Abstract
Standard-of-care treatment options provide an excellent prognosis for papillary thyroid cancers (PTCs); however, approximately 10% of cases are advanced PTCs, resulting in less than 50% 5-year survival rates. Understanding the tumor microenvironment is essential for understanding cancer progression and investigating potential biomarkers for treatment, such as immunotherapy. Our study focused on tumor-infiltrating lymphocytes (TILs), which are the main effectors of antitumor immunity and related to the mechanism of immunotherapy. Using an artificial intelligence model, we analyzed the density of intratumoral and peritumoral TILs in the pathologic slides of The Cancer Genome Atlas PTC cohort. Tumors were classified into three immune phenotypes (IPs) based on the spatial distribution of TILs: immune-desert (48%), immune-excluded (34%), and inflamed (18%). Immune-desert IP was mostly characterized by RAS mutations, high thyroid differentiation score, and low antitumor immune response. Immune-excluded IP predominantly consisted of BRAF V600E-mutated tumors and had a higher rate of lymph node metastasis. Inflamed IP was characterized by a high antitumor immune response, as demonstrated by a high cytolytic score, immune-related cell infiltrations, expression of immunomodulatory molecules (including immunotherapy target molecules), and enrichment of immune-related pathways. This study is the first to investigate IP classification using TILs in PTC through a tissue-based approach. Each IP had unique immune and genomic profiles. Further studies are warranted to assess the predictive value of IP classification in advanced PTC patients treated with immunotherapy.
Introduction
Thyroid cancer is the most frequent endocrine malignancy and the ninth most common cancer diagnosed globally in 2020 (Garcia-Alvarez et al. 2022). The incidence of cases in the United States has experienced a substantial three-fold increase between 1975 and 2013. This surge can largely be attributed to a significant rise in the diagnosis of PTC (Lim et al. 2017). PTC comprises about 75% of total thyroid cancers, and most of them exhibit favorable prognoses with survival rates of 95% after 35–40 years using standard treatment options, such as surgery and radioactive iodine (Viola et al. 2016). However, approximately 10% of PTCs are advanced and are difficult to treat with conventional treatment options, resulting in less than 50% 5-year survival. The clinical benefit of kinase inhibitors is limited due to drug resistance and toxicities. Immunotherapy is a promising approach for advanced thyroid cancer treatment, with proven clinical benefits in other cancer types (Naoum et al. 2018, Pan et al. 2020).
However, the administration of immunotherapy as a treatment option can present challenges due to the inconsistent response observed among patients, even when established biomarkers such as PD-L1 expression or tumor mutational burden (TMB) are considered. This variability has also been observed in clinical trials focusing on thyroid cancers (Arora et al. 2019, Garcia-Alvarez et al. 2022). Tumor-infiltrating lymphocytes (TILs) can be a potential biomarker, as immunotherapy involves exhausted T cells in the tumor microenvironment (TME) (Paijens et al. 2021). In a study of metastatic melanoma, the presence of CD8+ T cells within the tumor margin was predictive of response to PD-1 blockade, highlighting the importance of the spatial distribution of TILs (Tumeh et al. 2014). Furthermore, TILs had a general prognostic value in different types of cancer with treatment options other than immunotherapy, as they are the major effectors of anti-cancer immunity (Geng et al. 2015, Stanton & Disis 2016, Idos et al. 2020). However, utilizing TILs as a potential biomarker has been limited since it has to be manually calculated, leading to interobserver variability (Park et al. 2022b ).
To overcome the limitations associated with manual TIL calculation, we utilized an artificial intelligence (AI) model to quantify TIL densities in intratumoral and peritumor area. We analyzed the hematoxylin and eosin (H&E) slides of The Cancer Genome Atlas (TCGA) PTC cohort and classified tumors into three immune phenotypes (IPs). Our study is the first to investigate the implications of IP classification in PTC utilizing TILs via a tissue-based approach. Our study aimed to gain a better understanding of the TME of PTC and to investigate the potential predictive and therapeutic implications of IP classification according to the spatial distribution of TILs.
Methods
Immune phenotype classification
The Lunit SCOPE IO, an AI-powered analyzer of H&E slides, was used to classify patients into three IPs: immune-desert (’desert’ or ‘D‘), immune-excluded (‘excluded’ or ‘E‘), and inflamed (’inflamed’ or ‘I’). This AI model consists of two computer vision models – a cell detection model that locates lymphocytes and tumor cells and a tissue segmentation model that identifies cancer area (CA), cancer stroma (CS), or other irrelevant or background regions. The AI model was used for analysis in a previous publication (Jung et al. 2022).
The whole-slide images (WSIs) were divided into 1 mm2 grids to classify IPs. Each grid was annotated based on the spatial distribution of TILs using the TIL density cutoff value of 200/mm2. A grid was considered inflamed if the TIL density in CA exceeded the cutoff value; excluded if the TIL density in CA was below the cutoff value and the TIL density in CS exceeded the cutoff value; and desert if the TIL densities in both CA and CS were below the cutoff value. The final IP of the WSI was determined based on the proportion of grids with each IP. WSIs were classified as inflamed IP if the proportion of inflamed grids was greater than 1/3 (inflamed score (IS) ≥ 33.3%), excluded IP if the proportion of excluded grids exceeded 1/3 and the proportion of inflamed grids was less than 1/3, and desert IP if otherwise. We used the IS cutoff value of 33.3% for the purpose of assigning patients into one of these three IP categories.
Datasets, mutation analysis
Data from the PTC cohort in TCGA were utilized for the study (Cancer Genome Atlas Research 2014). Clinical data, mutation status, thyroid differentiation score (TDS), and mRNA expression levels were obtained through cBioportal (Cerami et al. 2012, Gao et al. 2013).
