1 Introduction
Primary central nervous system lymphoma (PCNSL) is an uncommon non-Hodgkin’s lymphoma confined to the central nervous system. Approximately 95% of PCNSLs are diffuse large B cell lymphoma (DLBCL), with the remaining 5% belonging to Burkitt, marginal zone, lymphoblastic, or T cell lymphomas [
1,
2]. This disease is most commonly detected among immunocompromised patients. However, its incidence has increased among healthy elderly individuals over the past decade [
3]. High-dose methotrexate (HD-MTX)-based chemotherapy is the main strategy for treating PCNSL, but this treatment only has a 30% 5-year overall survival (OS) rate, with many patients relapsing [
4]. Therefore, the genetic characteristics of PCNSL should be understood to develop effective treatments.
Unlike systemic DLBCL, studies focusing on PCNSL are limited due to its low incidence and difficulty in sample obtainment. With the development of modern high-throughput molecular techniques, several recurrent chromosomal abnormalities and mutations have been identified in PCNSL, including inactivating the mutation of
PRDM1 [
5] and activating the mutation of the oncogene
CARD11, which may induce the activation of the NF-κB pathway [
6] and Ig-H-BCL6 translocation [
7]. Braggio
et al. [
8] found the biallelic inactivation of the
TOX and
PRKCD genes in PCNSL samples using whole-exome sequencing (WES) but not in systemic DLBCL. Additionally, growing evidence shows that gene mutations are closely linked to the prognosis of PCNSL. For instance, using WES, Fukumura
et al. [
9] demonstrated that the focal deletions or somatic mutations of HLA genes are associated with the poor prognosis of PCNSL patients, while Zhou
et al. [
10] revealed that the mutations of the
CD79B and
GNA13 genes (members of the NF-κB signaling pathway) were related to inferior progression-free survival (PFS) in PCNSL patients. These findings provide novel insights for understanding the pathogenesis of PCNSL. However, the sample size of these studies was small due to the rare incidence of PCNSL. In addition, whole-genome sequencing (WGS) has rarely been applied due to its high cost. Therefore, larger sample sizes, especially Chinese PCNSLs, should be analyzed using WGS.
This study aimed to reveal the genetic characteristics of Chinese PCNSLs using WGS with a focus on structural variations (SVs), copy number variations (CNVs), and gene mutation profiles, and then to explore the relationship of these molecular aberrations with the patients’ clinicopathological features, response to HD-MTX-based polychemotherapies, and prognosis, and finally to explore the differences in gene mutation profiles between Chinese and other cohorts,
2 Materials and methods
2.1 Clinical samples
A total of 95 tumor specimens and 21 paired bone marrow tissues were obtained from 95 PCNSL patients between March 25, 2009 and June 22, 2020 for use in WGS. Following a quality assessment, 68 tumor specimens and 8 paired bone marrow specimens were sequenced successfully. All patients were newly diagnosed with PCNSL (DLBCL) without immune deficiencies, and the samples were collected prior to any type of treatment. Bone marrow tissues obtained from 8 of the 68 patients served as negative controls, and these tissues were not infiltrated. All patients were monitored up to March 1, 2021. General clinical characteristics, including age, sex, Hans’ classification, number of lesions, Eastern Cooperative Oncology Group (ECOG), Karnofsky performance status (KPS), International Extranodal Lymphoma Study Group scores (IELSG), Memorial Sloan–Kettering Cancer Center scores (MSKCC), lactic dehydrogenase (LDH) level, number of lesions, deep involvement, BCL2/MYC expression, lymphocyte count, therapeutic response, and survival time, were collected.
Experiments involving human clinical samples were conducted according to the Helsinki Declaration and were approved by the Institutional Ethical Standard Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine. Informed consent forms were signed by each patient.
