Cross-modality synthesis of EM time series and live fluorescence imaging

  1. Anthony Santella
  2. Irina Kolotuev  Is a corresponding author
  3. Caroline Kizilyaprak
  4. Zhirong Bao  Is a corresponding author
  1. Molecular Cytology Core, Memorial Sloan Kettering Cancer Center, United States
  2. Electron Microscopy Facility, University of Lausanne, Switzerland
  3. Developmental Biology Program, Memorial Sloan Kettering Cancer Center, United States

Abstract

Analyses across imaging modalities allow the integration of complementary spatiotemporal information about brain development, structure, and function. However, systematic atlasing across modalities is limited by challenges to effective image alignment. We combine highly spatially resolved electron microscopy (EM) and highly temporally resolved time-lapse fluorescence microscopy (FM) to examine the emergence of a complex nervous system in Caenorhabditis elegans embryogenesis. We generate an EM time series at four classic developmental stages and create a landmark-based co-optimization algorithm for cross-modality image alignment, which handles developmental heterochrony among datasets to achieve accurate single-cell level alignment. Synthesis based on the EM series and time-lapse FM series carrying different cell-specific markers reveals critical dynamic behaviors across scales of identifiable individual cells in the emergence of the primary neuropil, the nerve ring, as well as a major sensory organ, the amphid. Our study paves the way for systematic cross-modality data synthesis in C. elegans and demonstrates a powerful approach that may be applied broadly.

Editor's evaluation

This paper very nicely tackles a methodological problem in aligning different types of datasets (EM and light microscopy) to image embryonic nervous system development in the nematode C. elegans. The paper is important from a methodological standpoint, and also provides novel insights into nervous system development that will be of general interest.

https://doi.org/10.7554/eLife.77918.sa0

Introduction

Imaging modalities, both evolving established methods like electron microscopy (EM) (Helmstaedter, 2013; Chklovskii et al., 2010) and fluorescence microscopy (FM) (Sydor et al., 2015; Keller et al., 2015) and emerging methods like spatial sequencing (Takei et al., 2021) and micro CT (Ding et al., 2019) provide unprecedented streams of data. This data has revolutionized our ability to observe the development, structure, and function of the nervous system. However, interpretation and synthesis of this data remain challenging. Combining multiple modalities typically provides complementary information yielding a more holistic view of structural and functional dynamics across spatial and temporal scales (Ball et al., 2012; BRAIN Initiative Cell Census Network, 2021). Despite this, systematic atlasing that explicitly combines modalities (Vergara et al., 2021; Lalit et al., 2020) is limited, in part due to the technical difficulties in cross-modality image alignment. Most automated alignment methods such as those used in brain and whole organism atlases focus on appearance-based approaches (Vergara et al., 2021; Chiang et al., 2011; Oh et al., 2014; Chen et al., 2019; Kunst et al., 2019; Wang et al., 2020) which align areas with similar appearances. However, in cross-modality applications there is generally little shared appearance. Another class of alignment methods use anatomical landmarks, typically represented as points, and take known correspondences (Hunnicutt et al., 2016) or match points (Lalit et al., 2020; Heckscher et al., 2014; Yemini et al., 2021; Toyoshima et al., 2020; Yu, 2021; Long et al., 2009; Chaudhary et al., 2021; Preibisch et al., 2010) to compute an alignment. To account for complex variation of landmark positions (Toyoshima et al., 2020; Insley and Shaham, 2018), a variety of approaches are used to exploit relative position to improve matching accuracy (Lalit et al., 2020; Yu, 2021; Long et al., 2009; Chaudhary et al., 2021; Preibisch et al., 2010), but performance is limited with dense landmarks (typically 50–60% accuracy, at best). Efforts to build in additional information such as multi-color labeling to distinguish the landmarks, while technically difficult, can greatly improve matching performance (to 81–86%) (Yemini et al., 2021; Chaudhary et al., 2021).

Developmental data, which encompass dynamic anatomical changes over an extended time period, contain even more challenges. Different developmental processes show slight variation of developmental speed, creating heterochrony among anatomical structures. Between samples of comparable developmental stage, certain cell types and tissues/organs may or may not be present, causing a mismatch in the landmarks present. These potentially significant differences between samples complicate alignment and matching. More quantitatively, heterochrony in cell movement (from active migration or tissue shape change) may cause systematic positional variation that confounds approaches with naïve assumptions. Alignment using manually selected landmarks that are always present across samples partially bypasses these challenges (Wong et al., 2015). Efforts at alignment that largely ignore these discrepancies between individuals have also been attempted, with limited accuracy (<50%) (Lalit et al., 2020). However, generally, heterochrony and the associated challenges in developmental data are unaddressed.

Here, we address these limitations in a cross-modal study of C. elegans neural development. Specifically, we address two methodological difficulties, the challenge of annotating identities in developmental EM data, and that of collating the spatial detail of EM with the temporal detail of FM. We build an EM time series of C. elegans embryonic development and conduct cross-modality synthesis with time lapse FM to examine neural development with simultaneous high spatial and temporal resolution.

EM studies in C. elegans have made profound contributions to neuroscience including the first connectome White et al., 1997, sexual dimorphism (Cook et al., 2019), variability among individuals (Brittin et al., 2021; Witvliet et al., 2021), and principles of structural organization (Cook et al., 2019; Brittin et al., 2021; Witvliet et al., 2021; Moyle et al., 2021). However, these efforts are focused on the larval and adult nervous system, which leaves unanswered many questions about how the complex structure of the nervous system arises. Some EM work examines distinctive embryonic structures that can be readily identified such as the excretory system (Soulavie et al., 2018), amphid dendrite tip (Low et al., 2019), or the general appearance of the ring (Rapti et al., 2017). However, since neurons are typically identified in EM based on specialized adult morphologies it is challenging to identify them in the embryo during their emergence. It is primarily work using live FM that has examined the C. elegans nervous system during embryogenesis, providing insight into neurulation (Shah et al., 2017b; Barnes et al., 2020), organogenesis (Low et al., 2019; Fan et al., 2019), neuropil formation (Moyle et al., 2021; Rapti et al., 2017; Kennerdell et al., 2009; Shah et al., 2017a; Sengupta et al., 2021), synaptic specificity (Berghoff, 2021), as well as lineage differentiation and brain asymmetry (Chuang et al., 2007; Cochella and Hobert, 2012; Masoudi et al., 2021). Lineage tracing in FM establishes definitive identities for cells via ancestry even when they lack distinctive positions or morphologies (Shah et al., 2017a; Bao et al., 2006) and light sheet microscopy allows imaging to extend into embryonic motion (Wu et al., 2013). However, these efforts are limited by the lack of spatial resolution and availability of cell-specific markers needed to resolve individual cell shape. To fill in this gap, we generate an EM embryonic series, develop and apply a landmark-based co-optimization algorithm for cross-modality alignment in the presence of developmental variation. Furthermore, we synthesize the EM series and live FM series data representing over a dozen markers. The cross-modality synthesis reveals critical dynamic behaviors of identifiable individual cells across scales in the emergence of the primary neuropil, the nerve ring, as well as the sensory end of a major sensory organ, the amphid. Our study paves the way for systematic cross-modality data synthesis in C. elegans and demonstrates a powerful approach that may be applied broadly. We present the resulting annotated series as a resource for the community in an online format that facilitates additional community-based curation.

Results

Multi-modal developmental time series of C. elegans embryogenesis

We first image an EM time series of C. elegans embryogenesis with data at four classic stages, the bean, comma, 1.5-fold and 2-fold stages (Sulston et al., 1983; Figure 1a). The EM series spans a critical 150-min time window during which most tissues undergo major morphological changes: pharyngeal and intestinal lumens develop, distinctive morphologies develop in muscles and the excretory system, the epidermis spreads and closes over the embryo, and the body length doubles. In the nervous system, most of the major nerve tracts emerge. The nerve ring is not yet present at the first time point, but densely populated at the end of this period. As part of the effort to test and demonstrate the feasibility of EM time series acquisition, the bean and comma stage embryos are imaged with Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) while the 1.5- and 2-fold stage embryos are imaged with Array Tomography (AT). The lateral resolution ranges from 8.5 to 19.4 nm, and z spacing 25–85 nm (see Supplementary file 1a and Methods). Entire embryos were imaged, containing 576–601 cells including visible apoptotic cells (Supplementary file 1b).

Figure 1 with 2 supplements see all
EM and FM time series of C. elegans embryogenesis with complementary spatial and temporal resolution.

All scale bars indicate 5 μm. (a) Overview of EM data and their placement in the time course of C. elegans embryogenesis. Cartoons show body shape and the approximate physical sectioning plane orientation is shown in red. Orthogonal views of each dataset are shown below, with the view corresponding to physical sectioning outlined in red. f.c., first clevage or cell division. (b) An example FM series (see Figure 1—video 1) with a ubiquitous histone marker (red) and a cnd-1 promoter driven membrane marker labeling a subset of neurons (green). Representative time points are shown as max projection. Time stamps are min post first cell division (pfc).

3D time-lapse FM is widely used to study C. elegans embryogenesis, providing rich information for cross-modality synthesis with EM. We focus on two-color FM images from the WormGUIDES project (Moyle et al., 2021; Santella et al., 2015), where a ubiquitously expressed histone/nuclear marker is used to label and track every cell and a sparsely expressed cell membrane marker reveals dynamic cell behaviors such as neurite outgrowth (Figure 1b, Figure 1—figure supplement 1; Moyle et al., 2021; Barnes, 2020). These image series typically begin at the four-cell stage and continue for approximately 8 hr until the 1.75-fold stage and the onset of embryonic movement, with a temporal resolution of 60 or 75 s. The nuclear labels are used to trace the entire cell lineage and provide the position and lineage identity of every cell at every time point (Santella et al., 2014; Katzman et al., 2018) (see Figure 3—video 1).

In order to correlate a FM series to the EM series, we use the nuclear positions, which are also annotated for the EM series, as the common reference (i.e. a Rosetta stone) which allows us to identify cells in the EM. Given nuclei in both data sets we find the optimal 4D (3D+time) alignment between cells in the FM and EM series which conveys the lineage based identities to cells in the EM series. We then relate the dynamic cell behaviors as observed in the FM and EM series based on the correlated cell identities.

The co-optimization algorithm for cross-modality alignment of developmental images

We present a landmark-based approach for cross-modality image alignment, which accounts for developmental variation and functions well with challenging dense landmarks. Alignment here takes the form of assigning landmark identities from a set of labeled data to novel unlabeled data, this involves both the computation of a geometric alignment proper and matching. The landmarks, which could be of any anatomical scale that originate from any imaging modality and segmentation method, are represented as spatial point clouds. As detailed below, our algorithm involves an ensemble model of example data to represent the anatomical structure of landmarks in the labeled data and a co-optimization approach for alignment.

The ensemble model contains three parts (Figure 2a) in order to capture the regularity and developmental variations of the anatomy: the set of labeled data; a neighbor graph representing consistent estimated adjacencies (see Materials and methods) observed between landmarks across individual labeled datasets; and the set of variable landmarks whose presence is inconsistent among labeled data. First, the ensemble, like instance-based learning approaches that use training data to directly model a distribution, can better preserve the complex positional variation observed in the labeled data. Second, the consistent adjacencies graph summarizes the structural consistency in three-dimensional spatial relationships. The adjacency graph captures relative positions, which makes it insensitive to body size variation, or shifting between organs and tissues due to posture, sample distortion or heterochrony in development, since the invariant positional relationships within a given organ/tissue are preserved in the graph. The approach is further motivated by biology. Physical adhesion between cells and tissues underlie anatomical regularity. Consistent adjacencies over multiple samples enrich for the bona fide regularity caused by some underlying physical driver and removes coincidental adjacencies. Third, and most importantly, the set of inconsistent landmarks provides a data-driven approach to handle discrepancies in landmarks between labeled and unlabeled data, which has been a major challenge and source of error in using alignment to establish identity via matching. Inconsistent landmarks are likely discrepancies, which can be used to deliberately modify a labeled dataset by adding or removing these landmarks to reduce discrepancies and improve the success of alignment and identity assignment.

Figure 2 with 1 supplement see all
The co-optimization algorithm for cross-modality alignment of developmental data.

(a) Components of the ensemble model. Circles represent landmarks. Numbers in each circle mark the corresponding landmarks across individual labeled data (L1 to Ln). 1a and 1b denote a split of 1 (cell division or other separation of one landmark into two) and the empty circle indicates the disappearance of 6 (cell death or other transient landmark disappearance). Lines represent computed adjacencies, among which thick lines represent consistent adjacencies. (b) Co-optimization of landmark composition in a labeled dataset Li and alignment Ai between Li and unlabeled data. Li is iteratively modified to Li’, aligned, and scored to judge each modification. Modifications that improve score are accumulated to yield an optimized Li’ and corresponding alignment Aimax. (c) The overall process of deriving correlated identities for an unlabeled dataset. Each labeled dataset Li is optimized to generate a best corresponding Aimax. Voting per landmark is then performed over the set of labels across Aimax to achieve a final labeling Amax.

The co-optimization algorithm uses the ensemble model to simultaneously optimize modifications to the set of landmarks in labeled data and the alignment between the modified labeled and the unlabeled data (Figure 2b). Adding or removing inconsistent landmarks in labeled data Li creates a modified Li’ to be aligned with the unlabeled data. We score the resulting alignment by counting missing consistent adjacencies. This score is used as an objective function to drive optimization as we modify the set of inconsistent landmarks. The optimized set of landmarks ultimately yields an alignment with a maximal score, Aimax. We term this process by which we optimize to pick a set of modifications, scoring them by the alignment quality co-optimization. Finally, to utilize the multiple sets of labeled data in the ensemble model, the unlabeled data is independently aligned with each labeled dataset. From the alignment with each labeled dataset: Aimax, each landmark in the unlabeled data receives a predicted identity. A consensus Amax is arrived at by a vote among the set of predicted identities for each landmark, where the most common identity is taken (Figure 2c). The co-optimization algorithm is implemented with a correspondence-free nonlinear warping method for pre-alignment (Myronenko and Song, 2010; Figure 2—figure supplement 1a), Linear Assignment Problem (LAP) matching (Jonker and Volgenant, 1987), data-driven modification, and greedy optimization of labeled data (see Materials and methods).

