key: cord-102279-ena1usqv
authors: Long, Rory K. M.; Moriarty, Kathleen P.; Cardoen, Ben; Gao, Guang; Vogl, A. Wayne; Jean, François; Hamarneh, Ghassan; Nabi, Ivan R.
title: Super Resolution Microscopy and Deep Learning Identify Zika Virus Reorganization of the Endoplasmic Reticulum
date: 2020-06-23
journal: bioRxiv
DOI: 10.1101/2020.05.12.091611
sha: 
doc_id: 102279
cord_uid: ena1usqv

The endoplasmic reticulum (ER) is a complex subcellular organelle composed of diverse structures such as tubules, sheets and tubular matrices. Flaviviruses such as Zika virus (ZIKV) induce reorganization of endoplasmic reticulum (ER) membranes to facilitate viral replication. Here, using 3D super resolution microscopy, ZIKV infection is shown to induce the formation of dense tubular matrices associated with viral replication in the central ER. Viral non-structural proteins NS4B and NS2B associate with replication complexes within the ZIKV-induced tubular matrix and exhibit distinct ER distributions outside this central ER region. Deep neural networks trained to identify ZIKV-infected versus mock-infected cells successfully identified ZIKV-induced central ER tubular matrices as a determinant of viral infection. Super resolution microscopy and deep learning are therefore able to identify and localize morphological features of the ER and may be of use to screen for inhibitors of infection by ER-reorganizing viruses.

