Integrated Genomic and Tumor Microenvironment Subtyping Improved Risk Stratification in Primary Central Nervous System Lymphoma

ABSTRACT

Current prognostic models fail to capture the biological complexity of primary central nervous system lymphoma (PCNSL). We integrated whole‐genome sequencing and multiplex immunofluorescence in 68 treatment‐naïve patients to define four genomic subtypes (C1, C2, C3, and C4) with divergent survival (C4 worst: median overall survival [OS], 26 months). In parallel, a novel tumor microenvironment (TME) classification based on CD8⁺T/M2 macrophage ratio stratified patients into High (> 1.5), Intermediate (0.8–1.5), and Low (< 0.8) groups. Unexpectedly, the Intermediate TME group showed the poorest outcomes (5‐year OS: 10%). Integration revealed a lethal subgroup (C4 + Intermediate TME; 9.8% of cohort) with a median OS of 3.0 months (hazard ratio = 7.24, p = 0.006). Prognostic nomograms incorporating these subtypes showed promising discriminative performance in internal validation (C‐index > 0.78), but external validation is needed. Together, these findings identify a high‐risk biological subset and provide a hypothesis‐generating framework for future biomarker‐driven risk stratification and therapeutic discovery in PCNSL.

1. Introduction

Primary central nervous system lymphoma (PCNSL) is a rare and aggressive extranodal diffuse large B‐cell lymphoma (DLBCL) confined to the central nervous system. Despite high‐dose methotrexate (HD‐MTX)‐based therapy, prognosis remains poor, with 5‐year overall survival (OS) rates below 40% and marked heterogeneity in treatment response [1, 2]. Current clinical prognostic models, such as the International Extranodal Lymphoma Study Group (IELSG) and Memorial Sloan Kettering Cancer Center (MSKCC), rely solely on clinical parameters [3, 4], lacking molecular and tumor microenvironment (TME) inputs that may better capture biological complexity in the era of targeted therapies [5].

Genome‐wide expression analyses have demonstrated that PCNSL harbors distinct molecular signatures and greater heterogeneity compared to systemic DLBCL [6]. Genomic studies have identified frequent alterations in MYD88, CD79B, and other B‐cell receptor (BCR) signaling components [7]. Schmitz et al.'s genetic classification of systemic DLBCL (MCD, BN2, N1, EZB) identifies most PCNSL cases as belonging to the MCD subtype [8]. Verdin et al. defined four molecular clusters (CS1‐CS4) of PCNSL with divergent genomic and immune profiles [9]. Despite these advances, the genetic landscape of PCNSL remains incompletely characterized, and the functional interplay between genomic drivers and TME dynamics is still poorly understood.

The TME is a critical modulator of lymphoma progression and therapy response [10]. The brain's immune‐privileged status shapes a distinct TME [11], characterized by a predominance of CD8⁺ T cells and CD163⁺tumor‐associated macrophages (TAMs), albeit with lower overall immune infiltration than systemic DLBCL [12]. Frequent immune escape mechanisms (e.g., HLA loss, PD‐1/PD‐L1 upregulation) further reflect its highly immunosuppressive nature [13]. However, comprehensive TME characterization remains limited, and significant heterogeneity impedes biological understanding and clinical management [14, 15]. Although retrospective studies suggest prognostic roles for T cells and TAMs, reported associations are inconsistent, and no clinically applicable TME biomarkers have been robustly validated [16].

To address these gaps, we performed an integrated analysis of whole‐genome sequencing (WGS) and spatially resolved TME profiling in 68 treatment‐naïve patients with PCNSL. Our objectives were: (1) To define genomic subtypes using unsupervised clustering of driver alterations; (2) To establish a novel TME classification leveraging CD8⁺T/M2 macrophage ratios; (3) To evaluate the prognostic value of integrated genomic‐immune subtyping. This study establishes a hypothesis‐generating framework for risk stratification and precision therapy in PCNSL.

2. Methods

2.1. Patient Cohort

We retrospectively analyzed 68 immunocompetent patients with newly diagnosed PCNSL according to the 2016 World Health Organization criteria. The median age at diagnosis was 58 years (range, 24–80). These patients received treatment at the Second Affiliated Hospital of Zhejiang University between March 2009 and June 2020. All patients were treatment‐naïve at the time of diagnostic biopsy and subsequently received high‐dose methotrexate‐based induction therapy. We excluded individuals with immunodeficiency disorders, secondary CNS lymphoma involvement, or insufficient tumor tissue. Treatment response was evaluated using the International Primary CNS Lymphoma Collaborative Group response criteria. Clinical characteristics of the cohort are comprehensively detailed in Table S1.

2.2. Genomic Analysis

Genomic DNA was extracted from formalin‐fixed paraffin‐embedded (FFPE) tumor tissues and paired bone marrow aspirates (normal controls) using the QIAamp DNA FFPE Tissue Kit (Qiagen, Cat. No. 56404). Whole‐genome sequencing libraries were prepared from 200‐base pair fragmented DNA generated by acoustic shearing (Covaris M220). Sequencing was performed on the DNBSEQ‐T7 platform (MGI Tech), achieving a mean coverage depth of 58.82×, with a range of 23.23× to 96.61×. Raw sequencing data totaling 90 gigabytes per sample have been deposited in the National Genomics Data Center (Accession No. PRJCA009559).

Raw sequence reads underwent quality control assessment using FastQC software version 1.11.4. Adapter sequences and low‐quality reads (Phred score < 25 or length < 35 bp) were trimmed with Trimmomatic version 3.6. Processed reads were aligned to the human reference genome GRCh37 using BWA‐MEM algorithm version 0.5.9. Somatic mutations were called using GATK version 4.1.4.0, structural variants (SVs) were detected with Lumpy version 0.2.13, and copy number variations (CNVs) were analyzed using CNVkit combined with Nexus software version 5.0 (Biodiscovery).

