HOLTEP-CPU

Human Organ Level Toxicity Endpoints Predictor from China Pharmaceutical University (HOLTEP-CPU) is a collection of five models that could predict the specific toxicity endpoints of given molecular structures at the human organ level.

These five models are Chemprop (which is a deep learning model), RandomForest, SVM, XGboosting and LogisticRegression (which are machine learning models). The supported toxicity endpoints are carcinogenicity, cardiotoxicity, developmental_toxicity, hepatotoxicity, nephrotoxicity, neurotoxicity, reproductive_toxicity and skin_sensitization.

In addition, a transfer learning model based on Chemprop is also deployed with the help of pkuwangsw/COVIDVS. It has advantages in predicting skin_sensitization.

This repository contains the results of:

Hu Y, Ren Q, Liu X, et al. In Silico Prediction of Human Organ Toxicity via Artificial Intelligence Methods[J]. Chemical Research in Toxicology, 2023. https://doi.org/10.1021/acs.chemrestox.2c00411

Installation

• Clone repository

git clone https://github.com/Wenying-Yu-Lab/HOLTEP-CPU.git

• Change directory

cd HOLTEP-CPU

• (Optional) Add pkuwangsw/COVIDVS as a git submodule to apply the transfer learning model later

git submodule add https://github.com/pkuwangsw/COVIDVS.git COVIDVS

• Create a new conda environment

conda create -n holtep python=3.8

• Activate the environment

conda activate holtep

• Install cuda (11.3) and cudnn

conda install cudatoolkit 11.3 cudnn

• Install PyTorch (1.12.1)

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

• Install chemprop and other dependencies

conda install -c conda-forge rdkit

pip install git+https://github.com/bp-kelley/descriptastorus

pip install chemprop

Usage

General usage

usage: python -u predict.py [-h] [-m {Chemprop,Randomforest,SVM,XGboosting,LogisticRegression}] [-f FOLD_INDEX] [-s SAVE_DIR] {carcinogenicity,cardiotoxicity,developmental_toxicity,hepatotoxicity,nephrotoxicity,neurotoxicity,reproductive_toxicity,skin_sensitization} data_path

This tool predicts the specific toxicity endpoints of given molecular structures at the human organ level.

positional arguments:
  {carcinogenicity,cardiotoxicity,developmental_toxicity,hepatotoxicity,nephrotoxicity,neurotoxicity,reproductive_toxicity,skin_sensitization}
                        Specify the toxicity endpoint for prediction from the given list.
  data_path             Specify the file containing a series of SMILES data of the structures for prediction.

optional arguments:
  -h, --help            show this help message and exit
  -m {Chemprop,Randomforest,SVM,XGboosting,LogisticRegression}, --model {Chemprop,Randomforest,SVM,XGboosting,LogisticRegression}
                        Specify the model to be used from the given list (default: Chemprop).
  -f FOLD_INDEX, --fold FOLD_INDEX
                        Specify the fold index where the model is used (default: 1, max: 10).
  -s SAVE_DIR, --save-dir SAVE_DIR
                        Specify a directory to save results (default: ./predictions).

• Examples:
  • Do not forget to activate the environment before running predictions.

  conda activate holtep

  • Example 1. Model: Chemprop from fold 1 of "cardiotoxicity".

  python -u predict.py cardiotoxicity ./model_data/cardiotoxicity/test.csv

  • Example 2. Model: Randomforest from fold 3 of "developmental_toxicity".

  python -u predict.py -m Randomforest -f 3 developmental_toxicity ./model_data/developmental_toxicity/test.csv

• Specifications
  • Specification for input data
    • A file containing a list of SMILES data (*.smi) is supported
    • A *.csv file is supported with additional restrictions:
      • The column that contains the SMILES data should be the first column of the table.
      • If the table contains headers, then the column headers containing the SMILES data should be smiles (case insensitive).
  • Specification for toxicity endpoints
    • Available toxicity endpoints:
      • carcinogenicity
      • cardiotoxicity
      • developmental_toxicity
      • hepatotoxicity
      • nephrotoxicity
      • neurotoxicity
      • reproductive_toxicity
      • skin_sensitization
    • Toxicity endpoints are case sensitive.
  • Specification for models
    • Available models:
      • Chemprop (default)
      • Randomforest
      • SVM
      • XGboosting
      • LogisticRegression
    • Model names are case sensitive.

Apply tansfer learning model

• Steps
  • Activate the environment before running predictions

    conda activate holtep

  • Create a directory to save results (skip if it exists)

    mkdir predictions

  • Change directory

    cd COVIDVS

  • Generate the descriptors

    python generatorFeatures.py ../model_data/skin_sensitization/test.csv ../model_data/skin_sensitization/test-feat.npy 0

    • The first argument: data for prediction
    • The second argument: file path to save features (can be specified arbitrarily)
  • Prediction

    python predict.py --gpu 0 --test_path ../model_data/skin_sensitization/test.csv --features_path ../model_data/skin_sensitization/test-feat.npy --preds_path ../predictions/transfer_learning.csv --checkpoint_dir ../model_data/skin_sensitization/transfer_learning/up_AT_mouse_ip_LD50/fold_0/model_0/

    • --test_path: data for prediction
    • --features_path: feature file generated in the previous step
    • --preds_path: file path to save results (can be specified arbitrarily)
    • --checkpoint_dir: directory where the model.pt file is located (replace up_AT_mouse_ip_LD50 with a desired learner)
  • Change directory back

    cd ../

• Notice
  • Transfer learning model can only be used to predict skin_sensitization.
  • Refer to pkuwangsw/COVIDVS for more details.