Molecular features encoded in the ctDNA reveal heterogeneity and predict outcome in high-risk aggressive B-cell lymphoma

Genomic interrogation of circulating tumor DNA (ctDNA) provides clinically applicable and minimally invasive tools to profile cancers, including B-cell lymphomas.^1-3 Large B-cell lymphomas (LBCLs) are a heterogeneous group of cancers, of which diffuse large B-cell lymphoma (DLBCL) is the most common entity.⁴ In response to rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone (R-CHOP), 30% to 40% of the DLBCLs progress, resulting in high risk of death.⁵ Despite improved insights to decompose the biological and clinical complexity of DLBCL,^6-10 neither subtype-targeted^11,12 nor risk-adapted¹³ interventions have been established to be superior to R-CHOP, and the early identification of refractory patients remains inaccurate.^14,15 Inadequate dismantling of clinical and biological heterogeneity remains the main obstacle for improving outcomes in response to novel therapies, and there is a need for new clinically applicable tools to stratify patients with DLBCL more accurately. Here, precision medicine applications utilizing ctDNA are anticipated.

Cell-free DNA (cfDNA) has tremendous clinical potential as a liquid biopsy.^16,17 Besides opening avenues for molecularly targeted treatments,^18,19 the burden and kinetics measured in ctDNA concentration are of great interest for their clinical value. Pretreatment ctDNA burden has been reported as a surrogate for tumor burden with prognostic information in DLBCL.^19-21 Furthermore, early detection of refractoriness with dynamic measurements of ctDNA during therapy could lay the foundations for molecular response-adapted treatment decisions.^2,3,22-24 Albeit compelling, these early results have been obtained from heterogeneous patient cohorts treated with diverse therapies and utilizing plasma and/or serum samples.^{2,18,19,24,25} To date, the clinical value of ctDNA-based analyses has not been evaluated in the context of clinical trial patients with aggressive B-cell lymphoma and uniform therapy.

Despite the plethora of somatic events in the B-cell lymphoma genome, most of the ctDNA does not contain mutational information. Another approach that has proven to be useful in solid tumors is based on the analysis of cfDNA fragmentation patterns.^26,27 Upon its release mainly from apoptotic cells, cfDNA is fragmented, and nucleosomes, transcription factors, and other DNA-binding proteins prevent this cleavage, resulting in specific patterns of fragmentation. These fragmentation profiles reflect chromatin proteome occupancy maps and epigenetic fingerprints that can be used to detect and determine molecular profiles of cancer.^26,28,29 So far, these fragmentation landscapes and their potential clinical utility have remained unexplored in B-cell lymphomas.

Here, we have dissected the molecular determinants within the ctDNA that resolve heterogeneity and characterize trial patients with poor outcomes. We show how pretreatment ctDNA reflects the ecosystem of clinical disease and lymphoma biology that extends beyond conventional measures and tumor biopsies. We describe and use ctDNA hypermutation patterns and fragmentation disparities to show how complementary dynamic analysis of ctDNA accurately determines treatment responses and predicts survival. Taken together, we uncover novel features in the lymphoma ctDNA that we anticipate resulting in novel tools to noninvasively guide future treatment decisions.

The correlative study included 101 patients (aged 18 to 65) with high-risk (age-adjusted International Prognostic Index [aaIPI 2-3] and/or site-specific risk factors for central nervous system [CNS] recurrence) primary LBCL treated uniformly in a Nordic LBC-05 phase 2 trial (NCT01325194) (supplemental Table 1).³⁰ Plasma samples were collected according to the study protocol prior to therapy and at response evaluations (after 3 cycles and posttherapy) (Figure 1A). Additionally, 4 relapse plasma samples and 9 healthy controls were included. Diagnostic tumor tissue was available for sequencing from 39 patients. Matched whole-blood samples were used as germline controls for variant calling.

