Introduction

Shigella flexneri is a Gram-negative bacterium and the etiologic agent of bacillary dysentery, an intestinal infectious disease that kills more than 150 thousand people worldwide per year, disproportionally affecting children under 5 years old from low- and middle-income countries (Global Burden of Disease Collaborative Network., 2020). After the ingestion of a small number of bacteria (10-100), S. flexneri rapidly colonises the human intestinal epithelium. For this, S. flexneri uses a Type 3 Secretion System (T3SS) that injects bacterial effectors into the host cell to modulate key aspects of the infection process and to evade innate immune defences by the host (Schnupf & Sansonetti, 2019). Following uptake by host cells, S. flexneri breaks its vacuolar compartment to access the cytosol (Ray et al., 2010). Here, the virulence factor IcsA enables S. flexneri to hijack host actin and polymerise comet tails that render bacteria motile to disseminate to neighbouring cells and spread in the tissue (Bernardini et al., 1989; Egile et al., 1999).

Septins are highly-conserved GTP-binding proteins that form heteromeric complexes (hexamers, octamers) that assemble into non-polar filaments, bundles and rings (Martins et al., 2023; Soroor et al., 2021; Szuba et al., 2021; Woods & Gladfelter, 2021). Septins are involved in many physiological cellular processes (eg. cytokinesis, cilia formation), and play key roles during infection (Mostowy & Cossart, 2012; Robertin & Mostowy, 2020; Spiliotis & Nakos, 2021). Septins form collar-like structures at the plasma membrane during bacterial invasion (Boddy et al., 2018; Mostowy et al., 2009; Robertin et al., 2023), and ring-like structures surrounding membrane protrusions in cell-to-cell bacterial spread (Mostowy et al., 2010). In the cytosol septins are recruited to actin-polymerising bacteria, including S. flexneri and Mycobacterium marinum, as cage-like structures (Mostowy et al., 2010). In the case of S. flexneri, septins are recruited to μm-scale curvature at bacterial poles and the division site, which are enriched in the negatively charged phospholipid cardiolipin (Krokowski et al., 2018). Recent work has shown that IcsA promotes the recruitment of septins to S. flexneri, and that lipopolysaccharide (LPS) inhibits S. flexneri-septin cage entrapment (Lobato-Márquez et al., 2021; Mostowy et al., 2010). Septin cages prevent bacterial actin tail motility and therefore are considered as an antibacterial mechanism of host defence (Krokowski et al., 2018; Mostowy et al., 2010).

Intracellular infection is usually portrayed as a uniform, synchronised and highly orchestrated process. However, infection assays performed in controlled laboratory environments with isogenic populations display a high level of complexity (Avraham et al., 2015; Toniolo et al., 2021). It is particularly unknown how different subsets of bacteria (eg. motile vs. non motile bacteria) can colonise distinct intracellular microenvironments and become targeted by multiple cell-autonomous immunity mechanisms (eg. septin cages, autophagy). It has been proposed that this diversity of intracellular niches impacts bacterial physiology (Day et al., 2023; Gutierrez & Enninga, 2022; Santucci et al., 2021). For example, it may affect individual bacterial ability to acquire nutrients, to be exposed to varying antibiotic concentrations, or to elicit different responses and signalling pathways within the host cell. While the concept of phenotypic heterogeneity in prokaryotic clonal populations is well established (Ackermann, 2015), the concept is emerging also in eukaryotic cells where non genetic differences may for instance explain individual cell sensitivity and responses by cell-autonomous immunity (McDonald & Dedhar, 2024). It is therefore paramount to study host and bacterial cells at the single-cell level, to assess the individual responses that collectively contribute to the infection process.

Here, to dissect the complexity of S. flexneri infection in epithelial cells, we use high-content high-resolution microscopy coupled to automated image analysis. High-content microscopy enables the automated acquisition of thousands of microscopy images and therefore is useful to capture the heterogeneity of individual host and bacterial cells during the infection process for quantitative analysis (Aylan et al., 2023; Brodin & Christophe, 2011; Deboosere et al., 2021; Dramé et al., 2023; Fisch et al., 2019a; Lensen et al., 2023; Pylkkö et al., 2021). We analyse multiple parameters, such as morphological host cell features, the de novo synthesis of DNA and proteins in both host and bacterial cells, as well as activation of the bacterial T3SS. To investigate the antimicrobial potential of septin-mediated immunity, we develop a deep-learning based approach to automatically identify septin-S. flexneri interactions. Together, we highlight the power of high-content microscopy and artificial intelligence to reveal insights on both host and bacterial cells, and illuminate fundamental biology underlying the S. flexneri infection process.