2.3. Genomic Feature Selection and Consensus Clustering

We performed unsupervised consensus clustering of 271 curated genomic features including 256 non‐synonymous mutations, 4 SVs, and 11 CNVs (Table S2). These features were selected based on previously reported gene alterations relevant to PCNSL and DLBCL. This analysis utilized the R package ConsensusClusterPlus with 1000 iterations. The optimal number of molecular subtypes was determined as four clusters (k = 4) based on maximum consensus matrix stability and minimal area change in cumulative distribution function plots.

2.4. Multiplex Immunofluorescence Analysis of TME

FFPE tissue sections from 61 tumors underwent seven‐color multiplex immunofluorescence staining (alphaxbio). The antibody panel included CD8 for cytotoxic T cells (ZSGB‐Bio, RRID:AB‐2890107, 1:200), CD68 for pan‐macrophages (Alpha X Bio, AXB3010, 1:200), CD163 for M2 macrophages (ZSGB‐Bio, ZM0428, 1:200), PD‐1 for T‐cell exhaustion (ZSGB‐Bio, RRID:AB‐2921363, 1:100), SPP1 (Abcam, RRID:AB‐2894860, 1:2000), and CXCL9 (Abcam, RRID:AB‐3665738, 1:100). Detailed antibody specifications are provided in Table S3. Automated staining was performed on the AlphaXPainter X30 platform following manufacturer protocols. Whole‐slide imaging was conducted using a Zeiss Axioscan 7 scanner at 20× magnification.

Tumor regions were annotated based on CD20‐positive expression determined by immunofluorescence or immunohistochemistry. Cellular quantification was performed using HALO image analysis software version 3.5 (Indica Labs). We calculated immune cell densities as cells per square millimeter (cells/mm²) and proportions as percentages of all nucleated cells (ANC). Key immune subsets were defined as follows: M1‐like macrophages were identified as CD68⁺CD163⁻ cells, M2‐like macrophages as CD68⁺CD163⁺ cells. The CD8⁺T/M2 macrophage ratio was calculated by dividing the density of CD8⁺ T cells by the density of M2‐like macrophages. Patients were stratified into High (> 1.5), Intermediate (0.8–1.5), and Low (< 0.8) TME groups based on optimal receiver operating characteristic curve cutoffs for overall survival prediction. According to the prior studies on the immune landscape of solid tumor [17], PCNSLs were classified into “HOT,” “Intermediate” and “COLD” phenotypes by two pathologists (SM Wei and BZ Li).

Tumor regions were annotated based on CD20‐positive expression determined by immunofluorescence or immunohistochemistry. For comparison, “para‐cancerous” regions were defined as morphologically non‐neoplastic areas within the same biopsy section, confirmed by a pathologist (S.M. Wei) to lack CD20⁺ lymphoma cells and cytological atypia on hematoxylin–eosin staining. These peritumoral regions represent the best available internal control, although they may not be entirely equivalent to true normal brain tissue due to potential reactive changes. Cellular quantification was performed using HALO image analysis software version 3.5 (Indica Labs).

2.5. Penalized Regression Sensitivity Analysis

To assess the robustness of our multivariate Cox model against overfitting given the limited events‐per‐variable ratio (37 deaths for 11 candidate variables, 50 progressions for 10 candidate variables), we performed Lasso (L1‐penalized) Cox regression as a sensitivity analysis using the glmnet R package. Ten‐fold cross‐validation was used to select the optimal penalty parameter λ. Both λ.min (the λ that minimized partial likelihood deviance) and λ.1se (the largest λ within one standard error of the minimum) were examined. Variables with non‐zero coefficients at λ.min were considered potentially important, while those retained at λ.1se were deemed the most robust predictors.

2.6. Internal Validation of TME Cutoff Selection

To account for optimism bias in cutoff determination, we performed two internal validation procedures. First, bootstrap resampling with 1000 iterations was conducted. In each bootstrap sample (sampled with replacement from the original 61 patients), we re‐calculated the optimal CD8⁺T/M2 ratio cutoffs (High > 1.5, Intermediate 0.8–1.5, Low < 0.8) using the same log‐rank‐based maximization approach as in the original analysis. The resulting cutoffs were then applied to the original dataset to compute Harrell's C‐index for overall survival (OS) and progression‐free survival (PFS). The median C‐index from the 1000 bootstrap iterations is reported as the optimism‐corrected estimate, with 2.5th–97.5th percentiles as the 95% confidence interval. Second, 5‐fold cross‐validation was performed (10‐fold was not feasible due to the sample size of 61). The data were randomly partitioned into 5 equal folds. In each of the 5 iterations, 4 folds (≈49 patients) were used to determine the optimal cutoffs, which were then applied to the remaining held‐out fold to calculate the C‐index and log‐rank p value. The mean of the 5 iterations is reported. These analyses were implemented using custom R scripts.

2.7. Statistical Analysis

All statistical analyses were performed using R statistical software version 4.3.2 and SPSS version 27.0. Categorical variables were compared using Pearson's χ ² test or Fisher's exact test as appropriate. Continuous variables were compared using the Student's t‐test or one‐way ANOVA for normally distributed data; otherwise, the Mann–Whitney U test or Kruskal‐Wallis test was applied. Survival times were compared using the log‐rank test. Multivariate Cox proportional hazards regression models included variables with p < 0.1 in univariate analysis. We constructed prognostic nomograms based on independent predictors identified in multivariate analysis. Model performance was evaluated using concordance indices and calibration plots with 1000 bootstrap resamples for internal validation. Functional pathway enrichment analysis was conducted using DAVID bioinformatics resources version 6.8 with false discovery rate (FDR) correction (Benjamini‐Hochberg method) applied at a threshold of FDR < 0.05. A two‐sided p‐value < 0.05 defined statistical significance throughout the study.

