A CpG Methylation Classifier to Predict Relapse in Adults with T-Cell Lymphoblastic Lymphoma Free

T-cell lymphoblastic lymphoma (T-LBL) is a rare non-Hodgkin lymphoma and generally occurs in adolescents and young adults (1). Although classified as T-cell acute lymphoblastic leukemia (T-ALL) in the current WHO classification, T-LBL varies significantly in genetic profile, clinical presentations, subgroup stratification, and prognosis (2, 3). Unlike pediatric T-LBL, adults with T-LBL have a lower tolerance of intensive chemotherapy regimens, and, therefore, have lower complete remission (CR) rates and shorter survival (4). Intensive ALL-like regimens, such as the Berlin–Frankfurt–Muenster (BFM) protocol and the fractionated cyclophosphamide, vincristine, doxorubicin, and dexamethasone (hyper-CVAD) regimen, yield a CR rate of approximately 90% for T-LBL adults; however, relapse remains high after CR (40%–60%; refs. 5–7). Therefore, accurate assessment of high relapse risk after CR is important in providing useful insight into which subset of patients may require and benefit from more intensive chemotherapy and/or sequential hematopoietic stem cell transplantation (HSCT).

To date, only a few prognostic factors have been recognized in adult T-LBL (8–10). The minimal residual disease (MRD) and ¹⁸F-fluorodeoxyglucose PET scan also could not predict risk of relapse in patients with T-LBL treated with intensive therapy (11). Molecular factors such as NOTCH1/FBXW7 (N/F) mutations, N/K-Ras mutations, PTEN alterations, DNMT3A mutations, and T-cell receptor (TCR) gene rearrangements were also explored for predicting prognosis (11, 12). However, these risk factors have only been studied in small populations, and were not validated in independent centers. A formalin-fixed, paraffin-embedded (FFPE)-based miRNA signature has been developed to increase the predictive value of relapse of adult T-LBL (8). However, the miRNA-based signature would be influenced by tissue-specific dissimilarity. Moreover, the miRNA-based formulas and threshold values were detected by real-time PCR, which are not applied for other types of measurement platform data, and, therefore, have no universal applicability.

DNA methylation markers have rapidly emerged as predictors of cancer prognosis (13). However, few methylation markers have become established in T-LBL. High-throughput technologies allow us to identify reliable methylation markers. In this study, we delineated global genome-wide CpG methylation in patients with T-LBL and sought to identify and validate a CpG-based classifier for predicting relapse. A nomogram model was also established incorporating the CpG-based classifier, clinical variables, and genetic factors.

In this multicenter study, we collected 1,015 FFPE samples from patients with T-LBL who received treatment between January 1, 2003, and December 30, 2015, at Sun Yat-sen University Cancer Center (SYSUCC) and Guangdong General Hospital (GDGH), both in Guangzhou, China, and between January 1, 2006 and December 30, 2015, at 25 other medical centers across China. Total 549 were eligible for the final analysis (Supplementary Fig. S1), including 228 samples from SYSUCC and GDGH. We assigned 160 patients to the training cohort and 68 patients to the internal testing cohort using computer-generated random numbers. Furthermore, 321 samples from the 25 medical centers were assigned to the independent validation cohort. To generate CpG methylation profiles, we obtained 49 more FFPE samples from 21 relapsed and 28 relapse-free patients with T-LBL who received treatment between January 1, 2015 and January 1, 2016, at SYSUCC or GDGH (Supplementary Table S1).

All FFPE tissue samples were obtained by incisional lymph node biopsy and/or needle biopsy, and reassessed by three experienced pathologists independently (DX, LRH, and ML). The study was conducted in accordance with declaration of Helsinki. The study was approved by the institutional review board of each participating institution. All patients provided written informed content where appropriate.

