Dose-Volume Histogram (DVH) prediction is fundamental in radiation therapy that facilitate treatment planning, dose evaluation, plan comparison and etc. It helps to increase the ability to deliver precise and effective radiation treatments while managing potential toxicities to healthy tissues as needed to reduce the risk of complications. This paper extends recently disclosed research findings presented on AAPM (AAPM 65th Annual Meeting $\&$ Exhibition) and includes necessary technique details. The objective is to design efficient deep learning models for DVH prediction on general radiotherapy platform equipped with high performance CBCT system, where input CT images and target dose images to predict may have different origins, spacing and sizes. Deep learning models widely-adopted in DVH prediction task are evaluated on the novel radiotherapy platform, and graph neural networks (GNNs) are shown to be the ideal architecture to construct a plug-and-play framework to improve predictive performance of base deep learning models in the adaptive setting.