Detection of dendritic cell subsets in the tumor microenvironment by multiplex immunohistochemistry

Dendritic cells (DCs) are essential in antitumor immunity. In humans, three main DC subsets are defined: two types of conventional DCs (cDC1s and cDC2s) and plasmacytoid DCs (pDCs). To study DC subsets in the tumor microenvironment (TME), it is important to correctly identify them in tumor tissues. Tumor‐derived DCs are often analyzed in cell suspensions in which spatial information about DCs which can be important to determine their function within the TME is lost. Therefore, we developed the first standardized and optimized multiplex immunohistochemistry panel, simultaneously detecting cDC1s, cDC2s, and pDCs within their tissue context. We report on this panel's development, validation, and quantitative analysis. A multiplex immunohistochemistry panel consisting of CD1c, CD303, X‐C motif chemokine receptor 1, CD14, CD19, a tumor marker, and DAPI was established. The ImmuNet machine learning pipeline was trained for the detection of DC subsets. The performance of ImmuNet was compared with conventional cell phenotyping software. Ultimately, frequencies of DC subsets within several tumors were defined. In conclusion, this panel provides a method to study cDC1s, cDC2s, and pDCs in the spatial context of the TME, which supports unraveling their specific roles in antitumor immunity.


Introduction
Dendritic cells (DCs) are highly specialized antigen-presenting cells (APCs) that are known for their pivotal importance in the initiation and regulation of immune responses.More specifically, after antigen encounters and activation in the periphery, DCs secrete large amounts of proinflammatory cytokines and migrate to the lymph nodes.In the lymph nodes, DCs present antigens to T cells, inducing both T-cell activation and differentiation [1].Tumors can also be a source of antigens for DCs.Considering that an antitumor T-cell immune response is of paramount importance for tumor clearance, DCs are a crucial factor in initiating antitumor immunity [2,3].
DCs form a heterogeneous group of cells that can be divided into several DC subsets based on phenotype and function [4].In steady-state tissue and blood, three main DC subsets have been defined: two types of conventional (myeloid) DCs (cDC1 and cDC2) and plasmacytoid DCs (pDCs).For these subsets, varying roles in antitumor immunity have been proposed.
In brief, cDC1s are known to induce strong immunity against intracellular pathogens and tumors by interacting with CD8 T cells and eliciting CD4 Th1 cell responses [4,5].This proposed antitumorigenic function of cDC1s is confirmed in several tumor types in which an increased cDC1 gene signature is linked to improved overall survival [6].
Opposed to cDC1s, cDC2s preferably interact with CD4 T cells and mostly induce Th2 and Th17 responses [4].Melanoma patients who have a high intratumoral cDC2 frequency showed shorter progression-free survival compared with patients with a low intratumoral cDC2 frequency [7].On the contrary, patients with head and neck squamous cell carcinoma, breast cancer, and lung cancer showed a positive correlation between tumor-infiltrating cDC2s and survival [8][9][10][11].However, not all studies take into account the immunocompromised CD14 + cDC2 subgroup that was recently described in tumors [12,13].These CD14 + cDC2s induced lower T cell proliferation secreted higher levels of anti-inflammatory IL-10, and expressed higher levels of immunoinhibitory programmeddeath ligand 1 compared with their CD14 − equivalent in vitro.
Compared with conventional DCs, pDCs are less efficient in naïve T-cell priming but produce high levels of IFNs after pathogen encounter [14].Type I IFNs have direct effects on tumor cells by inhibiting proliferation and inducing apoptosis, but mainly exert their anticancer effect by stimulating antitumor immunity [15].For example, type I IFNs stimulate DC maturation and attract cytotoxic CD8 T cells and natural killer cells to the tumor microenvironment (TME) [4,15,16].Nevertheless, it has been shown that tumor-infiltrating pDCs often do not express co-stimulatory molecules [17].These pDCs usually show a tolerogenic phenotype, secrete IL-0, and induce regulatory T cells or CD4 T cell anergy [18,19].This might be the reason why tumor pDC infiltrates are correlated to decreased patient survival [20][21][22].Accordingly, a study of colorectal cancer in which tumorinfiltrating pDCs showed an activated phenotype, demonstrated a positive correlation between tumor-infiltrating pDCs and survival [21].
Currently, the role of the different DC subsets in tumor immunology is mainly investigated by gene expression profiling or flow cytometry on isolated cells or cell suspensions derived from tumor tissue.Unfortunately, with these techniques, all spatial information is lost.We consider this spatial information to be crucial in understanding the immunological processes within the tumor.For example, if immune cells are only located at the tumor-stroma border, this can indicate that immune cells do not have access to all tumor cells to induce an antitumor immune response.Tumor-stromal and intratumoral levels of T-cell infiltration have different correlations with cancer patient survival [23][24][25][26][27][28][29].This might apply to DCs as well.In addition, spatial data provides information on cell-cell communication between different immune cells.
Immunohistochemistry (IHC) is a technique that allows the analysis of immune cells in their spatial context.Previous research in our group used this technique to focus on lymphocyte and myeloid cells [30][31][32].To study DCs in general by IHC, CD11c is mostly used [33,34].However, CD11c is not specific to DCs as it is also expressed in humans by monocytes, macrophages, and other immune cells [35].In addition, by studying DCs in general, the specific role of each DC subset in tumor immunology remains unexplored.To our knowledge, this study provides the first standardized and optimized multiplex IHC (mIHC) pipeline to detect the three main DC subsets simultaneously in their spatial context.We report the process from marker selection to the mIHC workflow, including a recently developed cell quantification method.
The final marker selection resulted in a seven-color mIHC DC panel.Using this panel, pDCs, cDC1s, and cDC2s could be successfully detected in several tumor types.Through the implementation of CD19, cDC2s could be accurately discriminated from B cells (CD19 + /CD1c +/− ), and with CD14, immunocompromised CD14 + cDC2s could be identified.Next, the in-house developed machine learning pipeline "ImmuNet" was trained for accurate detection and quantification of cells.The development and validation of the ImmuNet pipeline is described before (article under revision, available in BiorXIV [36].Using this workflow, cDC1, cDC2, and pDC numbers were determined in tumor samples of several tumor types.
Our mIHC panel and analysis strategy could provide a reliable basis for future studies that aim to measure cDC1s, cDC2s, and pDCs in various tissue contexts.Furthermore, the design and validation method could be used as a set-up for panels of mIHC and other spatial techniques for the detection of immune cells in dense tissues.