The endoplasmic reticulum (ER) is a highly dynamic network composed of 30-100 nm ribosome-28 studded rough ER sheets and convoluted networks of smooth ER tubules (1, 2) . ER shaping proteins 29 such as the lumenal sheet spacer protein cytoskeleton-linking membrane protein 63 (CLIMP-63), 80 In order to study ER morphology following ZIKV infection, we first generated stable U87 glioblas- for 48 hours. Cells were fixed with 3% paraformaldehyde/0.2% glutaraldehyde to preserve ER archi-87 tecture (6, 7, 30, 31) and labeled for dsRNA, a marker for ZIKV replication factories (19). Maximum 88 Figure 1 . 3D STED microscopy reveals ZIKV-induced ER reorganization in human U87 glioblastoma cells. A) ERmoxGFP or Sec61β-GFP stably transfected U87 cells were mock-infected or infected for 48 hours with the PRVABC59 2015 ZIKV strain (MOI= 1). ER reporter GFP and immunostained dsRNA-labeled ZIKV replication factories were imaged by 3D STED microscopy. B) Fluorescence intensity of ERmoxGFP of infected cells using a spectrum heat map and a segmentation mask of the ER that colocalizes with dsRNA (grey), both generated on Imaris x64 9.2.1 (Imaris), are depicted. Yellow squares in the panels indicate the magnified ROIs shown in the adjacent panels. Quantification of the mean normalized ER density ((Intensity sum of mask/total cell intensity sum)/ (volume sum of mask/ total cell volume sum)) was performed for both dsRNA-positive and dsRNA-negative ERmoxGFP and Sec61β-GFP in ZIKV-infected cells by Imaris segmentation. Scale bar= 10 microns. 5 cells per biological replicate (N=3). Statistics were done using unpaired Student's T tests: **= P<0.01, Error bars represent SEM. were imaged by 3D STED microscopy. Magnified ROIs (yellow ROIs identified by red Roman numerals) show that the PER extends over 3-5 sections (210 nm step size) and CER >10 sections. Graph shows average PER and CER Z-height for each ERmoxGFP or Sec61β-GFP labeled cell. A Z-height cutoff of 1.26 microns (red line) was used to identify PER and CER objects. B) Segmented ER labeling from 48-hour ZIKV-or mock-infected ERmoxGFP or Sec61β-GFP stably transfected U87 cells (MOI= 1) was visualized using a Z-height spectrum heat map and CER (green; > 1.26 µm) and PER (red; < 1.26 µm) masks are shown. C) Volume percentage (left) and mean normalized density (right) of CER and PER masks between mock-and ZIKV-infected cells. 5 cells per biological replicate were analyzed for a total of N=3. ANOVA with post-hoc Tukey HSD: *= P<0.05, **= P<0.01, and ***= P<0.001. Error bars represent SEM. Scale bar= 10 microns. (green) and PER (red) masks overlaid with the dsRNA-positive ER mask (white) for ZIKV-infected ERmoxGFP and Sec61β-GFP transfected U87 cells. Enlarged images of ERmoxGFP transfected cells show 3x3 µm ROIs of the mock-infected CER (green box) and of the ZIKV-infected dsRNA-positive (red box) and dsRNA-negative (yellow box) CER shown in B. Graphs show the volume percent of the CER or PER region that contains dsRNA-positive ER (left) and the volume percent of the dsRNA-positive ER that resides within the CER mask. B) 2D images of ER (white) and dsRNA (red) labeling in 3x3 µm ROIs of the ZIKV-infected dsRNA-positive (red) and dsRNA-negative (yellow) CER and mock-infected (green) CER are shown above Imaris 3D surface rendering of 1x1 µm regions of the above ROIs. Graph shows mean normalized ER density for each of the three CER zones by Imaris segmentation and masking. 5 cells per biological replicate were analyzed for a total of N=3. ANOVA with post-hoc Tukey HSD: *= P<0.05, **= P<0.01, and ***= P<0.001. Error bars represent SEM. Scale bar: 10 microns (500 nm for zoomed ROIs). projections of 3D STED image stacks show high intensity ERmoxGFP and Sec61β-GFP labeling in a 89 CER region and low intensity labeling in PER tubules in mock-infected cells ( Figure 1A ), as reported 90 previously by diffraction limited confocal microscopy (3). Upon ZIKV infection, the CER reorganizes 91 to form an intensely labeled crescent-shaped region surrounding a lower intensity perinuclear re-92 gion ( Figure 1A) . Interestingly, the crescent-shaped ZIKV-induced perinuclear ER overlapped exten-93 sively with dsRNA ( Figure 1A ). Imaris Bitplane software fragments the ER into distinct segments 94 that can then be analyzed for different features, including reporter density, segment Z-height and 95 segment overlap with other labels, such as dsRNA. Density-based segmentation of the ERmoxGFP-96 and Sec61β-GFP-labelled ER of ZIKV-infected cells showed that the higher density crescent-shaped 97 CER region exhibited significant overlap with dsRNA-positive ER structures relative to the rest of 98 the ER ( Figure 1B ). This suggests that ZIKV dsRNA associates with an ER region of high density for 99 both lumenal and membrane ER reporters. 100 Figure 4 . Ultrastructural analysis of ZIKV-infected cerebral brain organoids. Transmission EM images of 50 nm thin sections of 48-hour mock-and ZIKV-infected cerebral brain organoids (MOI=1). Yellow boxes show ROIs shown of adjacent higher magnification images that highlight rough ER sheets in mock-infected and tubular matrices (convoluted membranes) in ZIKV-infected cells. Scale bars: 500 nm and 100 nm for zoomed image ROIs.

We then investigated the relation- Figure 2D ).