Adults 18 to 65 years of age with complete clinical data and follow-up information, diagnosed with T-LBL, were included in this study. For the classification of T-LBL versus T-ALL, two experienced physicians (XPT and QQC) separately reassessed the diagnosis of T-LBL by the 2008 WHO criteria (Supplementary Fig. S1; ref. 14). We excluded patients with T-LBL with more than 20% lymphoblasts in the bone marrow. We also excluded patients who did not achieve CR, or from whom samples were not taken at initial diagnosis, or contained less than 80% tumor cells, or RNA content less than 5 ng/L, or DNA content less than 1 μg. We excluded patients who were not treated with hyper-CVAD protocol or BFM protocol (15, 16). Bone marrow aspiration and standard blood biochemistries, and thoraco-abdominal-pelvic CT or PET scan were performed at baseline. We identified CR and relapse by CT, PET-CT, and/or bone marrow biopsy according to Cheson criteria (17). Plasma LDH levels >245 U/L were considered elevated. Performance status was assessed using Eastern Cooperative Oncology Group performance status (ECOG-PS) and clinical stage was defined using the Ann Arbor stage system. We assessed NOTCH1/FBXW7 mutations as described previously (8). We further excluded patients who did not receive central nervous system (CNS) prophylaxis with intrathecal chemotherapy (e.g., methotrexate, cytarabine, and corticosteroids), and high-dose systemic chemotherapy (e.g., methotrexate and cytarabine). The clinicopathologic characteristics (except Ann Arbor Stage) did not differ significantly between patients receiving HSCT and those without HSCT in the entire cohort (Supplementary Table S2). The clinicopathologic characteristics of patients with T-LBL who were treated with BFM protocol and those who received hyper-CVAD did not differ significantly in the entire cohort (Supplementary Table S3).

By Dobbin's method (18), the optimal sample size of microarray assay was first estimated to be 49, including 28 relapse-free and 21 relapse samples using the criterion of 1.5 as the estimated standardized fold change, and 0.10 as the tolerance. Methylation 850K Beadchip (Illumina) was used to screen significant CpGs. After primary filtration, the least absolute shrinkage and selector operation (LASSO) algorithm (19) and support vector machine–recursive feature elimination (SVM-RFE; ref. 20) were used for relapse-related CpG selection (Supplementary Table S4). In the training phase, relapse-related CpGs were further tested by pyrosequencing in 160 FFPE T-LBL samples from the training cohort. Relapse-related CpGs from both algorithms were entered to construct a CpG-based classifier for predicting relapse-free survival (RFS) by LASSO Cox regression analysis in the training cohort (21, 22). Four CpGs are selected by LASSO Cox regression analysis according to 1-SE criteria (optimal values). X-tile software (version 3.6.1; Yale University) was used to determine the optimal cutoff in the training cohort (23). In the validation phase, to estimate the reproducibility and validity of the classifier, we evaluated the CpG-based classifier in the internal testing cohort and the independent validation cohort. For reproducibility of the Illumina 850K Beadchip assay, we compared methylation levels by methylation array and pyrosequencing using the same group of 49 samples (Supplementary Fig. S2). For wide clinical application of the CpG classifier, we compared the methylation levels of single CpGs from the CpG-based classifier between FFPE tissues and fresh-frozen as well as fresh effusion samples by pyrosequencing (Supplementary Fig. S3). To examine the effect of tumor heterogeneity on prediction accuracy of the classifier, we compared the methylation levels of the four CpGs between different areas of the same tissue or different FFPE tissues from the same patient (Supplementary Fig. S3). Subgroup stratification analysis based on LDH level, CNS involvement, ECOG-PS, and NOTCH1/FBXW7 (N/F) status was also used to examine the performance of the classifier.

Genomic DNA was extracted from 49 FFPE samples using QIAamp DNA Mini Kit (Qiagen), following the manufacturer's manual and assayed with Infinium Human Methylation 850K Beadchip according to Illumina's protocol. To maximize the extraction of FFPE information, all DNA samples were restored and passed quality control test by Infinium FFPE DNA Restoration and QC Kit (Illumina; Supplementary Fig. S4). The signal value was processed using “RnBeads” package in R software (24) and normalized by Beta-Mixture Quantile dilation (BMIQ) method (Supplementary Figs. S4 and S5; ref. 25). The β values, from 0 (unmethylated) to 1 (totally methylated) were used to quantify methylation levels. After exclusion steps shown in Supplementary Fig. S6, a threshold value of 0.15 for the average β and the associated P < 0.05 were considered significant. The methylation microarray data are available at the Gene Expression Omnibus (GSE131677).

