Bacillus cereus is responsible for food poisoning and rare but severe clinical infections. The pathogenicity of strains varies from harmless to lethal strains. However, there are currently no markers, either alone or in combination, to differentiate pathogenic from non-pathogenic strains. The objective of the study was to identify new genetic biomarkers to differentiate non-pathogenic from clinically relevant B. cereus strains.A first set of 15 B. cereus strains were compared by RNAseq. A logistic regression model with lasso penalty was applied to define combination of genes whose expression was associated with strain pathogenicity. The identified markers were checked for their presence/absence in a collection of 95 B. cereus strains with varying pathogenic potential (food-borne outbreaks, clinical and non-pathogenic). Receiver operating characteristic area under the curve (AUC) analysis was used to determine the combination of biomarkers, which best differentiate between the "disease" versus "non-disease" groups.Seven genes were identified during the RNAseq analysis with a prediction to differentiate between pathogenic and non-pathogenic strains. The validation of the presence/absence of these genes in a larger collection of strains coupled with AUC prediction showed that a combination of four biomarkers was sufficient to accurately discern clinical strains from harmless strains, with an AUC of 0.955, sensitivity of 0.9 and specificity of 0.86.These new findings help in the understanding of B. cereus pathogenic potential and complexity and may provide tools for a better assessment of the risks associated with B. cereus contamination to improve patient health and food safety.