Title: AI-based quantification of tumor-infiltrating lymphocytes with integrative transcriptomics in ovarian clear cell carcinoma: JGOG3025-TR1/A1 study.

The JGOG3025 study ( NCT03159572 ), conducted by the Japanese Gynecologic Oncology Group (JGOG), prospectively collected clinical data and targeted DNA sequencing data for 51 genes from patients with epithelial ovarian cancer whose treatment commenced between March 2017 and March 2019 across 55 affiliated institutions [ 34 ]. The targeted 51 genes were listed in Supplementary Table 1. A total of 701 cases were included in the JGOG3025-whole cohort following central pathological review. Of these, 189 cases were diagnosed as clear cell carcinoma (CCC), comprising the JGOG3025-TR1 cohort. As a correlative study, JGOG3025-A1 study involved 28 centers and acquired H&E-stained slides from 400 cases, and the slides were digitally scanned with a NanoZoomer Digital Pathology System (Hamamatsu Photonics, Hamamatsu, Japan) at × 20 magnification (resolution of 0.5 mm per pixel). The slide with the largest tumor area was selected as the representative slide when multiple slides were available. Among them, 102 cases overlapped with the JGOG3025-TR1 cohort and 98 cases were incorporated into the present analysis after excluding four cases with insufficient tumor tissue (< 10 mm 2 ) (JGOG3025-A1 cohort). The Kyoto cohort consists of 38 WSIs from 38 patients with CCC from Kyoto university hospital. Clinical characteristics of the JGOG3025-TR1 and Kyoto cohorts are shown in (Table  1 ). Overall, a total of 136 patients had WSIs available for AI-based TIL quantification. The Washington cohort consisted of 198 CCC cases with available clinical and RNA sequencing data, which were obtained from the project repository ( https://github.com/kbolton-lab/Bolton_OCCC ) [ 28 ]. WSIs were not available for the Washington cohort. Table 1 Clinical characteristics of the JGOG3025-TR1 and Kyoto cohorts Cohort ALL Immune Non-immune Inflamed Non-inflamed Age Median(range) 56 (32–93) 54 (32–83) 56 (34–93) 53 (32–79) 57 (35–93) Stage I 148 57 91 30 63 II 21 7 14 3 8 III 43 26 17 9 12 IV 14 9 5 4 7 NA 1 0 1 0 0 PS 0 180 80 100 38 67  >  = 1 27 9 18 6 15 NA 20 10 10 2 8 Residual tumor 0 192 78 114 36 73    = 1 cm 24 14 10 6 13 Total 227 99 128 46 90 Clinical characteristics of the JGOG3025-TR1 and Kyoto cohorts Two state-of-the-art AI models were employed for nuclear detection and cell-type classification. CellViT is a vision transformer–based model with a convolutional backbone that was pretrained on large collections of histopathology tiles and subsequently trained on the PanNuke dataset, performs five-way classification into Neoplastic, Inflammatory, Connective/Soft tissue, Dead/Necrotic, and Epithelial cell types [ 30 , 31 ]. HoVer-NeXt is a convolutional neural network–based architecture that jointly predicts horizontal–vertical (HoVer) maps and semantic labels, and was trained on a modified version of the Lizard dataset for multiclass nuclear segmentation, providing seven-class output: Epithelial, Connective, Lymphocyte, Plasma, Neutrophil, Eosinophil, and Mitosis [ 33 ]. In this study, we used the publicly released pretrained weights and applied both models with their default configurations. Lymphocytes predicted by HoVer-NeXt were further classified as intratumoral TILs (iTILs) if located within tumor regions, and as stromal TILs (sTILs) if located within 1000 µm of the tumor boundary. Specifically, the 1000-µm (1-mm) peritumoral distance is recommended for ovarian cancer in a standardized TIL assessment framework [ 35 ], and was also adopted in our