Overlaying the CER and PER masks with the dsRNA-positive ER mask showed that the dsRNA-130 positive ER (>80% volume/volume) is predominantly included within the CER mask ( Figure 3A ). In-131 deed, only 10% of PER volume contains dsRNA while 35% of CER volume contains dsRNA for both 132 ER reporters ( Figure 3A ). Morphological comparison of the dsRNA-positive and -negative CER of 133 ZIKV-infected cells with the CER of mock-infected cells showed that the CER was composed of a 134 convoluted network of tubules for both the ERmoxGFP-and Sec61β-labeled ER ( Figure 3B ). 3D re-135 constructions confirmed that these regions were predominantly tubular with a few small sheet-like 136 structures, similar to the tubular matrix morphology of peripheral sheets (7). 3D voxel-based visu-137 alization and quantification showed that the density of ER tubular structures in the dsRNA-positive 138 ER is higher, for both the ERmoxGFP or Sec61β-GFP ER reporters, than in the dsRNA-negative CER 139 regions of ZIKV-infected cells or the CER of mock-infected cells ( Figure 3B ). The lower ER reporter 140 density reflects reduced spacing between tubules in the dsRNA-positive ER, suggesting that ZIKV 141 infection induces tubular matrix reorganization in a subdomain of the CER in U87 cells. Consis-142 tently, EM analysis of the microcephaly relevant cerebral brain organoid model (32) showed that 143 ZIKV-induced ER reorganization from perinuclear stacked rough ER sheets to a perinuclear, circular 144 region of convoluted smooth ER tubules (Figure 4 ). These results are consistent with ZIKV induction 145 of a perinuclear tubular matrix.

ER localization of ZIKV NS2B and NS4B structural proteins 147 We then labeled cells for ZIKV NS proteins NS2B and NS4B to assess their relationship to the ZIKV-148 induced tubular matrix. 3D STED analysis showed a predominant distribution of both NS2B and 149 NS4B to the CER and more particularly to the dense ZIKV-induced crescent-shaped tubular matrix 150 in ERmoxGFP transfected U87 cells ( Figure 5A ). While NS2B is predominantly associated with the 151 dsRNA-positive CER, NS4B labeling extended throughout the CER as well as to the PER ( Figure 5A ).

To quantify this, we identified NS2B-positive and NS4B-positive ER segments and determined their 153 overlap with total ER, CER and PER ( Figure 5A ). While NS4B was present at high levels on both PER 154 and CER segments, NS2B was enriched in the CER relative to the PER and presented a similar ER 155 distribution to dsRNA ( Figure 5A ).