2.2 Whole-genome sequencing (WGS)
Genomic DNA was extracted using a formalin-fixed and paraffin-embedded (FFPE) tissue kit (No. 56404, Qiagen, NV, Venlo, Netherlands) following the manufacturer’s instructions. The genomic DNA was cut into fragments of ~200 bp using a focused ultrasonicator (No. M220, Covaris, Woburn, MA, USA). After quality control and equimolar pooling, WGS was performed on the DNBSEQ-T7 sequencing instruments (MGI, Guangdong, China) with a sequencing data volume of 90 G and a mean sequencing depth of 58.82× (range: 23.23×–96.61×). FastQC (version 1.11.4) was used to assess the quality of the raw sequencing data. These data were then processed using Trimmomatic (version 3.6) to remove sequencing adapters and low-quality reads, that is, the joint sequence fragments of the 3′ end and low-quality fragments with a Q value of < 25 and fragments of < 35 bp. To prepare read alignments for analysis, all sequencing data were processed through the Broad Institute’s data processing pipeline. Reads were aligned to the Human Genome Reference Consortium build 37 (GRCh37) using BWA (version 0.5.9-tpx). The BAM files contained reads aligned to the human genome, and their quality scores were recalibrated using the Table Recalibration tool in the Genome Analysis Toolkit (version 4.1.4.0). Variant detection and analysis of the BAM files were performed using the Broad Institute’s Cancer Genome Analysis infrastructure program Mutect 2.
2.3 Data processing
SVs, deletions, duplications, inversions, and translations were annotated using Lumpy-0.2.13 and were screened based on the following filtering conditions: (1) deletions, duplications, and inversions sharing a common overlap of > 10% with the 8 negative control samples were removed; (2) translations involving the same two sites were filtered out; (3) SVs with a mutation frequency of < 15% were filtered out; and (4) SVs with no side or one side located at the gene sites were filtered out. The circos image (circlizeR, R software, R Core Team, Vienna, Austria) demonstrated fusions in different genes identified in ≥ 2 cases.
Following a comparison with GRCh37, the CNVs were analyzed using the CNV kit software package and Nexus software version 5 (Biodiscovery, El Segundo, CA). A copy number of 0 indicated loss, and a copy number of ≥ 3 indicated gain. The “Copynumber” package (version 1.26.0) of the R software was used to draw the frequency plot and heatmap.
Variants of the coding regions were screened based on the following filtering conditions: (1) mutations with mutation allele frequency (MAF) ≥ 0.001 in the 1000 Genomes Project or 1000 Genomes East Asian databases or with a MAF ≥ 0.001 in the Exome Aggregation Consortium (ExAC) or ExAC East Asian databases were removed; (2) mutations detected in the 8 negative controls were removed; (3) mutations with MAF < 0.001 in GenomAD and MAF < 0.001 in ExAC were reserved; (4) mutations with sites that were functionally annotated in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database were retained; and (5) mutations in genes not included in the union set or were derived from the full genetic database for cancer (MSK-Impact, Foundation One) or DLBCL-related gene lists were filtered out (Supplementary File 1). The “Maftools” package (version 2.2.10) of the R software was used to draw the horizontal histogram, which showed the genes with high mutation frequencies, and to perform the co-mutation analysis. The mutations in coding regions were annotated using Annovar (version 2017-07-17).
2.4 Functional enrichment analysis
Gene ontology (GO) analysis, which covered biological processes, cellular components, and molecular function terms, was performed using DAVID to evaluate the enriched functions of the CNVs and mutated genes. Fisher’s exact test was applied to detect the overlap between the genes and the GO annotation list beyond that which would be expected by chance. The KEGG database was used to understand the high-level functions and effects of biological systems. DAVID was also applied to assess the KEGG pathway enrichment of the mutated genes. A P value of < 0.05 was considered significantly enriched.
2.5 Establishment of a prognostic risk scoring system for PCNSL
To establish the training and test cohorts, the 68 samples were randomly divided into a training set and a test set with 51 (75%) and 17 samples (25%), respectively. The training set was subjected to 10-fold cross-validation to account for variability and provide risk estimates. All mutated genes and four clinicopathologic features (age, KPS, deep involvement, and LDH level) were included in the models.
Conventional logistic regression was used to train the models for assessing mortality risk. The methods included logistic regression with a forward selection of variables, and logistic regression with lasso regularization was carried out using the “Glmnet” package. Receiver operating characteristic (ROC) curves were used to estimate model discrimination by calculating the area under the curve (AUC). The test set was divided into low- and high-risk groups according to the cutoff value determined by the ROC curve. Afterward, the OS of the two risk groups was analyzed by Kaplan–Meier (K–M) curves with log-rank tests. A nomogram model was also established according to previously described methods [
11] to visually determine the value of this prognostic risk scoring system in predicting the OS of each PCNSL patient.
