Imaging the Architecture of Granulomas Induced by Mycobacterium tuberculosis Infection with Single-molecule Fluorescence In Situ Hybridization

Tuberculosis has historically been, and still is, a major cause of death from infectious diseases in many parts of the world. Long-recognized hallmarks of tuberculosis pathology are granulomas in the lung, which form around the sites of initial infection by Mycobacterium tuberculosis. They form when tissue-resident macrophages phagocytize the pathogen, while simultaneously recruiting other immune cells, including dendritic cells, monocytes, neutrophils, NK cells, and T and B cells to the site of infection via the release of chemokines and cytokines (1–3). Granulomas minimize damage to larger lung tissue by locally restricting both bacterial proliferation and inflammatory immune responses (1–3).

Although signatures of the immune response to M. tuberculosis are present in the peripheral blood as memory T cells (4, 5) and as transcriptomic markers (6), the most intimate and long-term contact between M. tuberculosis and the host immune system occurs within the confines of the granuloma (3). It is therefore informative to study interactions between the pathogen and the immune system through studies of the functional architecture of granulomas.

Because human granulomas are not readily accessible, several model animals, such as nonhuman primates, rabbits, mice, and zebrafish, are often used in these studies (1, 2, 7). These animal models simulate human disease to different extents and have different experimental benefits and limitations (1, 2, 7). M. tuberculosis infection in the nonhuman primate, macaque, closely mimics certain features of human disease progression, including latency and a granuloma architecture that is characterized by a necrotic core (1, 2, 7). Rabbits also form necrotizing granulomas similar to those found in human tuberculosis (7), whereas M. tuberculosis infection in common mice strains, such as C57BL/6 or BALB/c, generally does not develop necrotic lesions. However, in a number of situations, such as in mouse strain C3HeB/J, or in common mouse strains upon deletion of the NO synthase gene, or upon overexpression of cytokine IL-13, M. tuberculosis infection is reported to produce necrotizing granulomas (8–11).

Granuloma architecture has traditionally been studied using H&E staining and by immunofluorescence (12–14), although recently, other approaches such as multiplexed ion beam imaging by time of flight (MIBI-TOF) and in situ sequencing have been reported (9, 15). However, very few Abs are available for the proteins of model animals such as macaques and rabbits. An additional limitation of immunofluorescence is that secreted markers such as cytokines cannot be imaged in situ except in ex vivo settings where brefeldin A is used to prevent cytokine secretion. An attractive alternative is to image cells through fluorescence in situ hybridization (FISH) against expressed mRNAs. In a FISH variant called single-molecule FISH (smFISH), multiple small oligonucleotide probes are used to visualize the same mRNA, which allows for highly specific and sensitive detection of mRNA molecules (16, 17) (Fig. 1A). In smFISH, each molecule of target RNA becomes visible as a fine fluorescent spot when viewed under high magnification.

In this study, we demonstrate that the functional architecture of tuberculosis granulomas can be studied in high resolution using smFISH and its variants. We imaged mRNAs encoding a number of immunological markers in rabbit and macaque M. tuberculosis granulomas present in archived lung tissues. By combining smFISH with immunofluorescence we could image both mRNA and protein markers in the same sections. For mRNA markers that are expressed at very low levels, such as host cell surface receptors, we performed amplified smFISH (ampFISH), in which pairs of interacting hairpin probes were tiled across the length of the mRNA, where each pair created amplified signals via a hybridization chain reaction (HCR). Finally, we were able to image as many as eight different mRNA targets in the same tissue section, by performing iterative rounds of smFISH in which different targets are interrogated in successive cycles of hybridization, imaging, and signal removal. Using quantitative measures of the expression of different markers in the individual cells, we identified various cell types and charted their distributions in these complex tissues. These studies pave the way for transcriptome-based analysis of M. tuberculosis granuloma architecture, where the RNA markers identified from readily accessible RNA sequencing (RNA-seq) data can be used to study the distribution of different cell types in model animals without being hampered by the limitations of immunofluorescence.

New Zealand female white rabbits (Oryctolagus cuniculus) were infected with M. tuberculosis HN878 through a “snout-only aerosol-exposure system,” as described previously (18). This model has been demonstrated to produce progressive active tuberculosis with caseating necrotic and cavitary granulomas (18). At 4–12 wk postinfection, rabbits were euthanized, and their lungs were collected. Portions of the lungs were placed in 10% buffered formalin and kept in the refrigerator for at least 1 wk. The formalin-fixed lung tissue samples were then paraffin embedded and stored. These archived paraffin-embedded blocks were cut into 5-μm sections and used for FISH and immunofluorescence studies. The archived macaque tissues were contributed by D.K. from his previous studies.