A custom capture-based gene panel targeting the most common lymphoma driver genes, regions of somatic hypermutations, and immunoglobulin loci was used to detect ctDNA (supplemental Table 2). Target-enriched libraries from patients were sequenced to a median deduplicated depth 1557× with HiSeq2500 (Illumina) (supplemental Figure 1A-B). Additionally, whole-genome sequencing (WGS) (depth ∼10×) was performed for 8 overlapping pretreatment cfDNA libraries on the Novaseq6000 (Illumina) instrument.

ctDNA was quantified in log10-transformed haploid genome equivalents per mL of plasma (log10 hGE/mL), assuming 1 haploid genome weighs 3.3 picograms. The presence of ctDNA or minimal residual disease (MRD) was evaluated in tumor genotyped pretreatment samples and follow-up samples using an in-house developed R-package ‘ctDNAtools’ that utilizes an established Monte Carlo framework³¹ for testing MRD incorporating variant phasing patterns.³² MRD test was calibrated to 95% specificity using withheld healthy control cfDNA samples, and ctDNA detection was based on the MRD test positivity. Molecular responses were measured as log10-transformed fold-changes in ctDNA quantities between consecutive samples. Fragment length data were extracted from aligned sequencing data using the ‘ctDNAtools’ R-package. Detailed methodology is provided in the supplemental Methods.

Data can be found under accession number EGAS00001005835.

Statistical details accompany presented results in text, figure legends, figures, and supplemental results. P values < .05 were considered significant and significance of is indicated in figures as follows: ns, P ≥ .05; *P < .05; **P < .01; ***P < .001; ****P < .0001.

We detected ctDNA in the pretreatment plasma samples in 97 of 100 (97%) patients (Figure 1A-D; supplemental Figure 1; supplemental Tables 1-4). The quantified levels were comparable with the high-risk patient demographics reported previously in DLBCL^19,24 and similar between different countries (Figure 1E).

Elevated pretreatment ctDNA burden was associated with high metabolic tumor volume (MTV) and total lesion glycolysis (Figure 1F; supplemental Figure 2A). aaIPI and CNS-IPI were positively associated with ctDNA levels, but the levels did not differ between the patients with higher risk scores (Figure 1G; supplemental Figure 2B). Advanced stage, elevated lactate dehydrogenase and bone marrow involvement were associated with higher ctDNA levels, whereas age, poor performance status, bulky disease, or the presence of B-symptoms at diagnosis were not ( supplemental Figure 2C-I). The patients whose lymphoma was diagnosed from a needle biopsy had higher ctDNA levels than the patients diagnosed from a surgical biopsy (supplemental Figure 2J). ctDNA levels were similar between different histopathological entities and molecular subtypes (supplemental Figure 3A-D). Taken together, pretreatment ctDNA burden represents an accurate measure of tumor burden that reveals substantial heterogeneity in the trial patients with predefined high-risk inclusion criteria.

High pretreatment ctDNA burden translated to poor outcome on all measured endpoints (Figure 1H-I; supplemental Figure 4A-B; supplemental Table 5). Clinical risk factors were not associated with outcome, whereas the dismal impact of high ctDNA burden was evident also among the patients with advanced-stage or elevated aaIPI (Figure 1J; supplemental Figure 4C; supplemental Table 5). In multivariable analyses, the survival impact was independent of MTV, aaIPI, and CNS-IPI (Figure 1K; supplemental Figure 4D; supplemental Table 5). Together, these findings confirm the clinical significance of pretreatment ctDNA burden that extends beyond the established clinical risk groups.

Although high ctDNA level was a strong predictor for overall survival (OS), treatment failures were also recorded in the patients with low ctDNA burden (Figure 1L). However, while the fraction of lymphoma-related failures regressed per decreasing pretreatment ctDNA quartile, nonlymphoma related failures were more common among the patients with lower burden (supplemental Figure 5). The accumulation of nonlymphoma-related failures, including 3 treatment-related acute myeloid leukemias in the patients with below-median pretreatment ctDNA levels, suggests that dose deescalation could provide a clinical benefit for this subgroup.

To understand how treatment responses and clinical challenges are reflected by ctDNA kinetics during a uniform therapy, we interrogated the plasma samples collected at response evaluations for the presence of MRD (Figure 2A; supplemental Figure 6A-E). We compared the results with computer tomography (CT) and positron emission tomography (PET)-CT-based responses and outcomes.

