TREM2-Expressing Multinucleated Giant Macrophages Are a Biomarker of Good Prognosis in Head and Neck Squamous Cell Carcinoma

Squamous cell carcinoma (SCC) is one of the most frequent carcinomas and arises in several organs including lung, head and neck, esophagus, skin, and uterine cervix. Among these, head and neck SCC (HNSCC) is particularly common, with 890,000 new cases and 450,000 deaths globally in 2020. Its incidence is rising, expected to escalate by 20% by 2030, representing 1.08 millions of new cases per year (1). Current treatments include surgery, chemotherapy, and radiotherapy; alongside, immunotherapy is now approved for recurrent, metastatic, and unresectable carcinomas (2). However, more than 50% of patients experience recurrent or metastatic disease within 3 years of diagnosis (2) often facing numerous complications or disabilities even after recovery (3). Gold standard patient stratification is the tumor–node–metastasis staging where patients with equivalent staging receive similar treatment although they often hold heterogeneous outcomes. Such suboptimal stratification led to inadequate therapy with undertreatment compromising patient recovery or overtreatment increasing adverse events and unnecessary costs. Therefore, there is an unmet clinical need for novel individual prognostic biomarkers to guide therapeutic decisions.

A deeper understanding of cancer biology is essential for improving treatment. Recent studies have underscored the long under-appreciated role of the tumor microenvironment (TME) in shaping all stages of the disease (4). Within the TME, tumor-associated macrophages (TAM) are particularly abundant and influential (5). Recent single-cell RNA sequencing (scRNAseq) studies have described several conserved TAM transcriptional programs across tumors (6). Of particular interest, macrophages expressing the Triggering Receptor Expressed on Myeloid Cells-2 (TREM2) have been shown to exert immunosuppressive and protumoral functions in various cancers (7, 8).

Alongside conventional TAMs, a proportion of HNSCC also contains macrophages that exhibit a large cytoplasm and multiple nuclei, termed multinucleated giant cells (MGC; ref. 9). In SCC, macrophages and monocytes fuse to form MGC, performing a “foreign body reaction” against extracellular keratin produced by carcinomatous cells (9). While their presence is well documented in keratinizing SCC (10), their biology is not well understood in tumors and studies suggest varying impact of MGC on patient survival in esophageal and oral SCC (11–13). Conventional methods such as flow cytometry and scRNAseq struggle to analyze these cells due to their extraordinary size, leaving their clinical impact in SCC largely unexplored.

Herein, we used samples from several large cohorts of patients with HNSCC to ask about the association between MGC and survival following resection of their tumors. We investigated the interaction between chemotherapy, MGC, and patient outcomes and devised a methodology for clinical use adaptable to various SCC. Additionally, we employed spatial transcriptomic and proteomic approaches on tumor microscopic slides to dive further into the MGC-related TME and reveal the unique scRNA signature of MGC within keratin-rich tumor niches and their possible relationship to TREM2-expressing TAM.

We first characterized the presence and abundance of MGC in tumors from patients with HNSCC who underwent primary surgery (Fig. 1A). We used cohorts from The Cancer Genome Atlas (TCGA) and Gustave Roussy (GR; refs. 14, 15), consisting of 394 patients with oral cavity (Fig. 1B). We observed that 292 patients bore keratinizing tumors, while the remaining 102 did not (Fig. 1C and D). Within keratinizing tumors, we observed MGC that were aggregated in large granulomatous clusters in close contact with extracellular keratin (Fig. 1E). Interestingly, none of the non-keratinizing tumors contained MGC, while 33% of highly keratinizing tumors contained high MGC densities (Fig. 1F). Although MGC abundance was significantly associated with the keratinization level of the tumors, the presence of keratin did not guarantee the presence of MGC as 37% of patients with highly keratinizing SCC were devoid of MGC infiltration (Supplementary Fig. S1A and S1B).

We then asked whether the density of MGC in these tumors was related to patient outcomes. Using microscopic slides of HNSCC cohorts from TCGA (n = 110 patients) and from GR (n = 284 patients), we stratified patients into three groups (MGC^High, MGC^Int, and MGC^Low) according to the MGC density in their tumor (Supplementary Fig. S1C and S1D; Supplementary Table S1). In both cohorts, we saw that higher densities of MGC were associated with longer overall survival (OS; Fig. 1G and H) and progression-free interval (PFI; Supplementary Fig. S1E and S1F). In the combined TCGA and GR cohorts (n = 394 patients), we observed a positive association between MGC density and OS and PFI with a linear relationship between log-hazard of death and MGC density and the square root of MGC density and between log-hazard of event and the square root of MGC density (Supplementary Fig. S1G and S1H; Supplementary Tables S2–S4). A significant association was also observed when MGC density was categorized in the three groups (MGC^High, MGC^Int, and MGC^Low; Fig. 1I and J). Demographic and disease characteristics of the patients were well balanced between the three groups (Supplementary Table S2). In multivariate analyses adjusted for sex, age, p-stage, location, tobacco, and alcohol consumption, the density of MGC in the tumor (whether in three categories or with the square root of density as a continuous variable) was significantly associated with OS and PFI (Supplementary Tables S3 and S4). Importantly, high keratinization in the context of low/intermediate MGC densities was not associated with a better OS and PFI (Supplementary Fig. S1I–S1K; Supplementary Table S5), showing that it is not high keratinization per se that correlates with outcome.

