This is the official implementation of ThermoMPNN-D, a Siamese neural network designed to predict stability changes from protein double point mutations.
This work is an extension of ThermoMPNN (https://github.com/Kuhlman-Lab/ThermoMPNN), which is itself an extension of ProteinMPNN (https://github.com/dauparas/ProteinMPNN). For details, see our manuscript here.
To try out ThermoMPNN-D right in your browser, use the Colab notebook at this link.
First, clone the repository:
git clone https://github.com/Kuhlman-Lab/ThermoMPNN-D.git
cd ThermoMPNN-D
Then, install the python dependencies found in environment.yaml (I recommend using mamba):
mamba env create -f environment.yaml
Add ThermoMPNN to your PYTHONPATH so that python can find all the modules:
export PYTHONPATH=$PYTHONPATH:/path/to/ThermoMPNN-D
Finally, modify the local filepath information found in examples/configs/local.yaml to match your system.
Before running any ThermoMPNN-D scripts, just run mamba activate ThermoMPNN-D to load the necessary python packages.
We provide a script called v2_ssm.py which does inference on all possible single or double mutants in the protein. The output for this script is a CSV file with mutation and ddG values.
There is an important option called --threshold which dictates which mutations will get saved to disk. By default, ThermoMPNN will only save stabilizing mutations (ddG <= -0.5 kcal/mol), since this is fastest for saving to disk. To get all the mutations, including destabilizing mutations, set --threshold very high (e.g., 100).
The other useful option is --distance which is used for additive or epistatic predictions. This is the distance threshold used to filter for "nearby" residues that are likely to have epistatic interactions. A smaller value will lead to stricter filtering. Default is 15 A (based on Ca-Ca distance).
This is an updated version of single mutant ThermoMPNN that uses fewer parameters and proper batched inference for faster prediction. It should give similar results to the previously published ThermoMPNN models.
python v2_ssm.py --mode single --pdb 1VII.pdb --batch_size 256 --out 1VII
This sums the individual contributions from each single mutation without attempting to quantify epistatic coupling terms. Inference is faster than with the epistatic model since it just needs to add the terms rather than predict each one separately.
python v2_ssm.py --mode additive --pdb 1VII.pdb --batch_size 256 --out 1VII
This model attempts to capture epistatic interactions between double mutations, which requires running inference on every individual mutation. This is slower than the single or additive model but is still reasonably fast (<1 minute) due to some vectorizing and batching tricks.
python v2_ssm.py --mode epistatic --pdb examples/pdbs/1VII.pdb --batch_size 2048 --out 1VII
Note the higher batch size, which takes advantage of the lightweight prediction head to significantly speed up inference.
To repeat the benchmarks in the ThermoMPNN-D paper, see the benchmarks.ipynb notebook.
The datasets used in this study can be obtained from https://zenodo.org/records/13345274. Note that you may need to slightly modify the column names and/or filepaths to line up with what the benchmark script expects.
Training requires compatible CUDA drivers and an accessible GPU. Single mutant epochs should take 2-3 minutes on a V100 GPU, while epochs for epistatic models take a bit longer (8-10 minutes) due to data augmentation which provides a larger dataset. Training typically converges in 30-40 epochs.
python train_thermompnn.py ../examples/configs/local.yaml ../examples/configs/train_single.yaml
python train_thermompnn.py ../examples/configs/local.yaml ../examples/configs/train_epistatic.yaml
Metric curves can be logged using W&B if desired - simply un-comment the Project and name fields in train.yaml and hook up your W&B account.
This work is made available under an MIT license (see LICENSE file for details).
If this work is useful to you, please use the following citation:
@article {Dieckhaus2024.08.20.608844,
author = {Dieckhaus, Henry and Kuhlman, Brian},
title = {Protein stability models fail to capture epistatic interactions of double point mutations},
journal = {bioRxiv},
elocation-id = {2024.08.20.608844},
year = {2024},
doi = {10.1101/2024.08.20.608844},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2024/08/21/2024.08.20.608844},
eprint = {https://www.biorxiv.org/content/early/2024/08/21/2024.08.20.608844.full.pdf},
}