Diagnosis with Nanoscale Protein Distributions: Single-Molecule Fluorescence Localization Microscopy and Attention-Based Learning

Single-molecule (fluorescence) localization microscopy (SMLM) finds the position of markers for target proteins at approx. 10 nm precision. Diagnosis of some diseases currently relies on inspection of nanoscale morphology by electron microscopy (EM), an expensive and slow test with limited sample coverage. Nanoscale biological processes also underlie health and disease in general, and so there is a need for more efficient diagnostic methods. We demonstrate that SMLM of routine biopsy samples can be used to assist diagnosis via data classification models. We predict diagnosis of 20 patients with chronic renal diseases (focal segmental glomerulosclerosis or minimal change disease) with a mean area under the receiver operating characteristic curve of 0.97 in cross-validation, and balanced accuracy of 90%. We tested state-of-the-art pretrained feature extraction from image tiles at 0.045 microns per pixel, followed by training of weakly supervised, attention-based models. SMLM and automated analysis has the potential to save time to diagnosis and costs compared with EM, with greater sample coverage, as well as for finding new nanoscale biomarkers in other disease areas.