previous HGSC study [ 19 ]; because no CCC-specific recommendation is available, we applied the same threshold for consistency and comparability. TIL scores were calculated as the number of TILs divided by the corresponding area in each region. For 189 CCC cases in JGOG3025-TR1 cohort, total RNA was extracted from frozen tumor samples. RNA sequencing was performed on the Illumina NovaSeq 6000 platform with paired-end 150 bp reads. Samples were sequenced across 5 lanes of an S4 flow cell, generating on average 120 million reads and 18 gigabases per sample. For 38 CCC cases in Kyoto cohort, total RNA was extracted from frozen tumor samples. Sequencing was performed on an Illumina HiSeq 2500 platform with paired-end 100 bp reads, generating on average 70 million reads and 7 gigabases per sample. Raw reads were trimmed and preprocessed using Trim-Galore. The processed reads were aligned to the human reference genome GRCh38 using STAR (version 2.7.1a) and quantified with RSEM (version 1.3.1). The immune clustering was determined according to the methodology of the original publication [ 26 ]: the ssGSEA was performed based on immune-related pathways from the ImmPort database ( https://immport.niaid.nih.gov ), and Ward’s hierarchical clustering was performed on the z-normalized ssGSEA scores. The immune cell signatures employed in the ssGSEA were identical to those used in our previous study [ 19 ]. The other gene sets used in the analysis were downloaded from MsigDB ( https://www.gsea-msigdb.org/gsea/msigdb ) or obtained from original articles [ 36 , 37 ]. Tumor Immune Dysfunction and Exclusion (TIDE) scores were calculated as previously reported [ 38 ]. CIBERSORTx and xCell scores were calculated with default settings, as previously reported [ 39 , 40 ]. We performed consensus clustering to define immune and non-immune clusters within each cohort separately. Although we attempted batch correction across cohorts using ComBat-Seq, due to substantial inter-cohort differences (including platform and sample collection/processing conditions), batch correction did not achieve satisfactory harmonization. Therefore, to minimize batch-related confounding, we used rank-/proportion-based immune quantification methods (ssGSEA, CIBERSORTx, and xCell) rather than analyzing raw RNA-seq data for the integrated analysis of the different cohorts. Gene mutations were defined as originally reported [ 34 ]. MSI status was estimated from targeted DNA sequencing data, as previously reported [ 41 ]. TMB-high tumors were defined as tumors that had equivalent number of mutations as MSI-high tumors, as previously reported [ 41 ]. Statistical analyses were performed using Python 3.10. The correlation of continuous variables was assessed by Spearman's rank correlation coefficient. Patient characteristics were compared using Fisher’s exact test. Medians of continuous variables were compared between groups by the Mann‒Whitney U test or Wilcoxon signed-rank test. Cumulative survival probabilities were calculated using the Kaplan–Meier method. Univariate and multivariate Cox proportional hazards regression analysis and log-rank test were used to calculate p values, hazard ratios (HR), and 95% confidence intervals (CI). For Kaplan–Meier curves, patients were split by the upper tertile iTIL score into an immune-inflamed group and a non-inflamed group, as previously described [ 19 ]. Overall survival (OS) and progression-free survival (PFS) were defined as the time from initial treatment to death and recurrence or last follow-up. Two-sided p values < 0.05 were considered statistically significant in all analyses.