XCR1 targeting results in the accurate detection of cDC1s in tumor tissue
CD141 (also known as BDCA3 or thrombomodulin) is a wellknown marker of phenotype cDC1s by flow cytometry [37].Low expression of CD141 is also observed on other immune cells, such as a small percentage of cDC2s [4].Moreover, CD141 is highly expressed on endothelial and mesothelial cells and has been used as an immunohistochemical marker for different tumor types such as mesothelioma and urothelial carcinoma [38][39][40].Indeed, upon assessment, abundant staining in the different tissues is observed, which mainly are not cDC1s (Fig. 1A).Other proteins that are specifically involved in the interaction of cDC1s with CD8 T cells such as C-type lectin domain family 9 member A (CLEC9A) and X-C motif chemokine receptor 1 (XCR1) are used for cDC1 phenotyping as well [4].Upon testing CLEC9A and XCR1, CLEC9A staining was more prominent compared with XCR1.In the literature, it is described that the prevalence of cDC1s in tissue is low, theoretically matching the staining pattern of XCR1 [41].IHC antibodies to detect the human protein expression of CLEC9A and XCR1 in tissue are relatively new, antibody specificity for cDC1s has therefore not been tested before.For peripheral blood, CLEC9A is mainly expressed on cDC1s, but also some cDC2s and a small subset of monocytes, while XCR1 is exclusively expressed by cDC1s [42][43][44].To confirm these observations in tissue, double staining of CLEC9A or XCR1 with CD11c, a marker that is highly expressed on conventional DCs and most monocytes, was performed.Double staining revealed that in most cases, CLEC9A + cells did not show co-expression with CD11c, while XCR1 + cells always showed CD11c co-expression (Supporting information Fig. S1).Hence, XCR1 is used for cDC1 phenotyping.

cDC2s can be characterized in tumor tissue by a combination of CD1c-, CD19-, and CD14-recognizing antibodies
For cDC2 phenotyping, CD1c (also known as BDCA1) is the main marker used across different techniques [4].However, B cells can also express CD1c [45,46].The high-intensity staining pattern observed for CD1c matches the expected cDC2 abundance in tissue, while a low CD1c staining pattern was observed in B cell follicles of the tonsil (Fig. 1B) [41].This was confirmed in additional mIHC staining, in which high CD1c expression was observed on cDC2s (CD11c + CD19 − ) and low CD1c expression was observed on some B cells (CD19 + ) (Supporting information Fig. S2).CD19 was included in the mIHC panel to accurately distinguish between CD1c + B cells and CD1c + cDC2s (Supporting information Fig. S3A).CD19 was chosen over CD20 as it is expressed in B cells in various differentiation states [47].To identify the earlier described immunocompromised CD14 + cDC2 subset, CD14 was included in the mIHC DC panel as well (Supporting information Fig. S3B) [12,13].

Antibodies recognizing CD303 accurately identify pDCs in tissue
pDCs are usually detected through IL-3 receptor (CD123) and CD303 (also known as CLEC4C or BDCA2).Both receptors are involved in type I IFN secretion by pDCs [4,48,49].IHC with both CD303 and CD123 resulted in staining that seemed specific (Fig. 1C).High expression of CD123 can be observed on basophils, and low expression on eosinophils and monocytes in tissue, whereas CD303 is described to be exclusively expressed on pDCs [50,51].Therefore, CD303 became the marker of choice for pDC identification.