3. Results

3.1. Genomic Subtypes Predict Survival in PCNSL

We performed WGS on tumor samples from 68 treatment‐naïve patients with PCNSL and identified 271 genomic alterations including mutations, SVs, and CNVs (Figure 1A). Unsupervised consensus clustering delineated four molecular subtypes designated C1 through C4 with distinct genomic profiles (Figure 1B). These subtypes exhibited significantly divergent OS outcomes (p = 0.020) (Figure 1C). Patients with C4 tumors demonstrated the poorest prognosis, with a median OS of 26.0 months and a 5‐year survival rate of 16.7%. In contrast, genomic subtype C2 was independently protective for OS (HR 0.32, 95% CI: 0.11–0.94, p = 0.038), with median OS not reached and a 5‐year survival rate of 70.6%. No significant differences were observed among subtypes in clinical parameters (Table S4).

FIGURE 1.

Genomic Landscape and prognostic significance of new genomic subtypes in primary central nervous system lymphoma (PCNSL). (A) The study design including the workflow and the data composition for each cohort. (B) Cumulative distribution function (CDF) curves for consensus matrices of k = 2 through k = 8 (above). Heatmap of driver alterations across four genomic subtypes (below). Consensus clustering matrix (k = 4) for 68 treatment‐naïve patients with PCNSL based on 271 genomic features (mutations, SVs, and CNVs) delineated four new genomic subtypes (C1–C4). (C) Kaplan–Meier overall survival (OS) curves for C1‐C4 subtypes (log‐rank p = 0.020). Median OS: C1 (43 months), C2 (not reached), C3 (not reached), C4 (26 months). (D) Prevalence of molecular abnormalities per genomic subtype (top features displayed). (E) Enrichment of driver events in each subtype compared with the remainder of the cohort (Fisher's exact test or chi‐square test, p < 0.05).

Each genomic subtype displayed distinct driver alterations and pathway dysregulation, with varying prevalence of specific mutations and copy number alterations (Figure 1D,E). C1 tumors were enriched in DNA repair pathway defects featuring recurrent mutations in BRIP1 and MTAP coupled with chromosomal amplifications at 3p21.31 and 16q13. C2 tumors demonstrated apoptotic dysregulation driven by BCL2 and IRF4 mutations. C3 tumors exhibited hyperactivated BCR‐MAPK signaling with frequent alterations in CD79A, CARD11, BCL6, and EGFR. C4 tumors were defined by epigenetic dysregulation and mismatch repair (MMR) deficiency involving KMT2A, TET2, and MSH6 mutations.

We compared the prognostic performance of our C1–C4 genomic subtypes against two existing classification systems: the PCNSL‐specific Verdin subtypes (CS1–CS4) [9] and the Schmitz DLBCL subtypes (MCD/BN2/N1/EZB) [8], which were originally developed for systemic DLBCL (Table S5). For this comparison, we constructed univariate Cox models using each classification as the sole predictor of overall survival (Figure S1). Our classification achieved a C‐index of 0.634 (95% CI: 0.539–0.730), numerically higher than that of the Verdin system (0.544, 95% CI: 0.444–0.645) and the Schmitz system (0.571, 95% CI: 0.510–0.633). However, pairwise comparisons using the compareC package did not reach statistical significance (Our vs. Verdin, p = 0.114; Our vs. Schmitz, p = 0.287; Verdin vs. Schmitz, p = 0.588), likely due to the modest sample size and wide confidence intervals. In terms of model fit, our classification yielded an AIC of 275.51, which was substantially lower (indicating better fit) than that of the PCNSL‐specific Verdin classification (AIC = 284.20; ΔAIC = 8.69). The AIC of the Schmitz classification was 275.83, nearly identical to ours (ΔAIC = 0.32).

3.2. TME Classification Based on CD8 ⁺T/M2 Macrophage Ratio Predicts Survival

Of the 68 patients with WGS data, 61 had sufficient FFPE tissue for subsequent multiplex immunofluorescence analysis of the TME. Multiplex immunofluorescence analysis revealed that tumor cells constituted 61% of ANC, while immune subsets included CD8⁺ T cells, M1‐like macrophages, and M2‐like macrophages representing 7%, 11%, and 10% of cellular composition, respectively (Figure 2B). Univariate survival analysis identified strong associations. Higher CD8⁺ T cell proportion and density and higher M1‐like macrophage density were favorable prognostic factors for progression‐free survival (PFS) (Figure 2O–R). For OS, higher CD8⁺ T cell density and M1/M2 ratio were protective, while higher M2‐like macrophage proportion and density predicted poorer OS (Figure 2S–V). Subsequently, comparative analysis was performed between tumor lesions and morphologically non‐neoplastic peritumoral regions (referred to as “para‐cancerous” for brevity) within the same biopsy specimens. Significantly higher densities of M0 macrophages (1.42‐fold, p = 0.049) and M2 macrophages (1.33‐fold, p = 0.031) were observed within tumor lesions compared with these peritumoral regions (Figure 2K–N).

FIGURE 2.