CpG methylation levels in all three cohorts were determined by pyrosequencing as described previously (26). Briefly, genomic DNA was extracted with QIAamp DNA Mini Kit (Qiagen) and converted by EpiTect Bisulfite Kit (Qiagen). PCR was run using primers from PyroMark PCR Kit (Qiagen; Supplementary Table S5). cg00207389 primer was not designed and therefore excluded due to the density of CG sites. The CpG methylation value was calculated as [mC/(mC+C)]% using PyroMark Q24 instrument (Qiagen).

Univariate analysis was used to analyze correlation between variables and RFS/OS and variables with P < 0.1 were entered into multivariate analysis. Covariates included age (≤45 years vs. >45 years), sex, ECOG-PS (≤1 vs. ≥2), pleural and/or pericardial effusion, CNS involvement, mediastinal involvement, bone marrow involvement, LDH level (normal vs. elevated), Ann Arbor Stage (≤2 vs. >2), NOTCH1/FBXW7 status (mutated vs. wild-type), and four-CpG classifier (low vs. high). Log-rank test was used to compare survival curves. The package of “glmnet” and “e1071” were used to perform LASSO logistic regression and SVM-RFE algorithm, respectively, with R version 3.5.0 software. Survival was analyzed using the “survival” package in R software. RFS and overall survival (OS) were calculated from the date of diagnosis to the date of confirmed tumor relapse and death, respectively. Survival was censored at the date of death of other causes, or the date of the last follow-up visit. The accuracy of CpG-based classifier was assessed by time-dependent ROC curve analysis using the “survival ROC” package in R software. The performance of the nomogram was explored by calibration plots, using the “rms” package in R software. Statistical significance was set at P value less than 0.05.

The study flowchart is shown in Fig. 1 and the characteristics of patients with T-LBL in the training, internal testing, and independent validation cohorts are listed in Table 1. The median duration of follow-up was 42.0 (IQR 21.1–57.9) months in the training cohort, 42.5 (IQR 21.6–57.8) months in the internal testing cohort, and 37.7 (IQR 16.9–56.5) months in the independent validation cohort, respectively. Relapse occurred in 44.8% (246/549) patients, and the 1- and 3-year RFS were 76.9% and 54.7%, respectively. Furthermore, 40.3% (221/549) patients died and the 1- and 3-year OS were 85.8% and 60.4%, respectively (Supplementary Table S6).

Our microarrays assay identified 419 differential CpGs; 28 and 29 relapse-related CpGs were selected by LASSO logistic regression and SVM-RFE algorithm, respectively, and 13 relapse-related CpGs were selected simultaneously by both algorithms (Fig. 2A–C; Supplementary Table S4). To evaluate the reproducibility of the Illumina 850K Beadchip, we used pyrosequencing method to compare the methylation values in the same group of 49 samples. Consistent with other studies, the methylation value of three randomly selected CpGs between methylation microarray and pyrosequencing array exhibited good concordance (Supplementary Fig. S2). Hierarchical clustering of the 44 candidate CpG successfully separated the 49 samples into relapse and relapse-free clusters (except 2 relapse patients assigned to relapse-free clusters; Fig. 2D).

Forty-three candidate CpGs in the training cohort were quantified by pyrosequencing (Supplementary Table S7). Using LASSO Cox regression models, we built a four-CpG–based classifier (cg12373951, cg10323688, cg14569644, and cg01589972), corresponding to SRP receptor subunit beta (SRPRB), CUGBP Elav-like family member 2 (CELF2), homeobox C5 (HOXC5), and proline rich 36 (PRR36; Fig. 2E), for predicting RFS according to the weighted coefficients in the training cohort (Supplementary Fig. S7). The risk score was calculated as follows:

Risk score = (0.812 × methylation level of SRPRB) + (0.253 × methylation level of CELF2) + (1.044 × methylation level of HOXC5) − (0.163 × methylation level of PRR36).

