TaLiRAGen: target-aware ligand generation via retrieval-augmented large language models.

Generating ligands with high binding affinity for specific protein targets is a fundamental challenge in structure-based drug design. While generative models like VAEs and diffusion models show promise, they suffer from two key limitations: heavy dependence on target-specific training data and underutilization of vast biochemical knowledge. To overcome these issues, we propose TaLiRAGen, a no-training framework for target-aware ligand generation. Our method leverages general chemical knowledge embedded in Large Language Models (LLMs) through Retrieval-Augmented Generation (RAG), eliminating the need for target-specific model training. TaLiRAGen retrieves protein-ligand contexts from diverse repositories, integrates them through Chain-of-Thought (CoT)-augmented multi-turn prompting, and refines molecules via docking feedback. To enhance LLM’s ability in evaluating generated ligands, a unified evaluation metric is developed using evidence-theoretic normalization, jointly assessing binding affinity and drug-like properties. Critically, TaLiRAGen enables flexible, prompt-driven customization, successfully generating ligands that meet diverse structural requirements (e.g., LogP, ring count) for therapeutic targets. The codes and models are available in the supplement materials.