Machine Learning-based Prediction of Acute Oral Toxicity Across Structurally Diverse Pharmaceutical Compounds Using Molecular Descriptors
Year: 2024
Type: article
Abstract: Acute oral toxicity prediction is crucial for the initial phase of pharmaceutical safety testing and to minimize the reliance on costly experimental testing. This study identified, designed and tested machine learning models for the prediction of LD₅₀ for structurally diverse compounds using molecular descriptors. The analytical data set was divided into a training set and an independent test set after the data were preprocessed, the repeated molecular structures were removed, and molecular structures were consolidated. Multiple Linear Regression, Random Forest, Support Vector Regression, Gradient Boosting, XGBoost and k-Nearest Neighbours methods were evaluated and compared with five-fold cross-validation and test-set evaluation. The overall best performance was obtained by the Random Forest model that had an independent test R² of 0.558, RMSE of 0.630 and MAE of 0.456. PC analysis verified the chemical space diversity of the whole data set, where the first two principal components accounted for 52.62% of the variance. The permutation importance values selected SlogP_VSA12, MaxPartialCharge, MinEStateIndex, NumHeteroatoms, and MinPartialCharge as important predictors. The results presented suggest that machine learning methods based on descriptors can be useful for acute toxicity screening, but that there is a need for external validation and incorporation of biological, mechanistic, and assay-based information into future studies to achieve good prediction accuracy. The framework also placed a strong focus on validation, interpretability and applicability domain analyses to ensure greater reliability for chemically diverse compounds.
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: diamond
OpenAlex ID: W7223601503