Density-based optimization for unbiased, reproducible clustering applied to single molecule localization microscopy

Abstract

Single molecule localization microscopy (SMLM) has provided insight into the spatial organization of molecules at length scales below the diffraction limit of visible light. In SMLM data, density-based clustering approaches have proven to be valuable tools for probing the nanoscale structure of biological molecules, although little guidance is available for evaluating the accuracy of these results, which are often strongly dependent on user-input parameters. Here, we develop an efficient implementation of density-based cluster validation (DBCV) that can quantitatively evaluate clustering performance in SMLM-sized datasets without ground truth knowledge. We demonstrate that maximizing DBCV scores accurately identifies ground truth clustering in noisy, simulated datasets. By coupling DBCV score maximization with Bayesian optimization, we outline an optimization method, DBOpt, that selects unbiased input parameters for density-based clustering algorithms. We demonstrate that optimal input parameters can be selected for popular algorithms (DBSCAN, HDBSCAN, OPTICS) with minimal user input. Lastly, we show that DBOpt reports accurate feature sizes in 2D and 3D experimental datasets. Taken together, we propose an analysis pipeline that can be applied to a diverse array of experimental data that will improve the integrity and quality of cluster analyses in the broader scientific community.