IHC antibodies of the DC panel bind specifically to their epitope in tissue
To test the sensitivity and specificity of the chosen IHC antibodies for the mIHC DC panel, Chinese Hamster Ovary cells (CHO cells) were transfected with the selected markers and subjected to IHC and flow cytometry, using antibody clones suitable for each used technique (Fig. 2).CHO wildtype cells were used as a negative control.CHO wildtype cells did not show staining with the selected antibodies, while the transfected CHO cells did show staining with the relevant antibody (Fig. 2A).The varying levels of positivity observed between the different markers in the transfected CHO cells were also observed when these cells were analyzed with validated flow cytometry antibodies (Fig. 2B).Different staining intensities are presumably caused by differences in transfection efficiency per cDNA construct, rather than differences in the specificity of the IHC antibodies for their target.

The ImmuNet pipeline accurately detects and quantifies DC subsets in several tumor types
The selected DC markers were assembled and optimized into an mIHC DC panel according to our previously described procedure [52].mIHC was applied to several tumor types, in which the main DC subsets, CD14 + cells, and B cells could be visually identified (Fig. 3).Because DCs and CD14 + cells have a more complex morphology than lymphocytes, our recently in-house developed machine learning pipeline ImmuNet had to be trained on this dataset for the detection of DCs, CD14 + cells, and B cells [36].The error rate of ImmuNet was compared with the inForm software that is often used for the analysis of this type of data (Fig. 4).Errors are defined as illustrated in Fig. 4A.For our purposes, an error rate of 10% or lower for each cell type is strived for [36].
Initially, errors of both methods among validation annotations were calculated (Fig. 4B).Almost all error rates of ImmuNet were estimated to be below 10%, whereas the lowest estimated error rate of inForm was 19%.In general, error rates of all cell subsets were higher in inForm than in ImmuNet.Remarkably, the error rates of cDC1s and pDCs were 0% in ImmuNet, whereas 19% and 44%, respectively, in inForm.The highest error rate in both ImmuNet and inForm was observed for the CD14 + cDC2s.
An error type that was noted in inForm software was inaccurate cell segmentation [36].While some large cells are hypersegmented into multiple cells, some groups of multiple cells are hyposegmented as one large cell (Fig. 4C).In the first performance assessment, those errors were not always visible since annotations were scattered, and this type of error becomes most apparent when evaluating the phenotype of neighboring cells.Therefore, ROIs were drawn in tissue in which every cell was annotated.These were new annotations and were not included in the first method of error assessment.The number of annotated cells and cells detected by inForm or ImmuNet were then compared within these ROIs (Supporting information Fig. S4) and the absolute agreement was determined by Lin's concordance correlation coefficient (CCC; Fig. 4D).ImmuNet demonstrated the strongest CCC between detected and annotated total cell numbers (all cells with ImmuNet CCC = 0.994, 95% CI: 0.986-0.998;All cells with inForm CCC = 0.497, 95% CI: 0.-0.154-0.847) (Fig. 4D).Similarly, the CCCs of the individual cell types were always lower in inForm compared with ImmuNet (Fig. 4E).CCCs were lowest for cDC1s and CD14 + cDC2s (cDC1s with ImmuNet CCC = 0.664, 95%CI: −0.407-0.966;CD14 + cDC2s with ImmuNet CCC = 0.880, 95%CI: 0.455-0.979;cDC1s with inForm CCC = −0.095,95%CI: −0.722-0.682;CD14 + cDC2s with inForm CCC = −0.416,95%CI: −0.879-0.451).For inForm there were on average fewer cells detected than had been annotated, indicating false negative and/or hyposegmentation.