Correlated cell identities convey a single-cell view of EM series

We apply the co-optimization algorithm to align the FM and EM series of C. elegans embryogenesis, using individual cells/nuclei as landmarks. We use the FM series, where cells were tracked and lineage identities assigned, to generate the ensemble model for each developmental stage in the EM series and transfer cell identities to the EM (see Materials and methods). Due to the lack of lineaged FM data at the 2-fold stage, we align only the bean, comma and 1.5-fold stages in the EM series. Based on the timing of known developmental events in the FM series, the aligned EM data are timed as approximately 320, 345, and 380 min post first cell division (pfc), see Materials and methods for details. During this developmental window, 75 cell divisions and 76 deaths occur, while cell counts remain virtually constant (576-601) (Figure 3a). The ensemble FM model for each EM dataset (Supplementary file 1c) contains 39 labeled 3D FM images of embryos, each generating 5838–6170 pairwise adjacencies between landmarks, of which 439–504 (~8%) are consistent, as well as 27–34 inconsistent landmarks/cells (~5% of cells) (Figure 3a, Figure 3—figure supplement 1a, Figure 3—figure supplement 2a). While the scale of the ensemble models is fairly constant over the developmental stages, the consistent adjacency graph that guides alignment is dynamic (Figure 3, Figure 3—figure supplement 1a, Figure 3—figure supplement 2), only 146 adjacencies, about a third of consistent adjacencies, are observed at all three periods due to both inconsistent landmarks and embryo-wide cell movements during organ morphogenesis and body elongation (Barnes et al., 2020). We note our method uses only perfectly consistent landmarks but estimated adjacency captures more granular information about the consistency of cell positions, and has other potential applications in quantifying the consistency and variability in development (Figure 3—figure supplement 2b, c, Supplementary file 1).

Figure 3 with 3 supplements see all
Cross-modality alignment conveys a single-cell view of EM data.

(a) Illustration of the ensemble model for the comma stage. Landmarks (dots) and adjacencies (edges) in a single instance of labeled data. All adjacencies are shown with the consistent adjacencies highlighted in black. Inconsistent landmarks are larger and red. (b) A visualization of the comma stage EM data after alignment. Nucleus centroids are dots colored by major predicted tissue type with color key provided. Adjacencies are shown as gray lines. (c) Accuracy of predicted identities for each EM data at the single-cell and tissue levels. (d) Illustration of EM data annotated by identities. Tissue regions are shaded following the color scheme in b to highlight overall anatomical structure. The excretory canal cell is marked with a star with the nucleus of the excretory duct and part of the pore cell body visible below it. Individual neurons are numbered. Identities as follows are from alignment (-a) or manually confirmed (-c) 1 FLPR/AIZR parent-c, 2 ASER-c, 3 AVBR-c, 4 ASHR-c, 5 AWCR-c, 6 SIBDR-c, 7 AVKR-a, 8 AIYR-a, 9 SMBDR-c, 10 SMBVL-a, 11 AIML-a, 12 AVKL-a, 13 SMDDL-c, 14 FLPL/AIZL parent-a, 15 RMGL-a, 16 ALML/BDUL parent-c. Scale bar indicates 5 μm. (e) 3D reconstruction in comma stage EM data colored to maximize local contrast and showing cell body contours for seam cells (blue and orange), gut cells (alternating shades of red), germ line (cyan), and the excretory canal, duct and pore cells (red, green, and blue respectively) as well as nuclear contours for pharynx (green). (f) 3D reconstruction of the excretory system in the 1.5-fold stage EM data. Red, green, blue and cyan are the excretory canal, duct, pore and gland cell, respectively, as in e. Black arrows point to auto fusion in the duct and pore cell. (g) EM view of lumen (white arrow) and site of auto fusion (black arrow) in the excretory pore cell at 1.5-fold stage. Scale bar indicates 1 μm.

The result is a single-cell level alignment between the FM and EM series, with a predicted identity for each cell in the EM series (Figure 3b, Figure 3—figure supplement 1b), which we refer to as the correlated EM series. We assess the accuracy of the correlated cell identities based on a set of manually curated cell identities across organs and tissues (Figure 3—figure supplement 1c, 148 to 242 per embryo). We consider accuracy at two levels: individual cells and tissues (Figure 3c). Single cell accuracy, which is the agreement between manual and predicted identity, varies between 71% and 78%, while accuracy of tissue identities is 79–95%, both with a trend for decreasing accuracy for the later developmental stages. Among head neurons, the single cell accuracy is 69–78%. There is no comparable embryonic work, but this is higher than previous reported efforts that attempted to align and name all head neurons in young adults based purely on cell positions (50–68%) (Yemini et al., 2021). As an additional validation we use our approach to predict the names for all cells from anonymized nuclear positions at each stage in three additional FM embryos (see Materials and methods) and show comparable success (Supplementary file 1f).

These results are deposited on webKnossos (Boergens et al., 2017) as a community resource containing the EM images, annotated nuclear positions and cell identities. Data is publicly accessible and can be navigated within a convenient web interface that allows users to search for cell identities by name or query a cells identity by clicking on its centroid in the image. With a free account users can fork the annotation and edit names or reconstruct neurons or other structures. Currently, 56–70% of the cells have identities that are manually curated (Figure 3—figure supplement 1c), including 25–40% as cell identities (see Supplementary file 1e) and 16–34% as tissue/organ identities, greatly reducing potential errors in identities. Annotation will be continually updated based on community feedback and our own ongoing efforts at neuronal reconstruction, which we hope will make these online data sets an increasingly useful community resource. Researchers requiring raw data access are encouraged to contact the authors.

The correlated cell identities convey a single-cell view of the EM data for developmental dynamics at different scales (Figure 3d and e). As an example of how developmental succession can be followed over time we highlight tube formation in the excretory system. The excretory pore, canal and duct can be seen stacked on top of each other at the bean and comma (Figure 3d and e) stages. At 1.5-fold a continuous tube is visible that connects the canal to the outside environment which has formed through concerted auto fusion of these cells (Figure 3f and g). These events are captured with a clarity comparable with EM imaging taken specifically to study these events (Soulavie et al., 2018; Sundaram and Buechner, 2016).

Spatial and temporal organization of neuropil formation

Leveraging the co-optimization algorithm for alignment, we conduct cross-modality synthesis of diverse data sets documenting early NR development to elucidate the spatial and temporal organization underlying neuropil formation. The nerve ring (NR) is the primary neuropil in C. elegans. The adult NR consists of 181 neurites, where neurite topography and contacts constrain synapse formation and circuit function (Brittin et al., 2021; Witvliet et al., 2021; Moyle et al., 2021; Durbin, 1987). The NR has long been a prime model to study the general principles of brain structure and function. The NR emerges in the later half of embryogenesis (Rapti et al., 2017; Altun et al., 2021), but our understanding of the spatial and temporal organization of neurite outgrowths in this process is still poor. Efforts with FM revealed that the first neurites appear around the bean stage and a ring around the pharynx becomes visible by the comma stage (Figure 4a; Moyle et al., 2021; Rapti et al., 2017). FM with cell-specific markers and sparse labeling has led to critical insights: identification of pioneers (Moyle et al., 2021; Rapti et al., 2017; Kennerdell et al., 2009), the inside-out pattern of organization of outgrowth (Moyle et al., 2021), and even neurite relative placement through three color (Yip and Heiman, 2018) or combinations of two-color imaging (Figure 4b). However, lack of markers to systematically label the entire nervous system and lack of spatial resolution hinder further insights. EM on the other hand provides a global view of not only every neurite but also the surrounding tissues with high spatial resolution (Figure 4b). We scan through EM data to identify neurites, and back trace them to soma in order to obtain identities (see Materials and methods). We also correlate EM with a set of FM series that encompasses 12 strains with cell specific markers that label neurons individually or in small, isolated groups (Figure 1—figure supplement 1, Supplementary file 1d; Moyle et al., 2021; Barnes, 2020).

Figure 4 with 2 supplements see all
Spatial and temporal organization of neuropil formation.

All scale bars 1 μm. (a) FM cross section of a comma stage embryo with a ubiquitous histone marker (red) and a broadly expressed cell membrane label (green) (Barnes et al., 2020). Anterior view. Dotted line encircles the NR and routing of sublateral neurons into the ring. (b) Comparison of resolution in resolving NR in FM and EM. Two images of two-color FM image of developing neurites (red, green, and blue are labeled by the promoters of cnd-1, ceh-10, and zag-1, respectively). Neurite bundles are numbered anterior to posterior. EM showing the corresponding neurites and their placement in the ring at the comma stage. Neurites are labeled with names and colors of the corresponding bundle in FM. (c) Schematic of the NR (anterior view) showing first outgrowths, entry points, and approximate extent of neurites at bean and comma stages. Ovals represent neuronal soma and arrowheads the leading edge. Two arrowheads are shown for the amphid (dark blue) and sublateral (light blue). The earlier arrowhead indicates initial outgrowth, and the latter line and arrow indicates the entry points and approximate outgrowth extent at the bean stage. Moving dorsal to ventral, dorsal outgrowth is green, supralateral red, lateral orange, and ventral purple. See Figure 4—figure supplements 1 and 2 for corresponding EM. Open arrows and associated letters indicate the figure panels that show the location pointed to. (d) EM at comma stage showing the first outgrowth at the supralateral entry point with cells shaded red and named. (e) Comparison of initial outgrowth timing derived from FM and EM. Each dot represents a single neuron or small cluster of neurons. Dot color corresponds to membership in groups as pictured in panel c. Cells that do not clearly correspond to these major entry points are gray (See Supplementary file 1d for details). (f) Number of neurites visible in the cross sections of the ring in panel g. The location of this cross section is indicated by arrow b in panel c. Black arrow indicates bean stage. (g) Axial NR cross sections (neurites shaded cyan) near the lateral midline entry point (the location indicated by arrow g in panel c). (h) Precision in neurite outgrowth dynamics. A resliced view of the dorsal midline at comma stage shows the point left and right sublateral early neurites meet (white arrow). Colors match schematic in c.

Our analysis shows that the establishment of the NR is organized by eight entry points which include the dorsal midline, the ventral midline, and three left/right symmetric pairs in between, namely supralateral, lateral and sublateral (Figure 4c). The early time points in the correlated EM series reveal the first outgrowths that establish each entry point (Figure 4c and d, Figure 4—figure supplement 1a, b, c). Sublateral neurons, namely SIAD and SMDD, are the first neurons to initiate outgrowth into the ring, with their initial outgrowth captured at the bean stage (Figure 4—figure supplement 1b). This observation is consistent with these neurons serving as early pioneers for the NR (Moyle et al., 2021; Rapti et al., 2017; Kennerdell et al., 2009). In less than a half hour after this outgrowth (comma stage), all entry points have been populated with the corresponding early neurites. Moving from the dorsal midline to the ventral midline: at the dorsal midline, RMED and ALA have grown bilaterally (Figure 4—figure supplement 2b); supralateral neurons, namely AVD, AVH, and AVJ have entered and grown dorsally (Figure 4d, Figure 4—figure supplement 2c); lateral neurons, namely SMDV and RIV have also entered the ring almost reaching the dorsal midline (Figure 4—figure supplement 2d); and at the ventral midline, RIR, RMEV, and RIH have bilaterally advanced dorsally (Figure 4—figure supplement 2g). While the EM series provide a comprehensive and definitive set of entry points and first outgrowths due to its label-free nature, the FM series provide much finer temporal resolution of outgrowth dynamics. In terms of the timing of neurite initial outgrowth (Figure 4e, Supplementary file 1d), outgrowths observed as having occurred in EM by the bean stage span almost 30 min in the FM while outgrowths observed in EM by the comma stage span an hour as measured in FM. Most of the FM times are congruent with the upper bound of timing provided by EM. Further data would clarify if the occasional contradiction in timing is due to biological variability or technical issues in temporal alignment (see Materials and methods).

In addition to defining the first outgrowths and the initial outgrowth times, the cross-modality synthesis also reveals dynamics of subsequent neurite growth in the NR (Figure 4f). Assayed in a cross section near the lateral entry point (Figure 4g, arrow g in Figure 4c), the number of processes visible in the NR increase rapidly from 12 at the comma stage to 51 by the twofold stage. Outgrowth dynamics is not reducible to the timing of first outgrowth, subsequent extension also has striking temporal dynamics. For example, the first outgrowths into the amphid commissure, namely ASH, RIB, and AWC, show initial outgrowth at the bean stage, which make them one of the first neurons to grow out in the entire nervous system along with the sublateral neurons (Altun et al., 2021; Figure 4—figure supplement 1a, c). However, these neurites, and additional followers from the amphid neurons pause soon after their initial outgrowth (Moyle et al., 2021; Barnes, 2020) near the sublateral entry point (Figure 4—figure supplement 2a, e ) before they enter the NR (by the 1.5-fold stage).

Finally, the spatial resolution of the EM series highlights the precision of control in neurite outgrowth kinetics. For example, comparing the left- and right-side behavior of the first cells to grow into the NR, namely SIADL/R and SMDDL/R (Figure 4—figure supplement 1b, Figure 4—figure supplement 2f), suggests strong symmetry in behavior. Both sides have broken out at the bean stage, with similar sized outgrowths. The contralateral pairs then reach the midline simultaneously at the comma stage: their tips meet precisely beneath the midpoint of the ALA nucleus, which marks the dorsal midline (Figure 4h). This level of precision is not limited to the first sublateral outgrowths. Left and right counterparts among the lateral and supralateral first outgrowths have also reached nearly identical positions relative to the dorsal midline at the comma stage (Figure 4h). Quantification of the length of the most advanced sublateral neurite pairs SMDDL/R and SIADL/R shows left-right counterparts differ in both cases by less than 2 µm, a few percent of overall length. It is not clear how such level of precision is achieved, or if there is signaling to coordinate between the two sides at this stage though asymmetric sensitivity to guidance molecules has been shown in amphid axons (Grossman et al., 2013).

Cell-cell interactions and ultrastructure during amphid organogenesis

The amphid is a major sensory organ of C. elegans consisting of 12 neurons and two glial cells (the socket and sheath cells). The adult sensory end is a bundle of ciliated dendrite endings with drastically different morphologies surrounded by a tube-shaped glial channel. The glial cell extensions are in turn embedded between epidermal cells to create a sensory structure that is open to the environment (Altun et al., 2021; Ward et al., 1975; Doroquez et al., 2014). The dendrites grow via a distinctive mechanism, a collective retrograde extension from an initial rosette structure towed by the migrating epidermis (Low et al., 2019; Fan et al., 2019; Heiman and Shaham, 2009; Kunz et al., 2021). Formation of the embedded opening and elaboration of dendrite morphology occur in embryogenesis (Doroquez et al., 2014; Oikonomou et al., 2011; Nechipurenko et al., 2017). Molecular components necessary for correct glial (Oikonomou et al., 2011; Perens and Shaham, 2005; Oikonomou et al., 2012) and cilium (Bacaj et al., 2008) formation have been identified. However, the detailed steps in the embryonic assembly of this structure are less understood.