In addition, we asked whether tumor-infiltrating lymphocytes (TILs) density, tumor grade, and CD68-positive mononuclear macrophages were related to the prognosis value of MGC. MGC^High patients had higher levels of TILs as compared to MGC^Int and MGC^Low patients (Supplementary Fig. S2A–S2C) and in univariate analysis, TILs were significantly associated with OS (P value = 0.0002) and PFI (P value = 0.0008; Supplementary Tables S6 and S7). In addition, higher tumor grade was associated with higher risk of death and events (Supplementary Fig. S2D and S2E; Supplementary Tables S6 and S7). Importantly, in multivariate analysis adjusted for TILs and tumor grade, MGC remained significantly associated with survival and PFI (Supplementary Tables S3 and S4). Finally, CD68 mononuclear macrophages did not show significant association with risk of death and events (Supplementary Fig. S2F and S2G; Supplementary Tables S6 and S7).

Taken together, these data show that patients with HNSCC bearing keratinizing tumors are more likely to have positive outcomes in the presence of high MGC densities.

Next, we extended our investigation to include a cohort of patients with HNSCC (n = 52 patients) treated by three cycles of induction-chemotherapy (ICT) followed by surgery and adjuvant therapy (Fig. 1K and L). Patients were classified into three groups: no residual tumor and residual tumors having either high (MGC^High) or low/intermediate (MGC^Low/Int) MGC densities (Fig. 1L and M; Supplementary Fig. S3A and S3B). Interestingly, high MGC densities in surgical resections were more common after ICT compared to treatment-naive patients (Fig. 1N), suggesting that ICT might induce the formation of MGC. As observed in treatment-naive patients, higher MGC densities associated with longer OS, comparable to patients with no residual tumor (Fig. 1O). Pathological response was similarly correlated with MGC density: good/partial responders had higher MGC densities than poor responders (Fig. 1P).

In summary, MGC density is associated with good outcome in patients with HNSCC either in the presence or absence of ICT prior to surgical resection.

Effective stratification of patients with cancer is a mainstay of good clinical management, yet few novel biomarkers prove amenable to quick and easy measurement in hospital routine: for example, here, it would take a trained pathologist up to 30 minutes per slide to assess MGC density. Therefore, we aimed to establish a high-throughput automatic method to detect and enumerate MGC and compute the biomarker on whole-slide images (WSI) of patients with SCC. We trained two models on routinely hematoxylin and eosin (H&E)/hematoxylin, eosin, saffron (HES) stained WSI: one that detects tumor cells and one that detects MGC (Fig. 2A and B). More than 5,000 manually annotated MGC and more than 800,000 manually annotated carcinomatous cells from the TCGA and GR cohorts were used to train the MGC and tumor cells detection model. The automated biomarker was computed as the total number of detected MGC divided by the total number of detected tumor cells, across all available slides for each patient.

After training, we validated the performance of both detection models on 110 slides from TCGA cohort and 839 slides from GR cohort. Slides from the GR cohort used for training were excluded from the validation cohort. We found that the models performed comparably to pathologists, both for MGC in TCGA (average precision (AP) = 0.8188) and GR (AP = 0.7524) as well as for tumor cells in TCGA (AP = 0.688) and GR (AP = 0.679) cohorts. Additionally, the MGC detection model performed comparably to pathologists at the slide-level in TCGA (r² = 0.8792; P < 0.0001) and GR cohorts (r² = 0.7811; P < 0.0001; Fig. 2C–F). We then assessed the performance of the artificial intelligence (AI) versus the manual biomarker. On the 949 slides (representing 341 patients) that passed quality control (Supplementary Fig. S4A), the automatic MGC biomarker correlated with the manual MGC density in TCGA (r² = 0.7833; P < 0.0001) and GR (r² = 0.4036; P < 0.0001; Supplementary Fig. S4B and S4C) cohorts. When comparing the ranking of the predicted patient AI scores with the manually computed groups, the AI obtained a rank-biserial correlation of 0.929 in TCGA and 0.867 in GR. The area under the ROC curve (AUC) between the predicted patient AI scores and the manually computed MGC^Low versus MGC^Int/High was 0.990 in TCGA and 0.965 in GR, and 0.963 in TCGA and 0.909 in GR for MGC^High versus MGC^Low/Int. Accordingly, patients were efficiently stratified into MGC^High, MGC^Int and MGC^Low groups and the automatic MGC biomarker was associated with longer OS and PFI (Fig. 2G and H).

We then asked whether our automatic detection model could be used in Uterine cervix (CESC) and Larynx SCC. Initially, MGC predictions gave false positive MGC detections in both cervical and laryngeal SCC TCGA cohorts; due to tissue-specific features of the larynx and cervix that were absent in the oral cavity and so not included in the training data. We overcame this issue by calibrating the MGC detection model to increase its sensitivity. Doing so, we excluded the risk of false negatives, while leading to abundant (true and false) positives that were then manually corrected by a pathologist: although slower than full automation, this method was approximately 10 times faster compared to a pathologist-only approach (Supplementary Fig. S4D and S4E) and reached similar levels of accuracy (Supplementary Fig. S4F–S4I). Using this approach, we categorized patients into high and low MGC/tumor ratios and found that increased proportions of MGC again correlated with longer OS in both CESC and Larynx SCC (Fig. 2I and J) and with longer PFI in CESC (Supplementary Fig. S4J and S4K).

Altogether, our automatic cell detection approach was rapid and efficient, allowing easy assessment of the MGC-to-tumor cell ratio that acted as a clear biomarker of prognosis in patients with oral cavity, Uterine cervix, and Larynx SCC.

Having identified the relationship between MGC density and prognosis, we next wanted to investigate their biological function within the TME. We exploited Visium spatial transcriptomic technology (10x Genomics) to analyze the single giant cell transcriptomes of MGC, on formalin-fixed, paraffin-embedded (FFPE) HNSCC tumor sections (Fig. 3A and B). Such technology assesses transcriptomes on spots, directly on histological slides. Importantly, because the diameter of a Visium spot is 55 μm, this technology has a pseudo-single cell resolution. Here, we took advantage of this limit, as one spot corresponds roughly to the size of an individual MGC, allowing the definition of their specific transcriptomic signature.

