Multi Scale Diffeomorphic Metric Mapping of Spatial Transcriptomics Datasets

Michael I. Miller, Jean Fan, Daniel Jacob TwardWorkshops @ CVPR 2021 · DOI

Introduction

Spatial transcriptomics technologies, such as Multiplexed Error-Robust Fluorescence in situ Hybridization (MERFISH), provide unprecedented resolution by mapping the spatial coordinates of thousands of RNA molecules within intact tissue. However, this high-resolution data presents a significant computational challenge for registration—the process of aligning different tissue slices or samples from different animals. Unlike traditional image registration, spatial transcriptomics data consists of discrete point clouds (cells) with high-dimensional labels (gene expression), where no one-to-one correspondence between individual cells exists across samples. This paper introduces a multi-scale mathematical framework to compute smooth, non-rigid transformations that align these complex datasets by treating them as "varifolds."

Methodology

The core of the proposed approach is an extension of the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework, adapted for varifold representations.

Varifold Representation

Rather than treating cells as simple points, the authors represent the data as a sum of Dirac measures. This allows for a multi-scale analysis:

  • Fine Scale: Individual cells are represented with their specific cell-type labels.
  • Coarse Scale: The data is smoothed into a local mixture of cell types, characterized by the mean (average cell type) and local entropy (diversity of cell types). High entropy often indicates "mixing zones" between different biological regions, providing a robust signal for alignment.
Figure 1: Multi-scale representation of MERFISH data, showing raw signal (A, B) and cell-type assignments (C, D).

Diffeomorphic Mapping and Multi-scale Flows

The transformation is defined by velocity fields that evolve over time to deform a template image toward a target. By ensuring the mapping is "diffeomorphic," the algorithm guarantees that the transformation is smooth, invertible, and preserves the underlying topology of the tissue. Crucially, the fine-scale flow is treated as a refinement of the coarse-scale flow, ensuring that local cell alignments remain consistent with broader tissue-level movements.

Experimental Setup and Datasets

The framework was validated using a publicly available MERFISH dataset of the mouse preoptic hypothalamus. The dataset comprises 12 serial slices, each 10 microns thick and spaced 50 microns apart. The data includes spatial locations and expression profiles for 155 genes across thousands of cells, which were clustered into eight major types, including astrocytes, excitatory neurons, and inhibitory neurons.

Figure 4: Spatial distribution of eight major cell types in the mouse preoptic hypothalamus dataset.

Key Results and Insights

The model successfully computed optimal non-rigid alignments between neighboring brain sections, accounting for significant deformations of up to 140 microns. These displacements, often many times the size of a single cell, arise from physical tissue sectioning and natural anatomical variation.

Key insights from the study include:

  • Entropy as a Feature: Local cell-type entropy proved to be a powerful feature for characterizing tissue regions, aiding in the alignment of heterogeneous areas.
  • Weight Dynamics: The algorithm adjusts the "weight" of particles as the space is stretched or compressed, maintaining the "mass" of the signal and mimicking biological density changes.
  • Deformation Visualization: By using deformed grids, the authors visualized how the algorithm collapses or expands specific regions, such as the brain's ventricles, to achieve precise registration.

Limitations and Future Directions

While effective, the current approach relies on Reproducing Kernel Hilbert Space (RKHS) methods with quadratic complexity, which limits scalability for datasets containing millions of cells. Future work will focus on optimizing these computations using fast Fourier transforms or specialized kernels. Additionally, the authors plan to expand the framework to include subcellular scales (individual molecules) and macroscopic scales (whole-brain atlases), with the ultimate goal of building fully consistent 3D transcriptomic reconstructions.

Summary updated

BibTeX
@inproceedings{miller2021multi,
  author = {Michael I. Miller and Jean Fan and Daniel Jacob Tward},
  title = {{Multi Scale Diffeomorphic Metric Mapping of Spatial Transcriptomics Datasets}},
  booktitle = {Workshops @ CVPR},
  year = {2021},
  doi = {10.1109/cvprw53098.2021.00504},
}
Figure