An ensemble-averaged, cell density-based digital model of zebrafish embryo development derived from light-sheet microscopy data with single-cell resolution.

A new era in developmental biology has been ushered in by recent advances in the quantitative imaging of all-cell morphogenesis in living organisms. Here we have developed a light-sheet fluorescence microscopy-based framework with single-cell resolution for identification and characterization of subtle phenotypical changes of millimeter-sized organisms. Such a comparative study requires analyses of entire ensembles to be able to distinguish sample-to-sample variations from definitive phenotypical changes. We present a kinetic digital model of zebrafish embryos up to 16 h of development. The model is based on the precise overlay and averaging of data taken on multiple individuals and describes the cell density and its migration direction at every point in time. Quantitative metrics for multi-sample comparative studies have been introduced to analyze developmental variations within the ensemble. The digital model may serve as a canvas on which the behavior of cellular subpopulations can be studied. As an example, we have investigated cellular rearrangements during germ layer formation at the onset of gastrulation. A comparison of the one-eyed pinhead (oep) mutant with the digital model of the wild-type embryo reveals its abnormal development at the onset of gastrulation, many hours before changes are obvious to the eye.

030-A and AC254-060-A, Thorlabs, Newton, USA) the light beam is guided by a vertically deflecting galvo mirror scanner (8300K, Cambridge Technology, Bedford, USA) to an f-theta lens (S4LFT4375, Sill Optics, Wendelstein, Germany). To enable sample illumination from two opposite sides, the excitation beam is split by a 50/50 beamsplitter (NT49-004, Edmund Optics, Barrington, USA) into two identical illumination arms containing tube lenses and 5×/0.15 air objectives (Nikon, Tokio, Japan). For alignment of the deflected beam (upper illumination path in Fig. S1A), the beamsplitter is positioned on a translational stage (PT1, Thorlabs, Newton, USA) so that it can be moved along the straight beam optical axis (bottom illumination arm in Fig. S1a). For adjustment of the excitation beam waist to specific positions along the optical axis through the specimen, the illumination objectives are placed on translational stages (CT1, Thorlabs, Newton, USA) attached to a common cage system. Homogeneous light sheet illumination is achieved by a linear galvo mirror displacement over the field of view at the rate of 450 Hz. Consequently, for a typical frame acquisition over 40 ms, the light beam scans across the sample more than 35 times, which helps to avoid artefacts related to discrete galvo scanner displacement or camera rolling shutter mode.
The detection unit consists of a 16×/0.8w water dipping objective (Nikon), bandpass filter (Semrock BrightLine HC 525/50 and Semrock BrightLine HC 641/75, AHF, Tübingen, Germany), tube lens (Nikon), and sCMOS camera (Neo, Andor, Belfast, Northern Ireland). With the camera chip size of 16.6 × 14 µm 2 and the 16× magnification of the detection objective, the entire field of view is 1038 × 875 µm 2 , with a pixel size of 0.4 × 0.4 µm 2 , which fits well to the size of zebrafish embryos (up to 750 µm in diameter). Precise alignment of the focal plane of the detection objective with respect to the illumination light sheet is achieved by placing the objective with the firmly attached sample chamber on a linear stage equipped with a differential micrometer (M-462-X-SD and DM-13L, Newport, Irvine, USA). Opening on top of the sample chamber enables fast and easy sample replacement.
To synchronize XY stage movement with image acquisition during 3D stack recording, the translational stage is pre-programmed before the start of the measurement to perform the desired frame-to-frame displacement. Subsequently, program execution is triggered for each frame capture cycle by the camera fire output.
Sample mounting. Fluorescence microscopy measurements were performed on fertilized eggs obtained by overnight crossing of adult zebrafish Tg(h2afva:h2afva-GFP) and Tg(tdgf1m134/m134) lines containing the transgenic H2A-GFP fusion protein. Embryo mounting on our DSLM setup was adopted for optimal performance in the application to very early developmental stages of zebrafish ( Fig. 1c). Thus, short fluorinated ethylene propylene (FEP) tubes (BOLA-tube, outer/inner diameters: 1.6/0.8 mm, Bohlender GmbH, Grünsfeld, Germany or Thomafluid-High-Tech-tube, outer/inner diameters: 1.5/1.1 mm, Reichelt Chemietechnik GmbH, Heidelberg, Germany) were first flushed with 0.1 mg/ml BSA solution to prevent attachment of embryos to the tube walls. Then, a small amount of gel solution containing 1.5% low melting agarose (Sigma-Aldrich, München, Germany) with sparsely distributed 100-nm yellow-green fluorescent beads (Invitrogen, Carlsbad, USA) was soaked into the FEP tube and kept for a few minutes at room temperature to allow the agarose to harden. Afterwards, embryos with manually removed chorion were gently embedded into a solution containing 0.1% low melting agarose and soaked into the remaining part of the FEP tube so that the 1.5% agarose plug was underneath the embryo. Finally, the FEP tube was attached to the sample holder and placed inside the sample chamber. The sample chamber is a diver's helmet assembly of ~15 ml volume made from a stainless steel hollow cube. It has a quadratic cross section; three of its four openings are closed by sapphire windows, the fourth one contains the detection objective. To maintain a constant temperature of typically 26°C inside the sample chamber during the experiments, a heating foil (HK5590, Telemeter Electronic, Donauwörth, Germany) is attached to the bottom of the chamber and connected to a temperature controller (TC200, Thorlabs, Newton, USA). The temperature is monitored by a Pt100 resistance thermometer inserted into the sample medium inside the sample chamber and connected to the temperature controller.