We analyzed slides from nine patients: six MGC^High tumors and three MGC^Low (Fig. 3C–H). Unsupervised clustering revealed seven distinct clusters of cells that matched manual morphology-based annotations (Supplementary Table S8). When projected onto a Uniform Manifold Approximation and Projection for Dimension (UMAP) space, MGC formed a clear distinct cluster (Fig. 3I). Of note, MGC were the only cell type with significantly different abundance between MGC^High and MGC^Low tumor samples (Fig. 3J). To narrow MGC signature, spots only covering MGC were manually selected (Supplementary Fig. S5A) and were shown to overlap with the unsupervised analysis (Supplementary Fig. S5B). Differentially expressed genes (DEG) between annotated MGC spots and all non-MGC spots included CHIT1, FBP1, MMP9, SPP1, APOE, CHI3L1, CTSS, CTSB, CTSD, CTSZ, TYROBP, CD68, DC-STAMP, MARCO, and TREM2 confirming their macrophage nature (Fig. 3K; Supplementary Table S9). Gene Ontology (GO) analysis of this supervised signature showed upregulation of pathways involved in bone resorption, tissue remodeling and extracellular matrix disassembly, as well as macrophage activation, phagocytosis, antigen presentation, and ROS production (Supplementary Fig. S5C).

To validate this signature, we investigated the transcriptomic differences between MGC^High and MGC^Low tumors from TCGA patients with oral cavity SCC. We retrieved the bulk RNA sequencing data (bulk RNA-seq) for 108 patients (out of the initial 110 patients, two did not have RNA-seq data) for which we had quantified MGC density on WSI. DEG analysis showed a clear difference between the bulk signatures of MGC^High and MGC^Low tumors (Fig. 3L). GO analysis of TCGA RNA bulk signature also highlighted pathways common to the Visium-defined MGC signature, such as bone resorption, tissue remodeling, and macrophage differentiation (Supplementary Fig. S5D). GSEA analyses showed a significant enrichment of the spatial transcriptomic MGC signature in MGC^High TCGA patients (Fig. 3M; Supplementary Fig. S5E). Among genes that were more highly expressed in MGC^High tumors were macrophage-related genes (SPP1, MARCO, DC-STAMP, OSCAR) and MGC-specific genes (CHIT1, FBP1), which confirmed our spatial transcriptomic data (Supplementary Fig. S5F). Because the prognostic impact of CHIT1 (16) and TREM2 (8) was recently reported, we investigated their potential influence and showed that CHIT1 and TREM2 were associated with a longer OS in TCGA oral cavity SCC cohort (Supplementary Fig. S5G). Of note, RNA expression of CK5, CHIT1, and CD68 were validated at the protein level by immunohistochemistry (Supplementary Fig. S6A–S6C).

Taken together, these results revealed the core macrophage signature of MGC and suggested a potential antitumoral function through the expression of genes and proteins associated with phagocytosis, antigen presentation, keratin resorption, and tissue remodeling.

We then investigated the composition of the TME associated with MGC presence using the CosMx 64-plex human immuno-oncology protein panel (Nanostring) on four MGC^High and four MGC^Low patients. We identified nine general cell clusters (Supplementary Fig. S7A–S7D). MGC exhibited a specific macrophage signature (Supplementary Fig. S7D). MGC^Low patients showed similar clusters, except for MGC (Supplementary Fig. S7E–S7G). Within the immune population, we found 14 cell subsets in MGC^High patients (Fig. 4A–D) and 12 in MGC^Low patients (Supplementary Fig. S7H and S7I). Notably, TCM CD4 T cells increased in MGC^High patients, clustering together in close contact with B cells (Fig. 4D–G). No significant differences were seen in other cell subsets. However, we observed in MGC^High patients, a noticeable but nonsignificant trend toward increased infiltration of B cells and plasma cells that were clustering together (Fig. 4H and I).

Additionally, we observed a marked increased in a mononuclear macrophage subset sharing a signature and localization with MGC (Fig. 4D, G, and J; Supplementary Fig. S8), termed MGC-MNP (Fig. 4K). To characterize deeper MGC and MGC-MNP, we performed a second multiplex imaging panel containing CHIT1, CD163, CD68, and TREM2 antibodies, in nine patients (four MGC^High and five MGC^Low). MGC and surrounding MGC-MNP expressed CHIT1 and TREM2 proteins (Fig. 4L and M; Supplementary Fig. S9A–S9C; Supplementary videos S1 and S2). These TREM2-expressing MGC-MNP were clustering with MGC in keratin-rich niches (Fig. 4N), leading us to speculate that they might be MGC precursors.

Herein, we observed in two cohorts of patients with HNSCC from TCGA (n = 110 patients) and GR (n = 284 patients) that the density of MGC associates with a longer OS and PFI. As in treatment-naive patients (n = 394), the association was also observed in preoperative chemotherapy-treated (n = 52) patients. To improved patient stratification, we developed a deep-learning approach to automatically quantify the MGC-to-tumor-cell ratio on standard H&E/HES WSI, which proved as effective and faster than manual pathologist assessment. We then investigated the MGC-related TME using spatial transcriptomic and proteomic: we found an increase in CD4 TCM T cells and TREM2^High TAM that share a signature and localization with MGC in keratin-enriched tumor niches and correlate with MGC abundance.