Tumor microenvironment characterization of primary central nervous system lymphoma (PCNSL). (A) Typical imaging of immune subtypes stratified by the conventional HOT/Intermediate/COLD classification using multiplex immunofluorescence. (B) Cellular composition of the TME (mean percentage of all nucleated cells [ANC], n = 61). (C–J) Comparison of immune cell proportions and densities across HOT, Intermediate, and COLD subtypes (Kruskal‐Wallis test). (K–N) Immune cell density between tumor (T) and para‐cancerous (P) tissues in total, HOT, Intermediate, and COLD subtypes (paired Wilcoxon test). (O–R) Kaplan Meier Curves of immune metrics predicting progression‐free survival (PFS). (S–V) Kaplan Meier Curve of immune metrics predicting overall survival (OS). (*p < 0.05; ***p < 0.01; ns: Not significant).

Upon applying the conventional HOT, Intermediate, and COLD stratification [17] (Figure 2A), baseline clinical characteristics did not differ significantly among groups (Table S6). The HOT group had considerably higher proportions and densities of CD8⁺ T cells, higher proportions and densities of M0‐like macrophages, higher proportions of M1‐like macrophages, and higher proportions of M2‐like macrophages than the Intermediate group (Figure 2C–J). The COLD subtype displayed lower proportion/density of CD8⁺ T cells and lower density of M2‐like macrophages than the Intermediate group (Figure 2C–J). However, no significant differences in PFS or OS were observed across conventional TME groups. The CXCL9/SPP1 ratio has been proposed as a biomarker reflecting M1/M2 functional balance in solid tumors [18]. However, this ratio was not significantly associated with PFS or OS in our study (Figure S2).

We developed a novel TME classification system based on the spatial ratio of CD8⁺ T cells to M2‐like macrophages (Figure 3C). Stratification identified three prognostic groups: High ratio (> 1.5), Intermediate ratio (0.8–1.5), and Low ratio (< 0.8). Baseline clinical characteristics did not differ significantly among groups (Table S7). Unexpectedly, the Intermediate ratio group exhibited the worst clinical outcomes, with a median PFS of only 2.0 months (Figure 3A) and a 5‐year OS rate of 10.0% (Figure 3B). This contrasted sharply with the High ratio group, where median OS was not reached, and 5‐year survival reached 75.0%. Further investigation revealed that the poor prognosis in the Intermediate group was associated with markedly elevated proportions of PD‐1⁺CD8⁺ T cells compared with the Low ratio group, suggesting possible T‐cell dysfunction (Figure S3). This finding suggested a state of immune dysfunction where numerically adequate but functionally impaired T cells failed to control tumor progression.

FIGURE 3.

CD8⁺ T/M2 macrophage ratio‐based tumor microenvironment (TME) classification predicts survival. (A) Kaplan–Meier curves for progression‐free survival (PFS) stratified by CD8⁺T/M2 ratio: High (> 1.5, n = 12), Intermediate (0.8–1.5, n = 10), Low (< 0.8, n = 39). (B) Kaplan–Meier curves for overall survival (OS) stratified by CD8⁺T/M2 ratio. (C) Representative images of tumor microenvironment (TME) groups based on CD8⁺T/M2 ratio. (D) Concordance between the new TME classification and conventional HOT/Intermediate/COLD subtypes. (E) Association of the new TME classification with Han's classification (non‐GCB versus GCB).

To evaluate whether the prognostic performance of the CD8⁺T/M2 ratio‐based TME classification was affected by optimism bias from cutoff selection (Table S8), we performed bootstrap resampling (1000 iterations) and 5‐fold cross‐validation. The bootstrap‐corrected C‐index for OS was 0.618 (95% CI: 0.534–0.698, log‐rank p = 0.037), and for PFS was 0.606 (95% CI: 0.527–0.682, log‐rank p = 0.035). Five‐fold cross‐validation yielded mean C‐indices of 0.703 (OS, p = 0.036) and 0.692 (PFS, p = 0.023). These results confirm that the CD8⁺T/M2 ratio classification retains significant prognostic discrimination after correction for overfitting.

We then matched the new TME classification with the conventional classification and found that the new TME subtypes correlated strongly with conventional subtypes: 75.0% of the High group corresponded to HOT, while 53.8% of the Low group mapped to COLD (Figure 3D). A correlation trend with Han's classification was observed, with the majority of patients in the Low (79.5%) and Intermediate (75.0%) groups belonging to the non‐GCB subtype (Figure 3E).

3.3. Integrated Genomic‐Immune Profiling Reveals a High‐Risk Subgroup

Integration of genomic and TME classifications demonstrated remarkable cross‐modality associations (Figure 4A). Strikingly, 60% of C4 genomic subtype tumors co‐occurred with the Intermediate TME phenotype compared with only 7.8% in non‐C4 tumors (Figure 4B). Although most clinical parameters showed no differential distribution across genomic or TME subgroups, a trend suggested an association between Intermediate TME and POD24 progression (90.0% vs. 58.8% in non‐Intermediate TME, p = 0.079) (Figure 4C–F).

FIGURE 4.

Integrated genomic‐immune profiling identifies a high‐risk subgroup. (A) Cross‐tabulation of genomic subtypes and CD8⁺T/M2‐based TME groups (χ² = 19.3, p = 0.014). (B) Enrichment of C4 tumors within the Intermediate tumor microenvironment (TME) group (60% vs. 7.8% in non‐C4, p < 0.001). Distribution of clinical parameters across integrated subgroups: (C) IELSG score, (D) Han's subtype, (E) Chemotherapy response, (F) POD24 status. (G) Kaplan–Meier overall survival curve comparing genomic C4 subtypes with other clusters (median overall survival [OS]: 26.0 vs. 83.0 months; hazard ratio [HR] = 7.61, p = 0.006). (H) Kaplan–Meier overall survival curve for C4 tumors with Intermediate TME cohort versus other combinations (median OS: 3.0 vs. 43.0 months; HR = 7.24, p = 0.006). note: *p < 0.05.