Seven-color mIHC detects comparable DC subset frequencies in CD3-depleted PBMCs as elaborate flow cytometry marker panels
In other techniques, a wide range of markers is used simultaneously for DC subset phenotyping in cell suspensions.For current state of art multispectral mIHC, the marker panel is limited to seven colors, including a tumor marker and DAPI nuclear staining.To investigate whether seven markers can reliably identify DC subsets, cell material was divided and DC phenotyping by mIHC was compared with DC phenotyping by flow cytometry (Supporting information Fig. S6).Due to the very heterogenicity of tumor tissue, two randomly selected halves might be very different in cell composition and cell numbers.Very scarce DC subsets may not even be present in tumor material.Therefore, we chose peripheral blood mononuclear cells (PBMCs), a homogeneous cell suspension, to compare mIHC with flow cytometry.To increase the relative DC subset frequency, PBMCs (n = 7) were depleted of CD3 + cells.For DC detection by flow cytometry, an established gating strategy from our group was used (Supporting information Fig. S5A).For DC subset gating in the mIHC sam-ples (Supporting information Fig. S5B), cDC1s, CD14 − cDC2s, and pDCs were pregated on CD14 − CD19 − cells.cDC1s were then distinguished as XCR1 + CD303 − and pDCs as XCR1 − CD303 + .From the XCR1 − CD303 − population, CD1c-positive cells were identified as CD14 − cDC2s.CD14 + cDC2s and CD14 + cells were pregated on CD14 + CD19 − cells, where CD14 + cells were identified as CD1c-negative and CD14 + cDC2s as CD1c-positive.B cells were gated as CD14 − CD19 + .All DC subsets could be distinguished in CD3-depleted PBMCs (Fig. 6A).A favorable concordance between flow cytometry and mIHC was observed for the detection of all cells combined (Fig. 5B) (CCC = 0.995, 95% CI: 0.990-0.998).For the individual cell types, the highest concordance between flow cytometry and mIHC was observed for the detection of pDCs (CCC = 0.969, 95% CI: 0.902-0.991).The detection of CD14 + The percentage of each cell subset was calculated for all cell subsets that were identified by ImmuNet (CD14 + cells, B cells, pDCs, cDC1s, cDC2s, and CD14 + cDC2s).Percentages measured in both techniques are displayed in an X/Y plot on logit-transformed axes.(C) Lin's CCC and the 95% CI of each cell subset were measured in mIHC versus flow cytometry.Lin's CCC and its 95% CI were calculated using jackknifing with Fisher's Z transformation.CCC, concordance correlation coefficient; cDC, conventional dendritic cell; DC, dendritic cell; FFPE, formalin-fixed paraffin-embedded; mIHC, multiplex immunohistochemistry; PBMCs, peripheral blood mononuclear cells; pDCs, plasmacytoid dendritic cell.cells and B cells was also strongly concordant (CCC = 0.953, 95% CI: 0.861-0.985and CCC = 0.951, 95% CI: 0.863-0.983,respectively; Fig. 5C).A modest accordance was observed for cDC2s (CCC = 0.734, 95% CI: 0.210-0.931)and CD14 + cDC2s (CCC = 0.767, 95% CI: 0.301-0.937;Fig. 5C).cDC1 detection showed the lowest concordance between both techniques (CCC = 0.454, 95%CI: −0.125-0.802;Fig. 5C).

The distinct DC subset can be detected in various tumor types
To investigate whether the developed mIHC DC panel could be used to detect DC subsets in different tumor types, samples from several tumor types were analyzed.Representative images of each tumor type are shown in Supplementary information Fig. S7.The presence of the DC subsets appears to differ between tumor samples.We only tested a few samples, so we cannot draw any conclusions about differences between tumor types (Fig. 6).Remarkably, glioblastoma and renal cell carcinoma samples showed low numbers of total DCs, while lung adenocarcinoma and esophageal cancer samples, showed high numbers of total DCs.cDC2s were the most frequent tumor-infiltrating DC subset (mean = 33 cells/mm 2 ; SD = 4.47), followed by CD14 + cDC2s (mean = 18 cells/mm 2 ; SD = 3.47) and cDC1s (mean = 14 cells/mm 2 ; SD = 3.24).pDCs were the rarest tumor-infiltrating DC subset (mean = 4 cells/mm 2 ; SD = 3.72).Although these results are suggestive of differences between tumor samples and tumor types, these results need to be studied in representative cohorts in which factors such as tumor subcategory and clinical factors of the patients can also be included in the analysis.