2.6 Immunohistochemistry
A total of 68 paraffin-embedded newly diagnosed PCNSLs were studied using a rabbit anti-TMSB4X polyclonal antibody (Ag13914, Proteintech) at 1:400 dilution on an automated immunostainer (Bond, Leica, Wetzlar, Germany).
2.7 Statistical analysis
The Maftools (“clinicalEnrichment”) R package was applied to assess the relationship between the mutated genes and clinical characteristics via one-tailed Fisher’s exact tests. K–M curves with log-rank tests were used to analyze the relationship between the mutation profiles and the PFS and OS of patients with PCNSL. A P value of < 0.05 was considered significant.
3 Results
3.1 Patient characteristics and sequencing data
A total of 68 patients with newly diagnosed PCNSL were sequenced successfully and included in this study. All patients were identified as DLBCL subtype, with 26.5% (18/68) of germinal center B cell-like (GCB) subtype and 70.6% (48/68) of non-GCB subtype. Their clinicopathologic features are summarized in Tab.1. Their median age at diagnosis was 58 years (range: 24–80), 58.8% (40/68) were male and 41.2% (28/68) were female, 55.9% (38/68) had unifocal and 44.1% (30/68) had multifocal lesions, 54.4% (37/68) had an ECOG score of 1 to 2, and 45.6% (31/68) had an ECOG score of 3 to 4. After the diagnosis, 37 patients received HD-MTX-based chemotherapy only, 28 received HD-MTX-based chemotherapy combined with radiotherapy, and the other 3 were given HD-MTX-based chemotherapy followed by auto-transplantation. The genomic DNA isolated from tumor samples and 8 paired bone marrow control samples was subjected to WGS with a mean sequencing depth of 58.82× (range: 23.23×–96.61×).
3.2 SVs of the 68 PCNSL samples identified by WGS
SVs, including deletions, duplications, inversions, and translocations (break end), are common causes of gene functional alterations and have been identified in several kinds of cancers [
12–
14]. In this study, 23 338 SVs were identified in the 68 PCNSL samples (mean: 343, range: 1–858), with deletions being the most frequent events (Fig.1). To explore how SVs drive PCNSL occurrence, those genes related to SV breakpoints were identified across the whole genome (Fig.1). The genes marked with red were located in the frequent SV region and included
CACNA1B,
BCLAF1,
OPRM1, and
RSPO1. The top 35 genes containing SV breakpoints are shown in Fig.1, with the top 10 genes being
RAD51B,
CHD2,
PTPRD,
BCLAF1,
ETV6,
FGF12,
MS12,
NRXN3,
TP63, and
MAP3K13. Given that translations of
BCL6, BCL2, and MYC were reported in PCNSL, these rearrangements were also screened in this cohort.
BCL6 and
BCL2 rearrangements were found in 5 (7.4%) and 1 (1.5%) of the 68 samples, respectively, while no
MYC rearrangement was identified. This difference may be due to the use of a different study cohort.
Further analysis showed that CLCNKB-FAM131C, FAM234B-MIR7641-2, MYOF-WBP1L, and SUSD6-CCDC50 were the most frequent fusions, which were detected in 28, 19, 17, and 11 samples, respectively. The fusions of different genes detected in ≥ 2 samples are shown in Fig. S1. The other frequent fusions included SLC2A5-BTBD7, MRPS18A-B4GALT1, FAM238B-FAM238C, LLPH-DT-TMBIM4, KCNMA1-AS3, and KCNMA1.
To explore the clinical value of SVs in PCNSL, the associations between SV numbers and the OS or PFS of patients with PCNSL were assessed. These 68 patients were divided into an SV-high group (median SVs ≥ 359) and a SV-low group (median SVs < 359), with 34 patients in each group. The SV number did not significantly influence either the OS or PFS of PCNSL patients (Fig.1 and 1D).