For smFISH, 48 oligodeoxynucleotide probes were designed for each mRNA using the Stellaris RNA FISH probe designer tool (available at https://www.biosearchtech.com/support/tools/design-software/stellaris-probe-designer). All probe sequences and transcript IDs are provided in Supplemental Table I. These oligonucleotides were synthesized with 3′-amino groups, pooled in equimolar concentrations, coupled to aminoreactive fluorophores, and then purified by HPLC as described before (19). Primary probes for iterative smFISH did not possess amino groups and were obtained as a pool that contained equimolar amounts of each of oligonucleotide purchased from Integrated DNA Technologies (Coralville, IA) (Supplemental Table I). Readout probes were obtained with 3′-amino moieties and were coupled to either Texas Red or Cy5 and purified by HPLC in the same manner as the smFISH probes. The ampFISH acceptor probes for the mRNA encoding the immune recognition receptor protein CD3ε were prepared by click chemistry as described by Marras et al. (20).

To deparaffinize and hydrate the formaldehyde-fixed and paraffin-embedded (FFPE) tissue sections on slides, the slides were serially incubated for 10 min at room temperature in xylene (twice), 100% ethanol, 90% ethanol, 70% ethanol, and, finally, either in hybridization wash buffer (10% formamide in 2× saline-sodium citrate [SSC] buffer) or in Ag retrieval buffer (10 mM sodium citrate solution and 0.05% Tween 20 [pH 6.0]). For Ag retrieval, the sections were incubated in the Ag retrieval buffer in a decloaking chamber at 90°C for 40 min, followed by a slow return to room temperature. In some experiments, a proteinase K treatment was carried out after this step to increase the permeability of the crosslinked matrix. This consisted of equilibrating the section with proteinase K buffer (10 mM Tris-HCl, 0.5% SDS, 50 mM EDTA [pH 8.0]) and then incubating with proteinase K (Thermo Fisher Scientific, Waltham, MA) dissolved in this buffer at 0.2 μg/ml for 5 min at room temperature, followed by several washes in the same buffer.

Immunofluorescence-based imaging was performed by first blocking the section with RNase-free BSA (Ambion BSA [Thermo Fisher Scientific], 5 mg/ml in PBS) for 30 min, followed by incubation with an ionized calcium-binding adapter (IBA1)-specific primary Ab, ab5076 (Abcam, Waltham, MA), at 1:250 dilution. The slides were washed with PBS and incubated with an Alexa Fluor 488–labeled secondary Ab, ab150129 (Abcam), at 1:1000 dilution for 1 h. IBA1-negative control immunofluorescence experiments were carried out using a goat IgG isotype Ab, WH23-51, obtained from Abraham Pinter at the Public Health Research Institute (Newark, NJ). Hypoxia-inducible factor-1α (HIF-1α) protein was detected using an anti–HIF-1α Ab directly conjugated to DyLight 650 (MA5-16008, Thermo Fisher Scientific) for 2 h. All immunofluorescence steps were carried out at room temperature. In experiments where both protein and mRNAs were detected, immunofluorescence was performed first, followed by a second formaldehyde fixation (4% formaldehyde in 1× PBS) for 5 min.

For smFISH, the sections were equilibrated with hybridization wash buffer and then incubated overnight in a 37°C water bath with sufficient hybridization solution to cover the section. The hybridization solution contained 10% dextran sulfate (Sigma-Aldrich, St. Louis, MO), 1 mg/ml Escherichia coli tRNA (Sigma-Aldrich), 2 mM ribonucleoside vanadyl complex (New England Biolabs, Ipswich, MA), 0.02% RNase-free BSA (Thermo Fisher Scientific), 10% formamide, and 500 ng/ml for each probe set. After hybridization, the slides were washed twice with hybridization wash buffer at room temperature and then mounted with 0.17-mm-thick coverslips in oxygen-depleted mounting medium (0.4% w/v glucose, 2× SSC, 37 μg/ml glucose oxidase, 1% v/v catalase suspension [both from Sigma-Aldrich], and 1 μg/ml DAPI [Thermo Fisher Scientific]).

When TrueBlack (Biotium, Fremont, CA) was employed to suppress autofluorescence, it was used after immunofluorescence and or hybridization. TrueBlack was diluted at 1× concentration in 70% alcohol (in water) and applied to the section for 30 s. The sections were immediately washed using 2× SSC and 0.4% w/v glucose, mounted, and imaged the same day.

Images were acquired using an Axiovert 200M widefield epifluorescence microscope (Zeiss, Oberkochen, Germany) under the control of MetaMorph software (Molecular Devices, San Jose, CA) with a Prime sCMOS camera (Teledyne Photometrics, Tucson, AZ). A Lumen 200 metal arc lamp (Prior Scientific, Rockland, MA) was used as the excitation light source in combination with the following filter cubes. For each channel the excitation filter, the dichroic mirror, and the emission filter used in the filter cubes were, respectively, as follows: DAPI, 330wb80, 400dclp, 450DF65; fluorescein, 475AF20, xf2077, 530DF30; rhodamine, 546DF10, 555DRLP, 580DF30; Texas Red, 590DF10, 610DRLP, 630DF30; and Cy5, 655AF50, 692DRLP, 710AF40. The filters were obtained from Omega Optical (Brattleboro, VT) and Chroma Technologies (Bellows Falls, VT).