Discussion
In this study, we developed an mIHC panel that phenotypes cDC1s, cDC2s, CD14 + cDC2s, and pDCs in tumor tissue.XCR1, CD1c, CD303, CD14, and CD19 proved to be suitable markers for DC subset identification using mIHC.By combining this mIHC panel with the ImmuNet pipeline, DC subsets were automatically detected and quantified in various tumors.Automatic detection of DCs by ImmuNet showed lower error numbers than established automatic cell detection software.This indicates that ImmuNet is the preferred pipeline to pick up small differences in DC subset numbers between samples.To the best of our knowledge, this study provides the first standardized and optimized mIHC protocol for identifying the main DC subsets simultaneously in the spatial context of the TME.Studying DCs in this context allows us to further elucidate the biological role that the different DC subsets play in tumor immunology.
Marker selection for cDC2s and pDCs was straightforward.Selecting markers to identify cDC1s was challenging.The use of antibodies directed against CD141 or CLEC9A resulted in an unspecific staining pattern.While the abundant expression pattern of CD141 in tissue could be explained by its presence in endothelial cells, the cause of the unspecific expression of CLEC9A remains unclear.Since both cDC2s and CLEC9A + monocytes also express CD11c, the off-target was not caused by these cells as was previously hypothesized [42].The commercial availability of most DC-specific monoclonal antibodies for IHC is limited, so it could not be tested if unspecific staining was caused by this particular CLEC9A clone.Staining with XCR1 showed a pattern that corresponded more to the expected cDC1 prevalence in tissue [41,43].XCR1 can be internalized under immunosuppressive conditions such as in tumors or after stimulation, as was observed in experiments in vitro [6].In (m)IHC, intracellular expression can still be detected as the cells are cross-sectioned.However, if internalization is followed by degradation, XCR1 will not be expressed anymore.This means that there could be XCR1 − cDC1s in the tumor that will go undetected with our mIHC panel.All markers selected for DC subset phenotyping using the mIHC panel were also validated as DC subset-specific in single-cell RNA sequencing data of blood-derived DCs and/or (metastatic) lung cancer tissue [10,[53][54][55].
The largest restraint of mIHC is the number of markers that can be included.For studying DC function, maturation markers, such as CD86, or markers important for antigen presentation, like human leukocyte antigen DR, should be added.Human leukocyte antigen DR single staining was tested and revealed an abundant staining pattern, explained by its expression by all APCs, some tumor cells, and activated T cells [56] (Supporting information Fig. S3C).The limited number of markers within one panel also prevents the identification of other immune cells in the same section.Fortunately, there are possibilities to extend the mIHC DC panel using newer imaging systems such as the PhenoImager HT (Akoya Biosciences).This allows for two extra fluorophores to be added.However, this comes at the expense of the unmixing precision and imaging speed.Another option would be to replace CD19 with a different marker.CD1c + B cells were only occasionally found in tertiary lymphoid structures surrounding tumor tissue and were hardly seen in other parts of the TME.This is in accordance with literature [57].Another option to obtain more information on DCs or other immune cells is to work with consecutive slides from the same tissue block and optimize current techniques for stacked image analysis [58].In our experiments, slide thickness is 4μm.This could be suitable for revealing cell-cell interactions on adjacent slides, as the average size of mature DCs is 10-15 μm [59].To analyze smaller cells, such as T cells (7-10 μm), cutting thickness could be decreased to 1 μm [60].
ImmuNet performance in tissue was assessed by two different approaches.For the first approach, the error rate of DC pheno-typing was calculated for annotations made at manually selected tissue sites.Except for the CD14 + cDC2s, all error rates remained below the aimed cut-off error rate of 10% [36].CD14 + cDC2s were only annotated when there was co-expression of CD14 and CD1c at the complete membrane of a cell, to avoid phenotyping neighboring CD14 + cells and cDC2s.CD14 + cDC2s were rarely present in tissue (Supporting information Table S1).Their rarity together with the strict requirements for annotation, resulted in fewer annotations for teaching ImmuNet.This explains the larger error rate.
In the second approach, the concordance between the annotated phenotype and detected phenotype was calculated for completely annotated ROIs.Besides incorrect phenotyping, this manner of performance assessment also identifies incorrect cell segmentation.Here, the annotation and detection of CD14 + cDC2s showed a moderate concordance, even though cell numbers within the ROIs were low.The cell type that showed the lowest concordance between annotations and detected cells by ImmuNet within ROIs was the cDC1s.When comparing the differences between annotated and detected cDC1 frequencies, no clear pattern of either hypo-or hypersegmenation of cells by ImmuNet was observed.Since the numbers of cDC1s within the ROIs are very low, small inconsistencies have a strong negative influence on the CCC.For the initial error rate calculations, larger numbers of cDC1 annotations were used, and therefore small differences do not result in a high error rate in this method of performance assessment.
In CD3-depleted PBMCs, DC phenotyping by flow cytometry showed a remarkable concordance to DC phenotyping using mIHC.When evaluating this concordance, it should be taken into consideration that flow cytometry and mIHC not only differ in the number of markers used for cell identification but also have other fundamental differences.For technical reasons, different antibody clones are used for both techniques.In addition, in flow cytometry, single cells are analyzed, while in mIHC all cells are analyzed simultaneously.The marker expression of neighboring cells can influence cell phenotyping, making analysis of each cell membrane more difficult in mIHC.ImmuNet is therefore accurately trained for the analysis of individual cell membranes in tissue.When studying the concordance of each specific cell type between flow cytometry and mIHC, the lowest CCC was noted for cDC1s, which were the least prevalent DC type within CD3-depleted PBMCs (Supporting information Fig. S8).A major difference is that the cDC1 gating in flow cytometry is based on CD141 and CLEC9A positivity, while mIHC gating is based on XCR1 expression (Supporting information Fig. S5).We do not use XCR1 for flow cytometry gating as XCR1 can be internalized as described above [6].However, when comparing the percentage of CLEC9A + CD141 + cDC1s to the percentage of XCR1 + CD141 + cDC1s detected by flow cytometry in the CD3depleted PBMCs, differences are minor (Supporting information Fig. S9).
As stated before, when detected cell frequencies are low, a skewed image can arise as small differences can influence the CCC strongly.
In Fig. 6 we show that with our panel we can identify DC subsets in different tumor types.To correlate DC subset frequencies with tumor type or clinical characteristics, the mIHC panel should be applied in patient cohorts.It would be of interest to correlate the spatial information acquired by IHC to clinical parameters.For instance, it can be studied whether DC presence in certain locations in the TME is correlated to prognosis or response to treatment.On a cellular level, mIHC staining identifies cells in close proximity to each other as indicative of cell-cell interactions.The average distance between DCs and tumor cells or other immune cells could be calculated and used for further understanding of the biological processes.
Besides tumor samples, it will be interesting to use the mIHC DC panel on lymph nodes.DCs activate T cells in the lymph nodes, which is a crucial step in antitumor immunity.The ratio of the different DC subsets and the interaction that DCs have with other immune cells could be studied in tumor-draining lymph nodes.When applying the DC panel to a distinct tissue, small adaptations in antibody dilutions or additional annotations might be required (Supporting information Table S2).
The seven-color mIHC panel developed in this study is the first mIHC panel that accurately identifies cDC1s, cDC2s, and pDCs simultaneously in tissue.Automatic cell detection by ImmuNet was optimized and shown to be a superior method for DC detection compared with inForm.In future research, this mIHC panel can be used to elucidate the link between DCs in the TME and disease characteristics or response to treatment.