3.3 CNV profiles across the 68 PCNSL samples
The CNVs of the 68 samples were profiled to identify significant peaks that might contain potential driver genes. A total of 2952 CNVs were found in the samples (mean: 43, range: 8–493). Copy loss occurred in all 68 samples, while gains were detected in 53 samples (77.9%). The most common losses were those of 16p11.2 (97.1%), 6p11.2 (97.1%), 22q11.1 (91.2%), 2q11.1 (91.2%), 9q13 (89.7%), 2q11.2 (89.7%), 9p11.2 (86.8%), 1p36.21 (82.4%), 22q11.22 (77.9%), and 15q11.2 (75.0%), whereas the most common gains were those of 6q13 (22.1%), 16q13 (22.1%), 17q11.2 (22.1%), 3p21.31 (20.1%), 7p13 (19.1%), 9q13 (17.6%), 2q13 (17.6%), 19p13.2 (17.6%), 9p13.2 (17.6%), and 17p13.2 (17.6%) (Fig.2). Previously reported CNVs, such as deletions of 6p21, 6q, and 9p21.3 and gains of 7q, 11q, and 9p24.1 [
8,
15–
17], were also found in the cohort, with frequencies of 41.2%, 20.6%, 11.8%, 25.0%, 23.5%, and 2.9%, respectively.
GO enrichment analysis was performed to assess the biological functions of the 7791 genes related to the CNVs. The genes were significantly enriched in the biological process (BP) terms of innate immune response, natural killer cell activation involved in immune response, B cell proliferation, humoral immune response, and positive regulation of peptidyl−serine phosphorylation of STAT protein, the cellular component (CC) terms of nucleosome and intermediate filament, and the molecular function (MF) terms of type I interferon receptor binding and ribonuclease activity (Fig.2).
To reveal the role of CNVs in predicting PCNSL prognosis, the 68 samples were divided into a CNV-high group (median CNV: ≥ 24.5) and a CNV-low group (median CNV: < 24.5). Interestingly, patients with high CNV levels had lower OS (Fig.2) and PFS (Fig.2) compared with patients with low CNV levels, hence underscoring the vital role of CNV number in PCNSL.
3.4 Recurrently mutated protein-coding genes across the 68 PCNSL samples
The mutation profiles of coding regions in the 68 PCNSL patients were evaluated based on the WGS results. A total of 1424 mutated sites in 263 genes were detected (Supplementary File 2). The median number of gene alterations per patient was 2 (range: 1–46). The most common mutation type was missense mutation (Fig.3). The number of detected single-nucleotide-variants (SNVs) was approximately 11–12-fold the number of insertions and deletions (Fig.3). Among the 263 mutated genes, 36 frequently mutated in ≥ 10% of the samples, with the top 10 genes being IGLL5 (68%), PIM1 (68%), MYD88 (53%), CD79B (47%), BTG2 (40%), KMT2D (37%), TBL1XR1 (31%), PCLO (28%), HIST1H1E (26%), and BTG1 (25%) (Fig.3).
By checking the 263 mutation genes in the cancer hotspots, Genomic Data Commons, intogene, and OncoKB databases, 263 genes, except for LRRN3, SLITRK3, RAG2, ZNF608, HIST2H2AB, and TLDC2 were identified as cancer drivers, among which 81 (AR, B2M, BCL2, BTG1, BTG2, CARD11, CD79A, CD79B, CDKN2A, CREBBP, DTX1, DUSP2, HIST1H1E, HLA-B, IGLL5, IRF4, ITPKB, KLHL6, KMT2D, MPEG1, MYC, MYD88, PAX5, PIK3C2G, PIM1, TBL1XR1, TMSB4X, TP53, UBE2A, XPO1, TET2, SPOP, EP300, DDX3X, CXCR4, ATM, AMER1, EZH2, BRCA1, SMARCA4, BTK, PRDM1, MET, JAK3, CDK12, TBX3, ERBB4, MED12, ATR, FGFR4, PTCH1, NF1, STAG2, PBRM1, KDR, ARID1A, FAS, STAT3, FGFR2, BRAF, APC, CD58, EBF1, SPEN, ID3, GNA13, NOTCH2, CIITA, DOT1L, SOCS1, TNFAIP3, CDKN1B, NOTCH1, MEF2B, SGK1, CCND3, BCL6, STAT6, PTPN6, MAP2K1, and CDH1) were DLBCL driver genes.