The granulomas in tissue were identified by visual examination in the DAPI channel by their characteristic morphology, in which compressed and aggregated nuclei are surrounded by healthy alveolar tissue, and they were then imaged in the surrounding tissue using a ×20 objective (0.75 numerical aperture) using the scan function in the MetaMorph software or manually. Higher magnification z-stack images were acquired using a ×63 objective (1.3 numerical aperture). When imaging under high magnification, multiple z-sections separated by 0.2 μm were obtained and then compressed into maximum intensity merges. This ensured that all diffraction-limited spots corresponding to single mRNA molecules present in the volume of cells were accounted for (17).

The images were generally acquired in the differential interference contrast, DAPI, rhodamine, Texas Red, and Cy5 channels. No probe or Ab was used in the rhodamine channel, which served as a control for the detection of autofluorescence.

Iterative hybridization and imaging were performed in an open dish with a coverslip bottom. FFPE tissues were sectioned on a 25-mm coverslip instead of normal slides, and the relatively harsher initial steps of deparaffinization, hydration, and Ag retrieval were then carried out on the free-standing coverslips using coverslip racks, and then each coverslip was placed into an open-top imaging chamber (Warner Instruments, Holliston, MA). A pool of 358 primary probes (1.4 nM of each) was then hybridized to the tissue section on each coverslip overnight at 37°C in 300 μl of hybridization buffer. Excess probes were then removed by washing twice with hybridization wash buffer and then the imaging chamber along with the coverslip was placed on the microscope stage.

With the imaging chamber on the microscope stage, four cycles of the following steps were performed, where in each step 300 μl of different solutions was added and each step was preceded by the removal of the previous solution. After each cycle, the readout probe pair in step 2 below was changed. Step 1: equilibrate with readout probe hybridization solution (10% w/v ethylene carbonate, 10% dextran sulfate, 4× SSC) for 2 min. Step 2: hybridize with a pair of readout probes, one labeled with Texas Red and the other labeled with Cy5 (12 nM each) and dissolved in the readout probe hybridization solution (sequences provided in Supplemental Table I) for 15 min. Step 3: wash twice with hybridization wash buffer supplemented with 0.1% Tween 20 (Thermo Fisher Scientific). Step 4: equilibrate with glucose buffer (0.4% w/v glucose, 2× SSC) for 2 min. Step 5: equilibrate with oxygen-depleted mounting medium and image.

For cell segmentation we used images from the DAPI channel (Fig. 2), IBA1 staining (Fig. 3), maximum intensity merges of HIF-1α smFISH and immunofluorescence images (Fig. 4), and maximum intensity merges of IFN-γ smFISH and CD3ε ampFISH images (Fig. 5). Before the cell segmentation, the scanned images were stitched together and z-sections were merged. The “segmentation layers” mentioned above were incorporated into a multilayer image file (three layers for Fig. 2 and four layers for Figs. 4 and 5), which were then imported into a QuPath project (27). After specifying the entire image as the annotation area, custom scripts were used to call StarDist (26) or Cellpose (21) to find nuclear or cellular boundaries. Examples of the quality of segmentations are shown in Figs. 2, 4, and 5. The parameters were adjusted to achieve accurate segmentation of most objects while minimizing spurious segmentation. After the detection of all cells, the images were classified using average pixel intensity within the area of cells for each of the individual measurement channels, as described in Supplemental Fig. 2. From these single measurement classifiers, composite classifiers were created, which identified cells expressing combinations of markers. The composite images and vector graphic overlays were exported into Adobe Illustrator, where additional annotations and graphs were added, creating the final figures. The cell segmentations for the identification of IBA1-positive cells for cell density measurements reported in Fig. 3 and Supplemental Fig. 3 were performed with the Cellpose algorithm (21). Although this algorithm resulted in undercounting of IBA1-positive cells, it could clearly distinguish between the cell densities in the core and the periphery of different granulomas. The density measurements were performed using QuPath’s density map function on the segmented and classified objects.

There are two major challenges in the analysis of granulomas by smFISH. First, smFISH is usually performed with high-magnification objectives (×63 or ×100) that yield diffraction-limited spots for single mRNA molecules (17). However, to obtain a good overview of the distributions of different cell types in the granuloma, the imaging needs to be performed at lower magnification, which cannot resolve single-molecule spots, and has lower sensitivity. To address this issue, we imaged large regions by scanning granuloma sections using a ×20 objective with a high numerical aperture and then stitched these images together.

Second, the lung tissues exhibit high levels of autofluorescence in some cells, which obscures relatively weak smFISH signals. Autofluorescence is strongest in the green channels and diminishes at longer wavelengths. To address this issue, we avoided the fluorescein channel and the rhodamine channel for smFISH. Instead, the fluorescein channel was used for Ab staining, which yields stronger signals compared with smFISH, and the rhodamine channel was used for imaging autofluorescence. A second approach was to use TrueBlack, which was effective in reducing the autofluorescence significantly without negatively impacting the signals from the smFISH probes (Supplemental Fig. 1A).

