TIMEAS, a promising method for the stratification of testicular germ cell tumor patients with distinct immune microenvironment, clinical outcome and sensitivity to frontline therapies

Abstract

Purpose

With the heterogeneous genetic background, prognosis prediction and therapeutic targets for testicular germ cell tumors (TGCTs) are still unclear. We defined the tumor immune microenvironment activation status (TIMEAS).

Methods

We collected a total of 314 TGCT patients from four cohorts, including a 48-case microarray. A nonnegative matrix factorization algorithm was applied to identify the “immune factor”, derived the top 150 weighted genes to divide patients into immune and non-immune classes, and further separated the immune class into activated and exhausted subgroups by nearest template prediction. Tumor mutant burden, gene mutation, and copy number alteration were compared with our recently developed package “MOVICS”. A random forest algorithm was performed to establish a prediction model with fewer genes. Immunohistochemistry staining was performed to identify TIMEAS in the microarray.

Results

We constructed the TIMEAS in the TCGA-TGCT cohort and further validated it in the GSE3218 and GSE99420 cohorts. The immune class contained the activated status of T-lymphocytes, B-lymphocytes, and macrophages, while Treg cells and the WNT/TGFβ signature were more activated in the immune-suppressed subgroup. Patients in the immune-exhausted subgroup had the worst prognosis, and 22.9% of patients in the immune-activated subgroup had KRAS mutations, which might stimulate the response of the immune system and lead to a favorable prognosis. The immune-exhausted group benefited more from chemotherapy, while the immune-activated subgroup responded well to anti-PD-1/PD-L1 therapy. FSCN1 was validated as the target of the immune-exhausted microenvironment by immunohistochemistry.

Conclusion

TIMEAS classification can separate TGCT patients; patients in the immune-activated subgroup could benefit more from anti-PD-L1 immunotherapy, and those in the immune-exhausted subgroup are more suitable for chemotherapy.

Supplementary information

The online version contains supplementary material available at 10.1007/s13402-023-00781-1.

Introduction

Testicular germ cell tumor (TGCT) predominantly occurs in young men and occupies 98% of all advanced testicular tumors [1], but it is a rare type in the general population [2]. The number of TGCT patients has been increasing worldwide over the past few years, putting all-aged men at risk of developing testicular cancers. The data also showed that the incidence of TGCTs peaks in adulthood (84% developed between 15 and 44, while only 1% carcinogenesis occurred among boys under 15) [3]. The latest data also demonstrated the significantly distinctive incidence of TGCTs among different regions and races: the lowest incidence was reported in African and Asian countries, whereas the highest incidence was observed in North European countries, and whites are more susceptible to TGCTs than other ethnic groups [4]. According to TGCT incidence data from 1999 to 2012 obtained from 39 US cancer registries, the highest incidence rate was among non-Hispanic white men, with 6.97 per 100,000 man-years, while the lowest incidence rate was 1.20 per 100,000 man-years among Americans.

TGCTs should be defined as a group of heterogeneous diseases with multiple etiopathogeneses and various histological types. 90% of cancers of the testicle start in cells known as germ cells and are further separated into seminoma and non-seminoma. Non-seminoma is detailly consist of embryonal carcinoma, choriocarcinoma, yolk sac carcinoma, and teratoma. Different histological types have different biological behaviors, clinical signatures and prognoses. Seminoma, the most common type of TGCT, is highly sensitive to radiotherapy and chemotherapy and is related to a good clinical outcome [5], which makes TGCT the most curable tumor during the past two decades [6]. Non-seminomas are usually treated by surgery or chemotherapy, and the cure rate can reach 99% under early treatment. The prognosis of choriocarcinoma is significantly worse than that of other germ cell tumors, especially in the advanced metastatic stage. It has also been reported that the level of β-HCG is associated with prognosis [7]. In addition, the essential risk factors that may be involved in the development of TGCTs include cryptorchidism, disorders of sex development, hypo/infertility, contralateral germ cell tumors, and endogenous and exogenous hormones [8]. Moreover, young age is considered to be one of the most frequent factors that contribute to the generation of TGCTs [9]. Although several markers have been reported to have the ability to discriminate various types and might serve as potential therapeutic targets, the main diagnostic gold standards are still serum tumor markers and immunohistochemical markers [10].

Molecular subtype is a novel classification system based on gene signatures and robust statistical methods. It can robustly predict the occurrence and prognosis of tumors and select appreciative patients to accept chemotherapy or radiotherapy. PARKER J S et al. [11] developed a 50-gene risk model known as PAM50, which is beneficial for breast cancer prognosis and chemotherapy prediction. Choi et al. [12] identified 3 molecular subtypes of muscle-invasive bladder cancer based on whole-genome RNA sequencing profiles, and the novel gene signature can predict sensitivity to chemotherapy. Genetic susceptibility loci are considered an important risk factor [13]. However, limited data are available on the classification of TGCTs. In the current study, we tried to reveal the molecular features based on RNA sequencing data, which is from a different angle compared with the already existing histological types, and aimed to provide new insight into progression prediction or to help to search for novel clinical strategies.