GO enrichment and KEGG pathway analysis were then applied to assess the enriched signaling pathways of the 263 mutated genes. The GO analysis showed that these mutated genes were enriched in “BP” terms of regulation of phosphatidylinositol 3 kinase (PI3K) signaling and positive regulation of ERK1 and ERK2 cascades, MAPK cascade, and B cell receptor signaling pathway, the “CC” terms of nucleus and nucleoplasm, and the “MF” terms of PI3K activity, Ras guanyl nucleotide exchange factor activity, ATP binding, and protein binding (Fig. S2A). The KEGG analysis demonstrated that the mutated genes were enriched in the Rap1 signaling pathway, PI3K-Akt signaling pathway, signaling pathways regulating the pluripotency of stem cells, ErbB signaling pathway, and B cell receptor signaling pathway (Fig. S2B). Genes in the BCR/PI3K signaling pathway were significantly mutated, including the previously reported genes
CD79B,
CARD11, and
BTK [
18] (Fig. S2C). Therefore, the mutated genes might participate in PCNSL pathogenesis by modulating these pathways.
3.5 Differential mutation profiles of Chinese PCNSLs
The mutated genes identified in the present cohort using WGS were compared with those identified in previous studies that used WES, including the articles of Vater
et al. [
19], Bruno
et al. [
20], and Fukumura
et al. [
9]. A total of 195 genes were exclusively found in the present cohort compared with the above three cohorts (Fig.4). After integrating the genes exclusively detected in the present cohort with other 3 reports [
8,
10,
21], 138 genes were retained (Fig.4), of which 6 (
ROBO2,
KMT2C,
CXCR4,
MYOM2,
BCLAF1, and
NRXN3) were detected in ≥ 10% of patients in the present cohort. By comparing the mutation profile of the present cohort with that of 2 other Chinese cohorts [
10,
22], mutations in 72 genes were exclusively detected in the present cohort, and 28 genes were commonly mutated in the 3 Chinese cohorts (Fig.4). The mutation frequencies of
PIM1,
CD79B,
KMT2D,
IRF4,
CARD11,
PRDM1,
GNA13,
CREBBP, and
TP53 were ≥ 10% in the present cohort.
The mutation landscapes detected in Asian patients (including 3 Chinese cohorts [
10,
22] and 2 Japanese cohorts [
9,
21]) and non-Asian patients (including 1 American cohort [
20] and 1 German cohort [
19]) with PCNSL were then compared. A total of 406 genes were exclusively detected among Asian PCNSL patients (Fig.4), with 38 co-mutated genes identified in both the Japanese and Chinese cohorts, among which
KMT2A,
AR, and
ROS1 had a mutation frequency of ≥ 10% in the present cohort.
The mutated genes in the present cohort of PCNSLs were also compared with those in three reports targeting the system DLBCL [
23–
25] (Fig.4). A total of 19 genes (
RAG1,
FGF4,
MYCN,
BCORL1,
TGM7,
HSD3B1,
MTAP,
ESR1,
GATA4,
MPL,
FGF12,
FOXL2,
MAP3K13,
FGF6,
FGF10,
FGF3,
FGF14,
SOX2, and
FGF19) were found exclusively in the present cohort. These 19 genes were then compared with other common brain tumors, such as glioma [
26] and meningioma [
27,
28]. The mutation of only 1 of these genes (
BCORL1) was detected in glioma (Fig.4), thereby suggesting that the remaining 18 genes (
RAG1,
FGF4,
MYCN,
TGM7,
HSD3B1,
MTAP,
ESR1,
GATA4,
MPL,
FGF12,
FOXL2,
MAP3K13,
FGF6,
FGF10,
FGF3,
FGF14,
SOX2, and
FGF19) might be used for the differential diagnosis of PCNSL from system DLBCL, glioma, and meningiomas.
3.6 Mutation profiles associated with patients’ clinical features
The possible link between the mutated genes and the clinical features of PCNSL patients was then tested. The mutation frequencies of HIST1H1E, CARD11, and ROBO2 in the GCB group were significantly higher than those in the non-GCB group, while the mutation frequencies of MPEG1, DUSP2, and PRDM1 in the GCB group were significantly lower than those in the non-GCB group (Fig.5). The mutation frequencies of PCLO, CREBBP, PRDM1, ATM, and NOTCH2 showed significant differences among the high-, middle-, and low-risk groups based on their IELSG scores (Fig.5). In risk groups defined based on MSKCC scores, the BRCA2, CREBBP, and ERBB4 mutation frequencies showed noticeable differences across different risk groups (Fig.5). In addition, the mutation frequencies of MPEG1 and AR in the multifocal group were significantly higher than those in the unifocal group (Fig.5). The mutation frequencies of PCLO and AR were higher among patients with deep involvement than among those patients without deep involvement (Fig.5). IRF4 mutation was less common among patients with double expression of both BCL2 and MYC, while TMSB4X and BTG1 mutations were more common among these patients (Fig.5). An increased AR mutation frequency was significantly related to a lower lymphocyte count (≤ 2.0) (Fig.5). KMT2D mutation frequency was positively associated with LDH level (Fig.5). Additionally, KMT2A mutation was only detected in the KPS < 70 group (Fig.5). However, no significant differences in gene mutation frequency were observed based on age and ECOG score. These results linked the gene mutation profiles to the clinical features of PCNSL patients.
