m⁶A regulator expression profile predicts the prognosis, benefit of adjuvant chemotherapy, and response to anti-PD-1 immunotherapy in patients with small-cell lung cancer

We downloaded the training cohort somatic mutations and expression profiles (the international cohort) from Cbioportal (https://​www.​cbioportal.​org/​study/​summary?​id=​sclc_​ucologne_​2015). During data processing, all RNA-seq data was subjected to log2 transformation. The mean expression values were selected when targeting genes that had more than one probe. We chose the GSE40275 database to explore the expression profile of m⁶A regulators, both in normal and LS-SCLC tissues. This dataset was downloaded from the Gene Expression Omnibus (GEO) dataset (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​) via the GPL15974 platform. Two validation cohorts—including the Shanghai cohort (GSE60052) downloaded from the GEO dataset and the National Cancer Centre (NCC) cohort, collected from the National Cancer Centre of China from January 2009 to November 2017—were collected. The NCC cohort included 150 LS-SCLC samples with formalin-fixed paraffin-embedded (FFPE) archived tissues were collected during surgeries. All enrolled patients from the NCC cohort were pathologically reconfirmed, had no other tumours, and carried clinically confirmed diagnoses of LS-SCLC. The relapse-free survival (RFS) and overall survival (OS) were defined as the day of surgical removal to recurrence, metastasis, or latest follow-up and the day of surgical removal to the date of death or latest follow-up. The study was approved by the Ethics Committee Board of our institute. The demographic and clinicopathologic parameters of the 150 LS-SCLC samples are displayed in Table 1.

We included 14 patients with SCLC who received sequential chemotherapy and anti-PD-1 treatment in our hospital from April 2019 to January 2021. Their baseline biopsy FFPE specimens before immunotherapy were collected. The RECIST V1.1 Criteria were used to evaluate the response to therapy.

The 4-μm FFPE slides were subjected to immunochemical staining. After deparaffinization and rehydration with high-concentration ethanol and pure water, the slides were incubated in 3% H₂O₂ or 15 min to block endogenous peroxidase activity. Then, the slides were subjected to heat antigen retrieval and non-specific site blocking using an EDTA buffer (PD 9.0) and 10% standard serum, respectively. Next, the slides were incubated overnight at 4 °C. The final counterstaining was performed using secondary antibodies (CD8, Abcam, ab17147, 1:100) and the 3, 3′-diaminobenzidine (DAB, Dako, Glostrup, Denmark) and haematoxylin.

The digital pathological system (HALO) was utilized to quantify the density of CD8+ T cells on the whole slides. We scan the slides images at high resolution (× 400) using the Pannoramic MIDI II slide scanner (3DHISTECH). The tumour regions were identified by a trained pathologist (LYX). The “Membrane IHC Quantification” module was selected for absolute counting of CD8+ T cells on the CaseViewer_2.3.

Only biopsies with at least 70% tumour cells were collected, and ~ 30-μm sections were cut from the FFPE samples. The Ambion RecoverAll Total Nucleic Acid Isolation Kit for FFPE (Thermo Fisher, Waltham, MA, USA) was used to isolate the FFPE tissue total RNA. The NanoDrop 2000C spectrophotometer (Thermo Scientific, Waltham, MA, USA) was used to quantify the extracted RNA. Meanwhile, the extracted RNA with an A260/A280 ratio between 1.8 and 2.2 were chosen for the quantitative RT-PCR (qRT-PCR) analysis. We used 200 ng RNA of a 20-μL reaction to reverse transcription through the FastKing Reverse Transcription Kit (Tiangen Biotech, Beijing, China). Next, we used 1 μL cDNA for PCR reaction with quantiNova PCR Kits (Qiagen, Dusseldorf, Germany) using 7900HT Fast Real-Time PCR system (Applied Biosystems, Carlsbad, USA; Indianapolis, IN). The qRT-PCR analysis was performed on all validation and independent cohort samples. The 2^−ΔΔCt method was used to calculate the expressions of interested m⁶A regulators. The details of the target m⁶A regulators primer sequences for qRT-PCR are shown in Additional file 1: Table S1.