Materials and methods

Patient information summary

In this study, we collected a total of 314 TGCT patients for subsequent analysis, whose clinicopathological features, prognostic information, and RNA sequencing profile are all available. The training cohort was The Cancer Genome Atlas-Testicular germ cell tumor (TCGA-TGCT) cohort, including 132 patients. The GSE3218 cohort (74 patients) and the GSE99420 cohort (60 patients) were used as the external validation cohorts. We purchased human TGCT tissue microarrays (48 cases) from Taibosi Biotechnology Co., Ltd. (Xi’an, China). The Human Research Ethics Committees of The First Affiliated Hospital of Anhui Medical University approved the study following the Declaration of Helsinki.

Identification of the subtypes of TIMEAS

As described previously, the nonnegative matrix factorization (NMF) method was utilized to analyze the data of patients from the TCGA cohort. The NMF algorithm was disposed via the formula: V = W × H, where V represents the gene expression matrix, W represents the gene factor matrix, and H represents the sample factor matrix [14]. The immune enrichment score (IES) was generated by single-sample gene set enrichment analysis (ssGSEA). Based on the IES, we preset the numbers of factors as 5 to 10 to obtain the robust immune factor. When the total number of factors was 10 (k = 10), the IES of the eighth factor was higher; this factor was then named the “immune factor”.

The exemplar genes consisted of the top 150 weighted genes that could mostly represent the “immune factor”. According to patients’ different levels of exemplar genes, they were divided into immune and non-immune classes via the module “NMF Consensus”, and the subtyping was further modified via the multidimensional scaling (MDS) random forest method. Furthermore, based on the activated stroma signature, we separated patients of immune class into immune-activated and immune-exhausted subgroups by employing the nearest template prediction (NTP) method [15]. The activated stroma signature includes 24 stromal-associated template genes and 22 non-stromal template genes [15]. For each sample, distances to the template genes are calculated and class assigned based on the smallest distance. To define and verify the above-identified subtypes, we hand-curated gene signatures representing various immune cell types or immune status (Table S1) from the literature and databases. The immune class should contain a higher enriched score of signatures, including the immune enrichment score, immune cell subsets, immune signaling and other immunocyte-associated signatures. The immune-exhausted subgroup should contain a higher stromal enrichment score, such as Treg cells, the TITR signature, MDSCs, the WNT/TGFβ signature, TGFβ1 activated and the C-ECM signatures than the immune-activated subgroup. Non-immune, immune-activated and immune-exhausted subgroups were defined as TIMEAS subgroups. The external validation of the TIMEAS subgroups progressed in the GSE3218 and GSE99420 cohorts. Using the GenePattern module “NMFConsensus”, patients were categorized as immune or non-immune class based on the level of the 150 exemplar genes. Subsequently, using the NTP method, patients were categorized into immune-activated or immune-exhausted subgroups.

Characterize the molecular features

To further confirm the three immunophenotypes, we obtained several gene signatures that can represent immune cell types or host antitumor immunity, and the enrichment score of each gene set for patients was calculated via ssGSEA [16, 17]. The survival probability, TGCT TNM stages, and gene expression levels of several genes in patients with the three immunophenotypes were also analyzed. Tumor mutation burden (TMB), gene mutation waterfall curves, and copy number alteration (CNA) plots were calculated and compared based on our recently developed R package “MOVICS” [18]. The CNA difference among specific observed regions between different immune classes was further evaluated by GISTIC 2.0. The average G-score associated with the amplitude of the CNA and the frequency among each subtype was first calculated, and then the CNAs in specific chromosome areas with significant q-values were regarded as visible differences [19].

Four kinds of TGCT therapy drugs that are frequently used in the clinic, including cisplatin, gemcitabine, paclitaxel, and etoposide, were used to evaluate different effects among the three immunophenotypes. Subclass mapping analysis was employed to verify the therapeutic effect of immunotherapies in the immune class, and the data were from the MD Anderson melanoma cohort and IMvigor 210 cohort.

Dimensionality reduction of the TIMEAS prediction model

According to the random forest (RF) algorithm, the TIMEAS prediction model was created by utilizing the R package “varSelRF”. The exemplar genes among TIMEAS subgroups were obtained by utilizing the R package “limma”, after which informative genes were selected as input data for the RF model (fold change > 2 or < 0.5 with false discovery rate < 0.05). The optimal markers were selected through a backwards elimination procedure. Meanwhile, we set an out-of-bag (OOB) error as the minimization criterion, and the declining fraction of the criterion was set at 0.2. The ultimate gene group was defined as the best predictive model when the OOB error rate was lowest. The predictive model was verified in the TCGA-TGCT, GSE94420 and GSE3218 cohorts via receiver operating characteristic (ROC) curves.

