A supervised graph-based deep learning algorithm to detect and quantify clustered particles.

Area of Science:

• Biophysics
• Computational Biology
• Data Science

Background:

• Characterizing membrane-embedded proteins is crucial for understanding cellular functions.
• Current methods for analyzing protein topography often involve complex biophysical and numerical techniques.
• Dynamic clustering of particles, such as protein nanoclusters, presents a significant analytical challenge.

Purpose of the Study:

• To develop an end-to-end, human intervention-independent algorithm for detecting and quantifying dynamic particle clustering.
• To leverage Graph Neural Networks (GNNs) for analyzing high-dimensional single-molecule localization microscopy (SMLM) datasets.
• To provide a faster, parameter-independent, and reusable computational tool for biophysical analysis.

Main Methods:

• An algorithm comprising two concatenated binary Graph Neural Network (GNN) classifiers was designed.
• The GNNs were trained using simulated data, making the algorithm parameter-independent.
• The GNN-based algorithm was tested on simulated data and validated using experimental fluorescence microscopy data.

Main Results:

• The GNN-based algorithm successfully detected and quantified dynamic particle clustering.
• The method demonstrated superior speed compared to existing approaches for high-dimensional SMLM datasets.
• The algorithm can be implemented on standard desktop computers and trained GNN models are reusable.

Conclusions:

• This study presents the first application of GNNs for analyzing particle aggregation.
• The developed algorithm offers an efficient and accessible tool for studying nanoscopic particles, including membrane-associated protein nanoclusters in live cells.
• The GNN approach holds significant potential for advancing the characterization of protein dynamics and organization at the nanoscale.