3D-SBDD meets LLM: Towards FDA-Level Drug Design

ICML 2025 Workshop FM4LS Submission24 Authors

Keywords: Large Language Model; Structure Based Drug Design

Abstract: Structure-guided molecular generation plays a key role in early drug discovery by designing compounds for specific protein targets. While recent 3D generative models improve docking scores, they often generate chemically implausible structures outside drug-like space. To address this, we introduce the Molecule Reasonable Ratio (MRR), a new metric for assessing structural plausibility. We propose the Collaborative Intelligence Drug Design (CIDD) framework, combining 3D interaction modeling with large language models (LLMs) and Chain-of-Thought reasoning. CIDD generates spatially compatible, drug-like, and synthesizable molecules. On the CrossDocked2020 benchmark, it improves metrics like QED, SA, and MRR, and raises the success rate—from 15.72% to 34.59%—by balancing binding affinity and drug-likeness, showing the power of integrating geometry with symbolic reasoning.

Submission Number: 24