The correlated EM series confirms the dynamic cell behaviors revealed by FM during dendrite extension (Figure 5a; Fan et al., 2019) and adds structural details. For example, 3D reconstruction at the bean stage shows small protrusions from amphid neurons clustered as in the initial rosette and lodged in a dent at the anterior/leading edge of the epidermis (Figure 5b and c, Figure 5—figure supplement 1a, b, c). Notably, the socket cell, though engaged in the rosette earlier does not appear engaged in this focal location at the bean stage and only touches a handful of amphid neurons (Figure 5b, Figure 5—figure supplement 1a, b).

Figure 5 with 2 supplements see all
Cell-cell interactions and ultrastructure during amphid organogenesis.

All scale bars 1 μm. (a) FM view of amphid dendrite development with a cnd-1 promoter driven membrane label. Dashed line shows embryo contour. Dot indicates the location of dendrite tips, and arrow the socket and sheath cells. Dashed circle inside the embryo outlines the initial amphid rosette at ~280 min pfc. (b) 3D reconstruction of cells participating in the collective dendrite extension from bean stage EM. Epidermis: semi-transparent; amphid neurons: magenta; sheath cell: cyan; socket cell: pink. Arrows indicate protrusions on the socket cell. (c) The same reconstruction as in b with socket and epidermal cells removed. Black arrow highlights protrusion on the sheath cell. White arrow indicates focal point of amphid neuron protrusions. See also Figure 5—figure supplement 1a-c. (d) Partial max projection of FM image with a cnd-1 promoter driven membrane label showing socket cell (filled arrow) and more weakly labeled sheath cell (open arrow) posterior migration. (e) Schematic of the time course of sensory opening formation. Arrows indicate motion of cells. Colors match b. (f) Resliced view of EM at comma stage highlighting socket cell (tinted rose) shape and relative position to epidermal cells (green), sheath cell (cyan), and dendrites (magenta). Black arrow indicates contact between the socket cell and the exterior of the embryo. (g) EM at the 1.5-fold stage showing the sensory opening. Colors follow f. Arrow indicates the opening to the exterior of the embryo. (h) EM at twofold stage showing three sections of a dendrite tip. White arrows mark nine individual microtubule structures. (i) EM at twofold stage showing a bifurcated dendrite tip, each tip marked with an arrow.

More importantly, our cross-modality synthesis reveals critical details of cell behaviors and interactions in creating the multi-layered structure of the open sensory end (Figure 5e). The sheath cell starts on the anterior side of the amphid rosette (Figure 5a). Subsequently, the sheath cell body moves dorsal-posteriorly (Figure 5d). Lamellipodia-like protrusions from the sheath cell during this process are visible at the bean stage (Figure 5c). Meanwhile, the anterior portion of the sheath cell wraps itself around the collected mass of the dendrite tips, which is evident by the comma stage (Figure 5f). The socket cell exhibits more complex behaviors. One set of behaviors follows a similar course as the sheath cell. The cell body migrates posteriorly (Figure 5d), creating transient posterior protrusions (Figure 5—figure supplement 1d), while the anterior end remains close to the dendrite tips. Subsequently, after the sheath cell has wrapped the dendrites, the socket cell starts to wrap over the sheath cell (Figure 5e and f). The wrapping is complete by the 1.5-fold stage (Figure 5g; Low et al., 2019; Oikonomou et al., 2011). Distinctively, the socket cell must also interface with the epidermis to create the open sensory end (Figure 5g). Noteworthy details hint at the dynamic role of the socket cell in the process. During dendrite extension, the socket cell makes elongated protrusions, one of which appears to reach out and contact the advancing epidermal cell (Figure 5b). Later, the anterior end of the socket cell is juxtaposed between epidermal cells and exposed to the outside of the embryo like a plug (Figure 5f), which may play a role in creating the opening to the environment in the sensory end. The socket cell is fully surrounded by the epidermis by the 1.5-fold stage (Figure 5g, Figure 5—figure supplement 1e). Overall, the dynamic details from the cross-modality synthesis suggest that the strategy of amphid sensory end assembly is successive wrapping of interior structures with additional layers.

Furthermore, the series illustrates the elaboration of distinctive ultrastructure. Centrioles move to the dendrite tips between the bean and comma stages (Figure 5—figure supplement 1f; Li et al., 2017). By the twofold stage, centrioles have disappeared, presumably degraded or remodeled (Nechipurenko et al., 2017; Li et al., 2017), and microtubule bundles (9 bundles per dendrite) are visible growing into the short proto-dendrite tips (Figure 5h). Their appearance is essentially the same, aside from length, as adult structures (Doroquez et al., 2014). Meanwhile, the dendrite tips begin to elaborate their complex morphology. The shape of the dendrite tips remains simple up to the 1.5-fold stage. By the twofold stage, some diversity in length and shape is visible. Notably, two dendrites, namely ADF and ADL, show short bifurcated tips, each with its own set of microtubules branching from a central location (Doroquez et al., 2014; Figure 5i). Dendrites tips are not yet individually embedded within the sheath cell by the twofold stage.

Discussion

Our study demonstrates a powerful co-optimization algorithm to align and assign identity to landmarks. Because each cell in C. elegans is unique, with well-characterized molecular and morphological traits, it is often desirable to identify individual cells. For postembryonic cell identification, people combine cell position with the characteristic cell/neurite morphology to make a positive identification. In the embryo, before such morphology becomes prominent, cell identification has been challenging. In his seminal work, Sulston suggested that cells in the late embryo can be identified by careful examination of relative positions within small neighborhoods (Sulston et al., 1983), but he remained the only person who could do so systematically. Instead, researchers use live imaging-based lineage tracing for reliable cell identification (Bao et al., 2006; Schnabel et al., 1997). By paying attention to relative position through the adjacency graph, our algorithm essentially follows Sulston’s suggestion, and achieves unprecedented accuracy in cell identification. Additional information can be readily added to the adjacency graph to better score matched landmarks and improve the alignment, such as size, appearance and other features of the landmarks when available. Furthermore, while our study is focused on the wild type, one could use live FM to build the appropriate ensemble models to analyze mutants. It should also be possible to adapt our methods to partially imaged embryos. Thus, by eliminating the bottleneck in data acquisition as well as data interpretation, we envision a broad use of EM and cross-modality synthesis of C. elegans. As a proof of concept, we present the image data and identity annotation of our correlated WT EM series as an accessible public resource. Existing annotation provides a navigational aid and the seed of a community effort that we hope will fully validate annotation of the series. Extending our method to mutants is conceptually similar, requiring positional information from several lineaged FM data sets and an EM time series. This data could be efficiently assembled, at least in the case of highly penetrant mutants.

More broadly, the co-optimization algorithm provides a general solution for cross-modality and cross-scale alignment. Representation of landmarks as a point cloud is modality- and scale-free. In our study, we use individual cells/nuclei as landmarks due to the desired single-cell resolution for C. elegans biology, but the point cloud can easily represent a mixture of landmarks at different scales, enabling ambitious future work. With advances in deep learning and image analysis, landmarks can be extracted from complex organs such as the mammalian brain by learned region detectors (Chen et al., 2019) with increasing reliability. While the mammalian brain is orders of magnitude larger in terms of cells, it is worth noting that in state of the art atlases it is currently annotated with hundreds of regions (Wang et al., 2020), which makes alignment at this level of detail a similar scale problem to ours. We postulate that at an optimal level of landmark density consistent adjacencies provide an effective implicit model of anatomy that can maximize available prior information about structure and constrain accurate automated alignment. The data-driven modification in the co-optimization algorithm further facilitates such a vision for future automated alignment and annotation of complex brain images. It readily handles not only biological variability in landmarks (developmental or between individuals) but also technical variability such as errors in landmark extraction or systematic differences between modalities. This model only becomes more powerful as data accumulates. Ultimately, we imagine a multi-modal fusion process built on our approach will be capable of systematically collating information from multiple modalities and discovering consistent landmarks, likely at the level of organs and anatomical structures rather than single cells. Our approach provides spatial alignment robust to landmark variability. Aligning between individuals and over time, and incorporating molecular information to allow segmentation of aligned landmarks into spatio-temporal regions would automate the decision making involved in defining the regions of a digital atlas (Wang et al., 2020) making a full spatio-temporal atlas of developing anatomy possible.

Last but not least, or experience provides broader insights on the usefulness of various forms of whole embryo EM in C. elegans and beyond. Firstly, our post-hoc correlation of EM and FM data provides a useful alternative to true correlative EM a powerful, but complex to implement, approach for targeted EM based on the expression pattern of fluorescence markers (de Boer et al., 2015). The identity of landmarks obtained from our post-hoc correlation can be used to select relevant targets for re-imaging with AT at higher resolution, similar to the use of correlative methods in identifying regions for EM imaging. Dense correspondences could also in principle be used to visualize correlated FM data superimposed on EM. Secondly, our efforts provide broad insights on the relative ease and usefulness of AT and FIB-SEM in the C. elegans embryo. FIB-SEM excels in axial resolution and avoids alignment artifacts, making it easier to follow small neurites between planes. This is however balanced by greater difficulty in achieving good results due to exacting preparation requirements and slow acquisition of a limited area. As the embryo is close to the maximal size of the imaging area there is danger that only late in acquisition will it become apparent that parts of the embryo are found to be out of the ROI captured. In addition, it can be difficult to precisely stage the embedded embryos until time has been expended on imaging them. In contrast, AT is more artifact prone, but a more flexible and forgiving process. With AT it is possible to section first, then inspect sections and select appropriate samples for imaging. In our application, this is particularly useful for selecting appropriately staged embryos from sections containing multiple embryos before spending time on imaging. AT acquisition is less automated and post-acquisition internal stack alignment can be challenging and require extensive manual effort. However, largely because of its reduced risk of acquisition failure, we have found AT to be more reliable in yielding useful data.

Materials and methods

EM imaging

High-pressure freezing and freeze-substitution

Request a detailed protocol

A generous amount of E. coli was placed into 0.5 µm carriers filled with 20% BSA solution. Embryos from N2 worms were collected using a platinum pick from the seeded plates and settled in the carrier. The carriers were covered with a lid and frozen in a Wohlwend HPF Compact 02 (Wohlwend GmbH, Sennwald, Switzerland).

The freeze-substitution step was done in the freeze-substitution unit (AFS2; Leica Mikrosysteme GmbH, Vienna, Austria) following the previously described procedure (Kolotuev et al., 2010; Lee et al., 2018; Kolotuev et al., 2012). Briefly, frozen samples were transferred under liquid nitrogen to the freeze-substitution mix of 2% (w/v) osmium tetroxide, 0.1% (w/v), uranyl acetate 2% (v/v) H2O in acetone pre-cooled to –120°C. Samples were heated to –90 °C over a period of 2 hr and left at this temperature for 2 hr. Over 24 hr, the temperature was raised to –30 °C. Samples were left in this temperature for 2 hr. After they were washed once with acetone and three times with ethanol, the temperature was raised to +20 °C. The solvent was replaced with the mix of ethanol and Epon resin mix (Embed 812 resin EMS, 1420), infiltrating the samples with increasing concentrations 30%, 50%, 75% for three hours each exchange at ambient temperature. Three incubations and exchanges of 100% EPON were done over 12 hr. Samples were flat embedded using the two-step procedure (Kolotuev et al., 2012) and polymerized for 48 hr at 60 °C.

Block trimming

Request a detailed protocol

For accurate sectioning, it is important to orient samples correctly. Flat blocks facilitate recognition of a ROI and help in trimming. Delimited areas surrounding the sample were trimmed using a 90 diamond trim tool (Diatome, Biel, Switzerland). Care was taken to approach the sample as closely as possible and keep the top and bottom sides parallel to one another and the knife’s surface (Kolotuev, 2014). A buffer area of empty resin at the leading edge of the trimmed sample was left to allow space for alignment maneuvers during the subsequent steps.

Focused ion beam/scanning electron microscopy (FIB-SEM)

Request a detailed protocol

For FIB experiments, samples were oriented transversely or longitudinally already at the stage of trimming. Sample blocks with targeted trimmed rectangles ∼100×200 µm were glued on aluminum SEM specimen stubs and coated with a 30 nm layer of platinum. Stubs were introduced into FIB‐SEM (Helios Nanolab 650, Thermo Fischer Scientific) and further calculations were done inside the microscope chamber to target the ROI precisely to the embryo of interest. FIB/SEM acquisitions were performed using FEI Slice and View software. The FIB milling was done at 30 kV acceleration voltage, 790 pA current. Images were acquired with the in column detector (ICD, comma stage embryo) and mirror detector (MD, bean stage embryo) in immersion mode using an electron beam of 2 kV, 800 pA and 10 µs of dwell time.

Array tomography (AT)

Request a detailed protocol

For AT experiments, the ‘blind’ sectioning strategy does not require precise sample orientation. Several samples can be sectioned at once, and final assessment of age is done after sectioning.

Planar arrays of the section were generated using the ATS knife (Diatome, Biel, Switzerland) mounted on a Leica UC7 ultramicrotome (Leica Microsystems; Vienna, Austria). Rectangles of silicon wafer approximately 2 × 4 cm (Ted Pella, No 16015) were cut from the round plate glow discharged, cleaned, and placed into the knife boat. ‘Ribbons’ of ~300 consecutive sections, 70–100 nm thick, were subdivided into shorter segments and then aligned side by side on support (Kolotuev, 2014). After the sections transfer and water drains, the wafer was let dry slowly at the ambient temperature and subsequently incubated for 30 min at 60 C.

The wafers were transferred to the Helios 650 FIB-SEM (Thermo Fischer Scientific). The strategy of positional correlation to select the embryos of the desired stage was used. To find the anchor sections that are worthy of being analyzed in sequence, we chose the strategy of "leaping" between centrally located sections across the short horizontally aligned ‘ribbons’ to screen for the ROI. Acquisition parameters were similar to those used in FIB protocols. Images were acquired with mirror detector (MD) in immersion mode using an electron beam of 2 kV, 800 pA and dwell time of 5–10 µs.

EM image analysis and rendering

Request a detailed protocol

For alignment of slices within an AT EM data set, we used IMOD. Automatic IMOD registration was performed. On top of this manual adjustment was used only to correct occasional qualitatively misaligned slices. 3D tracing of sections was done with the 3dmod module of IMOD (Kremer et al., 1996) or TRACKEM (Cardona et al., 2012) with 3D rendering done in Daz3D Daz Studio, 2005.

