A novel approach to generate robust classification models to predict developmental toxicity from imbalanced datasets

Abstract

Computational models to predict the developmental toxicity of compounds are built on imbalanced datasets wherein the toxicants outnumber the non-toxicants. Consequently, the results are biased towards the majority class (toxicants). To overcome this problem and to obtain sensitive but also accurate classifiers, we followed an integrated approach wherein (i) Synthetic Minority Over Sampling (SMOTE) is used for re-sampling, (ii) genetic algorithm (GA) is used for variable selection and (iii) support vector machines (SVM) is used for model development. The best model, M3, has (i) sensitivity (SE) = 85.54% and specificity (SP) = 85.62% in leave-one-out validation, (ii) classification accuracy of the training set = 99.67%, (iii) classification accuracy of the test set = 92.59%; and (iv) sensitivity = 92.68, specificity = 92.31 on the test set. Consensus prediction based on models M3-M5 improved these percentages by 5% over M3. From the analysis of results we infer that data imbalance in toxicity studies can be effectively addressed by the application of re-sampling techniques.

TIRESIA: An eXplainable Artificial Intelligence Platform for Predicting Developmental Toxicity

“…For the feature representation of chemicals, a pool of 818 bidimensional molecular descriptors was calculated. In this respect, 212 were computed by using RDKit (2021.09.4 release), which include mostly constitutional, topological, and electro-topological indexes, and 606 from Mordred (1.2.0 version), which are autocorrelators widely employed in previous studies. , The entire list of descriptors is available as Supporting Information (File S3).…”

“…Being considered as a default method for feature selection, , the SVM (Support Vector Machine) classifier was employed to weigh feature importance within each round of a 10-fold cross-validation run. All the available 818 features were thus ranked according to their importance by SVM.…”

Review of machine learning and deep learning models for toxicity prediction

“…The in vitro testing of pregnant animals, preferably rats and rabbits, allows for the prediction of toxic effects in both the dams and their fetuses. 86,87 In addition to traditional in vivo methods, computational approaches, including ML models 2,9,17,18,88–91 and DL models, 18 have been used as alternative methods to assess several endpoints of reproductive toxicity such as sperm reduction, gonadal dysgenesis, abnormal ovulation, teratogenicity, and infertility growth retardation.…”

“…Since SVM can handle correlated descriptors and has good generalization performance, it has been widely used in the development of models for predicting toxicity of chemicals, with 304 models reported. 18,25–27,29,32,34,35,64–70,82,89,93,101,116,117…”

“…Since SVM can handle correlated descriptors and has good generalization performance, it has been widely used in the development of models for predicting toxicity of chemicals, with 304 models reported. 18,[25][26][27]29,32,34,35,[64][65][66][67][68][69][70]82,89,93,101,116,117 Decision tree (DT) is an upside-down tree-like classification and regression algorithm with the root on the top, leaf nodes at the bottom, and several layers of internal nodes in the middle. A path from the root to a leaf forms a branch which represents a series of decision rules used to classify chemicals.…”