Driver mutations were identified as BRAF V600E, RAS (HRAS/NRAS/KRAS), RET fusion, or NTRK fusion, which are known to commonly occur in PTC (Cancer Genome Atlas Research 2014). DNA repair gene mutations were investigated using 33 genes related to mismatch repair or homologous recombination (Chae et al. 2019). The clinical implications (oncogenicity) of identified DNA repair gene mutations were determined using OncoKb, a precision oncology knowledge database (Chakravarty et al. 2017).
TDS was calculated from the mRNA expression levels of 16 thyroid function genes (Cancer Genome Atlas Research 2014). It provides a measure of thyroid differentiation, with higher scores indicating more differentiated PTCs.
Mutation count/neoantigen count/TMB
CloudNeo pipeline was used to generate mutation count and predict neoantigen count as described in their paper (Bais et al. 2017). In brief, the Variant Call Format and Binary Alignment Map (BAM) files were acquired from the Genomic Data Commons database (Ellrott et al. 2018) and used as inputs for the CloudNeo pipeline which was run through Cancer Genomics Cloud platform (Bais et al. 2017). Total mutation count was defined as the number of non-synonymous mutations (missense mutations, frameshift indels, and inframe indels), which were considered for generating neo-peptides. Six HLA (human leukocyte antigen) types were predicted using the tumor whole exome sequence BAM file and OptiType (Szolek et al. 2014). Neo-peptides with strong MHC (major histocompatibility complex)-binding affinity were selected and subsequently underwent two filtration processes (’normal filter’ and ‘expression filter’) to derive the final neoantigen count (Bais et al. 2017). TMB was calculated as the total mutation count per megabase pairs covered by whole-exome sequencing.
Cytolytic score, expression of immunomodulatory molecules
Cytolytic score (CYT), defined as the geometric mean of granzyme A (GZMA) and perforin-1 (PRF1) mRNA expression levels, was used to measure immune activity (Park et al. 2022a ). Immunomodulatory gene expression levels were compared between groups using mRNA expression z-scores. Thirty-four immune stimulatory molecules and 22 immune inhibitory molecules were examined which included 11 immunotherapy target molecules: ICOS, CD40, TNFRSF4 (OX40), TNFRSF18 (GITR), TNFRSF9 (4–1BB), CTLA4, PDCD1 (PD-1), CD274 (PD-L1), LAG3, IDO1, andHAVCR2 (TIM-3) (Lu et al. 2017, Chae et al. 2019, Park et al. 2022a).
Cell type enrichment analysis
The xCell tool was utilized to predict the infiltrations of various cell types in the TME based on mRNA expression data (Aran et al. 2017). This web-based cell type enrichment analytics tool can predict infiltrations of 64 different types of immune and stromal cells.
Each cell type was classified as myeloid cell, lymphoid cell, stromal cell, hematopoietic stem cell, or epithelial cell (Park et al. 2022a ).
Gene set enrichment analysis
Gene set enrichment analysis (GSEA) was carried out using the Limma package in R to explore differentially expressed pathways between groups (Ritchie et al. 2015). The RNA sequence data were preprocessed and assessed for quality before being used to find differentially expressed genes (DEGs). The rankings of DEGs based on their log-transformed fold change and adjusted P-value were used for GSEA of 50 hallmark pathways (Liberzon et al. 2015, Subramanian et al. 2005). The GSEA was repeated ten times using a different seed number from 1 to 10 and a permutation number of 1000. Differentially expressed pathways with an adjusted P-value consistently less than 0.05 were considered statistically significant.
Statistics
Statistical analyses were performed using R 4.2.2 and GraphPad Prism 9.5.1. The comparison between multiple groups, such as different IPs or driver mutation statuses, was conducted using Kai square test or Kruskal–Wallis test, as appropriate. If the comparison between multiple groups was significant, further pairwise group comparisons were performed using Fisher’s exact test with Bonferroni correction or post hoc Dunn test with Bonferroni correction, as appropriate. The degree of correlation was assessed using Spearman’s rank correlation coefficient. To control the false discovery rate (FDR) when comparing the expression of immunomodulatory molecules or cell type enrichment analysis between groups, an FDR-adjusted P-value (q-value) less than 0.05 by the Benjamini–Hochberg method was used for statistical significance. The P-values in the figure were represented as *P < 0.05, **P < 0.01, and ***P < 0.001.
Results
The IP classification was associated with different clinical characteristics
Of the 496 PTC patients in the TCGA dataset, 469 patients were included in this study based on their availability of IP classification, mRNA expression, and xCell results. The IP distribution was as follows: 224 (48%) immune-desert, 158 (34%) immune-excluded, and 87 (18%) inflamed. Patients in desert IP were diagnosed at a higher age than those with other IPs, with a median age of 49 (P <0.001). There was no difference in the gender distribution of patients across different IPs (P = 0.292). Patients in excluded IP had a higher incidence of lymph node metastasis (P < 0.001). Table 1 provides more details about baseline characteristics.
Demographics and baseline characteristics of patients with papillary thyroid cancer.
