MARD-Mol: A Hybrid Autoregressive-Diffusion Paradigm for Coarse-Grained Molecular Modeling.

Abstract

Deep generative models have transformed drug molecule generation. However, molecules exhibit complex hierarchical structures, requiring models to simultaneously balance macroscopic topological coherence and microscopic chemical self-consistency. Although autoregressive and discrete diffusion paradigms are highly complementary, integrating their advantages within a unified architecture remains severely limited by traditional "atom-by-atom" fine-grained modeling. We propose MARD-Mol, a hybrid autoregressive-diffusion framework based on motif-inspired units. By elevating the representation granularity from atoms to motif-inspired units and introducing a dual-stream hierarchical attention mechanism, it couples inter-unit autoregressive global scaffold planning with intra-unit discrete diffusion generation. To support goal-directed drug discovery, we reformulate property optimization into an iterative "diagnose-and-repair" process, enabling targeted optimization of defective motifs while preserving the global scaffold. Extensive experiments demonstrate that MARD-Mol achieves an 86.0% Quality score in de novo generation and exhibits superior performance in fragment-constrained and multi-objective optimization, establishing a new paradigm for high-quality drug design. The source code and datasets used in this study are available at GitHub: https://github.com/CSUBioGroup/MARD-Mol.