3.7 Mutation profiles associated with patients’ treatment response and prognosis
The relationship between the mutation profiles and the patients’ treatment response and progression was then analyzed. The CARD11 mutation was detected in the response group (CR + PR) (21.3%) but not in the non-response group (SD + PD) (P = 0.0379) (Fig.6), whereas the KMT2A mutation frequency was lower in patients with two-year progression (4.1%) than in those without progression (26.3%) (P = 0.016) (Fig.6). Therefore, CARD11 mutation might be a potent marker for treatment response, and KMT2A mutation could predict the two-year progression of PCNSL.
The relationship between the mutation profiles and patients’ prognosis was then examined. A total of 36 genes with a mutation frequency of ≥ 10% were evaluated using K–M curves to understand their effects on the OS and PFS of PCNSL patients. Only TMSB4X mutation was significantly associated with lower OS compared with the wild-type TMSB4X group (Fig.6), while CD79B mutation was significantly associated with lower PFS (Fig.6). The mutation frequencies of 10 genes, namely, PCLO, PAX5, ZNF217, MSH6, EZH2, B2M, SOCS1, TNFAIP3, CDKN1B, and BCL6, showed significant differences among patients with and without TMSB4X mutation (Fig.6).
3.8 High expression of TMSB4X protein associated with poor prognosis
The expression of TMSB4X and its association with the prognosis of systemic DLBCL patients was initially examined. The expression of TMSB4X was higher among DLBCL patients than those with other common tumors both at the transcriptional level as shown by the GEPIA database (Fig.7) and the translation level as retrieved from the ProteinAtlas database (Fig.7). The TMSB4X transcriptional level of DLBCL in the GEPIA database was higher than that in the normal control group (Fig.7). According to the GEO database (GSE11392 cohort), the TMSB4X transcriptional level in PCNSL patients was obviously higher than that of nodal DLBCL (Fig.7), but no significant difference was found between PCNSL patients and extranodal DLBCL (Fig.7). To reveal the effect of TMSB4X mutation on its expression in PCNSL, the TMSB4X protein expression levels in PCNSL patients carrying mutated TMSB4X and wild-type TMSB4X were compared. Fig.7 shows the mutation schematic diagram of the TMSB4X gene in its domain region. The expression of TMSB4X protein in patients carrying TMSB4X mutation was slightly higher than that in patients carrying the wild-type TMSB4X (Fig.7). Typical IHC images of the TMSB4X protein were shown in Fig.7 and 7I. On the basis of the average integrated option density (AOD) of TMSB4X staining, the 68 PCNSL patients were divided into the low TMSB4X expression group (AOD ≤ 0.08) and high-risk (AOD > 0.08) group according to the cutoff value determined by the ROC curve. High TMSB4X expression was associated with lower OS in PCNSL patients compared with the low TMSB4X group (Fig.7).
3.9 Establishment of a new prognostic risk scoring system for PCNSL
A prognostic model was built to further clarify the clinical value of gene mutation profiles in PCNSL. Lasso regression and 10-fold cross-validation were used to select risk factors among the 263 mutated genes and 4 clinicopathologic features (age, KPS, deep involvement, and LDH level) (Fig.8). Following the training and validation, KPS and 6 genes (BRD4, EBF1, BTG1, CCND3, STAG2, and TMSB4X) were incorporated into the prognostic risk scoring system, which was validated by ROC analysis (AUC, 0.8182; 95% CI, 0.6356–0.9854; kappa, 0.7213) (Fig.8). The probability of high risk was calculated as follows = 1/[1+exp(–0.37993+0.1624KPS+0.63449BRD4+0.52019EBF1–0.41161BTG1+0.04921CCND3+0.23689STAG2–0.61347TMSB4X)]. The test set was divided into the low-risk (≤ 0.504) and high-risk (> 0.504) groups according to the cutoff value determined by the ROC curve. The K–M curve showed that the patients in the high-risk group had significantly lower OS than those in the low-risk group (Fig.8). Subsequently, KPS and the above six genes were applied to construct a nomogram model for predicting the 1-, 2-, and 3-year OS rates of PCNSL patients (Fig.8). The mutations of three of the six genes (BTG1, STAG2, and BRD4) showed interactions with other high-frequency mutations (Fig. S3).
