Background
With the discovery of immune checkpoint molecules, immunotherapy becomes more promising strategy for cancer patients to elicit clinical responses durably [
1]. It has been approved for the treatment of lung adenocarcinoma (LUAD), skin cutaneous melanoma (SKCM), and head and neck squamous cell carcinoma (HNSC). These tumors have high PD-L1 and CD8A expression levels [
2]. Nevertheless, immunotherapy only benefits a minor subset of patients for long-term survival [
3]. Identification of potentially therapeutic indexes linked to tumor prognosis and immunotherapy responses will remarkedly contributed to precision medicine.
Tumor microenvironment (TME) consists of immune and non-immune stromal components, both of which were reported to be closely associated with oncogenesis and malignant behaviors of tumors [
4]. The existing evidences demonstrate that abundant immune components in TME are positively associated with immunotherapy responses [
5]. To date, a range of algorithms have been developed to estimate the immune and stromal cell infiltration including CIBERSORT, TIMER, ESTIMATE and MCPcounter [
6‐
9]. These tools perform well in the estimation of TME cells, but not in the prediction of tumor prognosis and immunotherapy responses. Although ESTIMATE provides the immune and stromal infiltration scores [
8], LUAD, SKCM and HNSC were not included in the training datasets. There is still a lack of single index to reflect both immune and stromal activation signals associated with prognosis and immunotherapy responses.
Here we developed a novel immune and stromal scoring system named ISTMEscore, which followed a unique design: (1) Isolate TME signals associated with prognosis from bulk gene expression data; (2) Extract specific gene signatures from the above TME signals; (3) Calculate ISTME scores with single-sample gene set enrichment analysis (ssGSEA) algorithm [
10]. In addition, we collected 15 datasets with 2965 patients to train and validate our ISTMEscore system, and depicted the landscapes of immune and stromal cell infiltration, transcriptome, genome, prognosis and immunotherapy responses in patients with different ISTME scores. Finally, we compared ISTMEscore with previous TME indexes on prediction of TME status and cancer prognosis.
Discussion
In this study, we built the novel ISTMEscore system with unique workflow, and depicted the multi-dimensional landscape of different TME subtypes. Our TME subtypes could represent TME patterns, and were associated with clinical features and immunotherapy responses in LUAD, SKCM and HNSC. Additional analysis suggested that high collagen, matrix metalloproteinases, glycolysis and acid environment, VEGF signaling were the integral parts of stromal activation. Whether the stromal signals were involved in immune exhaustion remained to be further studied.
Although immunotherapy benefited LUAD, SKCM and HNSC patients, only a small proportion of patients had long-term survival [
5]. Identification of potentially sensitive population for ICIs helped to decrease medical expenses and improve quality of life. Our studies found that the LH type with malignant TME and the LL type with desert-like TME had low ICI responses (Fig.
7). Interestingly, the TME subtypes determined by transcriptome before treatment did not demonstrate significant association with immunotherapy response. In the 3 immunotherapy cohorts with biopsies taken before treatment, TME subtype was not significantly associated with ICI response: HH vs. HL vs. LH vs. LL: 100% vs. 94.7% vs. 94.7% vs. 78.6%, Fisher's test P = 0.2 in GSE93157 (Fig.
7I) [
25]; and HH vs. HL vs. LH vs. LL: 100% vs. 33.3% vs. 33.3% vs. 0%, Fisher’s test P = 0.26 in GSE67501 (Fig.
7J) [
26]; and HH vs. HL vs. LH vs. LL: 50% vs. 52.9% vs. 13.3% vs. 40%, Fisher’s test P = 0.09 in GSE35640 (Fig.
7K) [
27]. In GSE93157, there was a difference of progression-free survival only in subgroup comparisons of HL vs. LH (logrank-P = 0.056, Fig.
7L). Chen et al
. also found positive implications for dynamic monitoring of the immune microenvironment [
5]. The protein levels of CD3, CD4, CD8, PD-1, PD-L1 and LAG3 during the treatment could reflect the responses (all, P < 0.01) better than those before treatment. Moreover, Riaz et al
. found that patients with TME of immune activation or “hot tumor” were associated with high CR/PR rates in the group receiving ICIs in advance, whereas TME immune infiltration was not linked to ICI responses in the patients without prior immunotherapy [
24]. In stromal environment, the desmoplastic stroma was the physical barrier for tumor to resist immunotherapy and chemotherapy [
51]. Kim et al
. reported the single-cell sequencing analysis of the longitudinal samples from 20 triple-negative breast cancer patients during neoadjuvant chemotherapy [
52]. They found that degradation of ECM and angiogenesis signals were upregulated in the chemoresistant tumors. In preclinical studies, cancer-associated fibroblasts were reported to compensate immunotherapy through crosstalk with myeloid-derived suppressor cells and CD8+ T cells [
53,
54]. In this study, we found that the majority of patients, who switched to the LH and LL types during ICI therapy, were non-responders. Accordingly, longitudinal detection of TME-related indicators might be a better choice for patients with ICI treatment, which required studies with large sample size.
The HL patients had high immune activation, inhibitory angiogenesis and long OS, which was considered as balanced states of TME. As previous studies, there were synergistic effects between immune normalization and vascular normalization [
55]. VEGF induced immunosuppressive cells such as myeloid suppressive cells, tumor-related macrophages and Tregs, which developed immune exhaustion. On the other hand, the infiltration and activation of intratumoral effector T cells promoted the remodeling and normalization of vessels. The immune and vascular normalization might explain the results of IMpower150 study, in which the first-line immunotherapy combined with anti-angiogenic drugs benefited non-squamous NSCLC patients [
56]. The immune and vascular normalization seemed to correspond to the HL type in our study.
Compared with the existing methods (ESTIMATE and MCPcounter) [
8,
9], our scores showed advantages on the prediction of prognosis and immunotherapy response. However, the improvement in predictive performance was not large. B lineage and myeloid dendritic cell infiltration scores of MCPcounter were also good predictors of prognosis (Additional file
16: Table S8). For prediction of ICI responses, monocytic lineage infiltration scores also had excellent performance.
The effects of TMB on prognosis were controversial. Some clinical studies revealed opposite prognostic effects of TMB in NSCLC patients without immunotherapy. In LACE-Bio-II study with 908 NSCLC patients [
57], high TMB group (≥ 8 m/Mb) showed better OS, while the low TMB group (< 4 m/Mb) had worse prognosis (P = 0.016). However, another clinical study indicated that higher TMB (≥ 62 m/Mb) was correlated with worse OS in the 90 NSCLC patients (P = 0.0003) [
58]. TMB may not be a very robust prognostic marker due to lack of consideration on the threshold, complex effects of mutation, as well as the gene mutations from immune or stromal cells.
There were some limitations in this study: All datasets were retrospective, requiring prospective clinical trials to further verify. Whether our scores worked in other tumors, especially tumors lacking immune cell infiltration, remained to be further investigated. In addition, the link between our ISMEscore system and ICI response was not strong. Our system needed to be applied cautiously for the prediction of immunotherapeutic efficacy.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.