Naert T , Çiçek Ö , Ogar P , Bürgi M , Shaidani NI , Kaminski MM , Xu Y , Grand K , Vujanovic M , Prata D , Hildebrandt F , Brox T , Ronneberger O , Voigt FF , Helmchen F , Loffing J , Horb ME , Willsey HR , Lienkamp SS .

xenCAKUT - 891127 H2020 Marie Skłodowska-Curie Actions, KA5060/1-1 Deutsche Forschungsgemeinschaft, DK-076683-13 NIH HHS , Overlook International Foundation, National Institutes of Mental Health Convergent Neuroscience Initiative, 1U01MH115747-01A1 Psychiatric Cell Map Initiative, 310030_189102 Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung, Swiss National Centre of Competence in Research Kidney Control of Homeostasis, ERC-StrG DiRECT - 804474 Horizon 2020 Framework Programme, U01 MH115747 NIMH NIH HHS, P40 OD010997 NIH HHS , RC2 DK122397 NIDDK NIH HHS , R01 HD084409 NICHD NIH HHS , R24 OD030008 NIH HHS , R01 DK076683 NIDDK NIH HHS

                CAKUT
                polycystic kidney disease
                branchiootorenal syndrome

	Fig. 1. Deep learning for 2D and 3D phenotyping of altered pronephros development in Xenopus. (A) Schematic of unilateral TBX18 expression by injection of mRNA at the four-cell stage in X. laevis embryos. Kidney development was assessed on both sides at NF stage 38. (B) Fluorescence microscopy image of an LE-Lectin stained embryo. The gray rectangle indicates the region shown at higher magnification in C. Scale bar: 500 μm. (C) Overexpression of RFP or truncated TBX18 (G337V*) did not affect right-sided kidney morphology when compared with left non-injected side. In contrast, overexpression of TBX18 wild-type and A164T resulted in unilateral kidney hypoplasia. Scale bar: 500 μm. (D) A neural network (TubuleNet) was trained to assess pronephros morphology. Top to bottom: input image of an LE-lectin stained pronephros, the output as a segmentation mask, overlay of input image and mask, and skeletonized mask for feature extraction. Scale bar: 500 μm. (E) TubuleNet accurately segmented renal tubules across a wide phenotypic range (normal, hypoplastic, absent). Scale bar: 500 μm. (F) TubuleNet segmentations closely correlated with expert human annotators on unseen data. r, Pearson's correlation coefficient. (G) Plot of pronephros bounding box area per expressed construct shows high agreement between expert annotator and TubuleNet segmentation. L, left (uninjected) side; R, right (injected) side. Boxes and whiskers indicate interquartile range and variability outside the upper and lower quartiles. (H) mesoSPIM light-sheet microscopy in toto imaging of a wild-type embryo stained for LE-Lectin (cyan) and Atp1a1 (red). Scale bar: 200 μm. (I) The two channels were merged as input images (top), accurate segmentation by 3D-NephroNet in the volume. Scale bar: 200 μm. (J) 3D segmentation for feature extraction and volumetric measurements. Unilateral expression of TBX18 H524Y resulted in a reduction of 66% in tubule volume on the injected side. Scale bar: 40 μm.
	Fig. 2. Deep-learning analysis of a pkd1 crispant model for autosomal dominant polycystic kidney disease (ADPKD). (A) Immunofluorescence microscopy showed pronounced tubular cystogenesis upon mosaic inactivation of pkd1 in X. tropicalis. (B) Confocal laser scanning microscopy (CLSM) revealed epithelial thinning and luminal expansion in pkd1 crispants. CystNet3D was used to segment cysts in CLSM stacks (yellow). (C) mesoSPIM light-sheet microscopy in toto imaging of a wild-type embryo and a pkd1 crispant stained for LE-Lectin and Atp1a1. The near-isotropic mesoSPIM recordings were optically resliced on the xz-plane. 3D-NephroNet segmentations of pronephric tubules in mesoSPIM recordings. (D) Image processing pipeline for automated quantification of pronephros area and assessment of cystic index. Pronephric area was measured by 2D-NephroNet in input images. Segmentation masks were used to extract the pronephros, then processed by 2D-CystNet. 2D-CystNet outputs two softmax channels, each of which corresponded to either ‘normal’ or ‘cystic’ morphology. (E) Fully automated measurements of kidney tubule area and percentage of cystic area across three developmental stages of X. tropicalis. Area measurements of pronephric tubules were not significantly different when comparing tyrosinase (tyr) control crispants with pkd1 crispants. In contrast, measuring the percentage of cystic areas detected a significant difference between tyr and pkd1 crispants. (Kruskal–Wallis with Dunnett's multiple comparison: ns, not significant, **P<0.01, ***P<0.001, ****P<0.0001). Data are mean±s.e.m. Scale bars: 100 μm (white); 50 μm (gray).