Immunohistochemistry staining

We further validated the TIMEAS types via immunohistochemistry staining (IHC) with a TGCT microarray. PD-1 (anti-PD-1 antibody: Cat. ab182422, Abcam Inc., Massachusetts, USA) was chosen to distinguish the immune and non-immune class, while TGFβ2 (anti-TGFβ2 antibody: Cat. 251,481, ZENBIO Inc., Sichuan, China) was selected to represent stromal status and identify the immune-activated and immune-exhausted subgroups. The concrete procedure has been reported before [20, 21]. The different protein levels of FASN1 (anti-FASN1 antibody: Cat. 382,612, ZENBIO Inc., Sichuan, China) in three subtypes was also detected. All samples were assessed by diverse staining intensity, the scores for which were defined as 0 for negative, 1 for weakly positive, 2 for moderately positive, and 3 for strongly positive. The histochemistry score (H-Score) was calculated using the following formula: H-score = (percentage of weak intensity area ×1) + (percentage of moderate-intensity area ×2) + (percentage of strong intensity area ×3) [22].

Statistical analysis

Student’s t test and Wilcoxon rank-sum test were used to compare normally distributed and non-normally distributed continuous data, respectively. The Kruskal‒Wallis test was utilized when more than two factors were compared. The survival time of the TIMEAS groups was analyzed by the log-rank test and is shown by Kaplan‒Meier plots. Differences were deemed statistically significant when P < 0.05 was applied on both sides. All procedures were processed by GenePattern [23] and R version 4.1.2 (http://www.r-project.org). More details of the methods used in TIMEAS subtyping and characterization were described in our prior articles [16, 24].

Result

Creating immune-related factors and identifying TIMEAS subgroups

In the TCGA-TGCT training cohort, the gene expression files of 132 TGCT patients were virtually microdissected utilizing the NMF algorithm. We preset the total number of factors from 5 to 10, and when the total number was ten, the eighth factor was the robust one that enriched patients with high IES; then, we defined it as the “immune factor” (Fig. 1A). The first 150 weighted genes were viewed as exemplar genes (Table S2). They are correlated with myeloid leukocyte activation, adaptive immune response, lymphocyte activation, leukocyte differentiation, leukocyte migration, and response to interferon-gamma based on ontological analysis of biological processes (Fig. 1B).

Fig. 1.

Immune and non-immune subgroups were separated via the NMF algorithm. A Ten modules were identified via the NMF algorithm, and the eighth module (light green module) with the highest immune enrichment score was defined as an “immune factor”. B Exemplar genes were correlated with myeloid leukocyte activation, adaptive immune response, lymphocyte activation, leukocyte differentiation, leukocyte migration, and response to interferon-gamma. C Immune and non-immune classification modified by the MDS algorithm. D NMF weight, immune score, and immune enriched status between immune and non-immune subgroups. E Signaling pathways were activated in the immune class. F Signaling pathways activated in in non-immune class

Furthermore, using the NMF algorithm, patients in the TGCT cohort were separated into immune and non-immune classes based on the levels of exemplar genes. Subsequently, after MDS random forest was performed, we observed a more precise classification (Fig. 1C). A concrete description of the immune enrichment score, immune factor weight, immune clustering, and immune classes is shown in Fig. 1D.

To verify the classification, we chose some immune-associated signatures and calculated their IESs for each patient via ssGSEA. These signatures are correlated with immune cells, such as T-lymphocytes, B-lymphocytes, and macrophages etc. The results showed that patients in the immune-associated subgroup had a higher IES than the other patients (P < 0.05, Fig. 2A). KEGG signaling pathways in both immune and non-immune classes were analyzed by GSEA. Immune-associated pathways were more active in the immune class (Fig. 1E), while immune-non-associated pathways were more active in the non-immune class (Fig. 1F).

Fig. 2.

Molecular distribution, immune checkpoints and diverse prognosis among the three identified TIMEAS subgroups. A Molecular feature distribution of TIMEAS subgroups. TLS, tertiary lymphoid structure; CYT, cytolytic activity score; TITR, tumor-infiltrating Tregs; MDSC, myeloid-derived suppressor cell; C-ECM, cancer-associated extracellular matrix. B Gene expression of immune checkpoints and immune suppress markers among TIMEAS subgroups; C Distribution of TNM stages among TIMEAS subgroups. D OS of patients in the TIMEA subgroups. E Recurrence-free survival of TIMEAS subgroups