| n (%) | Desert | Excluded | Inflamed | P | ||
|---|---|---|---|---|---|---|
| All | 469 | 224 (48%) | 158 (34%) | 87 (18%) | ||
| Age | Total | 443 | 213 | 148 | 82 | |
| Median (years, range) | 46 (15-89) | 49 (15-88) | 44.5 (16-89) | 39 (17-72) | <0.001 | |
| Gender | Total | 443 | 213 | 148 | 82 | 0.292 |
| Male | 120 (27%) | 65 | 35 | 20 | ||
| Female | 323 (73%) | 148 | 113 | 62 | ||
| Stage | Total | 441 | 211 | 148 | 82 | |
| Stage I | 254 (58%) | 106 | 83 | 65 | ||
| Stage II | 49 (11%) | 32 | 13 | 4 | ||
| Stage III | 94 (21%) | 50 | 35 | 9 | ||
| Stage IV | 44 (10%) | 23 | 17 | 4 | ||
| T | Total | 442 | 212 | 148 | 82 | 0.403 |
| T 1/2 | 279 (63%) | 138 | 87 | 54 | ||
| T 3/4 | 163 (37%) | 74 | 61 | 28 | ||
| N | Total | 400 | 182 | 139 | 79 | <0.001 |
| N0 | 203 (51%) | 112 | 49 | 42 | ||
| N1 | 197 (49%) | 70 | 90 | 37 | ||
| M | Total | 244 | 118 | 79 | 47 | 0.229 |
| M0 | 236 (97%) | 112 | 77 | 47 | ||
| M1 | 8 (3%) | 6 | 2 | 0 | ||
| Driver mutation | Total | 380 | 176 | 136 | 68 | |
| BRAF V600E | a225 (59%) | a70 | 113 | 42 | <0.001 | |
| RAS | a50 (13%) | a46 | 3 | 1 | <0.001 | |
| NRAS | 33 | 30 | 2 | 1 | ||
| HRAS | 13 | 12 | 1 | 0 | ||
| KRAS | a4 | a4 | 0 | 0 | ||
| Fusion | 32 (8%) | 10 | 12 | 10 | 0.073 | |
| RET | 23 (6%) | 7 | 9 | 7 | ||
| NTRK | 9 (2%) | 3 | 3 | 3 | ||
| No driver mutation | 74 (19%) | 51 | 8 | 15 | <0.001 | |
| TDS | Total | 380 | 176 | 136 | 68 | <0.001 |
| Median (range) | −0.23 (−4.08 to 2.58) | 0.63 (−4.08 to 2.23) | −0.69 (−3.16 to 2.53) | −0.38 (−2.39 to 2.58) | ||
| Mutation count | Total | 383 | 183 | 133 | 67 | 0.18 |
| Median (range) | 6 (1-62) | 6 (1-46) | 6 (1-62) | 5 (1-32) | ||
| Neoantigen count | Total | 383 | 183 | 133 | 67 | 0.29 |
| median (range) | 1 (0 to 21) | 1 (0 to 11) | 1 (0 to 21) | 1 (0 to 10) | ||
| TMB | Total | 315 | 147 | 116 | 52 | 0.411 |
| Median (range) | 0.2 (0.03-2.05) | 0.23 (0.03-1.15) | 0.2 (0.03-2.02) | 0.2 (0.03-1.1) |
aOne coexisting BRAF/KRAS mutation.
TDS, thyroid differentiation score; TMB, tumor mutational burden.
The IP classification demonstrated unique mutational profiles
Among the 380 patients with available mutation status by whole genome sequencing, 225 (59%) had BRAF V600E mutation, 50 (13%) had RAS mutation, 32 (8%) had fusion genes (RET or NTRK), and 74 (19%) did not harbor any driver mutations (Table 1). The driver mutations were nearly mutually exclusive, with only one patient having coexisting mutations of BRAF V600E and KRAS. The number of BRAF V600E mutations was 70 in desert IP, 113 in excluded IP, and 42 in inflamed IP. The prevalence of BRAF V600E mutation was statistically different between the three IPs (P < 0.001), and subsequent pairwise comparisons between IPs also showed significant differences in all pairs of IPs (adjusted P, D–E < 0.001, D–I = 0.007, E–I = 0.004). The number of RAS mutations was 46 in desert IP, 3 in excluded IP, and 1 in inflamed IP. The prevalence of RAS mutation was statistically different between the three IPs (P < 0.001), and subsequent pairwise comparisons showed statistical differences between the following pair of IPs: D–E and D–I (adjusted P, D–E < 0.001, D–I < 0.001). The number of fusion genes was 10 in desert IP, 12 in excluded IP, and 10 in inflamed IP. The prevalence of fusion genes was not statistically different between the three IPs (P = 0.07). The number of patients without any driver mutations was 51 in desert IP, 8 in excluded IP, and 15 in inflamed IP. The prevalence of patients without any driver mutations showed a statistical difference between the three IPs (P < 0.001), and subsequent pairwise comparison showed statistical differences between the following pair of IPs: D–E and E–I (adjusted P, D–E < 0.001, E–I = 0.004).
There was a clear pattern of driver mutation distribution among the three IPs (Fig. 1). The desert IP had 70 (40%) BRAF V600E mutations, 46 (26%) RAS mutations, 10 (6%) fusion genes, and 51 (29%) cases without any driver mutations. The excluded IP was mainly composed of BRAF V600E mutations (113, 83%) with 3 (2%) RAS mutations, 12 (9%) fusion genes, and 8 (6%) cases without any driver mutations. The inflamed IP consisted of 42 (62%) BRAF V600E mutations, 1 (1%) RAS mutation, 10 (15%) fusion genes, and 15 (22%) cases without any driver mutations.
Distribution of driver mutations according to IPs. (A) Number of tumors with driver mutations in each IP. (B) Proportion of driver mutations in desert IP. (C) Proportion of driver mutations in excluded IP. (D) Proportion of driver mutations in inflamed IP. IP, immune phenotype; mut, mutation.
Citation: Endocrine-Related Cancer 30, 9; 10.1530/ERC-23-0110
The IP classification or driver mutation status did not result in significant differences in total mutation count, neoantigen, or TMB
The number of non-synonymous mutations varied from 1 to 62, with a median of 6. The number of neoantigens ranged from 0 to 21, with a median of 1. TMB ranged from 0.03 to 2.05, with a median of 0.2. However, there were no significant differences in the number of mutations, neoantigen, or TMB between IPs or driver mutation status (Table 1, Supplementary Fig. 1, see section on supplementary materials given at the end of this article).
Out of the 469 patients included in the study, 23 (5%) had mutations in DNA repair genes, but each patient had only one mutation. Of these mutations, only five (2 ATR, 2 ATM, 1 CHEK2) were classified as likely oncogenic, while the remaining 18 mutations (1 MSH2, 1 PMS1, 1 ATR, 3 ATM, 4 CHEK2, 1 BRCA1, 3 BRCA2, 1 BAP1, and 3 FANCD2) were of unknown oncogenic effect.