MGC accumulation is well documented in response to foreign body (17). Indeed, following its introduction, macrophages and monocytes fused to form foreign-body giant cells (17). In HNSCC, the carcinoma synthesizes the keratin that is seen as a foreign body by the immune system (10). However, it is not the presence of keratin per se, but rather of keratin-associated MGC that is associated with longer survival. Understanding why only a fraction of patients with keratinizing SCC have high MGC densities could reveal the missing steps of the foreign body reaction toward keratin, providing insights into the drivers of this antitumor program and paving the way for new treatments.

The presence of MGC in SCC after preoperative therapy has been long known (18) and is part of the histological scoring for pathological response following antitumor therapy (19). However, only recently a study in esophageal SCC highlighted MGC as the only factor associated with prolonged OS in multivariate analysis after preoperative therapy (20). In addition, two other recent studies suggest a positive impact of MGC in treatment-naive patients with SCC of the esophagus or tongue (12, 13). Here, we confirmed these observations and defined the first evidence-based threshold for the relationship between MGC density and patient risk, demonstrating its potential for informing therapeutic decisions.

Although the clinical relevance of MGC is emerging, their biology remains poorly understood. We analyzed the whole transcriptomes of individual MGC in HNSCC, revealing their specific “single giant cell RNA signature,” that share strong similarities with TREM2^High mononuclear macrophages. TREM2^High TAM are known to be protumoral in several cancers (8), including colorectal (21), breast (21–23), and lung (24) adenocarcinomas and are more abundant in patients with nonresponding melanomas treated by immunotherapy (25). Their deletion enhances the efficacy of immunotherapy in a model of ovarian adenocarcinoma (26). Interestingly, neither adenocarcinomas nor melanomas produce keratin. In contrast, we report a good prognosis for TREM2^High MGC in HNSCC—a carcinoma producing extracellular keratin. We found that TREM2^High TAM and MGC clustered in keratin-containing tumor niches, suggesting that keratin creates specific intratumoral niches leading to the accumulation of TREM2^High TAM, which then fuse into anti-tumoral MGC. These opposite prognosis values of TREM2^High macrophages across cancers highlight the specificity of each tumor type and underline that the protumoral and antitumoral features of macrophage populations cannot be generalized. Notably, it urges caution for current clinical trials of anti-TREM2 therapies in keratinizing SCC as they could hamper patient prognoses.

CHIT1 expression was shown to be associated with good prognosis in bulk RNA-seq data from three HNSCC cohorts (16) and elevated in the blood of good responder patients with HNSCC after radiotherapy (27). In addition, a population of CHIT1- and TREM2-expressing macrophages associated with complete response to neoadjuvant immunotherapy in patients with esophageal SCC (28). Here we confirmed the favorable impact of CHIT1 and TREM2 on OS and identified MGC as a major source of CHIT1 production via single giant cell analysis.

In line with previous results (29, 30), we showed that MGC^High patients had high levels of TILs. We then explored the contribution of each lymphoid cell subset using high-throughput multiplex analysis. We highlighted an increase in TCM CD4 T cells in MGC^High patients, which were found in proximity to B-cell clusters. Notably, a recent study highlighted the expansion of TCM CD4 T cells, B cells, and plasma cells in patients with HNSCC treated with a combination of anti-CTLA4 and anti-PD-L1 (31). Furthermore, tumor-infiltrating TCM CD4 T cells have been associated with favorable prognosis in OSCC (32). Collectively, the increase of these immune cell ecosystems, composed of TCM CD4 T cells, B cells, and plasma cells, may contribute to the favorable prognosis in patients with high MGC density.

The digitization of pathology laboratories opens the way for AI algorithms to improve diagnostic and prognostic (33). Manual quantification of MGC is time-consuming for overwhelmed pathologists (34) and impossible to apply in routine. Thus, we designed a deep learning model to quantify the MGC-to-tumor cell ratio automatically on routine diagnostic WSI stained by H&E/HES, without the need for additional costly techniques such as immunohistochemistry or genetic testing. This approach is faster, yet equally efficient as pathologists, and effectively stratifies patients. Pending further validations in additional HNSCC cohorts from different centers, this method could identify patients in need of therapeutic optimization based on predicted risk level.

Lastly, because SCC is widely spread across the body, we took advantage of our deep learning model to screen for the presence of MGC in laryngeal SCC and CESC. Consistent with previous esophageal publications (11, 20), MGC density indicated favorable prognosis in SCC of Larynx and Uterine cervix. Robust validation on external cohorts is required.

To conclude, TREM2^High MGC are major contributors to good prognosis in treatment-naive and preoperatively treated SCC, and their monitoring will be of tremendous value to improve the clinical management of patients suffering from HNSCC.

Data from 527 patients with HNSCC were retrieved from the TCGA website. Exclusion criteria were patients with only frozen sections, low quality slides, no diagnostic slides, absence of surgical resection, only biopsies, HPV-positive tumors, or tumor localization elsewhere than oral cavity: this led to exclusion of 417 patients, leaving data from 110 patients eligible for further analyses (Supplementary Table S1). These patients were considered “treatment-naïve” because they did not receive preoperative therapy. All slides were stained by H&E. We had access to the bulk RNA-seq data for 108 of these patients. These patients had a median follow-up of 34 months (interquartile range 21; 56 months).

Data from 307 patients with uterine cervix carcinoma and from 527 patients with HNSCC were retrieved from TCGA. Exclusion criteria for uterine cervix were patients without squamous cell carcinoma, frozen sections, low quality slides, and no diagnostic slides: this led to the exclusion of 118 patients, leaving data from 189 patients eligible for further analyses. Exclusion criteria for Larynx were frozen sections, low quality slides, no diagnostic slides, and tumor localization elsewhere than larynx: this led to exclusion of 427 patients, leaving data from 100 patients eligible for further analyses. All slides were stained by H&E.