Two distinct subgroups highlighted by different microenvironmental conditions

Fibroblasts, the extracellular matrix (ECM), and mesenchymal stromal cells (MSCs) are essential tumor stroma components that can support and connect cells. As tumors progress, activated stromal components cause genetic and epigenetic changes that affect tumor cells. Fibroblasts can regulate cancer-associated ECM (C-ECM), which affects the level of Tregs and immunosuppressive cells. MSCs, one kind of inherent regulator of tumors, can regulate the levels of PD-L1 and Treg cells [25]. Myeloid-derived suppressor cells (MDSCs) could act as the mirror of the immune status of the tumor environment, and TGFβ could act as the immunosuppressor of the tumor environment [26]. On this basis, based on activated stromal signatures, patients of immune class were separated into immune-activated and immune-exhausted subgroups. In the immune-exhausted subgroup, the expression levels of the stromal enrichment score, Treg cells, TITR signature, MDSC, WNT/TGFβ signature, TGFβ1 activated and C-ECM signatures were higher than those in the immune-activated subgroup. For patients in the TCGA-TGCT cohort, 39.39% (52/132) were separated into the immune-activated subgroup, while 12.88% (17/132) were separated into the immune-exhausted subgroup (Fig. 2A). TIM-3, TGFβ1, TGFβ 2, IL-11, and LAG3 have been reported to correlate with immune exhaustion status in previous reports [24, 27, 28]. Siglecs (sialic acid-binding immunoglobulin-like lectins) are mainly found on immune cells and are engaged in inhibitory cell signaling; Siglec-8 is the only molecule expressed on mast cells and eosinophils [29]. Then, we evaluated the gene expression of PD-1, PD-L1, PD-L2, IFN-γ, SIGLEC8, TIM3, TGFβ1, TGFβ2, IL-11, and LAG3 and discovered that the expression of TIM-3, TGFβ2 and IL-11 in the immune-exhausted subgroup was higher than that in the immune-activated subgroup, while the expression of PD-1, PD-L1, PD-L2, IFN-γ, and SIGLEC8 in the immune-exhausted subgroups was lower (Fig. 2B).

Clinical features among the three immunophenotypes were evaluated. We observed that there was no difference in the Race distribution among the TIMEAS subtypes, while non-immune class contained more patients with advanced tumor stage. The immune-activated subgroup comprised most seminoma cases, and the major histologic type for non-immune class and immune-exhausted subgroup were non-seminoma tumor (Fig. 2C; Table 1). We also observed that although non-immune class contained more patients with stage III, it still faced better clinical outcomes than the immune-exhausted subgroup, which supports our hypothesis that the immune-exhausted microenvironment leads to poor prognosis.

Table 1.

Distribution of clinical parameters among the three TIMEAS subtypes

Parameters	Non-immune (n = 63)	Immune-exhausted (n = 17)	Immune-activated (n = 52)	P value
Race (%)
Asian	2 (3.3%)	0 (0.0%)	2 (4.0%)	0.520
Black or African American	3 (5.0%)	2 (11.8%)	1 (2.0%)
White	55 (91.7%)	15 (88.2%)	47 (94.0%)
Stage (%)
Tis	24 (40.7%)	7 (43.8%)	15 (30.0%)	0.037*
Stage I	20 (33.9%)	4 (25.0%)	29 (58.0%)
Stage II	5 (8.5%)	4 (25.0%)	3 (6.0%)
Stage III	10 (16.9%)	1 (6.2%)	3 (6.0%)
Histologic type (%)
Mixed	3 (4.8%)	1 (5.9%)	4 (7.7%)	< 0.001*
Non-Seminoma	48 (76.2%)	13 (76.5%)	1 (1.9%)
Seminoma	12 (19.0%)	3 (17.6%)	47 (90.4%)

*, P < 0.05

For OS, we found that patients in the immune-exhausted subgroup had poorer OS than those in the other two subgroups (P = 0.019, Fig. 2D). Regarding recurrence-free survival, we observed that patients in the immune-activated subgroup had a longer recurrence-free survival time (P = 0.200, Fig. 2E). In the TCGA-TGCT cohort, approximately 75.81% of seminoma patients were separated into the immune-activated group, but the other 24.19% of seminoma patients belonged to the nonimmune or immune-exhausted groups and met the poor prognosis (Fig. S2). Therefore, we have to say the TIMEAS subgroups are not tightly consistent with histological types but a novel insight from the immune activation angle. After adjusting for clinical features, we revealed that histological type and TIMEAS were independent prognostic factors for the OS outcome of TGCTs (P = 0.0216, Table 2). The recurrence-free survival time is important for patients expecting and guiding clinical treatment. We observed that the 5-year recurrence-free rate was as high as 81.5%, while non-immune class was 72.0%, falling down to 68.2% in the immune-exhausted subgroup (Table 3).

Table 2.

Independent overall survival prognostic value of TIMEAS subtypes

TCGA-TCGT cohort
Parameters		HR	Lower 95% CI	Upper 95% CI	P value
Age		0.92	0.77	1.10	0.349
Stage	Tis	-	-	-
	Stage I	0.00	0.00	Inf	0.998
	Stage II	0.00	0.00	Inf	0.999
	Stage III	4.95	0.42	58.11	0.203
Histological type	Mixed	-	-	-
	Non-Seminoma	6.23E + 08	5.55E + 07	6.99E + 09	< 0.001*
	Seminoma	2.13E + 09	1.90E + 08	2.39E + 10	< 0.001*
TIMEAS	Activated + Non-Immune	-	-	-
	Exhausted	21.66	1.94	242.30	0.013*
GSE3218 cohort
Parameters		HR	Lower 95% CI	Upper 95% CI	P value
IGCCCG	Good	-	-	-
	Intermediate	0.438	0.19	1.009	0.052
	Poor	0.34	0.116	0.991	0.048*
Histological type	Mixed	-	-	-
	Non-Seminoma	0.485	0.166	1.416	0.185
TIMEAS	Activated + Non-Immune	-	-	-
	Exhausted	2.521	1.286	4.942	0.007*

*, P < 0.05

Table 3.

