Deep Learning-Based Event Classification of Mass Photometry Data for Optimal Mass Measurement at the Single-Molecule Level.

Abstract

Mass photometry (MP) is a powerful technique for studying biomolecular structures, interactions, and dynamics in solution. It detects and quantifies small reflectivity changes at a glass-water interface during protein (un)binding, with signals typically averaged over 100 ms. However, particle motion at the point of single-molecule measurement can compromise key metrics such as mass resolution, sensitivity, and concentration. We present a three-dimensional convolutional residual network trained via supervised learning to classify landing events based on their spatiotemporal dynamics. By analyzing 3D event thumbnails, our method isolates optimal single-molecule measurements, eliminating cumulative histogram artifacts and improving resolving power by up to a factor of 2. Validated across diverse experimental data sets, including resolved and partially resolved samples, and varying masses, concentrations, and integration times, our approach delivers robust performance under (sub)optimal conditions. Our approach also provides measurement-level data-driven feedback, facilitating high quality MP measurements in challenging scenarios.