3D, time-lapse FM datasets

Request a detailed protocol

As previously described (Moyle et al., 2021) time lapse images of fluorescently labeled C. elegans embryos were collected with spinning disc or diSPIM microscopy (Duncan et al., 2019), and lineaged with StarryNite (Bao et al., 2006; Santella et al., 2014; Boyle et al., 2006) to assign cell identities. The annotated nuclear centers serve as landmarks for ensemble model creation and series correlation.

Cross-modality alignment

Pre-alignment

Request a detailed protocol

Pre-alignment between labeled and unlabeled data allows the use of absolute spatial position and is performed in two steps.

First, an approximate affine least squares or Thin Plate Spline (TPS) (Zeller and Schempp, 1976) alignment (the latter used for the 1.5-fold embryo) is performed based on a small number of manually identified landmarks. The landmarks are identified in each EM dataset based on their characteristic positions and cell shapes. Cells from different regions of the embryo are chosen to highlight the body axes (Supplementary file 1g). The landmark set was evaluated visually per EM data set to ensure the point clouds of labeled and unlabeled data fully overlapped when aligned.

Second, the approximate alignment is refined using a coherent non-linear method GMM-CPD (Myronenko and Song, 2010). This method creates a regularized local warping function to align two point clouds by iteratively warping and matching points under an assumption that nearby points move similarly, which provides a more granular adjustment for non-linear shifts in body proportion and relative positioning of organ structures (Figure 2—figure supplement 1a).

Temporal staging

Request a detailed protocol

A coarse temporal staging is used to select the appropriate time window in the FM data from which to create the ensemble model. Stage is first visually estimated based on matching body axis elongation in the FM and EM datasets. At the comma and 1.5-fold stages this provides a very strong timing estimate. At the bean stage elongation is a weaker cue and we take advantage of the readily identifiable anterior gut cell Ealaa which undergoing mitosis in our bean stage EM embryo to refine our initial estimate. Isolated distinctive divisions are scattered throughout late embryogenesis making this a generally useful strategy. We then visualize the quality of spatio-temporal alignment by applying the pre-alignment as described above and plotting the EM aligned to a randomly selected FP data set. A visualization of manual correspondences under this pre-alignment is then used to adjust the temporal staging if alignment results appear to contain a systematic AP shift related to the elongating body axis.

We note that quantitative constraint violation scores from the alignment method described below can be used for more precise temporal alignment (Figure 2—figure supplement 1b) which would be useful in implementing fully automated methods of spatio-temporal alignment. A wide and smooth minima around the manual temporal alignment demonstrates that a minimization of constraint violations over different possible temporal alignments would provide a similar temporal alignment to our manual assessment. The absolute minima being offset by a few minutes from our manually assessed staging suggests the approach might even improve on manual time estimate. However, such a precise temporal alignment is not strictly needed for the purpose of choosing a relatively wide time window to create the ensemble model.

Ensemble model creation

Request a detailed protocol

For each EM stage an ensemble model encompassing adjacency, position and inconsistent landmark information is created to be used in alignment. Adjacency and positional information is collected within a local (13 min) window of time in each of the three FM embryos that contribute to the labeled data ensemble. Temporal window size was optimized to maximize identity prediction performance using one of the three labeled data embryos as a test case and the other two as labeled data. Adjacencies are predicted from landmark positions via a Gabriel graph (Gabriel and Sokal, 1969) for all landmarks in a labeled dataset. Gabriel graph edges represent a subset of Voronoi diagram faces, a common computational approximation of cell membranes, that match an additional geometric definition of stability. Consistent adjacencies are those present in all datasets within the ensemble. A list of inconsistent landmarks is compiled from all cell deaths and cell divisions that occur within the local temporal window.

Modification and co-optimization

Request a detailed protocol

We use our ensemble model to guide modifications of inconsistent landmarks in labeled data optimizing the exact set of labeled data landmarks along with alignment. This optimization uses missing consistent adjacencies in the unlabeled data, (measured on the predicted labeling of the unlabeled data which is created by the alignment) as an objective function in a process we term co-optimization.

Each possible modification to the labeled data represented by the list of inconsistent landmarks is tested. The hypothesis that a programmed cell death has already disappeared in the unlabeled sample is considered by removing that cell from the labeled data. For divisions the hypothesis that a recently divided cell in the labeled data has not divided in the unlabeled sample is considered by merging the two sibling cells into one parent cell at their midpoint. The hypothesis that a cell in the labeled data that will soon divide has already done so in the unlabeled dataset is entertained by splitting the parent cell into two cells sharing the same location. Note that while it would be possible to use prior information about the expected division axis to create more accurate daughter positions, this seems over complicated as the main goal is to prevent cascading errors due to qualitative discrepancies in the size of the set of landmarks present.

A simple, greedy, gradient descent process optimizes the labeled data. The optimization first considers the (limited in C. elegans) possible divisions and then deaths and accumulates changes that improve total adjacency constraint violations. For each of these hypotheses a match of the unlabeled dataset against the modified labeled data (Li’) is made and adjacency constraint violations are counted under the corresponding alignment (Ai). A proposed change is kept if it results in a significant (in our case ≥ 3) reduction in adjacency violations.

Linear assignment and scoring via consistent adjacencies

Request a detailed protocol

For a given pre-aligned unlabeled dataset and modified labeled dataset, matching is framed as a linear assignment problem (LAP) minimizing the total distance between pairs of points. Missing consistent adjacencies are then used to score these matching results to evaluate modifications and to identify and prohibit incorrect matches.

Given an alignment between a modified version of the labeled data and unlabeled data that alignment can be evaluated by computing the Gabriel graph for the unlabeled data and, using the predicted identities under the alignment, counting missing consistent adjacencies. The total number of missing consistent adjacencies for the alignment is used to score the modification process described above.

In addition, a high number of missing consistent adjacencies for an individual landmark’s predicted identity (summed over all expected adjacencies of that landmark only) indicates it is likely wrong (note this can be assessed only given the context of the global alignment, this cannot be assessed independently for a single landmark match). These individual match scores are used to identify and attempt to correct incorrect matches in an iterative process. The individual identity assignment with the worst score, (i.e. the most missing consistent adjacencies) if any exists, is made impossible (set to infinite cost) and LAP matching is repeated. This process is repeated for N iterations, and at the end of those iterations the global alignment with the lowest total violations is selected. Iterative error correction (200 steps) is performed after each modification, before judging its quality. Once a labeled dataset is fully co-optimized an additional 1400 iterations are performed and the best alignment and matching for this is Aimax.

Voting and reiteration

Request a detailed protocol

Voting picks an identity for each landmark from the set of predicted identities corresponding to an alignment of the unlabeled data against each labeled data set. The most common name out of these independent answers for each landmark in the unlabeled data is taken as its identity. This procedure can result in multiple cells in the unlabeled data being assigned the same identity. Although it would be possible to force a unique set of identities by using an additional linear assignment between landmarks and available identities scored by total votes for each identity per landmark, we do not perform this step as we found this marginally reduces accuracy (data not shown).

Highly reliable landmark names can be used to reinitialize the alignment more accurately improving matching results. Landmarks in the sample where all or nearly all assignments agree are most reliable, (see Figure 2—figure supplement 1c). Taking landmarks where the winning identity has a 75% supermajority in place of the TGMM automated nonlinear alignment we compute a new landmark based TPS alignment based only on these ‘known good’ matches. These reliable automatic matches are greater in number than initial manual landmarks and can potentially provide a better nonlinear pre-alignment than TGMM. We then repeat the exact same co-optimization based matching for only the subset of uncertain cases, (the certain cases are reinserted with their fixed identities when evaluating adjacency constraints). All reported results include one round of iterative re-alignment.

Understanding of the overall structure of these steps is aided by pseudo code illustrating how these steps fit together:

Pseudo-code of the algorithm:

1. Alignment per labeled data: For each set of labeled data Li=1..n

 • Two-step spatial pre-alignment between labeled and unlabeled data

  • Pre-Align via manually ID’d landmarks in Li and unlabeled data

  • Correspondence-less GMMCPD nonlinear pre-alignment

 • Modification and co-optimization of inconsistent landmarks:

  • Li max =Li; Aimax = LAP_with_iteration(Li max, unlabeled data)

  • For each inconsistent landmark present in Li

   • Make modification to Li max (adding or removing landmark) creating Li max’

   • Ai = LAP_with_iteration(Li max’,unlabeled data)

   • Score(Ai)=missing consistent adjacencies in Ai

   • If Score(Ai)<Score(Aimax)

    • Li max = Li max’; Aimax = Ai

2. Voting: For each landmark j in unlabeled data:

 • Pick most common identity for landmark j over the set of identities in Aimax i=1…n

 • If 75% supermajority then the identity of landmark j is considered confident

3. Reiteration: Repeat steps 1–2 above with the modifications:

 • Substitute TPS warping using confident identities in place of pre-alignment using manual landmarks and GMCCPD

 • Confident identities in unlabeled data remain fixed and are not recomputed

Algorithmic ablation results

Request a detailed protocol

To evaluate the contributions of distinct components of our method to performance we compare performance after ablating parts of the algorithm to our full algorithms’ performance (71–78%). Ablating all of the unique components of our approach leaves the prior method of an LAP matching based on distance (after pre-alignment) to each labeled example followed by voting, this has accuracy of 43–66%. The most novel component of our system the co-optimization of unstable landmarks makes a significant contribution to accuracy; if we skip this step single cell accuracy ranges from 67% to 75%.

Computational considerations

Request a detailed protocol

Our implementation has a significant runtime but could be significantly optimized if desired. Our co-optimization algorithm was implemented in Matlab. Alignment was performed on a Xeon Gold 6128 3.4 GHz CPU with 500 mb of memory (alignment is however not particularly memory intensive). Runtime was variable depending on the temporal stage and number of modifications to be considered but took about a day. The vast majority of run time is taken by the tens of thousands of iterations of linear assignment implicit in the modification and constraint optimization process. Results are fairly insensitive to the number of constraint iterations used. Reducing iterations and total computation time by 3/4 results in only a small decrease in accuracy. Given the small number of datasets computation is not a major issue and reported timing and accuracy results reflect a large number of iterations run after initial landmarks and parameters were optimized using runs with a low number of iterations. A more efficient approximate LAP implementation would significantly decrease computational costs as over 95% of computational time is spent in this function. Runtime would also be significantly reduced by an explicitly parallelized implementation as alignment of individual labeled datasets, currently sequential, is completely independent; this was not attempted.

Additional validation

Request a detailed protocol

To further assess our algorithm, we test its ability to name additional independent point clouds based on FM microscopy. In all experiments we use the same staging method, model, parameters, and same set of initial landmarks for pre-alignment as in EM experiments. Our additional test set contains 3 embryos imaged for 8 hr with a spinning disc confocal microscope. All embryos are lineaged and globally curated through 1.5-fold to ensure accurate cell identities for validation. Note that while our model is based on light sheet FM on uncompressed embryos, these additional test data sets are compressed while imaged, representing a challenge to alignment at least as significant (though distinct) as distortions from AT EM. We blind ourselves to identities except for the exact same pre-alignment landmarks used in the EM. Alignment is performed on a point cloud taken from the equivalent stage of each data set and accuracy is then assessed over all cells, since a complete ground truth is available. Result accuracy is comparable to EM and presented in Supplementary file 1f.

Manual curation

Manual curation of cell identities

Request a detailed protocol

A significant number of cells in each EM dataset were manually curated to establish their identities at the tissue or individual cell level. Tissue level annotation was created completely manually, moving one by one through each cell and annotating cells whose type could be readily assigned visually. This was possible based on overall morphology for the majority of pharynx, epidermis, gut, and body wall muscle cells, as well as some neurons.

Single-cell annotation

Request a detailed protocol

Verified individual cell identities fall into two conceptually distinct categories: a. landmarks used in pre-alignment b. additional curated identities used in evaluation and for navigation.

The majority of pre-alignment landmarks are individual cells within readily identifiable tissues such as the gut, body wall muscles, and epidermis. The individual identities of cells within these tissue blocks can be unambiguously established by counting cells along the body axes within the block of tissue types. We found this most easily done by visualizing the cell centroids in 3D within WormGUIDES for comparison to a labeled data set. First, we identify tissues in the EM images and mark these cell centroids in a 3D visualization. Then, taking the gut as an example we count our way down the AP axis in the 3D model, identifying which gut nucleus centroid is which gut cell by comparing position to an identical 3D centroid visualization of a FM data set with all identities. In addition, there are a few landmarks which have dramatically distinct appearances (such as the excretory cell) or occupy highly distinctive positions on the dorsal, ventral or lateral midlines, or both (such as AVL precisely on the ventral midline, directly anterior of the excretory cell, and ALA, on the midline, and sharing an elongated left-right shape with its sibling RMED). These landmarks are very difficult to confuse with any other cell and are used as initial landmarks.

Additional curated identities for evaluation were first established for highly distinctive cells not used in pre-alignment but identifiable with the same procedures described above. Then, moving essentially in a ring out from these highly reliable landmarks we examine labeled and unlabeled data sets as 3D point clouds and identify cells with consistent position relative to a landmark in labeled data sets and the corresponding cell in EM. At the stages examined we have found this kind of relative position to be very reliable, with all observed discrepancies being local displacements traceable to variability of nuclear position within extended cell bodies. These tentative positional identifications are further validated against EM morphology, tissue boundaries, and known neuronal appearance at this developmental stage, when this is available from FM experiments. At the comma and 1.5-fold stages individual neuronal shape provides strong hints for identification. Identified patches are expanded checking for inconsistencies until all cells in a local patch are identified. Though user errors are possible initially, once a patch of identities is extended sufficiently to span the area between highly distinctive cells/tissues, and all cells within an area are accounted for there is relatively little opportunity for human error. Total time spent on annotation varied between a few days for the least annotated data set and a bit more than a week for the best annotated comma data set.

Neuronal outgrowth

Request a detailed protocol

Assigning cell identities in FM: In all FM data the identities of cells were determined via lineage tracing (see details in Section 2 above). Individual embryos were lineaged to identify individual cells or small clusters of cells expressing the membrane bound GFP transgenes detailed in Supplementary file 1d. Note that due to mosaicism the outgrowth dynamics of single cells can sometimes be assessed when expression in a region in potentially broader. Expressing cells once identified have their initial outgrowth times recorded. Assessing initial outgrowth in FM: Neurite initial outgrowth was defined in time lapse FM recordings as the first frame in which a distinct outgrowth was visible on a neuron cell body. This was further formalized as a protrusion reaching, and continuing to reach in future, in the direction of final outgrowth. This is distinct from possible earlier contractility of the cell surface which does not display this directedness. In small clusters of expressing cells, the cell used to name the cluster was confirmed to have an elongated shape aligned with nearby expressing cells, suggesting the outgrowth observed comes from multiple cells sharing the same orientation. Initial outgrowth timings from distinct embryos were aligned to a common temporal reference frame based on a linear normalization using cell division timings as landmarks (Barnes, 2020).