Five- and ten-year recurrence-free survival for the three  TIMEAS subtypes

	Time	Survival rate	Lower 95% CI	Upper 95% CI
Non-Immune	5 years	72.00%	61.10%	84.80%
	10 years	68.60%	56.70%	82.90%
Exhausted	5 years	68.20%	48.60%	95.70%
	10 years	51.10%	26.40%	98.90%
Activated	5 years	81.50%	67.50%	98.30%
	10 years	67.90%	45.30%	100.00%

The TMB of patients was evaluated by calculating the number of nonsynonymous mutations per million bases. The results indicated that there were no evident differences in TMB between the immune and non-immune classes (Fig. 3A).

Fig. 3.

Genetic alteration among TIMEAS subgroups. A TMB levels of TIMEAS subgroups. B The distribution of KIT and KRAS gene mutations, ages, TNM stages, and histological types among TIMEAS subgroups. C Different KIT expression levels between wild-type (WT) KIT and mutated KIT. D Estimated IC50 of imatinib among TIMEAS subgroups. E Copy number alteration among TIMEAS subgroups. F G-scores of chromosome alterations in the nonimmune class. G Chromosome alterations in specific regions in TGCT patients in non-immune class

However, the specific genes (KIT and KRAS) were observed to have different genetic alteration landscapes among the three immunophenotypes. Notably, KIT mutations occurred in 15% of patients, while KRAS mutations occurred in 9% of patients. KIT and KRAS mutations predominantly occurred in the immune-activated subgroup; 22.9% of patients in the immune-activated subgroup had KRAS mutations, and 33.3% had KIT mutations, which might stimulate the response of the immune system and lead to a favorable prognosis (Fig. 3B, Table S3). Furthermore, regarding KIT mutation status, the mRNA expression level of mutated KIT was higher than that of WT KIT (wild-type KIT) (P < 0.001, Fig. 3C). Imatinib is a potent and selective inhibitor of KIT, and the IC50 of imatinib in the immune-activated subgroup was lower than that in the immune-exhausted and non-immune groups (Fig. 3D, P < 0.007), indicating that imatinib is more beneficial to patients in the immune-activated subgroup.

To further explore genetic alterations, we distinguished gained and lost CNA types among different subgroups and found that the immune-exhausted subgroup had a lower frequency of gained CNA as well as total CNA, while non-immune class had a higher frequency of CNA (Fig. 3E). Furthermore, we explored which chromosome site led to this result. The output was that gained CNA of non-immune class mainly occurred on 12p13.32, 12p11.21, 12p11.1, 12q15, and 4q12, while lost CNA mainly occurred on 18q23, 11q25, 4q22.1, 10q26.3, 9p24.3, 6p22.3, and 1p36.32 (Fig. 3F, G). In the immune-exhausted subgroup, the lost mutation occurred on 11q24.1 with statistical significance. In the immune-activated subgroup, the gained mutation occurred on 12p12.1, 12p13.31, and 8q11.23, while the lost mutation occurred on 11q24.3 and 10q26.3 with statistical significance (Fig. S1).

Validation of the three immunophenotypes in external cohorts

Moreover, to verify the TIMEAS types defined above, two external cohorts (GSE3218 and GSE99420) were employed with RNA sequencing profiles. Exemplar genes and stromal activation status defined previously were utilized to divide patients into TIMEAS subgroups. The exemplar genes, which are the 150 most weighted genes in the “immune factor” identified based on the TCGA-TGCT training cohort, were used to first distinguish the immune and non-immune subclasses, and then the NTP algorithm with activated stroma signature was applied to reveal the immune-activated and immune-exhausted subtypes. It was validated with the ssGSEA scores of Th17 cells, Treg cells, the TITR signature, MDSCs, the WNT/TGFβ signature, and the C-ECM signature among TIMEAS subgroups. As expected, patients in the immune class had higher IES and immune signaling signatures, especially the immune-exhausted subgroup. In the GSE3218 cohort, 36.49% (27/74) of patients belonged to non-immune class, while 20.27% (15/74) and 43.24% (32/74) belonged to the immune-activated and immune-exhausted subgroups, respectively (Fig. 4A). TIMEAS acted as the independent risk predictor after adjusting the impact of IGCCCG type and Histological type in GSE3218 cohort (P = 0.007, Table 2). In the GSE99420 cohort, 48.33% (29/60) of patients belonged to non-immune class, while 28.33% (17/60) and 23.33% (14/60) belonged to the immune-activated and immune-exhausted subgroups, respectively (Fig. 4D). For the GSE99420 cohort, seminoma patients were almost equally divided into three parts and belonged to the three subgroups of TIMEAS. The GSE3218 cohort contains 70 non-seminoma patients and 4 patients with mixed histological types, all the immune-activated patients are with non-seminoma histology (Fig. S2).

