High-quality real-world data is a critical element for training accurate machine learning models. Challenges such as privacy concerns and information constraints often result in incomplete or insufficient datasets, limiting the ability of models to fully capture underlying behaviors. Generative models have demonstrated great potential in synthesizing high-quality data, particularly for tabular datasets. Nevertheless, due to the complexity of subtle relationships between attributes in tabular data, generating high-quality synthetic tabular data remains a significant challenge. Inspired by computer-aided drug design, this chapter proposes a novel generative model, TabBMG, which integrates inter-attribute relationships within tabular data. It also introduces a unified attribute binning method to handle mixed data types. Experimental results demonstrate that data generated by TabBMG, when combined with graph neural networks, achieves superior classification performance compared to state-of-the-art methods.