Benchmarking explainable AI methods for toxicophore detection and toxicity prediction

Abstract

Recent studies have reported inconsistent behavior across explainable AI (XAI) methods in molecular property prediction, raising concerns about their reliability. This work investigates whether such inconsistencies arise from the XAI methods themselves or from the accuracy of the underlying predictive model. A high-accuracy model was first trained on deterministic functional-group labels, where all evaluated XAI methods consistently highlighted the correct atoms corresponding to the true structural motifs. The analysis was extended to mutagenicity prediction, where the methods again identified known toxicophores and chemically meaningful scaffolds. Model performance was then systematically degraded by introducing controlled amounts of label noise. As predictive accuracy decreased, agreement between XAI methods weakened gradually, and the highlighted features became less chemically relevant. When accuracy reached around 0.65, this trend changed, with a much sharper loss of agreement, indicating an explainability cliff. These findings underline the importance of assessing model accuracy before drawing conclusions from XAI outputs.