The appearance of microarray technologies makes itpossible to simultaneously monitor the expression levelsfor tens of thousands of genes. Bayesian Network (BN)is an important approach for predicting gene regulatorynetworks from expression data. However, two fundamentalproblems greatly reduce the effectiveness of current BNmethods. The first problem is much less samples than genes,it makes that to find a ”best” network is very difficult.The second is the excessive computational time, the searchspace of possible Bayesian network is very large becausegenes are numerous. In this paper, we introduce a newmethod to learn Bayesian networks which combines fuzzyclustering algorithm to reduce the search space. From theview of systems biology, modularity and hierarchy are keyfeatures of biological networks. This allows us to gainglobal network by assembling local components. A localcomponent is composed of genes with same function. Thosegenes may have same or very similar expression pattern.Besides, those inter-related genes may be bridge of differentfunctional modules. Consequently, in a process of learningof local networks, the gap between the number of samplesand genes is shortened, and the search space is reduced.Ourapproach is evaluated using artificial data and expressiondata measured during the yeast cell cycle. Results demonstratethat this approach can predict regulatory networkswith significantly improved accuracy and reduced computationaltime compared with existing BN approaches.