Results

Capturing the heterogeneity of S. flexneri infection using high-content microscopy

To understand the complexity and reveal the heterogeneity of S. flexneri infection in human epithelial cells, we dissected the infection process using high-content high-resolution microscopy (Figure 1A, B). We observed a mean of 11.5% of infected cells measured by fluorescence at 3h 40 min post infection (hpi) at MOI 100:1 (Figure 1C). These infected cells presented different burdens of bacteria following an exponential decay distribution (Figure 1D), reflective of the formation of infection foci (Ortega et al., 2019). With growing evidence of the cytosol being a non-homogeneous compartment containing concentration gradients, biomolecular condensates, protein complexes and concentrates (van Tartwijk & Kaminski, 2022), we tested whether S. flexneri may be exposed to different intracellular microenvironments. For this, we measured the distance of bacteria to the centroid of cells (Figure 1E) and observed a median distance of 8.74 µm. The median distance to the centroid of the cell significantly decreased with increasing bacterial burden (Figure 1E), suggesting that the microenvironment of bacteria is diverse and changes as bacteria replicate intracellularly and populate the central parts of host cytosol after uptake at the host cell periphery. To dissect the heterogeneity in host cells, we quantified an increase of 36.9% and 22.3% in the cellular and nuclear area of infected cells, respectively (Figure 1G, I), which was dependent on the infection burden (Figure 1H, J). To test whether this increase reflected an increase in volume, we measured the total amount of Hoechst fluorescence in the nucleus of the cell as a proxy for nuclear volume (Figure 1-figure supplement 1A) and observed a 29.4% increase in infected cells. To test if this effect was due to morphological rearrangements during infection, we treated cells with Latrunculin B to inhibit actin polymerisation. While cells treated with Latrunculin B decreased their cellular area (but not their nuclear area) (Figure 1-figure supplement 1B, C), the drug did not abrogate the difference in size between infected and uninfected cells. Taken together, these results highlight an increase in cellular volume during infection that is dependent on the infection burden.

Figure 1.

Epithelial cells reduce de novo DNA and protein synthesis upon infection

In order to understand the physiological state of enlarged epithelial cells infected with S. flexneri, we monitored host DNA replication and protein translation using Click-chemistry. Using a 1 h pulse of the clickable thymidine analogue Edu, we specifically visualized and quantified de novo DNA synthesis in the population of epithelial cells with high-content microscopy (Figure 2A, Figure 2-figure supplement 1). We identified the subpopulation of cells in S-phase by their high level of Edu incorporation. Consistent with this, this subpopulation of cells was fully abrogated in the presence of the S-phase blocker Aphidicolin (Pedrali-Noy et al., 1980) (Figure 2-figure supplement 2A, B). Approximately one third of uninfected cells (34.45%) were found in S-phase (Figure 2B, C), in agreement with S-phase lasting around one third of the cell cycle (∼8h out of ∼22h). Infected cells presented a similar proportion of cells in S-phase (35.89%), indicating that S. flexneri infection did not block transition of cells from G1 to S at the timepoint tested. We then assessed the amount of de novo DNA synthesis incorporated in cells in S-phase during infection. Surprisingly, we observed a reduction of Edu incorporation in infected cells (Figure 2D), which was dependent on the bacterial burden (Figure 2E). These results indicate that S. flexneri infection slows down DNA synthesis of epithelial cells.

Figure 2:

To assess protein translation, we used a 1h pulse of the clickable methionine analogue AHA, which specifically labels de novo synthesized proteins as AHA incorporation is reduced after treatment with the inhibitor of translational elongation Cycloheximide (Ennis & Lubin, 1964) (Figure 2-figure supplement 1C, D). During infection, we observed a reduction of protein translation in infected cells compared to non-infected cells (Figure 2F,G), which was also dependent on the bacterial burden (Figure 2H).