We first defined transcriptome-based immune clusters using RNA sequencing data. Hierarchical clustering was performed on immune pathway–related gene expression profiles, as described previously [ 26 ]. Consistent with previous studies, unsupervised clustering successfully reproduced two distinct subgroups across two cohorts (JGOG3025-TR1 and Kyoto cohorts) (Fig.  1 A). These two subgroups were designated as the immune and non-immune clusters, respectively, as originally reported. Previous study reported that the immune cluster is associated with poorer prognosis compared to the non-immune cluster [ 26 ]. Prognostic analysis in JGOG3025-TR1 and Kyoto cohorts revealed that the immune cluster had significantly worse OS compared to the non-immune cluster (hazard ratio, 2.54; 95% CI, 1.42–4.55; p  = 0.002), although the difference in PFS was not statistically significant (hazard ratio, 1.50; 95% CI, 0.95–2.36; p  = 0.083) (Fig.  1 B). Examination of patient characteristics within each cluster showed that the immune group had a significantly higher proportion of advanced-stage disease and residual tumor ( p  = 0.003 and p  = 0.041 by Fisher’s exact test, respectively) (Fig.  1 C). We therefore performed multivariate analyses incorporating transcriptome-based immune classification. The immune subtype was not associated with survival outcomes for both OS and PFS (hazard ratio, 1.37; 95% CI, 0.70–2.68; p  = 0.35 for OS. hazard ratio, 0.94; 95% CI, 0.56–1.56; p  = 0.81 for PFS) (Fig.  1 D). Fig. 1 Transcriptome-based immune subtypes in ovarian clear cell carcinoma. (A) Unsupervised clustering based on single-sample gene set enrichment analysis (ssGSEA) scores of immune-related gene signatures identified immune and non-immune subtypes, which were reproduced within each cohort (JGOG3025-TR1 and Kyoto). (B) Kaplan–Meier curves comparing overall survival (OS) and progression-free survival (PFS) between immune and non-immune subtypes. Both JGOG3025-TR1 and Kyoto cohorts are included. The horizontal dotted lines indicate median values. (C) Distribution of stage and residual tumor between immune and non-immune subtypes. The immune group showed a significantly higher proportion of advanced-stage disease and presence of residual tumor. (D) Multivariate Cox proportional hazards regression survival analysis for PFS and OS. The vertical line indicates the point of no effect. After adjustment for clinical covariates, the transcriptome-based immune cluster was not an independent prognostic factor. HR, hazard ratio; CI, confidence interval; PS, performance status Transcriptome-based immune subtypes in ovarian clear cell carcinoma. (A) Unsupervised clustering based on single-sample gene set enrichment analysis (ssGSEA) scores of immune-related gene signatures identified immune and non-immune subtypes, which were reproduced within each cohort (JGOG3025-TR1 and Kyoto). (B) Kaplan–Meier curves comparing overall survival (OS) and progression-free survival (PFS) between immune and non-immune subtypes. Both JGOG3025-TR1 and Kyoto cohorts are included. The horizontal dotted lines indicate median values. (C) Distribution of stage and residual tumor between immune and non-immune subtypes. The immune group showed a significantly higher proportion of advanced-stage disease and presence of residual tumor. (D) Multivariate Cox proportional hazards regression survival analysis for PFS and OS. The vertical line indicates the point of no effect. After adjustment for clinical covariates, the transcriptome-based immune cluster was not an independent prognostic factor. HR, hazard ratio; CI, confidence interval; PS, performance status To further validate this observation, we evaluated the Washington cohort [ 28 ]. Unsupervised clustering again successfully reproduced immune and non-immune subgroups in the Washington cohort (Supplementary Fig. 1A). The immune cluster was again associated with poorer prognosis (hazard ratio, 1.49; 95% CI, 0.96–2.32; p  = 0.074) (Supplementary Fig. 1B). In terms of patient characteristics, the Immune cluster again showed a higher prevalence of advanced-stage disease, although the difference was not statistically significant ( p  = 0.15) (Supplementary