ThermoNet is a computational method for quantitative prediction of the changes in protein thermodynamic stability ($\Delta\Delta G$) caused by single-point amino acid substitutions. The core algorithm of ThermoNet is an ensemble of deep 3D convolutional neural networks. ThermoNet was benchmarked against a large set of $\Delta\Delta G$ predictors and was shown to perform equally well as the best methods while effectively addressing prediction bias. If you find ThermoNet useful in your work, please consider citing our paper:

• Li, B., Yang, Y.T., Capra, J.A., and Gerstein, M.B. (2020). Predicting changes in protein thermodynamic stability upon point mutation with deep 3D convolutional neural networks. PLoS Comput Biol 16, e1008291.

ThermoNet depends on a few packages that are only available on Linux platforms, thus, it has only been tested on Linux platforms. To use ThermoNet, you would need to install the following software and Python libraries:

• Rosetta 3.10. Rosetta is a multi-purpose macromolecular modeling suite that is used in ThermoNet for creating and refining protein structures. You can get Rosetta from the Rosetta Commons website: https://www.rosettacommons.org/software
• HTMD 1.17. HTMD is a Python library for high-throughput molecular modeling and dynamics simulations. It is called in ThermoNet for voxelizing protein structures and parametrizing the resulting 3D voxel grids. We developed ThermoNet to work with HTMD 1.17, however, it was also shown to work with HTMD 1.22 (thanks to Dr. Toni Giorgino for testing it out). The latest tests showed that ThermoNet is compatible with HTMD 2.0.5.
• Keras 2.1.6 with TensorFlow 1.12.0 as the backend.

Predicting the $\Delta\Delta G$ using ThermoNet is done through a multi-step protocol. Each step relies on a specific third-party software that needs to be installed first. In the following, we outline the steps to install them.

Clone ThermoNet to a local directory.

git clone https://github.com/gersteinlab/ThermoNet.git

1. Go to https://els2.comotion.uw.edu/product/rosetta to get an academic license for Rosetta.
2. Download Rosetta 3.10 (source + binaries for Linux) from this site: https://www.rosettacommons.org/software/license-and-download
3. Extract the tarball to a local directory from which Rosetta binaries can be called by specifying their full path.

The free version of HTMD is free to non-commercial users although it does not come with full functionality. But to use it with ThermoNet, the free version is sufficient. You can either install it by following the instructions listed here, or by running

conda env create --name thermonet --file environment.yaml

The above command will create a conda environment and install all dependencies so that one can run ThermoNet to make input tensors.

This step is needed for running ThermoNet to make $\Delta\Delta G$ predictions. There are many resources out there that one can follow to install TensorFlow and Keras. I found it easiest to install them with the Anaconda Python distribution.

1. Get the Python 3.7 version Anaconda 2019.10 for Linux installer.
2. Follow the instructions here to install it.
3. Open anaconda-navigator from the command line. Go to Environments and search for keras and tensorflow, and install all the matched libraries.

Alternatively, one can create a conda environment to use Keras and TensorFlow on a high-performance computing cluster, i.e.:

# create conda environment for deep learning/neural networks
conda create -y -n tensorflow112 python=3.6 anaconda
source activate tensorflow112

# install libraries
pip install keras tensorflow-gpu==1.12

As previously mentioned, the ThermoNet $\Delta\Delta G$ method is a multi-step protocol. We outline the steps needed to make $\Delta\Delta G$ predictions of a given variant or a list of variants in the following. Note that ThermoNet requires that a protein structural model is available for the protein from which the variants are derived. In the case where an experimental structure is not available, in principle one can create a structural model through homology modeling. However, we did not test ThermoNet under such a scenario, thus, its performance when used with homology models is unknown.

1. Run the following command to refine the given protein structure XXXXX.pdb:

relax.static.linuxgccrelease -in:file:s XXXXX.pdb -relax:constrain_relax_to_start_coords -out:suffix _relaxed -out:no_nstruct_label -relax:ramp_constraints false

where relax.static.linuxgccrelease is the binary executable of the Rosetta FastRelax protocol for relaxing the given protein structure with the Rosetta energy function. This command will generate a PDB file named XXXXX_relaxed.pdb, which will be used in later steps.

2. Run the following command to create a structural model for each of the variants in the given list:

rosetta_relax.py --rosetta-bin relax.static.linuxgccrelease -l VARIANT_LIST --base-dir /path/to/where/all/XXXXX_relaxed.pdb/is/stored

where VARIANT_LIST is a given file in which each line is a given variant in the format XXXXX POS WT MUT. For example, 2ocjA 282 R W.