Collectively, the cell size increase together with the reduction of de novo DNA and protein synthesis in infected epithelial cells suggests that S. flexneri induces an arrest in the host cell cycle.

Design of a deep learning pipeline to automatically identify bacteria interacting with SEPT7

S. flexneri have been shown to interact with septins during infection. These interactions include septin ring-like structures surrounding phagocytic cups, autophagosomes, actin tails and protrusions, all being morphologically diverse structures that are normally scored by hand (Krokowski et al., 2018; Lobato-Márquez et al., 2023; Mostowy et al., 2010). To study septin-associated bacteria using high-content microscopy, we designed a tailored analysis pipeline for the automatic and unbiased identification of SEPT7-bacteria interactions (Figure 3). SEPT7 is a core component of septin hexamers and octamers, and is historically used to identify septin structures (Mostowy et al., 2010). The analysis pipeline includes two deep learning models based on Convolutional Neural Networks (CNN), which are effective in solving classification tasks in computer vision (Krizhevsky et al., n.d.; Lecun et al., 2015; Nielsen, 2015). As septins can surround individual and also multiple bacteria simultaneously (Figure 4-figure supplement 1A), we focused on individual bacteria. For this, we trained the first classification model to discriminate between “single” bacteria and “clumps”. The second classification model was designed to assign those single bacteria into “SEPT7-positive structures” or alternatively “negatives”, according to the morphology of SEPT7 recruitment around them.

Figure 3.

Figure 4.

The first classification model (discerning “single” versus “clumped” bacteria) together with its training and performance metrics is summarized in Figure 4. A large dataset of annotated bacteria (Figure 4A, B, Figure 4-figure supplement 1B) was used to train a Sequential model with a typical CNN architecture that was experimentally tested and optimised accordingly for best results (Figure 4C). This model provided a good fit of the training data (high accuracy, low loss), with similar predictive power for the validation dataset (Figure 4D, E). Performance of the model was further assessed with a confusion matrix where a separate annotated test dataset was challenged and correct number of predictions scored, showing that 96% of SEPT7 structures and 83% of clumps were correctly identified (Figure 4F). Finally, effective performance of the model was ensured with high values achieved in additional metrics testing for type I and II errors (precision, recall, F1-score) (Figure 4G).

The second classification task (identification of “SEPT7 positive” vs “negative” bacteria) was more challenging due to their intrinsic heterogeneity plus the presence of ubiquitious SEPT7 filaments that populate the vicinity of bacteria (but do not associate with bacteria per se). To address these challenges, we increased the amount of data and complexity of the second classification model (Figure 4H-J, Figure 4-figure supplement 1C) compared to the first (Figure 4A-C). This included more than 1200 images of S. flexneri associated to SEPT7, both from infections with wild-type bacteria or expressing AfaI to increase bacterial invasion. The training process achieved a high accuracy and low loss for the training and validation dataset (Figure 4K, L). Correct prediction of manually annotated SEPT7 positive structures was 80% as shown in the confusion matrix (Figure 4M), with performance metrics Precision, Recall and F1 score also around 0.8 (Figure 4N).

Characterization of the morphological heterogeneity of SEPT7 recruitment to S. flexneri

The training of the classification model to identify bacteria associated to SEPT7 required a large dataset of images. We observed diverse morphological conformations of SEPT7 signal around bacteria (Figure 5 A-C). We differentiated five frequent subcategories from a total of 855 cases. The first corresponds to most cases (66.7%), where SEPT7 forms rings that entrap bacteria transversally (Figure 5A, B, a). We also observed two categories where SEPT7 surrounds bacteria more homogeneously either in a tight (18.2%) or loose association (6.1%) (Figure 5A, B, b-c). Finally, we observed partial SEPT7 recruitment, with only one (5.0%) or two (4.0%) bacterial poles being targeted (Figure 5A, B, d-e). Given this diversity, we combined the entire dataset to summarise the probability of SEPT7 interactions with bacteria (Figure 5-figure supplement 1). As bacteria are found in any orientation inside epithelial cells, we included only those instances where the z-projection of the bacterial axis measured more than 2.5μm, ensuring they were positioned horizontally when imaged. Representations shown in Figure 5-figure supplement 1A, C and graphs in Figure 5-figure supplement 1B, D can be understood as a probability map of the area of influence of SEPT7 when associating with intracellular S. flexneri.