Fig. 1C). The immune subtype similarly showed no significant prognostic impact in multivariate analysis (hazard ratio, 1.10; 95% CI, 0.66–1.82; p  = 0.72) (Supplementary Fig. 1D). Next, we investigated iTIL and sTIL scores quantified from H&E slides. A schematic diagram illustrating the AI-based quantification of TILs is presented (Fig.  2 A). During preliminary evaluation, CellViT showed a consistent tendency to misclassify stromal fibroblasts as neoplastic cells throughout the dataset (Fig.  2 B). To address this, only cells predicted as “Neoplastic” by CellViT and simultaneously as “Epithelial” by HoVer-NeXt were defined as tumor cells. Tumor regions were delineated based on the spatial coordinates of these tumor cells (Fig.  2 C). Quantitatively, the tumor area after HoVer-NeXt correction was 84.7% of the CellViT-derived tumor area on average, indicating that 15.3% of the initially labeled tumor region was excluded (range, 1.6%–66.8%). Fig. 2 Artificial Intelligence (AI)-based cell-type inference and tumor region delineation on hematoxylin and eosin (H&E)–stained slides. (A) Schematic illustration of the analysis pipeline. Nuclear detection and cell-type classification were performed using two AI models (CellViT and HoVer-NeXt). (B) Representative images of cell-type inference performed by AI models. CellViT performs five-way classification (Neoplastic, Inflammatory, Connective/Soft tissue, Dead/Necrotic, and Epithelial), and HoVer-NeXt provides seven-class output (Epithelial, Connective, Lymphocyte, Plasma, Neutrophil, Eosinophil, and Mitosis). For visualization, only selected classes are colored: Neoplastic, Inflammatory, and Connective/Soft tissue for CellViT, and Epithelial, Connective, and Lymphocyte for HoVer-NeXt. Scale bars: 100 μm. (C) Representative images of tumor regions defined based on cell-type inference. Tumor regions defined using CellViT alone (left) and after integration with HoVer-NeXt (middle) are shown; the right panel is a higher-magnification view of the indicated area in the middle panel. Brightly colored areas indicate tumor regions. Combination of the two models enables more detailed tumor region segmentation Artificial Intelligence (AI)-based cell-type inference and tumor region delineation on hematoxylin and eosin (H&E)–stained slides. (A) Schematic illustration of the analysis pipeline. Nuclear detection and cell-type classification were performed using two AI models (CellViT and HoVer-NeXt). (B) Representative images of cell-type inference performed by AI models. CellViT performs five-way classification (Neoplastic, Inflammatory, Connective/Soft tissue, Dead/Necrotic, and Epithelial), and HoVer-NeXt provides seven-class output (Epithelial, Connective, Lymphocyte, Plasma, Neutrophil, Eosinophil, and Mitosis). For visualization, only selected classes are colored: Neoplastic, Inflammatory, and Connective/Soft tissue for CellViT, and Epithelial, Connective, and Lymphocyte for HoVer-NeXt. Scale bars: 100 μm. (C) Representative images of tumor regions defined based on cell-type inference. Tumor regions defined using CellViT alone (left) and after integration with HoVer-NeXt (middle) are shown; the right panel is a higher-magnification view of the indicated area in the middle panel. Brightly colored areas indicate tumor regions. Combination of the two models enables more detailed tumor region segmentation A total of 136 CCC cases with WSIs available from the JGOG3025-A1 and Kyoto cohorts were evaluated (Fig.  3 A). No significant correlations were found between iTIL scores and either clinical stage ( p  = 0.48) or residual tumor status ( p  = 0.79). First, we evaluated the association between iTIL score and gene expression data from the JGOG3025-TR1 and Kyoto cohorts. iTIL scores showed the strongest correlation with the T-cell signature (R = 0.32, p < 0.001), whereas no significant correlation was observed with the neutrophil signature (R = 0.15, p  = 0.08) (Fig.  3 B). Prognostic analysis revealed that, although not statistically significant, the immune-inflamed group defined by iTIL score showed a trend toward improved PFS (hazard ratio, 0.73; 95% CI, 0.39–1.36; p  = 0.32) and OS (hazard ratio, 0.86; 95% CI, 