We had access to samples from two previously published cohorts of patients with HNSCC treated at GR (14, 15). These patients did not receive preoperative therapy and were thus considered as “treatment-naive.” Exclusion criteria were patients with not enough representative pathological diagnostic slides, low quality slides, or the presence of neoadjuvant therapy: out of 419 patients, 284 were included (Supplementary Table S1). These patients had a median follow-up of 113 months (interquartile range 81; 150 months). We further had access to a retrospective cohort of 52 patients with HNSCC treated by ICT at GR. All FFPE blocks and slides came from the pathology laboratory of GR and were stained by HES. All patients gave a written informed consent. For each patient, the study protocol was approved by the institutional review board of GR (IRB-2023-293). All experiments were in accordance with the Declaration of Helsinki. The REporting recommendations for tumor MARKer prognostic studies were followed.

A pathologist quantified the MGC per mm² of tumor for each H&E/HES slide from each patient from TCGA (n = 110 slides) and GR cohorts (n = 1,304 slides). A second pathologist quantified the MGC per mm² of tumor for a subset of the patients (TCGA n = 10 slides; GR n = 20 slides): the counts from the two pathologists showed almost perfect agreement (r² = 0.9606, P value<0.0001 and weighted kappa = 0.876; Supplementary Fig. S10A and S10B).

MGC-related stratification was based on two thresholds for manual quantification: MGC^High (≥1 MGC/mm²), MGC^Int (≥0.2 MGC/mm² and <1 MGC/mm²), and MGC^Low (<0.2 MGC/mm²). The threshold of 0.2 was selected using the maximally selected rank statistics, on the OS endpoint in the combined TCGA and GR cohorts (n = 394 patients). The threshold of 1 was selected to obtain a group of more than 50 patients with very high MGC densities.

The annotations were carried out with QuPath version 0.4.3, an open-source software for digital pathology images analysis (35).

Stratification of patients treated by ICT into good, partial, and poor responders was made by a pathologist using the original grading system published by Braun and colleagues (19).

Two pathologists quantified the TILs on H&E/HES slide for each TCGA (n = 110 slides) and GR patients (n = 1,304 slides). TILs were evaluated according to a scoring method developed by the International Immuno-Oncology Biomarkers Working group (36). TILs were scored as the percentage of surface occupied by lymphocytes. Only the tumor and tumor-related stroma were included in the analysis. The percentage of TILs was assessed semiquantitatively from 0% to 100%. Challenging cases were reviewed by the two pathologists until an agreement was reached.

Two pathologists quantified the tumor grade on H&E/HES slides for each TCGA (n = 110 slides) and GR patients (n = 1,304 slides). Grade was evaluated according to the fifth edition of the WHO that is based on the similarity to healthy squamous epithelium as described by Broder in 1920 (37). Squamous cell carcinoma was graded in well (G1: grade 1), moderate (G2: grade 2), and poorly (G3: grade 3) differentiated tumors. Challenging cases were reviewed by the two pathologists until an agreement was reached.

We retrieved the 1,304 physical microscopic slides from the two cohorts of treatment-naïve patients from GR (14, 15). Multiple slides per patient were available (range 1–17). While most of the slides had a standard microscopic size (75 × 25 mm), several of them were larger (75 × 50 mm) and did not physically fit into the digital slide scanner. Therefore, all large microscope slides were removed from the analysis and 44 patients with large slides only were excluded from the digital analysis. Quality control was performed by a pathologist to exclude slides with common artifacts such as tissue folds, blurring or faint staining, resulting in the exclusion of nine additional patients. Overall, 839 slides were analyzed, representing tumors from 231 patients (of the original 284 included patients). Slides were digitized at 20× resolution, with an Olympus VS120.

Slides from TCGA cohort were already digitized and one slide per patient was quantified (n = 110 patients and n = 110 slides).

To train the MGC detection model, 16 oral cavity WSI from TCGA and 12 oral cavity WSI from the GR cohort were retrieved. A pathologist annotated a total of 5,000 MGC by drawing bounding boxes using QuPath software version 0.4.3. Rather than exhaustively annotating all MGC in a small number of slides, we opted to maximize the variability by nonexhaustively annotating some of the MGC on many slides. The annotation instructions were to draw bounding boxes around each MGC and then to draw an englobing bounding box that ensured that all MGC were exhaustively annotated in the latter. As such, all pixels outside the englobing boxes were not used to train the detection model. We randomly annotated additional negative patches, i.e., patches that contained no MGC, by drawing an englobing bounding box. All englobing bounding boxes were extracted with both tissue image and annotations using an in-house software at 20×, resulting in fields of view (FOV) with varying size and with exhaustively annotated MGC. All FOV of 20 WSI were randomly assigned to the training set and the remainder FOV from eight WSI to the validation set. All the training and validation slides from the GR cohort that were used to train the MGC detection model were discarded from the rest of the study.

The same protocol was applied to obtain ground-truth annotations for training the tumor detection model. In total, 814,606 tumor cells were collected from 15 WSI of TCGA tongue and 52 WSI of the GR cohort. Additionally, a pathologist exhaustively annotated the viable tumor cells in 107 WSI from the GR cohort, ensuring that there were not any tumor cells outside the annotated regions. Patches outside the annotated regions were added to the training data with an associated empty ground-truth since these patches do not contain tumor cells. The final training data contained 13,624 256 × 256 pixels patches with at least one tumor cell and 106,000 patches without tumor cells. All FOV of 47 WSI were randomly assigned to the training set and FOV from the remainder 20 WSI to the validation set. Here again, all the training and validation GR slides for the tumor detection model were discarded from the rest of the study.

