Abstract
Reproductive toxicity is a concern critical to human health and chemical safety assessment. Recently, the U.S. Food and Drug Administration announced plans to assess toxicity with artificial intelligence-based computational models instead of animal studies in "a win-win for public health and ethics." In this study, we used a reproductive toxicity dataset using Simplified Molecular Input Line Entry Specifications (SMILES) to represent 1091 reproductively toxic and 1063 non-toxic small-molecule compounds. A repeated nested cross-validation procedure was applied, in which the dataset was randomly partitioned into five distinct folds in the outer loop, each time, one fold serving as the test set. In the inner loop, a similar procedure was also repeated five times, with 12.5% each time serving as the validation set. We first evaluated the performance of classical machine learning (ML) methods such as Random Forest and Extreme Gradient Boosting on predicting reproductive toxicity, using standard model evaluation metrics including accuracy score (ACC), the area under the curve (AUC) of the receiver operating characteristics curve (ROC) and F1 score. Our analyses indicate that these methods' overall results were mediocre and insufficient for high-throughput screening. To overcome these limitations, we adopted the Communicative Message Passing Neural Network (CMPNN) framework, which incorporates a communicative kernel and a message booster module. Our results show that our ReproTox-CMPNN model outperforms the current best baselines in both embedding quality and predictive accuracy. In independent test sets, ReproTox-CMPNN achieved a mean AUC of 0.946, ACC of 0.857 and F1 score of 0.846, surpassing traditional algorithms to establish itself as a new state-of-the-art model in this field. These findings demonstrate that CMPNN's deep capture of multi-level molecular relationships offers an efficient and reliable computational tool for rapid chemical safety screening and risk assessment.