We sought to perform smFISH in combination with traditional immunofluorescence, as the latter can be used as a benchmark and to provide valuable controls. Given the pathogenicity of M. tuberculosis, it is necessary to formaldehyde fix the infected lung specimens under rather harsh conditions (10% formaldehyde for at least 7 d). These conditions lead to sequestration of protein epitopes as well as of RNAs, and it is necessary to treat the FFPE sections with harsh steps, including boiling for Ag retrieval, which can lead to loss of RNA. We optimized this process using stringent RNase-free conditions that allow for Ag retrieval while preserving the RNA. We then performed immunofluorescence against ionized calcium-binding adapter (IBA1), which is a protein marker for tissue-resident macrophages (22, 23) using an Ab, fixed the section again with 4% formaldehyde to ensure that the Ab remained bound, and we then performed smFISH using a set of Cy5-labeled probes (48 oligonucleotides) that were complementary to the cytokine mRNA encoding rabbit IFN-γ, which is a proinflammatory marker. The resulting images demonstrate the simultaneous detection of cells expressing a protein and an mRNA marker (Fig. 1B). Although the well-segregated single-molecule diffraction-limited spots characteristic of smFISH were not visible in low-magnification images, many such spots were visible in high-resolution images, particularly in cells that had lower levels of IFN-γ mRNA expression (and thus less crowding) (Fig. 1B). These images are consistent with the expectation that some cells (macrophages) in the granuloma will express IBA1 alone, some will express IFN-γ alone (T cells), and some will express both (activated macrophages). Alveolar macrophages are known to produce IFN-γ avidly upon stimulation in vitro (23) and have also been reported to do so in human granulomas (24). Although the specificity of the anti-IBA1 Ab has been documented earlier (23), we used an isotype control Ab as a negative control in parallel imaging of a second granuloma, which yielded no signals (Supplemental Fig. 1B).

We then imaged the expression of three other cytokine mRNAs, IL-10, IL-12, and IL-2, and found that many cells robustly express them within rabbit granulomas (Supplemental Fig. 1C). Several observations provided evidence of the specificity of the signals: 1) as in the case of IFN-γ mRNA above, high magnification of the cells expressing these cytokines displayed diffraction limited spots; 2) only a fraction of the cells in the granulomas are seen to express the mRNAs; 3) the cells expressing these cytokines were generally missing in the uninvolved healthy area around the granulomas (Supplemental Fig. 1D); and 4) to show that the signals emanate from the binding of probes to their specific targets, we used a “no-target” probe set as a control, which did not yield any signals (Supplemental Fig. 1E). This no-target probe set was designed to hybridize to mRNA encoding GFP, which was expected to be absent in rabbit lung tissue. These data, together with previously reviewed evidence of the specificity of the smFISH probe approach (25), strongly support the conclusion that the RNA signals in the granuloma arise specifically from the targeted mRNAs.

To demonstrate the detection of cell types by combinatorial expression of two mRNA markers and one protein marker, we included probes against TNF mRNA (another proinflammatory marker, formerly known as TNF-α) that were labeled with Texas Red in the experiment discussed above. A primary image showing the expression of the IBA1 protein, IFN-γ, and TNF mRNAs is shown in Fig. 2A.

To quantify the expression of each of the three markers in single cells within a section, we first computationally defined the locations and boundaries of cells in the section using DAPI staining of the nuclei (not shown in Fig. 2A). The location and boundaries of DAPI-stained nuclei were determined using a machine-learning algorithm, StarDist, that can accurately identify nuclei and demarcate their boundaries even in crowded spaces (26). The coordinates of the segmented nuclei were used to determine the approximate boundaries of cells around each nucleus by expanding the nuclear boundaries until they met with the cell boundaries of neighboring nuclei using QuPath (27). The accuracy of this method is depicted in Fig. 2B (upper left panel). We then measured the cell intensity (mean pixel intensity within each cell boundary) in each of the three channels corresponding to the immunofluorescence of IBA1 and the smFISH signals from TNF mRNA and IFN-γ mRNA. Thereafter, we classified cells into three primary classes, IBA1, TNF, and IFN-γ, based on the expression of individual markers above certain thresholds (Fig. 2B). The thresholds were chosen interactively using cell-intensity histograms in QuPath, such that most cells that can be visually identified as expressing the marker would be included in the class, and most cells not expressing the marker would be excluded from the class (Supplemental Fig. 2).

After identifying the cells that expressed the three markers individually, we identified those cells that expressed them in various combinations. A digital map of this section was created by presenting each cell category using a distinct color. This map was overlaid on the primary image (Fig. 2C). The number of cells in each category that we found in the section is presented in Fig. 2D along with the color keys. By interactively adjusting the opacity of the digital map, we could confirm that this map faithfully represented the underlying primary image for each of the three fluorescence channels. An examination of closeups of several different regions of this digital map and the underlying image enlargements confirmed the accuracy of the cell type assignments in the digital map (Supplemental Fig. 2C).

The expression of various marker combinations indicates that the cells expressing just IBA1 are likely to be naive macrophages; cells expressing IBA1 in combination with either or both cytokines are activated macrophages; cells not expressing IBA1 but expressing either or both cytokines are activated T cells; and cells expressing none of the markers are lung epithelial cells.