We trained the Fully Convolutional One-Stage (FCOS; ref. 38) detection architecture for both tumor cells and MGC automatic annotation. Contrary to region proposal models, detections of FCOS are made at the pixel level: each pixel notably produces four values that are reconstructed as a predicted bounding box and an additional probability value for each class—1 in our case for both tumor and MGC. During training, each proposed bounding box is first matched with one ground-truth bounding box or no bounding box. Then, a regression loss and a classification loss are computed based on the predicted box coordinates and its associated probability. We used the ResNet50 architecture (39) pretrained on ImageNet (40) as the backbone model. Both classification and regression heads comprised four ReLU-activated convolutional layers of kernel size 3 and stride 1, with batch normalization. Contrary to the original implementation of FCOS, image rescaling was discarded to input patches with the same magnification as the model. Data augmentation consisted of rotations, flips, shifts, and color jitter. For each input patch, the same data augmentation was performed to the associated ground-truth bounding boxes using the Albumentations (41) Python library version 1.2.1. The detection model parameters were stochastically updated using the Adam optimizer (42) from errors computed by the focal loss (43) cost function with a learning rate of 10⁻⁴, a regularization of 10⁻⁴ and a batch size of four patches. Training was conducted for up to 2,000 epochs on a NVIDIA A40 GPU. We selected the model snapshot with the maximum mean average precision at intersection-over-union of 0.5 (mAP@50) on the validation set for both tumor and MGC tasks for the remainder of our study. All deep learning implementation was done in Python 3.8, PyTorch (44) version 1.13, and Torchvision version 0.14.0.

We assessed the performances of the MGC and tumor cell detection model. Briefly, ground truth MGC and tumor cells (annotated by a pathologist) were compared to automatically detected MGC and tumor cells and the average precision was computed using the MMdetection package version 3.1.0 (arXiv 1906.07155). To assess further the performance of the MGC detection model, we compared the number of automatically detected MGC to the number of MGC that were quantified by a pathologist (Fig. 2C–F).

Once both tumor detection and MGC detection models were trained, the biomarker was computed for each WSI as follows. Patches of width 2,048 pixels were extracted from a WSI in a sliding window fashion with an overlap of 256 pixels for both sides. We used a non-zero overlap to remove border artifacts (such as false positives) arising from a lack of context at the borders. Each patch was forwarded into both the tumor and MGC models, producing two lists of predicting bounding box coordinates and their predicted probabilities. Bounding boxes predicted with a probability below 0.4 were discarded and we applied non-maximum suppression with a threshold of 0.6 intersection-over-union to remove highly overlapping predictions. We finally computed the proposed MGC biomarker as the ratio of the number of predicted MGC to the number of predicted tumor cells for each patient.

The Visium Spatial Gene Expression Slide, Visium FFPE Reagent Kit, and Visium Human Transcriptome Probe Kit (10x Genomics) were used to generate sequencing libraries. RNA quality was accessed for all samples using the DV200 method. For library construction, 5 μm cryosections of 6.5 × 6.5 mm from each FFPE sample were placed into the four capture areas of the Spatial Transcriptomics slides (10x Genomics). Samples were then deparaffinized, stained, imaged, and decrosslinked. Probe hybridization and ligation to RNA were then performed, followed by the single-stranded ligation product release and extension. Libraries were then prepared following the manufacturer’s instructions. All the libraries were sequenced with a 10-base index read (dual index), a 28-base Read1 containing cell-identifying barcodes and unique molecular identifiers (UMI), and a 90-base Read2 containing transcript sequences on an Illumina NovaSeq 6000.

All sections were processed individually using the Seurat package, with spots containing less than 500 UMI and 500 RNA counts were discarded from further analyses. Afterward, all slides were merged into one object used for DEG analysis and integration was based on the slide ID. For this, reciprocal principal component analysis (PCA) integration based on 30 first PCA dimensions, with k.anchor = 20 was performed. The Loupe software (10x Genomics) was used to explore the data.

To identify the DEG between MGC spots and non-MGC spots, we used the R package Seurat, and the volcano plot was plotted using the Enhanced Volcano R package. Only the genes with a log2 fold change (log₂FC) over 0.25 and an adjusted P value less than 0.05 were considered significant.

Bulk RNA-seq data of primary tumors from 108 out of the 110 patients with tongue SCC were used for survival analysis. Data were retrieved using TCGABiolinks R package, then normalized using the EDAseq package based on gene length; genes whose expression was zero in more than 25% of the samples were removed.

The 108 patients with available histological slides were split into three groups based on the presence of MGC as annotated by a pathologist. DEGs were calculated between groups using the R package DESeq2, and the volcano plot was plotted using the Enhanced Volcano R package. Only genes with a log₂FC higher than 0.25 and an adjusted P value less than 0.05 were considered significant.

The patients were separated based on the pathology score of MGC^High and MGC^Low. Differential expression analysis was performed with DESeq2, Logarithm fold changes were modified by the lfcShrink algorithm with the “Ashr” option. DEG cutoff was defined at an FDR less than 0.05 and an absolute log₂FC value greater than 1 in pairwise comparisons. R package clusterprofiler (4.12) was used to analyze the DEGs between different groups for GO analyses. For the Visium GO analysis, only genes with log₂FC superior to 1.5 were considered.

The analysis was made in the interactivenn website (45). The top DEG whose adjusted P value was lower than 0.05 and log₂FC higher than 0.25 were extracted from TCGA bulk RNA-seq data and from the Visium MGC signature.