The clinical value of the IELSG [
29] and MSKCC scoring systems [
30], which are two prognostic scoring systems developed specifically for PCNSL, was also assessed. The IELSG system showed a higher prognosis value in both OS and PFS (Fig. S4A and S4B) compared with the MSKCC system (Fig. S4C and S4D) but was only confined to predicting the low-risk group.
4 Discussion
With the development of precision medicine, the genetic features underlying PCNSL have received increasing attention. Next-generation sequencing provides a path to explore the genetic profiles of PCNSL but is far from sufficient. Recently, Radke
et al. [
31] from Germany compared the genomic and transcriptional landscapes of 51 CNS lymphomas (42 cases of PCNSL, 6 cases of secondary CNS lymphomas, and 3 cases of EBV + CNS lymphomas) with those of 39 follicular lymphoma and 36 DLBCL cases outside the CNS via WGS and RNA sequencing and concluded that PCNSL can be clearly distinguished from DLBCL due to its distinct expression profiles and IG expression and translocation patterns. In this study, the genetic landscape of 68 Chinese patients with PCNSL was revealed using WGS, which can completely profile the status of the genome. To the best of the authors’ knowledge, this study is the largest population study of PCNSL in China.
Mutations in the
IGLL5,
PIM1,
MYD88,
CD79B,
BTG2,
KMT2D,
TBL1XR1,
PCLO,
HIST1H1E,
BTG1,
DTX1, and
CARD11 genes were the most common coding region variations in the present cohort. In addition to the common mutations previously identified in PCNSL [
9,
15,
32], the mutations in the
ROBO2,
KMT2C,
CXCR4,
MYOM2,
BCLAF1, and
NRXN3 genes were exclusively detected in the present cohort with a ≥ 10% frequency rate, thus highlighting the high heterogeneity and complexity of PCNSL. The mutation landscapes detected in Asian patients (3 Chinese cohorts [
10,
22] and 2 Japanese cohorts [
9,
21]) and non-Asian patients (1 American cohort [
20] and 1 German cohort [
19]) with PCNSL were also compared. A total of 38 co-mutated genes were identified in the Japanese and Chinese cohorts, among which
KMT2A,
AR, and
ROS1 had a mutation frequency of ≥ 10% in the present cohort. The mutated genes detected in PCNSL, DLBCL, glioma and meningioma were also compared. A total of 18 genes, namely,
RAG1,
FGF4,
MYCN,
TGM7,
HSD3B1,
MTAP,
ESR1,
GATA4,
MPL,
FGF12,
FOXL2,
MAP3K13,
FGF6,
FGF10,
FGF3,
FGF14,
SOX2, and
FGF19, were detected only in PCNSL, of which 6 (
FGF3/
6/
10/
12/
14/
19) belonged to the fibroblast growth factors (FGFs) family, which contains > 20 members that encode secreted polypeptides and act through tyrosine kinase receptors [
33]. Accumulated evidence shows that FGFs are involved in the progression of multiple types of tumors, such as non-small cell lung cancer (NSCLC) [
34], ovarian cancer [
35], and pancreatic cancer [
36]. Noticeably, targeting FGF signaling is a promising therapeutic strategy [
37,
38]. Results indicated that the mutations in the genes of the FGF family might be used for the differential diagnosis of PCNSL from DLBCL and other brain tumors.
Evidence also shows that CNVs and mutation profiles are associated with the prognosis of many kinds of cancers [
39,
40]. In this study, the
CD79B mutation and a high level of CNV were significantly associated with the lower PFS of PCNSL. Similarly, Zhou
et al. [
10] revealed that the
CD79B mutation was related to inferior PFS among PCNSL patients. In addition, the
TMSB4X mutation and a high level of CNV were significantly associated with lower OS.
