Abstract
Automated neuron identification in C. elegans is a critical bottleneck in systems neuroscience. While the nematode's eutely—having a fixed number of 302 neurons—theoretically allows for a universal brain map, inter-lab variability in imaging conditions has historically prevented the generalization of automated tools. We present WormID, a harmonized repository of 118 whole-brain datasets from five different laboratories. By standardizing data formats and implementing color-space histogram matching, we developed a statistical atlas that improves automated identification accuracy from 41% to 63%. This unified dataset also reveals fundamental biological principles of neuron positioning, demonstrating that gangliar boundaries and developmental lineage are primary determinants of spatial organization.
Introduction
In C. elegans research, whole-brain imaging is the gold standard for observing neural dynamics across an entire organism. However, manual identification of neurons in 3D fluorescent volumes is an arduous task, often requiring years of expertise and hours of labor per specimen. The introduction of "NeuroPAL"—a multi-color genetic barcode—provided a structural framework for identification, but automated algorithms remained fragile when applied to data outside their training set. This "data silo" problem stems from differences in microscopy, exposure settings, and sample preparation across labs. This study addresses these challenges through rigorous data harmonization and the creation of a community-wide statistical atlas.
Methods and Data Harmonization
To bridge the gap between disparate datasets, we developed a pipeline to convert heterogeneous lab-specific data into a unified corpus. Central to this was the extension of the Neurodata Without Borders (NWB) format via a new extension, ndx-multichannel-volume, which supports 5D data (multichannel, volumetric, time-series).
A major technical hurdle was color inconsistency. We implemented a histogram matching technique to align the color distribution of any new sample to a reference NeuroPAL manual, which independently improved identification accuracy by approximately 9%. These standardized datasets were then used to train a bipartite graph matching algorithm (the Hungarian algorithm) on a massive multi-lab corpus of 104 worms, rather than the small, single-lab cohorts used previously.
The WormID Statistical Atlas
The resulting WormID repository (available at wormid.org and hosted on the DANDI archive) comprises 118 datasets. This includes both structural NeuroPAL images and functional GCaMP calcium recordings from immobilized and freely-moving worms. To facilitate the use of this atlas, we mapped 3D neuron positions into a standardized coordinate system.
This mapping allows for the creation of a "Butterfly plot," which visualizes the angular and longitudinal positions of neurons across the population. This standardized view is essential for the atlas to provide robust "Top 5" identification suggestions, which reach an accuracy of ~82%.
Biological Determinants of Neuron Positioning
Beyond engineering improvements, the unified dataset allowed us to investigate the rules governing neuron placement:
- Gangliar Constraints: Neurons located within rigid structures, such as the pharynx, exhibit significantly less positional variability than those in looser somatic areas. This suggests that identification algorithms should adopt a hierarchical approach, identifying the "neighborhood" or ganglion before the specific cell.
- Connectivity and Proximity: Our data supports "Peter’s Rule," showing that neurons within 40 μm of each other have significantly higher chemical synaptic weights, suggesting physical proximity is a strong predictor of functional connectivity.
- Developmental Lineage: We observed that "sister" neurons (cells sharing a final division) remain close together regardless of birth timing. This indicates that their relative positions are established early in development and are largely resistant to post-embryonic growth drifts.
Results and Impact
The transition from lab-specific training to a harmonized multi-lab corpus resulted in a 1.6x reduction in error, jumping from 41% to 63% accuracy on unseen data. By providing 118 open-access datasets with persistent DOIs, WormID serves as a benchmark for the development of next-generation machine learning models for cell segmentation and tracking.
Limitations and Future Work
While WormID represents a significant leap forward, it currently lacks direct synaptic connectivity data found in Electron Microscopy (EM) studies. Future iterations of the atlas aim to incorporate neuronal shape, size descriptors, and dynamical activity correlations to push automated identification accuracy toward 100%.
Summary updated