Abstract
Recent studies have primarily focused on introducing novel frameworks to enhance the predictive power of toxicity prediction models by refining molecular representation methods and algorithms. However, these methods are inherently complex and often pose challenges in understanding and explaining, leading to barriers in their regulatory adoption and validation. Therefore, it is necessary to select the optimal model, considering not only model performance but also interpretability. This study aimed to identify the optimal combination of molecular fingerprints (pattern-based versus algorithm-based) and machine learning algorithms (simple versus complex) for developing explainable toxicity prediction models through an comprehensive investigation of the ToxCast/Tox21 bioassay data set. For 1092 ToxCast/Tox21 assays, five molecular fingerprints (MACCS, Morgan, RDKit, Layered, and Patterned) and six algorithms (MLP, GBT, Random Forest, kNN, Logistic Regression, and Naïve Bayes) were used to train the models. Results showed that 35 models revealed acceptable performance (F1 score or accuracy is 0.8 or higher). Among the combinations, either MACCS or Morgan, paired with Random Forest, demonstrated robust performance compared with other molecular fingerprints and algorithms. MACCS and Random Forest are valuable, even when prioritizing interpretability. Consequently, the MACCS-Random Forest combination model based on four assays, targeting G protein-coupled receptor and kinase, were identified and they can be used to discern specific structural features or patterns in chemical compounds, offering explainable insights into toxicity-related chemical structures. This study indicates the importance of not disregarding the utilization of simple models when assessing both predictivity and interpretability within the context of chemical feature-based Tox21 data analysis.