Epidermal Cell Surface Structure and Chitin–Protein Co-assembly Determine Fiber Architecture in the Locust Cuticle

The geometrical similarity of helicoidal fiber arrangement in many biological fibrous extracellular matrices, such as bone, plant cell wall, or arthropod cuticle, to that of cholesteric liquid mesophases has led to the hypothesis that they may form passively through a mesophase precursor rather than by direct cellular control. In search of direct evidence to support or refute this hypothesis, here, we studied the process of cuticle formation in the tibia of the migratory locust, Locusta migratoria, where daily growth layers arise by the deposition of fiber arrangements alternating between unidirectional and helicoidal structures. Using focused ion beam/scanning electron microscopy (FIB/SEM) volume imaging and scanning X-ray scattering, we show that the epidermal cells determine an initial fiber orientation, from which the final architecture emerges by the self-organized co-assembly of chitin and proteins. Fiber orientation in the locust cuticle is therefore determined by both active and passive processes.

wavelet transform, the wavelet family and the width of the Gaussian filter were visually optimized to obtain a corrected stack with the least contrast and information loss and the largest stripe removal.
Finally, to improve the signal to noise ratio, the image stacks were denoised using the skimage.restoration non-local denoising filters of the scikit-image python library (v. 0.14.0). Again, the filtering algorithms and parameters were chosen by visually optimizing the corrected images to obtain the largest noise removal with the lowest information loss.
Segmentation of structures of interest was performed semi automatically using a combination of Amira 3D (FEI, USA) and a custom python code implementing state of the art 2D and 3D machine learning methods based on fully convolutional networks using Keras and GPU accelerated 2 Tensorflow. For the 2D segmentation between 20 and 30 representative 256 x 256 regions of the image stack were sparsely annotated using Amira 3D. The images and ground truth labels were augmented using the Keras ImageDataGenerator class by allowing a rotation and shearing. The augmented images were used to train a U-net like network which consists, for the contracting path, of the repeated application of 3 x 3 convolutions (padding = same) followed by a ReLU activation and 2 x 2 max pooling for downsampling. For the expanding path, every step consists of a 2 x 2 upconvolution concatenated with the corresponding layer of the contracting path and the repeated application of 3 x 3 convolutions followed by a ReLU activation. A dropout of 0.25 after each max pooling and upconvolution operations was applied. A total of 4 max pooling and upconvolutions were used (depth of the network = 4). On the final layer, a sigmoidal activation was employed.
The input images and the corresponding ground truth labels were used to train the network on a workstation with a single Nvidia Quadro M5000 GPU with 8GB of memory with the adam optimization algorithm, a learning rate of 10 -4 and binary cross entropy as a loss function. Once the loss dropped below 1%, the training was stopped and the network was used to infer the labels on the whole image stack. Although the network would predict the microvilli with a high accuracy, many of the images would contain a small number of predicted labels not coinciding with the objects of interests. These were manually deleted from the image stack using Amira 3D.
For the 3D segmentation, a similar workflow was used: first, 2 representative sub volumes (512 x 512 x 512) px 3 of the image stack were selected and densely manually annotated using Amira 3D. From these 2 volumes, we extracted randomly 30 smaller volumes (92 x 92 x 92) px 3 . Each of these 30 volumes was randomly transformed 3 times to generate 3 'augmented' volumes using one of the following transformations: identity, reflection with respect to the x or y axis, 10% zoom in or out and 10% dilation or compression on a random axis (x, y or z). The generated augmented volumes were randomly cut to a final size of (64 x 64 x 64) px 3 not containing boundary artifacts associated with the transformation (like for zooming out and compressing). The final 90 volumes (64 x 64 x 64) px 3 were used to train a 3D network with an architecture slightly modified with respect to the one used for the 3 2D segmentation. The basic building blocks of the network remained unchanged, but we used convolution kernels of (5 x 5 x 5) px 3 , a network's depth of 2 instead of 4 and the layers in the contracting and expanding path were not concatenated.
The network was trained with the same hardware and software implementation as in two dimensions. In this case, the training was stopped once the loss reached a value of 0.05 and the network used to infer the whole volume. Because of the limitation of the GPU memory size (8GB), the inference was run by tiling the whole volume in tiles of (128 x 128 x 128) px 3 overlapping for half of their size. For each tile, only the inferred labels in the central (64 x 64 x 64) px 3 were used to generate the final predictions on the original image stack. Because the training sub-volumes were taken from the image stack the inference was run on, the quality of the predictions in those regions was superior to the others. Nevertheless, large portions of the stack neighboring that were selected for the training exhibited also accurate predictions. Also, in this case, the predicted labels were checked with Amira 3D and the wrongly predicted labels were manually deleted. The Drishti volume exploration and representation tool (v.2.6.5) was used for volume rendering [2]. Calibration, integration and data analysis of the 2D diffraction patterns were performed using the software DPDAK [3]. DPDAK was also used for background removal and peak-fitting the XRD patterns. For each peak, several fits were performed where the fitting parameters (q position, intensity and peak width) were first varied and then fixed. We used Lorentzian peak-shape for fitting in q space (neglecting micro-strain contribution) and Gaussian peak-shape for fitting in azimuthal   and C correspond to features depicted in Figure 6E, F and G in the main text.