Background
Primary bone cancer is a rare malignant tumor that originates in the bones and accounts for far less than 1% of all cancers [
1,
2]. Osteosarcoma is the most frequent malignancy of primary bone cancer and commonly occurs in children and teenagers [
3]. Currently, comprehensive treatment strategies for osteosarcoma include surgery, chemotherapy, radiotherapy and targeted therapy [
4]. Most often, chemotherapy is the most frequently used strategy for patients with osteosarcoma before surgery. Neoadjuvant chemotherapy shrinks the tumor and makes surgery easier in patients with osteosarcoma [
5]. Before the introduction of chemotherapy, approximately 90% of patients with osteosarcoma developed distant metastases, which is the leading cause of death for patients with osteosarcoma [
6]. The advent of chemotherapy has dramatically improved the 5-year overall survival (OS) rate from less than 20 to 60% in patients with osteosarcoma [
7]. However, therapy resistance and chemoinsensitivity remain major challenges for osteosarcoma treatment. Despite numerous clinical trials attempting to improve the outcomes of poor responders by modifying chemotherapy regimens and intensifying postoperative therapy, their prognosis remains poor [
8‐
10].
The response to chemotherapy significantly influences the incidence of local recurrence and prognosis of osteosarcoma [
11]. Patients with osteosarcoma who have a good response to chemotherapy represent a subset of patients with osteosarcomas exhibiting innate sensitivity to treatment, resulting in a superior response and good prognosis. Patients who display a poor response to chemotherapy have a much higher risk of recurrence and poor outcomes even after complete resection of the primary tumor [
12]. Tumor necrosis induced by chemotherapy remains the only reliable method to assess the effects of chemotherapy [
13]. However, the tumor necrosis rate is only able to be detected in surgically resected specimens, and histological response monitoring during the course of chemotherapy treatment is impossible. Thus, developing a predictive signature that discriminates patients who will respond to chemotherapy before treatment may help to improve the efficacy of chemotherapy in patients with osteosarcoma.
Recently, with the development of deep sequencing techniques and gene microarrays, bioinformatic analyses of genome-wide expression profiles have been broadly performed to distinguish patients who will or will not respond to chemotherapy [
14‐
16]. Currently, multigene signatures for the prediction of the chemotherapy response have been widely employed using gene expression analysis with gene microarrays or deep sequencing techniques [
17]. Here, we strived to develop a predictive model based on gene expression profiles for predicting the response to chemotherapy in patients with osteosarcoma. The tumor immune microenvironment (TIME) plays a critical role in regulating tumor progression and the response to chemotherapy [
18]. An increasing number of studies have reported that a higher level of intratumor immune cell infiltration is associated with a better response to chemotherapy [
19‐
21]. Therefore, we reasoned that a predictive signature based on IRGs would likely be more strongly predictive of chemotherapy responsiveness in patients with osteosarcoma.
In this study, we attempt to construct a signature with the capability to predict the chemotherapy response and prevent patients who are not sensitive to chemotherapy from receiving ineffective treatment. In this study, we systematically surveyed tumor-infiltrating immune profiles between good responders and poor responders employing CIBERSORT and ssGSEA using data from the TARGET and GEO databases. We explored the predictive value of IRGs for the chemotherapy response. Finally, we established a novel immune gene-based signature that may be clinically useful for predicting responsiveness to chemotherapy in patients with osteosarcoma. Furthermore, we aimed to obtain a deeper understanding of the underlying mechanism of chemosensitivity and chemoresistance that could lead to the development of personalized medicine for patients with osteosarcoma.
Methods
Data acquisition and preprocessing
The Therapeutically Applicable Research to Generate Effective Treatments (TARGET) program applies a comprehensive genomic approach to determine molecular changes that drive paediatric cancers. The goal of TARGET data is to facilitate the discovery of therapeutic targets for paediatric cancers and facilitate the rapid translation of those findings into clinical applications. Gene Expression Omnibus (GEO) is a public database repository of high-throughput gene expression data and other functional genomics datasets.