Fig. 4.

Reproduce the TIMEAS subgroups in external TGCT cohorts. A Reproduce the TIMEAS subgroups in 74 TGCT patients from the GSE3218 cohort and feature signatures. B Survival probability of patients among TIMEAS subgroups in the GSE3218 cohort. C Survival probability of patients between exhausted and non-exhausted subgroups in the GSE3218 cohort. D Patients in the GSE99420 cohort were separated into TIMEAS subgroups. Reproduce the TIMEAS subgroups in 60 TGCT patients from the GSE99420 cohort and feature signatures

Remarkably, based on the results obtained above, patients in the immune-exhausted subgroup of these two cohorts showed the worst OS among TIMEAS subgroups, while the data between non-immune and immune-activated groups were approximate (Fig. 4B, P = 0.19). Then non-immune and immune-activated subgroups were viewed as non-exhausted subgroup. The difference in OS between exhausted and non-exhausted subgroups was almost marginally significant (Fig. 4C, P = 0.069).

The immune-exhausted subgroup benefits more from chemotherapy, while the immune-activated subgroup benefits more from anti-PDL1 therapy

To explore the responses of the TIMEAS types to different anti-tumor therapies, we compared the IC50 values of four common chemotherapeutics (cisplatin, gemcitabine, paclitaxel, and etoposide) in the TCGA-TGCT cohort, GSE3218 cohort, and GSE99420 cohort. Notably, the IC50 of chemotherapies in the immune-exhausted subgroup of all cohorts was lower than that in the immune-activated and non-immune subgroups (Fig. 5A). The results indicated that the immune-exhausted group benefited more from chemotherapies.

Fig. 5.

Identify the appropriate chemotherapies and immunotherapies for patients from different TIMEAS subgroups. A IC50 of cisplatin, gemcitabine, paclitaxel, and etoposide among TIMEAS subgroups in TCGA-TGCT, GSE3218, and GSE99420 cohorts. B Response of patients in the activated and non-activated subgroups to immunotherapy evaluated by the TIDE algorithm. C Prediction of the response to CTAL4, PD-1, and PD-L1 between the activated and non-activated subgroups

Based on the transcriptomic data, the TIDE algorithm was utilized to evaluate the immunotherapy responders and explore the predictive effect of “immune factor”. The output showed that the percentage of immunotherapy responders was 21% in the immune-activated subgroup, which was higher than non-activated subgroup (4%) (Chi-square test, P = 0.003, Fig. 5B).

To confirm that patients in the immune-activated subgroup responded well to immunotherapy, we employed Submap analysis to display the responders distributed in the activated and non-activated subgroups in the MD Anderson cohort and IMvigor210 cohort. In the MD Anderson cohort, only patients in the immune-activated subgroup responded well to anti-PD-1/PD-L1 therapy, and some patients of non-activated subgroup did not respond to anti-CTLA4 therapy. In the IMvigor210 cohort, patients in the immune-activated subgroup responded well to anti-PD-1/PD-L1 therapy, while patients of non-activated subgroup did not (Fig. 5C).

Dimensionality reduction of the TIMEAS prediction model

TIMEAS was created by a two-step process. Patients were first classified into immune and non-immune classes according to 150 exemplar genes via NMF consensus. Then, patients in the immune class were ultimately classified into immune-activated and immune-exhausted subgroups based on stromal-associated genes via the NTP method. In the clinic, panels with fewer genes are more applicable. Therefore, 25 immune-related genes were chosen by the RF algorithm to create a predictive model of immune and non-immune classes in the TCGA training cohort, while three exhausted-related genes were chosen for immune-activated and immune-exhausted subgroups (Table S4). These 25 immune-related genes could distinguish patients between immune and non-immune classes in the GSE99420 validation cohort with 0.952 accuracy and 0.921 in the GSE3218 validation cohort (Fig. 6A). The three exhausted-related genes (BCAT1, FSCN1, TNFRSFF10B) could distinguish patients between immune-activated and immune-exhausted subgroups in the GSE99420 validation cohort with 0.807 accuracy and 0.796 in the GSE3218 cohort (Fig. 6B). The concrete distribution of patients in each cohort is shown in Fig. 6C and D. We also uploaded the R code and the Rdata on GitHub (https://github.com/AHMUJia/TIMEAS) to help the readers generate the TIMEAS subgroups in the external cohort for further study.

Fig. 6.

