A photo-switchable assay system for dendrite degeneration and repair in Drosophila melanogaster

Significance We introduced a versatile system for photo-switchable caspase-3 activation in developing and mature Drosophila dendrite arborization neurons to induce degeneration and monitored ensuing dendrite degeneration, protection, repair, regeneration, and cell death. Using this assay system, we observed the protection afforded by Wallerian degeneration slow (WldS) upon photo-switchable caspase-3–induced neurodegeneration and further demonstrated that WldS does not improve dendrite regrowth or regeneration in development and adulthood, respectively. Because of the ease and flexibility to systematically induce neuronal injury, this assay system should facilitate uncovering the underlying cellular and molecular mechanisms by doing genetic screens. This photo-switchable assay system can provide physiologically relevant insights because caspase-3 is involved in the developmental pruning of axon and dendrite, injury-induced neurodegeneration, and neurodegenerative diseases.


Illumination setups
For illuminating larvae, we used a homemade 40 cm x 10 cm x 15 cm carbon box to house three 30 cm long and 8mm wide Blue 3528 LED strips, (single row of 60 LEDs per meter, Environmental Lights) stick on the ceiling of the box in parallel and connected by wires. The power of the light 15 cm away from the LED strips, where larvae were kept, is 0.91 mW/cm 2 .
For illuminating adult flies, three 7.5 cm long and 10 mm wide Blue 3528 LED strips, (single row of 240 LEDs per meter, Environmental Lights) were attached to a piece of glass in parallel and connected by wires. This piece of glass with LED strips was kept 2 cm away from the flies during illumination. To maintain the temperature, we install aluminum heat sinks with a 120x120x38mm 24V DC industrial cooling case fan (Wathai) turned on throughout the illumination period. The power of the light 2 cm away from the LED strips, where adult flies were kept, is 29.5 mW/cm 2 .

In vivo time-lapse imaging
Live imaging of larvae was performed as described before (7). Larvae were anesthetized with diethyl ether for 5-8 minutes (Acros Organics) before being mounted in glycerol on top of a thin patch of agarose. For adult flies, we use the carbon dioxide source as an anesthetic throughout the imaging sessions. After images were acquired using a Leica SP5 microscope with a 20X oil objective (NA 0.75), larvae or flies were returned to yeasted grape juice agar plates or yeasted food vials. Sum slices for Z-projection were generated using ImageJ software and used for dendrite structure prediction as described later.
Deep learning based-automatic dendrite structure prediction For the deep learning based-model for automatic dendrite structure segmentation, we followed the U-Net architecture specified in the original study (8) by modifying the channel number of the final segmentation map from 2 to 1 since we only predicted dendrite structure versus background. Each training data consisted of a maximum intensity Z-projection image of one neuron manually cropped by drawing an ROI, paired with the manually segmented dendrite structure (mask) generated using the plugin, "simple neurite tracer", in ImageJ. For the larval model, we generated 29 sets of imagemask pairs for training and 8 sets for validation. For the adult model, we generated 65 sets of image-mask pairs for training and 17 sets for validation. These datasets were all generated inhouse. We trained our model on a Quadro P5000 Graphics processing unit (GPU) with 16 GB random-access memory (RAM) in a Dell Precision 7920 Tower with Dual Intel Xeon Gold 6136 central processing units (CPUs) (3.0/3.7 GHz), having 12 cores and 128 GB RAM. The operating system was Windows 10. We have tested our system on Mac and Windows operating systems. Two data augmentation strategies were used to increase the robustness of our model. First, an area of 512x512 pixels was randomly cropped from each input 1024x1024 training image and the associated mask. Then the cropped image and mask were randomly flipped horizontally and vertically with a probability of 0.5. We used the sum of binary cross-entropy and Dice loss (defined as 1 -Dice coefficient) as the loss function and trained the model with Adam optimizer at learning rate 1e-4 for 500 epochs. The best model evaluated by Dice loss using the validation dataset was chosen for the downstream analysis. Our larval model achieved the Dice loss at 0.13 and 0.16 for training and validation datasets, respectively and our adult model achieved the Dice loss at 0.12 and 0.17 for training and validation datasets, respectively.
A threshold of 0.5 was used to binarize segmentation maps generated by the model. We found a high correlation (R 2 = 0.98) in the total dendrite length of larval neurons between modelpredicted segmentation and manual reconstruction, while tip numbers only showed a moderate correlation (R 2 = 0.45). This was because tip number was more sensitive to discontinuity and small fragments occasionally found in model-predicted segmentation masks. Therefore, we included a 3-step post-processing procedure to exclude small fragments and reduce the discontinuity in the segmented dendrite structure. First, small objects with areas less than 10 pixels (7.75 μm) were discarded. Second, dilation with a cross-shaped structuring element (connectivity=1) was used to fill in the gaps. Finally, skeletonization using the skeletonize function from Python scikit-image package was applied to obtain the final segmentation for the downstream morphology quantification. With post-processing to fill in gaps and remove small fragments, we observed a dramatic increase in the correlation of tip numbers (R 2 = 0.97) and a slight increase in total dendrite length (R 2 = 0.99). The post-processing procedure also greatly improve the correlation of tip numbers (R 2 = 0.78 to R 2 = 0.99) and total dendrite length (R 2 = 0.98 to R 2 = 0.94) for the predictions made with adult models.
This deep learning based-automatic dendrite structure prediction system can be applied to predict the structures of other types of neurons using the exiting models or applied to create models with a new set of training datasets as long as one can separate the individual neurons at the manual ROI selection step.