The tumors with higher TDS were more frequently observed in the desert IP
The median TDS for all samples was −0.23, ranging from −4.08 to 2.58. The desert IP showed the highest median TDS among the three IPs (median 0.63 with a range of −4.08 to 2.23), while the excluded IP and inflamed IP showed a negative median TDS value (Table 1). The difference in TDS between the three IPs was statistically significant (P < 0.001), with significant differences in the following pair of IPs: D–E and D–I (adjusted P, D–E < 0.001, D–I < 0.001) (Fig. 2A).
TDS and CYT according to IPs or driver mutations. Median with interquartile range. (A) TDS according to IPs. (B) TDS according to driver mutations. (C) CYT according to IPs. (D) CYT according to driver mutations. CYT, cytolytic score; IP, immune phenotype; mut, mutations; TDS, thyroid differentiation score.
Citation: Endocrine-Related Cancer 30, 9; 10.1530/ERC-23-0110
TDS also varied by driver mutation status (P < 0.001). BRAF V600E mutated samples had a negative median TDS value (−0.696), while samples with other driver mutation statuses had a positive median value. There were statistically significant pairwise differences in TDS between groups (P-values in Fig. 2B), except in the comparison between RAS mutation and no driver mutations (Fig. 2B).
The tumors in the inflamed IP had higher CYT levels, regardless of driver mutations
The median CYT was highest in the inflamed IP, followed by excluded IP and desert IP (median CYT: 192, 59, 40, respectively). The difference in CYT between the three IPs was statistically significant (P < 0.001), and further pairwise comparisons showed a significant difference between all pairs of IPs (adjusted P, D−E = 0.001, D−I < 0.001, E−I < 0.001) (Fig. 2C).
There were differences in CYT based on driver mutation status (P < 0.001), with RAS mutations having the lowest median CYT compared to other mutations. There were statistically significant differences in CYT between most pairs of groups (P-values in Fig. 2D), except for BRAF mutation vs Fusion and BRAF mutation vs no driver mutations (Fig. 2D).
To better understand the relationship between IPs, driver mutation status, and CYT, we analyzed the data by subgrouping samples based on their driver mutation status. The analysis revealed that regardless of driver mutation status, the inflamed IP consistently contained samples with high CYT (Supplementary Fig. 2A, B and C).
The analysis of all samples showed a negative relationship between TDS and CYT (spearman r = −0.173, P < 0.001). However, when TDS was examined in each IP subgroup, only the desert IP displayed a statistically significant correlation between TDS and CYT (spearman r = −0216, adjusted P = 0.011) (Supplementary Fig. 2D).
Inflamed IP was associated with higher expression levels of both stimulatory and inhibitory immunomodulatory molecules
The expression of immunomodulatory molecules was significantly different between the three IPs (q < 0.05). The inflamed IP had the highest expression of both stimulatory and inhibitory immunomodulatory molecules (34/34 immune stimulatory molecules, 16/22 inhibitory molecules, q < 0.001), while the desert IP had the lowest expression (Fig. 3A, Supplementary Fig. 3). CTLA4, CXCL9, ICOS, and GZMA were among the most statistically significant immunomodulatory molecules that had the highest median expression in inflamed IP (q < 0.001) (Fig. 3B). Excluded IP showed the highest median expression for only four immunomodulatory molecules (TGFB1, TGFB2, TGFBR1, and PRDM1) (q < 0.001) (Fig. 3C), while desert IP had the highest median expression for only two immunomodulatory molecules (TGFBR2 and TGFBR3) (q < 0.001) (Fig. 3D).
mRNA expression of immunomodulatory molecules according to IPs. The color denotes each IP: Inflamed IP as red, excluded IP as blue, and desert IP as brown. (A) Heatmap. (B) The four most statistically significant molecules which showed the highest median value in inflamed IP. (C) The four statistically significant molecules which showed the highest median value in excluded IP. (D) The two statistically significant molecules which showed the highest median value in desert IP. IP, immune phenotype.
Citation: Endocrine-Related Cancer 30, 9; 10.1530/ERC-23-0110
All immunotherapy target molecules were most highly expressed in inflamed IP.
Inflamed IP had significantly higher infiltrations with immune-related cells compared to the other IPs
The eight cell types that showed the most statistical differences (q < 0.001) in infiltrations are as listed in the order of significance: activated dendritic cell (aDC), B cells, DC, CD8+ central memory T cells, CD4+ memory T cells, epithelial cells, CD8+ T cells, and class-switched memory B cells. The inflamed IP had the highest median value for all these cell types except for epithelial cells (Fig. 4B). The excluded IP had the highest median value for epithelial cells, sebocytes, and immature DC. The desert IP had the highest median value for neurons, osteoblasts, and CD4+ central memory T cells. Immune-related cell types, such as myeloid and lymphoid cells, were highly infiltrated in inflamed IP, while cell types unrelated to immune response were highly infiltrated in desert IP (Fig. 4A).
Infiltrations of different cell types according to IPs. The color denotes each IP: Inflamed IP as red, excluded IP as blue, and desert IP as brown. (A) Heatmap. (B) The 8 cell types which showed the most statistically significant difference in infiltrations; all the cell types except epithelial cells showed the highest median value in inflamed IP, epithelial cells showed the highest median value in excluded IP. aDC, activated DC; CD4+ Tem, CD4+ effector memory T cell; cDC, conventional DC; CLP, common lymphoid progenitor; CMP, common myeloid progenitor; DC, dendritic cell; GMP, granulocyte-macrophage progenitor; HSC, hematopoietic stem cell; iDC, immature DC; IP, immune phenotype; MEP, megakaryocyte-erythroid progenitor; MPP, multipotent progenitor; MSC, mesenchymal stem cell; mv endothelial cells, microvascular endothelial cells; NK cells, natural killer; NKT, NK T cell; pDC, plasmacytoid DC; Tgd cells, gamma delta T cell; Th1 cells, type 1 helper T cell; Th2 cells, type 2 helper T cell; Tregs, regulatory T cell; ly endothelial cells, lymphatic endothelial cells..
