Achieving ultra-high-density, ultra-high-precision single molecule localization microscopy with whole-video deep learning architectures

The National Institute of General Medical Sciences awarded Xgenomes Corp. $349,280 on September 15, 2025, under the Biomedical Research and Research Training program (CFDA 93.859) to develop machine learning algorithms for single molecule localization microscopy (SMLM) video analysis.

Xgenomes will build three deliverables over the Phase I project period. In Aim 1, the recipient will create an SMLM video simulator to generate training data for machine learning models. In Aim 2, Xgenomes will develop a machine learning-based grouping algorithm that improves the speed and accuracy of analysis on existing localization data. In Aim 3, the recipient will construct a machine learning model to analyze whole SMLM videos at once, enabling detection density an order of magnitude higher than current algorithms achieve. The work will be made freely available to the academic research community.

Performance occurs in Cambridge, Massachusetts. The period of performance runs from September 15, 2025, through September 14, 2027. The award is classified as a Project Grant.

Abstract

Superresolution microscopy is an incredible family of approaches that have provided easier observability of nanometer-scale features using visible-light fluorescence which was previously limited to ~200-300 NM. The highest resolution of these techniques is single molecule localization microscopy (smlm), in which an otherwise dark field has random, sparse fluorescent flashes. These flashes may be achieved either through a photophysical process on a stationary fluorophore (storm/palm) or through a kinetic process where a mobile fluorophore is transiently fixed to a target (paint). A point-source diffraction-limited flash can be localized with precision on the order of 1-20 NM. However, the requirement for sparsity means that the fluorophore blinking rates need to be carefully optimized, the acquisition time for a video can be quite long, and most localization methods scale poorly with the number of localizations. These and other issues are challenging for routine usage of smlm, including in the single-molecule diagnostic assay and sequencing platform that xgenomes is developing. In this phase i project, xgenomes will build machine learning algorithms to analyze whole smlm videos to identify emitter positions at ultra-high precisions and ultra-high densities. In aim 1, we will create the smlm video simulator necessary to generate training data. In aim 2, we will build a machine learning-based grouping algorithm that can work with existing localization data and primarily improve the speed and accuracy of analysis. In aim 3, we will build a machine learning model to analyze whole smlm videos at once and pushing the density up an order of magnitude higher than current algorithms can achieve. This work will be made freely available for use to the academic research community. The successful completion of these aims will create a leap forward in smlm and will unlock new uses for theses technologies.