The patients who fulfilled the following inclusion criteria were selected from the TARGET and GEO databases: 1) gene expression was detected using a microarray or high-throughput sequencing, and 2) the histological response to chemotherapy was recorded. One hundred thirty-seven patients who fulfilled the abovementioned criteria were selected from four datasets (TARGET, GSE14827, GSE87437 and GSE39055). The TARGET cohort was used as the training dataset to construct the signature for the prediction of chemotherapy response. The GSE14827 and GSE87437 cohorts served as validation datasets. The GSE14827 and GSE87437 cohorts were annotated according to the Affymetrix Human Genome U133 Plus 2.0 Array platforms. The “sva” R package containing the “Combat” function was applied to remove batch effects. Clinical information of patients with osteosarcoma was obtained from the TARGET and GSE39055 cohorts.
Histological evaluation of the chemotherapy response
The chemotherapy response was assessed by comparing tumor necrosis in the biopsy specimens or surgically resected specimens. Tumor necrosis was graded as follows: grade IV (100% necrosis), grade III (90–99% necrosis), grade II (50–89% necrosis), and grade I (0–49% necrosis). In this study, patients with a tumor necrosis rate less than 90% were defined as poor responders, and patients showing a tumor necrosis rate of 90% or greater were defined as good responders.
Evaluation of tumor-infiltrating immune cells and the immune infiltration level
The CIBERSORT algorithm was used to quantify the proportions of immune cells in a mixed cell population in both the training and validation cohorts [
22]. The normalized gene expression data were analysed to calculate the abundance ratio matrix of 22 immune cell types in each sample, including macrophages (M1 macrophages, M2 macrophages, and M0 macrophages), T cell types (T follicular helper (Tfh) cells, resting memory CD4 T cells, activated memory CD4 T cells, γδ T cells, CD8 T cells, Tregs, and naïve CD4 T cells), resting natural killer cells, activated NK cells, resting mast cells, activated mast cells, memory B cells, resting dendritic cells, activated DCs, naïve B cells, monocytes, plasma cells, neutrophils and eosinophils [
23]. The CIBERSORT results of samples with
p < 0.05 indicated that the inferred fractions of immune cell populations were accurate and were eligible for further analysis. The CIBERSORT output estimates were normalized, and immune cell type fractions were summed to one. Spearman’s correlation analysis was performed to identify the relationship of each immune cell type.
Single-sample gene set enrichment analysis (ssGSEA) was applied to quantify the enrichment levels of the 29 immune signatures in each osteosarcoma sample [
24]. Next, the ESTIMATE algorithm was used to calculate the immune cell infiltration level (immune score) for each osteosarcoma sample with the “ESTIMATE” R package [
25]. In addition, patients with osteosarcoma were divided into high and low immune infiltration subtypes based on the median value of immune scores.
Identification of differentially expressed genes and functional enrichment analyses
The differentially expressed genes between good responders and poor responders were screened using the “limma” R package. A p value < 0.05 and | log FC| ⩾ 0.5 were set as the cut-off criteria.
Functional enrichment analyses were performed to investigate the possible molecular mechanisms of differentially expressed genes using the “clusterProfiler” package. Terms identified in GO and KEGG analyses with a false discovery rate (FDR) < 0.05 were considered statistically significant and were visualized using the “digest” and “GOplot” packages.
Construction of an immune-related signature for the prediction of the chemotherapy response
The comprehensive list of IRGs was downloaded from the Immunology Database and Analysis Portal (ImmPort) database (
https://immport.niaid.nih.gov), which shares immunology data and provides a list of IRGs for cancer researchers.
Least absolute shrinkage and selection operator (LASSO) logistic regression analysis was used to identify and select the optimal genes in the training cohort with the “glmnet” R package. The parameter λ selection in the LASSO model was tuned using ten-fold cross-validation. Next, a binary logistic regression analysis was carried out to discover the genes with predictive value for the chemotherapy response, and the selected genes with a nonzero coefficient were combined into a gene signature. Each gene in the signature has a regression coefficient (β) for predicting chemotherapy response that was calculated as follows: index = (expression of gene 1)*β1 + (expression of gene 2)*β2 + (expression of gene 3)*β3...… + (expression of gene n)*βn, where n indicates the number of the gene.
Principal component analysis (PCA) was used to examine the distribution of samples. The PCA plot was drawn across the first two principal components. A receiver operating characteristic (ROC) curve was generated to obtain the area under the curve (AUC) accuracy and sensitivity in the training and validation cohorts using the “pROC” R package and to evaluate the discriminative power of the signature.