Citation: Endocrine-Related Cancer 30, 9; 10.1530/ERC-23-0110
The distribution of different cell types within each IP was analyzed and presented in Supplementary Fig. 4. Among 13 myeloid cell types, 7 cell types (activated DC, DC, macrophage M1, macrophages, plasmacytoid DC, conventional DC, and monocytes) showed the highest median value in inflamed IP, followed by 3 cell types (immature DC, eosinophils, and mast cells) in excluded IP, 1 cell type (macrophage M2) in desert IP and 2 cell types (basophils and neutrophils) without significant differences between IPs.
Among 21 lymphoid cell types, 10 cell types (B cells, CD8+ central memory T cells, CD4+ memory T cells, CD8+ T cells, class-switched memory B cells, CD4+ naive T cells, naive B cells, plasma cells, memory B cells, and CD8+ naive T cells) showed the highest median value in inflamed IP, followed by 1 cell type (natural killer T cell) in excluded IP, 1 cell type (CD4+ central memory T cells) in desert IP, and 9 cell types (pro B cells, CD4+ effector memory T cell, type 1 helper T cells, CD4+ T cells, regulatory T cells, natural killer cells, CD8+ effector memory T cells, type 2 helper T cells, and gamma delta T cells) without significant differences between IPs.
Among 14 stromal cells, 8 cell types (osteoblast, microvascular endothelial cells, endothelial cells, lymphatic endothelial cells, pericytes, skeletal muscle, mesenchymal stem cells, and preadipocytes) showed the highest median value in desert IP, followed by 2 cell types (mesangial cells and fibroblasts) in inflamed, 2 cell types (myocytes and adipocytes) in excluded IP, and 2 cell types (smooth muscle and chondrocytes) without significant differences between IPs.
Among 9 hematopoietic stem cells, 3 cell types (hematopoietic stem cell, megakaryocyte-erythroid progenitor, multipotent progenitor) showed the highest median value in desert IP, followed by 1 cell type (granulocyte-macrophage progenitor) in inflamed IP, and 5 cell types (common lymphoid progenitor, platelets, common myeloid progenitor, erythrocytes, megakaryocytes) without significant differences between IPs.
Among 7 epithelial cells, 3 cell types (epithelial cells, sebocytes, and keratinocytes) showed the highest median value in excluded IP, followed by 2 cell types (neurons and hepatocytes) in desert IP, 1 cell type (astrocytes) in inflamed IP, and 1 cell type (melanocytes) without significant differences between IPs.
Inflamed IP consistently showed higher expression of immune-related pathways when compared to other IPs
The GSEA method allowed for pairwise comparison of differentially expressed pathways. Comparing inflamed IP with excluded IP revealed eight pathways that were highly expressed in inflamed IP, including allograft rejection, interferon-α response, interferon-γ response, complement, inflammatory response, IL-6 JAK STAT3 signaling, KRAS signaling up, and IL2 STAT5 signaling (Fig. 5A). When comparing inflamed IP with desert IP, 15 differentially expressed pathways were identified. All 15 pathways were highly expressed in inflamed IP, which included the same 8 pathways from the previous comparison plus additional 7 pathways (TNFA signaling via NFKB, apoptosis, G2M checkpoint, E2F targets, epithelial mesenchymal transition, coagulation, and p53 pathway) (Fig. 5B). Comparing excluded IP with desert IP revealed 22 differentially expressed pathways, with 20 pathways highly expressed in excluded IP (including the same 15 differentially expressed pathways from inflamed IP vs desert IP plus apical junction, mitotic spindle, estrogen response late, estrogen response early, and hypoxia) and 2 pathways highly expressed in desert IP (oxidative phosphorylation and adipogenesis Fig. 5C).
Gene set enrichment analysis. The color denotes each IP: Inflamed IP as white, excluded IP as gray, and desert IP as black. (A) Inflamed IP vs us excluded IP. (B) Inflamed IP vs desert IP. (C) Excluded IP vs desert IP. IP, immune phenotype.
Citation: Endocrine-Related Cancer 30, 9; 10.1530/ERC-23-0110
Discussion
Our study found that each IP in PTC, based on the spatial distribution of TILs, had unique immune and genomic characteristics. Interestingly, our results suggest that approximately 18% of PTC exhibit immune characteristics associated with a favorable response to immunotherapy in other types of cancer (Tumeh et al. 2014, Chen & Mellman 2017). A previous study on PTC patients showed that having a BRAF V600E mutation or low TDS is related to increased immune cell infiltrations and PD-L1 expression (Na & Choi 2018). However, our study revealed that the inflamed IP, which exhibited high antitumor immune activity, was not distinguishable from other IPs by BRAF V600E mutation or TDS.
Of note, we examined the percentage of patients with an inflamed IP using an IS cutoff value greater than 20% to compare it with other cancer types. The proportion of patients classified as having an inflamed IP using this criterion was 25% (116/469) of the total patient population, which was lower when compared to melanoma (56.3%), renal cell carcinoma (52.9%), and non-small cell lung cancer (NSCLC, 33.7%) (Shen et al. 2022). This difference may be attributed to the varying immunogenicity between different cancer types.
Desert IP was characterized by having a high TDS and harboring the majority of RAS mutations, which are known to create an immunosuppressive environment. Previous studies have demonstrated that high TDS is associated with lower antitumor immune response in PTC (Na & Choi 2018), and RAS-mutated thyroid cancer exhibits downregulations of MHC molecules and immunomodulatory molecules (Charoentong et al. 2017). Furthermore, studies on other cancer types, including melanoma and colorectal cancer, have also shown that RAS mutations are associated with a less active immune response (Thomas et al. 2015, Janssen et al. 2022). In desert IP, these factors led to lower immune activity, as evidenced by low CYT and low expressions of immunomodulatory molecules. TGFBR2 and TGFBR3,the two immunomodulatory molecules with the highest median value in desert IP, are also related to immune suppression (Shi et al. 2022). Somewhat evident from the IP classification process for desert IP involving low TILs in H&E slides, desert IP showed greater infiltrations of non-immune-related cells compared to immune-related cells (e.g., antigen-presenting cells or effector cells). Of the immune-related cells, macrophage M2 showed the highest median value in desert IP, which is consistent with the immunosuppressive TME of desert IP.