Patient tissue material
Tonsil tissue and tumor samples of lung adenocarcinoma and melanoma were used for testing potential DC markers in chromogenic IHC.Further optimization of the fluorescent mIHC panel was performed in tonsil tissue.
The final mIHC panel was tested on tissue microarrays (TMAs) containing tissue of 15 different tumor types.Per tumor type, tissue from two or three different patients was selected for the TMA.Each TMA core was 2 mm in diameter and was aimed to be obtained at the tumor-stroma border.
All tissue samples were anonymized and randomly chosen based on availability.According to Dutch legislation before 2020, the used tissue material is coded formalin-fixed and paraffin-embedded (FFPE) left-over material, for which patients have not objected against the use in (clinical) research.

Cell culture and transfection
CHO cells (85050302, Sigma Aldrich) were transfected with cDNA constructs to confirm the sensitivity and specificity of the candidate IHC antibodies for the DC panel as described before [52].cDNA constructs encoding CD1c, CD303, XCR1CD14, and CD19 (OHu03695D, OHu05871D, OHu29470D, OHu26942, and OHu27320D, respectively; all from Genscript) were used.Three million CHO cells were seeded in a T75 flask and cultured for 24 h to obtain confluency of around 70% at the time of transfection.CHO cells were transfected using the Lipofectamine TM LTX with PLUS TM Reagent kit (A12621, Thermofisher Scientific) according to the manufacturer's instructions.Transfection efficiency was measured by flow cytometry 48 h after transfection and varied between 6-42%, depending on the cDNA construct that was transfected.About 5-15 million transfected CHO cells were processed based on the AgarCyto cell block preparation for IHC analysis [61].An adaptation to the protocol was that cells were resuspended in 30-40 μL 2.25% liquid agarose, equal to the volume of the cells.These cells were then transferred to a dispomold embedding device (7 × 7 × 5 mm, 7820, Bio Optica) and solidified at 4°C for 10 min.This agar-cell template was then dissolved in formalin before further processing and paraffin embedding.

PBMC isolation and CD3 depletion
To validate the sensitivity of the developed mIHC DC panel, we compared the results of mIHC with DC detection by flow cytometry in PBMCs that were enriched for DCs through T-cell depletion.PBMCs were isolated from buffy coats of seven healthy donors (Sanquin Nijmegen) as described before [62].CD3 depletion was performed immediately, or after cryopreservation.CD3 depletion was performed by magnetic-activated cell sorting using anti-CD3-conjugated microbeads (130-050-10, Miltenyi).From these CD3-depleted PBMCs, ∼15 million cells were processed by AgarCyto cell block preparation for IHC analysis as described above and ∼1 million cells were used for each flow cytometric staining.

Flow cytometry
Flow cytometry was used for confirmation of successful transfection of the CHO cells and validation of the mIHC panel in CD3depleted PBMCs.
CD3-depleted PBMCs were first incubated with Fixable Viability Dye eFluor780 (65-0865-14, Thermofisher Scientific) at 4°C for 30 min, to distinguish between living and dead cells.Thereafter, cells were washed with protein blocking agent (PBS + 1% BSA and 0.1% sodium azide) and blocked in protein blocking agent + 2% human serum (H4522, Sigma-Aldrich) for 10 min at 4°C.Next, both CD3-depleted PBMCs and CHO cells were incubated with fluorochrome-conjugated primary antibodies at 4°C for 30 min (Supporting information Table S3).Samples were acquired using the BD FACSLyric System (BD Biosciences).Data were analyzed in FlowJo software (v10, Three Star).The gating strategy is shown in Supporting information Fig. S5A.Our exper-iments and analysis are in adherence with the "Guidelines for the use of flow cytometry and cell sorting in immunological studies" [63].