Quantification of dendrite structure
Using the skeletal images, we performed Sholl analysis of dendrite branches to determine the complexity of the dendrite structure. The crossing continuous circles were separated by 0.76μm on either manually traced or predicted dendrite arbors. To determine the percentage of territory covered, we measured the territory covered by the dendrite arbor of the neuron of interest using ROI selection tools in ImageJ and divided it into the total area of the hemisegment of the body wall. We defined a cell as "survived" if the neuron has more than 2 tips (at least more than one neurite) and if the average dendrite length (total dendrite length/total tip numbers) is over 10 pixels (7.75 μm) to filter out small fragments that mostly contributed by the remain axons or debris. To reduce the batch-by-batch variations, we normalized the quantifications to the controls for each batch before combining all data. For comparison between different conditions, the number was normalized to the averaged number in dark (control). The results are normalized to the controls for each set of experiments before combining.

Statistical Tests
The Student's t test was used to compare between two groups. One-way ANOVA with Tukey's post hoc test was used for comparisons of multiple groups. The Kruskal-Wallis rank sum test with Dunn's post hoc test further adjusted by the Benjamini-Hochberg False Discovery Rate (FDR) method was used for multiple comparisons of nonparametric samples. Two-way ANOVA with Tukey's post hoc test was used for comparisons of interaction between two factors.           caspase-LOV and labeled by ppk-CD4-tdTom following 10 min illumination with EtOH (control, 3 columns on the left), 0.5 mM RU486 (3 columns in the middle), or 1 mM RU486 (3 columns on the right). The same neuron was imaged before drug treatment (-1 d) and at 1-3 d after illumination started (from top to the bottom row). We observed different local degeneration events following 10 min illumination in the control group and the group with mild drug treatment. We also observed dendrite regeneration in neurons treated with 0.5 mM RU486. The sites of local degeneration were indicated by red arrowheads and the dendrite addition (A) was marked by red solid arrowheads. Local degeneration included dendrite branch severing (S), dendrite branch fragmentation (F), dendrite blebbing (B), dendritic debris clearing (C), and engulfment of dendritic debris after breakdown (E) as noted next to the red arrowheads. The same region of dendrite arbors compared over time was outlined with a dashed green line. Scale bars =100 μm.