Chemistry-Enhanced Diffusion-Based Framework for Small-to-Large Molecular Conformation Generation.

Area of Science:

• Computational Chemistry
• Machine Learning
• Structural Biology

Background:

• Predicting 3D molecular structures at the quantum chemistry level is computationally intensive.
• Existing machine learning methods struggle with large molecules, requiring significant computational resources.

Purpose of the Study:

• Introduce StoL, a novel diffusion-model-based framework for rapid, knowledge-free generation of large molecular structures.
• Enable the assembly of large molecules from small-molecule data without prior exposure to similar structures.

Main Methods:

• StoL decomposes input molecules (SMILES) into chemically valid fragments.
• A diffusion model, trained on small molecules, generates 3D structures for these fragments.
• Fragments are then assembled into diverse conformations using a LEGO-style approach.

Main Results:

• StoL achieves rapid generation of large molecular structures.
• The fragment-based strategy eliminates the need for large-molecule training data.
• Generated structures demonstrate high scalability, transferability, and broad configurational coverage, validated by DFT calculations.

Conclusions:

• StoL offers a scalable and transferable solution for predicting large molecular 3D structures.
• The knowledge-free, fragment-based diffusion model approach accelerates molecular structure generation.
• StoL provides chemically rational structures with broad configurational coverage, overcoming limitations of current methods.