After 3 therapy cycles, 36% (21 of 58) of the samples were positive for MRD (MRD^mid+) (Figure 2A-B; supplemental Figure 7A). Patients with MRD^mid+ were characterized by a lower rate of complete responses (CR) according to midstaging CT in comparison with MRD-negative cases (MRD^mid-) (Figure 2C). MRD^mid- translated to a favorable trend toward better survival, whereas CT-based CR did not (Figure 2D; supplemental Figure 7B-F). Notably, refractory patients were characterized by MRD^mid+ and elevated ctDNA levels at midstaging (≥median) (supplemental Figure 7G). Although limited by sample size, these data suggest that ctDNA evaluation at midstaging could be a feasible tool to identify patients with favorable responses and outcomes.

At the end-of-therapy response evaluation, ctDNA was detected in 14% (10 of 71, MRD^end+) of the patients with extremely poor outcomes, and primary refractory diseases were characterized by a rebound in ctDNA concentration (Figure 2E-F; supplemental Figure 8A-E). In contrast, only 5 of the 18 patients with PET positivity experienced progression (PET+, Deauville score = 4 to 5), among whom MRD^end+ could predict 4 of the 5 progressions with 100% specificity (Figure 2G; supplemental Figure 8F-H). Additionally, 10 patients underwent confirmatory biopsies from their ambiguous fluorodeoxyglucose-avid lesions, among whom MRD positivity could capture the only patient with biopsy-confirmed progression (9 tissue biopsies negative for viable lymphoma) (Figure 2H). Taken together, these data confirm the prognostic impact of MRD measurements and demonstrate how end-of-therapy examination of ctDNA could be used to complement the poor specificity of PET-CT in response evaluation.

Measuring ctDNA quantities, kinetics, and MRD relies on the detection of somatic mutations. To explore mutation-independent translational utilities of ctDNA, we dissected the fragmentation landscapes captured by our study. First, by comparing the fragment profiles captured by WGS and our targeted approach in 8 overlapping pretreatment plasma samples, we confirmed that target enrichment does not distort fragment length patterns that depend on the genomic regions analyzed (supplemental Figure 9).

Pooled analysis of the pretreatment fragmentation data revealed that fragment size distributions differed between the fragments bearing mutant and reference sequences (Figure 3A). Mutated fragments were, in general, shorter, favoring submononucleosomal and subdinucleosomal lengths, and they were more enriched for dinucleosomal lengths than their reference counterparts (Figure 3A; supplemental Figure 10A). As expected, within a few patients with exceptionally high variant allele frequencies, the fragment profiles of mutated and wild-type (WT) fragments were nearly identical, as most of their cell-free DNA was lymphoma-derived (supplemental Figure 10B). Similar differences were detected between healthy controls and patient samples at diagnosis (supplemental Figure 10C). Together these data show that lymphoma ctDNA and other cfDNA fragments in plasma differ for their composition of fragment lengths.

Further exploration of the pretreatment samples revealed that there were also differences in the complete and mutated fragmentation patterns between the patients (Figure 3B, supplemental Figure 10D). Intriguingly, some pretreatment fragmentation patterns were characterized by a more prominent dinucleosomal abundance, whereas the others had a more prominent shift toward submono- and subdinucleosomal lengths (Figure 3B; Examples I and II). A small group of pretreatment samples was distinguished by an evident bifurcation of the mononucleosomal fragments that reflected the very short nature and high abundance of ctDNA fragments in these samples (Figure 3B; Example III). These findings establish that fragmentation patterns also differ between lymphoma patients.

To understand the disparities in fragmentation landscapes, we ran a principal component analysis (PCA) of the fragmentation profiles of all the cfDNA samples in our study (n = 242). Decomposition of fragmentation patterns revealed that the samples from different treatment phases differed for the major components in the PCA (Figure 3C; supplemental Figure 11A). Interestingly, we observed that the end-of-therapy cfDNA profiles of cured and relapsing patients differed for the major principal components (Figure 3D; supplemental Figure 11B-D). This led us to hypothesize that end-of-therapy fragmentation patterns contain information that could be used for mutation-free relapse prediction.