TMSB4X encodes a 43-amino-acid short peptide implicated in multiple biological activities, including inflammation, cell survival, and metastasis [
41,
42]. The mutation of
TMSB4X was also previously found to be a recurrent variation in PCNSL [
19] and a target of aberrant somatic hypermutation in systemic DLBCL [
43]. However, whether or not
TMSB4X affects the prognosis of PCNSL remains unknown. Therefore, the role of
TMSB4X in PCNSL warrants further research.
The IELSG [
29] and MSKCC scoring systems [
30] are two prognostic scoring systems developed specifically for PCNSL. However, these scoring systems do not include genetic profiles. In the present cohort, the IELSG system showed a higher prognosis value than the MSKCC system but was only confined to predicting the low-risk group. In this study, a prognostic model combining KPS with six mutated genes (
BRD4,
EBF1,
BTG1,
CCND3,
STAG2, and
TMSB4X) was developed. Among these genes,
EBF1 (early B cell factor 1) and
BTG1 are tumor suppressors [
44,
45].
EBF1 belongs to a family of four highly conserved DNA binding transcription factors that are involved in the differentiation and maturation of B-progenitor lymphoblasts. Specifically,
EBF1 is required for the expression of
PAX5 (paired box 5) during B cell development, and the deletion of
EBF1 leads to the pathogenesis, drug resistance, and relapse of B-progenitor acute lymphoblastic leukemia (ALL) [
45]. Hodkinson
et al. [
46] recently revealed that
EBF1 mutation was not detected in responders (ibrutinib plus nivolumab) in relapsed DLBCL, follicular lymphoma, or Richter’s transformation patients, thereby suggesting that
EBF1 mutation may be linked to poor response to ibrutinib and nivolumab.
CCND3 is a potential driver of Burkitt lymphomagenesis [
47] and mantle cell lymphoma (MCL) [
48], a treatment-refractory subtype of malignant lymphoma.
BRD4 (bromodomain containing 4) is a transcriptional and epigenetic regulator that is preferentially located at the super-enhancer regions of several crucial enhances, such as
C-MYC,
BCL-XL, and
BCL-6, and can lead to cancer development [
49,
50].
I-BET151, a BRD4 inhibitor, has been reported to suppress cell proliferation in MCL.
BRD4 can target genes involved in the BCR signaling pathway, including
BLNK (B cell linker),
PAX5, and
IKZF3 (IKAROS family zinc finger 3) [
51], thereby suggesting that
BRD4 may be a potential treatment target in lymphoma. The
STAG2 gene encodes a cohesion complex subunit that regulates chromatid separation, mutations of which can result in enhanced chromosomal stability and be detected in acute leukemias [
52] and adult T cell lymphoma [
53]. Results of this work might provide a new molecular classification method for predicting the prognosis of PCNSL, which needs to be further verified in clinical practice.
This study has several limitations, one of which lies in the lack of ideal normal controls (such as paraneoplastic normal brain tissue) and the use of only 8 paired bone marrow tissues as negative controls. To minimize the error induced by this small negative control size, the mutation sites and CNVs were filtered by GRCh37 and/or the GenomAD, ExAC, and 1000 Genomes Project databases. Moreover, FFPE samples were used in this study. The DNA extracted from these samples is problematic for mutation testing, especially for amplicon-based massively parallel sequencing, owing to DNA fragmentation and the artificial C:G > T:A SNVs caused by the deamination of cytosine to uracil. However, the FFPE samples used in this study were pretreated with uracil DNA glycosylase, which can eliminate uracil-containing DNA molecules [
54]. Moreover, the PPFE samples were submitted to WGS and not to amplicon-based massively parallel sequencing, and the proportion of C > T SNVs (51%) detected in the present cohort is no greater than that detected in a previous report (58%) [
55].
In conclusion, this study further describes the genomic landscape of Chinese PCNSL using WGS, including SVs, CNVs and gene mutations in PCNSL, and explores their clinical values. The TMSB4X mutation and a high level of CNV are significantly associated with lower OS, while the CD79B mutation and a high level of CNV are significantly associated with lower PFS. A prognostic model containing KPS and six mutated genes was also built, by which the 1-, 2-, and 3-year OS of PCNSL patients can be predicted accurately and conveniently. This study depicts the genomic characteristics of Chinese PCNSLs, thereby enriching the present understanding of the genetic pathogenesis of PCNSL.
PDF
(8853KB)