To define the microenvironments within a granuloma where different cell types dominate, we created plots of the density of different cell types and overlaid them onto the digital cell map (Fig. 2E). These plots indicated that the solid core of this granuloma (which is a relatively young granuloma at 4 wk postinfection) is largely populated by activated macrophages and some activated T cells, whereas its periphery contains activated T cells, naive macrophages, and lung epithelial cells. The multifunctional cells that express all three markers are restricted to the core of the granuloma. Similar microenvironmental features have recently been observed by MIBI-TOF imaging of human granulomas (15).

To demonstrate the efficacy of these techniques in distinguishing different types of granulomas, we imaged and analyzed a few morphologically diverse rabbit granulomas with respect to IBA1 protein and IFN-γ mRNA (Fig. 3, Supplemental Fig. 3A–C). We identified IBA1-positive cells using the Cellpose algorithm that uses cell surface markers for cell segmentation (21), determined the expression of IFN-γ mRNA in each identified cell, and then classified them into two categories, IBA1⁺IFN-γ⁺ and IBA1⁺IFN-γ⁻. Small solid granulomas typically had a layer of compacted cells that segregated the core of granulomas from their periphery. This layer served as the boundary of solid granulomas. The cores of necrotic granulomas were acellular due to cell death and were identified using amorphous DAPI signals. The outer boundaries of necrotic granulomas were identified by high cellular compaction compared with surrounding healthy lung tissue. These boundaries are indicated by broken white lines in the images (Fig. 3, Supplemental Fig. 3A–C).

The densities of macrophages in the core and the periphery of the younger solid granulomas were very similar (Fig. 3, Supplemental Fig. 3A). In contrast, in the core of more mature necrotic granulomas there was a near complete absence of macrophages (and that of other recognizable cells due to cell death). In the periphery of the necrotic core, the macrophage density was higher and similar to that of solid granulomas. The macrophages inside the cores of solid granulomas were dominated by IBA1⁺IFN-γ⁺ cells (91–96%) whereas in the region outside of the core their percentage was lower (51–64%). The area immediately surrounding the necrotic core in large granulomas often appeared to be composed of clusters of several “mini” solid granulomas (Fig. 3A). These mini granulomas were identified computationally by finding regions of high density of IBA1⁺IFN-γ⁺ cells (Fig. 3, Supplemental Fig. 3C). The mini granulomas are likely to be secondary granulomas that emerge after the original ones are necrotized.

These studies show that the cell density measurements of macrophages expressing cytokine mRNAs can be used to distinguish granulomas in different stages of development.

Over the years, mRNAs have been extensively used as surrogates for proteins in many RT-PCR, microarray, and RNA-seq studies. However, stabilization or decay of the proteins after translation (28), and stochasticity in mRNA synthesis (29), can create a discordance between mRNA and protein amounts in single cells. A particularly interesting case is HIF-1α, which is an oxygen-sensing transcription factor that is degraded under normal oxygen concentrations, but becomes stable under hypoxic conditions, such as those occurring during microbial infections (30). HIF-1α serves as a key metabolic regulator of immune cell activation. In addition to stabilizing the HIF-1α protein, microbial infections, including M. tuberculosis infections (23), stimulate HIF-1α mRNA synthesis by producing transcription factor NF-κB, which in turn induces transcription from the HIF-1α gene (31–34). We explored the relative abundance of HIF-1α mRNA and the HIF-1α protein within granulomas by imaging both simultaneously.

By imaging HIF-1α protein with an Ab and by imaging the HIF-1α mRNA with smFISH probes, we found that ∼10% of the cells in rabbit granulomas express HIF-1α (Fig. 4). As expected in the high-magnification imaging (×63 objective), smFISH signals were visible as discrete spots, whereas the protein signals were more diffused. To explore the relative abundance of HIF-1α mRNAs and HIF-1α proteins in these cells, we segmented the cells using the combined sum of the fluorescence in the protein and mRNA channels, rather than the DAPI fluorescence, as was done in Fig. 2, and then determined the average pixel fluorescence intensity within the cell boundaries in each of the two channels. We found that out of 148 cells that we analyzed from three fields within one granuloma, 31 expressed only HIF-1α protein, 5 expressed only HIF-1α mRNA, and 111 expressed both (Fig. 4B). The mRNA and protein fluorescence signals in the cells that expressed both were correlated (correlation coefficient 0.66). The frequency of HIF-1α–expressing cells varied from field to field within the same granuloma, but the mRNA and protein correlation remained about the same. Because in M. tuberculosis–infected cells HIF-1α mRNA synthesis is stimulated and HIF-1α protein is stabilized (23, 31, 32), the correlation of the HIF-1α mRNA and the HIF-1α protein in these cells is indicative of the presence of M. tuberculosis at the site, and also supports the supposition that mRNAs can serve as reasonable surrogates for proteins.