Nine FFPE blocks of HNSCC (four MGC^High patients and five MGC^Low patients) were stained by multiplex immunofluorescence. The Experimental and Translational Pathology Platform of GR performed the stainings using a preset routine protocol. Briefly, FFPE blocks were cut into 3-μm-thick sections. Multiplex staining was performed on a Bond RX (Leica Biosystems). Slides were baked in high pH for 20 minutes at 100°C for epitope retrieval. Then, multiple rounds of staining were performed. Each round included endogenous peroxidase blocking for 10 minutes, nonspecific sites blocking for 5 minutes, primary antibody incubation, secondary HRP-labeled antibody incubation for 10 minutes and OPAL reactive fluorophore incubation for 10 minutes to covalently label the primary epitope, and heat denaturation of antibodies. The sequence of the antibodies with the associated Opal dyes is as follow: (i) rabbit anti-TREM2 (clone D814C; Cell Signaling, reference 91068), 1/400 dilution, 1 hours, 37°C, with OPAL 520, 1/100 dilution; (ii) rabbit anti-CHIT1 (polyclonal; Biorbyt, reference orb377995), 1/50 dilution, 30 minutes, ambient temperature, with OPAL 570, 1/100 dilution; (iii) mouse anti-CD68 (clone KP1; DAKO, reference M0814), 1/1,000 dilution, 1 hours, ambient temperature, with OPAL480; (iv) mouse anti-CD163 (clone 10D6; Diagnostic BioSystems, reference Mob 460-05), 1/600 dilution, 15 minutes ambient temperature, with OPAL690, 1/100 dilution. Nuclei were visualized by a final incubation with the Spectral DAPI (Akoya Biosciences) for 5 minutes. Slides were mounted with mounting medium for fluorescence. Finally, images were acquired on the Polaris 2 (Akoya Biosciences).

Eight FFPE blocks of HNSCC (four MGC^High patients and four MGC^Low patients) were stained by multiplex immunofluorescence. The stainings were performed with the CosMx SMI 64-plex Human Immuno-Oncology Protein Panel. For each patient, two serial slides were cut: one slide stained by immunofluorescence and one slide stained by HES.

HNSCC tissue samples were cut into 500-μm-thick sections using a custom-made tissue slicer. Samples were permeabilized and blocked using 0.3% Triton-X-100 (Sigma) in PBS (Gibco) including 5% rat serum (Sigma) overnight at room temperature. Same buffer was used for immunostaining with anti-TREM2 followed by AlexaFluor-conjugated Fab Fragment Goat Anti-Rabbit IgG (Jackson ImmunoResearch), anti-CD68 (clone Y1/82A; BioLegend), and Hoechst (Biolegend) for 3 days at room temperature. After rigorous washing, samples were cleared for 3 hours using Ce3D (BioLegend). Imaging was performed using a Zeiss LSM 880 confocal microscope with a Plan-Apochromat 20×/1,0 objective at Nyquist rate. Individual channels were deconvolved using Huygens 22.10 (Scientific Volume Imaging) at default settings. Visualizations were done using ZEISS arivis Pro/Vision4d (Zeiss).

A pathologist manually annotated the tumor, the keratin, and the technical artifact regions on the nine immunofluorescence images, using QuPath v0.4.3 (35). In addition, a small proportion of TREM2-expressing mononuclear macrophages and MGC were manually annotated by a pathologist. These annotations were used to train a pixel classifier based on a random forest model. The classifier model was applied to predict the presence of TREM2-expressing MGC and mononuclear macrophages, in the tumor regions of the nine immunofluorescence images. Measurements of the areas and distances of these regions to the closest keratin-rich regions were exported and analyzed with a homemade R script (https://www.r-project.org) to produce graphics of surface ratios and distances to keratin regions.

A pathologist manually annotated the MGC and the tumor regions. Cell annotation was based on the combination of protein expression and cell localization on the HES slide. Images were processed and analyzed using the SOPA pipeline (46). Annotation of the cells was performed in two layers, focusing on the immune fraction of the slide. For the neighborhood analysis, we used the squidpy analysis (47).

Among the 284 patients from GR, 264 of them had available FFPE blocks. One representative slide per patient was selected by a pathologist. Patients without available FFPE blocks were excluded from the analysis (n = 20). Slides were stained with anti-CD68 antibody (clone KP1, DAKO, reference M0814), 1/1,000 dilution, 1 hours, 36°C.

For each patient analyzed by Visium, three serial slides were stained with anti-CHIT1 antibody (polyclonal; Biorbyt, reference orb377995), 1/50 dilution, 30 minutes, 36°C; anti-CD68 antibody (clone KP1; DAKO, reference M0814), 1/1,000 dilution, 1 hour, 36°C;and anti-CK5 antibody (clone SP27; Roche, reference 760-4935), prediluted antibody (0,17 µg/mL), 32 minutes, 36°C.

The Experimental and Translational Pathology Platform of GR performed the stainings using a preset routine protocol. Briefly, FFPE blocks were cut into 3-μm-thick sections. Staining was performed on an ULTRA BenchMark (Roche Diagnostics). After deparaffinization, slides were baked in CC1 solution (pH 8, 20 minutes, 95°C) for epitope retrieval. Endogenous peroxidase activity was inhibited with peroxidase blocking reagent (UltraView DAB kit). Slides were then incubated with a primary antibody followed by a secondary antibody. 3,3′-Diaminobenzidine tetrahydrochloride was used for revelation and hematoxylin was used for counterstaining (UltraView DAB kit). Finally, slides were mounted with mounting medium (Pertex) and glass coverslips.

A pathologist annotated the tumor region in each slide, and the nontumoral regions were excluded from the analysis. The intensity of the CD68 staining per tumor surface was quantified. The intensity of CD68 in MGC was removed from the analysis. The analysis was performed with QuPath v0.4.3.

