Coarse-to-Fine: a Hierarchical Diffusion Model for Molecule Generation in 3D

Bo Qiang,Yuxuan Song,Minkai Xu,Jingjing Gong,Bowen Gao,Hao Zhou,Weiying Ma,Yanyan Lan

2023-05-26

Abstract:Generating desirable molecular structures in 3D is a fundamental problem for drug discovery. Despite the considerable progress we have achieved, existing methods usually generate molecules in atom resolution and ignore intrinsic local structures such as rings, which leads to poor quality in generated structures, especially when generating large molecules. Fragment-based molecule generation is a promising strategy, however, it is nontrivial to be adapted for 3D non-autoregressive generations because of the combinational optimization problems. In this paper, we utilize a coarse-to-fine strategy to tackle this problem, in which a Hierarchical Diffusion-based model (i.e.~HierDiff) is proposed to preserve the validity of local segments without relying on autoregressive modeling. Specifically, HierDiff first generates coarse-grained molecule geometries via an equivariant diffusion process, where each coarse-grained node reflects a fragment in a molecule. Then the coarse-grained nodes are decoded into fine-grained fragments by a message-passing process and a newly designed iterative refined sampling module. Lastly, the fine-grained fragments are then assembled to derive a complete atomic molecular structure. Extensive experiments demonstrate that HierDiff consistently improves the quality of molecule generation over existing methods

Biomolecules,Artificial Intelligence,Machine Learning

What problem does this paper attempt to address?

The paper aims to address several key issues in 3D molecular generation. Specifically:

1. **Problems with existing methods**: Current methods typically generate molecules at atomic resolution, ignoring intrinsic local structures such as rings, leading to poor quality of generated structures, especially when generating large molecules. Additionally, fragment-based molecular generation strategies, while promising, are difficult to adapt to non-autoregressive 3D generation tasks due to the combinatorial optimization problem.

2. **Proposed new method**: The paper proposes a Hierarchical Diffusion Model (HierDiff) that operates in a coarse-to-fine manner. It first generates a coarse-grained molecular geometry (where each node represents a fragment of the molecule), then decodes it into fine-grained fragments, and finally assembles them into a complete atomic molecular structure. This method aims to retain the validity of local fragments without relying on autoregressive modeling.

3. **Specific goals**: Improve the quality of generated molecules, particularly generating real molecules with better drug-like properties and their conformations, making them closer to real conformations. Experimental results show that HierDiff can generate high-quality molecules with more stable substructures.

In summary, the main goal of this paper is to design a new molecular generation method in 3D space to overcome the limitations of existing methods and improve the quality and realism of generated molecules.