0.41–1.83; p  = 0.70) (Fig.  3 C), which was in contrast to the prognostic findings based on transcriptome-based immune classification. This favorable trend for the immune-inflamed subtype was consistently observed in subgroup analyses for both OS and PFS (Supplementary Fig. 2). Importantly, the benefit was significant for PFS in patients with residual tumor (hazard ratio, 0.21; 95% CI, 0.07–0.59; p  = 0.003) and for OS in those with advanced-stage disease (hazard ratio, 0.41; 95% CI, 0.17–0.99; p  = 0.05), indicating that iTIL score may provide prognostic discrimination particularly in high-risk populations. Multivariate analyses were performed incorporating pathology-based classification. The inflamed phenotype was a statistically significant favorable prognostic factor for both OS and PFS (hazard ratio, 0.31; 95% CI, 0.13–0.76; p  = 0.012 for OS. hazard ratio, 0.31; 95% CI, 0.15–0.65; p  = 0.002 for PFS) (Fig.  3 D). Fig. 3 Pathology-based TIL phenotypes and their association with immune transcriptional profiles and prognosis. (A) Clinical characteristics of patients in JGOG3025-TR1 and Kyoto cohorts. (B) Correlation between iTIL scores and gene expression profiles using ssGSEA scores calculated from immune cell-type gene signatures. The R value indicates the Spearman correlation coefficient. *P < 0.05, **P < 0.01, and ***P < 0.001. (C) Kaplan–Meier curves comparing OS and PFS between pathology-based inflamed and non-inflamed subtypes. The horizontal dotted lines indicate median values. (D) Multivariate Cox proportional hazards regression survival analysis for PFS and OS. The vertical line indicates the point of no effect. After adjustment for clinical covariates, the inflamed phenotype became an independent prognostic factor Pathology-based TIL phenotypes and their association with immune transcriptional profiles and prognosis. (A) Clinical characteristics of patients in JGOG3025-TR1 and Kyoto cohorts. (B) Correlation between iTIL scores and gene expression profiles using ssGSEA scores calculated from immune cell-type gene signatures. The R value indicates the Spearman correlation coefficient. *P < 0.05, **P < 0.01, and ***P < 0.001. (C) Kaplan–Meier curves comparing OS and PFS between pathology-based inflamed and non-inflamed subtypes. The horizontal dotted lines indicate median values. (D) Multivariate Cox proportional hazards regression survival analysis for PFS and OS. The vertical line indicates the point of no effect. After adjustment for clinical covariates, the inflamed phenotype became an independent prognostic factor We further investigated the discrepancy between transcriptome-based immune classification and histopathology-based classification. Although nearly half of the transcriptomic immune subtype cases were also classified pathologically inflamed (29/59, 49.2%), a subset of transcriptomic immune but pathologically non-inflamed cases showed minimal TIL infiltration on H&E sections (Supplementary Fig. 3A and B). Conversely, although most non-immune cases were non-inflamed by pathology (60/77, 78.0%), certain non-immune yet inflamed cases showed prominent TIL infiltration histologically (Supplementary Fig. 3A and C). These findings suggest that transcriptomic immune profiling does not necessarily correspond with morphological evidence of immune infiltration, but may rather reflect the intratumoral heterogeneity of immune cell distribution. We next focused on the immune group, which was characterized by elevated overall immune-related gene expression, to explore how inflamed and non-inflamed subtypes differ in relation to TIL abundance across various immune-related pathways. The inflamed group exhibited higher activated CD8 T cell scores and significantly greater expression of CTLA4-related pathway genes in ssGSEA ( p  = 0.041), as well as elevated IFNG-related scores (Fig.  4 A). Conversely, the non-inflamed group had higher TIDE Dysfunction scores and higher ssGSEA scores for HALLMARK_HYPOXIA. Fig. 4 Gene expression analysis within the immune group. (A) ssGSEA scores of immune signatures and Tumor Immune Dysfunction and Exclusion (TIDE) scores between inflamed and non-inflamed subtypes. (B) Cell fractions calculated