Despite very careful embryo mounting, the animal-to-vegetal pole axis was often tilted from the normal to the surface of the 1.5% agarose layer supporting the embryo. As cells migrate toward the vegetal pole, their movement becomes slowed once they encounter the 1.5% agarose surface. The cells located in the tilt direction experience the frictional effect first, whereas cells on the opposite side are still exposed to 0.1% agarose solution for some time until they reach the bottom. This asymmetry in cell movement shifts the pole axis toward the point at which the embryo touches the agarose plug. Of note, the simultaneous bi-directional illumination induces some scattering on the opposite (with respect to the excitation) side of the sample due to beam divergence. Foci displacement, however, also helps to reduce this effect, specifically for zebrafish embryos during early stages of development, when a thin cell layer covers the yolk sphere. Furthermore, we have tested the stability of the microscope by monitoring positions of fluorescent beads over many hours under conditions simulating measurements with zebrafish embryos. The axial and lateral displacements of the bead positions over 12 h were less than (1.2 ± 0.3) µm and (3.5 ± 0.5) µm, respectively, i.e., less than the optical resolution of the microscope.
Data management. Long-term DSLM data storage is supplied by the Large Scale Data Facility (LSDF) at the Steinbuch Centre for Computing (SCC) at the Karlsruhe Institute of Technology (KIT), which provides an excellent infrastructure with a huge storage capacity (>300 TB for the experiments reported here and >6 PB in total), secure backup of the data in a tape library, fast and parallel access to big data sets and computing facilities. Convenient data handling on the LSDF is assured by the DataBrowser software, which extracts the OME data from each uploaded file and added this information to a metadata database. Moreover, the software provides the possibility to perform standardized image processing, e.g., multi-view fusion, maximum intensity projection (MIP) video compilation or cell nuclei segmentation, directly on the data stored at the LSDF by utilizing a distributed computing environment based on the Apache™ Hadoop® cluster. The cluster consists of 25 nodes, each of them equipped with 2 quad-core Intel Xeon E5520 CPUs @ 2.27GHz and 36 GB of memory. All methods are implemented in custom C++ software using the Insight Toolkit SDK (www.itk.org).
Image processing. Cell nuclei segmentation was performed on each 3D image stack individually. Seed points are detected by finding local extrema in the 26neighborhood of each pixel on a Laplacian-of-Gaussian (LoG) space-scale maximum projection, which was iteratively calculated using LoG filtered images.
For every detected seed point, a cuboid with side lengths proportional to the radius of the respective blob is cropped out and processed independently, effectively utilizing parallelization to speed up the entire segmentation algorithm. views were fused if their distance was smaller than the minor axis. Objects were combined by a weighted sum of the feature vectors. The weighting was performed according to the distance from the detection objective, i.e., closer objects contributed more to a fused object, to preserve feature quality of objects that were detected in sharp image regions while reducing the weight of nuclei that were detected in low quality image regions.
To reveal the migration pattern of individual cells during gastrulation, temporal correspondences of the detected nuclei have been identified using a nearest neighbor tracking algorithm implemented in the Matlab toolbox Gait-CAD.
During the first iteration, the maximum displacement speed of associated objects between two time points was set to 0.2 µm/s; objects missing in a single frame and reappearing in a subsequent one were assigned to the same trajectory. With this procedure, we achieved an average frame-by-frame association rate of 89.1% including initial time points where nuclei detection was hardly possible. In the second iteration, incomplete tracks were fused over a larger time span using an additional nearest neighbour approach to connect start and end positions of the identified tracks in the spatio-temporal domain. The maximum temporal distance was constrained to ±300 s, with a permitted displacement speed of 0.17 µm/s based on the average nucleus movement. We note that complete lineage reconstruction at the single cell level was not possible over the entire time interval; still, trajectories were revealed spanning several hours.