Assigning cell identities to neurites in EM: To identify individual neurite cross-sections labeled in figures, we manually traced neurites back to a cell body. Once neurites were traced back to cell bodies, nuclei were identified with the procedure described in section Additional validation above.

Assessing initial outgrowths in EM: Every cell in the vicinity of the ring was assessed for visible neurite outgrowth (though not all cells are pictured in Figure 4e). A cell was scored as broken out if it contained a significantly elongated structure extending in the direction expected from adult morphology. There is some ambiguity when observing static images between initial outgrowths and cells with elongated tear drop shapes that might precede outgrowth. The examples discussed have neurite outgrowths long enough to be unambiguously neurites, rather than tapered cell bodies. Other more ambiguous cells, other sublaterals, are arguable and may or may not have neurites at the bean stage. Temporal staging of samples was performed as described in section Temporal staging above.

Data availability

EM data has been made available on WebKnossos, and source code made available on https://github.com/zhirongbaolab/CellIDAlignment, (copy archived at swh:1:rev:7d5f9853cfc6f8def19e02a0a22a22a2c519e2ac).

The following data sets were generated
    1. Santella A
    2. Kolotuev I
    3. Kizilyaprak C
    4. Bao Z
    (2021) webKnossos
    ID 1712,526,428,0,2.2,576. C. elegans bean stage EM.
    1. Santella A
    2. Kolotuev I
    3. Kizilyaprak C
    4. Bao Z
    (2021) webKnossos
    ID 399,796,779,0,4.716,424. C. elegans comma stage EM.
    1. Santella A
    2. Kolotuev I
    3. Kizilyaprak C
    4. Bao Z
    (2021) webKnossos
    ID 940,634,77,0,1.129,249. C. elegans 1.5-fold stage EM.
    1. Santella A
    2. Kolotuev I
    3. Kizilyaprak C
    4. Bao Z
    (2021) webKnossos
    ID 1173,591,151,0,0.621,1130. C. elegans two-fold stage EM.

References

  1. Book
    1. Barnes K
    (2020)
    Spatial and Temporal Organization of Brain Morphogenesis at Single-Cell Resolution a Study in C. Elegans
    Weill Cornell Graduate School of Medical SciencesWeill Cornell.
  2. Book
    1. Durbin R
    (1987)
    Studies on the Development and Organisation of the Nervous System of Caenorhabditis Elegans
    Cambridge University.
  3. Conference
    1. Lalit M
    2. Handberg-Thorsager M
    3. Hsieh YW
    4. Jug F
    5. Tomancak P
    (2020)
    in Computer Vision – ECCV 2020
    Workshops Lecture Notes in Computer Science Ch. Chapter 30. pp. 458–473.
    1. Myronenko A
    2. Song X
    (2010) Point Set Registration: Coherent Point Drift
    IEEE Transactions on Pattern Analysis and Machine Intelligence 32:2262–2275.
    https://doi.org/10.1109/TPAMI.2010.46

Decision letter

  1. Oliver Hobert
    Reviewing Editor; Columbia University, Howard Hughes Medical Institute, United States
  2. Claude Desplan
    Senior Editor; New York University, United States
  3. Andrew D Chisholm
    Reviewer; University of California, San Diego, United States

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Cross-modality Synthesis of EM Time Series and Live Fluorescence Imaging" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor (Oliver Hobert) and Claude Desplan as the Senior Editor. The following individual involved in the review of your submission has agreed to reveal their identity: Andrew D Chisholm (Reviewer #1).

All three reviewers agree that this is very nice work that we would like to see published in eLife. However, all three reviewers feel that the manuscript would very much profit from significant editorial revisions that emphasize the biological findings more and also contextualizes a number of these findings more effectively. You are also much encouraged to move material from the Supplement into the Main text, particularly those figures that provide novel biological insight. You will find below a quite extensive list of suggestions for editorial changes and we encourage you to consider them well.

Reviewer #1 (Recommendations for the authors):

Some issues in the presentation of the manuscript could be improved prior to publication. There appears to be mis-numbering of the supplemental figures and/or their callouts in the text. Although I have a theory as to the correct numbering, this created considerable confusion in the review.

The cell identification in the EM datasets is a combination of manual identification and automated identification based on the alignment algorithm. The manual identification process could be clarified, as it is extensive (140-240 manual IDs per data set). Table S5 summarizes some 'landmarks' used in pre-alignment, could the authors expand on what makes these neurons or other cells 'distinctive/distinctively shaped'? This table might be of more utility if fully written out at the single-cell level. It would be worth explaining what constitutes the 'ground truth' for these manual IDs.

Presentation and wording:

(1) The title states 'cross modality synthesis' of EM and FM imaging, however elsewhere it may be more accurately described as 'image alignment' or 'cross modality analysis'.

(2) Why is the EM data a 'pseudo time series? Pseudotime implies that the time dimension is inferred from non-temporal data, yet the EM data form a sparse time series (4 time points, though described as 'dense' in results).

(3) What is the definition of 'breakout' in the context of neurite growth?

(4) The C. elegans-specific term 'hypoderm' is not widely used; for a general audience 'epidermis' can be used as it means the same thing.

Reviewer #2 (Recommendations for the authors):

Strengths

– This is a first attempt for a combined study of C. elegans embryos by fluorescent and electron microscopy, by aligning – registering images without the use of correlative electron microscopy.

– This study introduces a landmark-based alignment assigning identities from a set of labeled data to unlabeled data, with co-optimization of unstable landmarks contributing to accuracy.

– This work provides a comprehensive analysis that largely supports and sometimes expands on previously studied processes of neuropil and sensory channel formation in C. elegans embryos.

– The authors provide publicly available data of their EM sections in webKnossos open resource, which is provided in amenable form can be very advantageous for the community.

Weaknesses

– The technical strength of the approach compared to others is not sufficiently demonstrated. There is a sparse actual assessment of the method's robustness/variability across samples and in comparison to other methods, in terms of data acquisition/analysis. Including comparative analysis of EM sample replicates for each time point, and comparisons across alternative methodologies (algorithms, electron microscopy) will allow substantiating the approach's superiority and its feasibility for future large-scale use, suggested by the authors in the study.

– Regarding data accessibility, the study sometimes lacks a detailed, comprehensive presentation of data acquisition, analysis, and underlying assumptions/caveats. This refers to instances such as the use of landmark fluorescent markers with multicellular expression to follow single-cell processes, the image registration/cell-identification using landmarks with expression patterns that are not presented in detail, explanations about process tracing and verification in EM sample analysis. The authors invite for community curation/ feedback of a publicly available EM resource, yet the provided resolution and formats (in terms of cell identification, tracing) may not facilitate curation.

– The study sometimes lacks comprehensive, inclusive contextualizing of new findings in relation to knowledge previously published by the authors and by other specialists. That refers to discussed developmental events of lineage progression guiding sample analysis, and findings of excretory pore and neuropil formation, pioneer-neuron concepts, amphid channel assembly.

I list my comments in order of appearance in the manuscript, which should be addressed for publication.

Suggestions for improved or additional experiments, data or analyses:

86-87 "As part of the effort to test and demonstrate the feasibility of EM time series acquisition[…]": The authors should discuss any potential variability/ differences arising from using different EM techniques to analyze different timepoints 1,2 and 3,4. Since the two techniques are used "to demonstrate feasibility" it appears key to discuss (later in the manuscript) the comparison of the two approaches in terms of their use in such project, i.e. effort/ implementation of sample preparation/data analysis, the influence of resulting resolution/stitching to cell identity registration, other pros/ cons. More importantly, if a single embryo is analyzed per timepoint, this would be one of my main critiques of the manuscript. At least 2 samples per timepoint will allow assessing the method's feasibility, robustness, reproducibility, variability.

92 "sparsely expressed cell membrane marker […] (Figure 1b, S1)": For the detailed understanding of C. elegans specialists, the authors should provide in supplementary data the exact lineage-derived expression pattern of used markers similarly to the Lineagomics project (https://epic.gs.washington.edu/), or else cite studies that provide these.

160-162 "Correlated cell identities […] using individual cells/nuclei as landmarks": It appears key that the authors present the variability in cell/nuclei positioning across embryos (and for different cell types) used for FM-lineaging. Especially given the use of these positions as landmarks on a single EM sample per timepoint.

165 "Based on timing of known developmental events in the FM series, the aligned EM data are timed": What are the known developmental events? This is not adequately or comprehensively explained. The authors should expand as necessary for a broader audience.

168 "We assess the accuracy of the correlated cell identities": Here and elsewhere in the manuscript, it appears important to compare the accuracy of cell identification and process tracing to other techniques of identification in electron microscopy studies such as CLEM, connectomics studies, or studies of EM-FM registration in other organisms (i.e. in Vergara, 2021 where nuclear-alignment results in 99.0% accuracy and stepwise image registration uses STAR Methods).

187-190 Figure 3g: "auto-fusion of the excretory canal, duct, pore and gland cell.": The images and notions concerning this topic and data should be explained in detail. Text and images lack context and details necessary even to C. elegans readers (unless they are specialists of this structure). Relevant studies should be discussed and cited (Soulavie et al., 2017,PMID: 29717108).

201 "These results are deposited on webKnossos.": The accessibility of the data is commendable. The authors are encouraged to explain what possibilities of annotations/ data-browsing are allowed in the interface. Given the study emphasizes describing fine cell morphologies, providing images with higher resolution and traces of the studied cells/processes across sections would allow better data visualization and integrating community work/ feedback.

215 "appear around the bean stage and a ring around the pharynx becomes visible by the comma stage": Earlier paper showing this should be cited here (Rapti et al., 2017, PMID: 28846083).

246 Figure 4c: "NR showing pioneers, entry points, the approximate extent at bean and comma."

Given there is one schematic in 4c, it is unclear if it corresponds to bean or comma. This should be explained. Entry points/growth can be different between bean and comma stages, being 25min apart.

246,262, 272 The authors often use the term pioneers to denote early growing neurites in the specific entry points. However, in other neuroscience studies including C. elegans studies (also cited by the authors), the term pioneers is used for early entering axons that are functionally important for entry of others. AVJ-AVD (referred to as supralateral commissure pioneers) ASH, RIB, and AWC (referred to as amphid commissure pioneers) have not been shown functionally important. The authors should refer to early entering neurites and not pioneers. Otherwise, pioneer roles should be established with ablation/functional studies, or other studies showing functional importance should be cited accordingly.

252 "Figure 4e: Each dot represents a single neuron or small cluster of neurons. (lim-4:red, unc-86:green, zag-1:blue, ttx-3:magenta, egl-13:cyan, cnd-1:black, ceh-37:gray, ceh-10:purple).": The expression patterns of the used labels are missing here, or please indicate the published source.. Given their expertise in lineaging, it is important that the authors provide (supplementary) the lineage-related expression pattern of used markers as in Lineagomics project (https://epic.gs.washington.edu/). Secondly, at least some of these reporters do not label single cells (i.e. mgIs18, Table S4 labels AIY but also SMDD as per Bertrand and Hobert, 2009; Rapti et al., 2017). Moreover, many of these labels used (Table S4) appear to be non-integrated arrays (egl-13, ceh-37, zag-1, lim-4?) and thus are prone to mosaicism. It is unclear if/how the breakout times of specific neurons are identified in the single-cell resolution. In FM, are these identified by lineaging these markers and identifying processes in single-cell resolution? In EM, are these identified through registration of cell identities and EM-tracing of single-cell axons? The authors should explain. Replying to the above is key to allow contextualizing the findings with other published information on neuropil assembly, for a better understanding of the reader.

255 and 260 "Figure 4f […] Number of neurites visible in g across the EM series.": The authors should explain in detail what each dot represents in EM timepoints: is this the total number of neurites organized in bundles or also includes single neurites visible across EM sections? Does it refer to neurites across the whole embryo or the ones contributing to the NR neuropil? In the latter case, are the NR-contributing neurites defined by cell identities or by tracing in the NR?

279-280 "a sensory structure that is open to the environment […] The dendrites grow via a distinctive mechanism, a collective retrograde extension": Authors should cite earlier papers demonstrating this structure by EM, including in embryos! (Ward et al., 1975,PMID: 1112927; Oikonomou, et al., PLoS Biol. 2011) (Low et al., Development, 2019, PMID: 30683663). The first paper demonstrating the mechanism of retrograde extension should also be cited (Heiman, Shaham, 2009, PMID: 19344940). I find it puzzling that these are missing.

281 “Formation of the embedded opening and elaboration of dendrite morphology occur in embryogenesis but are less understood.”: In fact, mechanisms of the channel’s formation are significantly studied. A series of relevant publications investigate this process, with mechanistic findings (Perens and Shaham, 2006 PMID: 15935778; Oikonomou, et al., PLoS Biol. 2011 PMID: 21857800; Oikonomou et al., 2012 PMID: 22138055; Wang et al., 2017 PMID: 28803967; Bacaj, Lu, Shaham, 2008 PMID: 18245347; Low et al., 2019, PMID: 30683663). These should be discussed and cited.

361 "Last but not least, our post hoc correlation of EM and FM data provides a useful alternative to true correlative EM. Correlative EM is a powerful, but complex to implement […]": For this first demonstration of cell identification with post hoc FM-to-EM correlation, it appears important to validate the method with CLEM of one sample. An embryo expressing a cell landmark marker with sparse labeling (i.e. Pttx-3) could be prepared by correlative EM so as to identify the labeled cells (AIY,SMDD) by CLEM labeling and by the new co-optimization algorithm. This will allow evaluating the accuracy of nuclei estimation. It should be hopefully feasibly, given the author's previous expertise/use of Correlative EM (Kolotuev et al., 2009, PMID: 19807690; Kolotuev et al., 2012,PMID: 22857930; Burel et al.,2018, PMID: 29802150).

339 "While our study is focused on the wild type, one could use live FM to build the appropriate ensemble models to analyze mutants": Using this methodology to study mutants is an exciting perspective! Yet, WT-mutant comparisons require analysis of more than a single individual per genetic condition. The study in its current form does not allow to evaluate the feasibility of larger-scale usage. Also, EM analysis of one sample per condition/timepoint does not inform on possible variability across samples of one timepoint. Performing EM analysis of a second sample for at least one timepoint per technique will allow addressing the issues.

