Spatial Clustering of Molecular Localizations with MIRO

MIRO (Multifunctional Integration through Relational Optimization) is a geometric deep learning framework that enhances clustering algorithms by transforming complex point clouds into structured, compact representations. It enables more robust clustering of single-molecule localization data using recurrent graph neural networks (rGNNs).

How it works?

MIRO learns to pull together localizations belonging to the same structure, producing spatially compact, well-separated clusters. This transformation enables standard algorithms like DBSCAN to perform significantly better — especially in challenging scenarios involving varying densities, blinking artifacts, or multiple cluster types.

Key Features

• Improved Clustering Performance: MIRO increases the efficiency of existing clustering algorithms by transforming point clouds into an optimized format.
• Simplified Parameter Selection: By enhancing the differentiation among clusters and their separation from the background, MIRO streamlines parameter selection for clustering methods like DBSCAN.
• Single-Shot and Few-Shot Learning: MIRO’s single- or few-shot learning capability allows it to generalize across scenarios with minimal training, making it highly efficient and versatile.
• Multiscale Clustering: MIRO’s recurrent structure allows for identifying patterns at different scales.
• Broad Applicability: MIRO is effective across datasets with diverse cluster shapes and symmetries.

Installation

To install MIRO and its dependencies:

1. Make sure you have Python 3.9 or higher installed.

2. Clone the repository to your local machine:

   git clone https://github.com/DeepTrackAI/MIRO.git
   

3. Install the necessary dependencies:

   pip install -r requirements.txt
   

MIRO is included as part of deeplay, a modular framework for deep learning.

📘 Tutorials

Explore MIRO's capabilities through interactive Jupyter notebooks:

• Benchmark: Reproduce MIRO's performance on the benchmark datasets from Nieves et al. Follow the Benchmark Tutorial to train your own model or load pretrained ones.

• Single-Shot Learning: See how MIRO achieves impressive results even when trained on a single cluster. Try it yourself in the Single-Shot Tutorial.

• Multiscale Clustering: Perform simultaneous clustering of nested structures with the Multiscale Tutorial.

• Multishape Clustering and Classification: Learn how MIRO can cluster and classify structures of different shapes using the Multishape Tutorial.

Citation

If you use MIRO in your research, please cite:

@article{pineda2024spatial,
  title={Spatial Clustering of Molecular Localizations with Graph Neural Networks},
  author={Pineda, Jes{\'u}s and Mas{\'o}-Orriols, Sergi and Bertran, Joan and Goks{\"o}r, Mattias and Volpe, Giovanni and Manzo, Carlo},
  journal={arXiv preprint arXiv:2412.00173},
  year={2024}
}