Gene set enrichment analysis
We applied a gene set enrichment analysis (GSEA) between the good response and poor response groups separated by the immune-related predictive signature using clusterProfiler and the enrichplot R package. Two functions (gseGO and gseKEGG) were applied to identify the enriched terms in Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases, and p < 0.05 and FDR < 0.25 were considered statistically significant in the GSEA.
Statistical analysis
In this study, all statistical analyses were conducted using R software (version 3.6.5). Continuous variables were compared using Student’s t-tests. Survival analyses were conducted using the Kaplan–Meier method with the log-rank test by the “survival” R package. All statistical analyses were two-sided, and statistical significance was defined at p values < 0.05. For the hierarchical clustering analysis, Pearson’s correlation coefficients and the averaging method were used, and the results were shown in heat-maps.
Discussion
Chemotherapy is an important treatment option that can improve the prognosis of patients with osteosarcoma. High-dose methotrexate, doxorubicin, and cisplatin are the backbones of chemotherapy regimens [
33]. However, a significant proportion of patients still do not respond to chemotherapy and experience serious side effects [
34]. Therefore, the ability to accurately predict the response to chemotherapy is crucial for the development of appropriate treatments for osteosarcoma. To date, the use of genome-wide expression data to predict the response to chemotherapy in patients with osteosarcoma is still in its infancy. In this study, we developed and validated a five immune-related gene signature for the prediction of the chemotherapy response in patients with osteosarcoma. The identified immune-related predictive signature, including CD70, TNFRSF9, EGFR, PDGFD and S100A6, will facilitate the development of personalized therapy for osteosarcoma and provide new insights into the role of the tumor immune microenvironment in regulating patients’ responses to chemotherapy.
The tumor immune microenvironment is a complex network consisting of immune cells, cytokines and fibroblasts that plays important roles in both cancer treatment and prognosis [
35,
36]. In recent years, accumulating evidence has documented the critical role of the immune microenvironment in the determination of the therapeutic response to chemotherapy. Research suggests that tumor-infiltrating immune cells are related to the chemotherapy response in various tumors [
20,
37,
38]. For example, Denkert et al. reported that immune cell infiltration was relevant to the response to neoadjuvant chemotherapy and the prognosis after adjuvant chemotherapy in patients with breast cancer [
37]. In the present study, we revealed that high-level immune cell infiltration was strongly correlated with a good response to chemotherapy, and CD8 T cells, T follicular helper (Tfh) cells and regulatory T cells were increased in the good responder group. Previous studies have suggested that Tfh cells induce antitumor immunity by activating the effector functions of CD8 T cells [
39], and CD8 T cells are the key cell type responsible for the elimination of tumor cells during cancer immunosurveillance [
40]. Therefore, the finding that higher levels of CD8 T cells and Tfh cells were strongly associated with a good response was not surprising. Furthermore, chemotherapy has the ability to evoke immunogenic cell death [
41,
42], which may partially explain the positive role of CD8 T cells and Tfh cells in the good responder group [
43]. In addition, these results suggested that the combination of chemotherapy with immunotherapy may be beneficial for patients with osteosarcoma exhibiting poor response to chemotherapy. Our results revealed that tumor-infiltrating lymphocytes play a crucial role in the tumor response to chemotherapy. Cytotoxic drugs have the potential to increase immunogenicity by triggering immunogenic cell death, which provides therapeutic opportunities for combinations with immunotherapy. Doxorubicin and cyclophosphamide have the powerful capacity to activate immunogenic cell death and are thus attractive for use in combination with immunotherapy [
44]. In addition, Wang et al. reported that the combination of an anti-PD-L1 antibody and doxorubicin enhances the antitumor response of the immune system by inhibiting doxorubicin-induced PD-L1 overexpression in osteosarcoma [
45].
Recently, gene signatures based on genome-wide expression profiles using genome-wide microarrays or high-throughput sequencing have attracted increasing attention and displayed great potential in predicting chemotherapy responsiveness in various tumors, including breast cancer [
46], lung adenocarcinoma [
47] and colorectal cancer [
17]. With the goal of developing an immune-related gene signature that will identify patients who might benefit from chemotherapy, we analysed gene expression in the TARGET database, where 218 differentially expressed genes were identified between good responders and poor responders. The KEGG analysis indicated that these genes were involved in the TGF-beta signalling pathway and cytokine-cytokine receptor interaction pathway. These results supported the hypothesis that the immune cell infiltration of tumors was associated with the chemotherapy response in patients with osteosarcoma. In addition, the top 10 KEGG categories, including the Hippo signalling and TGF-beta signalling pathways, were also involved in the resistance of various cancer cells to different chemotherapeutic drugs [
48,
49].