	Fig. 3. Localized renal cysts in pkd1 crispants and phenotypic characterization of a pkd2 knockout line. (A) High-resolution mesoSPIM imaging was used to generate a topological skeleton via DiameterNet and a distance transformation was used as a measure of local tubular diameter. Scale bars: 50 μm. (B) Immunofluorescence microscopy showed pronounced tubular dilation in pkd2 knockout X. laevis. Scale bars: 300 μm (white); 100 μm (gray). (C) mesoSPIM light-sheet microscopy in toto imaging of pkd2+/+ and pkd2−/− embryos (dorsal view). Scale bars: 300 μm (white); 100 μm (gray). (D) Bright-field images of pkd2+/+, pkd2+/− and pkd2−/− animals. EmbryoNet masks were used for skeletonization. Scale bar: 1 mm. (E) Fully automated measurements of kidney tubule area and percentage of cystic area using 2D-NephroNet and 2D-CystNet (Kruskal–Wallis with Dunn's multiple comparison: ns, not significant, ****P<0.0001). Fully automated measurements of embryo length (Longest Shortest path skeletonized EmbryoNet, white, ***P<0.001) and a measure of embryo curliness (Euclidian distance between A and B, blue, ****P<0.0001). Data are mean±s.e.m.
	Fig. 4. Deep-learning analysis of neural phenotypes in a dyrk1a-depleted embryos. (A) TelenNet for automated segmentation of the telencephalon (forebrain) from whole-mount β-tubulin immunofluorescence stainings imaged by wide-field microscopy of dyrk1a CRISPR/Cas9-edited embryos. TelenNet independently segmented the telencephalon on either side of the midline. (B) Telencephalon area as determined by manual assessment of two independent expert human annotators and TelenNet. As TelenNet was trained by annotator 1, automated measurements were similar to those of annotator 1 (ANOVA, ns; Kruskal-Wallis, ns). (C) Cross-correlation between TelenNet and each independent annotator. (D) BrainNet achieved multiclass segmentation of various brain regions (telencephalon, diencephalon, mesencephalon and rhombencephalon) bilaterally. (E-G) Chained image processing pipeline for cell counting in the telencephalon. (E) TelenNet was fine-tuned to recognize telencephalons in PCNA-stained embryos and the resulting masks were used to isolate the left and right telencephalons from the pHH3-channel. (F) A cell-counting model (ProliNet) identified pHH3+ cells. (G) The number of proliferating (pHH3+) cells in the telencephalon of harmine-treated embryos (unpaired t-test: ****P<0.0001) and the percentage of telencephalon area covered by pHH3+ cells (unpaired t-test: ****P<0.0001) were significantly increased. (H) mesoSPIM light-sheet in toto imaging of a dyrk1a unilateral (right) morphant stained for β-tubulin and counterstained with DAPI. (I) A sparse annotation approach was used to segment the Xenopus brain using 3D-BrainNet. The telencephalon is pseudo-colored in blue. Scale bars: 500 μm (white); 100 μm (gray).
	Fig. 5. Volumetric analysis of craniofacial abnormalities induced by retinoic acid inhibition and in six1−/− X. tropicalis embryos. (A) Schematic of the Xenopus craniofacial cartilages. AlcianNet achieved multiclass segmentation of craniofacial elements from Alcian Blue stained X. tropicalis embryos. (B) mesoSPIM light-sheet in toto imaging of a wild-type embryo stained for Col2a1. CranioNet based 3D-segmentation of craniofacial cartilages (yellow). Autofluorescence (Autofluo, red) recorded at 488 nm excitation. (C) Three-dimensional quantitative phenotyping of X. tropicalis revealed a dose-dependent response to the retinoic acid inhibitor BMS-453. (D) Quantification of the normalized cartilage surface revealed a BMS-453 dose-response curve (ANOVA, P<0.0001; Holm–Šidák multiple comparison, ***P<0.001). (E) Morphological differences between six1 heterozygous and homozygous knockout embryos. The arrows indicate the collapsed dysmorphic Meckel's and ceratohyal cartilages in six1 knockouts. (F) Quantification revealed lower gross cartilage volume of six1 knockouts, a decreased distance between quadrates (red arrows) and an increase in the ceratohyal angle (dashed blue line) (unpaired t-test: *P<0.05, **P<0.01, ***P<0.001). Data are mean±s.d. Scale bars: 200 μm.
