Microscopy deep learning predicts virus infections and reveals mechanics of lytic-infected cells

•Artificial intelligence identifies HSV- and AdV-infected cells without specific probes.•Imaging lytic-infected reveals nuclear envelope rupture AdV dissemination.•Live cell imaging neural networks presciently pinpoint cells.•Lytic-infected nuclei have mechanical properties distinct from non-lytic nuclei. Imaging across scales disease mechanisms in organisms, tissues, cells. Yet, particular infection phenotypes, such as virus-induced lysis, remained difficult to study. Here, we developed modalities deep learning procedures identify herpesvirus adenovirus (AdV) infected virus-specific stainings. Fluorescence microscopy of vital DNA-dyes live-cell revealed learnable patterns transferable related viruses the same family. Deep predicted two major outcomes, (nonspreading) lytic (spreading) infections, up about 20 hr prior lysis. Using these predictive algorithms, had levels green fluorescent protein (GFP)-tagged virion proteins but enriched faster, collapsed more extensively upon laser-rupture than nuclei, revealing impaired Our algorithms may be used infer phenotypes emerging viruses, enhance single biology, facilitate differential diagnosis infections. Virus infections give rise a wide range with complex biology. Depending on virus, can lead cycle growth enhancement or arrest, swelling shrinkage, membrane blebbing contraction, organelle alterations, collapse secretory pathway, mitochondrial aggregation condensation, increasingly studied virus resolution (Belov, 2016Belov G.A. Dynamic lipid landscape picornavirus replication organelles.Curr. Opin. Virol. 2016; 19: 1-6Crossref PubMed Scopus (18) Google Scholar; den Boon et al., 2010den J.A. Diaz A. Ahlquist P. Cytoplasmic viral complexes.Cell Host Microbe. 2010; 8: 77-85Abstract Full Text PDF (237) Roulin 2014Roulin P.S. Lotzerich M. Torta F. Tanner L.B. van Kuppeveld F.J. Wenk M.R. Greber U.F. Rhinovirus uses phosphatidylinositol 4-phosphate/cholesterol counter-current for formation compartments at ER-golgi interface.Cell 2014; 16: 677-690Abstract (114) Suomalainen 2020Suomalainen Prasad V. Kannan Cell-to-cell genome-to-genome variability transcription tuned by cycle.J. Cell Sci. 2020; 134 (jcs.252544)PubMed White 2014White M.K. Pagano J.S. Khalili K. Viruses human cancers: long road discovery molecular paradigms.Clin. Microbiol. Rev. 27: 463-481Crossref (83) Scholar). Alterations pro- anti-viral comprise metabolic changes, shifts glucose, nucleotide metabolism, well apoptotic, necroptotic, lysis (Delgado 2010Delgado T. Carroll P.A. Punjabi A.S. Margineantu D. Hockenbery D.M. Lagunoff Induction Warburg effect Kaposi's sarcoma is required maintenance latently endothelial cells.Proc. Natl. Acad. U S 107: 10696-10701Crossref (140) 2018Lotzerich Boucke Witte R. Georgiev O. 3C protease suppresses apoptosis triggers caspase-independent death.Cell Death Dis. 2018; 9: 272Crossref (19) Munger 2008Munger J. Bennett B.D. Parikh Feng X.J. McArdle Rabitz H.A. Shenk Rabinowitz J.D. Systems-level flux profiling fatty acid synthesis target antiviral therapy.Nat. Biotechnol. 2008; 26: 1179-1186Crossref (432) Prusinkiewicz Mymryk, 2019Prusinkiewicz M.A. Mymryk Metabolic reprogramming host infection.Viruses. 2019; 11: 141Crossref (28) While programmed death triggered dissemination has strong anti-inflammatory effects, virus-controlled typically proviral, releases large amounts progeny virions, enhances inflammation, disseminates virions neighboring other organisms. The onset overshadows transmission, shown recently (Georgi 2020aGeorgi Andriasyan Murer L. Hemmi S. Yu Grove Meili N. Kuttler Yakimovich al.The FDA-approved drug Nelfinavir inhibits cell-free, not cell-associated transmission adenovirus.Antimicrob. Agents Chemother. 