We deployed a specialized immunological database that may provide insights into the important role of IRGs in the response to chemotherapy to screen potential biomarkers and further explore the role of immune function in chemotherapy. Among these 218 differentially expressed genes, 24 genes overlapped with IRGs from the immunological database. A LASSO regression analysis was applied to retrieve the proportion of intersecting genes and optimize feature selection, which has been broadly used for characteristics of cancer therapy and diagnosis [
50], and to facilitate the construction of a practical predictive signature. Finally, a five-gene panel comprising TNFRSF9, CD70, EGFR, PDGFD and S100A6 was included in the predictive signature. Among these five IRGs, TNFRSF9, also known as CD137 or 4-1BB, is an inducible costimulatory receptor known to be expressed on the surface of activated T cells and natural killer (NK) cells. Preclinical results from various tumor models suggested that targeting 4-1BB with agonist antibodies leads to tumor clearance and durable antitumor immunity [
51]. In our study, TNFRSF9 was expressed at higher levels in good responders, which may partially explain why these patients are sensitive to chemotherapy. PDGFD is a proangiogenic factor that regulates many cellular processes, including cell proliferation, migration, invasion and angiogenesis [
52]. Several studies have reported that PDGFD overexpression is an independent predictor of chemotherapy resistance in ovarian cancer [
53] and colorectal carcinoma [
54], consistent with our results. In contrast to our expectations, CD70 and S100A6, which were expressed at high levels in good responders in our study, were identified as chemoresistance genes in several studies [
55,
56]. In those studies, increased CD70 and S100A6 expression was associated with resistance to chemotherapy and poor survival. Therefore, variability in the response to chemotherapy strongly depends on the genetic and epigenetic profiles of the tumor in these respects.
Subsequently, a chemotherapy predictive signature that consisted of five IRGs was successfully validated as a prediction model in independent GEO datasets. Man et al. and Ochi et al. reported similar results from studies developing a multigene predictive model to classify the chemotherapy response in a few patients with osteosarcoma [
57,
58]. Cao et al. and Xiao et al. constructed a multigene prognostic risk model for osteosarcoma. However, no overlapping genes were identified between these gene sets and the 5 IRGs presented in our study. These results indicated a wide range of heterogeneity related to the prognosis and predictive capacity of osteosarcoma. In addition, our study focused on immune cell infiltration and immune-related genes and used different training strategies and algorithms; thus, a comparison of the predictor gene lists among these studies was difficult. We would like to highlight that our study revealed a positive correlation between chemosensitivity and immune cell infiltration in good responders. Immunotherapy has become an established pillar of cancer treatment. The introduction of IRGs may potentially identify new molecular targets for cancer immunotherapy. Therefore, a worthwhile approach would be to evaluate the predictive capacity of immune-related markers as a basis for the identification of new molecular targets for cancer immunotherapy in patients with osteosarcoma.
However, our study had several limitations. First, our study did not include a sufficiently large sample size to comprehensively explore the relationship between IRGs and the chemotherapy response of patients with osteosarcoma. Second, our study was limited by the retrospective nature of the included datasets. Although the five IRG signature has been validated in independent datasets, the prediction of the response to chemotherapy in patients with osteosarcoma has been a more complex issue. Factors such as tumor heterogeneity, the sample size and use of different chemotherapeutic agents that may affect the chemotherapy response of patients with osteosarcoma impede the ability to develop and standardize predictive gene signatures. In this study, the PCA results suggested that the predictive signature seemed to have better predictability for distinguishing patients with a poor response in the training cohort. One explanation for the discrimination is that the sample size is relatively small. To some extent, the discriminatory power depends on the size of the sample. An insufficient sample size has a low probability of detecting a statistically significant difference between treatment groups. Another reason is that tumor heterogeneity may affect the immune-related signature to discriminate between patients with different chemotherapy responses. Therefore, the predictive signature requires further validation in a prospective study with a large sample size.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.