SuperResNET: Model-Free Single-Molecule Network Analysis Software Achieves Molecular Resolution of Nup96

Yahongyang Lydia Li, Ismail M. Khater, Christian Hallgrimson, Ben Cardoen, Timothy H. Wong, Ghassan Hamarneh, Ivan Robert NabiAdv. Intell. Syst 2025 · DOI

Abstract

SuperResNET is an integrated machine learning-based analysis software designed for visualizing and quantifying 3D point cloud data from single molecule localization microscopy (SMLM). By conceptualizing localization data as a spatial network, the software performs denoising, multi-scale segmentation, and modularity analysis to identify molecular-scale features without the bias of pre-defined structural templates. Applied to dSTORM data of the nuclear pore protein Nup96, SuperResNET GUI successfully resolved individual molecules at a distance of 11.1 nm, matching cryo-electron microscopy benchmarks. This model-free approach allows for the discovery of structural diversity in situ, providing a robust tool for biological discovery in pleomorphic subcellular environments.

Introduction

Single-Molecule Localization Microscopy (SMLM) techniques, such as STORM and PALM, have transformed cell biology by providing coordinate-based "point cloud" data of proteins at the nanoscale. However, translating these millions of coordinates into meaningful biological insights remains a significant computational challenge.

Traditional analysis methods often rely on template matching or particle averaging, where the software is instructed to look for specific shapes, such as the octagonal symmetry of the Nuclear Pore Complex (NPC). While effective for highly ordered structures, these methods introduce "template bias," potentially overlooking structural variations or incomplete biological assemblies. Furthermore, the stochastic "blinking" of fluorophores—often treated as noise to be filtered—contains dense spatial information that is frequently discarded. SuperResNET GUI was developed to address these limitations by treating SMLM data as a spatial network, enabling molecular resolution without the need for prior structural models.

Methodology: The Spatial Network Approach

SuperResNET GUI is a MATLAB-based tool that redefines SMLM data as a graph where each localization is a node and edges connect points within a specific proximity. This network-centric view allows for a sophisticated multi-step analysis pipeline:

Denoising and Segmentation

The software utilizes Ripley’s H-function to compare experimental data against a generated random network, effectively filtering out background noise. Following denoising, SuperResNET employs multi-scale segmentation using mean-shift and DBSCAN algorithms. This identifies "blobs" or clusters at varying proximity thresholds, allowing the software to distinguish between large-scale assemblies and smaller sub-components.

Feature Extraction and Modularity

For every identified cluster, the software calculates 30 distinct features, including size, anisotropy (shape), topology (hollowness), and network metrics like the characteristic path length. The core innovation of SuperResNET is its modularity analysis. By examining the internal connectivity of a single segmented cluster, the software can detect if that cluster is composed of multiple distinct modules—effectively identifying individual molecules within a single protein assembly.

Validation: Resolving the Nup96 Octagon

To validate the software, the researchers utilized publicly available 2D dSTORM data of Nup96, a component of the NPC known for its highly predictable 8-fold symmetric structure.

Turning "Noise" into Resolution

A key finding of the study is that retaining multiple blinks from the same fluorophore—rather than merging them—provides the network density required for modularity analysis. These extra localizations allow the software to resolve molecules that are closer together than the nominal localization precision of the microscope.

Molecular Resolution Results

SuperResNET GUI successfully segmented both complete nuclear pores and isolated Nup96 "corners." The modularity analysis of these corners identified two distinct sub-clusters at a mean distance of 11.13 nm. This result is remarkably consistent with the 12 nm distance established by cryo-electron microscopy, demonstrating that SuperResNET can achieve molecular resolution from dSTORM data without template-based guidance.

Discussion and Biological Implications

Because SuperResNET is model-free, it excels at identifying structural heterogeneity. In the Nup96 dataset, the software used K-means clustering on extracted features to automatically distinguish between complete octagonal pores and isolated corners (Group 1 blobs). A template-matching approach might have discarded these isolated corners as artifacts, but SuperResNET quantifies them as legitimate biological variations, offering a more accurate representation of the cellular environment.

The software confirmed the known dimensions of the NPC, identifying adjacent corners at approximately 42 nm and opposing corners at 107 nm. Overall, the GUI provides a predicted resolution of 13.3 ± 0.1 nm, effectively reaching the molecular scale.

Conclusion and Future Directions

SuperResNET GUI represents a shift away from template-dependent analysis toward a more flexible, network-based discovery model. While the success of the modularity analysis depends on high labeling density and fluorophore blinking, it provides a powerful means to study pleomorphic organelles where no structural template exists. Future applications may include the analysis of even higher-resolution data from technologies like MINFLUX, further pushing the boundaries of in situ structural biology.

Summary updated

BibTeX
@article{li2025superresnet,
  author = {Yahongyang Lydia Li and Ismail M. Khater and Christian Hallgrimson and Ben Cardoen and Timothy H. Wong and Ghassan Hamarneh and Ivan Robert Nabi},
  title = {{SuperResNET: Model-Free Single-Molecule Network Analysis Software Achieves Molecular Resolution of Nup96}},
  journal = {Adv. Intell. Syst.},
  year = {2025},
  doi = {10.1002/aisy.202400521},
}
Figure