Chromogenic and mIHC staining
FFPE material of tumor tissue, tonsil tissue, TMAs, CHO cells, and CD3-depleted PBMC suspensions were cut into 4μm thick sections.mIHC staining was performed automatically with the Opal 7-Color Automation IHC Kit (NEL821001KT, Akoya Biosciences) on the Leica Bond system (BOND-Rx Fully Automated IHC and ISH, Leica Biosystems) as described before [30,64].Immunofluorescent staining was counterstained using DAPI nuclear staining (NEL797B001KT, PerkinElmer) at room temperature for five minutes, and enclosed in Fluoromount-G (0100-01, Southern Biotech).Chromogenic staining was performed manually as described before [65].Antibodies used for staining are listed in Supporting information Table S4.
The optimal order of the antibodies and Opal dye combinations within the multiplex was determined with the method described by Gorris et al (Supporting information Table S2) [52].

Tissue imaging and analysis
Multispectral imaging was performed on the Vectra 3 imaging system (PerkinElmer) at 20× magnification.This was followed by unmixing the multispectral images in inForm Advanced Image Analysis software (inForm 2.4.8,PerkinElmer), using libraries made from single stained tissue.For further analysis of the mIHC staining, either inForm software or the ImmuNet pipeline was used [36].
For inForm analysis, a selection of 15-20 representative multispectral images was used to train tissue segmentation, cell segmentation, and phenotyping.These training settings were saved in an algorithm that was used on all multispectral images of the same experiment in a batch analysis.Batch analysis was optimized by manually checking the success of the network in images that were not used for training.If inForm phenotyping in these images did not match our expert opinion about phenotyping, these images were added as new inForm training images.Four rounds of these optimizations were performed.
The ImmuNet pipeline, originally developed for lymphocyte detection, was compared with other programs for automatic cell detection in tissue (inForm) and validated by comparison with another technique (flow cytometry) [36].The ImmuNet pipeline is suitable for other mIHC panels and adapting it for the DC panel was relatively simple.Similar to the inForm training procedure, a few rounds of training were performed and at the end of each round, the performance of ImmuNet was evaluated.To train ImmuNet, thousands of annotations were made with a custom  S1).For DC phenotyping in PBMCs, 2806 additional annotations were made in the PBMC dataset.Images from both the PBMC dataset and from datasets of tumor tissue were used to train a separate model.ImmuNet-recognized cells were converted into Flow Cytometry Standard (FCS) files.These FCS files contained the following information on the data: X and Y coordinates, marker expression prediction by the model (pseudo marker expression) of CD1c, CD303, XCR1, CD14, CD19, and DAPI, and staining intensity extracted from the composite of CD1c, CD303, XCR1, CD14, CD19, DAPI, and the autofluorescence channel (the average intensity in the disk of radius two, centered on the coordinates of a predicted cell).FCS files are available upon request.The marker expression predicted by the model was used for phenotyping of the cells using the gating strategy of Supporting information Fig. S5B in FlowJo Software (v10, Three Star).

Statistics
The performance of the phenotyping models of inForm and ImmuNet was assessed with two different methods.First, the error rate of the validation annotations was calculated.Errors in the model are annotations that were either not recognized as a cell or annotations that were assigned to a different cell subset within a 3.5 μm radius of the annotation (Fig. S4A).The phenotype of a cell was determined using the standard threshold of 0.4 for pseudomarker expression established before [36].The error rate was calculated per cell subset as (# errors that occurred for this specific cell subset/# all annotations for this subset) × 100.Annotations used for error rate calculations were not included in the training of the models but were extracted from tissue that the models were optimally trained on.Second, ten ROIs were drawn within the tissue, in which each cell was annotated.Within all ROIs, the annotations of each cell subset were compared with cells detected by inForm or ImmuNet by computing Lin's CCC and its 95% CI.Due to the small sample size, estimates and CIs were computed by jackknifing with Fisher's Z transformation [66].
To study the correlation between flow cytometry and mIHC in CD3-depleted PBMCs, the abundance of each cell subset was first calculated as a proportion of all cells that could be phenotyped with mIHC (B cells, CD14 + cells, cDC1s, cDC2s, CD14 + cDC2s, and pDCs).This proportion was then logit transformed for further analysis.Flow cytometry and mIHC were compared by computing Lin's CCC and its 95% CI, using jackknife resampling with Fisher's Z transformation.The CCC and its CI range between −1 (maximal discordance) via 0 (maximal concordance) to 1 (strong concordance).Concordances between 0 and 1 were interpreted as follows: <0.4-poor, 0.4-0.59-modest,0.6-0.75good,and ≥0.75-remarkable.Statistics were calculated in R studio.

Figure 2 .
Figure 2. IHC and flow cytometry of CHO cells transfected with cDNA constructs encoding DC markers.(A) Chromogenic DAB IHC of the selected DC markers on CytoAgar of CHO wildtype cells (top panel) and CHO cells transfected with cDNA encoding the indicated marker (bottom panel).Original magnification of the images ×20.The scale bar equals 30 μm. (B) Flow cytometry plots of the DC markers in CHO wildtype (red) and transfected CHO cells (black).Antibody clones that are used for chromogenic DAB IHC and flow cytometry are indicated above each result.BV, brilliant Violet; CHO cells, Chinese hamster ovarian cells; DAB, 3,3 -diaminobenzine; DC, dendritic cell; IHC, immunohistochemistry; PerCP, peridinin-chlorophyllprotein complex; SSC, sideward scatter.

