MolCluster: An Unsupervised Framework for Multiscale Molecular Representations with Physically Consistent Resolution...

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Traditional coarse-grained (CG) modeling faces challenges with diverse chemical structures and fixed mapping rules.
• Supervised CG methods are limited by small labeled datasets and lack of resolution control for multiscale modeling.

Purpose of the Study:

• To develop an unsupervised model, MolCluster, for extracting customizable coarse-grained representations.
• To enable precise control over mapping resolution for diverse molecular systems.

Main Methods:

• Integration of graph neural networks and community detection for unsupervised CG representation extraction.
• Implementation of a predefined group pair loss for target group preservation.
• Utilization of a bisection strategy for customizable resolution control.

Main Results:

• MolCluster demonstrates superior performance in CG mapping and bead type prediction compared to traditional and supervised models on the MARTINI2 dataset.
• The label-free pretraining strategy significantly enhances MolCluster's effectiveness.

Conclusions:

• MolCluster offers a novel, unsupervised approach for chemically consistent and customizable CG mapping.
• This model shows significant potential for applications in polymers, proteins, and complex multiscale systems.