LOMAR: LOcalization Microscopy Analysis in R

M. Theiß, Alvis Brāzma, Virginie Uhlmann, Jean-Karim HèrichébioRxiv 2022 · DOI

TL;DR

This paper introduces LOMAR, an R package designed to analyze Single-Molecule Localization Microscopy (SMLM) data, which is inherently represented as point sets. LOMAR offers novel methods for comparing and clustering these point sets, particularly addressing the challenge of biological heterogeneity in SMLM data by leveraging topological data analysis and advanced registration techniques.

What Problem Does This Paper Address?

Single-Molecule Localization Microscopy (SMLM) provides incredibly high-resolution images by pinpointing the locations of individual fluorescent molecules. This data is naturally represented as sets of 3D coordinates (point sets). A common analysis technique is "particle averaging," where many instances of the same molecular complex are aligned and averaged to reveal its structure. However, SMLM data often suffers from biological variability, meaning not all instances of a complex are identical. This heterogeneity can obscure the true structure during averaging. To overcome this, researchers need to group similar instances before averaging, which requires robust methods to measure structural similarity between point sets, even in the presence of noise and missing data. Existing tools often lack interoperability or are not specifically tailored for the unique challenges of SMLM data.

Methods

LOMAR tackles this problem by providing a comprehensive suite of tools within the R programming language. Key methods include:

  • Data Input and Preprocessing: LOMAR can ingest SMLM data from text files and offers filtering and clustering (DBSCAN) options. It can also convert point sets to images for visualization.
  • Point Set Registration: This is crucial for aligning different instances of a molecular complex. LOMAR implements several established methods like Iterative Closest Point (ICP) and Coherent Point Drift (CPD), alongside two novel approaches:
  • WGMMreg: An algorithm to register Gaussian Mixture Models (GMMs) by minimizing the Wasserstein distance, which is designed to be more robust to noise and capture relative point positions better than L2-based methods.
  • JRMPC (Joint Registration of Multiple Point Clouds): This method efficiently registers many point sets simultaneously by modeling them as originating from a common GMM, making it suitable for large datasets.
  • Topological Data Analysis (TDA): LOMAR incorporates persistent homology, a TDA technique that characterizes the "shape" of data by identifying topological features like holes and voids at multiple scales. This is achieved by converting point sets into simplicial complexes and generating persistence diagrams.
  • Similarity Measures: LOMAR derives similarity between point sets in several ways:
  • Using registration errors from the implemented algorithms.
  • Treating point sets as probability distributions and using measures like the Wasserstein distance.
  • Comparing their persistence diagrams using sliced Wasserstein distance or persistence scale-space kernels.

Datasets

The paper demonstrates LOMAR's capabilities using:

  • Real SMLM data: Specifically, 3D nuclear pore complex (NPC) data obtained via STORM.
  • Simulated SMLM data:
    • Simulated 3D instances of NPCs with varying structural characteristics (ring diameter, inter-ring distance) to test clustering performance.
  • Simulated NPC data with varying symmetries (8-fold vs. 9-fold) to evaluate robustness to class imbalance.

Key Insights and Ideas

  • Biological heterogeneity is a major challenge: Simply averaging all SMLM instances can mask important biological variations. LOMAR's ability to group similar structures before averaging is critical for accurate analysis.
  • Topological Data Analysis offers a powerful, shape-based similarity measure: Persistence diagrams, derived from TDA, capture the multi-scale "porosity" of structures, providing a complementary view to geometric registration methods. This approach is robust to noise and coordinate system differences.
  • Wasserstein distance is a valuable tool for comparing probability distributions (like GMMs): WGMMreg's use of Wasserstein distance for GMM registration offers an improvement over L2-based methods by better accounting for relative point positions.
  • Joint registration is essential for large datasets: JRMPC's ability to register many point sets simultaneously is a significant computational advantage over pairwise methods, enabling analysis of larger biological samples.
  • TDA-based clustering outperforms registration-based methods in certain scenarios: For simulated NPC data, TDA-based clustering demonstrated higher accuracy in separating classes based on subtle structural differences and in handling imbalanced datasets.

Key Results

  • LOMAR successfully ingested and processed real and simulated SMLM data.
  • The JRMPC method registered 356 NUP107 nuclear pores in under 20 minutes, significantly faster than previous all-vs-all registration approaches.
  • In clustering simulated NPC data, TDA-based approaches achieved accuracies of 91% for separating classes based on ring diameter and inter-ring distance, and 99% for separating classes based on symmetry (8-fold vs. 9-fold), outperforming registration-based clustering in the latter case.
  • The paper provides visual examples (Figures 1-4) illustrating the registration, TDA, and clustering capabilities of LOMAR.

Limitations and Future Work

  • Computational Cost: The current implementation of WGMMreg can be computationally prohibitive for very large point sets, though downsampling can mitigate this.
  • Rigid Registration: LOMAR currently only implements rigid registration (translation and rotation). The authors note that detecting deformations would be biologically relevant and could be a future extension.
  • Anisotropic Data: CPD may not be suitable for anisotropic SMLM data.
  • Future Directions: The authors suggest that the generic nature of LOMAR's functions means they can be applied to any 2D or 3D point sets, opening avenues for application beyond SMLM. Further development could include more advanced registration techniques and deeper integration of TDA for more complex structural analyses.

Summary updated

Abstract

Summary In single molecule localization microscopy data (SMLM), individual instances of a macromolecular complex come in the form of point sets. Particle averaging, which combines localization data from a large number of instances, is often used to overcome experimental noise and obtain a refined view of the underlying structure. However, SMLM point sets are often heterogeneous due to biological variations in the structure they represent and must be partitioned into groups with similar structure before averaging which calls for being able to compute structurally-relevant similarity measures between sets of points. Here we introduce LOMAR (LOcalization Microscopy Analysis in R) a software package for the R programming language that enables comparison of point sets through implementation of several point sets registration methods and similarity measures derived from topological data analysis. We demonstrate use of the package on real and simulated SMLM data of nuclear pore complexes.
BibTeX
@misc{thei2022lomar,
  author = {M. Theiß and Alvis Brāzma and Virginie Uhlmann and Jean-Karim Hèriché},
  title = {{LOMAR: LOcalization Microscopy Analysis in R}},
  year = {2022},
  doi = {10.1101/2022.05.30.493957},
  howpublished = {bioRxiv},
}
Figure