Firstly, we screened out the m⁶A regulators with prognostic significance based on the optimal cut-off point in the international cohort. We used the least shrinkage and selection operator (LASSO) Cox model to determine the most prognostic m⁶A regulators; the minimum criteria were chosen during the analysis process. Lastly, the final m⁶A score equation was accomplished based on the expression of the five chosen m⁶A regulators and corresponding Cox regression coefficients.

R version 3.5.1 (https://​www.​r-project.​org) was used for all data analysis and image generation. The 30 m⁶A regulator protein-protein interaction analysis was conducted using the STRING interaction database and visualized using the Cytoscape software. The optimum cut-off survival analysis was completed using the “surv_cutpoint” function of the “survminer” R package. The Kaplan-Meier curve model was used to determine the prognostic value of the m⁶A regulator-based signature in the training and validation sets. The R package “survival” was employed to determine if the m⁶A score was an independent prognostic predictor for SCLC. The time-dependent receiver operating characteristic (ROC) curve analysis was created with the “survivalROC” R package. The Mann–Whitney U test was chosen to calculate the CD8+ T cell density between high- and low-score patients. A P-value less than 0.05 was considered statistically significant.

After reviewing the latest relevant literature [23‐26], we finally identified 30 m⁶A regulators, including 11 methyltransferases (METTL3, METTL14, METTL16, METTL5, WTAP, VIRMA, RBM15, RBM15B, ZC3H13, CBLL1, and ZCCHC4), 2 demethylases (ALKBH5 and FTO), and 17 binding proteins (YTHDF1, YTHDF2, YTHDF3, YTHDC1, YTHDC2, HNRNPA2B1, HNRNPC, FMR1, EIF3A, IGF2BP1, IGF2BP2, IGF2BP3, ELAVL1, G3BP1, G3BP2, PRRC2A, and RBMX) (Fig. 1A; Additional file 1: Table S2). Firstly, we explored the incidence of somatic mutations for m⁶A regulators in LS-SCLC. Mutations were present in 20 of 88 samples (22.73%; Fig. 1B). FMR1 displayed the highest mutation frequency, while approximately 14 m⁶A regulators exhibited no mutations within the LS-SCLC samples, including demethylases ALKBH5 and FTO (Fig. 1B). We identified co-occurrent mutations between METT3 and YTHDC2 and between HNRNPC and IGF2BP3 (Additional file 2: Fig. S1).

To determine if this genetic alteration affected the expression pattern of m⁶A regulators in LS-SCLC samples, we explored the mRNA expression of regulators between LS-SCLC and normal lung specimens. Principal component analysis revealed markedly different expression patterns of m⁶A regulators in LS-SCLC and normal samples (Fig. 1C). We also generated a heatmap for different expression profiles of these m⁶A regulators in LS-SCLC and normal tissues (Fig. 1D). The regulators’ expression details—between LS-SCLC and normal groups—are shown in Fig. 1E. Notably, almost all methyltransferases and binding proteins were upregulated in LS-SCLC; however, the two demethylases tended to exhibit lower expression in LS-SCLC than their normal counterparts. These results suggested that there may be abundant m⁶A modifications in LS-SCLC and significant heterogeneity in the m⁶A regulator genetic profile expression between LS-SCLC and normal lung tissues. Disordered m⁶A regulator expression may be involved in tumorigenesis and SCLC development.

Various m⁶A regulators cooperatively promote tumour development. Therefore, we also tried to explore the expression relationships for the 30 m⁶A regulators in SCLC. Notably, we detected remarkable correlations among the expressions of methyltransferases, demethylase, and binding proteins (Additional file 2: Fig. S2). Several significant positive correlations were also identified, including a correlation coefficient between KIAA1429 and YTHDF3 as high as 0.820 (Fig. 2A). Our protein-protein interaction network analysis determined that these 30 m⁶A regulators were in frequent communication (Fig. 2B). Importantly, the methyltransferases exhibited the most frequent interactions. Various methyltransferases formed a protein complex to perform biological functions. Therefore, the crosstalk among multiple m⁶A regulators may be actively involved in the SCLC development and progression.

