Toward a More Accurate 3D Atlas of C. elegans Neurons

Michael Skuhersky, Tailin Wu, Eviatar Yemini, Edward S. Boyden, Max TegmarkbioRxiv 2021 · DOI

Abstract

Determining cell identity in volumetric images of tagged neuronal nuclei is an ongoing challenge in contemporary neuroscience. Frequently, cell identity is determined by aligning and matching tags to an “atlas” of labeled neuronal positions and other identifying characteristics. Previous analyses of such C. elegans datasets have been hampered by the limited accuracy of such atlases, especially for neurons present in the ventral nerve cord, and also by time-consuming manual elements of the alignment process. We present a novel automated alignment method for sparse and incomplete point clouds of the sort resulting from typical C. elegans fluorescence microscopy datasets. This method involves a tunable learning parameter and a kernel that enforces biologically realistic deformation. We also present a pipeline for creating alignment atlases from datasets of the recently developed NeuroPAL transgene. In combination, these advances allow us to label neurons in volumetric images with confidence much higher than previous methods. We release, to the best of our knowledge, the most complete C. elegans 3D positional neuron atlas, encapsulating positional variability derived from 7 animals, for the purposes of cell-type identity prediction for myriad applications (e.g., imaging neuronal activity, gene expression, and cell-fate).

Introduction

The nematode C. elegans is a cornerstone of neuroscience, being the only organism with a fully mapped connectome of its 302 neurons. However, identifying these specific neurons in a living, moving worm under a microscope is surprisingly difficult. For decades, researchers relied on a "canonical" atlas derived from electron microscopy data published in 1986. This legacy atlas has major flaws: it is largely 2D, constructed from a mosaic of different specimens, and fails to account for how much neuron positions vary between individuals. Furthermore, existing automated alignment tools often fail when faced with "real-world" microscopy issues, such as missing neurons (dropout), partial views (cropping), or the non-linear way a worm’s body stretches and bends.

Methods

The researchers addressed these challenges through a combination of high-quality data acquisition and algorithmic innovation.

The NeuroPAL Atlas and Pipeline

The team utilized the NeuroPAL transgene, which "barcodes" neurons with specific colors, allowing for deterministic identification. By imaging seven different worms, they created a 3D atlas that uses median positions to account for individual variability. The overall workflow involves constructing this atlas and then aligning new, unlabeled datasets to it.

The automated pipeline for constructing the NeuroPAL atlas and aligning unlabeled neurons.

Straightening and Canonicalization

To standardize comparisons across different animals, a "tangent circle" method was developed. This algorithm inscribes 1,000 circles into the worm's 2D hull to mathematically "untwist" the animal into a standardized 3D coordinate system.

The straightening pipeline: from raw fluorescence to a standardized 3D coordinate system using tangent circles.

Generalized-Mean (GM) Alignment

The core algorithmic contribution is the Generalized-Mean (GM) alignment algorithm. Unlike the standard Coherent Point Drift (CPD) method, which can get distracted by outliers, the GM algorithm uses a tunable parameter ($\gamma$). When $\gamma$ is negative, the algorithm ignores poorly matched pairs and focuses on high-confidence matches—a "soft clamping" effect that makes it incredibly robust to messy data. Biological realism is further modeled using a Gaussian mixture field, ensuring that if one neuron moves, its neighbors move in a similar direction, reflecting the physical constraints of biological tissue.

Key Results

The NeuroPAL-derived atlas consistently outperformed existing baselines, including the OpenWorm model and the original White et al. (1986) reconstruction, across all testing scenarios.

  • Superior Accuracy and Dropout Resilience: In simulations where neurons were randomly removed, the GM-realistic algorithm maintained 100% accuracy even with only 80 out of 300 neurons remaining. In contrast, existing CPD methods saw accuracy plummet as soon as neurons went missing.
  • Robustness to Cropping: The GM algorithm excels at aligning partial data (e.g., head-only images). While standard algorithms try to stretch head-data to fit a whole-body atlas, the GM algorithm simply ignores the missing body neurons and aligns the head perfectly.
  • Biological Noise Handling: The method remained stable even under significant non-rigid deformation, whereas rigid alignment methods failed once simulated biological stretching exceeded 15 microns.

Discussion and Future Work

While the new atlas represents a major leap forward, some limitations remain. The atlas currently lacks the two CAN cells because they express markers in non-neuronal tissues (like muscle), which can occlude imaging. Additionally, because worms are imaged on agar pads under a coverslip, they suffer from slight physical compression along the left-right axis, which may introduce minor biases in the 3D coordinates. The team hopes this atlas and alignment pipeline will become the standard for functional imaging, allowing researchers to automatically track the activity of specific, identified neurons while a worm performs complex behaviors.

Summary updated

Abstract

Abstract Determining cell identity in volumetric images of tagged neuronal nuclei is an ongoing challenge in contemporary neuroscience. Frequently, cell identity is determined by aligning and matching tags to an “atlas” of labeled neuronal positions and other identifying characteristics. Previous analyses of such C. elegans datasets have been hampered by the limited accuracy of such atlases, especially for neurons present in the ventral nerve cord, and also by time-consuming manual elements of the alignment process. We present a novel automated alignment method for sparse and incomplete point clouds of the sort resulting from typical C. elegans fluorescence microscopy datasets. This method involves a tunable learning parameter and a kernel that enforces biologically realistic deformation. We also present a pipeline for creating alignment atlases from datasets of the recently developed NeuroPAL transgene. In combination, these advances allow us to label neurons in volumetric images with confidence much higher than previous methods. We release, to the best of our knowledge, the most complete C. elegans 3D positional neuron atlas, encapsulating positional variability derived from 7 animals, for the purposes of cell-type identity prediction for myriad applications (e.g., imaging neuronal activity, gene expression, and cell-fate).
BibTeX
@misc{skuhersky2021toward,
  author = {Michael Skuhersky and Tailin Wu and Eviatar Yemini and Edward S. Boyden and Max Tegmark},
  title = {{Toward a More Accurate 3D Atlas of <i>C. elegans</i> Neurons}},
  year = {2021},
  doi = {10.1101/2021.06.09.447813},
  howpublished = {bioRxiv},
}
Figure