The agrochemical and pharmaceutical industries rely heavily on pathological and biochemical data from non-human mammals to assess the toxicity of new molecules and their potential effects on human health. This process is expensive and resource-intensive, often involving thousands of animals per molecule. Despite rigorous testing, significant human toxicity is sometimes only discovered in late-stage development or even after product deployment, posing serious public health risks. The identification of early biomarkers of toxicity is crucial for mitigating these risks and improving the efficiency of drug development. Toxicogenomics, applying genomic methods to predict adverse effects, is a promising approach. Advances in computing and the availability of curated datasets, such as the Toxicogenomics Project-Genomics Assisted Toxicity Evaluation System (TG-GATEs) database, provide opportunities to develop cost-effective computational methods for predicting toxicity. These methods can expedite data analysis, reduce the need for large-scale animal studies, and reduce time to market for safe products. However, challenges remain, including the high dimensionality of gene expression data and the need for robust and accurate predictive models. The large number of gene profiles generated from a limited number of samples necessitates the use of data reduction techniques. Traditional statistical methods often yield extensive gene lists unsuitable for laboratory testing. Machine learning (ML) techniques such as feature selection and classification offer a potential solution by reducing the number of variables and improving predictive accuracy. Supervised classification models have shown promise in identifying discriminative gene signatures across microarray data. This study aimed to develop a robust ML framework for feature selection, feature ranking, and predictive analysis applicable to liver toxicity, applying it to the TG-GATEs dataset and validating it using the Microarray Quality Control (MAQC)-II study.

Previous studies have explored the use of machine learning methods for predicting biological endpoints related to drug toxicity, particularly liver toxicity. However, the predictive ability of these models has been limited due to factors such as systematic noise in gene expression experiments, high numbers of features in the gene signature, low predictive performance, and poor validation of identified biomarkers. The use of supervised classification predictive models has been investigated for identifying discriminative gene signatures across multiple microarray data platforms. Existing studies have employed various machine learning techniques, but often faced challenges in achieving high predictive ability, highlighting the need for innovations in data analysis and modeling pipelines. The large number of genes involved and the inherent noise in microarray data lead to models that don't effectively generalize to new datasets and often suffer from overfitting issues. This necessitates careful feature selection and robust model validation strategies. Furthermore, a balance needs to be struck between model complexity and generality to achieve reliable predictions in the context of toxicity testing.

This study utilized gene expression data from the TG-GATEs database for male rats, encompassing 42 chemical compounds at various dose levels and time points (single dose: 3, 6, 9, and 24 h; repeat dose: 4, 8, 15, and 29 days). The Affymetrix Rat 230 2.0 microarray platform was used, providing expression values for 31,099 genes. Data normalization was performed using Robust Multi-array Average (RMA). The ethinyl estradiol (EE) dataset was used to determine the optimal dose and time point for feature selection. Clinical pathology parameters (alkaline phosphatase, total bilirubin, body weight, liver weight, triglycerides) were analyzed to identify the earliest time point showing significant liver damage (necrosis). Differential gene expression analysis was conducted using the limma package in R to identify genes significantly altered by EE exposure. Hierarchical clustering and principal component analysis (PCA) were used to group genes with similar expression patterns and visualize the data. The 24 h high-dose exposure data from the TG-GATEs dataset was then used for feature selection. The chosen feature selection methods involved three different approaches: marginal screening (Mann-Whitney, t-test, DCor), wrapper methods (Boruta, Recursive Feature Elimination (RFE) with random forest (RF) and support vector machine (SVM)), and embedded methods (RF, Elastic Net, Lasso, Ridge Regression Cross Validation (RidgeCV), and SVM). A comprehensive pipeline was implemented to integrate feature selection and classification. The top N features were selected for each method, and predictive modeling was performed using logistic regression, RF, SVM, Lasso, and ElasticNet. The MAQC-II-NIEHS dataset (GSE16716) was used as an independent validation set to evaluate model performance, utilizing the area under the curve (AUC) of the receiver operating characteristic (ROC) and F1 score as quantitative performance metrics. Tenfold cross-validation was performed, with all compounds grouped in the same fold to avoid bias. Parameter tuning for the ML algorithms was performed using GridSearchCV to optimize model performance. Performance was assessed based on AUC, F-statistics, and Matthews Correlation Coefficient (MCC).

Analysis of the EE dataset indicated that liver necrosis was the most consistent early apical change observed across different doses and time points. Hierarchical clustering of differentially expressed genes revealed distinct clusters with varying expression kinetics and functions. Based on PCA analysis and the identification of a robust gene expression program at 24 h, this time point was chosen for subsequent feature selection. A total of 31,099 genes from the 24-hour time point were analyzed to identify genes correlated to liver necrosis. Analysis of the AUC values for different numbers of features suggested that a subset of 10 genes provides optimal prediction performance. Tenfold cross-validation using the TG-GATEs dataset as a training set and the MAQC-II dataset as an independent validation set identified the top-performing feature selection and classification method combinations: Mann-Whitney paired with RF had the highest F1 score (0.91), AUC (0.91), sensitivity (0.85), specificity (0.97) and MCC (0.82), demonstrating high accuracy and predictive power. The ten genes selected by this method include: *Scly, Dcd, RGD1309534, Slc23a1, Bhmt2, Tkfc, Srebf1, Ablim3, Extl1, and Cyp39a1*. Five of these genes (*Scly, Slc23a1, Dcd, Tkfc, and RGD1309534*) were consistently among the top contributors across multiple feature selection methods (DCor, Boruta, RFE_RF, Mann-Whitney). Several of the identified genes are known to be involved in metabolic processes, detoxification, and transcriptional regulation; some are also implicated in liver carcinogenesis.

This study successfully developed and validated a machine learning-based predictive model for identifying early gene biomarkers of liver toxicity in rats. The use of multiple feature selection and classification methods, coupled with an independent validation dataset, enhanced the robustness and generalizability of the identified gene signature. The selection of 10 genes minimized the risk of overfitting, improving the practical applicability of the findings. The consistent identification of specific genes across different feature selection methods supports their importance as biomarkers for liver necrosis. The high accuracy of the prediction models, as evidenced by the AUC, F1 score, sensitivity, specificity, and MCC values, highlights the potential for translating these findings into accelerated toxicity testing protocols. The identified genes' involvement in metabolism, detoxification, and transcriptional regulation aligns with the known biological mechanisms of liver injury and carcinogenesis. The study's results suggest that the 10-gene signature could be used as a valuable tool for predicting liver necrosis—and consequently, liver cancer risk—in rodent models following chemical exposure. The identification of a small, robust set of biomarkers has significant implications for reducing the time and cost associated with conventional toxicity testing. The use of multiple feature selection and classification methods improved the model's reliability and generalizability.

This study successfully identified a 10-gene signature that accurately predicts liver necrosis in rats following 24-hour toxicant exposure. The use of multiple machine learning techniques and an independent validation dataset ensured the robustness and generalizability of the findings. The identified genes are involved in key metabolic and regulatory processes, and some are implicated in liver carcinogenesis. This signature offers a promising approach for accelerating toxicity testing and reducing reliance on extensive animal studies. Future research could focus on validating this signature in other animal models and human cell lines and exploring its predictive potential for other forms of liver injury.

This study focused solely on male rats and liver necrosis as the endpoint. The generalizability of the findings to other species, genders, and types of liver injury remains to be determined. The TG-GATEs dataset may have inherent biases that could affect the model's performance. Further studies are needed to validate the identified gene signature in other independent datasets and to evaluate its predictive accuracy under various experimental conditions.