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

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Deep generative models are revolutionizing drug discovery.
• Molecules possess complex hierarchical structures requiring models to balance global topology and local chemistry.
• Existing models struggle to integrate autoregressive and diffusion methods due to fine-grained atom-by-atom generation.

Purpose of the Study:

• To develop a unified framework integrating autoregressive and diffusion models for drug molecule generation.
• To address the limitations of atom-by-atom modeling by using higher-level representations.
• To enable goal-directed drug discovery through property optimization.

Main Methods:

• Proposed MARD-Mol, a hybrid autoregressive-diffusion framework utilizing motif-inspired units.
• Introduced a dual-stream hierarchical attention mechanism for coupled scaffold planning and generation.
• Implemented an iterative 'diagnose-and-repair' process for targeted property optimization.

Main Results:

• MARD-Mol achieved an 86.0% Quality score in de novo molecule generation.
• Demonstrated superior performance in fragment-constrained and multi-objective optimization tasks.
• Established a new paradigm for high-quality drug design.

Conclusions:

• MARD-Mol effectively integrates autoregressive and diffusion models using motif-inspired units.
• The framework enables efficient and targeted optimization for drug discovery.
• Represents a significant advancement in generative models for high-quality molecule design.