First, patients with C4 genomic subtype alone already exhibited significantly worse overall survival compared to non‐C4 patients (median OS 26.0 vs. 83.0 months; HR = 7.61, p = 0.006) (Figure 4G). This lethal intersection defined a biologically distinct subgroup comprising 9.8% of the cohort (n = 6). Patients with combined C4/Intermediate‐TME features suffered catastrophic outcomes, with a median OS of merely 3.0 months. This represented a 7.24‐fold increased mortality risk compared with other genomic‐TME combinations. The survival curve for this high‐risk subgroup diverged markedly within the first 6 months after diagnosis (Figure 4H). This distinct biological subset represents an urgent therapeutic priority requiring mechanistic investigation and tailored interventions.

3.4. Multivariate Analysis Confirms TME and Genomic Classifications as Independent Prognostic Factors

In the univariate survival analysis of clinical parameters for PFS, IELSG stratification, Han's classification, and chemotherapy response were strongly associated with PFS (Table 1). For OS, IELSG stratification, Han's classification, chemotherapy response, and POD24 emerged as strong predictors (Table 1).

TABLE 1.

Univariate and multivariate analyses of risk factors for survival.

Prognostic factor		No (%)	PFS			OS
			Univariate analysis	Multivariate analysis		Univariate analysis	Multivariate analysis
			p (log‐rank test)	p value (cox‐regression)	Relative risk (95% CI)	p (log‐rank test)	p value (cox‐regression)	Relative risk (95% CI)
Age	> 60	33 (48.5)	0.851			0.529
	≤ 60	35 (51.5)
ECOG	2–4	64 (94.1)	0.625			0.068		NA a
	0–1	4 (5.9)
Number of lesions	Unifocal	38 (55.9)	0.832			0.238
	Multifocal	30 (44.1)
Deep‐Involvement	Yes	35 (51.7)	0.216			0.228
	No	33 (48.5)
LDH	Elevated	8 (11.8)	0.938			0.747
	Normal	60 (88.2)
IELSG stratification	Low risk	12 (17.6)	0.018	0.149	1.00 (Reference)	0.008	0.376	1.00 (Reference)
	Middle risk	41 (60.3)		0.070	2.37 (0.93–6.02)		0.164	4.47 (0.54–36.81)
	High risk	15 (22.1)		0.066	2.57 (0.94–7.02)		0.213	3.91 (0.46–33.58)
Han's classification	Non‐GCB	48 (70.6)	0.004	0.025	2.47 (1.12–5.44)	0.032	0.110	2.23 (0.83–5.96)
	GCB	18 (26.5)
Chemotherapy response	Not Response	17 (25)	< 0.001	< 0.001	49.75 (10.36–239.03)	< 0.001	0.020	2.69 (1.17–6.19)
	Response	51 (75)
POD24	Yes	40 (58.8)	NA			< 0.001	0.020	4.01 (1.24–12.96)
	No	28 (41.2)
CD8+T/M2 ratio	> 1.5	12 (19.7)	0.005	0.015	Reference	0.0052	0.219	1.00 (Reference)
	0.8–1.5	10 (16.4)		0.004	5.59 (1.75–17.83)		0.084	3.70 (0.84–16.33)
	< 0.8	39 (63.9)		0.288	1.58 (0.68–3.64)		0.142	2.61 (0.73–9.39)
Genomic clusters	Cluster 1	25 (36.8)	0.210			0.020	0.004	1.00 (Reference)
	Cluster 2	17 (25.0)					0.022	0.28 (0.10–0.83)
	Cluster 3	14 (20.6)					0.785	0.87 (0.33–2.31)
	Cluster 4	12 (17.6)					0.035	2.76 (1.08–7.08)

Abbreviations: CI, confidence interval; ECOG: eastern cooperative oncology group; GCB: germinal center B‐cell–like; IELSG: international extranodal lymphoma study group; LDH: lactate dehydrogenase; OS: overall survival; PFS: progression‐free survival; POD24: progression of disease within 24 months.

^a ECOG performance status was not included in the final multivariate OS model because only 4 patients (5.9%) had ECOG 0–1, leading to quasi‐complete separation and an unstable hazard ratio estimate. Lasso penalized regression also eliminated ECOG at the λ.1se level, supporting its exclusion.

All variables with p < 0.1 in the univariate analysis (clinical parameters, new genomic clusters, and new TME classification) were included in the multivariate Cox proportional hazards model to identify independent prognostic factors. In the multivariable analysis for PFS, Han's subtype, lack of response to chemotherapy, and TME subtypes emerged as independent prognostic factors for shorter PFS (Table 1). For OS, lack of chemotherapy response, POD24, and new genomic subtype independently predicted inferior OS (Table 1).

3.5. Lasso Penalized Regression Identifies Robust Prognostic Factors for OS and PFS

To assess the robustness of our multivariate Cox models against overfitting, we performed Lasso Cox regression including all candidate variables that showed association with each endpoint. For OS, 11 variables were entered; for PFS, 10 variables were entered.