Figure 3 .
Figure 3. Seven-color mIHC staining identifying cDC1s, cDC2s, CD14 + cDC2s, pDCs, CD14 + cells, and B cells in different tumor types.Sections containing invasive margin and tumor tissue were stained with mIHC using optimized antibody dilutions and optimal order of antibodies within the multiplex.Pictures on top, original magnification x20.Red boxes in the upper panel indicate regions on which was zoomed in in the lower panel, to show each cell subset.In the lower panel, the composite and each separate fluorescence channel are shown for this location.Channels that are highlighted in green were interpreted as positive and determined the decision to designate the cell to a certain cell subset.Sizes of the scale bars are noted in the image.cDC, conventional dendritic cell; mIHC, multiplex immunohistochemistry; pDC, plasmacytoid dendritic cell.

Figure 4 .
Figure 4. Performance of ImmuNet and inForm in recognizing DC subsets, CD14 + cells, and B cells.(A) Errors of the model are annotations that wereeither not recognized as a cell or annotations that were assigned to a different cell subset within a 3.5 μm radius (white circle) of the annotation (small black outlined circle in the color corresponding to the annotated phenotype).Cells that are detected are displayed as a larger colored circle (corresponding to the phenotype) with a white circle in the middle.(B) Error rates with their 95% CI are displayed.Error rates were calculated per cell subset as follows: errors divided by all annotations made for this cell type.The 95% CI was computed by the Wilson/brown method.The dotted line indicates an error rate of 10%, which was the chosen cut-off value for successful cell detection.Annotations were made on a TMA containing several tumor types.(C) Hypersegmenation and hyposegmentation of cDC2s by inForm software.(D) Correlation between the number of annotated immune cells versus detected immune cells by inForm and ImmuNet within seven ROIs on a log10 scale.Lin's CCC of all cells together is displayed within both graphs.(E) Lin's CCC and the 95% CI of each cell subset annotated versus detected by inForm or ImmuNet.Lin's CCC and its 95% CI were calculated using jackknifing with Fisher's Z transformation.All scale bars equal 10 μm.Abbreviations: CCC, concordance correlation coefficient; CI, confidence interval; cDC, conventional dendritic cell; DC, dendritic cell; ROI, region of interest; TMA, tissue microarray.

Figure 5 .
Figure 5. Identification of DC subsets, CD14 + cells, and B cells by flow cytometry and 7-color mIHC.(A) The mIHC DC panel was applied to a CytoAgar of CD3-depleted PBMCs in which cDC1s, cDC2s, CD14 + cDC2s, pDCs, CD14 + cells, and B cells can all be visualized.Original magnification of the image 20×.Sizes of the scale bars are noted in the image.(B)The percentage of each cell subset was calculated for all cell subsets that were identified by ImmuNet (CD14 + cells, B cells, pDCs, cDC1s, cDC2s, and CD14 + cDC2s).Percentages measured in both techniques are displayed in an X/Y plot on logit-transformed axes.(C) Lin's CCC and the 95% CI of each cell subset were measured in mIHC versus flow cytometry.Lin's CCC and its 95% CI were calculated using jackknifing with Fisher's Z transformation.CCC, concordance correlation coefficient; cDC, conventional dendritic cell; DC, dendritic cell; FFPE, formalin-fixed paraffin-embedded; mIHC, multiplex immunohistochemistry; PBMCs, peripheral blood mononuclear cells; pDCs, plasmacytoid dendritic cell.

Figure 6 .
Figure 6.Evaluation of the different DC subsets in several tumor types.A TMA slide containing two or three random samples of each tumor type was stained with the mIHC DC panel.Individual tumor samples are shown in the columns.Marker expression is displayed as marker-positive cells/mm 2 on a log10 scale.Tumor types are sorted based on the lowest presence of total DCs (left) to the highest presence of total DCs (right).Regions of cell analysis contained tumor tissue only, without invasive margin.DC, dendritic cell; DLBCL, diffuse large B cell lymphoma; mIHC, multiplex immunohistochemistry, TMA, tissue microarray.
web browser-based annotation tool.For unbiased evaluation, the whole dataset was split into training and validation sets by tile.Training sets were used for training of the model and validation sets were used for evaluation of the model, in a way that model performance was assessed only on images unseen during training.Validation data amounted to about 20% of all annotations.It was ensured that both training and validation datasets had a similar distribution of different cell types.If ImmuNet performance was unsatisfactory for certain cell types, the error cases were manually evaluated to identify reasons for incorrect predictions, and additional annotations were made.The final DC model used for DC subset detection and phenotyping in tissue was trained on 85760 patches (63 × 63 pixels) sampled from 10278 annotations (Supporting information Table