disable logging.
  # other pl.Trainer flags...

A recommended set of training flags with multi-GPU and mixed precision training is as follows:

$ python train.py \
    --task cc \
    --model_name bert_gvp \
    --freeze_bert True \
    --num_nodes 1 \
    --gpus 8 \
    --bs 4 \
    --lr 1e-4 \
    --max_epochs 200 \
    --early_stopping_patience 10 \
    --precision 16 \
    --accelerator ddp \
    --num_workers 32

Data preparation

Scripts and utilility functions used for preparing protein sequence and structure data are available in data. The python scripts parse PDB files to extract the coordinates of the 4 key atoms from amino acids to generate json records compatible to the LM-GVP and competitive models.

Datasets used in this study are from publicly available resources:

GO datasets
TAPE datasets:
- Fluorescence: s3://songlabdata/proteindata/data_pytorch/fluorescence.tar.gz
- Protease stability: s3://songlabdata/proteindata/data_pytorch/stability.tar.gz

Model interpretation and mechanistic analyses

Codes and notebooks used for model analyses are available in the following directories:

Citation

Please cite the following preprint:

@article {LMGVP21,
	author = {Wang, Zichen and Combs, Steven A. and Brand, Ryan and Calvo, Miguel Romero and Xu, Panpan and Price, George and Golovach, Nataliya and Salawu, Emannuel O. and Wise, Colby J. and Ponnapalli, Sri Priya and Clark, Peter M.},
	title = {LM-GVP: A Generalizable Deep Learning Framework for Protein Property Prediction from Sequence and Structure},
	elocation-id = {2021.09.21.460852},
	year = {2021},
	doi = {10.1101/2021.09.21.460852},
	publisher = {Cold Spring Harbor Laboratory},
	abstract = {Proteins perform many essential functions in biological systems and can be successfully developed as bio-therapeutics. It is invaluable to be able to predict their properties based on a proposed sequence and structure. In this study, we developed a novel generalizable deep learning framework, LM-GVP, composed of a protein Language Model (LM) and Graph Neural Network (GNN) to leverage information from both 1D amino acid sequences and 3D structures of proteins. Our approach outperformed the state-of-the-art protein LMs on a variety of property prediction tasks including fluorescence, protease stability, and protein functions from Gene Ontology (GO). We also illustrated insights into how a GNN prediction head can guide the protein LM to better leverage structural information. We envision that our deep learning framework will be generalizable to many protein property prediction problems to greatly accelerate protein engineering and drug development.Competing Interest StatementThe authors have declared no competing interest.},
	URL = {https://www.biorxiv.org/content/early/2021/09/21/2021.09.21.460852},
	eprint = {https://www.biorxiv.org/content/early/2021/09/21/2021.09.21.460852.full.pdf},
	journal = {bioRxiv}
}

Security

See CONTRIBUTING for more information.

License

This library is licensed under the MIT-0 License. See the LICENSE file.