Comprehensive machine learning based study of the chemical space of herbicides

Davor Orsolic, Vesna Pehar, Tomislav Smuc, Visnja Stepanic
Laboratory for Machine Learning and Knowledge Representation, Rudjer Boskovic Institute

General Information

Several modeling approaches were tested for characterisation of structure-activity relationships according to manually currated MoA/SoA and weed selectivity labels. This repository contains code for hyperparamter tuning with 10x10-fold cross-validation and training of optimized models (randomForest, XGBoost, SVM, NaiveBayes) on whole training sets.

Datasets

HRAC

HRAC dataset contains 346 synthetic herbicides available in original HRAC list, and extended sample of 163 herbicides collected from literature and free-access online databases - with additional 131 phytotoxic natural products collected from literature. Chemical space for HRAC classification approach was represented with 166-bit MACCS structural fingeprints

Weed Selectivity

Dataset used for weed selectivity inference was collected from a subset of 221 herbicides, respectively, with assigned labels. Weed selectivity datasets were all preprocessed and standardized (x - mean(x)) / sd(x) prior to upload.

Requirements

R version 3.6.3  
randomForest 4.6-14  
caret 6.0-84  
klaR 0.6-15  
xgboost 0.90.0.2

Instructions

1. Clone github repository:

git clone https://github.com/davoors/Herbicide-Classification.git

2. For HRAC or Weed-Selectivity approach go to "models" folder and run each model seperately with "Rscript" function, and pretrained models will be saved in the same folder as a .rds file:

Rscript svm_classifier.R

3. To test pretrained model over a set of test sets, run a Predict.R script - Output files will be saved in the output folder for each dedicated approach (ex: /output/svm/logD or ./logP):

Rscript svm_pred.R