Figure 5.

SEPT7 is recruited to S. flexneri with increased T3SS activity

Septins have previously been shown to recognise growing bacterial cells in vitro and during infection of HeLa cells, using pharmacologic and genetic manipulation (Krokowski et al., 2018; Lobato-Márquez et al., 2021). To test the metabolic state of SEPT7-associated S. flexneri at 3h 40 min post infection (hpi) we used our high-content imaging and analysis pipeline (Figure 6). We measured S. flexneri de novo synthesis of DNA and proteins using a 1 h pulse Click-IT Edu and Click-IT AHA. Only active bacteria incorporated these analogues, as the antibiotic Nalidixic acid (inhibitor of the bacterial DNA gyrase (Sugino et al., 1977)) and Rifampicin (inhibitor of the bacterial RNA polymerase (Ezekiel & Hutchins, 1968)) significally reduced their incorporation, respectively (Figure 6-figure supplement 1A-D, Figure 6-figure supplement 1E,F). During infection, bacteria associated with SEPT7 showed heterogeneus levels of DNA synthesis and protein translation similar to SEPT7-negative bacteria (Figure 6A-H).

Figure 6.

S. flexneri pathogenicity relies on its T3SS. To understand effector secretion at the individual bacterial level, we used S. flexneri expressing the Transcription-based Secretion Activity Reporter (TSAR) (Campbell-Valois et al., 2014). This reporter comprises a short lived fast-maturing superfolder GFP (GFPmsf2) that is expressed in conditions where the T3SS is active, such as in broth upon addition of congo red (Figure 7-figure supplement 1A, B) (Parsot et al., 1995) and during infection (Figure 7). We observed heterogeneity in TSAR signal (Figure 9A,B). T3SS activation was higher in bacteria found in cells with a low bacterial burden (less than 10 intracellular bacteria) (Figure 7A,C). This suggests the highest level of activation occurs in bacterial pioneers that colonise neighbouring cells, in agreement with (Campbell-Valois et al., 2014). We then assessed T3SS activation in S. flexneri associated with SEPT7 using our deep learning analysis pipeline (Figure 7D-H). We detected a significant increase in the percentage of TSAR-positive S. flexneri and the overall intensity of TSAR signal compared to SEPT7-negative bacteria (Figure 7G,H). These include SEPT7-associated protrussions and cage-like structures (Figure 7E, Figure 7-figure supplement 1C). These data underscore the antimicrobial role of septins in associating with actively pathogenic bacteria.

Figure 7.

Discussion

Unravelling the intrinsic heterogeneity of host and bacterial cells underlying the infection process is paramount to design effective therapeutic strategies. High-content and high-throughput image analysis are ideal tools to capture this heterogeneity when coupled to automatic phenotypic image analysis (Eliceiri et al., 2012; Mattiazzi Usaj et al., 2016; Pegoraro & Misteli, 2017). Here we use high-content high-resolution microscopy and automated image analysis to quantitatively dissect the heterogeneity of S. flexneri in epithelial cells at the single-cell and population level at unprecedented subcellular detail.

We discover that S. flexneri induces important physiological changes in epithelial cells after 3h40min post infection. These include a marked decrease of DNA synthesis in cells in S-phase, a collective shutdown of protein translation, and a larger cellular and nuclear area, all in a bacterial-burden dependent manner. Collectively, these observations suggest S. flexneri infection induces host cell stress and a halt in the normal progression of the cell cycle. Consistent with this, S. flexneri has been shown to induce genotoxic stress (Bergounioux et al., 2012), inhibit translation and induce formation of stress granules (Zhang et al., 2024), and block the G2/M phase transition (Iwai et al., 2007; Wang et al., 2019). In the natural S. flexneri niche of the human gut, such an arrest would prevent maturation and shedding of infected epithelial cells, thereby promoting efficient gut colonisation (Iwai et al., 2007).