For OS, at the λ.min penalty level, six variables retained non‐zero positive coefficients: POD24 (coefficient 1.238), chemotherapy response (0.336), Han's classification (0.135), ECOG (0.062), genomic clusters (0.050), and age (0.018). All other candidate variables, including IELSG stratification and CD8⁺T/M2 ratio, were compressed to zero (Table S9; Figure S4A,B). At the more stringent λ.1se level, only POD24 remained with a non‐zero coefficient (0.387). This indicates that POD24 is the single most robust predictor of overall survival in our cohort, while the independent contributions of IELSG stratification and genomic clusters—though statistically significant in the conventional multivariate Cox model—are of smaller magnitude and were not retained under the strictest penalization, likely due to limited sample size (37 events) and the conservative nature of λ.1se.

For PFS (51 events), Lasso at λ.min retained seven variables with non‐zero coefficients. Chemotherapy response showed the largest positive coefficient (3.225), followed by Han's classification (0.733), CD8⁺T/M2 ratio (0.417), genomic clusters (0.055), and ECOG (0.050). Deep involvement (−0.228) and LDH (−0.003) had negative coefficients but with very small magnitudes. At λ.1se, only two variables survived penalization: chemotherapy response (coefficient 1.445) and Han's classification (0.024); the CD8⁺T/M2 ratio was eliminated (Table S10; Figure S5A,B). These results confirm chemotherapy response and Han's classification as the most robust predictors of PFS, whereas the independent prognostic value of the CD8⁺T/M2 ratio‐although significant in the conventional Cox model (HR 5.59 for Intermediate vs. High, p = 0.004)‐ was not sustained under the most conservative penalty, suggesting a moderate effect size that requires validation in larger cohorts.

Taken together, the Lasso sensitivity analyses reinforce the dominant prognostic roles of POD24 for OS and chemotherapy response plus Han's classification for PFS, while providing a more nuanced view of other predictors under penalization.

3.6. Prognostic Nomograms for Clinical Translation

To facilitate clinical translation, we constructed validated nomograms integrating significant predictors (Figure 5A,D). The PFS model incorporated Han's subtype, CD8⁺T/M2 ratio classification, and chemotherapy response, achieving a bootstrap‐validated concordance index of 0.795 (Figure 5G). The time‐dependent area under the curve (AUC) for the PFS nomogram was 0.905 at 1 year, 0.855 at 3 years, and 0.866 at 5 years (Figure 5B). The OS model incorporated POD24, genomic subtype, and chemotherapy response, with a concordance index of 0.781 (Figure 5H). The corresponding AUCs for the OS nomogram were 0.889 at 1 year, 0.806 at 3 years, and 0.894 at 5 years (Figure 5E). Calibration plots showed acceptable agreement between predicted and observed probabilities at 1, 3, and 5 years (Figure 5C,F). However, these performance metrics should be interpreted cautiously, as they were derived from the same dataset used to develop the nomograms. External validation in independent cohorts is required before any clinical application can be considered.

FIGURE 5.

Prognostic nomograms and validation for progression‐free survival (PFS) and overall survival (OS). (A) Nomogram for predicting 1‐, 3‐, and 5‐year PFS rates incorporating independent predictors: Han's subtype, CD8⁺T/M2 macrophage ratio, and chemotherapy response. (B) Time‐dependent receiver operating characteristic (ROC) curves for the PFS nomogram (1‐/3‐/5‐year area under the curve [AUC]: 0.905/0.855/0.866). (C) Calibration plot for the PFS nomogram (bootstrap = 1000 resamples). (D) Nomogram for predicting 1‐, 3‐, and 5‐year overall survival (OS) incorporating predictors: Presence of progressive disease at 24 months (POD24), genomic subtype, and chemotherapy response. (E) Time‐dependent ROC curves for the OS nomogram (1‐/3‐/5‐year AUC: 0.889/0.806/0.894). (F) Calibration plot for the OS nomogram. (G) Time‐dependent, Bootstrap‐validated C‐index for PFS. (H) Time‐dependent, Bootstrap‐validated C‐index for OS. Note: These nomograms are proof‐of‐concept and require external validation.

4. Discussion

This integrated genomic and TME study establishes a novel classification framework that significantly advances risk stratification in PCNSL. We define four molecular subtypes with distinct therapeutic vulnerabilities and identify a CD8⁺T/M2 macrophage ratio‐based TME classification that unexpectedly reveals intermediate immune infiltration as the poorest prognostic state. Critically, the integration of these systems uncovers a lethal C4/Intermediate‐TME subgroup representing nearly 10% of patients with PCNSL who experience catastrophic outcomes with current therapies.

While the Schmitz and Verdin classifications have advanced our understanding of DLBCL and PCNSL biology, their prognostic performance in our PCNSL cohort was limited (Figure S1). A formal comparison of discrimination and model fit showed that our C1‐C4 system performed numerically better than the PCNSL‐specific Verdin system (C‐index: 0.634 vs. 0.544; AIC: 275.5 vs. 284.2). The improvement in AIC (ΔAIC = 8.69) indicates a meaningfully better model fit. In our study, we established four genomically distinct PCNSL subtypes through integrated WGS profiling. Based on literature from other cancer types, these subtypes demonstrate clinically actionable divergence (Figure 6). C1 tumors (DNA repair defects) might be considered for future preclinical evaluation of PARP inhibitors [19], given the presence of 3p/3q amplifications. C2 tumors (apoptosis‐dysregulated) harbor BCL2/IRF4 alterations, raising the hypothesis that BCL2 inhibitors [20] or immunomodulatory agents [21] could be explored in model systems. For C3 tumors (BCR‐MAPK hyperactivated), CD79A/CARD11 mutations provide a biological rationale for testing BTK inhibitors [22, 23] in PCNSL. Finally, C4 tumors (epigenetic/MMR‐deficient) with KMT2A/TET2 alterations might be candidate for hypomethylating agents [24, 25, 26], and their mismatch repair deficiency suggests a potential susceptibility to PD‐1 blockade [27] a hypothesis that requires dedicated investigation in PCNSL. Critically, the poor prognosis of the C4 subtype (median OS 26.0 months) and its enrichment in the lethal C4‐Intermediate TME intersection (median OS 3.0 months) necessitate subtype‐specific therapeutic escalation. This classification thus transcends descriptive taxonomy toward a clinically actionable framework for precision trials.