GSEA analysis revealed that desert IP showed consistently lower expression of immune, DNA replication related, KRAS signaling, and epithelial mesenchymal transition pathways when compared to other IPs. In comparison, the desert IP showed higher expression of oxidative phosphorylation and adipogenesis pathways, which are associated with an immunosuppressive TME related to M2 macrophage activity or infiltrations (Oshi et al. 2021, Park et al. 2022a).
Excluded IP was characterized by being mostly comprised of BRAF V600E mutation and having moderate antitumor activity and a higher rate of lymph node metastasis. It was differentiated from inflamed IP by having lower CYT, although both IPs had low TDS. BRAF V600E mutated tumors with low CYT were included in either excluded IP or desert IP, whereas those with high CYT were included in inflamed IP. The differing distribution of the BRAF V600E mutation across samples may reflect the heterogeneity of this mutation in terms of its prognostic value and its potential to differentiate tumors, as suggested in earlier studies (Cancer Genome Atlas Research 2014). TGFB1, TGFB2, TGFBR1, and PRDM1 were the immunomodulatory molecules that showed the highest median value in excluded IP. BRAF V600E mutation is known to increase the rate of metastasis through TGFB, which can explain the higher rate of lymph node metastasis in excluded IP (Riesco-Eizaguirre et al. 2009). Additionally, TGFB contributes to the development of abundant stromal elements that can restrict the access of immune cells to the tumor parenchyma (Chen & Mellman 2017). This may explain why excluded IP harbors the highest proportion of BRAF V600E mutation as excluded IP was defined as a group of tumors with higher TIL density in stroma compared to cancer epithelium.
Excluded IP was characterized by moderate immune activity, as demonstrated by the expression of immunomodulatory molecules and infiltration of immune cells, such as NKT, iDC, mast cells, and eosinophils. This immune activity was situated between that of desert IP and inflamed IP.
Pathway enrichment analysis revealed that excluded IP had a relatively lower expression of immune-related pathways when compared to inflamed IP, but a relatively higher expression of immune-related pathways when compared to desert IP.
Inflamed IP did not exhibit a specific driver mutation status or TDS. However, it displayed the highest cytolytic activity and expressed both stimulatory and inhibitory immunomodulatory molecules, including those targeted by immunotherapy. In addition, inflamed IP showed the highest infiltration of immune-related as well as an enrichment of immune-related pathways, indicating a strong antitumor response. The antitumor immune response has been proposed to be induced by immune cells recognizing neoantigens of cancer (Arora et al. 2019). However, our study found no significant difference in total mutation count, neoantigen count, or TMB between the three IPs, possibly due to the low mutation count in PTCs compared to other immunogenic cancers, such as NSCLC (median total mutation count: 6 in PTC vs 232 in NSCLC) (Chae et al. 2019). The low mutation count in PTCs may be related to the underlying tumorigenic mechanism, which is driven by driver mutations instead of somatic mutations or DNA repair gene mutations. The detection of 23 mutations in DNA repair-related genes, with the majority having unknown oncogenic effects, is in line with the findings of a previous study on the TCGA PTC cohort that showed no significant variation in mutation density associated with DNA repair-related mutations (Cancer Genome Atlas Research 2014). The driver mutation status was unable to account for the strong antitumor response observed in inflamed IP, as revealed in our subgroup analysis (Supplementary Fig. 2). One potential explanation could be the reduced infiltration of DCs in other IPs, resulting in insufficient T-cell activation and infiltration compared to inflamed IP (Spranger et al. 2016). Further research is needed to understand the underlying mechanism that triggers the robust antitumor immune response in inflamed IP.
According to GSEA, inflamed IP showed consistently high expression of immune-related pathways as well as KRAS signaling pathway. This is consistent with findings from a previous study using the TCGA NSCLC cohort (Park et al. 2022a ). Based on our results, this subset of patients might be a good candidate for the investigation of treatment with immunotherapy.
Our study has some limitations. First, analyzing a single biopsy sample may not accurately reflect tumor heterogeneity, potentially leading to variability in IP classification when using different biopsy samples from the same tumor. Second, the TCGA cohort primarily included surgical samples and excluded clinically aggressive cancers (Cancer Genome Atlas Research 2014). Third, using mRNA expressions to infer infiltrations of different cell types and PD-L1 expressions may not accurately reflect the actual infiltrations of cells or the expression of PD-L1 proteins. While PD-L1 gene expression can indicate a relative abundance of expression, it may not precisely align with the current biomarker, which involves detecting PD-L1 protein expression through immunohistochemical staining.
In conclusion, this study is the first to investigate the implications of IP classification on PTC through a tissue-based approach using TILs. We used AI technology to classify tumors based on TILs from the H&E slide, minimizing interobserver variability. This approach offered two clear advantages over earlier studies that used gene expression data. First, we were able to identify and measure the density of actual TILs. Second, we could extract data on the spatial distribution of TILs which have been shown to be important in antitumor immunity. The results indicated that the three IPs had unique immune and genomic profiles, which were characterized by different driver mutation status, TDS, or antitumor immune activity. Most importantly, inflamed IP identified a subgroup of patients who may benefit from immunotherapy. Further investigation is necessary to confirm the predictive implications of IP classification in advanced PTC patients undergoing immunotherapy.
Supplementary materials
This is linked to the online version of the paper at https://doi.org/10.1530/ERC-23-0110.