	Fig. 6. Deep learning is applicable to various imaging modalities. (A) EmbryoNet-ISH accurately segmented colorimetric whole-mount in situ hybridization (WISH) stained embryos. (B) Segmentation masks were used to extract, crop and register the in situ signal of stage 26 X. laevis embryos. Unsupervised clustering (n=63) identified distinct co-expression groups. (C) Example images confirmed similar expression patterns within co-expression groups; divergent expression patterns are visualized in multichannel images. (D) mesoSPIM recording of a stage 45 embryo stained for hnf1β using hybridization chain reaction FISH (HCR v3.0), LE-lectin and anti-Atp1a1. DAPI was used as a counterstain. Various morphological structures were segmented in 3D using U-Net models, revealing hnf1β to most strongly expressed in the proximal tubular segments PT2/3 (purple label) and pancreas. (E) Schematic and enlarged view of the pronephros segmentation. (F) mesoSPIM recording of the adult kidney of an induced Slc12a3/NCC-cre-ERT2Tg/+:TdTomato-floxTg/+ reporter mouse to visualize the distal convoluted tubules (DCT). DCT-Net segmented single DCTs and maintained separation of DCTs in close proximity. DCT-Net 3D-segmentations highlighted the spatial distribution of DCT in the renal cortex and permitted feature extraction such as volume measurements. Scale bars: 200 μm (white); 40 μm (gray).
	Fig. S1. TubuleNet segmentations closely copied expert human annotations. (A-C) TubuleNet and expert annotations for tubule length, pronephros width (bounding box area x) and pronephros height (box area y). Shown on the right is raw input data as used. Crop outs reflect the manual segmentation and TubuleNet segmentations. Deviations in pronephros height and width when directly comparing the measurements between manual segmentation and TubuleNet due to differential methodology of applying the bounding box are apparent. (D) Correlation of the median value of the pronephros area between TubuleNet and U-net segmentations across each setup (left and right pronephros separately) (n = 18) (pearson r = 0.99). (E) Correlation of the difference in means across the left/right pronephros across each setup (n = 9) (pearson r = 0.94). (F) For eight setups the Dunn's multiple comparison mean rank difference was calculated to an RFP injection control. This was performed for both TubuleNet measurements and expert annotations. Matched mean rank differences for identical comparisons were compared between TubuleNet and expert annotations revealing profound correlation (n = 9) (pearson r = 0.99). ( see next image for DEF)
	Fig. S1. TubuleNet segmentations closely copied expert human annotations. ( continued) D) Correlation of the median value of the pronephros area between TubuleNet and U-net segmentations across each setup (left and right pronephros separately) (n = 18) (pearson r = 0.99). (E) Correlation of the difference in means across the left/right pronephros across each setup (n = 9) (pearson r = 0.94). (F) For eight setups the Dunn's multiple comparison mean rank difference was calculated to an RFP injection control. This was performed for both TubuleNet measurements and expert annotations. Matched mean rank differences for identical comparisons were compared between TubuleNet and expert annotations revealing profound correlation (n = 9) (pearson r = 0.99).
	Fig. S2. 3D-NephroNet. (A) U-Net training logs. Blue lines (left to right) are locally estimated scatterplot smoothing (LOESS) of validation loss. (B) Comparison between Imaris intensity threshold segmentation (threshold level increased until full segmentation of the left kidney was achieved) and 3D-NephroNet. Threshold segmentation was performed on the single LEL-lectin channel, the single Atp1a1 channel and the LEL-lectin/ Atp1a1 composite. 3D-NephroNet was run on the LEL-lectin/Atp1a1 composite. A largest-two blob filter was applied on the 3D-NephroNet segmentation. This embryo is unseen test data.