FIGURE 6.

Implications of the PCNSL genomic subtypes for pathogenesis and therapy.

TME dynamics play a crucial role in PCNSL progression and chemotherapy response [28]. In PCNSL, TAMs and T‐lymphocytes dominate the microenvironment [12]. Current TME classifications of PCNSL predominantly rely on immune cell density thresholds (e.g., HOT/Intermediate/COLD) [13, 17] or polarization markers (e.g., M1/M2 ratios [29], CXCL9/SPP1 balance [18]), metrics that failed to predict survival in our cohort. These limitations motivated our CD8⁺T/M2 ratio‐based stratification, which uncovered a non‐linear prognostic relationship: the Intermediate group (CD8⁺T/M2 = 0.8–1.5) exhibited the worst outcomes (median OS 31 months, 5‐year OS rate of 10.0%), inferior even to the Low group (CD8⁺T/M2 < 0.8; median OS 35 months, 5‐year OS rate of 42.7%). We attribute this to a “perfect storm” of immune dysfunction. First, CD8⁺ T cell dysfunction: Intermediate tumors harbored markedly elevated PD‐1⁺CD8⁺ T cells, which were functionally impaired effectors unable to control tumor progression despite numerical adequacy [12]. Second, M2 macrophage‐mediated suppression: The high M2 density in the Intermediate group (mean 452 cells/mm²) secreted IL‐10/TGF‐β that paralyzed CD8⁺ T cytotoxicity [30]. This “immunosuppressive equipoise” allowed M2 to inhibit CD8⁺ T cells, yet the CD8⁺ T‐cell response remained insufficient to overcome suppression [31, 32], explaining why profoundly immune‐deserted (Low group) tumors outperformed Intermediate ones. Critically, this ratio has direct implications for PD‐1 blockade. The High group (CD8⁺T/M2 > 1.5) represents an “inflamed” PCNSL phenotype with abundant PD‐1⁺CD8⁺ T cells, potentially representing suitable candidates for PD‐1/PD‐L1 monotherapy. In contrast, the Intermediate group may require combination strategies, as PD‐1 blockade alone might not be sufficient to reverse the suppressive effects of M2 macrophages, and the elevated PD‐1⁺CD8⁺ T cells in this group may not represent a fully reversible exhaustion state. Our data support the addition of CSF1R inhibitors to remodel the TME [33]. The Low group may require priming with epigenetic modulators (e.g., azacitidine in C4 tumors) before immunotherapy [24, 25, 26].

In PCNSL, traditional prognostic models such as the IELSG and MSKCC scores have been used to predict outcomes. However, their accuracy decreases in the current era of immunotherapy [5]. As a result, there is increasing interest in identifying alternative prognostic markers that can offer a more comprehensive and accurate prediction of patient outcomes [34, 35]. One such marker is POD24, which has demonstrated significant prognostic implications in various lymphoma subtypes. In DLBCL, patients with POD24 had a 5‐year OS rate of 19%, compared with 87.6% in those without POD24 [36]. Similarly, in follicular lymphoma, POD24 was associated with poorer prognosis, with a 5‐year OS rate of 50% in patients with POD24, compared with 90% in those without POD24 [37]. In PCNSL, a recent study reported that patients with POD24 had a markedly lower 5‐year OS rate compared with those who did not progress early (25% vs. 96.7%) [38]. Our study confirmed that POD24 is a strong and independent poor prognostic factor for OS in patients with PCNSL, with a 5‐year OS rate of 21.1% versus 85.7%. Our study further developed a nomogram model that incorporates POD24, providing a quantitative tool for predicting 1‐, 3‐, and 5‐year OS rates in PCNSL. The model's high concordance indices (C‐index) further validate its clinical utility. This underscores the importance of incorporating dynamic disease progression metrics like POD24, which clinical parameters alone cannot address, into clinical decision‐making.

Our study defines a molecularly distinct high‐risk PCNSL subgroup characterized by concurrent C4 genomic alterations and intermediate TME features, which experienced catastrophic outcomes, with a median OS of only 3.0 months. This biologically aggressive variant demonstrates converging epigenetic instability and immunosuppressive pathophysiology that collectively drive therapeutic resistance. The profoundly poor prognosis emerges from synergistic interactions between C4‐defining molecular lesions and TME dysfunction. Mutations in epigenetic regulators including KMT2A (histone‐modifying enzyme) and TET2 (DNA demethylase) disrupt chromatin architecture and transcriptional programs, facilitating immune evasion through impaired antigen presentation machinery [39, 40]. Concurrent mismatch repair defects via MSH6 inactivation accelerate genomic instability while depleting immunogenic neoantigens, creating dual barriers to checkpoint immunotherapy [41, 42]. These genomic aberrations intersect with the dysfunctional intermediate TME state in which PD‐1‐mediated T‐cell exhaustion coincides with M2 macrophage polarization, establishing self‐reinforcing immunosuppressive circuits that paralyze antitumor immunity [43, 44, 45]. To overcome this therapeutic challenge, we propose a mechanistically grounded dual‐pathway intervention strategy. First, epigenetic priming using hypomethylating agents such as azacitidine may reverse KMT2A/TET2‐mediated transcriptional silencing and restore tumor immunogenicity [25, 26]. Second, combined CSF1R inhibition and PD‐1 blockade could simultaneously deplete immunosuppressive macrophages and reinvigorate exhausted T cells [33]. Early‐phase trials in Hodgkin lymphoma demonstrate enhanced response rates when epigenetic modulators precede checkpoint immunotherapy, supporting sequential administration in this high‐risk subgroup [46]. Future clinical investigations should prioritize validating this combinatorial approach specifically for patients with C4/Intermediate‐TME PCNSL.

