An effective fragment-based dual conditional diffusion framework for molecular generation

TL;DR

This paper introduces a new method for generating molecules that better fits 3D protein structures while maintaining chemical validity and drug-like properties.

Contribution

A dual conditional diffusion framework that separately models scaffold and R-group generation with structural and chemical constraints.

Findings

FDC-Diff outperforms existing methods on multiple SBDD benchmarks in terms of chemical validity and spatial compatibility.

The model generates pharmacologically relevant molecules by integrating curated reaction rules and physical-chemistry refinement.

Decomposing molecule generation into scaffold and R-group stages improves both global topology and local interaction modeling.

Abstract

Fragment-based molecular generation has emerged as a promising paradigm in structure-based drug design (SBDD), deriving effective compounds with advanced properties, including chemical validity, synthetic feasibility, pharmacological relevance, etc. However, existing approaches often struggle with generating molecules which can both conform to 3D structural constraints and retain chemical plausibility. This is largely due to the fact that prior works often treat scaffolds and R-groups of molecules indiscriminately, overlooking the distinct semantic roles played by scaffolds and R-groups. Specifically, the scaffold serves as the rigid structural backbone that determines the global geometric topology and binding pose, whereas R-groups act as functional substituents responsible for fine-tuning local physicochemical interactions. Therefore, in this work, we propose fragment-based dual conditional diffusion (FDC-Diff), a novel dual conditional diffusion framework that integrates chemical priors and structural cues for fragment-based molecular generation. Unlike traditional de novo methods that generate atoms sequentially, FDC-Diff decomposes the molecule generation process into two semantically complementary stages. Given the protein pocket and an initial fragment, in the first stage, a spatially constrained scaffold is constructed to capture the global molecular topology. In the second stage, R-groups onto the obtained scaffold are elaborated to capture local semantics to further refine molecular properties. To ensure synthetic accessibility, initial fragments and scaffold-modification hierarchy are derived from curated reaction rules, and a physical-chemistry-inspired refinement step is applied to optimize final conformations. Experimental results on multiple SBDD benchmarks demonstrate that FDC-Diff achieves state-of-the-art performance in terms of comprehensive evaluations. Furthermore, our model excels at producing chemically valid, spatially compatible, and pharmacologically relevant molecules, suggesting its potential as a feasible tool for fragment-based drug design.

Figures