64: e01002-e01020Crossref (4) Scholar, Georgi 2020bGeorgi Turcatti G. A high-content image-based screen clinical compounds against adenovirus.Sci. Data. 7: 265Crossref (1) been observed both enveloped nonenveloped viruses. It exploited applications oncolytic herpes simplex (HSV), measles Newcastle vaccinia reovirus, parvovirus, coxsackie polio (reviewed Lawler 2017Lawler S.E. Speranza M.C. Cho C.F. Chiocca E.A. Oncolytic cancer treatment: review.JAMA Oncol. 2017; 3: 841-849Crossref (230) detect ultrastructural analyses, which provide snap-shots insights, although time-lapse fluorescence single-cell promises directly visualize events Pied Wodrich, 2019Pied Wodrich H. cycle.FEBS Lett. 593: 3419-3448Crossref (7) Puntener 2011Puntener Engelke M.F. Ruzsics Z. Strunze Wilhelm C. Stepwise loss core V during entry into cells.J. 2011; 85: 481-496Crossref (58) Wild 2019Wild Leisinger de Oliveira A.P. Doehner Schraner E.M. Fraevel Ackermann Kaech Nuclear impairment facilitated 1 Us3 kinase.F1000Res. 198Crossref 2012Yakimovich Gumpert Burckhardt C.J. Lutschg V.A. Jurgeit Sbalzarini I.F. Cell-free passive mass transfer culture simulated computer model.J. 2012; 86: 10123-10137Crossref (42) Yamauchi 2008aYamauchi Y. Kiriyama Kimura Nishiyama Herpes induces extensive modification dynamic relocalisation mitotic apparatus (NuMA) interphase 121: 2087-2096Crossref (12) key technology life sciences. Many its original limitations, including sensitivity, sample penetration depth, spatiotemporal resolution, toxicity resolved improvements optics chip technology, fluorophore chemistry (Sahl 2017Sahl S.J. Hell S.W. Jakobs nanoscopy biology.Nat. Mol. Biol. 18: 685-701Crossref (403) Wang 2018Wang I.H. Imaging, tracking computational analyses egress cytoskeleton.Viruses. 10Crossref (34) 2018Witte Concepts light viruses.Viruses. (20) In recent years, it enhanced artificial (AI). Particularly methods allow data analysis, often par capabilities (Dechter, 1986Dechter Learning while Searching Constraint-Satisfaction Problems. C.S.D. University California, Cognitive Systems Laboratory, 1986http://www.aaai.org/Papers/AAAI/1986/AAAI86-029.pdfGoogle LeCun 2015LeCun Bengio Hinton learning.Nature. 2015; 521: 436-444Crossref (27824) AI are biomedical sciences, example image recognition, restorations low sampling, object segmentation (Christiansen 2018Christiansen Yang Ando Javaherian Skibinski Lipnick Mount E. O'Neil Shah Lee A.K. al.In silico labeling: predicting labels unlabeled images.Cell. 173: 792-803.e19Abstract (191) Falk 2019Falk Mai Bensch Cicek Abdulkadir Marrakchi Bohm Deubner Jackel Seiwald al.U-Net: counting, detection, morphometry.Nat. Methods. 67-70Crossref (372) Fisch 2019Fisch Clough B. Wright Bunyan Howell Mercer Frickel Defining host-pathogen interactions employing an workflow.eLife. e40560Crossref (0) 2020Lee J.Y. Sadler N.C. Egbert R.G. Anderton C.R. Hofmockel K.S. Jansson J.K. Song H.S. predicts microbial self-organized patterns.Comput. Struct. 1259-1269Crossref (3) Ounkomol 2018Ounkomol Seshamani Maleckar M.M. Collman Johnson G.R. Label-free prediction three-dimensional images transmitted-light microscopy.Nat. 15: 917-920Crossref (124) Silver 2017Silver Schrittwieser Simonyan Antonoglou I. Huang Guez Hubert Baker Lai Bolton al.Mastering game Go knowledge.Nature. 550: 354-359Crossref (2762) Weigert 2018Weigert Schmidt U. Boothe Muller Dibrov Jain Broaddus Culley al.Content-aware restoration: pushing limits 1090-1097Crossref (201) Yakimovich, 2019Yakimovich mSphere influence: biology.mSphere. 4: e00315-19Crossref (2) Recently, pattern recognition (ANNs). One first implementations ANN was multilayer perceptron (MLP). Inspired architecture animal visual cortex (Hubel Wiesel, 1968Hubel D.H. Wiesel T.N. Receptive fields functional monkey striate cortex.J. Physiol. 