3. Run the following command to generate tensors (parameterized 3D voxel grids), which will be the input of the 3D deep convolutional neural networks:

gends.py -i VARIANT_LIST -o test_direct_stacked_16_1 -p /path/to/where/all/XXXXX_relaxed.pdb/is/stored --boxsize 16 --voxelsize 1

This command will generate a file called test_direct_stacked_16_1.npy that stores the parameterized 3D voxel grids for each variant in the given list. This file will be used as input to the ensemble of 3D deep convolutional neural networks. If the purpose is to make predictions for reverse mutations, then add the --reverse flag to the above command so that it reads as follows

gends.py -i VARIANT_LIST -o test_reverse_stacked_16_1 -p /path/to/where/all/XXXXX_relaxed.pdb/is/stored --boxsize 16 --voxelsize 1 --reverse

4. Run predict.py:

for i in `seq 1 10`; do predict.py -x test_direct_stacked_16_1.npy -m models/ThermoNet_ensemble_member_${i}.h5 -o test_predictions_${i}.txt; done

In the following, we outline the steps to predict $\Delta\Delta G$ for a list of single-point mutations of the p53 tumor suppressor protein.

1. Change the directory to examples where you'll find the PDB file 2ocjA.pdb and a file named p53_mutations.txt which contains a list of p53 single-point mutations.

2. Run the Rosetta FastRelax application to relax the PDB structure. The purpose of doing this is to remove potential energetic frustrations (clashes, bad residue geometries, etc.) in the PDB structure and to create a good reference state for creating mutant structures.

relax.static.linuxgccrelease -in:file:s 2ocjA.pdb -relax:constrain_relax_to_start_coords -out:suffix _relaxed -out:no_nstruct_label -relax:ramp_constraints false

This step usually takes a few minutes depending on the size of the protein (it may take hours to relax a large protein, for example, one that has >500 residues) and will generate a file named 2ocjA_relaxed.pdb. If your protein is very big, say 1000+ residues, you can add the flag -default_max_cycles 200 to make this relax run shorter.

3. Now create a directory called 2ocjA and move 2ocjA_relaxed.pdb to this directory.

mkdir 2ocjA
mv 2ocjA_relaxed.pdb 2ocjA

4. Run the following command to create a structural model for each of the variants in the given list.

rosetta_relax.py --rosetta-bin relax.static.linuxgccrelease -l p53_mutations.txt --base-dir ./

This step will create a PDB file for each mutation in the given list. For example, it will create a file named 2ocjA_relaxed_Q104H_relaxed.pdb for the mutation 2ocjA 104 Q H.

5. Run the following command to rename all PDB files.

rename _relaxed_ _ *_relaxed_*.pdb

The purpose of renaming the files is to make the file names conform to the requirement of the gends.py utility.

6. Generate a dataset containing the tensor for each mutation. Note that you'll need to run this step in the HTMD conda environment.

gends.py -i p53_variants.txt -o p53_direct_stacked_16_1 -p ./ --boxsize 16 --voxelsize 1

7. Run predict.py to make $\Delta\Delta$ predictions.

for i in `seq 1 10`; do predict.py -x p53_direct_stacked_16_1.npy -m ../models/ThermoNet_ensemble_member_${i}.h5 -o p53_predictions_${i}.txt; done

8. Now take the average of the predictions from these 10 models as the final $\Delta\Delta G$ prediction.

1. ThermoNet

• Li, B., Yang, Y.T., Capra, J.A., and Gerstein, M.B. (2020). Predicting changes in protein thermodynamic stability upon point mutation with deep 3D convolutional neural networks. PLoS Comput Biol 16, e1008291.

2. 3D convolutional neural networks

• Torng, W., and Altman, R.B. (2017). 3D deep convolutional neural networks for amino acid environment similarity analysis. BMC Bioinformatics 18, 302
• Jimenez, J., Skalic, M., Martinez-Rosell, G., and De Fabritiis, G. (2018). K-DEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks. Journal of Chemical Information and Modeling 58, 287-296

3. Rosetta

• Leaver-Fay, A., Tyka, M., Lewis, S.M., Lange, O.F., Thompson, J., Jacak, R., Kaufman, K., Renfrew, P.D., Smith, C.A., Sheffler, W., et al. (2011). ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules. Methods Enzymol 487, 545-574
• Tyka, M.D., Keedy, D.A., Andre, I., Dimaio, F., Song, Y., Richardson, D.C., Richardson, J.S., and Baker, D. (2011). Alternate states of proteins revealed by detailed energy landscape mapping. J Mol Biol 405, 607-618

4. Protein folding and thermostability

• Li, B., Fooksa, M., Heinze, S., and Meiler, J. (2018). Finding the needle in the haystack: towards solving the protein-folding problem computationally. Crit Rev Biochem Mol Biol 53, 1-28
• Guerois, R., Nielsen, J.E., and Serrano, L. (2002). Predicting changes in the stability of proteins and protein complexes: A study of more than 1000 mutations. Journal of Molecular Biology 320, 369-387

5. HTMD

• Doerr, S., Harvey, M.J., Noe, F., and De Fabritiis, G. (2016). HTMD: High-Throughput Molecular Dynamics for Molecular Discovery. Journal of Chemical Theory and Computation 12, 1845-1852