To pursue a predictor of progression by utilizing end-of-therapy fragmentation disparities, we extracted features for machine learning-based classifier from fragment density histograms in fragment size windows identified by their correlation with major principal components of the PCA (Figure 3E; supplemental Table 6). The classifier estimate for recurrence probability was in general concordant with the mutation-based MRD approach but disagreed for some patients, and in 5-fold crossvalidation, the classifier performance was even better than the MRD-based prediction (Figure 3F-G; supplemental Figure 11E-F). The independent methods complemented each other and together predicted the majority of the failures (Figure 3H; supplemental Figure 11G). Importantly, the fragment classifier could augment prediction in samples with borderline negative MRD results and PET positivity (supplemental Figure 11G-H). As a proof-of-principle, our discoveries demonstrate that mutation-independent means based on fragmentome disparities can complement mutation-based MRD detection in predicting survival.

Having learned how quantitative, kinetic, and structural features in ctDNA predict outcome in LBCL, we next sought to scrutinize the genetic content of ctDNA and its translational impact. First, we compared the mutational repertoires between concurrent liquid and tumor tissue biopsies at diagnosis (Figure 4A). The number of recorded mutations between tumor-educated and tumor-naive cases did not differ, and most mutations recorded in tumor tissue were also detected in plasma (Figure 4A; Case#1; supplemental Figure 12A-B). In some patients, however, we observed that the diagnostic biopsy represented a subclone diverged from the major systemic clone in the ctDNA (Figure 4A; Case#2). Importantly, in some relapsing patients, these site-specific or systemically subclonal mutations were not present in the subsequent relapse ctDNA or tissues (Figure 4B; supplemental Figure 12C-D). These data underline that a tissue biopsy may not mutation-wise represent the most clinically significant clone, which potentially compromises MRD measurements relying only on tumor-educated reporter mutations.

We next examined the driver gene landscape in the pretreatment ctDNA. As expected, the overall catalog of mutated driver genes in plasma represented the heterogeneous mutational landscape described in LBCLs (Figure 4C). We observed that high ctDNA burden arose from heterogeneous backgrounds, yet we recorded a striking association between TP53 mutations and high ctDNA levels and MTV (Figure 4D; supplemental Figure 12E). Consequently, TP53 mutations were identified as a surrogate for poor OS (Figure 4E). This finding reveals that some established biomarkers for dismal outcomes may interact with tumor burden at diagnosis.

To investigate phenotyping properties encoded in ctDNA, we studied the recorded mutational landscape for underlying patterns of driver genes and mutational signatures. Explorative analysis reproduced MCD/C5-like and EZB/C3-like DLBCL subgroups, and interestingly, distinguished 2 clusters enriched for T-cell/histiocyte-rich and primary mediastinal B-cell lymphomas, suggesting that genomic distinction of subtypes and histological entities can be achieved noninvasively (supplemental Figure 13A). Subsequently, we interrogated activation-induced cytidine deaminase (AID)-related hypermutation signatures that we have previously found to distinguish activated B-cell (ABC)- and germinal center B-cell (GCB)-like DLBCLs,³³ and confirmed a striking difference in the abundance of RCH and TW signatures between these subtypes (Figure 4F; supplemental Figure 13B-E). As expected, ABC-related genomic drivers CD79B, PIM1, and GRHPR were enriched in lymphomas with RCH-enriched signature (henceforth post-GC signature), whereas G-protein signaling-related GCB drivers GNA13 and GNAI2 occurred in TW-enriched lymphomas, demonstrating that hypermutation signatures can be used to complement driver gene-based assessment of cell-of-origin (COO) noninvasively (Figure 4G). Importantly, the post-GC signature identified DLBCL patients with inferior survival, while the subtyping based on immunohistochemistry, digital gene expression, or mutational clustering did not reach statistical significance (supplemental Figure 13F). Furthermore, the post-GC signature and its surrogate, CD79B mutations, recognized patients with poor outcomes independent of pretreatment ctDNA burden, suggesting that accurate noninvasive phenotyping can provide COO-related additional prognostic information to tumor burden at diagnosis (Figure 4H; supplemental Figure 13G-H).