by CIBERSORTx and xCell between inflamed and non-inflamed subtypes. Boxes in the box plot represent interquartile ranges, and horizontal lines represent 5th and 95th percentile ranges, with a notch for the median. P values were calculated using the Mann–Whitney U-test Gene expression analysis within the immune group. (A) ssGSEA scores of immune signatures and Tumor Immune Dysfunction and Exclusion (TIDE) scores between inflamed and non-inflamed subtypes. (B) Cell fractions calculated by CIBERSORTx and xCell between inflamed and non-inflamed subtypes. Boxes in the box plot represent interquartile ranges, and horizontal lines represent 5th and 95th percentile ranges, with a notch for the median. P values were calculated using the Mann–Whitney U-test We further evaluated immune cell composition using two deconvolution approaches, CIBERSORTx and xCell (Fig.  4 B). The inflamed group exhibited higher CD8 T cell scores by xCell, with a similar trend observed using CIBERSORTx. M2 macrophages tended to be enriched in the non-inflamed group, whereas M1 macrophages were more abundant in the inflamed group; accordingly, the M1/M2 ratio was significantly higher in the inflamed group by xCell ( p  = 0.002), with concordant results from CIBERSORTx. Neutrophils were enriched in the non-inflamed group, and analysis of the NLR revealed higher values in this group by CIBERSORT, again consistent with xCell results. Moreover, the StromaScore calculated by xCell was significantly higher in the non-inflamed group ( p  = 0.018). Previous studies have indicated that MSI-H and TMB-H tumors in clear cell carcinoma are associated with increased TIL infiltration [ 42 – 44 ]. In the JGOG3025-TR1 cohort (n = 189), three cases were identified as TMB-H, including one MSI-H tumor. All three were assigned to the immune group by transcriptome-based classification. For the two TMB-H cases with available WSIs, both demonstrated elevated iTIL and sTIL scores, and were categorized as Inflamed according to pathology-based classification (Supplementary Fig. 4A). These observations, however, should be interpreted as exploratory given the small number of TMB-H cases. Because ARID1A mutations have also been implicated in immune response [ 45 , 46 ], we further examined ARID1A status. No differences in TIL scores were observed between ARID1A -mutated and wild-type tumors (Supplementary Fig. 4B), and no significant correlations were found with immune cell ssGSEA scores except for plasma cell (Supplementary Fig. 4C). The ARID1A mutations were present in 48 cases (25.4%), with slightly higher frequency in the immune group (32.5%) than in the non-immune group (19.8%); however, this difference did not reach statistical significance ( p  = 0.063) (Supplementary Fig. 4D).

To our knowledge, JGOG3025-TR1 cohort is one of the largest multi-omics datasets that includes targeted DNA sequence and whole transcriptome sequence in CCC. In addition, diagnostic H&E slides were also collected, and the spatial distribution and prognostic relevance of TILs were investigated utilizing a single-cell–resolution cell classification model. Our findings suggest a possible discordance between immune phenotyping based on histopathology and that based on gene expression profiling. Several transcriptome-based classification systems have been proposed for CCC, and their prognostic significance has been reported. In context, Huang et al. analyzed 44 early-stage CCC cases using panel transcriptome sequencing, classifying tumors as “immune-hot” or “immune-cold”, and found that the immune-hot subtype was associated with poorer prognosis [ 23 ]. However, the immune-hot group included a significantly higher proportion of stage II cases, raising the possibility of stage-related confounding. Similarly, Ye et al. classified 50 CCC cases into immune and non-immune groups, reporting poorer prognosis in the immune group [ 26 ]. Yet in that study, the immune group also included a greater proportion of advanced-stage cases (70% vs. 44%). In our larger cohorts, we replicated the immune/non-immune classification and confirmed a similar trend toward poorer prognosis in the immune subtype (Figs. 1 A and 1B). Nevertheless, the