The paper introduces an interaction-aware 3D molecular generative framework, DeepICL, designed to enhance the generalizability of deep generative models in drug design. The framework leverages universal patterns of protein-ligand interactions as prior knowledge to improve the performance of generative models, which often struggle with limited data and produce less innovative designs with unfavorable interactions for unseen target proteins. DeepICL is structured into two main stages: interaction-aware condition setting and interaction-aware 3D molecular generation. It uses a variational auto-encoder (VAE) architecture to sequentially add ligand atoms based on the 3D context of a binding pocket and predetermined interaction conditions. The framework is evaluated through various metrics, including binding pose stability, affinity, geometric patterns, diversity, and novelty, demonstrating its effectiveness in designing ligands for unseen targets. Additionally, the framework is applied to site-specific interaction conditioning, successfully designing ligands that selectively bind to mutant proteins while sparing wild-type proteins. The results highlight the importance of incorporating prior knowledge in deep generative models to improve their generalizability and applicability in structure-based drug design.The paper introduces an interaction-aware 3D molecular generative framework, DeepICL, designed to enhance the generalizability of deep generative models in drug design. The framework leverages universal patterns of protein-ligand interactions as prior knowledge to improve the performance of generative models, which often struggle with limited data and produce less innovative designs with unfavorable interactions for unseen target proteins. DeepICL is structured into two main stages: interaction-aware condition setting and interaction-aware 3D molecular generation. It uses a variational auto-encoder (VAE) architecture to sequentially add ligand atoms based on the 3D context of a binding pocket and predetermined interaction conditions. The framework is evaluated through various metrics, including binding pose stability, affinity, geometric patterns, diversity, and novelty, demonstrating its effectiveness in designing ligands for unseen targets. Additionally, the framework is applied to site-specific interaction conditioning, successfully designing ligands that selectively bind to mutant proteins while sparing wild-type proteins. The results highlight the importance of incorporating prior knowledge in deep generative models to improve their generalizability and applicability in structure-based drug design.