To continue learning from hypermutable patterns in ctDNA, we next dissected the landscape of phased variants that are linked somatic mutations in a shared haplotype (ie, same allele) of cancer cells. Overall, we recorded phased variants in 96 of 99 (97%) of the patients most commonly affecting well-established targets of AID in the immunoglobulin loci and BCL6 gene untranslated regions³⁴ (Figure 4I). Phased variants in other hypermutable genes BCL2, SOCS1, and PIM1 were common, and they occurred in a mutually exclusive subtype-related fashion (Figure 4I). Notably, phased hypermutations of BCL2 were associated with BCL2 translocations, suggesting that they could be used to noninvasively detect BCL2 translocations (Figure 4J; supplemental Figure 14A-C).

Conversely, we observed that some adjacent somatic events were not phased (ie, derived from different haplotypes) in the ctDNA and provided direct evidence of convergent evolutionary patterns between different lymphoma clones. In the pretreatment plasma of Case#3, for example, we recorded multiple clones having independently acquired C-terminus truncating mutations of BCL10 that result in loss of apoptosis-inducing capability while retaining the CARD domain for NF-κ-B activation³⁵ (Figure 4K-L). Interestingly, in another case, these nonphased patterns concerned another CARMA complex proto-oncogene CARD11 (supplemental Figure 14D-F). Altogether, these findings demonstrate that besides their utility in MRD interrogation, hypermutation patterns in plasma provide an untapped resource to assess molecular subtype, oncogene addiction, and clonal heterogeneity in B-cell lymphomas.

Despite clinical high-risk demographics, the overall outcome of the patients was excellent (Figure 1B-C). To improve our understanding of patients who relapsed or had refractory disease, we scrutinized their ctDNA for molecular and genomic characteristics (Figure 5A).

Nonresponding DLBCLs were characterized by continuous ctDNA positivity with an end-of-therapy rebound in concentration (Figure 5A). All 3 primary refractory cases with complete sample series carried TP53 mutations, 2 were high-grade GCB lymphomas with concurrent MYC and BCL2 alterations, and 1 was an MCD-like ABC DLBCL (Figure 5B). Interestingly, in PD#1 and PD#3, TP53 mutations appeared to be coupled with the loss of heterozygosity at diagnosis, whereas in PD#2 we observed a shift in mutational repertoire with enrichment of TP53 mutation at progression coupled with a change in involved sites of the disease (Figure 5C-D). Despite being associated with refractory disease, the majority of TP53-mutated cases eventually reached molecular remissions, albeit TP53 loss was strongly associated with MRD positivity at midstaging (Figure 5E). Together with the observed elevated pretreatment ctDNA levels, the findings add a novel molecular understanding of the high-risk nature of TP53-mutated lymphomas^36-39 and may provide insight why MRD^mid+ was not significantly associated with outcome in our study (Figure 2D).

Apart from the 2 treatment-refractory cases with concurrent BCL2 and MYC aberrations, the outcomes of double-hit (DHIT) lymphomas were favorable (Figure 5F). Among DHITs, primary refractoriness was distinguished by concurrent TP53-loss coupled to high pretreatment ctDNA burden, whereas DHITs with lower burden, even those with TP53-loss,⁴⁰ responded to treatment and remained alive at the last follow-up (Figure 5F-G). These data suggest that high pretreatment ctDNA burden can distinguish poor prognosis even among rare high-risk biological entities.

Plasma samples of the patients who initially achieved CR but progressed during follow-up contained reemerged ctDNA with mostly pretreatment matching mutational repertoire (4 patients) (Figure 5A). Interestingly, in 2 of these patients, explorative variant calling of the plasma samples at relapse revealed emerged MYC single nucleotide variations (SNVs) that were not detectable in plasma at diagnosis (Figure 6A-B). Moreover, in Rel#4, these mutations emerged linked in phase with an ancestral subclonal MYC lesion (S197T), demonstrating that phasing patterns can advance during clonal evolution of lymphoma (Figure 6C). In the tissue biopsies at diagnosis and relapse, we identified a shared rearrangement known to cooccur with MYC SNVs,^36,41 and a marked increase in MYC protein levels in the relapse tissue (Figure 6D; supplemental Figure 15A-B). These findings suggest that upon translocation, MYC is subjected to mutation pressure, and the emerging point mutations associate with chemorefractory status of a late-relapsing clone.