To understand how bacteria are targeted by cell-autonomus immune defences, such as septin cage entrapment, it is necessary to investigate phenotypic characteristics of targeted bacteria and the heterogeneous recruitment of cell-autonomous immune factors. While traditional image segmentation tools fail to enable this task, rapidly evolving developments in the field of computer vision are democratising the use of deep learning and CNNs for multiple applications (Berg et al., 2019; Hung et al., 2020; Stringer et al., 2020; von Chamier et al., 2021; Weigert et al., 2020) including infection biology (Davidson et al., 2021; Fisch et al., 2019b, 2021; Touquet et al., 2018). However, the analysis of complex datasets still requires a tailored approach, as shown here for bacterial interactions with the septin cytoskeleton (which is ubiquitious in the cell and forms diverse assemblies whose precise identifcation requires high resolution). We successfully implement a CNN based approach for the automated classification of bacteria associated with SEPT7 assemblies, enabling the analysis of thousands of bacterial cells while saving time and preventing human bias.

A large dataset of images depicting S. flexneri – SEPT7 interactions was acquired to train the deep learning algorithm and to characterize the diversity of these structures. We observe that the most frequent septin assembly is in the form of rings transversally surrounding the bacterial surface. We also detected alternative morphologies, including homogeneous SEPT7 labelling surrounding the bacterial surface (tight vs loose), as well as the discrete recruitment of SEPT7 to one or two bacterial poles. While previous work has shown that septin cage assembly around S. flexneri can start with the discrete recruitment of septins to one or two bacterial poles (Krokowski et al., 2018), it is not yet known if the different septin assemblies we describe here correspond to different stages or fates of the same structure or whether it marks separate biological events.

We use deep learning analysis to interrogate intracellular bacteria and show that S. flexneri associated with SEPT7 present heterogenous DNA and protein synthesis. In addition, we show that bacteria associated with SEPT7 have increased T3SS activity (as compared to bacteria not associated with SEPT7) indicating that SEPT7 is recruited to actively pathogenic bacteria. T3SS activity has been shown to be higher in motile bacteria (Campbell-Valois et al., 2014) which reinforces the antimicrobial role of septins in counteracting actin tail motility. Altogether, our deep learning tool to analyse septin-bacteria interactions opens the door to future studies on interactions between bacteria and components of cell-autonomous immunity.

Concluding remarks

Heterogeneity in a controlled lab environment is only a small representation of the broader heterogeneity underlying the clinical infection process, considering the genetic background of the patient, interactions with microbiota, etc. In the future, our pipeline can be applied to investigate septin-bacteria interactions among different Shigella spp. or clinically relevant strains, as well as other intracellular bacteria (eg. M. marinum, Staphylococcus aureus, Pseudomonas aeruginosa). In addition, our pipeline can be adapted to identify and score other cytoskeletal structures (such as actin tails) or membrane organelles (such as autophagosomes) associated with intracellular bacteria. Ultimately, we envision that our automated microscopy workflows will help to screen pharmacologic or genetic libraries in a multiparametric manner and understand their impact on host and bacterial cells at the single-cell and population level. In this way, novel host and bacterial factors underlying heterogeneity and bacterial pathogenesis could be identified, leading to transformative therapeutic strategies to combat bacterial infection in humans.

Materials and methods

Reagents

The following antibodies were used: rabbit anti-SEPT7 (1:100 dilution, #18991, IBL), Alexa-488-conjugated anti-rabbit antibody (1:500, #10729174, ThermoFisher Scientific), Alexa-647-conjugated anti-rabbit antibody (1:500, #A27040, ThermoFisher Scientific). The following dyes were used: Alexa-647-conjugated phalloidin (1:500, #10656353, ThermoFisher Scientific), Hoechst (1:500, #H3570, ThermoFisher Scientific), 0.01% (w/v) congo red (#C6767, Sigma-Aldrich). The following drugs were used: 50 μg/mL aphidicolin (#A0781 Sigma-Aldrich), 50 μg/mL cycloheximide (#2112 Cell Signalling), 2 μg/mL nalidixic acid (#N8878 Sigma-Aldrich), and 100 μg/ml rifampicin (#R3501 Sigma-Aldrich).

