DeepFPlearn⁺

Link molecular structures of chemicals with multiple targets.

DeepFPlearn⁺[1] transforms chemical SMILES or InChIkeys into molecular graph representations for use in a graph neural network (GNN) that predicts an association to a target molecule, e.g., a receptor protein.

DeepFPlearn⁺ is an extension of deepFPlearn[2], which uses binary fingerprints to represent the molecule's structure computationally.

Setting up Python environment

The DFPL package requires a particular Python environment to work properly. It consists of a recent Python interpreter and packages for data-science and neural networks. The exact dependencies can be found in the requirements.txt (which is used when installing the package with pip) and environment.yml (for installation with conda).

You have several ways to provide the correct environment to run code from the DFPL package.

1. Use the automatically built docker/Singularity containers
2. Build your own container following the steps here
3. Setup a python virtual environment
4. Set up a conda environment install the requirements via conda and the DFPL package via pip

In the following, you find details for option 1., 3., and 4.

Docker container

You need docker installed on you machine.

In order to run DFPL use the following command line

docker run --gpus GPU_REQUEST registry.hzdr.de/department-computational-biology/deepfplearn/deepfplearn:TAG dfpl DFPL_ARGS

where you replace

• TAG by the version you want to use or latest if you want to use latest available version)
• You can see available tags here https://gitlab.hzdr.de/department-computational-biology/deepfplearn/container_registry/5827. In general a container should be available for each released version of DFPL.
• GPU_REQUEST by the GPUs you want to use or all if all GPUs should be used (remove --gpus GPU_REQUEST if only the CPU should)
• DFPL_ARGS by arguments that should be passed to DFPL (use --help to see available options)

In order to get an interactive bash shell in the container use:

docker run -it --gpus GPU_REQUEST registry.hzdr.de/department-computational-biology/deepfplearn/deepfplearn:TAG bash

Singularity container

You need Singularity installed on your machine. You can download a container with

singularity pull dfpl.TAG.sif docker://registry.hzdr.de/department-computational-biology/deepfplearn/deepfplearn:TAG

• replace TAG by the version you want to use or latest if you want to use latest available version)
• You can see available tags here https://gitlab.hzdr.de/department-computational-biology/deepfplearn/container_registry/5827. In general a container should be available for each released version of DFPL.

This stores the container as a file dfpl.TAG.sif which can be run as follows:

singularity run --nv dfpl.TAG.sif dfpl DFPL_ARGS

• replace DFPL_ARGS by arguments that should be passed to DFPL (use --help to see available options)
• omit the --nv tag if you don't want to use GPUs

or you can start a shell script (look at run-all-cases.sh for an example)

singularity run --nv dfpl.sif ". ./example/run-multiple-cases.sh"

It's also possible to get an interactive shell into the container

singularity shell --nv dfpl.TAG.sif

Note: The Singularity container is intended to be used on HPC cluster where your ability to install software might be limited. For local testing or development, setting up the conda environment is preferable.

Set up DFPL in a python virtual environment

From within the deepFPlearn directory call

virtualenv -p python3 ENV_PATH
. ENV_PATH/bin/activate
pip install ./

replace ENV_PATH by the directory where the python virtual environment should be created. If your system has only python3 installed -p python3 may be removed.

In order to use the environment it needs to be activated with . ENV_PATH/bin/activate.

Set up DFPL in a conda environment

To use this tool in a conda environment:

1. Create the conda env from scratch

   From within the deepFPlearn directory, you can create the conda environment with the provided yaml file that contains all information and necessary packages

   conda env create -f environment.yml

2. Activate the dfpl_env environment with

   conda activate dfpl_env

3. Install the local dfpl package by calling

   pip install --no-deps ./

Prepare data

DFPL can calculate fingerprints of chemical structures from SMILES or INCHI representation. Therefore, e.g. CSV input-files need to contain a "smiles" or "inchi" which is then used to calculate the fingerprints. There is an example CSV file in the tests/directory directory and when you're training using the DFPL package, it will load the input files and add fingerprints. You can test the conversion

import dfpl.fingerprint as fp

fp.importDataFile("tests/data/smiles.csv")

If you're data is in CSV format, has a header row, and contains a "smiles" or an "inchi" column, you can use it as input for training as it is. However, if you're data is in a different format, you can use function in the fingerprint module to import it correctly.

The tests/data/inchi.tsv contains data in TSV format without a header row which makes it impossible to identify how to import it automatically. You can use the import_function argument to tell importDataFile how it can turn your data into a Pandas DataFrame that contains, e.g. an "inchi" column. After that DFPL can calculate and add the fingerprints to the DataFrame

import pandas as pd
import dfpl.fingerprint as fp

data = fp.importDataFile(
    "tests/data/inchi.tsv",
    import_function=(lambda f: pd.read_table(f, names=["toxid", "inchi", "key"]))
)

You can store the converted data as a "pickle" file which is a binary representation of the Pandas dataframe and can be used directly as input file for the DFPL program. The advantage is that the fingerprint calculation needs to be done only once and loading these files is fast.

data.to_pickle("output/path/file.pkl")

Note that the file-extension needs to be "pkl" to be identified correctly by DFPL. Also, you might want to look at the convert_all function in the fingerprint module that we use to convert different data-files all at once.

Running deepFPlearn

Here you will find example code for running deepFPlearn in all five modes:

• train
• predict
• traingnn
• predictgnn
• convert

The input data for each of these modes can be found in the example/data folder. The pre-computed output of the train mode can be found in the assets of the release, for the predict mode it is stored in the respective example/results_predict folder. Trained models that are used in the prediction mode are stored in the models folder.