immune group in our cohort likewise contained significantly more advanced-stage cases (Fig.  1 C), and the transcriptome-based immune classification did not retain prognostic significance in multivariate analysis (Fig.  1 D). Previous reports have shown that advanced-stage CCC is characterized by increased exhausted T cells expressing lymphocyte-activation gene 3 (LAG-3) [ 47 ], suggesting that transcriptome-based differences in immune profiling may partly reflect disease stage. The pathology-based inflamed group tended to have better prognosis, and in multivariate analysis, the inflamed phenotype emerged as an independent prognostic factor. Favorable outcomes in CD8-positive, TIL-high CCC cases have also been reported in previous studies [ 20 , 21 ]. Yet those studies were based on restricted assessments of tissue cores or tissue microarrays. Contrarily, our study evaluated whole H&E slides with cutting-edge AI methods and reached consistent conclusions, providing more robust evidence. Of note, because CCC-specific validated TIL thresholds are currently unavailable, we adopted an empirical upper-tertile cutoff consistent with our prior HGSC study, which warrants external validation in independent CCC cohorts. Collectively, these findings underscore a potential discrepancy between pathology-based and transcriptome-based immune classifications and indicate that pathology-based immune profiling may more directly reflect patient outcomes. Gene expression analysis of the immune group revealed intriguing findings. Pathology-based inflamed cases appeared to predominantly reflect CD8 T cell activity, whereas non-inflamed cases were characterized by adverse prognostic features, including decreased M1/M2 ratios and elevated NLRs. Previous studies have shown that stromal macrophages in CCC are predominantly M2 polarized [ 48 ], which aligns with our finding of significantly higher StromaScores in the non-inflamed group. Other reports have suggested that low-risk CCC cases exclude tumor-associated macrophages from the tumor margin [ 49 ], and that macrophage-rich CCC is associated with poor prognosis [ 50 ], supporting the notion that macrophages may play a role in immune suppression in CCC. The higher Dysfunction scores observed in the non-inflamed group further underscore differences in the immune landscape. Collectively, these compositional and functional differences may partially explain the discrepancy between pathology-based and transcriptome-based immune classifications. Bulk transcriptomic “immune-hot” signals may capture inflammation driven by stromal and myeloid components and/or dysfunctional (exhausted) T-cell programs, rather than reflecting effective cytotoxic T-cell infiltration within tumor regions. This study has the strength of large sample size with comprehensive genetic and transcriptomic profiles. In addition, this study quantitatively assessed TILs using H&E-stained slides from a large, prospectively collected, and relatively recent ovarian cancer cohort with detailed clinical information. However, a key limitation is that the mechanism underlying the observed discrepancy between pathology-based and transcriptome-based immune classifications remains unclear, and validation in larger cohorts is warranted. Integrating spatial transcriptomics with AI-based pathological quantification would offer an attractive future direction to help bridge the gap between bulk transcriptomes and morphologic phenotypes. Another limitation is that the cohorts largely comprised patients treated with conventional platinum-based chemotherapy, and associations with PARP inhibitors or ICIs could not be assessed. As several trials have demonstrated promising efficacy of ICIs in CCC [ 12 , 13 , 51 – 53 ], further studies are needed to explore potential predictive biomarkers. In conclusion, pathology-based immune classification was more strongly associated with prognosis than transcriptome-derived clusters in CCC, underscoring potential differences between these modalities. These findings provide new insights into the tumor immune microenvironment in CCC and highlight the importance of integrating histological and molecular approaches to immune profiling.

