Quantifying the extent to which points are clustered in single-molecule localization microscopy (SMLM) data is crucial for understanding molecular spatial relationships. Existing methods often struggle with large datasets, sample heterogeneity, or require subjective parameter choices. This paper presents a fast and accurate supervised machine-learning approach using neural networks trained on simulated clustered data. The network classifies millions of points from SMLM datasets, with potential for adding classifiers for different cluster subtypes. The output enables measurements of cluster area, shape, and point density. The approach is demonstrated on simulated and experimental data of kinase Csk and adaptor PAG in human T cell immunological synapses.

David J. Williamson, Garth L. Burn, Sabrina Simoncelli, Juliette Griffié, Ruby Peters, Daniel M. Davis, Dylan M. Owen