Predictive Modeling of DNA Damage Outcomes: Classification of Mutational Determinants Using Augmented Machine Learning Techniques

The mutational outcome of DNA damage as a direct result of constant chemical assault is governed by major factors, including the structure and nature of damage, replication, and repair machinery <i>in vivo</i>. The role of the size of the adduct, adduct-flanking bases, and the type of polymerase involved in the replication pathway is prominently seen through existing <i>in vitro</i> and <i>in vivo</i> studies. In this work, machine learning methods have been developed to predict the critical parameters for the mutational outcome of the adducts when they encounter polymerase in a particular sequence context. We carried out the analysis with three different classification models: Logistic Regression (LR), Decision Tree (DT), and Support Vector Machine (SVM). Using the literature data, mutational results of covalent DNA adducts and abasic sites were used to train the classification models. Following this, we used a generative network method with the available information on the structure of the DNA damage, polymerase, and sequence context to generate synthetic data that accurately mirrors the real data. Further, we employed an Extreme Gradient Boosting Classifier to identify the parameter that most influences the DNA mutational outcome. Metrics such as Accuracy, Sensitivity, Precision, F1 score, and AUC value have been used to evaluate the performance of classifier methods. The proposed Bootstrapped-Variational Autoencoder (BT-VAE) model enhanced the overall prediction accuracy of classifiers by 40%. The SVM model delivered the best performance across all classification metrics in predicting mutational outcomes among the three classification models evaluated. By providing the size of the carcinogen/covalent DNA adduct, polymerase, and flanking base as input, the proposed BT-VAE framework can predict the mutational outcome (match or mismatch for covalent DNA adducts and adenine or nonadenine for abasic site), an additional tool for <i>in vivo</i> and <i>in vitro</i> studies in the field of toxicology.