ProLaTherm: Protein Language Model-based Thermophilicity Predictor

In this repository, we publish four parts related to protein thermophilicity prediction:

• Code to get predictions using ProLaTherm, the first protein language model-based thermophilicity predictor
• A new benchmark dataset for protein thermophilicity prediction consisting of three significantly updated datasets from literature and newly collected data
• Code to run our protein thermophilicity prediction optimization framework including several feature- and sequence-based prediction models
• Results of the hyperparameter optimization we conducted with respect to the publication given below

Predict using ProLaTherm

If you want to get thermophilicity predictions for your protein sequences using ProLaTherm, you first have to prepare a .fasta file containing sample ids and amino acid sequences. We expect the standard fasta format with the sample id, and the amino acid sequence following in the next line, see prolatherm/assets/dummy_data.fasta for an example.

Then you have two possibilities to run ProLaTherm:

• Create a Docker container using the Dockerfile we provide and run ProLaTherm within this tested working environment
• Run ProLaTherm directly on your machine (e.g. within a virtual environment to prevent Python package version changes) after installing all required packages using our requirements.txt

Both possibilities are outlined below. In either case, make sure to have a minimum of 20GB memory. As you only run inference, a GPU is not necessary, but if you want to use one it should have at least 16GB memory. Using the provided .fasta-file, the pipeline will run on the command line and create a .csv-file containing the results.

Docker workflow

Docker needs to be installed and running on your machine, see the Installation Guidelines at the Docker website On Ubuntu, you can use docker run hello-world or docker --version to check if Docker works (Caution: add sudo if you are not in the docker group).

If you want to use GPU support, you need to install nvidia-docker-2, see this Installation Guide. and a version of CUDA >= 11.2 (see this CUDA Installation Guide. To check your CUDA version, just run nvidia-smi in a terminal.

1. Open a Terminal and navigate to the directory where you want to set up the project

2. Clone this repository

    git clone https://github.com/grimmlab/ProLaTherm.git
   

3. Navigate to prolatherm/Docker after cloning the repository

    cd ProLaTherm/prolatherm/Docker
   

4. Build a Docker image using the provided Dockerfile tagged with the IMAGENAME of your choice

    docker build -t IMAGENAME .
   

5. Run an interactive Docker container based on the created image with a CONTAINERNAME of your choice

    docker run -it -v /PATH/TO/REPO/FOLDER:/REPO_DIRECTORY/IN/CONTAINER -v /PATH/TO/DATA/DIRECTORY:/DATA_DIRECTORY/IN/CONTAINER -v /PATH/TO/RESULTS/SAVE/DIRECTORY:/SAVE_DIRECTORY/IN/CONTAINER --name CONTAINERNAME IMAGENAME
   

   • Mount the directory where the repository is placed on your machine, the directory where your phenotype and genotype data is stored and the directory where you want to save your results using the option -v.
   • You can restrict the number of cpus using the option cpuset-cpus CPU_INDEX_START-CPU_INDEX_STOP.
   • Specify a gpu device using --gpus device=DEVICE_NUMBER if you want to use GPU support.

   Let's have a look at an example. We assume hat you created a Docker image called prolatherm_image, your repository and data is placed in (subfolders of) /myhome/, you want to save your results to /myhome/ (so /myhome/ is the only directory you need to mount in your container), you only want to use CPUs 0 to 9 and GPU 0 and you want to call your container prolatherm_cont. Then you have to run the following command:

    docker run -it -v /myhome/:/myhome_in_my_container/ --cpuset-cpus 0-9 --gpus device=0 --name prolatherm_cont prolatherm_image
   

6. Navigate to the directory where the repository is placed within your container and to the prolatherm subfolder

    cd /REPO_DIRECTORY/IN/CONTAINER/ProLaTherm/prolatherm
   

7. Run ProLaTherm with giving the full path to the .fasta-file (default: prolatherm/assets/dummy_fasta.csv) and directory where you want to save your results file (default: repository folder)

    python3 run_prolatherm.py -df /FULL/PATH/TO/FASTA/FILE -sd /FULL/PATH/TO/RESULTS/SAVE/DIR
   

In case you have problems in getting inference running on your GPU, we provide an option --no_gpu, which you can set to True when calling python3 run_prolatherm.py

That's it! The .fasta-file will be processed in batches of 10 samples, you will see the current status on the command line and in the end a .csv-file containing the results will be created.

Run ProLaTherm directly on your machine

Instead of using Docker, you can run ProLatherm directly on your machine. We recommend to set up a Python virtual environment for that purpose, see here: https://docs.python.org/3/library/venv.html ProLaTherm was developed and tested in Python 3.8, so please use this Python version

For our tutorial below, we assume that you know how to work with a virtual environment and are working within such an environment (or you decided against it).