	Fig. S3. Efficient genome editing of pkd1 results in cystogenesis in X. tropicalis embryos. (A) Genome editing efficiencies for three distinct gRNAs targeting pkd1 as quantified by Sanger sequencing and trace deconvolution approaches (3 pools of 5 embryos per gRNA, N = 15). (B) Development of cystic kidneys in X. tropicalis embryos after bilateral injection of pkd1 gRNAs in two-cell embryos. (C) Kidney areas were manually measured revealing only for pkd1 gRNA 2 a significant increase. (Kruskal-Wallis: p<0.05; Dunn's multiple comparisons: * p<0.05) (D) U-Net training logs for 3D-CystNet. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes
	Fig. S4. pkd1 crispants do not show gross morphological abnormalities. (A) mesoSPIM recordings in the DAPI channel reveal no gross differences in embryos between non-injected, tyr crispants and pkd1 gRNA crispants. Nevertheless, obvious kidney cystogenesis can be observed in the LE/Lectin-A5 channels. (B) A chained U-Net approach to investigate possible broad abnormalities in structures originating from the vegetal-ventral blastomere. First, an Embryonet (IOU: 0.87) is used to isolate embryos from low magnification bright-field stereomicroscopy. Isolated single embryos are then processed by OrganNet (IOU: 0.75) to provide area measurements of 4 anatomical regions: The head (pink), the somites (blue), the eye (green) and then intestines (red). (see next image for Fig. S4 CDE)
	Fig. S4. pkd1 crispants do not show gross morphological abnormalities. (continued ) (C-D) U-Net training logs for EmbryoNet and OrganNet. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes. (E) Fully automated measurements using a Fiji macro chaining EmbryoNet and OrganNet allows extraction of somite and intestine area of single embryos from low-magnification stereomicroscopy. This reveals no significant differences in sizes of two anatomical structures, originating from the CRISPR/Cas9-targeted ventral-vegetal lineage, when comparing non-injected, tyr injected and pkd1 injected embryos demonstrating absence of gross abnormalities in pkd1 crispants.
	Fig. S5. 2D-NephroNet, a deep learning solution for kidney segmentation in X. tropicalis embryos. ( panels A-C) (A-B) U-Net training logs: IOU and F1 scores across different training dataset sizes (5, 15, 25 and 35 per state). State is defined as either hypoplastic, normal or cystic X. tropicalis kidneys. Orange line is the LOESS of the three technical repeats, dashed lines are the LOESS of each technical repeat separately. (C-D) U-Net training logs: 2D-NephroNet models trained with a larger training dataset for more iterations and with a smaller dataset for less iterations. (E) Cross- correlation of model 1 (panel C) and model 2 (panel D) to two independent human experts using 45 unseen test images, stratified evenly across the three states.
	Fig. S5. 2D-NephroNet, a deep learning solution for kidney segmentation in X. tropicalis embryos. (continued: panel D.) (A-B) U-Net training logs: IOU and F1 scores across different training dataset sizes (5, 15, 25 and 35 per state). State is defined as either hypoplastic, normal or cystic X. tropicalis kidneys. Orange line is the LOESS of the three technical repeats, dashed lines are the LOESS of each technical repeat separately. (C-D) U-Net training logs: 2D-NephroNet models trained with a larger training dataset for more iterations and with a smaller dataset for less iterations. (E) Cross- correlation of model 1 (panel C) and model 2 (panel D) to two independent human experts using 45 unseen test images, stratified evenly across the three states.
	Fig. S7. High-resolution mesoSPIM (using MVPLAPO2XC objective) reveals localized tubular cysts interspersed with undilated epithelia in pkd1 crispants. (A) mesoSPIM imaging with a 2x objective clearly showcases local cystogenesis in pkd1 crispants. (B) U-Net training logs for 3D-NephroNet-PKD1. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (C) 3D-NephroNet-PKD1 for 3D reconstruction of kidneys from high-resolution mesoSPIM data (acquired with 2x objective) showcasing the spectrum of phenotypes occurring in pkd1 crispants.
	Fig. S8. Cystogenesis in pkd1 crispants is characterized by a phenotypic spectrum ranging from tubular cysts interspersed with undilated epithelia to fully cystic kidneys, reflecting the mosaic nature of CRISPR/ Cas9 genome editing. [Panel A only ] (A) U-Net training logs for DiameterNet. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes (B) DiameterNet was deployed on maximum projections of single kidneys imaged using high-resolution MesoSPIM. The resulting segmentation map was employed for medial axis skeletonization allowing heatmapping of the distance between each pixel within the DiameterNet mask to background pixels with the value zero (C) Significant increase in the average tubular dilation comparing the kidneys on the injected side of pkd1 crispants to kidneys on the non-injected side. Tubular dilation is an index calculated as the average of each non-zero pixel in the heatmap shown in C. Mann-Whitney test, **p<0.01. (D) Histogram plots demonstrating the increase in distance from pixels within DiameterNet to background, a measure for tubular dilation around that pixel, when comparing the injected side of pkd1 crispants to either the non-injected side or both sides of tyrosinase control crispants. see following images for Fig. S8 B-C
	Fig. S8. Cystogenesis in pkd1 crispants is characterized by a phenotypic spectrum ranging from tubular cysts interspersed with undilated epithelia to fully cystic kidneys, reflecting the mosaic nature of CRISPR/ Cas9 genome editing. [Panel B and C ] (A) U-Net training logs for DiameterNet. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes (B) DiameterNet was deployed on maximum projections of single kidneys imaged using high-resolution MesoSPIM. The resulting segmentation map was employed for medial axis skeletonization allowing heatmapping of the distance between each pixel within the DiameterNet mask to background pixels with the value zero (C) Significant increase in the average tubular dilation comparing the kidneys on the injected side of pkd1 crispants to kidneys on the non-injected side. Tubular dilation is an index calculated as the average of each non-zero pixel in the heatmap shown in C. Mann-Whitney test, **p<0.01. (D) Histogram plots demonstrating the increase in distance from pixels within DiameterNet to background, a measure for tubular dilation around that pixel, when comparing the injected side of pkd1 crispants to either the non-injected side or both sides of tyrosinase control crispants. see previous images for Fig. S8 A
	Fig. S9. Generalized edema can be triggered in pkd1 crispants at stage 45 by targeting both kidneys. (A) Targeting both the left and the right ventral blastomere allows bilateral gene editing using pkd1 gRNA2. (B-C) When both kidneys are targeted pronounced general edema occurs, showcasing a functional consequence of renal malfunction and fluid retention in early development.