In the forgoing experiments, cytokine mRNAs could be detected readily using smFISH, because they are abundantly expressed in activated immune cells. However, we found that mRNAs encoding surface receptors, the traditional phenotypic markers of lymphocytes, yielded rather faint signals when low magnification was used. Therefore, we used ampFISH to demonstrate the detection of mRNAs encoding CD3ε (which is the ε-chain of CD3) that serves as a marker of T cells, as it is a component of the TCR complex. In ampFISH, a pair of hairpin probes is used. Their binding to adjacent locations on a target mRNA sequence engenders a conformational reorganization in one of the two probes (Fig. 5A) (20). This reorganization reveals a previously sequestered sequence that initiates a HCR, depositing multiple copies of fluorescent labels at the site. This signal amplification can be further increased by tiling multiple pairs of ampFISH probes across the length of a target mRNA (Fig. 5B).

We designed 22 pairs of ampFISH probes complementary to the mRNA encoding CD3ε, and we used them in combination with a set of 48 traditional smFISH probes complementary to IFN-γ mRNA. For these experiments, we imaged an archived necrotic granuloma obtained from a rhesus macaque infected with M. tuberculosis (14). As mentioned before, these granulomas mimic human disease very closely. After hybridization with all of the probes in a common hybridization mixture, followed by removal of the excess probes by washing, HCR was performed to create ampFISH signals in a fluorescence channel that was distinct from the fluorescence channel used for detecting the smFISH probe labels. Color-combined images for the IFN-γ smFISH signals and the CD3ε ampFISH signals are presented in Fig. 5.

These images show that single cells that express the two markers are scattered around the periphery of the central necrotic region of the granulomas (Fig. 5C). High-magnification images of the cells expressing both markers show that the signals from both markers are spot-like and present in only a small subset of the cells (Supplemental Fig. 3D). Inside the necrotic region, cells are present in various degrees of cellular degradation, as indicated by the absence of discrete nuclei in DAPI images (not shown) and by a loss of cellular morphology. Therefore, we did not use DAPI signals for the identification of cell boundaries as we did in the earlier experiment. Instead, we used the sum of the smFISH signals and the ampFISH signals as the indicator of cell segmentation. After identifying all of the cells that expressed these markers, we determined the fluorescence intensity within each cell in each channel and classified each cell as being either IFN-γ⁺, CD3ε⁺, or both IFN-γ⁺ and CD3ε⁺, using QuPath. These classified cells are indicated by colored circles (which in this case are larger in size than the cells) that are overlaid on the primary image in Fig. 4C, and the distribution of their fluorescence is depicted in Fig. 4D. An enlargement of a region of the granuloma, indicated by the blue square in Fig. 4C, is also presented to demonstrate the accuracy of this cell classification. In these enlargements, images from each of the two channels are presented in separate grayscale images with overlays indicating the cell classifications. This colocalization analysis reveals that a small fraction of T cells (CD3ε⁺) also express IFN-γ, indicating that they are in a state of activation. We found that only ∼4% of the CD3ε⁺ cells were expressing IFN-γ (Supplemental Fig. 3H), which agrees with earlier flow cytometry data from dissociated cells obtained from macaque granulomas (35). Those cells that express IFN-γ but do not express CD3ε are likely to be predominantly activated macrophages. A similar pattern was apparent in several other macaque granulomas of different sizes that we analyzed (Supplemental Fig. 3D–H).

Using probe sets labeled with distinguishable fluorophores, it is possible to image up to four mRNAs simultaneously in cultured cells. However, the autofluorescence in lung tissues limits this approach to only two or three clearly distinguishable fluorophores. A powerful strategy to achieve higher levels of multiplexing is to perform iterative hybridization, in which multiple mRNAs are interrogated in successive cycles of binding target-specific probes, imaging, and then removal of the probes. Using this approach, 33 mRNA species have successfully been imaged in brain sections (36). Two groups have achieved dramatically higher levels of multiplexing (100–1000 s of target mRNAs) by using pools of probes for all target mRNAs in a common hybridization reaction (MERFISH [37] and seqFISH [38]). These “primary target-binding probes” are themselves not labeled with dyes but contain address-tag sequences that in turn can be detected using secondary “readout probes” (Fig. 6A). All target mRNAs can be detected by performing multiple cycles of imaging, where, in each cycle, several readout probes labeled with distinguishable fluorophores are bound, followed by imaging the probes, and then by dissociation of the probes from their targets (Fig. 6A). High depths of multiplexing are achieved by combining multiple rounds of smFISH with color coding of the probes, where each target is identified by a combination of colors, rather than by a unique color (37, 38). Given the limitations inherent in M. tuberculosis granuloma imaging (high autofluorescence, heavy crosslinking, age of archived samples, and low magnification), we used a simplified version of iterative smFISH with a modest level of multiplexing, where pooled primary probes were used and multiple rounds of smFISH were performed, but implementing it without color coding.