The majority (>55%) of dsRNA-labeled puncta were associated with NS2B or NS4B, consistent 157 with the presence of both these NS proteins in the ZIKV-induced tubular matrix. In contrast, a 158 minority of NS2B (~25%) and NS4B (~10%) spots overlapped with dsRNA spots ( Figure 5B ). In the 159 dsRNA-positive CER, the highly punctate NS2B labeling differed from a more reticular NS4B label-160 ing. These two ZIKV NS proteins therefore exhibit distinct distributions within the ZIKV-induced 161 tubular matrix when not associated with dsRNA replication complexes ( Figure 5B ). Together with 162 the differential distribution of NS2B and NS4B within the ER as a whole ( Figure 5A ), these results 163 highlight that these two ZIKV NS proteins do not associate exclusively with replication factories and 164 suggest that following synthesis of the ZIKV polyprotein, NS2B and NS4B undergo distinct biosyn-165 thetic pathways before reuniting in ER-associated replication complexes. are required by non-deep learning methods; and 2) provide the ability to move beyond simple 178 classification to inspect discriminative regions (i.e. subregions of the ER within each cell). 179 We therefore applied deep neural networks to identify and distinguish the morphological fea-180 tures of the ER of ZIKV-infected cells. A pipeline outlining our approach is shown in Figure 6A . We 181 train a CNN using 2D frames (each representing a single Z-frame) from 3D STED volumes of ER- Figure 5 . ER distribution of ZIKV NS2B and NS4B proteins. A) Mock-or ZIKV-infected ERmoxGFP (red) stably transfected U87 cells were immunostained for ZIKV NS2B or NS4B (green) and dsRNA (white). Merged images show NS2B or NS4B (green) overlaid with the CER, PER or dsRNA-positive ER (white). Graphs show quantification of volume percent of NS4B, NS2B and dsRNA ER regions relative to total ER, CER or PER. B) Mock-or ZIKV-infected ERmoxGFP stably transfected U87 cells were immunostained for ZIKV NS4B or NS2B (green) and dsRNA (red). White squares show ROIs of adjacent zoomed images in which white arrowheads show colocalization between NS protein and dsRNA puncta. Graphs show percent of dsRNA puncta overlapping NS4B or NS2B puncta (left) and percent of NS4B or NS2B puncta overlapping dsRNA puncta. 5 cells per biological replicate (N=3) with each dot representing a cell. ANOVA with post-hoc Tukey HSD: *=P<0.05, **=P<0.01, and ***=P<0.001. Error bars represent SEM. All images are maximum projections from 3D STED stacks. Scale bar= 10 microns. ROI scale bar = 2 microns. rate of networks when dealing with small target datasets. Certain filters (combinations of weights) 189 learned on the first dataset (i.e. ImageNet) may still be useful for classifying a second dataset 190 (i.e. STED); as a result, less weight updates are needed before achieving good performance. Us-191 ing a pretrained VGG16 as our base model we obtained a 20% boost in test accuracy, compared 192 against a random weight initialization. Using ERmoxGFP labelled ER alone, the CNN was able to 193 distinguish between ZIKV-and mock-infected cells with an 82% accuracy ( Figure 6B, Figure 6 . Deep learning classification pipeline: Pretrained convolutional neural network accurately predicts labels of 2D frames from 3D STED volumes. A) Leave-one-out cross validation is successively applied to each cell. This prevents information from 2D frames leaking between training, validation and test sets. During training, network uses 2D frames from 55 cells (specifically, 44 for training and 11 for validation). The trained CNN then predicts a class label (i.e. ZIKV-infected or mock-infected) to each 2D frame of the remaining test cell. Class Activation Maps are also generated for each 2D frame belonging to the test set. B) CNN performance reported on a cell basis and across 2D frames. Normalized confusion matrices report the total number of predicted labels (ZIKV-infected or mock-infected) over the total number of ground truth labels. For example, 79% of all mock-infected 2D frames were predicted correctly by the CNN (top right). Predicted cell labels correspond to the majority label of predicted frame labels for each cell (top left). When excluding frames beyond the cell with reduced ERmoxGFP signal, performance metrics increase both in terms of cell label predictions (bottom left) and individual frame label predictions (bottom right). the remaining cell. This process is outlined in Figure 6A , and is repeatedly applied using each cell show poor accuracy to predict class label (Supp Fig 1) . When considering those frames containing 212 ERmoxGFP signal intensity greater than the median, we achieved 84% accuracy and 86% sensitiv-213 ity. On a per frame basis, accuracy for all frames was 78% and sensitivity 79% that increased to 214 81% and 84%, respectively, when considering frames expressing ERmoxGFP greater than median This suggests that the CAMs used to identify both ZIKV-and mock-infected cells correspond to high 249 ER density regions localized over the cell (see Figure 7B ) and that VGG16 is using differences in the 250 ER label (ERmoxGFP) to identify slices as either ZIKV-or mock-infected. Figure 8B, left) . Conversely, regions identified by the thresholded mock CAM have increased ER 257 density for TN compared to TP cells ( Figure 8B, right) . We then calculated the average Euclidean Figure 7C ) and the precise nature of the features that the CNN uses to 290 discriminate between ZIKV-and mock-infected cells remains to be determined. CAM localization 291 analysis shows that the neural network uses the same CER region that we have observed to be 292 modified upon ZIKV infection. Deep learning therefore has the ability to identify the ER morpho-293 logical changes associated with ZIKV infection.

