This article introduces an interaction-aware 3D molecular generative framework, DeepICL, for interaction-guided drug design. The framework leverages universal patterns of protein-ligand interactions as prior knowledge to enable generalizable drug design. It addresses the limitations of existing generative models that struggle with generalization due to limited data, often resulting in less innovative designs with unfavorable interactions for unseen targets. DeepICL uses a 3D conditional generative model to design ligands that fulfill specific interaction conditions, enabling structure-based drug design for any protein. The framework consists of two main stages: (1) interaction-aware condition setting and (2) interaction-aware 3D molecular generation. In the first stage, interaction conditions are determined by analyzing protein atoms of a given binding site. These conditions are used to guide the generation of ligands in the second stage, where atoms are sequentially added inside a protein pocket based on the interaction condition. The model is trained using a small set of ground-truth crystal structures from the PDBbind database, demonstrating its ability to generalize from limited data. The framework was tested on various tasks, including ligand elaboration and de novo ligand design. The results showed that DeepICL can generate ligands with high binding stability, affinity, and diversity. The model was also able to design ligands that selectively bind to specific mutant proteins, such as a mutant Epidermal Growth Factor Receptor (EGFR), while sparing the wild-type EGFR. This demonstrates the framework's ability to design ligands with desired interaction patterns. The study highlights the importance of incorporating prior knowledge in deep learning models for data-deficient scientific problems. By leveraging the universal nature of protein-ligand interactions, DeepICL achieves high generalizability and can generate ligands with desired properties. The framework's ability to design ligands with specific interaction patterns and its success in generating novel scaffolds and structures with high synthetic accessibility make it a promising tool for structure-based drug design.This article introduces an interaction-aware 3D molecular generative framework, DeepICL, for interaction-guided drug design. The framework leverages universal patterns of protein-ligand interactions as prior knowledge to enable generalizable drug design. It addresses the limitations of existing generative models that struggle with generalization due to limited data, often resulting in less innovative designs with unfavorable interactions for unseen targets. DeepICL uses a 3D conditional generative model to design ligands that fulfill specific interaction conditions, enabling structure-based drug design for any protein. The framework consists of two main stages: (1) interaction-aware condition setting and (2) interaction-aware 3D molecular generation. In the first stage, interaction conditions are determined by analyzing protein atoms of a given binding site. These conditions are used to guide the generation of ligands in the second stage, where atoms are sequentially added inside a protein pocket based on the interaction condition. The model is trained using a small set of ground-truth crystal structures from the PDBbind database, demonstrating its ability to generalize from limited data. The framework was tested on various tasks, including ligand elaboration and de novo ligand design. The results showed that DeepICL can generate ligands with high binding stability, affinity, and diversity. The model was also able to design ligands that selectively bind to specific mutant proteins, such as a mutant Epidermal Growth Factor Receptor (EGFR), while sparing the wild-type EGFR. This demonstrates the framework's ability to design ligands with desired interaction patterns. The study highlights the importance of incorporating prior knowledge in deep learning models for data-deficient scientific problems. By leveraging the universal nature of protein-ligand interactions, DeepICL achieves high generalizability and can generate ligands with desired properties. The framework's ability to design ligands with specific interaction patterns and its success in generating novel scaffolds and structures with high synthetic accessibility make it a promising tool for structure-based drug design.