	Fig. S10. pkd2 crispant model for autosomal dominant polycystic kidney disease (ADPKD). (A) Schematic of CRISPR/Cas9 injection in the vegetal-ventral blastomere of an 8-cell stage in X. laevis embryos. (B) Stereomicroscopy of LE-Lectin/A5 stained embryos reveals normal kidney development in non-injected and tyrosinase control embryos. Injections of two independent gRNAs targeting pkd2, existing as a non-duplicated gene on the X. laevis L. chromosome, leads to localized renal cyst formation. (C) mesoSPIM light-sheet microscopy revealing cystogenesis on the injected (left) side of pkd2 crispants, which is absent on the non-injected (right) side or in tyr crispant controls.
	Fig. S11. 2D-NephroNet and 2D-CystNet transfer learning. (A) Transfer learning finetuning adapting a pre- trained 2D-NephroNet towards towards pkd2 X. laevis embryos. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes (B) Transfer learning finetuning adapting a pre-trained 2D- CystNet towards towards pkd2 X. laevis embryos.
	Fig. S12. Analysis of pkd2 mutants. (A) Fully automated measurements reveal significant differences in embryo size (EmbryoNet total area) (Analysis shown in detail in main figure 3D. (B) EmbryoNet-PKD2 training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score.
	Fig. S13. U-Net deep learning for brain phenotyping. ( Panels A and B only) (A) TelenNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (B) BrainNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes. (C) TelenNet transfer learning to PCNA-stained embryos training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (D) ProliNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. See next images for Fig. S13 CDE
	Fig. S13. U-Net deep learning for brain phenotyping. ( Panels CDE only) (A) TelenNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (B) BrainNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes. (C) TelenNet transfer learning to PCNA-stained embryos training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (D) ProliNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. See previous images for Fig. S13 AB
	Fig. S14. U-Net deep learning for facial phenotyping.( Panel A and B) (A) AlcianNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes. (B) FaceNet training logs. Blue lines (left to right) are LOESS of validation loss and IOU across classes. (C) Embryos were subjected to four different concentrations of BMS-453 and a DMSO control and facial photographs were acquired. FaceNet was deployed on this unseen data and the masks were used to quantify orofacial area (blue), face height (blue - bounding box y), face width (blue - bounding box x), eyes (yellow), distance between eyes (bounding box edge yellow left eye to bounding box edge yellow right eye) and mouth area (red). See next image for Fig. S14. C
	Fig. S14. U-Net deep learning for facial phenotyping. (Panel C only) (A) AlcianNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU across classes and F1 score across classes. (B) FaceNet training logs. Blue lines (left to right) are LOESS of validation loss and IOU across classes. (C) Embryos were subjected to four different concentrations of BMS-453 and a DMSO control and facial photographs were acquired. FaceNet was deployed on this unseen data and the masks were used to quantify orofacial area (blue), face height (blue - bounding box y), face width (blue - bounding box x), eyes (yellow), distance between eyes (bounding box edge yellow left eye to bounding box edge yellow right eye) and mouth area (red). see previous image for Fig. S14.A-B
	Fig. S15. CranioNet. (A) CranioNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (B-C) Transfer learning finetuning adapting a pre-trained CranioNet towards BMS-453 treated embryos. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. In panel C double the training data was used as in panel B, using the same validation dataset. (D) Transfer learning finetuning adapting a pre-trained CranioNet towards six1 embryos. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score.