Train

The train mode is used to train models to predict the association of molecular structures to biological targets. The encoding of the molecules is done based on molecular fingerprints (we chose 2048 as the fp length). The training data contains three targets and you may train models for each using the following command:

python -m dfpl train -f example/train.json

Training with the configurations from the example/train.json file will take approximately 4 minutes on a single CPU. The trained models, training histories, respective plots, as well as the predictions on the test data are stored in the example/results_train folder as defined in the example/train.json file (this can be changed).

• configFile or -f: The path to the train.json options file that provides various configurations for training the model.
• outputDir: The directory where the trained models and other output files will be saved.
• trainAC: Whether to train an autoencoder model.
• aeType: Whether to train a deterministic or variational autoencoder
• split_type and aeSplitType: The type of data splitting used for training. Options are scaffold_balanced, random, or molecular weight.
• fpType: The type of fingerprint used.
• fpSize: The size of the fingerprint.
• fnnType: The type of Feedforward Neural Network used.
• optimizer: The optimization algorithm used.
• lossFunction: The loss function used.
• epochs: The number of training epochs.
• batchSize: The size of the training batch.
• learningRate: The learning rate for the optimizer.
• l2reg: The L2 regularization parameter.
• wabTracking or aeWabTracking: Whether to use Weights and Biases (WANDB) for tracking.
• wabTarget: The target metric for WANDB tracking.

Predict

The predict mode is used to predict the from molecular structures. Use this command to predict the provided data for prediction:

python -m dfpl predict -f example/predict.json

The compounds are predicted with the provided model and results are returned as a float number between 0 and 1.

• inputFile: The path to the input file containing the molecules to be predicted.
• outputDir: The directory where the output file will be saved.
• outputFile: The name of the output file containing the predicted values.
• ecModelDir: The directory where the autoencoder model is saved.
• ecWeightsFile: The name of the file containing the weights of the autoencoder. This is not needed for predicting with an AR model.
• fnnModelDir: The directory where the FNN model is saved.
• compressFeatures: Whether to compress the features using the autoencoder or not.
• trainAC: Whether to train a new autoencoder or use a pre-trained one.
• trainFNN: Whether to train a new FNN model or use a pre-trained one.

Traingnn

python -m dfpl traingnn -f example/traingnn.json

The traingnn mode is used to train models using a graph neural network to predict the chemical-biological associations from molecular structures. If the training data contains multiple targets the GNN will perform nultilabel classification.

The trained models, training histories, and respective plots, as well as the predictions on the test data, are stored in the example/results_traingnn folder as defined in the example/traingnn.json file (you may change this). Similar and even more options are offered via the GNN model. Go to chemprop/args.py to take a peek and set your options.

Some basic arguments include:

• data_path: Specifies the path to the dataset in CSV format. This is where the model will read the training data from.

• save_dir: The directory where the trained model and any generated output files will be saved. Make sure the directory exists or can be created.

• epochs: The number of training epochs.

• num_folds: The number of folds to use for cross-validation. Cross-validation is a resampling procedure used to evaluate machine learning models on a limited data sample.

• metric: The primary metric used for model evaluation. In this case, it's "accuracy".

• loss_function: Specifies the loss function to be minimized during training. Here, it's set to " binary_cross_entropy".

• split_type: Defines the strategy used for splitting the data into training and test sets. The "random" option will randomly split the data. The "scaffold_balanced" based on the chemical scaffold (It also logs stats of the split). The molecular_weight puts the small molecules in training and large in the test set based on their MW.

• dataset_type: Specifies the type of problem you are solving; i.e classification.

• smiles_columns: The name of the column in the CSV file that contains the SMILES strings of the molecules.

• extra_metrics: A list of additional metrics that you want to compute for model evaluation. These could include metrics like "balanced_accuracy", "(AUC)", "F1 score", "Matthews Correlation Coefficient (MCC)", " recall", "specificity", and "precision".

• hidden_size: Specifies the size of the hidden layers in the neural network. A larger size may capture more complex representations but could risk overfitting.

Predictgnn

python -m dfpl predictgnn -f example/predictgnn.json

The predictgnn mode is using trained GNN model to predict these chemical-biological target associations from molecular structures using a graph neural network.

Again some basic arguments:

• test_path: The path to the dataset containing smiles in CSV format.

• checkpoint_path: Specifies the path to the trained model checkpoint file, usually with a .pt extension. This is the model that will be used for making predictions.

• save_dir: The directory where the prediction results will be saved. Make sure the directory exists or can be created.

• saving_name: The name of the CSV file where the predictions will be saved. This file will be created in the save_dir.

Convert

The convert mode is used to convert .csv or .tsv files into .pkl files for easy access in Python and to reduce memory on disk.

References

[1] Kyriakos Soulios, Patrick Scheibe, Matthias Bernt, Jörg Hackermüller, and Jana Schor. deepFPlearn⁺: Enhancing Toxicity Prediction Across the Chemical Universe Using Graph Neural Networks. Submitted to a scientific journal, currently under review.

[2] Jana Schor, Patrick Scheibe, Matthias Bernt, Wibke Busch, Chih Lai, and Jörg Hackermüller. AI for predicting chemical-effect associations at the chemical universe level—deepFPlearn. Briefings in Bioinformatics, Volume 23, Issue 5, September 2022, bbac257, https://doi.org/10.1093/bib/bbac257

For questions or comments please reach out: jana.schor@ufz.de