Protocol for 3D surface texture modeling and quantitative spectral decomposition analysis in Drosophila border cell clusters

Abstract

Drosophila border cell clusters model collective cell migration. Airyscan super-resolution microscopy enables fine-scale description of cluster shape and texture. Here we describe how to convert Airyscan images of border cell clusters into 3D models of the surface and detect regions of convex and concave curvature. We use spectral decomposition analysis to compare surface textures across genotypes to determine how genes of interest impact cluster surface geometry. This protocol applies to border cells and could generalize to additional cell types. For complete details on the use and execution of this protocol, please refer to Gabbert et al.¹.

Keywords: Biophysics; Cell Biology; Developmental biology; Microscopy.

Graphical abstract

Figure 1

Acquire Airyscan images through Zen (A) Full screen view of Zen during Airyscan imaging setup. (B) Zen opens with a view of the Locate tab. (C) Switching to the Acquisition tab enables changes to the image acquisition settings. The LSM Mode is switched to Airyscan and the imaging range matches the fluorophore of interest. A range of about 496–566 nm clearly captures GFP signal. (D) View of the Channels window. Only the track(s) of interest is selected. Laser Power and Master Gain are adjusted for imaging. (E) View of the Acquisition Mode window. Frame size is decreased during initial set up but is increased to Optimal settings during sample image acquisition. Scan area can be adjusted. (F) Histogram of signal range at current imaging settings. The histogram covers about half of the total range, illustrating optimal imaging settings. (G) View of the Airyscan Detector. Checking the “Detector” box shows the detector. This illustrates if the detector is properly calibrated or if it requires alignment.

Figure 2

Map the surface of the cluster using ImSAnE (A) Subhyperstack maker in FIJI. This window allows for the selection of specific channels and slices of interest to include in the analysis. (B) Show Info window in FIJI. This window provides a list of image specifications and details associated with the image including dimensions and voxel size. (C) The Dataset Properties window in Ilastik allows for the adjustment of properties when importing compressed h5 files. Ilastik interprets axes in a different order than the other softwares utilized in this protocol, so the axes order must be transposed during data import and export with Ilastik. (D) Features selection window in Ilastik. The choice of features corresponds to what area and what characteristics the pixel classification software will consider during training. It is recommended to include all areas and characteristics for accurate results. As the data is in 3 dimensions, this training is performed in 3D. (E) Training pixel classification software in Ilastik. The yellow labels signify pixels to classify as part of the object of interest and the blue labels signify pixels to exclude. This labeling should be performed on a range of z-slices and should help clarify ambiguous areas. (F) Running “Live Update” in Ilastik after preliminary label markups shows a predicted segmentation of the object based on the current labeling information. Viewing regions of uncertainty in live mode can highlight regions that require further labeling. (G) Image Export Options window in IIlastik. Similar to importing data into Ilastik, exporting data from Ilastik also requires transposing of the axis order for downstream analysis in additional softwares. For use in MATLAB, the order is switched back to “zxyc” before export. (H) 3D point cloud model in MATLAB. The data exported from Ilastik and imported back into MATLAB is represented as a 3D point cloud. This model should clearly illustrate the 3D surface of the cluster without additional objects present or disruptions of the surface such as flat regions or inclusion of background areas. This model can rotate in space.

Figure 3

Generate 3D meshes in MeshLab (A) Poisson-disk Sampling window in MeshLab. After importing the h5 file with the point cloud from MATLAB, Poisson-disk sampling is performed on the mesh. The number of samples should be chosen with attention to the complexity of the surface; more complex surfaces should have higher sample numbers (∼10000), while smoother or rounder surfaces should have lower sample numbers (∼5000). Starting at 7000 is generally an efficient choice. Base Mesh Subsampling should be selected and then click “Apply”. (B) If the resulting mesh following surface reconstruction is dark, this means that the surface is inverted from the actual cluster. This is resolved by inverting the face mesh. (C) The surface should appear light if the border cell cluster surface has the correct side facing outward. (D) After verifying that the surface is not inverted, then the file is exported. In the Saving Options window in MeshLab, select ‘Normal’ under ‘Vert’, then click ‘OK’.

Figure 4

Spectral decomposition of border cell cluster shape Decomposing the surface into spherical harmonics provides a quantitative measure of shape. (A and B) Mapping each mesh to a sphere using conformalized mean curvature flow provides a measure of protrusion from a reference surface. (A) The mesh triangulation of the border cell cluster surface acquired earlier maps to a sphere in a manner that preserves angles of the triangulation — a ‘conformal’ map. (B) Subtracting each mesh vertex’s radial coordinate $R\left(\theta ,\phi \right)$ from the radius of the sphere $R_0$ provides a measure of radial distance $\delta r$. Note that radial distance measurement is patterned on the sphere using the mapped configuration’s spherical coordinates $\left(\theta ,\phi \right)$, so that $\delta r$ gives the radial displacement that each vertex acquires while mapping the spherical mesh to the true surface geometry. (C) We then decompose this signed distance field on the sphere into components of increasingly fine spatial scale using the spherical harmonics as a set of basis functions. The coefficients $a_l^m$ provide a measure of spectral weight for each pattern of deformation from a spherical state. For clarity, we show only one pattern ($m=0$) for each index $l$.

Figure 5

Expected outcomes Comparing shape spectra across conditions reveals the effect of septin perturbations on border cell cluster shape at different spatial scales. (A) Our analysis yields a measure of spectral weight. We chose to add the weights indexed by each index $m$, which ranges between $-l<m<l$ for each index $l$, so that $A_l={\Sigma }_{m=-l}^l\left|a_l^m\right|$. For different conditions, the spectral weights have the same general trend, but differ quantitatively in the amount of spectral weight across different shape indices $l$. (B) Increasing septin expression increases spectral weight at high values of $l$ (fine texture of the surface). (C) Increasing septin expression reduces spectral weight for $l=1$, which is a measure of unilateral protrusion of the mesh relative to a spherical reference geometry. (D) Septin expression is correlated with greater surface roughness. Error bars represent standard error, and shaded regions in (A-B) represent standard deviations. n = 4 for control, 10 for knockdown, and 3 for overexpression. ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001 when analyzed by one-sided t tests.