ThermoMPNN - GitLite

KO | EN

gitlite — search

#hacktoberfest #go #machine-learning #deep-learning #linux #macos #java #awesome #kubernetes #awesome-list #pytorch #security #mcp #open-source

ThermoMPNN

(★ 264)

GNN trained to predict changes in thermodynamic stability for protein point mutants

파일 탐색기

최종 버전 다운로드 (.zip)

- - profiles_settings.xml
  - Project_Default.xml
- .gitignore
- .name
- enzyme-stability.iml
- misc.xml
- modules.xml
- vcs.xml
- bootstrap.py
- custom_inference.py
- rosetta_RaSP_metrics.py
- SSM.py
- thermompnn_benchmarking.py
- - fireprot_HF.csv
  - fireprot_test.csv
  - mega_test.csv
  - myoglobin_clean.csv
  - s669_clean_dir.csv
  - ssym-5fold_clean_dir.csv
  - ssym-5fold_clean_inv.csv
- - fireprot_train.csv
  - fireprot_val.csv
  - mega_train.csv
  - mega_val.csv
- fireprot_clusters.tsv
- fireprot_splits.pkl
- mega_splits.pkl
- megascale_clusters.tsv
- megascale_vs_fireprot_homologues.m8
- s669_vs_megascale_homologues.m8
- 2OCJ.pdb
- inference.sh
- ThermoMPNN_inference_2OCJ.csv
- train.sh
- - thermoMPNN_scheme.pdf
- - thermoMPNN_scheme.svg
- thermoMPNN_default.pt
- 1_clean_fireprot_dataset.py
- 2_seqalign_fireprot_dataset.py
- 3_add_metadata_fireprot_dataset.py
- 4_consolidate_pH_entries.py
- download_pdb.py
- split_fireprot.ipynb
- split_megascale.ipynb
- v_48_002.pt
- v_48_010.pt
- v_48_020.pt
- v_48_030.pt
.gitignore
cache.py
config.yaml
datasets.py
environment.yaml
LICENSE
local.yaml
model_utils.py
protein_mpnn_utils.py
README.md
train_thermompnn.py
transfer_model.py

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ThermoMPNN
ThermoMPNN is a graph neural network (GNN) trained using transfer learning to predict changes in stability for protein point mutants.

![ThermoMPNN Scheme](https://raw.githubusercontent.com/Kuhlman-Lab/ThermoMPNN/main/images/SVG/thermoMPNN_scheme.svg)

For details on ThermoMPNN training and methodology, please see the accompanying [paper](https://doi.org/10.1073/pnas.2314853121).

UPDATE (August 2024): A new ThermoMPNN model has been released for prediction of ddG for double mutant pairs at a new repo, [ThermoMPNN-D](https://github.com/Kuhlman-Lab/ThermoMPNN-D). For details, see our new [preprint](https://www.biorxiv.org/content/10.1101/2024.08.20.608844v1)!

UPDATE (September 2024): An experimental ThermoMPNN model has been added for indel (insertion/deletion) ddG prediction. This can be found at [ThermoMPNN-I](https://github.com/Kuhlman-Lab/ThermoMPNN-D/tree/ThermoMPNN-I). We include a new [Colab notebook](https://github.com/Kuhlman-Lab/ThermoMPNN-D/blob/ThermoMPNN-I/ThermoMPNN-I.ipynb) for ThermoMPNN-I as well. Note that this model has only had limited _in silico_ validation.

## Citing this work
If you use the code, please cite:
```
@article{
doi:10.1073/pnas.2314853121,
author = {Henry Dieckhaus  and Michael Brocidiacono  and Nicholas Z. Randolph  and Brian Kuhlman },
title = {Transfer learning to leverage larger datasets for improved prediction of protein stability changes},
journal = {Proceedings of the National Academy of Sciences},
volume = {121},
number = {6},
pages = {e2314853121},
year = {2024},
doi = {10.1073/pnas.2314853121},
URL = {https://www.pnas.org/doi/abs/10.1073/pnas.2314853121},
eprint = {https://www.pnas.org/doi/pdf/10.1073/pnas.2314853121},
}
```

## Colab Implementation
For a user-friendly version of ThermoMPNN requiring no installation, use this [Colab notebook](https://colab.research.google.com/drive/1OcT4eYwzxUFNlHNPk9_5uvxGNMVg3CFA#scrollTo=i06A5VI142NT).

## Installation
To install ThermoMPNN, first clone this repository
```
git clone https://github.com/Kuhlman-Lab/ThermoMPNN.git
```
Then use the file ```environment.yaml``` install the necessary python dependencies (I recommend using mamba for convenience):
```
mamba env create -f environment.yaml
```
This will create a conda environment called ```thermoMPNN```.

NOTE: installing from the .yaml file may cause pytorch to install incorrectly (CPU version instead of GPU version).

Alternate installation instructions:
```
# create and activate environment
mamba create -n thermoMPNN python=3.10
mamba activate thermoMPNN

# install pytorch related packages
mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 pytorch-lightning -c nvidia -c pytorch -c conda-forge

# install all other packages
mamba install joblib omegaconf pandas numpy tqdm mmseqs2 wandb biopython -c bioconda -c conda-forge -c anaconda

# check for proper GPU recognition (should return True)
python -c 'import torch; print(torch.cuda.is_available())'
```
Note: if you are planning to do any model training or complicated inference (i.e., from a CSV), you will need to update the ```local.yaml``` file to reflect dataset locations on your local system so that ThermoMPNN can find the data it needs. This step can be skipped if only running ```custom_inference.py```.

## Inference
There are a few different ways to run inference with ThermoMPNN all located in the ```analysis``` directory.

### From a PDB
The simplest way is to use the ```custom_inference.py``` script to pass a custom PDB to ThermoMPNN for site-saturation mutagenesis.

### From a CSV and many PDBs
For larger batches of predictions, it is recommended to set up a **CustomDataset** object by inheriting from the **ddgBenchDataset** class in the ```datasets.py``` file, then add this dataset to the ```SSM.py``` script to get aggregated predictions for the whole dataset.

### For benchmarking purposes
The ```thermompnn_benchmarking.py``` is set up to score different models on a **CustomDataset** object or one of the datasets used in this study. An example inference SLURM script is provided at ```examples/inference.sh```.

## Training
The main training script is ```train_thermompnn.py```. To set up a training run, you must write a ```config.yaml``` file (example provided) to specify model hyperparameters. You also must provide a ```local.yaml``` file to tell ThermoMPNN where to find your data. These files serve as experiment logs as well.

Training ThermoMPNN requires the use of a GPU. On a small dataset (<5000 data points), training takes <30s per epoch, while on a mega-scale dataset (>200,000 data points), it takes 8-12min per epoch (on a single V100 GPU). An example training SLURM script is provided at ```examples/train.sh```.

### Splits and Model Weights
For the purpose of replication and future benchmarking, the dataset splits used in this study are included as ```.pkl``` files under the ```dataset_splits/``` directory.

ThermoMPNN model weights can be found in the ```models/``` directory. The following model weights are provided:
```
- thermoMPNN_default.pt (best ThermoMPNN model trained on Megascale training dataset)
```

### Dataset Availability
The datasets used in this study can be found in the following locations:

Fireprot: https://doi.org/10.5281/zenodo.8169288
Megascale: https://doi.org/10.5281/zenodo.7401274
S669: https://doi.org/10.1093/bib/bbab555
SSYM, P53, Myoglobin, etc: https://protddg-bench.github.io