Metabolomics combined with machine learning reveals candidate metabolites and a putative working network associated with cereulide production in emetic Bacillus cereus.

Cereulide, a heat-stable toxin produced by Bacillus cereus, is recognized as a major foodborne risk. However, the metabolic features associated with cereulide production and their biological relevance remain insufficiently characterized. This study aimed to prioritize candidate metabolites associated with cereulide-related phenotypes and to explore their potential pathway-level relationships. An integrated analytical pipeline combining RF-CNN-assisted feature prioritization, pathway enrichment, time-series profiling, and temporal association analysis was developed to support both classification performance and biological interpretability. Using fivefold cross-validation, nine candidate metabolites were prioritized. Based on pathway enrichment and temporal analyses across three growth stages, histidinol, L-leucine, and biotin were highlighted as candidate metabolites associated with cereulide-related phenotypes. A putative working network describing their potential relationships was then proposed. In exogenous supplementation experiments, L-leucine and biotin were associated with reduced cereulide production, whereas histidine showed the opposite trend, providing phenotypic support for the biological relevance of the related pathways under the tested conditions. Overall, this study establishes a metabolomics-guided strategy for prioritizing candidate metabolites and provides biologically interpretable and testable hypotheses for exploring the potential metabolic basis of B. cereus-associated phenotypes.