1968; 195: 215-243Crossref (4070) Scholar), MLPs layers fully connected nodes (artificial neurons), where each layer's output input succeeding layers. Each neuron MLP represents non-linear function trainable weights applied signal generate output. trained known perform better conventional machine methods, time being notoriously hard train prone overfitting. introduction convolutional (CNNs) addresses this issue exploiting fact that broken down successively simpler features (convolutions). CNNs composed three basic building blocks – layers, pooling dense Convolutional apply random convolution filters sub-regions derive feature map, nonlinear applied. This layer helps extract images. down-sample maps extracted reduces dimensionality preserving important information. Dense essentially usually final refinement CNN yield ultimate classification based previous distinguish reusable network architecture, makes versatile retraining addressing similar, yet unrelated problems new (Rawat Wang, 2017Rawat W. classification: comprehensive review.Neural Comput. 29: 2352-2449Crossref current state-of-the-art residual framework shorten training make them accurate. They called (ResNets) (He 2015He Zhang X. Ren Sun recognition.arXiv. (1512.03385)Google ResNets widely vision scientists, mimic functionality brain, neurons arranged cortical receive axonal dendritic connections receiving neuron, thereby skipping Such skip connection strategies ResNet further mitigate overfitting address problem vanishing gradient. present ResNet, termed ViResNet, detects HSV-1 very high accuracy. ViResNet qualitatively different classical learning, implicitly morphological stained DNA-binding agent Hoechst 33342 imaged (Latt Stetten, 1976Latt S.A. Stetten Spectral studies 33258 bisbenzimidazole dyes useful detection deoxyribonucleic synthesis.J. Histochem. Cytochem. 1976; 24: 24-33Crossref 2017Yakimovich Huttunen Zehnder Coulter L.J. Gould Schneider Kopf McInnes Inhibition poxvirus gene expression genome bisbenzimide derivatives.J. 91: e00838-17Crossref We demonstrate learned effectively annotate serotypes included set network. abundant double-stranded DNA respectively. replicate assemble capsids nucleus (Berk, 2013Berk A.J. Adenoviridae: their replication.in: Knipe Howley P.M. Fields Virology. Lippincott Williams & Wilkins, 2013: 1704-1731Google Roizman Knipe, 2001Roizman Fundamental Lippincott-Raven), 2001: 1123-1183Google causes persistent acute mucosal tissue latency phenotype trigeminal ganglia nervous system (Steiner 2007Steiner Kennedy P.G. Pachner A.R. neurotropic viruses: varicella-zoster.Lancet Neurol. 2007; 6: 1015-1028Abstract (341) Poorly controlled encephalitis, conjunctivitis, Bell's palsy, eczema, genital skin lesion. AdVs, turn, infect respiratory, ocular digestive tracts, blood persist lymphoid intestines tonsils (Greber, 2020Greber Adenoviruses - infection, pathogenesis therapy.FEBS 594: 1818-1827Crossref 2013Greber Arnberg Wadell Benko Kremer E.J. pathogens therapeutics: report 10th International meeting.Cell 2013; 16-23Crossref (16) Flatt, 2019Greber Flatt J.W. Adenovirus entry: immunity.Annu. 177-197Crossref Lion, 2014Lion immunocompetent immunocompromised patients.Clin. 441-462Crossref (370) Upon immunosuppression, they spread intestinal epithelial unknown pathways cause severe morbidity mortality (Kosulin 2016Kosulin Geiger Vecsei Huber W.D. Rauch Brenner Wrba Hammer Innerhofer Potschger al.Persistence reactivation adenoviruses gastrointestinal tract.Clin. Infect. 22: 381 e381-388Abstract (63) high-fidelity recognizes prospective cells, staining live imaging. These allowed us physical contrasting non-lytic-infected results extend trafficking (Burckhardt 2011Burckhardt Schoenenberger Drifting motions receptor CAR immobile integrins initiate uncoating exposure.Cell 10: 105-117Abstract Chenouard 2014Chenouard Smal Chaumont Maska Gong Cardinale Carthel Coraluppi Winter al.Objective comparison particle methods.Nat. 281-289Crossref (496) Gazzola 2009Gazzola Bayati Koumoutsakos stochastic model microtubule motors describes vivo cytoplasmic transport adenovirus.PLoS 2009; 