Composite Machine Learning Strategy for Natural Products Taxonomical Classification and Structural Insights

Overview

This repository contains the source data and python scripts for taxonomical multi-class classification of natural products (NPs) into five kingdoms (animal, bacteria, chromista, fungi, and plant). Training data is obtained from the LOTUS database (https://lotus.naturalproducts.net/download).

The generated results and figures are presented in the manuscript titled “Composite Machine Learning Strategy for Natural Products Taxonomical Classification and Structural Insights".

Environment

The python packages in the conda environment for performing the taxonomical classification and structural insights are provided the yaml file (environment.yml).

How to use the pre-trained model for taxonomical classification of natural products from SMILES

1. Setup environment (environment.yml).

conda env create -f environment.yml
conda activate chemenv   

2. Go to 4_quickstart/.
3. Edit (test_data.csv) to input the SMILES to be predicted.
4. Run the code under the notebook (Prediction.ipynb).
5. Access predictions from (test_data_preds.csv). Note that the example notebook runs the 3 pre-trained composite models GCNN-SVM, GCNN-LGBM, and GCNN-XGB.

Directory structure

• 0_data/: Contains part of the source data used in the analysis. The rest has been uploaded to Figshare. (https://figshare.com/projects/Composite_Machine_Learning_Strategy_for_Natural_Products_Taxonomical_Classification_and_Structural_Insights/203637)
• 1_dataset_preparation/: Contains code for data curation and reproduction of figures.
• 2_models/: Contains retrieved and organized datasets for analysis and modeling.
  • 2.1_GCNN_model/: Graph Convolutional Neural Network Model
  • 2.2_SVM_models/: Composite GCNN-SVM (Support Vector Machine) Model
  • 2.3_LGBM_models/: Composite GCNN-LGBM (Light Gradient Boosting Machine) Model
  • 2.4_XGB_models/: Composite GCNN-XGB (Extreme Gradient Boosting) Model
• 3_structural_analysis/: Contains Python scripts for data processing, visualization, and structural analysis.
  • 3.1_tsne_analysis/: t-distributed Stochastic Neighbor Embedding (t-SNE) analysis of 3 molecular fingerprints (MAP4, MPN, Last_FFN)
  • 3.2_tsne_comparison/: Python script to reproduce comparison graphs of the 3 molecular fingerprints
  • 3.3_critical_substructures/: Python script to reproduce critical substructures of the 5 kingdoms
• 4_quickstart/: Contains examples for use of pre-trained composite GCNN-SVM model.

Instructions

In the folder 0_data/

• machine learning dataset ('np_5classes.csv')
• NP Atlas test dataset ('npatlas_dataset.csv')

In the folder 1_dataset_preparation/

• Run the script 1_data_curation.py to process the lotusUniqueNaturalProduct.bson (10.6084/m9.figshare.25745637) and curate the dataframe df_curated.csv (10.6084/m9.figshare.25745325) for data exploration.
• Run the script 2_data_exploration_preparation.py to process df_curated.csv (10.6084/m9.figshare.25745325) and generate the machine learning dataset np_5classes.csv for taxonomical classification. The machine learning dataset np_5classes.csv consists of the non-isomeric SMILES of 133,092 natural products (NPs) from five different kingdoms ([animal, bacteria, chromista, fungi, plant] = [0,1,2,3,4]). The exploratory figures will also be plotted.

In the folder 2_models/2.1_GCNN_model/

• Run the script 1_train_GCNN_model.py to train a GCNN model on the machine learning dataset.
• Run the script 2_LOTUS_MPN_lastFFN_fingerprint.py to generate the MPN (lotus_MPN.csv) and last_FFN (lotus_last_FFN.csv) fingerprints (10.6084/m9.figshare.25745448) of 133,092 NPs from the LOTUS database.
• Run the script 3_NPAtlas_screening.py to classify 13,136 NPs (bacteria and fungi) from the NPAtlas dataset (npatlas_dataset.csv) using pre-trained GCNN model, predictions are saved as npatlas_preds.csv. The MPN (npatlas _MPN.csv) and last_FFN (npatlas _last_FFN.csv) fingerprints (10.6084/m9.figshare.25745421) of 13,136 NPs of are also generated.

In the folder 2_models/2.2_SVM_models/ and for LGBM and XGB models

• Run the script 1_train_model_SVM_MPN.py to construct a SVM model (model_SVM_MPN) (10.6084/m9.figshare.25745634) based the MPN fingerprints of the machine learning dataset (lotus_MPN.csv) (10.6084/m9.figshare.25745448).
• Run the script 2_train_model_SVM_last_FFN.py to construct a SVM model (model_SVM_last_FFN) (10.6084/m9.figshare.25745598) based the last_FFN fingerprints of the machine learning dataset (lotus_ last_FFN.csv) (10.6084/m9.figshare.25745448).
• Run the script 3_SVM_prediction_NPAtlas.py to classify 13,136 NPs from NPAtlas dataset (npatlas_dataset.csv) using the constructed GCNN-SVM models (model_SVM_MPN and model_SVM_last_FFN).

In the folder 3_structural_analysis/3.1_tsne_analysis/

• Run the script tsne_analysis.py to plot and compare the Kullback-Leibler (KL) divergence and Davies-Bouldin (DB) score of 3 different molecular fingerprints (MAP4, MPN and last_FFN) of the 133,092 NPs from LOTUS database. The KL divergence and DB scores are plotted as function of the perplexity parameter in t-Distributed Stochastic Neighbor Embedding (t-SNE) analysis.

In the folder 3_structural_analysis/3.2_tsne_comparison/

• Run the script tsne_map.py to t-sne mapping of 3 different molecular fingerprints (MAP4, MPN and last_FFN) of the machine learning dataset np_5classes.csv of 133,092 NPs.

In the folder 3_structural_analysis/3.3_critical_substructures/

• Run the script critical_substructures.py to plot the top ten critical substructures of each kingdom from the machine learning dataset np_5classes.csv of 133,092 NPs.

Notes

GCNN models are based off Chemprop (https://github.com/chemprop/chemprop)

Training dataset obtained from the LOTUS initiative - Adriano Rutz, Maria Sorokina, Jakub Galgonek, Daniel Mietchen, Egon Willighagen, Arnaud Gaudry, James G Graham, Ralf Stephan, Roderic Page, Jiří Vondrášek, Christoph Steinbeck, Guido F Pauli, Jean-Luc Wolfender, Jonathan Bisson, Pierre-Marie Allard (2022) The LOTUS initiative for open knowledge management in natural products research. eLife 11:e70780. https://doi.org/10.7554/eLife.70780

Literature benchmarking was done using SVM / MAP4 fingerprint architecture from the manuscript - Capecchi, A., Reymond, JL. "Classifying natural products from plants, fungi or bacteria using the COCONUT database and machine learning" J. Cheminform 13, 82 (2021).(https://github.com/reymond-group/Coconut-TMAP-SVM)