1. Open a Terminal and navigate to the directory where you want to set up the project

2. Clone this repository

    git clone https://github.com/grimmlab/ProLaTherm.git
   

3. Navigate to prolatherm/Docker after cloning the repository

    cd ProLaTherm/prolatherm/Docker
   

4. Install all required Python packages using our requirements.txt

    pip3 install -r requirements.txt
   

5. Navigate to the prolatherm subfolder of your repository

    cd ..
   

6. Run ProLaTherm with giving the full path to the .fasta-file (default: prolatherm/assets/dummy_fasta.csv) and directory where you want to save your results file (default: repository folder)

    python3 run_prolatherm.py -df /FULL/PATH/TO/FASTA/FILE -sd /FULL/PATH/TO/RESULTS/SAVE/DIR
   

In case you have problems in getting inference running on your GPU, we provide an option --no_gpu, which you can set to True when calling python3 run_prolatherm.py

That's it! The .fasta-file will be processed in batches of 10 samples, you will see the current status on the command line and in the end a .csv-file containing the results will be created.

Data

Here, we publish our new benchmark dataset for protein thermophilicity prediction consisting of significantly updated datasets from literature and newly collected data. For details on the data collection and preprocessing, see our publication below.

Within the folder data, one can find three subfolders:

• datasets_w_datasplits: preprocessed and feature-engineered datasets used for our scientific paper (see below) as .csv-files and so-called index_files that contain indices for our datasplits to ensure reproducibility. Both files are generated by our optimization pipeline, when providing fasta files containing thermophilic and non-thermophilic proteins
• fasta_files: .fasta-files containing our benchmark data split in files containing thermo and non_thermo species, both as _fulldata and _preprocessed (CD-Hit with threshold of 40%, removing sequences with lengths below and above 5th and 95th percentile, see our publication for details)
  • Files with prefix ProtThermoPred_shared_organisms_: protein sequences of species that occur within the three previously published datasets and our newly collected data
  • Files with prefix ProtThermoPred_unique_organisms_: protein sequences of species that only occur in our newly collected data
  • Further .fasta-files with prefix ProtThermoPred_ (not containing key words shared_organisms and unique_organisms): protein sequences of our full dataset without a species-specific split
  • Subfolder update_of_benchmark_datasets: .fasta-files containing significantly updated datasets from literature
• meta_files: .csv-files with meta information such as species and UniProt history for all datasets

Protein Thermophilicity Prediction Optimization Framework

To run our optimization framework, we highly recommend our workflow using Docker due to its easy-to-use interface and ready-to-use working environment within a Docker container.

This optimization framework is based on easyPheno, our phenotype prediction optimization framework. So in case you are interested in further details, easyPheno's comprehensive documentation including installation guides, tutorials and much more, might be helpful: https://easypheno.readthedocs.io/

Requirements for the optimization framework

Docker needs to be installed and running on your machine, see the Installation Guidelines at the Docker website On Ubuntu, you can use docker run hello-world or docker --version to check if Docker works (Caution: add sudo if you are not in the docker group).

If you want to use GPU support, you need to install nvidia-docker-2, see this Installation Guide. and a version of CUDA >= 11.2 (see this CUDA Installation Guide. To check your CUDA version, just run nvidia-smi in a terminal.

CAUTION: The models transformer and bigbird require a minimum GPU memory of 48GB.

Setup of the optimization framework

1. Open a Terminal and navigate to the directory where you want to set up the project

2. Clone this repository

    git clone https://github.com/grimmlab/ProLaTherm.git
   

3. Navigate to thermpred/Docker after cloning the repository

    cd ProLaTherm/thermpred/Docker
   

4. Build a Docker image using the provided Dockerfile tagged with the IMAGENAME of your choice

    docker build -t IMAGENAME .
   

5. Run an interactive Docker container based on the created image with a CONTAINERNAME of your choice

    docker run -it -v /PATH/TO/REPO/FOLDER:/REPO_DIRECTORY/IN/CONTAINER -v /PATH/TO/DATA/DIRECTORY:/DATA_DIRECTORY/IN/CONTAINER -v /PATH/TO/RESULTS/SAVE/DIRECTORY:/SAVE_DIRECTORY/IN/CONTAINER --name CONTAINERNAME IMAGENAME
   

   • Mount the directory where the repository is placed on your machine, the directory where your phenotype and genotype data is stored and the directory where you want to save your results using the option -v.
   • You can restrict the number of cpus using the option cpuset-cpus CPU_INDEX_START-CPU_INDEX_STOP.
   • Specify a gpu device using --gpus device=DEVICE_NUMBER if you want to use GPU support.

