To adapt to the hierarchical structural property of Gene Ontology, the standard Apriori algorithm is modified into a novel algorithm, RuleGO, which mines association rules of GO function classes and gene expression difference. The inputs of RuleGO are one set of differential expressed genes and another set of non-differential expressed genes, and the outputs of RuleGO are association rules linking GO function combinations to gene differential expression. Rules mined by RuleGO may guide insights into gene expression difference at the functional level, towards the clarification of the process of pathological changes or the mechanism of medicine. Both RuleGO and OntoExpress are applied to the datasets of colon cancer and adenocarcinoma, and RuleGO turned out to be more powerful to mine relevant function rules than OntoExpress. The experimental results also reveal that rules with both high significance and high support mostly involve more than one gene function classes, suggesting that considering the combination of multiple gene function classes may be more resonable in gene expression analysis than taking into account only a single gene function class.