Verify the accuracy of the dimensionality-reduced predictive model. A The accuracy of the predictive model in distinguishing immune-enriched and immune-zero subgroups. B The accuracy of the predictive model in distinguishing immune-activated and immune-exhausted subgroups. C Patients in the immune-enriched and immune-zero subgroups derived from RF compared with the actual values. D Patients of immune-activated and immune-exhausted subgroups derived from RF compared with actual

FSCN-1 was chosen to identify the subtypes of TIMEAS

PD-1 and PD-L1 have been reported to identify immune and non-immune classes, while TGFβ2 and IL-11 can identify immune-activated and immune-exhausted subgroups [24]. Therefore, we explored the connection between the three exhausted-related genes and these four genes, and the interrelation between each gene is shown in Fig. 7A. Furthermore, we assessed the expression of the three exhausted-related genes and found that the immune-exhausted subgroup had higher expression of all three genes than the immune-activated subgroup, and the expression of FSCN1 was highest among these three genes in the immune-exhausted subgroup (Fig. 7B). We found that patients with higher FSCN1 expression had shorter OS. (P = 0.0096, Fig. 7C).

Fig. 7.

FSCN1 was chosen as the biomarker of immune exhaustion and validated via IHC. A Correlation expression of PD-1, PD-L1, TGFβ2, TNFRSF10B, BCAT1, FSCN1, and IL-11. B The expression of BCAT1, FSCN1 and TNFRSF10B among TIMEAS subgroups. C Survival probability of patients with high or low FSCN1. D IHC results of PD-1, TGFβ2, and FSCN1 among TIMEAS subgroups. E The subgroups and clinical features of microarray cases. F H-scores of PD-1 levels among TIMEAS subgroups. G H-scores of TGFβ2 levels among TIMEAS subgroups. (H) H-scores of FSCN1 levels among TIMEAS subgroups

Therefore, we chose PD-1 as the marker of immune class and TGFB2 as the distinguishing factor of the immune-activated or immune-exhausted subgroup. We evaluated the TIMEAS subgroups by IHC based on TGCT tumor microarray, which contained 35 cases of seminoma and five cases of lymphoma. Then, IHC was performed to detect the expression of PD-1, TGFβ2, and FSCN1. H-scores of PD-1 and TGF β2 were further used to separate patients into TIMEAS subgroups (Fig. 7D). Patients were separated into immune or non-immune class based on the median H-score of PD-1, and stromal scores represented by TGFβ2 were also separated by the median H-score. Finally, 11 cases belonged to the immune-activated subgroup, 9 belonged to the immune-exhausted subgroup, and the remaining 20 cases were regarded as non-immune class (Fig. 7E). PD-1 protein level is higher in both activated and exhausted subgroups than non-immune class (P < 0.001, Fig. 7F), while the level of TGFβ2 in exhausted subgroup is higher than activated subgroup (P < 0.001, Fig. 7G), and the level of FSCN1 protein in exhausted subgroup was also higher than activated subgroup (P = 0.0008, Fig. 7H).

Based on the results shown in Fig. 7, we confirmed TIMEAS subtyping based on IHC staining for the PD-1 and TGFβ2 proteins. The immune-exhausted specific marker FSCN1 was also validated in the IHC-based TIMEAS subtyping of the TGCT microarray.

Discussion

TGCTs, the most common malignant tumors occurring in the testis, are most common in young men, especially those aged 30–35 [30]. Reviewing recent decades, the trend of incidence is increasing. Generally, they are diagnosed by identifying different histological subgroups [31]. However, they often represent complex histology, so a careful microdissection is needed to explore genes correlated with particular histology [32]. TGCTs appear to have high sensitivity to chemotherapies, especially the commonly used cisplatin, but approximately 30% of patients are resistant to standard chemotherapy, and several cases of acute toxicity and chronic side effects have been reported [31, 33]. Recent studies have shown that immune infiltration is correlated with the histology and prognosis of TGCT patients, and immunotherapy based on anti-PD-1/PD-L1 is beneficial to TGCT patients [31, 34].

NMF, one of the most robust methods for clustering and feature selection, is a linear dimensionality technique applied to non-negative data with applications such as audio source separation and image analysis [35, 36]. The NTP method can make class predictions by assessing the prediction confidence of each patient’s gene expression data, and it has been applied for clinical classification and outcome prediction [37]. In this study, NMF and NTP algorithms were utilized to develop a robust classifier of TIMEAS. First, 132 patients in the TCGA-TGCT training cohort were separated into immune and non-immune classes according to the expression levels of 150 exemplar genes via the NMF algorithm. Second, patients of immune classes were separated into immune-activated and immune-exhausted subgroups according to the activated stroma signature via the NTP algorithm. The TIMEAS types were characterized by confirmative signatures or immune signaling pathways. Signatures of T cells, B cells, IFN, and CYT were higher in the immune class [38], while stromal enrichment scores and several signatures associated with immune-exhausted status were higher in the immune-exhausted subgroup [16]. Additionally, activated KEGG signaling pathways associated with immunity were enriched in the immune class. These signatures and signaling pathways indicated that the classifier could well dissect patients into TIMEAS subgroups.