OS and PFI were the two clinical endpoints used for the analysis. Log-rank test and log-rank test for trend were performed. Multivariable survival analyses were also performed using Cox model and P value of Wald test was presented. Linear relationship between the log-hazard and MGC density and square root of MGC density were tested with a smoothing spline, using the pspline function of coxph in R survival package. For survival curves graphical representation, data were truncated at 2,000 days; however, tests were performed with all the length of the follow-up.

Clinical data were retrieved from the GDC Data Portal (https://portal.gdc.cancer.gov/). OS and PFI were retrieved from an integrated TCGA pan-cancer clinical data resource (48) in which the authors recommended using OS and PFI as clinical endpoints. PFI was defined as follows: “1 for patient having a new tumor event whether it was a progression of disease, local recurrence, distant metastasis, new primary tumors all sites, or died with the cancer without new tumor event, including cases with a new tumor event whose type is N/A. 0 for censored otherwise” (48).

Clinical data of treatment-naive patients were retrieved from the two ancillary annotated cohorts (14, 15). PFI was defined as follows: “events for progression or 2^nd HNSCC or death by HNSCC or death of unknown cause with HNSCC”. GR clinical data of ICT patients were retrieved from the GR patient’s database.

WSI were generated from publicly available cohorts (TCGA) and from institutional cohorts (GR). The digital WSI from TCGA are publicly accessible through the NIH Genomic Data Commons Data Portal https://portal.gdc.cancer.gov/. The digital WSI generated from GR cohorts are not publicly available due to general data protection regulations and institutional guidelines. Availability could be made upon request to the lead contact after execution of a data transfer agreement with GR.

Data from TCGA can be retrieved from their website.

All generated Visium RNA-seq data are available and have been deposited in the functional genomics data collection (Array Express) under accession number E-MTAB-14409 (https://www.ebi.ac.uk/biostudies/arrayexpress/studies/E-MTAB-14409?key=dcc24ea4-8654-4f29-a4b1-26cdb071a1a5).

Availability could be made upon request to the lead contact after execution of a data transfer agreement with GR.

The source code of the bulk RNA-seq data, Visium data, and CosMx data analyses is available on GitHub (https://github.com/AhmedAmineAnzali/MGC_Paper_Analysis). The source code of the AI platform is available on GitHub (https://github.com/mgc-ai-cancer-discovery/mgc-ai-cancer-discovery). All methods and software packages used in the study have been documented and explained in ways accessible to the broader scientific audience.

Further information and requests for resources and code should be directed to and will be fulfilled by the corresponding author.

G. Gessain reports grants from Institut National du Cancer—PRT-K 2022 and Prix Fonds Recherche 2021 de la Société Française de Pathologie during the conduct of the study; in addition, G. Gessain has a patent EP24305278.4 pending. M. Lerousseau reports other support from Spotlight Medical outside the submitted work; in addition, M. Lerousseau has a patent EP24305278.4 pending to Gustave Roussy. A. Auperin reports grants from Government: PRT-K 2022 during the conduct of the study; other support from MerckSD outside the submitted work. S. Uderhardt reports grants from Deutsche Forschungsgesellschaft DFG during the conduct of the study. F. Ginhoux reports a patent for Products and Uses Thereof for Predicting the Sensitivity of a Subject to Cancer Immunotherapy pending. No disclosures were reported by the other authors.

G. Gessain: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. A.-A. Anzali: Data curation, formal analysis, methodology, writing–review and editing. M. Lerousseau: Software, formal analysis, investigation, methodology, writing–review and editing. K. Mulder: Software, formal analysis, investigation, methodology, writing–review and editing. M. Bied: Formal analysis, investigation. A. Auperin: Formal analysis, funding acquisition, investigation, methodology, writing–review and editing. D. Stockholm: Formal analysis, investigation, methodology. N. Signolle: Formal analysis. F. Sassi: Data curation, formal analysis. M.E. Marques Da Costa: Resources, data curation. A. Marchais: Resources, data curation. A. Sayadi: Formal analysis. D. Weidner: Formal analysis. S. Uderhardt: Formal analysis. Q. Blampey: Formal analysis, methodology. S.R. Nakkireddy: Formal analysis, methodology. S. Broutin: Resources. C.-A. Dutertre: Formal analysis. P. Busson: Resources. T. Walter: Funding acquisition, methodology. A. Marhic: Resources. A. Moya-Plana: Resources. J. Guerlain: Resources. I. Breuskin: Conceptualization, resources, project administration. O. Casiraghi: Conceptualization, resources, project administration. P. Gorphe: Conceptualization, resources, funding acquisition, project administration, writing–review and editing. M. Classe: Conceptualization, resources, project administration. J.-Y. Scoazec: Conceptualization, resources, funding acquisition, writing–review and editing. C. Bleriot: Conceptualization, resources, formal analysis, supervision, investigation, project administration, writing–review and editing. F. Ginhoux: Conceptualization, resources, supervision, funding acquisition, investigation, methodology, project administration, writing–review and editing.

G. Gessain was supported by the grant “Année-Recherche en Médecine”, the grant for DUERTECC (“Diplôme Universitaire Européen de Recherche Translationnelle Et Clinique en Cancérologie”) and the “Prix Fonds Recherche 2021 de la Société Française de Pathologie”. S. Uderhardt is supported by DFG (project-IDs 448121430, 405969122, 505539112), the Hightech Agenda Bavaria and by an ERC starting grant (project-ID 101039438). Volumetric tumor imaging was performed on a DFG-funded confocal microscope (project-ID 261193037). F. Ginhoux is supported by the Fondation Gustave Roussy and Fondation ARC—“Recruiting International Leaders 2020”. This work was supported by the “Programme de Recherche Translationnelle en Cancérologie 2022 Institut National du Cancer PRT-K 2022”. T. Walter received funding from ITMO Cancer (20CM107-00).