Application of Molecular Descriptors in the In Vivo Genotoxicity Prediction Using Machine Learning

Resumo

Genotoxicity represents a significant public health concern, being associated with mutations and genomic instability. In vivo assays are considered more representative of biological complexity and are essential for predictive risk assessment. The use of molecular descriptors in computational models offers an efficient alternative to traditional experimental testing, contributing to savings in both time and resources. The objective of this paper is to evaluate the effectiveness of classical molecular descriptors in predicting in vivo genotoxicity using machine learning algorithms and robust performance metrics.A total of 2,223 substances with in vivo results were selected from public databases (GENE-TOX, CCRIS, and ECHA). Compounds were represented by SMILES strings, with structures verified and harmonized. A total of 366 molecular descriptors were calculated using the RDKit library, encompassing topological indices, electronic properties, structural complexity metrics, and features related to three-dimensional conformation. Additionally, structural alerts recognized by regulatory agencies were identified through substructure matching. Models were trained and evaluated using stratified 10-fold cross-validation, employing Random Forest, ExtraTrees, and LGBMClassifier algorithms. The primary metric was the F1-score, complemented by accuracy and ROC-AUC. Models trained exclusively with in vivo data and using only molecular descriptors demonstrated consistent performance. The ExtraTreesClassifier model achieved an F1-score of 80.4%, accuracy of 80.6%, and ROC-AUC of 86.0%. The molecular descriptors alone proved highly effective in predicting in vivo genotoxicity. These findings highlight the relevance of in vivo data and classical descriptors as reliable tools in predictive toxicity models, supporting advances in the screening of safe compounds for pharmaceutical and environmental applications.