ZIKV infection is characterized by re-organization of the ER to create replication factories and con-296 voluted ER membranes involved in viral replication, whose ultrastructure has been elegantly char- given ROI, ER density is defined as total ERmoxGFP intensity within the ROI divided by ERMoxGFP area inside the ROI. ER density for each ROI defined by the CAM is then normalized by the ER density of the whole cell. A) ER density of ROIs defined by CAM thresholds from 10-95% with increments of 5% is compared across 4 subgroups: ZIKV-infected 2D frames correctly predicted to be infected (true positives); mock-infected 2D frames correctly predicted to be uninfected (true negatives); ZIKV-infected 2D frames incorrectly predicted to be uninfected (false negatives); mock-infected 2D frames incorrectly predicted to be infected (false positives). B) ER densities of 80% CAMs ROIs compared across subgroups. C) Euclidean distances between center of mass of 80% CAMs ROIs and weighted center of mass of ERmoxGFP signal.

(15, 42).

The ER is a morphologically complex organelle, containing smooth ER tubules and ribosome-309 studded rough ER sheets identified ultrastructurally by EM since over 60 years (43 networks that correspond to previously described PER tubular matrices (7). While we were unable 315 to detect ER sheets by super-resolution analysis of cultured U87 cells, EM of brain organoids shows 316 the transformation of ER sheets to convoluted membranes upon ZIKV infection. This suggests that 317 organoid structures present more highly developed ER structures than cultured cells; application 318 of 3D live cell super-resolution analysis (44) to this model of the developing fetal brain, composed 319 of a heterogenous population of cell types, may lead to better definition of complex ER structures 320 and their dynamic transitions in response to stress, such as viral infection. Nevertheless, the fixed 321 cell 3D STED analysis applied here demonstrates that convoluted membranes associated with ZIKV 322 replication derive from tubular matrix reorganization in the CER.

The ZIKV-induced CER-localized, high ER density tubular matrices are enriched for dsRNA. As formance for 3D super-resolution microscopy data sets that will be of service to other researchers 360 applying deep learning approaches to super-resolution microscopy.

The interpretability of artificial intelligence is an evolving field and we believe that interpretable 362 methods, such as Grad-CAM (53), are important tools for the understanding of deep neural net-363 works applied to exploratory data sets. This approach has now allowed us to identify features of 364 discriminatory regions, and has not, to our knowledge, been applied to subcellular morphology, Figure 10 . Network Performance Analysis A) Performance metrics are reported across predictions of frame labels and cell labels, where cell labels correspond to the majority label of predicted frame labels for each cell.

Results are reported using a given selection criteria: using all frames (rows 1, 4), using only frames with normalized ERmoxGFP signal greater than mean normalized ERmoxGFP signal (rows 2, 5) or greater than median normalized ERmoxGFP signal (rows 3, 6). Mean and median thresholds are computed on a cell basis. The rug plot (above x-axis) visualises distribution of z-frames. B) Accuracy reported across corrected z-frames, z=0 is where normalized ERmoxGFP Intensity Sum is maximal. C) Normalized ERmoxGFP Intensity sum plotted against corrected z-frame. Dashed lines indicate the median (orange) and mean (red) Normalized ERmoxGFP Intensity sum computed across all frames.

Microtubules and the endoplasmic reticulum are highly inter-516 dependent structures

Rough sheets and smooth tubules

Mechanisms determining 519 the morphology of the peripheral ER

A class of membrane proteins shaping the 521 tubular endoplasmic reticulum

Dynamic 523 nanoscale morphology of the ER surveyed by STED microscopy

Reticulon and CLIMP-63 control nanodomain organization of pe-525 ripheral ER tubules

Increased spatiotemporal reso-527 lution reveals highly dynamic dense tubular matrices in the peripheral ER

Architecture and biogenesis of plus-strand RNA virus replication fac-530 tories

RIG-I-like receptors: cytoplasmic sensors for non-self RNA

Molecular mechanism of signal 534 perception and integration by the innate immune sensor retinoic acid-inducible gene-I (RIG-I)

Hostile Takeover: Hijacking of Endoplasmic Reticulum Function 537 by T4SS and T3SS Effectors Creates a Niche for Intracellular Pathogens