	Fig. S16. EmbryoNet-ISH. (A) EmbryoNet-ISH training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. (B) Cross-correlation of EmbryoNet-ISH to a human expert using unseen test images from different stages of X. tropicalis development.
	Fig. S17. Part1, U-Net training logs for models used in X. tropicalis reconstruction (Fig. 5D-E) and mouse whole kidney imaging (Fig. 5F). [Panels ABC only] (A-E) VoluNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. See next image for Fig S17 D and E
	Fig. S17. Part1, U-Net training logs for models used in X. tropicalis reconstruction (Fig. 5D-E) and mouse whole kidney imaging (Fig. 5F). [Panels D-E only] (A-E) VoluNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. See previous image for Fig S17 panels A B and C
	Fig. S18. Part2, U-Net training logs for models used in X. tropicalis reconstruction (Fig. 5D-E. (A-E) VoluNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score.
	Fig. S19. (A) EmbryoNets were trained on two images shown. [Panels A B and C]. EmbryoNet 1 – trained on two images containing embryos from clutch 1, EmbryoNet 2 – trained on two images containing embryos from clutch 1, EmbryoNet 3 – trained one 1 image containing embryos from clutch 1 and one image containing embryos from clutch 2 (B-C) All three EmbryoNets were validated on an unseen image containing embryos from either clutch 1, embryos from clutch 2 and embryos from clutch 1 and clutch 2 intermixed 1:1. Shown is the correlation matrix (B) and the results of each EmbryoNet on validation data (C). (D) EmbryoNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. see next image or Fig. S19 Panel D
	Fig. S19. (A) EmbryoNets were trained on two images shown. [PAnel D only]EmbryoNet 1 – trained on two images containing embryos from clutch 1, EmbryoNet 2 – trained on two images containing embryos from clutch 1, EmbryoNet 3 – trained one 1 image containing embryos from clutch 1 and one image containing embryos from clutch 2 (B-C) All three EmbryoNets were validated on an unseen image containing embryos from either clutch 1, embryos from clutch 2 and embryos from clutch 1 and clutch 2 intermixed 1:1. Shown is the correlation matrix (B) and the results of each EmbryoNet on validation data (C). (D) EmbryoNet training logs. Blue lines (left to right) are LOESS of validation loss, IOU and F1 score. see previous image fro Fig. S19 panels ABC

Akerberg, Deep learning enables automated volumetric assessments of cardiac function in zebrafish. 2019, Pubmed

Akerberg, Deep learning enables automated volumetric assessments of cardiac function in zebrafish. 2019, Pubmed
Baert, Hereditary polycystic kidney disease (adult form): a microdissection study of two cases at an early stage of the disease. 1978, Pubmed
Bergmann, Polycystic kidney disease. 2018, Pubmed
Boyle, Estimating Global Burden of Disease due to congenital anomaly: an analysis of European data. 2018, Pubmed
Brinkman, Easy quantitative assessment of genome editing by sequence trace decomposition. 2014, Pubmed
Buhmann, Automatic detection of synaptic partners in a whole-brain Drosophila electron microscopy data set. 2021, Pubmed
Campanella, Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. 2019, Pubmed
Choi, Mapping a multiplexed zoo of mRNA expression. 2016, Pubmed
Choi, Third-generation in situ hybridization chain reaction: multiplexed, quantitative, sensitive, versatile, robust. 2018, Pubmed
Chung, Structural and molecular interrogation of intact biological systems. 2013, Pubmed
Coppenrath, Generation of a new six1-null line in Xenopus tropicalis for study of development and congenital disease. 2021, Pubmed , Xenbase
Dang, Autism-associated Dyrk1a truncation mutants impair neuronal dendritic and spine growth and interfere with postnatal cortical development. 2018, Pubmed
Deniz, Quantitative Phenotyping of Xenopus Embryonic Heart Pathophysiology Using Hemoglobin Contrast Subtraction Angiography to Screen Human Cardiomyopathies. 2019, Pubmed , Xenbase
Dubey, Modeling human craniofacial disorders in Xenopus. 2017, Pubmed , Xenbase