   Let's have a look at an example. We assume hat you created a Docker image called thermpred_image, your repository and data is placed in (subfolders of) /myhome/, you want to save your results to /myhome/ (so /myhome/ is the only directory you need to mount in your container), you only want to use CPUs 0 to 9 and GPU 0 and you want to call your container thermpred_cont. Then you have to run the following command:

    docker run -it -v /myhome/:/myhome_in_my_container/ --cpuset-cpus 0-9 --gpus device=0 --name thermpred_cont thermpred_image 
   

Your setup is finished!

Run the optimization framework with prepared data

You are at the root directory within your Docker container, i.e. after step 5 of the above-described setup.

If you closed the Docker container that you created at the end of the installation, just use docker start -i CONTAINERNAME to start it in interactive mode again. If you did not create a container yet, go back to above-described setup.

1. Navigate to the directory where the repository is placed within your container

    cd /REPO_DIRECTORY/IN/CONTAINER/ProLaTherm
   

2. Run thermpred (as module). By default, thermpred starts the optimization procedure for 10 trials with XGBoost and a 5-fold nested cross-validation using the ProtThermPred_fulldataset.csv we provide in data/datasets_w_datasplits.

    python3 -m thermpred.run --save_dir SAVE_DIRECTORY
   

   That's it! You can now find the results in the save directory you specified. By default, if you do not specify a save_dir a results folder will be created at the top directory of your repository.

3. To get an overview of the different options you can set for running thermpred, just do:

    python3 -m thermpred.run --help
   

Feel free to test thermpred, e.g. with other prediction models.

CAUTION: If you want to run the optimization for prolatherm, you first have to generate .h5-files containing the embeddings. This will need up to 20GB of memory on your machine! To generate such a file, you need to run python3 -m thermpred.generate_pretrained_embeddings from the directory where your repository is placed in the Docker container. To check the different options, e.g. to change the dataset for which we generate the embeddings by default, please run python3 -m thermpred.generate_pretrained_embeddings --help.

Run the optimization framework using your own .fasta-files

To run the optimization framework with your own data, you have to prepare two .fasta-files: one containing the thermophilic and one containing the non-thermophilic proteins. We expect the standard .fasta-format with the identifier of the sample, followed by a line break and the amino acid sequence given in the single letter code in the next line.

You are at the root directory within your Docker container, i.e. after step 5 of the above-described setup.

If you closed the Docker container that you created at the end of the installation, just use docker start -i CONTAINERNAME to start it in interactive mode again. If you did not create a container yet, go back to above-described setup.

1. Navigate to the directory where the repository is placed within your container

    cd /REPO_DIRECTORY/IN/CONTAINER/ProLaTherm
   

2. Run thermpred (as module). By default, thermpred starts the optimization procedure for 10 trials with XGBoost and a 5-fold nested cross-validation using your newly created dataset.

    python3 -m thermpred.run -dd /DIRECTORY/CONTAINING/YOUR/FASTAFILES -sd /SAVEDIRECTORY/FOR/CREATED/DATASET -fth FULL_NAME_OF_THERMOPHILIC_FASTA_FILE -fnth FULL_NAME_OF_NON_THERMOPHILIC_FASTA_FILE -nnd NAME_OF_THE_NEW_DATASET
   

You only have to invoke the data preprocessing once. For the next run, you can use the above-described workflow for prepared data, and need to specify the name of your new dataset as the dataset to use.

Useful Docker commands

The subsequent Docker commands might be useful when using thermpred. See here for a full guide on the Docker commands.

• docker images: List all Docker images on your machine
• docker ps: List all running Docker containers on your machine
• docker ps -a: List all Docker containers (including stopped ones) on your machine
• docker start -i CONTAINERNAME: Start a (stopped) Docker container interactively to enter its command line interface

Results hyperparameter optimization

We further included a detailed overview of the hyperparameter optimzation for each prediction model, both for the nested cross-validation and setting with a test set containing species not present in training and validation data described in the publication given below. There is an .xlsx-file for each phenotype which has different sheets. One sheet gives an overview on the results. Further, there is one sheet for each prediction model with the results. Beyond that, there is a sheet with the runtime overview and tried hyperparameter combination for each outerfold and prediction model.

Contributors

This pipeline is developed and maintained by members of the Bioinformatics lab lead by Prof. Dr. Dominik Grimm:

• Florian Haselbeck, M.Sc.
• Maura John, M.Sc.
• Jonathan Pirnay, M.Sc.

Citation

When using parts of this repository, please cite our publication:

Superior Protein Thermophilicity Prediction With Protein Language Model Embeddings
F Haselbeck, M John, Y Zhang, J Pirnay, JP Fuenzalida-Werner, RD Costa, and DG Grimm
NAR Genomics and Bioinformatics, 2023
https://doi.org/10.1093/nargab/lqad087

Keywords: Protein Thermophilicity Prediction, Protein Language Model, Machine Learning