ClusterNet: Classifying Single-Molecule Localization Microscopy Datasets with Graph-Based Deep Learning of...

Area of Science:

• Biophysics
• Computational Biology
• Microscopy

Background:

• Single-molecule localization microscopy (SMLM) provides high-resolution insights into protein organization.
• Classifying SMLM data by sample type is crucial for automated analysis but lacks robust methods for larger structures.
• Existing methods struggle with classifying complex SMLM point-cloud datasets beyond individual protein clusters.

Purpose of the Study:

• To develop an advanced deep learning pipeline for classifying SMLM point-cloud data.
• To enable automated recognition and grouping of SMLM data based on sample type.
• To address limitations in classifying large-scale protein organization structures in SMLM.

Main Methods:

• A novel graph-based deep learning pipeline was implemented for SMLM data classification.
• The pipeline integrates features from individual protein clusters and their spatial arrangement (supracluster structure).
• Explainability tools (Uniform Manifold Approximation and Projection, SubgraphX) were used to interpret classification drivers.

Main Results:

• The method achieved 99% accuracy on a benchmark DNA-PAINT dataset, outperforming previous approaches.
• High classification accuracy was demonstrated on a challenging SMLM dataset from colorectal cancer tissue.
• Analysis confirmed the significant contribution of supracluster structure to accurate sample classification.

Conclusions:

• The developed graph-based deep learning pipeline offers a powerful tool for SMLM data classification.
• This method enhances the ability to analyze protein organization differences across various sample types, including disease states.
• Understanding supracluster structure is vital for accurate SMLM data interpretation and classification.