341-343 "We present the image data and identity annotation of our correlated EM series as an accessible public resource […] seed of a community effort that we hope will fully validate annotation"

The public accessibility of the resource is commendable. It would be important to understand what type of data-browsing and annotations are allowed by the interface and the resolution of the data provided. In order for effective browsing, 3D data could be available with the identified cell identities annotated on the data. Process tracing of the identified cells could also be provided on the browsed/ downloadable data. In order for validation and future use of this resource by the community, the data would need to be provided in a downloadable format with the required resolution.

354 "The data-driven modification in the co-optimization algorithm further facilitates such a vision for automated alignment and annotation of complex brain images.": It is unclear to the broader audience to which degree this is feasible, what advances the current method requires. The authors should explain what limitations or difficulties may be faced for the implementation of the approach in more complex images.

Figure S7d "socket cell over three successive frames in the same cnd-1 promoter"

By Lineagomics (https://epic.gs.washington.edu/) cnd-1 is expressed in many more cells than the amphid socket, including cells partaking in the amphid channel (AMsh sheath, AWC, ASG). AMso and AMsh could be imaged distinguishably by performing mosaic experiments or photoconversion or using less broad promoters. If not, it is unclear how the authors distinguish in FM between processes of all these cells in Figure 5d and Figure 7d, given the FM resolution. Because of this, it is unclear what the FM imaging is adding to the EM data and conclusions. The authors should explain.

Recommendations and corrections for text and figures:

62 "EM studies [contributed to] principles of structural organization"

Despite this being a more comprehensive study of embryonic tissues by EM, few past studies in C. elegans embryos contributed principles of structural organization (Soulavie et al., 2017, PMID: 29717108, Rapti et al., 2017, PMID: 28846083, Low et al., 2019, PMID: 30683663). It appears appropriate that such studies are cited here.

98 “we use the nuclear positions, which are also annotated for the EM series, as the common reference (i.e., a Rosetta stone)”: Authors should expand and explain adequately/ comprehensively as necessary for a broader audience.

185 Aligning color-codes to use the same in Figure 3e and b-c-d, will allow the following cell identities.

189 EM images can be shown both with and without color-coding to better visualize cell fusion.

213 “The NR emerges in the later half of embryogenesis”: Earlier paper showing this should be cited here (Rapti et al., 2017, PMID: 28846083).

217 “even neurite topography through combinations of two-color imaging”: Earlier studies with C. elegans two-color FM imaging of neurites (including in the embryo!) can be cited (Heiman, Shaham, 2009, PMID: 19344940, Rapti et al., 2017, PMID: 28846083, Moyle et al., 2020, PMID: 33627875)

220 “We scan through EM data to identify neurites, and backtrace them to soma”: The authors should explain the exact approach used for process tracing (manual, semi-/ automatic, etc).

240-250 Figure 4b-4e. Color-coding is not consistent between models, FM, EM images, i.e. sublateral pioneers are red in Figure 4b EM but light blue in Figure 4c, and Plim-4 which labels sublateral pioneers is red in Figure 4e. Consistent color-coding would help the reader.

Figure 4c, S5a It is unclear if the drawn arrows’ length is proportional to the actual neurite total growth (across EM sections of the same embryo)? For example, what is the relative length of SIA/SMD and ASH, in the model (Figure S5a) and in the actual EM (Figure S5f,g)?

268-271 “comparing the left- and right-side behavior of the first cells to grow into the NR, namely SIADL/R and SMDDL/R, suggests strong symmetry in behavior. Both sides have broken out at the bean stage, with similar-sized outgrowths […] their tips meet precisely beneath the midpoint of the ALA […].”

This is in line with previous studies, indicating such symmetries and ALA’s position, these studies should be cited (Rapti et al., 2017, PMID: 28846083; Insley and Shaham, 2018, PMID: 29590193; Moyle et al., 2021, PMID: 33627875). In order to establish symmetry beyond previous observations, it appears key to providing quantitative data, calculating the length of bilateral processes by tracing them in EM series.

282 The first paper demonstrating this dynamic process should be cited (Heiman, Shaham, 2009, PMID: 19344940).

Figure S5B Shading of EM figure in Box b doesn’t allow neurite visualization. Also (box B), how it is distinguished whether neurons extend neurites or have elongated cell bodies? It would help to provide the EM images with and without color-coding (and larger sizes) to facilitate visualization.

274 “It is not clear […] if there is signaling to coordinate between the two sides.”

Some mechanisms related to the regulation of left-right symmetry of axon guidance are studied in Grossman, Giurumescu, Chisholm, 2013, PMID: 23979582. The paper should be cited.

Figure S3, S4, S5, S6, S7 are mislabeled, 2 figures are abelled S3 and legends do not correspond. Also, scale bars are missing in many supplementary figures and when present, there is no reference of scale bar size in the legend. These should be corrected accordingly.

411 Some alignment of 1.5/2fold AT-EM data (available in webKnossos) is “jittery”, may any additional image registration help eliminate this, for the better visualization of data?

Reviewer #3 (Recommendations for the authors):

The method is well-executed and thoroughly explained, and this manuscript shines in the temporal and spatial precision by which developmental processes are uncovered and described as a result. However much of these biological discoveries were buried in the supplementals. To create more relevance and excitement for this journal’s audience, the authors should re-write parts of the manuscript, with a stronger emphasis on what the limitations of current imaging methods are, what cross-modality analyses improve upon, and particularly by bringing focus to the biological discoveries and putting them into the main article.

In addition to the inherent challenges of EM-FM cross-comparison, in this article, the cross-modality alignment method overcomes two distinct issues: the challenges associated with manual annotation of embryo developmental EM series; as well as integration of spatial resolution of EM with the temporal resolution with FM. These concepts are briefly mentioned in the introduction and throughout the article, but do not become apparent to a non-specialist reader until the end. These current challenges should be further emphasized in the introduction when describing these imaging methods to highlight the benefits of the analysis method presented.

Note: supplemental figures are mis-numbered (there are two Figure S3s, and the following figures do not match figure captions). I will refer to supplemental figures as they are numbered in the figure captions.

Most of Figures S5, S6 and S7 should be integrated into or added as additional main figures. These data beautifully highlight both the temporal and spatial resolution achieved by the cross-modality method used.

How is “adjacency” defined between nuclei that are not physically touching? Are there specific parameters for what distance between cell nuclei counts as adjacent?

Line 236: what is meant by “breakouts observed in the EM at the comma stage span an hour?”. Is this referring to lack of temporal resolution? This statement is confusing because Figure 4e and Table S4 referenced doesn’t seem to show this to be the case.

https://doi.org/10.7554/eLife.77918.sa1

Author response

Reviewer #1 (Recommendations for the authors):

Some issues in the presentation of the manuscript could be improved prior to publication. There appears to be mis-numbering of the supplemental figures and/or their callouts in the text. Although I have a theory as to the correct numbering, this created considerable confusion in the review.

Apologies for the supplemental figure mis-numbering which has been corrected. We have also revised and attaching supplemental figures to the primary figures to better integrate them into the main text.

The cell identification in the EM datasets is a combination of manual identification and automated identification based on the alignment algorithm. The manual identification process could be clarified, as it is extensive (140-240 manual IDs per data set). Table S5 summarizes some ‘landmarks’ used in pre-alignment, could the authors expand on what makes these neurons or other cells ‘distinctive/distinctively shaped’? This table might be of more utility if fully written out at the single-cell level. It would be worth explaining what constitutes the ‘ground truth’ for these manual IDs.

We have revised Methods with a new section 4 focused on Manual Curation to clarify identity validation. We clarify that manual annotation of EM includes two conceptually distinct kinds of annotation. A. Landmarks used in prealignment, uniquely identifiable relatively trivially by appearance (and in some cases by counting along the AP axis within an easy to identify class) or by adjacency to a trivially recognizable cell. B. Curated cell identities recognized either in the same manner, or with greater effort by counting outward in rings from initial landmarks and confirming by morphology or exhaustive accounting. We have significantly expanded on our description in methods of how these identities are established. In addition, we also revise our presentation of these two forms of annotation. Landmark Table S5 (now referenced on reformatting as Supplementary File 1g) has been supplemented with a new Supplementary File 1e providing a more detailed summary of what cells are annotated. A list of confirmed and predicted identities for all cells remains in supplemental data and available on webKnossos.

Presentation and wording: (1) The title states ‘cross modality synthesis’ of EM and FM imaging, however elsewhere it may be more accurately described as ‘image alignment’ or ‘cross modality analysis’.

We retain ‘cross modality synthesis’ in the title and have adjusted text to be more consistent in using this term to describe the synthesis of the EM and FM data sets in the biological demonstrations, as distinct from the ‘alignment’ process itself which enables that synthesis, for which we now consistently use the term alignment.

(2) Why is the EM data a ‘pseudo time series? Pseudotime implies that the time dimension is inferred from non-temporal data, yet the EM data form a sparse time series (4 time points, though described as ‘dense’ in results).

We had used pseudo to differentiate our EM series, assembled from multiple individuals, from a true time lapse data set, but have removed the modifier as well as dense as they seem more likely to confuse rather than clarify our meaning.

(3) What is the definition of ‘breakout’ in the context of neurite growth?

By breakout we mean the event of initial outgrowth of a neurite from a neuronal cell body. We have revised the wording to “initial outgrowth”.

(4) The C. elegans-specific term ‘hypoderm’ is not widely used; for a general audience ‘epidermis’ can be used as it means the same thing.

We have replaced hypoderm with epidermis throughout for clarity.

Reviewer #2 (Recommendations for the authors):

Strengths

– This is a first attempt for a combined study of C. elegans embryos by fluorescent and electron microscopy, by aligning – registering images without the use of correlative electron microscopy.

– This study introduces a landmark-based alignment assigning identities from a set of labelled data to unlabeled data, with co-optimization of unstable landmarks contributing to accuracy.

– This work provides a comprehensive analysis that largely supports and sometimes expands on previously studied processes of neuropil and sensory channel formation in C. elegans embryos.

– The authors provide publicly available data of their EM sections in webKnossos open resource, which is provided in amenable form can be very advantageous for the community.

Weaknesses

– The technical strength of the approach compared to others is not sufficiently demonstrated. There is a sparse actual assessment of the method’s robustness/variability across samples and in comparison to other methods, in terms of data acquisition/analysis. Including comparative analysis of EM sample replicates for each time point, and comparisons across alternative methodologies (algorithms, electron microscopy) will allow substantiating the approach’s superiority and its feasibility for future large-scale use, suggested by the authors in the study.

– Regarding data accessibility, the study sometimes lacks a detailed, comprehensive presentation of data acquisition, analysis, and underlying assumptions/caveats. This refers to instances such as the use of landmark fluorescent markers with multicellular expression to follow single-cell processes, the image registration/cell-identification using landmarks with expression patterns that are not presented in detail, explanations about process tracing and verification in EM sample analysis. The authors invite for community curation/ feedback of a publicly available EM resource, yet the provided resolution and formats (in terms of cell identification, tracing) may not facilitate curation.

– The study sometimes lacks comprehensive, inclusive contextualizing of new findings in relation to knowledge previously published by the authors and by other specialists. That refers to discussed developmental events of lineage progression guiding sample analysis, and findings of excretory pore and neuropil formation, pioneer-neuron concepts, amphid channel assembly.

We thank the reviewer for their comments, in response to which we have further contextualized our findings, clarifying the methods used in our biological case studies and the relationship of these studies to prior literature. We have also elaborated on the performance of our computational method and implications for the study of spatio-temporal consistency in development. Details follow.

I list my comments in order of appearance in the manuscript, which should be addressed for publication.

Suggestions for improved or additional experiments, data or analyses:

86-87 “As part of the effort to test and demonstrate the feasibility of EM time series acquisition[…]”: The authors should discuss any potential variability/ differences arising from using different EM techniques to analyze different timepoints 1,2 and 3,4. Since the two techniques are used “to demonstrate feasibility” it appears key to discuss (later in the manuscript) the comparison of the two approaches in terms of their use in such project, i.e. effort/ implementation of sample preparation/data analysis, the influence of resulting resolution/stitching to cell identity registration, other pros/ cons). More importantly, if a single embryo is analyzed per timepoint, this would be one of my main critiques of the manuscript. At least 2 samples per timepoint will allow assessing the method’s feasibility, robustness, reproducibility, variability.

Regarding lessons concerning EM techniques from our efforts, we note the EM methods used are well established and cited where appropriate, a full technical comparison is beyond the scope of this paper. However, we have added a section to Discussion further contextualizing the pros and cons of the two techniques in the context of the worm embryo and cross-modality analysis.

Regarding the need for multiple data points per stage there are two goals served by replication: I. Establishing the reproducibility of our alignment process itself. II. Investigating the consistency of the biological processes observed.

Regarding I, we have added additional quantification of our alignment method’s accuracy in predicting cell identity. These are derived by aligning additional independent FM embryos against the same three FM embryos used to name the EM data. Since data sets are represented in all cases by abstract point clouds this experiment provides a valid quantitation of our algorithm’s success in identifying cells in a larger set of embryos. Furthermore, the cell names in these independent FM data sets are derived from lineaging and as such provide an orthogonal ground truth, which is more certain and definitive than the manually curated cell identities in the main EM alignment experiments. Finally, it is worth nothing that these FM embryos were imaged compressed between coverslips (unlike the 3 FM embryos used as the model or the EM embryos). The compression creates nonlinear deformations throughout the embryo and success in identifying cells in these compressed embryos further demonstrates the robustness of our method. These additional experiments are now mentioned in results, described in Methods, and performance in them added as Supplementary File 1f.

Regarding II, the biological consistency of events, we believe that in terms of temporal consistency as well as coarse level spatial consistency (e.g. cell soma positions), one should turn to FM for data coverage that cannot be afforded by EM in the foreseeable future. In terms of fine spatial consistency, we agree that one wants additional EM experiments, and we are in the process of doing this, but we also feel that this is outside the scope of the current study, which seeks to establish the feasibility of the approach and provide an initial time series for the community.

92 “sparsely expressed cell membrane marker […] (Figure 1b, S1)”: For the detailed understanding of C. elegans specialists, the authors should provide in supplementary data the exact lineage-derived expression pattern of used markers similarly to the Lineagomics project (https://epic.gs.washington.edu/), or else cite studies that provide these.

The global expression pattern of these markers is not used to ID cells, and in some cases has not been fully established. Rather, in every case, FM identities are established by lineaging expressing cells in the specific data set utilized. We have clarified this in the new Methods section 4.2.

160-162 “Correlated cell identities […] using individual cells/nuclei as landmarks”: It appears key that the authors present the variability in cell/nuclei positioning across embryos (and for different cell types) used for FM-lineaging. Especially given the use of these positions as landmarks on a single EM sample per timepoint.