The optimal cutoff value was determined to be 1.05 by X-tile plots (Supplementary Figs. S8 and S9) and stratified the training cohort into the high-risk group (64 patients, 40%) and low-risk group (96 patients, 60%). The high-risk group had significantly shorter RFS (HR = 4.869; 95% CI, 2.914–8.135; P < 0.001) and OS (HR = 4.662; 95% CI, 2.629–8.267; P < 0.001) in the training cohort (Fig. 3). Time-dependent ROC analysis showed that the CpG-based classifier exhibited good predictive accuracy (Fig. 3). The AUC of the classifier at 3 years was 0.820 (95% CI, 0.752–0.888).

The classifier was validated in the internal testing cohort and the independent validation cohort. At a cutoff of 1.05, the high-risk group had significantly shorter RFS in both the internal testing cohort (HR = 3.621; 95% CI, 1.644–7.793; P < 0.001) and the independent validation cohort (HR = 3.765; 95% CI, 2.679–5.293; P < 0.001) versus the low-risk group (Fig. 3). Moreover, the classifier achieved similar and stable predictive accuracy in the internal testing cohort (AUC at 3 years 0.817; 95% CI, 0.716–0.918) and the independent validation cohort (0.753; 95% CI, 0.698–0.808). The classifier also showed good predictive value in OS in the three cohorts (HRs ranging from 3.034 to 4.662; AUC at 3 years ranging from 0.742 to 0.801, Fig. 3). Patient characteristics (except ECOG-PS in the independent validation cohort) did not differ between the high-risk and low-risk group (Table 1). The stratification analysis showed that our CpG-based classifier was still a statistically significant predictive model when stratified by clinicopathological risk factors (except for CNS involvement) and genetic risk factors (Supplementary Figs. S10 and S11).

Significant variables from univariate analysis (Supplementary Figs. S12 and S13) were entered into multivariate analysis, which revealed that LDH level, ECOG-PS, CNS involvement, N/F status, and the four-CpG classifier were independent prognostic factors of RFS and OS. After adjustment, the four-CpG–based classifier remained a significant and independent prognostic factor of RFS and OS in the combined training and testing cohort and independent validation cohort (Supplementary Tables S8 and S9). Moreover, our CpG-based classifier showed higher predictive accuracy than any clinicopathological/genetic risk factors or single CpG alone (log-rank test, all P < 0.05; Supplementary Fig. S14).

To estimate the reproducibility and applicability of CpG-based classifier, we assayed methylation value of the four CpG sites in different types of tissues (FFPE tissues, fresh-frozen tissues, and fresh effusion tissues). The four-CpG methylation value of fresh-frozen tissues and fresh effusion tissues exhibited good consistency with that of FFPE tissues (all P < 0.001, all R > 0.94, Pearson correlation analysis; Supplementary Figs. S3A and S3B). To determine whether tumor heterogeneity affected the risk score, we compared the methylation level of the four candidate CpGs between different areas of the same FFPE samples and different FFPE tissues from the same patient. Tumor heterogeneity did not affect the reliability of our classifier according to Pearson correlation analysis (all P < 0.001, all R > 0.95; Supplementary Figs. S3C and S3D). Furthermore, we assayed methylation values of the four CpG sites and their corresponding mRNA levels of purported gene targets in fresh-frozen tissues, confirming that the methylation of those four CpGs could reflect the mRNA expression levels of the purported gene targets (all P < 0.001, Pearson correlation analysis; Supplementary Fig. S15)

We generated a nomogram to predict 3-year RFS based on significant predictors of RFS via multivariate analysis of the training and internal testing cohort (the developing set). The predictors included the CpG-based classifier, LDH level, CNS involvement, ECOG-PS, and N/F status (Fig. 4A). The calibration plots showed an acceptable agreement in the two cohorts between nomogram prediction and actual observation (Supplementary Fig. S16). The AUC was 0.839 in the developing set and 0.776 in the independent validation cohort at 3 years, and both AUCs showed significantly higher predictive accuracy than the CpG-based classifier, clinicopathologic risk factor, or genetic risk factor alone (all P < 0.05, Fig. 4B; Supplementary Fig. S17).

