Over the past decades, a plethora of research efforts was spent on helping practitioner discover defects using various software dependency structures. However, it still remains unclear whether these various dependency structures will present almost similar accuracies in quality assurances tasks, such as defect prediction and localization. We are also uncertain whether the combination of these different dependency structures can improve the prediction/localization results. In this paper, we answer this question from the view of software architecture. Three well-studied program dependencies are selected to construct the software architecture. And DRSpace, a state-of-the-art reverse engineering technique, is employed to investigate these three structures. We conducted our study on 117 Apache open projects involving 643,079 revision commits and 101,364 bug reports. The experiment results are surprising - these three structures present significant differences on the software architectural design, and playing completely different roles in defect prediction/localization. Furthermore, what excites us is that the combination of these three structures can effectively and efficiently improve the accuracy of defect prediction/localization. In summary, our work provides a new perspective to solve the data mining problem from software architecture in defect prediction/localization. Our quantitative and qualitative conclusions also advance our knowledge of the relationship between software quality and architectural design.