Endoplasmic Reticulum: The Favorite Intracellular Niche for 539

Pathogen-endoplasmic-reticulum interactions: in through the 541 out door

Immunolocalization of the dengue virus nonstructural 543 glycoprotein NS1 suggests a role in viral RNA replication

Human Coronavirus: Host-Pathogen Interaction

Zika Virus Associated 547 with Meningoencephalitis

18. Fauci AS, Morens DM. Zika Virus in the Americas-Yet Another Arbovirus Threat

Ultrastructural 553 Characterization of Zika Virus Replication Factories

Cytoarchitecture of Zika virus infection in 555 human neuroblastoma and Aedes albopictus cell lines

Virus in Human Fetal Neural Progenitors Persists Long Term with Partial Cytopathic and Limited 558

Molecular aspects of Dengue virus replication

Genome Sequence of a Zika Virus Strain Isolated from the Serum of an Infected Patient in Thailand 563 in 2006

Rewiring cellular networks by members 565 of the Flaviviridae family

The Host Protein Reticulon

Is Utilized by Flaviviruses to Facilitate Membrane Remodelling

Virus Perturbs Mitochondrial Morphodynamics to Dampen Innate Immune Responses. Cell Host 570 Microbe

Composition and three-572 dimensional architecture of the dengue virus replication and assembly sites

Zika NS2B is a crucial factor recruiting NS3 to the 575 ER and activating its protease activity

A palette of 577 fluorescent proteins optimized for diverse cellular environments

Ligand-induced redistribution of a human KDEL receptor from the Golgi 580 complex to the endoplasmic reticulum

A conserved ER targeting motif in three families of lipid binding 582 proteins and in Opi1p binds VAP

Using brain organoids to understand Zika virus-584 induced microcephaly

Imagenet classification with deep convolutional neural 588 networks

Very deep convolutional networks for large-scale image recognition. 590 arXiv preprint arXiv

Rethinking Model Scaling for Convolutional Neural Networks

Skin 594 lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on 595 biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic)

A dataset for breast cancer histopathological 598 image classification

Learning Deep Features for Discriminative 600

Computer Vision and Pattern Recognition

Evaluating CNN-based semantic food segmentation 602 across illuminants

A novel 604 segmentation framework for uveal melanoma in magnetic resonance imaging based on class acti-605 vation maps

Coronavirus infection, ER stress, apoptosis and innate immunity. Front Micro-607 biol

Studies on the endoplasmic reticulum. I. Its identification in cells in situ

Applying systems-level 611 spectral imaging and analysis to reveal the organelle interactome

Atlastin En-613 doplasmic Reticulum-Shaping Proteins Facilitate Zika Virus Replication

ER-shaping atlastin 616 proteins act as central hubs to promote flavivirus replication and virion assembly

Crystal structure of unlinked NS2B-NS3 619 protease from Zika virus

Flaviviral NS4b, chameleon and jack-in-the-box roles in viral 621 replication and pathogenesis, and a molecular target for antiviral intervention

Zika Virus NS4A and NS4B Proteins Dereg-624

ulate Akt-mTOR Signaling in Human Fetal Neural Stem Cells to Inhibit Neurogenesis and Induce

Knowledge transfer 627 for melanoma screening with deep learning

Texture analysis for muscu-630 lar dystrophy classification in MRI with improved class activation mapping. Pattern recognition

Breast cancer histology images classification: Training from scratch or transfer learn-633 ing? ICT Express

Grad-CAM: Visual Explanations 635 from Deep Networks via Gradient-based Localization

Form follows function: the importance of endoplas-637 mic reticulum shape

Defining host-pathogen 639 interactions employing an artificial intelligence workflow. eLife

SARS-641 coronavirus replication is supported by a reticulovesicular network of modified endoplasmic retic-642 ulum

Ca2+ signaling machinery is present 644 at intercellular junctions and structures associated with junction turnover in rat Sertoli cells