Biomolecular assemblies, such as clusters, condensates, and aggregates, play crucial roles in cellular function. Their morphology (geometry, topology, size, shape, and internal structure) significantly impacts their properties and functions. Examples include metabolons, signalosomes, membrane-less organelles, and protein aggregates implicated in neurological disorders. Single-molecule localization microscopy (SMLM) techniques like STORM, PALM, PAINT, DNA-PAINT, and REPLOM offer nanoscale resolution but lack universally applicable analytical tools. Current analysis often relies on manual annotation or system-specific approaches, which are time-consuming, resource-intensive, and lack generalizability. Machine learning offers a potential solution, but supervised algorithms require extensive annotated data, while unsupervised methods often lack robustness and adaptability to diverse data sets. This research addresses this gap by introducing SEMORE, a novel unsupervised machine learning pipeline designed for universal and accurate analysis of super-resolution microscopy data.

Existing methods for analyzing SMLM data face several limitations. Supervised machine learning approaches, while accurate, require extensive manual annotation and are often dataset-specific. Unsupervised methods like OPTICS and DBSCAN, while more general, suffer from the limitations of a one-size-fits-all approach and often require considerable manual intervention for parameter tuning, hindering their application to the heterogeneity inherent in biological systems. Furthermore, existing methods typically do not effectively incorporate temporal information, making analysis of dynamic processes challenging. The need for a robust, generalizable method for extracting and quantifying protein assembly geometry and kinetics across various SMLM experimental setups has motivated the development of SEMORE.

SEMORE is an unsupervised machine learning pipeline comprising two main modules: a clustering module and a morphological fingerprinting module. The clustering module accepts x, y, and optionally t (time) coordinates from various SMLM techniques. It employs a multi-step density-based clustering approach (DBSCAN or HDBSCAN, chosen adaptively based on data density) to identify protein assemblies. A crucial component is the temporal refinement module, which uses time-dependent clustering to dissect spatially overlapping structures. A smart density filter removes false-positive detections. The morphological fingerprinting module then quantifies each identified assembly using over 40 features categorized into four groups: circularity, symmetry, graph network statistics, and geometric properties (Supplementary Table 1). These features capture the structural diversity of protein assemblies. The Euclidean distance is used for all distance calculations. Statistical distributions (Gaussian and Poisson) are used for data analysis and modeling. Data transformation includes standardization and min-max scaling. Performance metrics such as accuracy, precision, recall, and F1 score are used to evaluate the model. The hyperparameters of the clustering module, primarily those of DBSCAN/HDBSCAN, are largely agnostic and self-parameterize based on the input data. The spatial localizations are standardized and projected to ensure consistent feature computation. Specific methods used for simulations include Gaussian KDE for isotropic growth, Monte Carlo for steric hindrance, and a directed diffusion scheme for fibril growth. Simulated data types included isotropic, sterically driven, and fibril growth, with various noise levels and types. Experimental data included time-resolved insulin aggregation (REPLOM), dSTORM images of nuclear pore complexes (NPCs) and fibroblast growth receptor 1 (FGFR1), and live-cell PALM data of ryanodine receptors (RyRs) and Syntaxin 1a (Sx1a).

SEMORE demonstrated high accuracy in classifying simulated protein assemblies with diverse morphologies (isotropic, sterically driven, fibril) even under noisy conditions (F1 scores 84-98%). It accurately identified three types of morphologies in stress tests, significantly outperforming DBSCAN/HDBSCAN. The morphological fingerprinting module effectively separated the three simulated growth types, even capturing a fourth cluster representing noise detections, achieving >99% accuracy after noise correction (Fig. 3). Further analysis using UMAP and DBSCAN revealed the ability to distinguish between branching and non-branching fibrils within the fibril cluster (Fig. 3, Supplementary Fig. 13). SEMORE accurately tracked morphological changes over time in simulated dynamic data, achieving an average offset of just -13 frames (Supplementary Fig. 6). The model correctly classified morphologies even in high-density regions, reaching an F1 score >98% (Supplementary Fig. 11). SEMORE performed well with sparse structures, achieving >90% accuracy with as few as 4 detections at biologically relevant noise-to-signal ratios (Supplementary Fig. 7 & 8). It also accurately segmented degenerating structures, demonstrating its ability to analyze dynamic shrinkage or depolymerization (Supplementary Fig. 10). In experimental data, SEMORE successfully identified 139 individual, often overlapping, insulin aggregates, classifying them into isotropic and anisotropic growth pathways (Fig. 4). It also successfully analyzed dSTORM data of NPCs, extracting 285 segmentations and identifying individual NPCs, overlapping NPCs, and noise (Fig. 4), yielding an NPC diameter of 114 ± 18 nm, consistent with previous reports. SEMORE also accurately analyzed experimental data sets of FGFR1, Syntaxin 1a and RyRs (Supplementary Figs. 17-19). The impact of simulated blinking was minimal on SEMORE's performance, with segmentation accuracies above 90% (Supplementary Fig. 9). SEMORE effectively tracked the temporal evolution of morphology, revealing transitions between diverse morphologies (Supplementary Fig. 15). Overall, SEMORE demonstrated high performance and generalizability across various experimental systems, spatial scales, and temporal resolutions.

SEMORE addresses a significant bottleneck in SMLM data analysis by providing a robust, agnostic, and unsupervised machine learning framework. Its modular design, incorporating time-aware clustering and morphological fingerprinting, enables efficient and accurate analysis of diverse SMLM data. The use of a comprehensive set of features maximizes applicability across different biological systems, and the feature ranking facilitates mechanistic insights. SEMORE's ability to process both static and dynamic data, its resilience to noise and blinking artifacts, and its minimal human intervention requirements significantly enhance the usability of SMLM technology. The creation of morphological feature libraries holds promise for mapping assembly characteristics to their identity and function, advancing biological understanding and facilitating future supervised learning efforts. The results from diverse experimental data sets demonstrate SEMORE's ability to operate across a wide range of spatial and temporal scales and experimental conditions, making it a versatile tool for the SMLM community.

SEMORE presents a significant advancement in SMLM data analysis. Its unsupervised, agnostic, and modular design allows for rapid and accurate analysis of various SMLM datasets, addressing a major bottleneck in the field. The combination of time-aware clustering and morphological fingerprinting provides both quantitative and qualitative insights into protein assembly morphologies and dynamics. Future work could involve expanding the feature set of the fingerprinting module, creating libraries of protein assembly morphologies, and incorporating diffusional analysis for motion-aware clustering. SEMORE's versatility makes it a valuable tool for researchers working with SMLM data, accelerating data analysis and potentially leading to new biological discoveries.

SEMORE is currently optimized for laterally stable assemblies, although the temporal refinement module handles some lateral movement. While blinking has a minimal effect, it is recommended to correct for common aberrations like blinking, vibrations, and spherical aberrations before using SEMORE. Future versions will incorporate more sophisticated motion-aware clustering for applications with significant lateral movement. The development of standardized ground truth data across different SMLM platforms and experimental paradigms would facilitate more robust comparisons and benchmarking of analytical methods.