The developing set and the independent validation cohort were categorized into four subgroups using total scores according to maximization of prognostic difference (score: 0–38.5, 38.6–138.4, 138.5–190.7, and ≥190.8); each subgroup represented a distinct prognosis (Fig. 5A and B; Supplementary Fig. S18). Four therapeutic schedules (BFM, BFM followed by HSCT, hyper-CVAD, and hyper-CVAD followed by HSCT) yielded comparable RFS or OS of 549 patients (Fig. 5C). Meanwhile, patients with a nomogram score ≥138.5 receiving HSCT had longer RFS and OS versus those not receiving HSCT after intensive chemotherapy (Fig. 5D and E; Supplementary Figs. S19–S21). Patients receiving BFM followed by HSCT had significantly longer OS (HR = 2.017; 95% CI, 1.022–3.981; P = 0.043) but comparable RFS (HR = 1.589; 95% CI, 0.800–3.158; P = 0.186) versus those receiving hyper-CVAD followed by HSCT (Fig. 5E). These results indicated that the nomogram successfully stratified risk of patients and could help clinicians identify patients who benefited from a certain therapeutic schedule.

In this study, we successfully built a four-CpG–based classifier in the training cohort via methylation microarrays and further validated the classifier in the internal testing cohort and the independent validation cohort. The classifier was of good predictive value of RFS and OS for adults with T-LBL, and performed better than conventional clinical and genetic risk factors. We also built a nomogram based on the classifier and other clinical and genetic predictors of relapse, which has good clinical applications.

Until now, there has been no study on the relation between genome-wide CpG profiles and prognosis of T-LBL. Only scarce reports showed aberrant gene methylations in T-LBL tissues (12). SOCS3 hypermethylation enhances JAK–STAT signaling and contributes to T-LBL development (27). In this study, we used microarray analysis to assess the predictive value of CpGs for patients with T-LBL. To avoid losing important markers, we used two distinct algorithms to select 44 relapse-related CpGs in the discovery phase. Four CpGs most intimately associated with RFS were finally identified by LASSO Cox regression in the training phase. All four CpGs in our classifier were identified by LASSO logistic regression, and SRPRB and CELF2 were not identified using SVM-RFE, which may be due to the different mathematical principles of the two methods. LASSO regression shrinks the coefficient estimates towards zero, whereas SVM-RFE uses the maximum internal principle in SVM to eliminate the feature recursively. The predictive value of the CpG-based classifier was further validated in the internal testing cohort and the independent validation cohort. Moreover, the predictive value of the classifier was compared with single CpGs, confirming the effectiveness of the combined strategy.

The four CpG markers in our classifier showed novel association with T-LBL. According to Methylation 850K annotation, cg12373951 and cg14569644 were located at 5′UTR region, cg10323688 and cg 01589972 were located at TSS (transcription start site)1500 region. All four CpGs were associated with CIMP (CpG Island Methylator Phenotype) status. Upregulated SRPRB increases cell apoptosis rate and decreases NF-κB expression and phosphorylation in pancreatic ductal adenocarcinoma (28). CELF2 expression is increased in response to T-cell signaling through the combined regulation of transcription and mRNA stability (29). CELF2 controls lymphoid enhancer-binding factor 1 (LEF1) alternative splicing and aberrant expression of LEF1 is implicated in tumorigenesis and cancer cell proliferation, migration, and invasion (30, 31). HOXC5 is involved in lymphomagenesis and is a key upstream regulator of Wnt/BMP signaling (32, 33). HOXC5 upregulation can be detected in cells representing mature stages of lymphoid cell differentiation, but not in leukemic cells (with activated hTERT expression) representing immature stages of lymphoid differentiation (34). The function of PRR36 remains unknown. Further studies on SRPRB, CELF2, HOXC5, and PRR36 will provide new insights into T-LBL. The regulation of the protein metabolic process, and the cellular differentiation and development process involving PI3K-AKT, HIF-1, insulin, and RNA transport signaling pathways should be elucidated to understand the mechanism of T-LBL.