Emmert-Streib, An Introductory Review of Deep Learning for Prediction Models With Big Data. 2020, Pubmed
Exner, Xenopus leads the way: Frogs as a pioneering model to understand the human brain. 2021, Pubmed , Xenbase
Falk, U-Net: deep learning for cell counting, detection, and morphometry. 2019, Pubmed
Gilbert, Bilineal inheritance of PKD1 abnormalities mimicking autosomal recessive polycystic disease. 2013, Pubmed
Goda, Genetic screens for mutations affecting development of Xenopus tropicalis. 2006, Pubmed , Xenbase
Graving, DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning. 2019, Pubmed
Gómez-de-Mariscal, DeepImageJ: A user-friendly environment to run deep learning models in ImageJ. 2021, Pubmed
Göckler, Harmine specifically inhibits protein kinase DYRK1A and interferes with neurite formation. 2009, Pubmed
Günel, DeepFly3D, a deep learning-based approach for 3D limb and appendage tracking in tethered, adult Drosophila. 2019, Pubmed
Hakim, WorMachine: machine learning-based phenotypic analysis tool for worms. 2018, Pubmed
Hellsten, The genome of the Western clawed frog Xenopus tropicalis. 2010, Pubmed , Xenbase
Hemmati-Brivanlou, Localization of specific mRNAs in Xenopus embryos by whole-mount in situ hybridization. 1990, Pubmed , Xenbase
Hoff, ANKS6 is a central component of a nephronophthisis module linking NEK8 to INVS and NPHP3. 2013, Pubmed , Xenbase
Jinek, A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity. 2012, Pubmed
Kaminski, Direct reprogramming of fibroblasts into renal tubular epithelial cells by defined transcription factors. 2016, Pubmed , Xenbase
Kennedy, Median facial clefts in Xenopus laevis: roles of retinoic acid signaling and homeobox genes. 2012, Pubmed , Xenbase
Kennedy, Quantification of orofacial phenotypes in Xenopus. 2014, Pubmed , Xenbase
Kirst, Mapping the Fine-Scale Organization and Plasticity of the Brain Vasculature. 2020, Pubmed
Kroll, A simple and effective F0 knockout method for rapid screening of behaviour and other complex phenotypes. 2021, Pubmed
Kälin, An in vivo chemical library screen in Xenopus tadpoles reveals novel pathways involved in angiogenesis and lymphangiogenesis. 2009, Pubmed , Xenbase
LeCun, Deep learning. 2015, Pubmed
Lienkamp, Vertebrate kidney tubules elongate using a planar cell polarity-dependent, rosette-based mechanism of convergent extension. 2012, Pubmed , Xenbase
Liu, Emerging Imaging and Genomic Tools for Developmental Systems Biology. 2016, Pubmed
Loffing, Altered renal distal tubule structure and renal Na(+) and Ca(2+) handling in a mouse model for Gitelman's syndrome. 2004, Pubmed
Lu, Perinatal lethality with kidney and pancreas defects in mice with a targetted Pkd1 mutation. 1997, Pubmed
Mangos, The ADPKD genes pkd1a/b and pkd2 regulate extracellular matrix formation. 2010, Pubmed
McNamara, Husbandry, General Care, and Transportation of Xenopus laevis and Xenopus tropicalis. 2018, Pubmed , Xenbase
Menezes, Murine Models of Polycystic Kidney Disease. 2013, Pubmed
Metzner, A high throughput zebrafish chemical screen reveals ALK5 and non-canonical androgen signalling as modulators of the pkd2-/- phenotype. 2020, Pubmed
Moreno-Mateos, CRISPRscan: designing highly efficient sgRNAs for CRISPR-Cas9 targeting in vivo. 2015, Pubmed , Xenbase
Naert, Methods for CRISPR/Cas9 Xenopus tropicalis Tissue-Specific Multiplex Genome Engineering. 2018, Pubmed , Xenbase
Naert, Genotyping of CRISPR/Cas9 Genome Edited Xenopus tropicalis. 2018, Pubmed , Xenbase
Naert, CRISPR/Cas9 disease models in zebrafish and Xenopus: The genetic renaissance of fish and frogs. 2018, Pubmed , Xenbase
Naert, CRISPR/Cas9 mediated knockout of rb1 and rbl1 leads to rapid and penetrant retinoblastoma development in Xenopus tropicalis. 2016, Pubmed , Xenbase
Naert, RBL1 (p107) functions as tumor suppressor in glioblastoma and small-cell pancreatic neuroendocrine carcinoma in Xenopus tropicalis. 2020, Pubmed , Xenbase
Naert, Maximizing CRISPR/Cas9 phenotype penetrance applying predictive modeling of editing outcomes in Xenopus and zebrafish embryos. 2020, Pubmed , Xenbase