Some limitations warrant consideration, including the single‐center retrospective design and the limited sample size for subgroup analyses. External validation of the classification framework in independent, multicenter cohorts is therefore essential. Future investigations should elucidate the molecular mechanisms underlying the convergence of the C4 genomic subtype and the Intermediate TME phenotype and evaluate the efficacy of subtype‐specific therapeutic strategies identified in this study. In the Lasso sensitivity analysis, the λ.1se criterion retained only POD24 as a predictor, whereas IELSG stratification and genomic clusters – both significant in the conventional Cox model – were compressed to zero. This discrepancy is expected given the small sample size (37 events) and the conservative nature of λ.1se, which tends to produce overly sparse models. We therefore place greater weight on the λ.min results and the conventional Cox model, which are supported by clinical plausibility and prior literature. In the Lasso sensitivity analysis for PFS, the CD8⁺T/M2 ratio‐ an independent predictor in the conventional Cox model – was not retained at the λ.1se level, whereas chemotherapy response and Han's classification were. This discrepancy likely reflects the moderate effect size of the CD8⁺T/M2 ratio relative to the dominant predictors, combined with the conservative nature of λ.1se in small‐to‐moderate sample sizes. Therefore, we interpret the Lasso results as confirming the robustness of chemotherapy response and Han's classification for PFS, while the role of CD8⁺T/M2 ratio, though statistically significant, requires external validation. Furthermore, we used PD‐1 as a single marker for T‐cell dysfunction due to limited tissue availability. Definitive characterization of T‐cell exhaustion would require co‐staining for TOX, TIM‐3, or LAG‐3, which was not feasible in this retrospective cohort. Therefore, our findings regarding PD‐1⁺CD8⁺ T cells should be interpreted as suggestive of possible dysfunction rather than conclusive evidence of exhaustion. Finally, the nomograms presented in this study are proof‐of‐concept tools. Although we performed internal bootstrap validation to correct for optimism, this does not replace external validation. The reported C‐indices and calibration are likely optimistic, and the nomograms should not be used for clinical decision‐making until they have been validated in independent PCNSL cohorts. Our primary goal is to provide a hypothesis‐generating framework to guide future prognostic model development.

In conclusion, our study introduces an integrated genomic and TME classification system that enhances prognostic stratification in PCNSL. The integration of genomic subtyping and TME evaluation offers a hypothesis‐generating framework for future investigations. Prospective external validation studies are urgently needed to determine whether the proposed nomograms and risk stratification can be translated into clinical practice.

Author Contributions

Wenbin Qian, Yun Liang, Shumei Wei: conceptualization. Yurong Huang, Teng Yu, Qunyi Guo, Xueli Jin: methodology. Aiqi Zhao, Wen Lei, Xian Li, Baizhou Li: investigation. Chaoyi Li, Shanshan Guo: visualization. Wenbin Qian, Yun Liang: supervision. Xianggui Yuan, Qian Luo: writing – original draft. Wenbin Qian, Yun Liang, Shumei Wei: writing – review and editing.

Funding

Noncommunicable Chronic Diseases‐National Science and Technology Major Project (No. 2023ZD0501300, WBQ); National Natural Science Foundation of China (No. 82350104 and 82400217, WBQ); and Natural Science Foundation of Zhejiang Province of China (No. BY24ZH080013, XGY).

Ethics Statement

The study received ethical approval from the Institutional Review Board of Zhejiang University (Approval No. 2020–568).

Consent

Written informed consent was obtained from all participants specifically authorizing the use of their biomaterials for future biomarker research.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Validation and Prognostic Relevance of Existing Molecular Classifications in the PCNSL Cohort.

Figure S2: CXCL9/SPP1 Characterization in Macrophages.

Figure S3: Immune Contexture Underlying Inferior Prognosis of the Intermediate CD8⁺T/M2 Ratio Group.

Figure S4: Lasso Cox regression analysis for overall survival (A) Cross‐validation error (partial likelihood deviance) versus log(λ).

(B) Lasso coefficient trajectories for the 11 candidate variables as a function of log(λ). POD24 shows the largest coefficient (1.238), followed by chemotherapy response (0.336), Han's classification (0.135), ECOG (0.062), genomic clusters (0.050), and age (0.018).

Figure S5: Lasso Cox regression analysis for progression‐free survival.

Table S1: Characteristics of patients with PCNSL.

Table S2: Genomic features in consensus clustering analysis.

Table S3: Multiplex Immunofluorescence Staining Materials and Conditions for Antibody Panel in the PCNSL Study.

Table S4: Clinical characteristics across novel genomic subtypes.

Table S5: Comparison of C‐index and AIC among genomic classification systems.

Table S6: Clinical characteristics across conventional HOT/Intermediate/COLD stratification.

Table S7: Clinical characteristics of patients across CD8+/M2 ratio groups.

Table S8: Internal validation of CD8⁺T/M2 ratio‐based TME classification.

Table S9: Lasso Cox regression coefficients for OS.

Table S10: Lasso Cox regression coefficients for PFS.

Data Availability Statement

All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Raw sequencing data have been deposited in the National Genomics Data Center (NGDC), China (No. PRJCA009559).