We agree that characterization of variability is important, and a topic of great interest to us. Variability is a complex topic with spatial and temporal aspects and many perspectives on whether, for example, cell position, at each developmental stage is unimodal and can be meaningfully represented by a mean. The field of development has struggled with this problem and to our knowledge, not come up with succinct or definitive answers. If this is possible, it will require more example data than we have currently available.

This said, in revisions we elaborate on ways that our computational approach might further an understanding of variability. We argue that variability in adjacency is a natural framework in which to characterize positional variability. While we acknowledge our adjacency measure is a computational approximation of the underlying reality of cell contacts, it has the same strong advantages here as in our alignment method such as simplicity and invariance to absolute position. Any other characterization of variability will have its own limitations and caveats.

While our alignment method uses only universally present adjacencies in alignment, there is significant, more granular variation in how consistently adjacencies are observed. We briefly present this adding an additional supplemental figure and supplemental data characterizing the quantitative variability in adjacency at each stage. We illustrate that this model of position can be used to characterize changes between stages and to capture groups of cells with related behaviors over time. This approach may serve as a promising framework for future work.

165 “Based on timing of known developmental events in the FM series, the aligned EM data are timed”: What are the known developmental events? This is not adequately or comprehensively explained. The authors should expand as necessary for a broader audience.

We further clarify our approach for temporal staging in methods section 3.2 and now explicitly reference this in the main text. Briefly, we use overall elongation at comma and 1.5 fold stages. We supplement this with a key cell division at the earlier bean stage where body morphology is less reliable. We also note that though choosing the correct temporal window for alignment is critical, it is not necessary to establish timing to within a few minutes, but rather only to pick a reasonable model (assembled over a 13 min sliding window) to align against.

168 “We assess the accuracy of the correlated cell identities”: Here and elsewhere in the manuscript, it appears important to compare the accuracy of cell identification and process tracing to other techniques of identification in electron microscopy studies such as CLEM, connectomics studies, or studies of EM-FM registration in other organisms (i.e. in Vergara, 2021 where nuclear-alignment results in 99.0% accuracy and stepwise image registration uses STAR Methods).

We agree accuracy is critical and have added additional evaluation of the accuracy of our alignment method in establishing C. elegans cell identities, as detailed above.

Regarding other alignment methods the initial mansucript discusses the performance of alternative computational algorithms. We present their reported accuracy for comparison given it would be difficult, and not necessarily fair to adapt each method to our data given different models and assumptions.

Overall, we would argue that manual identification based on morphology and position is the near universal default approach and there are few other baselines to compare to. This includes Vergara 2021 where cell types are identified by morphology even in the context of extensive aligned expression data. The 99% accuracy in this paper cited is for nuclear detection rather than identity alignment. We did not find any results on identity alignment in that paper.

CLEM could in theory provide ground truth for a subset of cell identities, albeit at the cost of a technically complex method that might compromise the quality of ultrastructure imaging. However, we feel our additional computational experiments provide a better assessment of alignment quality. This is significantly more scalable and allows validation of every cell’s prediction as opposed to a small subset of cells.

187-190 Figure 3g: “auto-fusion of the excretory canal, duct, pore and gland cell.”: The images and notions concerning this topic and data should be explained in detail. Text and images lack context and details necessary even to C. elegans readers (unless they are specialists of this structure). Relevant studies should be discussed and cited (Soulavie et al., 2017,PMID: 29717108).

We have added the reference requested and edited the text here to briefly describe the events and to clarify the context that our point is not to investigate this event, but rather to demonstrate the ability to pick a subtle developmental event, and follow it clearly over time in our series. The example provides a demonstration of how the data can be used, including in the field of cell biology. As we don’t have significant new findings to report here discussion remains compact.

201 “These results are deposited on webKnossos.”: The accessibility of the data is commendable. The authors are encouraged to explain what possibilities of annotations/ data-browsing are allowed in the interface. Given the study emphasizes describing fine cell morphologies, providing images with higher resolution and traces of the studied cells/processes across sections would allow better data visualization and integrating community work/ feedback.

We expand briefly on our description of the webKnossos interface, highlighting relevant functionality, the ability to browse images and annotation together and for users to ‘fork’ and re-annotate the data set. We also expanding on our hope to extend annotation with contributions from the community.

Every method of data distribution has its limitations, but we hope webKnossos will maximize the ability of users to quickly find areas of interest, create, and share annotations.

We have also added a note to the manuscript asking readers to contact the authors for direct access to raw image data.

215 “appear around the bean stage and a ring around the pharynx becomes visible by the comma stage”: Earlier paper showing this should be cited here (Rapti et al., 2017, PMID: 28846083).

We apologize and add this citation.

246 Figure 4c: “NR showing pioneers, entry points, the approximate extent at bean and comma.”

Given there is one schematic in 4c, it is unclear if it corresponds to bean or comma. This should be explained. Entry points/growth can be different between bean and comma stages, being 25min apart.

We clarify the caption of Figure 4c. It now states that the first arrows in the amphid and sublaterals indicate first outgrowth at bean and the later paths and arrows (and the whole path on other neurons) represent approximate comma entry points and outgrowth extent. Entry to the ring proper has not occurred at bean, which represents the first neuronal outgrowths.

246,262, 272 The authors often use the term pioneers to denote early growing neurites in the specific entry points. However, in other neuroscience studies including C. elegans studies (also cited by the authors), the term pioneers is used for early entering axons that are functionally important for entry of others. AVJ-AVD (referred to as supralateral commissure pioneers) ASH, RIB, and AWC (referred to as amphid commissure pioneers) have not been shown functionally important. The authors should refer to early entering neurites and not pioneers. Otherwise, pioneer roles should be established with ablation/functional studies, or other studies showing functional importance should be cited accordingly.

We clarify this removing pioneer and replacing this with early, as our goal is to document first outgrowths along the path of the ring, rather than establish functional guidance ‘pioneering’.

252 “Figure 4e: Each dot represents a single neuron or small cluster of neurons. (lim-4:red, unc-86:green, zag-1:blue, ttx-3:magenta, egl-13:cyan, cnd-1:black, ceh-37:gray, ceh-10:purple).”: The expression patterns of the used labels are missing here, or please indicate the published source.. Given their expertise in lineaging, it is important that the authors provide (supplementary) the lineage-related expression pattern of used markers as in Lineagomics project (https://epic.gs.washington.edu/). Secondly, at least some of these reporters do not label single cells (i.e. mgIs18, Table S4 labels AIY but also SMDD as per Bertrand and Hobert, 2009; Rapti et al., 2017). Moreover, many of these labels used (Table S4) appear to be non-integrated arrays (egl-13, ceh-37, zag-1, lim-4?) and thus are prone to mosaicism. It is unclear if/how the breakout times of specific neurons are identified in the single-cell resolution. In FM, are these identified by lineaging these markers and identifying processes in single-cell resolution? In EM, are these identified through registration of cell identities and EM-tracing of single-cell axons? The authors should explain. Replying to the above is key to allow contextualizing the findings with other published information on neuropil assembly, for a better understanding of the reader.

In FM:

a. Identities are established by lineaging individual cells in each dataset, see response above and clarification in Methods.

b. Some points represent small clusters. We make this clearer in the manuscript as a caveat. We note, however, that in these (few) cluster cases two pieces of corroborative evidence exist. 1. We examine the cell soma shape – these have a tear drop shape indicating that the cell of interest appears to participate in the bundle of outgrowth near its pointed end. 2. Outgrowth for that specific cell was observed in EM around that time. This is clarified in Methods.

c. Mosaicism is present in these data sets. Because of this cell soma are confirmed as abelled in the specific data sets used and identity is established by lineaging, not expression status. In a number of cases loss of expression in neighboring, potentially expressing cells is used to achieve single cell resolution. This is clarified in Methods.

In EM:

Yes, identities are assigned by registration and then confirmed by morphology and relative position as described in expanded Methods.

255 and 260 “Figure 4f […] Number of neurites visible in g across the EM series.”: The authors should explain in detail what each dot represents in EM timepoints: is this the total number of neurites organized in bundles or also includes single neurites visible across EM sections? Does it refer to neurites across the whole embryo or the ones contributing to the NR neuropil? In the latter case, are the NR-contributing neurites defined by cell identities or by tracing in the NR?

We have clarified more explicitly that 4f is quantifying only neurites visible in the specific cross sectional location of the ring indicated.

279-280 “a sensory structure that is open to the environment […] The dendrites grow via a distinctive mechanism, a collective retrograde extension”: Authors should cite earlier papers demonstrating this structure by EM, including in embryos! (Ward et al., 1975,PMID: 1112927; Oikonomou, et al., PLoS Biol. 2011) (Low et al., Development, 2019, PMID: 30683663). The first paper demonstrating the mechanism of retrograde extension should also be cited (Heiman, Shaham, 2009, PMID: 19344940). I find it puzzling that these are missing.

We apologize for the omission of (Heiman, Shaham, 2009, PMID: 19344940 inadvertently deleted during initial manuscript revisions.

We have added Ward et al., here to provide a more comprehensive crediting of work defining the anatomy of the amphid structure. Low et al., is already cited in a what seems to us a more appropriate location in the successive sentence, in the context of the retrograde extension.

We have added Oikonomou et al., 2011 which we apologize for overlooking the more directly relevant context suggested below and the later relevant context of the essentially adult like structure being present in late embryogenesis.

281 “Formation of the embedded opening and elaboration of dendrite morphology occur in embryogenesis but are less understood.”: In fact, mechanisms of the channel’s formation are significantly studied. A series of relevant publications investigate this process, with mechanistic findings (Perens and Shaham, 2006 PMID: 15935778; Oikonomou, et al., PloS Biol. 2011 PMID: 21857800; Oikonomou et al., 2012 PMID: 22138055; Wang et al., 2017 PMID: 28803967; Bacaj, Lu, Shaham, 2008 PMID: 18245347; Low et al., 2019, PMID: 30683663). These should be discussed and cited.

In the text we clarify our statement to refer to limited knowledge of the early processes of formation for the structure of the sensory organ in the embryo, which we believe remains correct.

As noted above we have added Oikonomou 2011, and an additional publication we apologize for overlooking Nechipurenko 2017 to acknowledge prior knowledge regarding embryonic development of the overall sensory structure.

We have also added suggested citations to credit work characterizing necessary molecular machinery for amphid formation: Perens and Shaham 2006 Oikonomou 2012, Wang et al., 2017 for glial formation and Bacaj et al., 2008 for cilia formation.

Low 2019 is already cited in what we feel is a more appropriate context of the overall formation process.

361 “Last but not least, our post hoc correlation of EM and FM data provides a useful alternative to true correlative EM. Correlative EM is a powerful, but complex to implement […]”: For this first demonstration of cell identification with post hoc FM-to-EM correlation, it appears important to validate the method with CLEM of one sample. An embryo expressing a cell landmark marker with sparse abelling (i.e. Pttx-3) could be prepared by correlative EM so as to identify the abellin cells (AIY,SMDD) by CLEM abelling and by the new co-optimization algorithm. This will allow evaluating the accuracy of nuclei estimation. It should be hopefully feasibly, given the author’s previous expertise/use of Correlative EM (Kolotuev et al., 2009, PMID: 19807690; Kolotuev et al., 2012,PMID: 22857930; Burel et al.,2018, PMID: 29802150).

As we mention above we feel that we have addressed this issue with additional algorithm evaluation on FM data sets. While the experiment described would be possible, as we discuss above CLEM validation is neither necessary nor (in general) a scalable method of definitive cell identification. The effort of collecting a CLEM dataset seems out of scope in this project. This seems particularly true as cells can be directly identified manually with great certainty in an in toto 3D EM data set.

339 “While our study is focused on the wild type, one could use live FM to build the appropriate ensemble models to analyze mutants”: Using this methodology to study mutants is an exciting perspective! Yet, WT-mutant comparisons require analysis of more than a single individual per genetic condition. The study in its current form does not allow to evaluate the feasibility of larger-scale usage. Also, EM analysis of one sample per condition/timepoint does not inform on possible variability across samples of one timepoint. Performing EM analysis of a second sample for at least one timepoint per technique will allow addressing the issues.

Establishing our method for a mutant (similar to the WT) requires: lineageing a few embryos from FM, and generate an EM time series. This would be a qualitative step forward, because it is more challenging, in general, to interpret mutant EM cell identity.

We agree that this is perhaps only appropriate for high penetrant mutants, given the scale of labor involved in acquiring sufficient EM of low penetrance mutants. Scaling this up to many mutants is an ambitious and aspiring goal. We suspect this will require further advances in speed and automation of EM.

We elaborate on this briefly in the revised Discussion.

See point 1 above for discussion of WT replication.

341-343 “We present the image data and identity annotation of our correlated EM series as an accessible public resource […] seed of a community effort that we hope will fully validate annotation”

The public”accessibility of the resource is commendable. It would be important to understand what type of data-browsing and annotations are allowed by the interface and the resolution of the data provided. In order for effective browsing, 3D data could be available with the identified cell identities annotated on the data. Process tracing of the identified cells could also be provided on the browsed/ downloadable data. In order for validation and future use of this resource by the community, the data would need to be provided in a downloadable format with the required resolution.

Please see point 7 above.

354 “The data-driven modification in the co-optimization algorithm further facilitates such a vision for automated alignment and annotation of complex brain images.”: It is unclear to the broader audience to which degree this is feasible, what advances the current method requires. The authors should explain what limitations or difficulties may be faced for the implementation of the approach in more complex images.

We have expanded our discussion to clarify this hypothetical application. We elaborate the aspect of this larger problem our approach solves (spatial alignment with variable landmarks) and what additional data (identity information, most likely molecular information from spatial sequencing or FM) and processing techniques would be necessary to achieve our vision of automated spatio temporal atlas construction.

Figure S7d “socket cell over three successive frames in the same cnd-1 promoter”

By Lineagomics (https://epic.gs.washington.edu/) cnd-1 is expressed in many more cells than the amphid socket, including cells partaking in the amphid channel (Amsh sheath, AWC, ASG). Amso and Amsh could be imaged distinguishably by performing mosaic experiments or photoconversion or using less broad promoters. If not, it is unclear how the authors distinguish in FM between processes of all these cells in Figure 5d and Figure 7d, given the FM resolution. Because of this, it is unclear what the FM imaging is adding to the EM data and conclusions. The authors should explain.

Cnd-1 indeed labels a wide array of cells, including a number of amphids. These cells do not effect interpretation of the images:

Most of these cells are not primarily part of this context, overlap between amphid dendrites and the socket and sheath cells is minimal, while FM lacks the resolution to resolve the details of amphid dendrite and glial interaction, the glial cell bodies can be visualized without interference from the dendrite tips.