5: e1000623Crossref Helmuth 2007Helmuth novel supervised trajectory algorithm types motion 159: 347-358Crossref (73) 2009Helmuth Shape reconstruction subcellular structures images.J. 167: 1-10Crossref Sommer Gerlich, 2013Sommer Gerlich D.W. Machine biology teaching computers recognize phenotypes.J. 126: 5529-5539Crossref work here provides resource analyzing challenging topic owing rapid terminal nature death. To explore power content implemented pipeline scoring morphology mode DNA-dye (Figures 1A 1B ). noticed Hoechst-stained appeared uninfected presence throughout multi-round cycles did affect plaques (Figure S1A). wild-type AdV-C2 immuno-staining, GFP-expressing replication-competent AdV-C2-dE3B-GFP, AdV-C2-dE3B-dADP, AdV-C2-V-GFP, HSV-1-VP16-GFP, C12 (Crameri 2018Crameri Bauer Caduff Walker Steiner Franzoso F.D. Gujer Kucera Zbinden al.MxB interferon-induced restriction factor herpesviruses.Nat. Commun. 1980Crossref (47) Glauser 2010Glauser D.L. Seyffert Strasser Franchini Laimbacher Dresch Vogel Buning Salvetti al.Inhibition type adeno-associated rep depends combined ATPase/helicase activities.J. 84: 3808-3824Crossref (15) 2008bYamauchi Kubota Usukura UL14 tegument efficient alpha transinducing VP16 capsids.J. 82: 1094-1106Crossref (29) For AdV, acquired 72 post (pi), 48 pi, reflecting kinetics were computationally segmented individually annotated according signal, processed obtain individual S1B). procedure generated 2000 per condition. dataset using modified cutting-edge ResNet-50 Scholar) named ViResNet. classifiers, one S1C). accommodate resized 224 x pixels, adapted classes, additional hidden untrained initialized ImageNet weights. accuracies 95% 94% considerably higher k-nearest neighbors (k-NN), support vector (SVM), decision tree all test (Table 1). classifiers precision values 0.94 recall 0.92, respectively, much those determined namely 0.6 0.58 average HSV-1, receiver operating characteristic (ROC) curves k-NN, SVM, Figure S1D. 2-class then assess status population level. allows visualization quantification enables assessment plaque 1B). challenge performed AdV-C5-IX-FS2A-GFP, serotype accuracy 93% S1E).Table 1Accuracy, precision, performance detecting AdV-C2- HSV-1-infected compared methodsAccuracy=TP+TNTP+TN+FP+FNPrecision=TPTP+FPRecall=TPTP+FNAdV-C2:ViResNet0.950.940.92k-NN0.640.620.61SVM0.630.630.58Decision tree0.640.620.59HSV-1:ViResNet0.940.920.93k-NN0.640.610.56SVM0.610.580.59Decision tree0.650.620.61TP, TN, FP, FN denote true positives, negatives, false negatives predictions correspondingly. Open table tab TP, sum, accurately identified plaquing populations, outperformed methods. Notably, identification possible absence staining, previously possible. next employed analyze AdV-C2-GFP-V. incorporates 40 GFP-V fusion particle, could purified tracked (Puntener Infection HeLa stably expressing histone H2B-mCherry AdV-C2-GFP-V multiplicity (MOI) 0.2 showed newly synthesized accumulated 30 pi 2A, Video S1). At 39 localization changed, indicated increased diffuse cytosol, followed appearance clusters cytoplasm. largely devoid H2B-mCherry, consistent notion made particles, histones (Ostapchuk 2017Ostapchuk Zheng Hearing VII dispensable assembly essential infection.PLoS Pathog. 13: e1006455Crossref (24) After clusters, cytoplasm progressively lost GFP-V, also albeit extents, indicative cellular least partial disintegration 2A 2B, Videos S1 S2). 5.5% until 45 2C). confirmed EM 35 clustered particles 2D). Together show induced comprising several hundred extracellular medium. Towards harnessing assessing quantifying spreading efficiency high-throughput MOI 4 days. Monolayers A549 AdV-C2-GFP-V, AdV-C2-dE3B-GFP-dADP AdV-C2-dE3B-GFP lacks E3B region encoding internalization degradation RID? RID?, 14.7K, involved protecting premature tumor necrosis (Gooding 1988Gooding L.R. Elmore L.W. Tollefson A.E. Brady Wold W.S. 14,700 MW E3 cytolysis