We prepared a pool of probes against eight different rabbit mRNAs that were likely to be expressed by myeloid and lymphoid cells in rabbit granulomas: NOS2, CD163, CD3, IL-10, GZMH, S100A8/9, IFN-γ, and ARG1. The probe pool consisted of a total of 359 probes with 36–47 probes being complementary to one mRNA species (Supplemental Table I). All probes for each mRNA species contained an address tag sequence appended to each end. The address tags were identical for the same mRNA species but were distinct from the address tags of the other mRNA species. The address tags do not bind to the mRNA and remain single stranded when the target-specific region of the probe is bound to the mRNA. The presence of the address tag at the target could be determined by using fluorescently labeled readout probes complementary to the address tags (Fig. 6).

To ensure that the primary probes remain bound while repeated cycles of readout probe binding and removal are carried out, we designed relatively long (22- to 25-nt-long) target-specific regions in the primary probes, and we used 20-nt-long low-melting temperature readout probes. The primary probes form RNA-DNA hybrids, and the readout probes form DNA-DNA hybrids that are inherently weaker. We included 10% formamide in the medium used to form the primary probe hybrids, and we included 30% formamide in the medium to remove the readout probes. The 30% formamide destabilizes the weak hybrids formed between the readout probes and the address tags, but not between the primary probes and the target mRNAs. To ensure that the readout probes would be able to bind rapidly (within 15 min), we used only three types of nucleotides in the design of the probes, and we used an ethylene carbonate buffer for readout probe binding. Each of these techniques independently accelerates hybridization reactions (39–41). The readout probes were designed to detect only their cognate address tags and to not cross-react with the other address tags. This feature of the readout probe set was confirmed by in vitro fluorescence DNA melting assays performed before their deployment in situ.

The pool of mRNA-specific primary probes was hybridized to the tissue sections in an overnight hybridization reaction. After removing the excess primary probes, we interrogated the eight target mRNAs in four successive rounds of hybridization and imaging with pairs of labeled readout probes. In each round, one Texas Red–labeled readout probe complementary to the address tag of the one mRNA and another Cy5-labeled readout probe that was specific to the address tag of the second mRNA were hybridized to the sample. After imaging the first pair of mRNA targets in each channel, the first pair of readout probes was removed and the readout probes for the second set of mRNA targets were added. Four rounds of hybridization with a total of eight readout probes yielded images for all eight mRNA targets (Fig. 6B–G).

Each round of imaging lit up a specific set of cells in the granuloma, with most cells exhibiting little or no background. We found that many cells expressed multiple markers, although the relative expression levels of different markers varied from cell to cell. High-magnification imaging of a portion of the granuloma in round 4, which imaged IFN-γ and ARG1 mRNAs, exhibited spot-like signals that are characteristic of smFISH. This image shows that even though many cells express both markers, some cells exhibit only IFN-γ signals (Fig. 6D). This observation points to the specificity of each probe set in the pair and illustrates the diversity of cell types within this milieu.

A current limitation of our imaging apparatus is that there is a significant spatial shift in the tissue between cycles and the same cells cannot be reliably located in different cycles. This problem arises due to the requirement of scanning the tissue and then stitching the scanned tiles into a larger image. However, the expression of two markers that are imaged in each round can be localized with high precision within the same cells. An analysis of the expression of pairs of markers that were imaged together in the same cycles is presented in Supplemental Fig. 4. This analysis shows that the pattern of expression of each pair changes from one round of hybridization to the next. To show that the readout probes are removed while the primary probes remain bound in each round, we imaged the section after the third round without adding a readout probe set. The results of this experiment showed that the cellular fluorescence intensity is reduced to background levels after removal of the readout probes and becomes high again in the fourth round (Fig. 6F, 6G).

Finally, the open-dish imaging format was exploited to locate M. tuberculosis within the granuloma. We imaged M. tuberculosis using rhodamine-auramine staining in the same section after completion of the iterative FISH (Fig. 6E). The M. tuberculosis staining carried out in this manner was done after the mRNA imaging cycles because the tissue is damaged in the process.

We demonstrated the imaging of many mRNA species expressed in immune cells in archived tuberculosis lung granulomas obtained from rabbits and macaques. These results indicate that despite the challenges of tissue-intrinsic autofluorescence, extensive formaldehyde crosslinking, and the age of the specimens, sufficient mRNAs survive in granulomas to enable their detection by smFISH. In the case of sparsely expressed mRNAs, we used a tiled ampFISH approach, which provided enhanced signals for overcoming the impact of autofluorescence and for obtaining detectable signals in low-magnification settings. Furthermore, it was possible to perform ampFISH in multiplex utilizing conventional smFISH.

Although the general specificity of smFISH and ampFISH has been established earlier (17, 20, 25), our results show that targeted mRNAs can accurately be detected in the context of granulomas. In particular, probe sets for an RNA that is not present in the tissue did not yield any signals, and when Hif-1α protein and Hif-1α mRNA were probed using an Ab and smFISH probes, the signals appeared in the same cells (Fig. 4). Furthermore, our finding that in macaque granulomas there are only a very small fraction of CD3ε⁺ cells expressing IFN-γ recapitulates the results of an earlier flow cytometric study of T cells dissociated from macaque granulomas (Fig. 4), supporting the specificity of smFISH signals.

