Molecular clustering and fragmentomics on circulating tumor DNA improve outcome prediction in diffuse large B cell lymphoma

Diffuse large B-cell lymphoma (DLBCL) can be classified into molecular clusters that reflect the genetic and clinical heterogeneity of the disease. These subtypes have been identified on tissue biopsy that might not accurately capture the anatomical variability of DLBCL. The analysis of ctDNA in DLBCL is an accurate and non-invasive approach to retrieve molecular information not otherwise detectable in tissue biopsies. Moreover, the quantification of baseline ctDNA stratifies DLBCL outcomes, but its integration with molecular clustering has not been evaluated yet. The aims of the study were to evaluate the role of ctDNA in detecting DLBCL molecular clusters and to assess whether molecular clustering combined with ctDNA levels may improve outcome prediction. A cohort of 166 newly diagnosed DLBCL patients treated with R- CHOP and provided with ctDNA from plasma and germline gDNA from granulocytes (N=77/166 also with gDNA from LNF biopsies), were enrolled in the study. The LyV4.0 CAPP-seq assay consisting of a panel of 59 genes relevant to DLBCL and the LymphGen clustering tool were used. Clustering analysis allowed to classify 42.8% of patients based on ctDNA mutational profile. More precisely, 13.3% patients were classified as EZB, 11.4% as BN2, 8.4% as MCD, 6.0% as ST2 and 3.6% as A53. MCD patients exhibited a shorter PFS compared to other patients (4-years PFS: 40% vs 66.6%, p=0.032). The LymphGen tool assigned 46.5% of patients on the tissue biopsy. The concordance rate between ctDNA and LNF molecular clustering was 95.8%. Clusters ST2 and BN2 were associated with an excellent outcome after R-CHOP. Patients with ctDNA load< 2.5log10hGE and identified as ST2/BN2 showed superior outcome compared to patients with ctDNA load >2.5log10hGE and not assigned to ST2/BN2 clusters (4-years PFS and OS: 77.4% and 94.4% vs 39.9% and 46.5%, p=0.017 and p=0.002, respectively). Compared to ctDNA levels only, the addition of BN2/ST2 cluster improved the C statistics of the model (0.62 vs 0.59 for PFS and 0.68 vs 0.63 for OS). The assignment to cluster BN2/ST2 maintained an independent association with an improved PFS (HR 0.37, 95% CI 0.14- 0.95, p=0.04) and OS (HR 0.07, 95% CI 0.01-0.55, p=0.01) when adjusted for ctDNA levels, IPI and COO. Finally, patients with no detectable mutations on ctDNA or not assigned to a specific cluster, were stratified based on their fragmentome. Subnucleosomal group had less favourable outcome compared to nucleosomal group (4-years PFS and OS: 25.9% and 44.4% vs 76.9% and 87.3%; both p<0.001). The prognostic impact of subnucleosomal cfDNA was maintained in multivariate analysis both for PFS (HR 3.12, 95% CI 1.36-7.15, p=0.007) and OS (HR 3.06, 95% CI 1.17-8.02, p=0.023) when adjusted for ctDNA levels.