More concerning is the adjacency of the sheath and socket, both cnd-1 labeled. Changes in the socket, as highlighted in Fig5d and Figure 5 S1d can be visualized because of its approximately 1.5x greater brightness over the adjacent sheath cell. This makes it highly unlikely the signal visualized originates in the dimmer sheath cell. This is more apparent in the included supplemental video. In addition, we point out that this is corroboration of morphology seen in the EM (though easy to overlook without the dynamic context of FM, an example of what FM adds) and not stand-alone evidence.

We have attempted to clarify this in the figure caption.

Recommendations and corrections for text and figures:

62 “EM studies [contributed to] principles of structural organization”

Despite this being a more comprehensive study of embryonic tissues by EM, few past studies in C. elegans embryos contributed principles of structural organization (Soulavie et al., 2017, PMID: 29717108, Rapti et al., 2017, PMID: 28846083, Low et al., 2019, PMID: 30683663). It appears appropriate that such studies are cited here.

We agree, it was not our intention to imply no EM of the embryo has ever been useful previously and we have added a line here citing these as examples where insight has been gained from readily identified structures. We have also added Nechipurenko et al., 2017: https://pubmed.ncbi.nlm.nih.gov/28411364/, additional relevant embryonic EM work in our discussion of the amphid cilia.

98 "we use the nuclear positions, which are also annotated for the EM series, as the common reference (i.e., a Rosetta stone)": Authors should expand and explain adequately/ comprehensively as necessary for a broader audience.

We have rephrased and expanded here to clarify our approach which is further elaborated throughout the paper.

185 Aligning color-codes to use the same in Figure 3e and b-c-d, will allow the following cell identities.

We agree consistent colors are in general helpful, and we have attempted to achieve this.

In 3 B,D which use one consistent color scheme the goal is to highlight the distinct global positions of different tissue types. In 3e,f,g a different, but internally consistent color scheme is used to highlight contrast between adjacent cells. Given limited choices some colors are used in both of these necessary schemes, we have attempted to clarify the color logic in the revised figure caption.

189 EM images can be shown both with and without color-coding to better visualize cell fusion.

We have added an unannotated version of the image to aid clarity.

213 "The NR emerges in the later half of embryogenesis": Earlier paper showing this should be cited here (Rapti et al., 2017, PMID: 28846083).

We have added this citation.

217 "even neurite topography through combinations of two-color imaging": Earlier studies with C. elegans two-color FM imaging of neurites (including in the embryo!) can be cited (Heiman, Shaham, 2009, PMID: 19344940, Rapti et al., 2017, PMID: 28846083, Moyle et al., 2020, PMID: 33627875).

We have attempted to clarify in the text that we are not referring to multi color imaging in general (though this is indeed a vital technique) but specifically referencing the idea of triangulating relative process placement, with the goal of highlighting the different resolutions of information about this available in EM and FM views of the same neurites in Figure 4b. Here we add a reference to Yip and Heiman 2018 pMID 30117807 as another relevant example of what we mean, (while using 3 colors in one pass) but do not feel that the other suggested references, which we cite elsewhere in more directly relevant contexts elsewhere in the manuscript, are apropos here.

220 "We scan through EM data to identify neurites, and backtrace them to soma": The authors should explain the exact approach used for process tracing (manual, semi-/ automatic, etc).

We clarify in Methods that we perform manual backtracing of neurites to soma to establish neurite identity.

240-250 Figure 4b-4e. Color-coding is not consistent between models, FM, EM images, i.e. sublateral pioneers are red in Figure 4b EM but light blue in Figure 4c, and Plim-4 which labels sublateral pioneers is red in Figure 4e. Consistent color-coding would help the reader.

See point 22 above for general comments.

Coloring in panel e was based on source strain, we agree that this is confusing given it is used only in this panel and have recolored data to match the color of panels c,d,h to minimize potential confusion. This is noted in the revised caption.

Figure 4b is in a distinct color scheme to highlight the grouping of the neurites in the EM at the level of the accompanying FM data. We have retained our original color scheme here.

Figure 4c, S5a It is unclear if the drawn arrows' length is proportional to the actual neurite total growth (across EM sections of the same embryo)? For example, what is the relative length of SIA/SMD and ASH, in the model (Figure S5a) and in the actual EM (Figure S5f,g)?

We clarify in the figure legend that the drawing reflects the approximate extent of neurite outgrowth

268-271 "comparing the left- and right-side behavior of the first cells to grow into the NR, namely SIADL/R and SMDDL/R, suggests strong symmetry in behavior. Both sides have broken out at the bean stage, with similar-sized outgrowths […] their tips meet precisely beneath the midpoint of the ALA […]".

This is in line with previous studies, indicating such symmetries and ALA's position, these studies should be cited (Rapti et al., 2017, PMID: 28846083; Insley and Shaham, 2018, PMID: 29590193; Moyle et al., 2021, PMID: 33627875). In order to establish symmetry beyond previous observations, it appears key to providing quantitative data, calculating the length of bilateral processes by tracing them in EM series.

We agree the general notion of symmetry in the nervous system is not new. Our contribution regards the scale and resolution at which it remains symmetric. Our observation is, to our knowledge, the first documentation of developmental symmetry at this level of detail.

To add nuance to our observation, we have measured the lengths of the processes of early neurons (and potential functional pioneers) SIAD and SMDD, we have added a sentence referencing these results which suggest the outgrowth between the left and right is controlled to within a few percent of neurite length.

282 The first paper demonstrating this dynamic process should be cited (Heiman, Shaham, 2009, PMID: 19344940).

As mentioned above we have added this reference.

Figure S5B Shading of EM figure in Box b doesn't allow neurite visualization.

We have added unshaded versions of each panel in these supplemental figures to aid visualization.

Also (box B), how it is distinguished whether neurons extend neurites or have elongated cell bodies? It would help to provide the EM images with and without color-coding (and larger sizes) to facilitate visualization.

We have added a Methods section clarifying our treatment of neurite outgrowths, including how we make the distinction between elongated cell bodies and neurite outgrowth in EM. The distinction is easier in FM data where dynamic information informs interpretation of cell shape. The full EM data sets are provided for independent assessment as well as unshaded versions of panels as mentioned above.

274 "It is not clear […] if there is signaling to coordinate between the two sides."

Some mechanisms related to the regulation of left-right symmetry of axon guidance are studied in Grossman, Giurumescu, Chisholm, 2013, PMID: 23979582. The paper should be cited.

We apologize for this oversight and have added this citation as an example of asymmetric guidance molecule influence.

Figure S3, S4, S5, S6, S7 are mislabeled, 2 figures are labeled S3 and legends do not correspond. Also, scale bars are missing in many supplementary figures and when present, there is no reference of scale bar size in the legend. These should be corrected accordingly.

We apologize for this error in numbering, and for the omission of scale bar dimensions in supplemental figure captions this has been corrected. All image panels have scale bars, in some cases shared between adjoining montage frames within a panel.

411 Some alignment of 1.5/2fold AT-EM data (available in webKnossos) is "jittery", may any additional image registration help eliminate this, for the better visualization of data?

We agree that the alignment of the AT data sets could be improved with additional manual annotation. We intend to improve this in future which will make for a more attractive visualization. Until then, the quality of slice alignment is sufficient to support the most critical tasks, browsing the stack, and following cell bodies and neurites. We have added some additional documentation of our alignment approach in methods, essentially automatic registration, with manual adjustment used only to correct qualitatively misaligned slices.

Reviewer #3 (Recommendations for the authors):

The method is well-executed and thoroughly explained, and this manuscript shines in the temporal and spatial precision by which developmental processes are uncovered and described as a result. However much of these biological discoveries were buried in the supplementals. To create more relevance and excitement for this journal's audience, the authors should re-write parts of the manuscript, with a stronger emphasis on what the limitations of current imaging methods are, what cross-modality analyses improve upon, and particularly by bringing focus to the biological discoveries and putting them into the main article.

We thank the reviewer for their helpful suggestions and have attempted to contextualize our approach more in the text. We have also brought supplemental figures into the main article as secondary figures. Individual points are addressed in greater detail below.

In addition to the inherent challenges of EM-FM cross-comparison, in this article, the cross-modality alignment method overcomes two distinct issues: the challenges associated with manual annotation of embryo developmental EM series; as well as integration of spatial resolution of EM with the temporal resolution with FM. These concepts are briefly mentioned in the introduction and throughout the article, but do not become apparent to a non-specialist reader until the end. These current challenges should be further emphasized in the introduction when describing these imaging methods to highlight the benefits of the analysis method presented.

We have attempted to clarify these two distinct contributions in the introduction.

Note: supplemental figures are mis-numbered (there are two Figure S3s, and the following figures do not match figure captions). I will refer to supplemental figures as they are numbered in the figure captions.

Apologies for the mislabeling of supplemental figures, this has been corrected.

Most of Figures S5, S6 and S7 should be integrated into or added as additional main figures. These data beautifully highlight both the temporal and spatial resolution achieved by the cross-modality method used.

All supplemental figures have been moved into the main text as secondary figures associated with manuscript primary figures or placed in textual form in the Methods section.

How is "adjacency" defined between nuclei that are not physically touching? Are there specific parameters for what distance between cell nuclei counts as adjacent?

We have added an explicit pointer in the main text to the Methods section which describes how our adjacency graph is computed as the Gabriel Graph, a stable subset of Voronoi faces.

Line 236: what is meant by "breakouts observed in the EM at the comma stage span an hour?". Is this referring to lack of temporal resolution? This statement is confusing because Figure 4e and Table S4 referenced doesn't seem to show this to be the case.

We clarify in the text that we mean that cells observed as having grown out in the EM comma embryo span ~an hour in the FM timings of their initial outgrowth.

https://doi.org/10.7554/eLife.77918.sa2

Article and author information

Author details

  1. Anthony Santella

    Molecular Cytology Core, Memorial Sloan Kettering Cancer Center, New York, United States
    Contribution
    Conceptualization, Investigation, Methodology, Software, Visualization, Writing - original draft, Writing - review and editing
    Contributed equally with
    Irina Kolotuev
    Competing interests
    No competing interests declared
  2. Irina Kolotuev

    Electron Microscopy Facility, University of Lausanne, Lausanne, Switzerland
    Contribution
    Conceptualization, Data curation, Investigation, Methodology, Resources, Supervision, Writing - original draft, Writing - review and editing
    Contributed equally with
    Anthony Santella
    For correspondence
    irina.kolotueva@unil.ch
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1433-8048
  3. Caroline Kizilyaprak

    Electron Microscopy Facility, University of Lausanne, Lausanne, Switzerland
    Contribution
    Data curation, Methodology, Resources, Writing - review and editing
    Competing interests
    No competing interests declared
  4. Zhirong Bao

    Developmental Biology Program, Memorial Sloan Kettering Cancer Center, New York, United States
    Contribution
    Conceptualization, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing - review and editing
    For correspondence
    baoz@mskcc.org
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2201-2745

Funding

National Institutes of Health (R01GM097576)

  • Zhirong Bao

National Institutes of Health (R24OD016474)

  • Zhirong Bao

National Institutes of Health (P30CA008748)

  • Zhirong Bao

Chan Zuckerberg Initiative (2019-198110 (5022))

  • Anthony Santella

Faculty of Biology and Medicine of the University of Lausanne

  • Irina Kolotuev
  • Caroline Kizilyaprak

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

Thanks to Bruno Humbel, Jean Daraspe, Helmut Gnaegi, Tilman Franke for advice and assistance with EM, Chris Brittin, Li Fan, Kris Barnes and all the members of the WormGUIDES consortium for advice regarding C. elegans neurodevelopment. Particular thanks to Chris Brittin, Daniel Colón-Ramos, Hari Shroff, Mei Zhen, and Ryan Christensen for advice and comments on the manuscript.

Senior Editor

  1. Claude Desplan, New York University, United States

Reviewing Editor

  1. Oliver Hobert, Columbia University, Howard Hughes Medical Institute, United States

Reviewer

  1. Andrew D Chisholm, University of California, San Diego, United States

Version history

  1. Preprint posted: February 13, 2022 (view preprint)
  2. Received: February 15, 2022
  3. Accepted: June 5, 2022
  4. Accepted Manuscript published: June 6, 2022 (version 1)
  5. Version of Record published: June 21, 2022 (version 2)

Copyright

© 2022, Santella, Kolotuev et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 664
    Page views
  • 166
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: PubMed Central, Crossref, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Anthony Santella
  2. Irina Kolotuev
  3. Caroline Kizilyaprak
  4. Zhirong Bao
(2022)
Cross-modality synthesis of EM time series and live fluorescence imaging
eLife 11:e77918.
https://doi.org/10.7554/eLife.77918

Further reading

    1. Developmental Biology
    2. Neuroscience
    Igor Y Iskusnykh, Nikolai Fattakhov ... Victor V Chizhikov
    Research Article

    Development of the nervous system depends on signaling centers – specialized cellular populations that produce secreted molecules to regulate neurogenesis in the neighboring neuroepithelium. In some cases, signaling center cells also differentiate to produce key types of neurons. The formation of a signaling center involves its induction, the maintenance of expression of its secreted molecules, and cell differentiation and migration events. How these distinct processes are coordinated during signaling center development remains unknown. By performing studies in mice, we show that Lmx1a acts as a master regulator to orchestrate the formation and function of the cortical hem (CH), a critical signaling center that controls hippocampus development. Lmx1a co-regulates CH induction, its Wnt signaling, and the differentiation and migration of CH-derived Cajal–Retzius neurons. Combining RNAseq, genetic, and rescue experiments, we identified major downstream genes that mediate distinct Lmx1a-dependent processes. Our work revealed that signaling centers in the mammalian brain employ master regulatory genes and established a framework for analyzing signaling center development.

    1. Developmental Biology
    2. Evolutionary Biology
    Salvatore D'Aniello, Stephanie Bertrand, Hector Escriva
    Feature Article

    Cephalochordates and tunicates represent the only two groups of invertebrate chordates, and extant cephalochordates – commonly known as amphioxus or lancelets – are considered the best proxy for the chordate ancestor, from which they split around 520 million years ago. Amphioxus has been an important organism in the fields of zoology and embryology since the 18th century, and the morphological and genomic simplicity of cephalochordates (compared to vertebrates) makes amphioxus an attractive model for studying chordate biology at the cellular and molecular levels. Here we describe the life cycle of amphioxus, and discuss the natural histories and habitats of the different species of amphioxus. We also describe their use as laboratory animal models, and discuss the techniques that have been developed to study different aspects of amphioxus.