Bacterial strains and culture conditions

The bacterial strains and plasmids described in this study are listed in Sup. Table 1. Shigella flexneri 5a str. M90T mCherry was used throughout the manuscript unless indicated otherwise unless indicated otherwise. S. flexneri strains were grown in trypticase soy broth (TCS) agar containing congo red to select for red colonies, indicative of a functional T3SS. Conical polypropylene tubes (#CLS430828, Corning) containing 5 ml of TCS were inoculated with individual red colonies of S.flexneri and were grown ∼16 h at 37 °C with shaking at 200 rpm. The following day, bacterial cultures were diluted in fresh prewarmed TCS (1:50 or 1:100 v/v), and cultured until an optical density (OD, measured at 600 nm) of 0.6.

Cloning

Primers used in this study were designed using NEBuilder Assembly Tool (https://nebuilder.neb.com/#!/). To produce S. flexneri expressing AfaI and GFP, the pACAfaI plasmid was constructed (plasmids characteristics and primer sequences are located in Supplementary Table 1). Briefly, the afaI operon together with the amp^Rgene was amplified from S.flexneri afaI strain using the primers afaI-operon-fw and afaI-operon-rv. The ori p15A was amplified from pAC-P_lac using the primers p15A-ori-fw and p15A-ori-fw. Both DNA fragments were assembled using the Gibson assembly to generate the plasmid pAC-AfaI-AmpR. In a second step, the AmpR gene from pAC-AfaI-Amp was substituted by the CmR gene and generate pAC-AfaI-Cm. This was achieved by amplification of the CmR gene from pNeae2 using the primers CmR-fw CmR-rv and amplification of the backbone of pAC-AfaI-Amp with primers afaI-p15A-fw, afaI-p15A-rv, followed by Gibson assembly. Gibson assembly was performed at 50 °C for 30 min using the HiFi DNA Assembly Master Mix (#E2621L, New England Biolabs). Resulting plasmid pAC-AfaI-Cm was transformed into S. flexneri GFP by electroporation.

Mammalian cell culture

HeLa (ATCC CCL-2) cells were grown at 37 °C and 5% CO₂ in Dulbecco’s modified Eagle medium (DMEM, GIBCO) supplemented with 10% fetal bovine serum (FBS, Sigma-Aldrich).

Infection of human cells

Hela cells were seeded in 96-well microplates (PhenoPlates, Perkin Elmer) at a confluency of 10⁴ cells/well 2 days prior to infection. Alternatively, 9 × 10⁴ HeLa cells were seeded in 6-well plates (Thermo Scientific) containing 22 × 22 mm glass coverslips 2 days before the infection. Bacterial cultures were grown as described and cell cultures were infected with S. flexneri strains as described previously (Mostowy et al., 2010). Briefly, Hela cells were infected with S. flexneri by spin-inoculation at 110 g for 10 min at a multiplicity of infection (MOI, bacteria:cell) of 100:1. Then, microplates were placed at 37 °C and 5% CO₂ for 30 min. Infected cultures were washed 3 × with phosphate-buffered saline (PBS) pH 7.4 and incubated with fresh DMEM containing 10% FBS and 50 mg/mL gentamicin at 37 °C and 5% CO₂ for 3 or 4 h. In order to obtain more instances of S. flexneri associated to SEPT7 in Figure 4H, one of the datasets was obtained after infection with S. flexneri expressing the adhesin AfaI (Labigne-Roussel & Falkow, 1988) and GFP to increase host cell invasion at a MOI of 10 in the absence of spin inoculation.

Click chemistry

To label de novo DNA and protein synthesis, Click-iT™ EdU Cell Proliferation Kit for Imaging (#C10337 Thermofisher Scientific) and Click-iT™ AHA Protein Synthesis HCS Assay (#C10289 Thermofisher Scientific) were used, respectively, according to the manufacturer instructions. In the case of AHA, cells were cultured in glutamine, cystine and methionine-free DMEM (#21013024 Thermofisher Scientific) for 1 h prior to the treatment. 10μM EdU and 25μM AHA were supplemented to live bacterial or mammalian cell cultures for 1h. Cells were fixed for immunofluorescence as indicated below. Click azide/alkyne reactions were performed after permeabilization and prior to antibody staining as indicated below, according to the manufacturer’s instructions.