In addition, we built a nomogram including LDH level, CNS involvement, EOCG-PS, N/F status, and the CpG-based classifier to predict individual relapse risks. The performance of the nomogram was verified in the independent validation cohorts. To the best of our knowledge, this is the first nomogram for predicting survival of patients with T-LBL that is based on a large database. Both physicians and patients could perform an individualized survival prediction through this easy-to-use scoring system.

To date, the standard chemotherapy protocol for adult T-LBL has not been established. The BFM protocol and hyper-CVAD regimens are most commonly used in adult T-LBL. It is still unknown which chemotherapeutic protocol is superior. Moreover, HSCT is often used to consolidate CR1 after induction therapy, but which subset of patients that could benefit from which type of chemotherapy followed by HSCT remains unclear (35). In this study, using the nomogram, we showed that BFM followed by HSCT provided a survival benefit to patients with a nomogram score ≥138.5. This tool could allow for better identification of patients who benefit from a therapeutic schedule.

The limitations of our study should be noted. All the specimens were obtained from patients in China. Further studies are required to determine if our CpG-based classifier can be applied to other ethnic groups. The biological mechanism of the candidate markers such as SRPRB, CELF2, HOXC5, and PRR36 are still unknown; such studies may suggest treatment strategies. Although our retrospective study may be subject to biases, due to the scarcity of the disease and inclusion of a large-scale of adults with T-LBL in this study, our retrospective study still has a strong guiding value in treatment of adults with T-LBL.

In conclusion, our CpG-based classifier is a useful predictive biomarker for predicting relapse of T-LBL in adult patients. A nomogram incorporating the CpG-based classifier may help predict individual relapse risk and help clinicians manage patients with T-LBL.

No potential conflicts of interest were disclosed.

The key raw data have been uploaded onto the Research Data Deposit public platform (RDD), with the approval RDD number of (RDDA2019001073). The microarray data have been deposited online under accession number (GSE131677).

Conception and design: Q.-Q. Cai, X.-P. Tian, D. Xie

Development of methodology: Q.-Q. Cai, X.-P. Tian, D. Xie

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Q.-Q. Cai, X.-P. Tian, N. Su, L. Wang, W.-J. Huang, Y.-H. Liu, X. Zhang, H.-Q. Huang, T.-Y. Lin, S.-Y. Ma, H.-L. Rao, M. Li, F. Liu, F. Zhang, L.-Y. Zhong, L. Liang, X.-L. Lan, J. Li, B. Liao, Z.-H. Li, Q.-L. Tang, Q. Liang, C.-K. Shao, Q.-L. Zhai, R.-F. Cheng, K. Ru, X. Gu, X.-N. Lin, K. Yi, Y.-R. Shuang, X.-D. Chen, W. Dong, C. Sun, W. Sang, H. Liu, Z.-G. Zhu, J. Rao, Q.-N. Guo, Y. Zhou, X.-L. Meng, Y. Zhu, C.-L. Hu, Y.-R. Jiang, Y. Zhang, H.-Y. Gao, Z.-J. Xia, X.-Y. Pan, L. Hai, G.-W. Li, L.-Y. Song, T.-B. Kang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Q.-Q. Cai, X.-P. Tian, L. Wang, N. Su, D. Xie

Writing, review, and/or revision of the manuscript: Q.-Q. Cai, X.-P. Tian, N. Su, W.-J. Huang, D. Xie

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Q.-Q. Cai, X.-P. Tian, D. Xie

Study supervision: Q.-Q. Cai, D. Xie

This work was supported by grants from National Key R&D Program of China (2017YFC1309001, 2016YFC1302305, to D. Xie); National Natural Science Foundation of China (81603137, 81973384, to X.-P. Tian; 81672686, to Q.-Q. Cai); Special Support Program of Sun Yat-sen University Cancer Center (PT19020401, to Q.-Q. Cai).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.