Declaration of interest
1. Young Kwang Chae -Research Grant: Abbvie, BMS, Biodesix, Freenome, Predicine -Honoraria/Advisory Boards: Roche/Genentech, AstraZeneca, Foundation Medicine, Neogenomics, Guardant Health, Boehringher Ingelheim, Biodesix, Immuneoncia, Lilly Oncology, Merck, Takeda, Lunit, Jazz Pharmaceutical, Tempus, BMS, Regeneron, NeoImmunTech, Esai.
Funding
Jeffrey H. Chuang acknowledges support from NIH grants R01CA230031 and P30CA034196. Other authors involved in this work did not receive any specific grant from any funding agency in the public, commercial, or not-for-profit sector.
Data availability
All data except immune phenotype classification reported in this manuscript is publicly available or can be found in the original articles mentioned in the list of references. Immune phenotype classification of tumors is uploaded as supplementary data.
Acknowledgements
We thank Dr Jonghanne Park for his assistance with the gene set enrichment analysis.
References
Aran D, Hu Z & & Butte AJ 2017 xCell: digitally portraying the tissue cellular heterogeneity landscape. Genome Biology 18 220. (https://doi.org/10.1186/s13059-017-1349-1)
Arora S, Velichinskii R, Lesh RW, Ali U, Kubiak M, Bansal P, Borghaei H, Edelman MJ & & Boumber Y 2019 Existing and emerging biomarkers for immune checkpoint immunotherapy in solid tumors. Advances in Therapy 36 2638–2678. (https://doi.org/10.1007/s12325-019-01051-z)
Bais P, Namburi S, Gatti DM, Zhang X & & Chuang JH 2017 CloudNeo: a cloud pipeline for identifying patient-specific tumor neoantigens. Bioinformatics 33 3110–3112. (https://doi.org/10.1093/bioinformatics/btx375)
Cancer Genome Atlas Research Network 2014 Integrated genomic characterization of papillary thyroid carcinoma. Cell 159 676–690. (https://doi.org/10.1016/j.cell.2014.09.050)
Cerami E, Gao J, Dogrusoz U, Gross BE, Sumer SO, Aksoy BA, Jacobsen A, Byrne CJ, Heuer ML, Larsson E, et al.2012 The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discovery 2 401–404. (https://doi.org/10.1158/2159-8290.CD-12-0095)
Chae YK, Anker JF, Oh MS, Bais P, Namburi S, Agte S, Giles FJ & & Chuang JH 2019 Mutations in DNA repair genes are associated with increased neoantigen burden and a distinct immunophenotype in lung squamous cell carcinoma. Scientific Reports 9 3235. (https://doi.org/10.1038/s41598-019-39594-4)
Chakravarty D, Gao J, Phillips SM, Kundra R, Zhang H, Wang J, Rudolph JE, Yaeger R, Soumerai T, Nissan MH, et al.2017 OncoKB: a precision oncology knowledge base. JCO Precision Oncology 2017. (https://doi.org/10.1200/PO.17.00011)
Charoentong P, Finotello F, Angelova M, Mayer C, Efremova M, Rieder D, Hackl H & & Trajanoski Z 2017 Pan-cancer immunogenomic analyses reveal genotype-immunophenotype relationships and predictors of response to checkpoint blockade. Cell Reports 18 248–262. (https://doi.org/10.1016/j.celrep.2016.12.019)
Chen DS & & Mellman I 2017 Elements of cancer immunity and the cancer-immune set point. Nature 541 321–330. (https://doi.org/10.1038/nature21349)
Ellrott K, Bailey MH, Saksena G, Covington KR, Kandoth C, Stewart C, Hess J, Ma S, Chiotti KE, Mclellan M, et al.2018 Scalable open science approach for mutation calling of tumor exomes using multiple genomic pipelines. Cell Systems 6 e7. (https://doi.org/10.1016/j.cels.2018.03.002)
Gao J, Aksoy BA, Dogrusoz U, Dresdner G, Gross B, Sumer SO, Sun Y, Jacobsen A, Sinha R, Larsson E, et al.2013 Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal. Science Signaling 6 pl1. (https://doi.org/10.1126/scisignal.2004088)
Garcia-Alvarez A, Hernando J, Carmona-Alonso A & & Capdevila J 2022 What is the status of immunotherapy in thyroid neoplasms? Frontiers in Endocrinology (Lausanne) 13 929091. (https://doi.org/10.3389/fendo.2022.929091)
Geng Y, Shao Y, He W, Hu W, Xu Y, Chen J, Wu C & & Jiang J 2015 Prognostic role of tumor-infiltrating lymphocytes in lung cancer: a meta-analysis. Cellular Physiology and Biochemistry 37 1560–1571. (https://doi.org/10.1159/000438523)
Idos GE, Kwok J, Bonthala N, Kysh L, Gruber SB & & Qu C 2020 The prognostic implications of tumor infiltrating lymphocytes in colorectal cancer: a systematic review and meta-analysis. Scientific Reports 10 3360. (https://doi.org/10.1038/s41598-020-60255-4)
Janssen JBE, Medema JP, Gootjes EC, Tauriello DVF & & Verheul HMW 2022 Mutant RAS and the tumor microenvironment as dual therapeutic targets for advanced colorectal cancer. Cancer Treatment Reviews 109 102433. (https://doi.org/10.1016/j.ctrv.2022.102433)
Jung HA, Park KU, Cho S, Lim J, Lee KW, Hong MH, Yun T, An HJ, Park WY, Pereira S, et al.2022 A Phase II study of nivolumab plus gemcitabine in patients with recurrent or metastatic nasopharyngeal carcinoma (KCSG HN17-11). Clinical Cancer Research 28 4240–4247. (https://doi.org/10.1158/1078-0432.CCR-22-1238)
Liberzon A, Birger C, Thorvaldsdottir H, Ghandi M, Mesirov JP & & Tamayo P 2015 The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Systems 1 417–425. (https://doi.org/10.1016/j.cels.2015.12.004)
