Unifying community whole-brain imaging datasets enables robust neuron identification and reveals determinants of neuron position in C. elegans

We develop a data harmonization approach for C. elegans volumetric microscopy data, consisting of a standardized format, pre-processing techniques, and human-in-the-loop machine-learning-based analysis tools. Using this approach, we unify a diverse collection of 118 whole-brain neural activity imaging datasets from five labs, storing these and accompanying tools in an online repository WormID (wormid.org). With this repository, we train three existing automated cell-identification algorithms, CPD, StatAtlas, and CRF_ID, to enable accuracy that generalizes across labs, recovering all human-labeled neurons in some cases. We mine this repository to identify factors that influence the developmental positioning of neurons. This growing resource of data, code, apps, and tutorials enables users to (1) study neuroanatomical organization and neural activity across diverse experimental paradigms, (2) develop and benchmark algorithms for automated neuron detection, segmentation, cell identification, tracking, and activity extraction, and (3) share data with the community and comply with data-sharing policies. [Display omitted] •NWB file format extended to support C. elegans whole-brain imaging data•Production of a diverse data corpus using harmonization tools at wormid.org•Neuron ID algorithms improve in performance and do not need lab-specific retraining•Neuronal atlas reveals biological determinants of neuron positioning The C. elegans community continually produces whole-brain imaging datasets that could in theory be reused for new scientific inquiry. Nevertheless, because the data are produced using disparate equipment and techniques and stored in ad hoc formats, it is challenging to assimilate and reuse this growing collection. We developed a standard format and data harmonization methods to assimilate over 100 animals from five labs into a single data corpus. Using this corpus, we trained algorithms to produce a robust, lab-agnostic neural identification atlas system that does not require re-training for new labs and discovered biological factors that influence neural positioning. Training machine-learning algorithms on diverse data improves performance. Sprague et al. develop a standardized format and harmonization tools to unify a corpus of C. elegans whole-brain imaging data from five labs. Neuron ID algorithms show improved performance and generalization, obviating individual lab retraining, and the harmonized atlas yields neurobiological insights.