Comprehensive Benchmark Study of Diffusion-Based 3D Molecular Generation Models.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep diffusion models are advanced tools for de novo molecular generation in 3D.
• These models use forward noising and reverse denoising to learn molecular geometries and properties.
• Systematic performance comparisons of these models are limited.

Purpose of the Study:

• To benchmark nine state-of-the-art diffusion-based 3D molecular generative models.
• To evaluate model performance on both 2D structural and 3D geometric metrics.
• To identify challenges and advances in deep diffusion models for 3D molecular generation.

Main Methods:

• Comprehensive benchmark of nine diffusion-based 3D molecular generative models.
• Models trained on QM9 and GEOM-Drugs datasets.
• Evaluation using four metric types: 2D structural and 3D geometric features.

Main Results:

• Most models perform significantly worse on 3D metrics than 2D metrics.
• Generated 3D structures often deviate from energy-minimized references, indicating challenges in spatial modeling.
• Model effectiveness decreases for larger, more complex molecules.
• MiDi and EQGAT-diff demonstrated superior performance, with MiDi being particularly robust.

Conclusions:

• Deep diffusion models show promise for 3D molecular generation but face challenges in geometric accuracy.
• Current models struggle with precise 3D spatial modeling, especially for complex molecules.
• MiDi and EQGAT-diff represent leading approaches, but further advancements are needed for next-generation models.