Next, we investigated the clinical significance of m⁶A regulators in patients with LS-SCLC based on the optimal cut-off point derived from the international cohort. Most regulators were significantly associated with survival (Fig. 2C). Some regulators exhibited pro-carcinogenic properties, such as RBM15, RBM15B, ALKBH5, IGF2BP3, and PRRC2A. Some regulators functioned as tumour suppressors, including METTL5, YTHDC2, and G3BP1. Higher expression levels of these regulators often conveyed a favourable clinical prognosis. Given that most regulators exhibited clinical significance, we felt that an m⁶A regulator-based prognostic signature (m⁶A score) may be a useful molecular model for LS-SCLC. Therefore, using the LASSO Cox model, we included the above 22 regulators and determined five significant candidates—G3BP1, METTL5, ALKBH5, IGF2BP3, and RBM15B—for the subsequent m⁶A score creation (Fig. 2D, E).

Based on the expression levels of these five regulators and corresponding coefficients, we constructed the m⁶A score system for patients with LS-SCLC. The formula is as follows: m⁶A score = (0.0906 × G3BP1 expression) + (0.4096 × METTL5 expression) − (0.6365 × ALKBH5 expression) − (0.0912 × IGF2BP3 expression) − (0.0660 × RBM15B expression). We calculated the m⁶A scores for each patient and classified them into high- (m⁶A score ≥ − 1.271) and low-score (m⁶A score < − 1.271) groups according to the optimal cut-off point (Fig. 3A). The principal component analysis revealed high heterogeneity between the high- and low-score groups (Fig. 3B). The Kaplan-Meier survival curve results indicated that high-score patients suffered significantly worse OS (Fig. 3C, hazard ratios (HR) 5.19, 95% confidence interval (CI) 2.75–9.77, P < 0.001). To evaluate the prognostic performance of the m⁶A score, a time-dependent ROC analysis was conducted. The m⁶A score achieved area under the curve (AUC) values of 0.672, 0.812, and 0.793 for predicting OS within the international cohort at 1, 3, and 5 years, respectively (Fig. 3D). Further ROC analysis revealed that the m⁶A score performed better than clinicopathological characteristics for predicting OS (Fig. 3E). The C-index of the m⁶A score reached 0.881. This indicated that the prognostic accuracy of the m⁶A score was also higher than other clinicopathological parameters (Fig. 3F).

To further assess the reliability and robustness of the classifier, we incorporated another two independent cohorts of 197 samples for validation, including the Shanghai cohort (N = 47) and the NCC cohort (N = 150). The cohorts’ clinicopathologic features are presented in Table 1. Using the same formula, the two cohorts were divided into low- and high-score groups. Firstly, we tested the m⁶A score in the Shanghai cohort. As expected, the signature showed excellent repeatability and stability during validation (Fig. 4A). The high-score patients in the Shanghai cohort suffered shorter OS than low-score patients (Fig. 4B, P = 0.006). The AUCs were 0.652, 0.733, and 0.731 for predicting 1-, 3-, and 5-year OS in the Shanghai cohort, respectively (Fig. 4C). In the Shanghai cohort, both the m⁶A score (C-index = 0.862) and SCLC staging (C-index = 0.880) accurately predicted OS for patients with LS-SCLC (Fig. 4D).

The clinical applicability of the m⁶A score was further evaluated in the FFPE specimens from the NCC cohort. Here, the low-score patients tended to significant better clinical outcomes in terms of OS (Fig. 4E, F, P < 0.001), and the m⁶A score achieved AUCs of 0.794, 0.691, 0.686 at 1-, 3-, and 5-year OS, respectively (Fig. 4G). Also, the C-index of the m⁶A score for OS was up to 0769 and higher than other factors in the NCC cohort (Fig. 4H).

We evaluated the predictive performance of the m⁶A score for RFS in patients with SCLC (Fig. 4I). The high m⁶A score was also predictive of poorer RFS in the NCC cohort (Fig. 4J, P < 0.001). The AUCs of m⁶A score for 1-, 3-, and 5-year RFS predictions were 0.713, 0.682, and 0.695, respectively, and the C-index was 0.748 in the NCC cohort (Fig. 4K, L). Thus, the m⁶A score was superior to the TNM system and sufficiently reliable to predict prognosis in patients with SCLC—both for OS and RFS.