The TMBs of TIMEAS are approximate, indicating that the mutation levels of TIMEAS have no evident difference. KIT and KRAS have been reported to be well-known mutated oncogenes found in TGCTs [33]. KIT encodes the tyrosine kinase transmembrane receptor and is essential in testicular development, and mutations in KIT mainly occur in seminomas [39, 40]. Activating KRAS mutations contributes to hyperactivation of the MAPK pathway and the PI3K-AKT pathway, and they are correlated with tumorigenesis in many cancers [41]. In our study, KIT and KRAS mutations predominantly occurred in the immune-activated subgroup, the histological type was mainly seminoma, and the tumors were mostly in stage I. Imatinib is a KIT inhibitor that is most beneficial for KIT exon 11 deletion mutations and has been applied to metastatic gastrointestinal-stromal-tumor [42, 43]. The IC50 of imatinib in the immune-activated subgroup was lower, indicating that patients in the immune-activated subgroup could benefit more from imatinib, which was correlated with the higher KIT mutation. It has been known for decades that gain of chromosome 12 in TGCT, usually presented as isochromosome 12p, leads to the pathogenesis of TGCT [33, 44]. Graham Bignell et al. [45] also found losses of chromosomes 13q, 18q, 11q, and 4q. These results could explain the higher number of mutations and better therapeutic effect of imatinib in the immune-activated subgroup.

Cisplatin, gemcitabine, paclitaxel, and etoposide are four kinds of chemotherapy drugs that are beneficial for TGCT patients [30, 46, 47]. The IC50 values of these chemotherapy drugs were evaluated in both the training and external cohorts. The immune-exhausted subgroup had a lower IC50 among all groups, while the immune-activated subgroup had a slightly higher IC50, which could indicate that patients in the immune-exhausted subgroup show higher sensitivity to chemotherapies than patients in the immune-activated subgroup. Therefore, therapies for immune-activated subgroups are worthy of further exploration. The results of the TIDE and Submap algorithms showed that more patients in the immune class showed a response to immunotherapy, especially anti-PDL1 therapy. This result proves that immunotherapy is a new direction for immune-activated TGCTS.

A smaller group of genes in the prediction model is more suitable for clinical application. Hence, we performed dimensionality reduction of the TIMEAS prediction model and obtained a smaller number of genes, which could help divide patients into different subgroups with high sensitivity and specificity. BCAT1, FSCN1, and TNFRSF10B were selected to identify immune-activated and immune-exhausted subgroups. The expression of FSCN1 was not only higher than that of the other genes but also obviously different from that of the other subgroups. FSCN1 has been reported to be a possible therapeutic target in human tongue squamous cell carcinoma [48], a novel prognostic biomarker for renal cell cancer patients after nephrectomy [49], and a crucial predictor of early-stage breast cancer [50]. Therefore, FSCN1 was selected for further exploration. TGCT patients with high FSCN1 expression had a lower survival probability. Comparing the PD-1, TGFβ2, and FSCN1 levels of H-scores among TIMEAS types, we found that the H-scores of patients in the immune-exhausted subgroup were all higher than those in the immune-activated subgroup, which could well explain the lower OS of this subgroup. Testicular tumors are very diverse and one of the most challenging parts for pathologists. With recent advances in molecular techniques and a large amount of clinical samples, new genetic information and sequencing data have been presented in recent decades, which can help to reacquaint the molecular characteristics and point out new directions of tumor pathogenesis. In the current study, immune-activated and immune-exhausted subgroups were revealed, and FSCN1 was regarded as a marker of immune exhaustion, which can help pathologists identify TGCT samples with non-activation immune environment features and poor prognosis, which can support the current histology system.

Taken together, a novel classifier was defined and validated to divide 266 TGCT patients into TIMEAS subgroups, and it could be applied in the clinic after dimensionality reduction. Patients in the immune-activated subgroup could benefit more from anti-PD-L1 immunotherapy. Although patients in the immune-exhausted subgroup showed high sensitivity to chemotherapy, they had a low survival probability. FSCN1 could be a potential immune suppress gene for TGCT patients, so a corresponding inhibitor might have some efficacy to save the immunotherapy failure patients.

Supplementary information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, MJL, GJJ and LCZ; methodology, MJL, GR, LXF, and ZXS; formal analysis, MJL, LXF, YFR and WHT; writing the original draft, LX, LY and HZY; visualization, MJL, LX, LXF and WHT; funding acquisition, HZY, ZXS and LCZ; supervision, LCZ and ZXS.

Funding

This work was supported by the National Natural Science Foundation of China [grant numbers: 82170787, 82071637, 81973145]; the Supporting Project for Distinguished Young Scholars of Anhui Colleges [grant number: gxyqZD2019018]; the National Key R&D Program of China (2019YFC1711000), and the Key R&D Program of Jiangsu Province [Social Development] (BE2020694).

Data availability

The raw data for this study were generated at the corresponding archives. Derived data supporting the findings are available from the corresponding author [LCZ] upon reasonable request.

Declarations

Ethics approval and consent to participate

Ethical approval for the microarray was obtained from the Ethics Committee of the First Affiliated Hospital of Anhui Medical University (PJ-2022-06-36), and patient consent for the retrospective cohorts was waived. As the other data used in this study are publicly available, no ethical approval was needed.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Conflicts of interest

The authors have no conflicts of interest.