Lim H, Devesa SS, Sosa JA, Check D & & Kitahara CM 2017 Trends in thyroid cancer incidence and mortality in the United States, 1974–2013. JAMA 317 1338–1348. (https://doi.org/10.1001/jama.2017.2719)
Lu L, Barbi J & & Pan F 2017 The regulation of immune tolerance by FOXP3. Nature Reviews. Immunology 17 703–717. (https://doi.org/10.1038/nri.2017.75)
Na KJ & & Choi H 2018 Immune landscape of papillary thyroid cancer and immunotherapeutic implications. Endocrine-Related Cancer 25 523–531. (https://doi.org/10.1530/ERC-17-0532)
Naoum GE, Morkos M, Kim B & & Arafat W 2018 Novel targeted therapies and immunotherapy for advanced thyroid cancers. Molecular Cancer 17 51. (https://doi.org/10.1186/s12943-018-0786-0)
Oshi M, Tokumaru Y, Angarita FA, Lee L, Yan L, Matsuyama R, Endo I & & Takabe K 2021 Adipogenesis in triple-negative breast cancer is associated with unfavorable tumor immune microenvironment and with worse survival. Scientific Reports 11 12541. (https://doi.org/10.1038/s41598-021-91897-7)
Paijens ST, Vledder A, De Bruyn M & & Nijman HW 2021 Tumor-infiltrating lymphocytes in the immunotherapy era. Cellular and Molecular Immunology 18 842–859. (https://doi.org/10.1038/s41423-020-00565-9)
Pan C, Liu H, Robins E, Song W, Liu D, Li Z & & Zheng L 2020 Next-generation immuno-oncology agents: current momentum shifts in cancer immunotherapy. Journal of Hematology and Oncology 13 29. (https://doi.org/10.1186/s13045-020-00862-w)
Park J, Cho HG, Park J, Lee G, Kim HS, Paeng K, Song S, Park G, Ock CY & & Chae YK 2022a Artificial intelligence-powered hematoxylin and eosin analyzer reveals distinct immunologic and mutational profiles among immune phenotypes in non-small-cell lung cancer. American Journal of Pathology 192 701–711. (https://doi.org/10.1016/j.ajpath.2022.01.006)
Park S, Ock CY, Kim H, Pereira S, Park S, Ma M, Choi S, Kim S, Shin S, Aum BJ, et al.2022b Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as complementary biomarker for immune checkpoint inhibition in non-small-cell lung cancer. Journal of Clinical Oncology 40 1916–1928. (https://doi.org/10.1200/JCO.21.02010)
Riesco-Eizaguirre G, Rodriguez I, De La Vieja A, Costamagna E, Carrasco N, Nistal M & & Santisteban P 2009 The BRAFV600E oncogene induces transforming growth factor beta secretion leading to sodium iodide symporter repression and increased malignancy in thyroid cancer. Cancer Research 69 8317–8325. (https://doi.org/10.1158/0008-5472.CAN-09-1248)
Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W & & Smyth GK 2015 Limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Research 43 e47. (https://doi.org/10.1093/nar/gkv007)
Shen J, Choi Y-L, Lee T, Kim H, Chae YK, Dulken B, Bogdan S, Huang M, Fisher GA, Park S, et al.2022 The inflamed immune phenotype (IIP): a clinically actionable artificial intelligence (AI)-based biomarker predictive of immune checkpoint inhibitor (ICI) outcomes across >16 primary tumor types. Journal of Clinical Oncology 40 2621–2621. (https://doi.org/10.1200/JCO.2022.40.16_suppl.2621)
Shi X, Yang J, Deng S, Xu H, Wu D, Zeng Q, Wang S, Hu T, Wu F & & Zhou H 2022 TGF-beta signaling in the tumor metabolic microenvironment and targeted therapies. Journal of Hematology and Oncology 15 135. (https://doi.org/10.1186/s13045-022-01349-6)
Spranger S, Luke JJ, Bao R, Zha Y, Hernandez KM, Li Y, Gajewski AP, Andrade J & & Gajewski TF 2016 Density of immunogenic antigens does not explain the presence or absence of the T-cell-inflamed tumor microenvironment in melanoma. Proceedings of the National Academy of Sciences of the United States of America 113 E7759–E7768. (https://doi.org/10.1073/pnas.1609376113)
Stanton SE & & Disis ML 2016 Clinical significance of tumor-infiltrating lymphocytes in breast cancer. Journal for ImmunoTherapy of Cancer 4 59. (https://doi.org/10.1186/s40425-016-0165-6)
Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al.2005 Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences of the United States of America 102 15545–15550. (https://doi.org/10.1073/pnas.0506580102)
Szolek A, Schubert B, Mohr C, Sturm M, Feldhahn M & & Kohlbacher O 2014 OptiType: precision HLA typing from next-generation sequencing data. Bioinformatics 30 3310–3316. (https://doi.org/10.1093/bioinformatics/btu548)
Thomas NE, Edmiston SN, Alexander A, Groben PA, Parrish E, Kricker A, Armstrong BK, Anton-Culver H, Gruber SB, From L, et al.2015 Association between NRAS and BRAF mutational status and melanoma-specific survival among patients with higher-risk primary melanoma. JAMA Oncology 1 359–368. (https://doi.org/10.1001/jamaoncol.2015.0493)
Tumeh PC, Harview CL, Yearley JH, Shintaku IP, Taylor EJ, Robert L, Chmielowski B, Spasic M, Henry G, Ciobanu V, et al.2014 PD-1 blockade induces responses by inhibiting adaptive immune resistance. Nature 515 568–571. (https://doi.org/10.1038/nature13954)
Viola D, Valerio L, Molinaro E, Agate L, Bottici V, Biagini A, Lorusso L, Cappagli V, Pieruzzi L, Giani C, et al.2016 Treatment of advanced thyroid cancer with targeted therapies: ten years of experience. Endocrine-Related Cancer 23 R185–R205. (https://doi.org/10.1530/ERC-15-0555)