We additionally explored the prognostic significance of the m⁶A score in relationship to various clinicopathological features—including sex, age, and smoking history. Because the sample size of the Shanghai cohort was small, we only performed clinical subgroup analyses on the international and NCC cohorts. As illustrated in Additional file 2: Fig. S3-S4, in the international cohort, low-score cases achieved longer OS and RFS across all clinical subtypes, including male, female, smoker, older (age ≥ 60), and younger (age < 60) (P < 0.05). The same results were obtained during the NCC cohort validation (Additional file 2: Fig. S3-S4, P < 0.05).

To confirm whether the m⁶A score is an independent predictor of prognosis in SCLC, we carried out univariate and multivariate Cox regression analyses on three independent cohorts. Sex and the m⁶A score were significantly related to OS in the international cohort; staging and the m⁶A score were also correlated with the prognosis in the Shanghai cohort. Age and the m⁶A score were significantly associated with survival in the NCC cohort (Fig. 5A, P < 0.05). Moreover, after integrating these clinical parameters into the multivariate Cox regression analyses, the m⁶A score was the only stable, independent prognostic indicator for patients with SCLC across all three cohorts (Fig. 5B, P < 0.05). Additionally, after multivariable adjustment by clinicopathological features, the m⁶A score remained a significant independent prognostic factor for RFS in the NCC cohort (Additional file 1: Table S3).

Considering the decisive role of m⁶A regulators in chemotherapy resistance, we further explored whether the m⁶A score could predict ACT treatment benefit. In the international and NCC cohorts, 42 and 129 cases underwent ACT, respectively. The m⁶A score divided 23 and 19 of 42 patients into high- and low-score groups in the international cohort (Fig. 6A), respectively, and divided 75 cases into the high-score group and 54 cases into the low-score group NCC cohort (Fig. 6D). Those low-score patients benefited considerably from ACT and achieved much longer OS than the high score cases in either cohort (Fig. 6A, D, both P < 0.001). Additionally, ROC curves showed that the AUCs of m⁶A score for predicting ACT OS benefit were 0.768, 0.901, and 0.82, and 0.807, 0.68, and 0.67 in the international cohort and NCC cohort for 1-, 3-, and 5-year, respectively (Fig. 6B, E). Meanwhile, the C-index of the m⁶A score for OS was also higher than other clinicopathological characteristics and as high as 0.956 and 0.750 in the two cohorts, respectively (Fig. 6C, F). In the NCC cohort, high-score cases suffered shorter RFS than the low-score ones (Fig. 6F, P < 0.001). The m⁶A score also achieved a reliable predictive ability to stratify different RFS statuses for patients with ACT. For AUCs of 0.708, 0.683, 0.66 at 1-, 3-, and 5-year RFS, the corresponding C-index was up to 0.734 (Fig. 6G, H). Collectively, the m⁶A score was able to identify those patients with SCLC most likely to benefit from ACT.

Previous studies have demonstrated that m⁶A regulators relate to anti-tumour immune effects and tumour immune microenvironment (TIME) characterizations [27]. Since the m⁶A score is based on various m⁶A regulators, we decided to probe the relationship between the m⁶A signature and TIME features. Considering the centrality of CD8+ T cells in TIME, we explored the relationship between CD8+ T cell infiltration and the signature in SCLC. Using strict quality controls, we finally incorporated 117 FFPE samples in the NCC cohort. The density of CD8+ T cells in the tumour regions of SCLC was detected and calculated using the HALO digital pathological platform. Each patient’s m⁶A score was also matched. Representative pictures of CD8+ T cell distribution from high- and low-score groups are displayed in Fig. 7A. Low-score patients tended to have more CD8+ T cell infiltration than high-score patients (Fig. 7B). In addition, the m⁶A score was negatively correlated with CD8+ T cell density in SCLCs (Fig. 7C, R = − 0.34, P < 0.001).

We collected the pre-treatment samples from 14 patients with SCLC who received anti-PD-1 treatment to investigate the relationship between the m⁶A score and responses to immunotherapy. The overall response rate was 35.71%. Interestingly, patients with low scores seemed to benefit more from immunotherapy, while those with high scores tended to be resistant to immunotherapy (Fig. 7D, E). Meanwhile, ROC analyses indicated that the m⁶A score could predict non-responders with an AUC of 0.8. The m⁶A score showed superior performance than age or sex for identifying non-responders to immunotherapy (Fig. 7F).