Ovarian cancer remains one of the most lethal gynecological malignancies worldwide [ 1 , 2 ]. Clear cell carcinoma (CCC) is a subtype of epithelial ovarian cancer that is distinguished biologically and clinically [ 3 – 6 ]. CCC is more prevalent in Asian populations and is often diagnosed at a younger age, frequently arising in association with endometriosis and harboring characteristic molecular alterations such as ARID1A and PIK3CA mutations [ 7 , 8 ]. In addition to these genomic features, CCC shows distinct metabolic adaptations linked to mitochondrial dysfunction and resistance to oxidative stress [ 3 , 4 ]. While patients with early-stage CCC have a favorable prognosis, those with advanced or recurrent disease have a poor prognosis due to limited sensitivity to standard platinum-based chemotherapy [ 9 ]. Recent studies have suggested heterogeneous immune-related features in CCC, including associations between ARID1A alterations and/or PD-L1 expression and the tumor immune microenvironment [ 10 , 11 ]; several clinical trials have indicated that CCC is a promising target for immune checkpoint inhibitors (ICIs) [ 12 , 13 ], highlighting the need for more precise prognostic stratification and a deeper understanding of the immune landscape of CCC. Tumor-infiltrating lymphocytes (TILs) are recognized as a key factor in antitumor immunity and have been associated with prognosis across multiple cancer types [ 14 , 15 ]. The extent of TIL infiltration and its association with clinical outcomes in ovarian cancer have been reported to vary by histological subtype [ 16 – 18 ]. High-grade serous carcinoma (HGSC), the most common subtype, exhibits the highest levels of TIL infiltration, and numerous studies have established intraepithelial CD8 + TILs as a favorable prognostic factor [ 14 , 15 , 19 ]. Conversely, CCC is typically characterized by an “immunologically cold” microenvironment with sparse TIL infiltration [ 16 ]. The prognostic significance of TILs in CCC remains controversial; while some studies suggest a favorable association [ 20 , 21 ], others report insignificant impact [ 11 , 22 ]. Importantly, despite these previous investigations, a systematic and quantitative evaluation of TILs in CCC, using standardized methodology and large, well-annotated cohorts, has not yet been conducted. Notably, in transcriptomic immune subtyping, studies have paradoxically reported worse outcomes in the “immune-hot” subtype compared to the “immune-cold” subtype [ 23 – 26 ]. Several studies employed multiplex gene expression panels and found that the immune-hot group was associated with higher recurrence rates [ 23 ]. Another study used RNA sequencing to define immune subtypes and similarly reported poor outcomes in the immune group compared to the non-immune group [ 26 ]. Conversely, other studies have suggested better survival in immune-hot tumors relative to immune-cold tumors [ 27 ]. This inconsistency may reflect differences in cohorts, platforms, and analytic pipelines, as well as the fact that bulk transcriptomic signatures capture averaged immune activity and are influenced by tumor purity and spatial heterogeneity. In addition, a variety of gene expression–based classification systems have been proposed [ 25 , 28 ]; however, none have been clinically validated in CCC, and it remains unclear how transcriptomic immune subtypes relate to morphology-based immune profiling in this disease [ 7 , 29 ]. Recent advancements in artificial intelligence (AI)-based digital pathology have enabled simultaneous nuclear segmentation and cell-type classification on hematoxylin and eosin (H&E)–stained whole-slide images (WSIs) [ 30 – 32 ]. These methods allow objective, single-cell–level quantification of TILs and their spatial distribution within tumor and stromal compartments, thereby overcoming some limitations of conventional visual assessment such as inter-observer variability and coarse semiquantitative scoring. In this context, AI-assisted pathology is an attractive modality for evaluating the immune microenvironment, and we hypothesized that AI-assisted pathology–based TIL quantification would provide more robust prognostic stratification than transcriptome-based immune classification in CCC, potentially addressing contrasting reports. In the present study, we newly established large bulk transcriptome cohorts of CCC (JGOG3025-TR1 and Kyoto datasets) that allow for comprehensive immune profiling at the gene expression level. Using these cohorts, we performed detailed AI-based quantification of TILs on digitized H&E slides with two state-of-the-art cell classification models [ 30 , 33 ], and analyzed their association with survival outcomes. We also investigated the relationship between pathologically quantified TILs and transcriptome-based immune subtypes.

Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 7631 KB) Supplementary file1 (DOCX 7631 KB)