Immunofluorescence and fluorescence microscopy

Bacteria, coverslips or 96-well microplates containing adherent infected or uninfected cells were washed three times with PBS pH 7.4 and fixed 15 min in 4% paraformaldehyde (in PBS) at RT. Fixed cells were washed three times with PBS pH 7.4 and subsequently permeabilized 5min with 0.1% Triton X-100 (in PBS). Cells were then washed three times in PBS and incubated 1h with primary antibodies diluted in PBS supplemented with 0.1% Triton X-100 and 1% bovine serum albumin. Cells were then washed three times in PBS and incubated 45 min with anti-rabbit secondary antibodies diluted 0.1% Triton X-100 (in PBS), and Hoestch and Alexa-conjugated phalloidin where indicated. 96-well microplates were washed three times in PBS and preserved in 0.01% sodium azide (#10338380 Thermo Scientific) in PBS pH 7.4. Stained bacteria cultures and coverslips were washed three times with PBS pH 7.4 mounted on glass slides with ProLong™ Diamond Antifade Mountant (#P36970I Invitrogen). Fluorescence microscopy on stained bacteria was performed using a using a 63×/1.4 C-Plan Apo oil immersion lens on a Zeiss LSM 880 confocal microscope driven by ZEN Blacksoftware (v2.3). Microscopy images were obtained using z-stack image series taking 8–16 slices. Fluorescence microscopy on infected or uninfected cells was performed using a ZEISS Plan-APOCHROMAT 20× / 0.95 Autocorr Objective or a ZEISS Plan-APOCHROMAT 50×/1.2 water immersion lens coupled to a 0.5x tubelens on a Zeiss CellDiscoverer 7 with Airyscan detectors driven by ZEN Blue software (v3.5). Microscopy images were obtained using z-stack image series taking 32 slices. Confocal images were processed using Airyscan processing (Weinerfilter) using “Auto Filter” and “3D Processing” options.

Flow cytometry

5 x 10⁴ individual bacterial cells were analysed using flow cytometry with an LSRII flow cytometer (BD Biosciences). The data was analysed using FlowJo software, version 10.7.1. The median fluorescence of the bacterial population was determined and plotted for each staining.

Analysis pipelines

Number of bacteria per cell, as well as fluorescence of bacteria, nuclei and mammalian cells was analysed using Cellprofiler (v.4.0.7) (Carpenter et al., 2006). Custom made segmentation and deep learning pipelines for the classification of septin cages were written in ImageJ and Python using Keras.

Statistics

All graphs were plotted and statistical analysis were performed using Prism. n.s: non-significant, *: p-value < 0.05, **: p-value < 0.01, ***: p-value < 0.001).

Data availability

Due to the size of the data, representative datasets were deposited in Dryad (https://doi.org/10.5061/dryad.6wwpzgn5z). Materials and additional raw microscopy data can be obtained from the corresponding authors upon requests.

Code availability

Custom ImageJ macros, phyton scripts and Jupyter notebooks were annotated and deposited in Github (https://github.com/ATLopezJimenez/Toolset-high-content-analysis-of-Shigella-infection).

Acknowledgements

We thank members of the Mostowy and Brian Ho labs for helpful discussion. We thank François-Xavier Campbell-Valois for providing TSAR plasmids. We thank the instructors of the EMBL course Deep Learning for Image Analysis, in particular Prof. Pejman Rasti. A.T.L.J. was funded by the Swiss National Science Foundation Early Postdoc.Mobility Fellowship (P2GEP3_188277) and the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska - Curie grant agreement no. H2020-MSCA-IF-2020-895330. D.B. was supported by the Deutsche Forschungsgemeinschaft (DFG) Walter Benjamin Programme (BR 6637/1-1). G.Ö.G. is funded by the Human Frontier Science Program (LT000436/2021-L). Research in S. Mostowy laboratory is supported by a Wellcome Trust Senior Research Fellowship (206444/Z/17/Z) and European Research Council Consolidator Grant (772853 - ENTRAPMENT).

Supplementary Figure Legends

Figure 1-figure supplement 1.

Figure 2-figure supplement 1.

Figure 4-figure supplement 1.

Figure 5-figure supplement 1.

Figure 6-figure supplement 1.

Figure 7-figure supplement 1.