Recently, several high-content analyses have been performed on M. tuberculosis granulomas that also align well with our observations. Gideon et al. (42) dissociated cells from macaque M. tuberculosis granulomas and performed single-cell RNA-seq on them, and McCaffrey et al. (15) labeled human and macaque M. tuberculosis granulomas with a panel of 36 metal-labeled Abs followed by MIBI-TOF imaging. Both studies show that diverse immune cell types make up the bulk of the cell mass in granulomas, with macrophages and T cells being the most dominant constituents. This MIBI-TOF study further showed that myeloid cells are concentrated in the core of the granuloma and lymphatic cells are concentrated in a cuff that surrounds the core. This organization concurs with our observations in solid rabbit granulomas (e.g., see Fig. 3). In another recent study, Sawyer et al. (43) performed multiplex immunofluorescence studies on human M. tuberculosis lung granulomas using Abs against the major immune lineage markers for macrophages, B cells, CD8⁺ T cells, and CD4⁺ T cells. They reported that human M. tuberculosis lesions contain numerous nonnecrotizing granulomas in addition to well-developed necrotizing granulomas. Our observation in necrotic rabbit granulomas, where multiple mini solid granulomas were seen to surround a large necrotic granuloma, are reminiscent of this organization (Fig. 3). A second observation made by the Sawyer et al. study was that the macrophages and the T cells are present in the cellular layer immediately surrounding the necrotic core, which is very similar to our observations in the macaque necrotic granulomas (Fig. 5, Supplemental Fig. 3D–H).

An important aspect of the present work is the digital identification of cells based on combinations of the markers that they express. We segmented cells using publicly accessible machine learning algorithms, StarDist and Cellpose, implemented in the QuPath environment. This digital identification of cells based on multiple cell markers allows unbiased quantitative morphometric analysis, such as microenvironment analysis through cell type density maps, which will allow explorations of cell-to-cell interactions that give rise to structures such as granulomas. This method can easily be adapted for use in situations where the number of markers and cell types is considerably larger, such as in spatial transcriptomic studies with commercial platforms and is applicable to spatial analysis in general.

Granulomas harbor a great variety of immune cells, including macrophages differentiated along a spectrum of proinflammatory and anti-inflammatory states, and T cells present in various stages of activation and differentiation (2, 44). Furthermore, the observations that macrophages, dendritic cells, and T cells undergo shifts in their metabolic program upon encountering pathogens (45–47) point to the additional diversity of cell states. Classical phenotypic surface markers are not sufficient to capture this diversity. Accordingly, four markers that could identify the major immune cell types in the Swayer et al. study mentioned above were not sufficient to adequately cluster and classify the full spectrum of human granulomas. Therefore, imaging platforms with higher levels of multiplexing are needed.

Although in principle our platform fulfills this need, we encountered a significant limitation during our multiplex imaging of eight mRNAs when we performed four rounds of iterative hybridization. We were not able to locate the same cells between successive rounds of hybridization because of mechanical shifts, which prevented us from determining the expression profiles of all eight genes in the same single cells. This issue can be addressed by high-precision registration of images obtained from successive rounds of hybridization by embedding fiduciary beads within the tissue matrix and using the locations of those beads for image registration (41). With this improvement, it will be possible to profile single cells in the granuloma for all imaged markers. The number of markers can be increased from 8 to ∼30 species of mRNA by performing 15 rounds of hybridization using a common epifluorescence microscope. A computational analysis of single-cell RNA-seq data shows that >90% of cell types can be identified by the expression of 30 genes or less (48). This throughput is still modest compared with 100s of mRNA species that recent spatial transcriptomic technologies are able to achieve (49); however, probe sets for model animals such as rabbits and macaques are not yet available on current commercial platforms.

In conclusion, with our platform it will be possible to obtain sets of suitable mRNA markers for cell type and subtype identification from single-cell RNA-seq data, which are readily available, and to then develop probe sets rapidly for them, and to use the resulting images to ascertain the distributions of those cells within a tissue. Furthermore, because this process can be applied to any species, not just to the model animals used in the current study, it will be possible to perform such analyses on human granulomas when they are available. Future studies of granuloma architecture that use these techniques will allow for a better understanding of how the tussle between the proinflammatory and anti-inflammatory activities of various cell types within granulomas can occasionally lead to sequestration of M. tuberculosis into long-term nonreplicating latency, whereas at other times they lead to dissemination of replicating bacteria into the larger tissue, causing active disease.

Rutgers University receives royalties from the sale of prelabeled smFISH probes by LGC Biosearch Technologies, which markets them as Stellaris probes. A portion of these proceeds is distributed to S.T.’s laboratory for research support and to him personally. These proceeds do not influence the conclusions of this research. Part of the research described in this article is described in patents filed by Rutgers University. The other authors have no financial conflicts of interest.

We thank Ryan Dikdan and Fred Russell Kramer for careful reading of the manuscript, Salvatore A.E. Marras for help in the preparation of probes, and Yuri Bushkin for stimulating discussions.