Nakayama, Simple and efficient CRISPR/Cas9-mediated targeted mutagenesis in Xenopus tropicalis. 2013, Pubmed , Xenbase
Nasr, Endosome-Mediated Epithelial Remodeling Downstream of Hedgehog-Gli Is Required for Tracheoesophageal Separation. 2019, Pubmed , Xenbase
Nauli, Polycystins 1 and 2 mediate mechanosensation in the primary cilium of kidney cells. 2003, Pubmed
Pan, Deep Learning Reveals Cancer Metastasis and Therapeutic Antibody Targeting in the Entire Body. 2019, Pubmed
Reynaud, Guide to light-sheet microscopy for adventurous biologists. 2015, Pubmed
Royer, ClearVolume: open-source live 3D visualization for light-sheet microscopy. 2015, Pubmed
Ruf, A gene locus for branchio-otic syndrome maps to chromosome 14q21.3-q24.3. 2003, Pubmed
Saberi-Bosari, Deep learning-enabled analysis reveals distinct neuronal phenotypes induced by aging and cold-shock. 2020, Pubmed
Sanggaard, Branchio-oto-renal syndrome: detection of EYA1 and SIX1 mutations in five out of six Danish families by combining linkage, MLPA and sequencing analyses. 2007, Pubmed
Satterstrom, Large-Scale Exome Sequencing Study Implicates Both Developmental and Functional Changes in the Neurobiology of Autism. 2020, Pubmed
Schindelin, Fiji: an open-source platform for biological-image analysis. 2012, Pubmed
Schnoz, Deletion of the transcription factor Prox-1 specifically in the renal distal convoluted tubule causes hypomagnesemia via reduced expression of TRPM6 and NCC. 2021, Pubmed
Schottenfeld, Zebrafish curly up encodes a Pkd2 ortholog that restricts left-side-specific expression of southpaw. 2007, Pubmed
Shaidani, Animal Maintenance Systems: Xenopus tropicalis. 2020, Pubmed , Xenbase
Shaidani, Obtaining Xenopus laevis Embryos. 2021, Pubmed , Xenbase
Shaidani, Obtaining Xenopus laevis Eggs. 2021, Pubmed , Xenbase
Stringer, Cellpose: a generalist algorithm for cellular segmentation. 2021, Pubmed
Szenker-Ravi, RSPO2 inhibition of RNF43 and ZNRF3 governs limb development independently of LGR4/5/6. 2018, Pubmed , Xenbase
Tavares, Sobp modulates the transcriptional activation of Six1 target genes and is required during craniofacial development. 2021, Pubmed , Xenbase
Todorov, Machine learning analysis of whole mouse brain vasculature. 2020, Pubmed
Tomer, Advanced CLARITY for rapid and high-resolution imaging of intact tissues. 2014, Pubmed
Tomlinson, Chemical genomics identifies compounds affecting Xenopus laevis pigment cell development. 2009, Pubmed , Xenbase
Villoutreix, What machine learning can do for developmental biology. 2021, Pubmed
Vivante, Mutations in TBX18 Cause Dominant Urinary Tract Malformations via Transcriptional Dysregulation of Ureter Development. 2015, Pubmed
Vize, Multichannel wholemount fluorescent and fluorescent/chromogenic in situ hybridization in Xenopus embryos. 2009, Pubmed , Xenbase
Voigt, The mesoSPIM initiative: open-source light-sheet microscopes for imaging cleared tissue. 2019, Pubmed
Wallingford, We Are All Developmental Biologists. 2019, Pubmed
Wang, Targeted gene disruption in Xenopus laevis using CRISPR/Cas9. 2015, Pubmed , Xenbase
Wen, 3DeeCellTracker, a deep learning-based pipeline for segmenting and tracking cells in 3D time lapse images. 2021, Pubmed
Willsey, Katanin-like protein Katnal2 is required for ciliogenesis and brain development in Xenopus embryos. 2018, Pubmed , Xenbase
Willsey, The neurodevelopmental disorder risk gene DYRK1A is required for ciliogenesis and control of brain size in Xenopus embryos. 2020, Pubmed , Xenbase
Willsey, Parallel in vivo analysis of large-effect autism genes implicates cortical neurogenesis and estrogen in risk and resilience. 2021, Pubmed , Xenbase
Yang, Single-cell phenotyping within transparent intact tissue through whole-body clearing. 2014, Pubmed
Zhang, Automatic Segmentation and Cardiac Mechanics Analysis of Evolving Zebrafish Using Deep Learning. 2021, Pubmed
Zhao, Cellular and Molecular Probing of Intact Human Organs. 2020, Pubmed
van der Walt, scikit-image: image processing in Python. 2014, Pubmed
von Chamier, Democratising deep learning for microscopy with ZeroCostDL4Mic. 2021, Pubmed