Mid-treatment CSF ctdna and MYD88 clearance outperform PET-CT in predicting response and survival to orelabrutinib-based induction in newly diagnosed PCNSL: A prospective biomarker study

Abstract Background Primary central nervous system lymphoma (PCNSL) is an aggressive extranodal non-Hodgkin lymphoma with limited tools for dynamic treatment response monitoring. ^18F-FDG PET-CT, while commonly used in systemic lymphomas, offers suboptimal predictive value in PCNSL due to CNS-specific limitations. Cerebrospinal fluid (CSF)-derived circulating tumor DNA (ctDNA) offers a promising, noninvasive approach to assess molecular response, but its clinical utility in PCNSL remains underexplored. Methods We prospectively enrolled 22 newly diagnosed PCNSL patients (2021–2024), treated with six cycles of orelabrutinib, rituximab, and high-dose methotrexate (ORM). Tumor and matched baseline CSF samples were sequenced; follow-up CSF was collected at cycle 5 day 1 (C5D1). Somatic mutations, variant allele frequencies (VAFs), and molecular clearance were analyzed. Treatment responses were evaluated per international PCNSL criteria. Interim PET-CT was performed after four cycles. Correlations between molecular response and clinical endpoints—including progression-free survival (PFS), overall survival (OS), and objective response rate (ORR)—were compared to PET-CT findings. Results Median age was 64 (range 41–79); 63.6% had non-GCB subtype, and 90.9% deep parenchymal involvement. ORR after four cycles was 86.4%, including 40.9% CR and 45.5% PR. Post-cycle 6, ORR was 89.5%, with 78.9% CR. Median time to response was 2.6 months; 2-year DoR 72.4%. Median follow-up was 18.9 months; estimated 2-year PFS and OS were 62.5% and 75.2%, respectively. Treatment was generally well tolerated; grade ≥3 adverse events occurred in 22.7%, mostly infections. No treatment-related deaths or therapy discontinuations were reported. Somatic mutations or CNAs were detected in 100% of tumor tissues and 95% of baseline CSF samples. A total of 93 mutations (50.5%) were shared between tumor and CSF, with an additional 43 detected exclusively in CSF. Frequent mutations included MYD88 (84.2% in CSF), PIM1, KMT2D, CD79B, and CARD11. Tumor tissue VAFs were significantly higher than those in CSF (p<0.001). Lower baseline VAF and significant VAF reduction at C5D1 correlated with CR (p=0.020 and p<0.001, respectively). Notably, MYD88 clearance at C5D1 strongly predicted CR (p=0.004) with an AUC of 0.900. Patients with KMT2D mutations had a trend toward lower response rates (83.3% in PR/SD vs. 33.3% in CR, p=0.063). Among 18 patients with paired baseline and C5D1 CSF samples, ctDNA clearance was significantly associated with response. Patients achieving ctDNA negativity at C5D1 had a CR rate of 92.9% (13/14) vs. 20.0% (1/5) for ctDNA-positive cases (p=0.006). Dynamic ctDNA clearance between baseline and C5D1 further predicted CR (92.3% vs. 20.0%, p=0.008), with a predictive AUC of 0.862. Compared to PET-CT, ctDNA clearance demonstrated superior sensitivity (92.3% vs. 53.8%) and negative predictive value (80.0% vs. 45.5%). ctDNA response preceded radiographic response by approximately 8 weeks. Survival analyses showed significantly longer PFS among patients achieving CR vs. non-CR (p=0.009), and among those with objective response (CR/PR) vs. non-responders (p=0.002). OS was also improved in responders (p=0.048). Both ctDNA (p=0.024) and MYD88 (p=0.014) clearance at C5D1 were associated with longer PFS, with ROC AUCs of 0.740 and 0.786, respectively. Multivariate Cox regression confirmed MYD88 clearance as an independent predictor of PFS. In contrast, PET-CT response was not significantly associated with either PFS or OS. ConclusionORM induction is effective and well tolerated in newly diagnosed PCNSL. Early CSF ctDNA and MYD88 clearance at mid-treatment offer superior predictive and prognostic value compared to PET-CT. These molecular biomarkers enable real-time risk stratification and may guide early treatment adaptation. Incorporating CSF ctDNA monitoring into clinical workflows has the potential to improve response prediction and long-term outcomes in PCNSL.