Lastly, in REL#5 with pretreatment cfDNA WGS available, we observed exceptional sequencing depths of specific chromosome 8 sequences that assembled and completely solved the structure of an amplified extrachromosomal circular element of DNA (‘double minute’) containing an intact MYC gene (Figure 6E-G; supplemental Figure 15C, supplemental Table 7). The diagnostic biopsy had FL3B histology, a complex karyotype with tetraploid chromosome number, and BCL2 and BCL6 translocations (Figure 6H-I). According to fluorescence in situ hybridization (FISH), MYC locus signals, however, corresponded to tetrasomy of chromosome 8 with no rearrangement detected despite the plasma-recorded double minute carried hybridization sequences for FISH detection (Figure 6G,I). Therefore, to search for any evidence of the presence of this rearrangement in the diagnostic tissue, we reexamined MYC FISH in available archival samples and were able to pinpoint scattered and extremely few occasional cells with double minute corresponding FISH signals (Figure 6J). On the protein level, MYC immunoreactivity was estimated at 40% (Figure 6K). Despite complete metabolic response to therapy (PET-CT, Deauville score = 1), the response according to liquid biopsy was suboptimal, and 25 months after the initial diagnosis, the patient experienced progression with a transformation to triple-hit high-grade B-cell lymphoma (HGBL-TH) (Figure 6L-M). In the tissue biopsy at relapse, MYC immunoreactivity had dramatically increased (95% positive), and fluorescent interrogation of MYC locus revealed a pattern corresponding to innumerable amplified double minute signals detected in plasma WGS already at diagnosis (Figure 6N-O). This analysis reveals that the FL3B in the initial diagnostic biopsy had already undergone systemic transformation to HGBL-TH lymphoma at baseline with high ctDNA burden and TP53-loss, giving rise to progression and ultimately death from lymphoma.

Since the addition of rituximab, precision medicine approaches in aggressive B-cell lymphomas have not been successful as a part of primary therapy.^5,42 This is likely due to confounding clinical and biological complexity that compromises clinical trials and is poorly captured by clinical estimates and tumor biopsies. By dissecting liquid biopsies of aggressive B-cell lymphoma patients, we discovered novel quantitative, mutational, and fragmentation patterns in the ctDNA that can resolve undisclosed heterogeneity and identify the patients with incomplete responses and inferior outcomes following a uniform therapy.

Our findings anticipate a pivotal role for pretreatment ctDNA analysis in future clinical trial designs, treatment decisions, and translational research in DLBCL. First, despite meeting the same inclusion criteria for high risk, we found substantial variation in pretreatment ctDNA burden between the patients, which translated into different outcomes even within individual clinical and biological risk groups. To overcome this clinically relevant plane of heterogeneity, a ctDNA burden-based index could provide an objective and generalizable tool for clinical trial inclusion^21,43 and evaluation of their external validity. Furthermore, a ctDNA-based normalization could overcome selection biases arising from different diagnosis-to-treatment times^21,43 and procedures of tissue biopsy, for example, both of which are linked with high-risk in DLBCL.⁴⁴ Second, since baseline ctDNA levels determine both high- and low-risk patients, therapy stratification could result in improved outcomes. For example, patients with low ctDNA burden could benefit from therapy deescalation,⁴⁵ whereas consolidation could improve the outcomes of patients with a high burden. Third, genomic analysis of pretreatment ctDNA provides an accurate systemic diagnosis that can culminate into diagnosis-changing discoveries that are not necessarily obtained from the most readily accessible site of tissue biopsy. As such, spatial heterogeneity can compromise the emergence of precision medicine approaches relying on tumor tissue for the detection of actionable mutations. Establishing these concepts and their further understanding warrants international efforts, prospective trials, and additional translational studies.

Treatment decisions based on pretreatment ctDNA-educated approaches could be complemented with serial analysis of ctDNA for response evaluation and surveillance as conceptualized by Kurtz and colleagues.²² Our findings validate and build on previously reported concepts,^19,31 including novel ctDNA tools for precision oncology, such as phasing-aware MRD testing⁴⁶ and fragmentation-based relapse prediction. With further development, these independent conceptual sources of input could be applied to detect incomplete responses and progression early and accurately. These applications could also save resources, diminish potentially harmful invasive procedures, and reduce unnecessary overtreatment caused by current response evaluation methods lacking specificity. Under continuous development, more sensitive ctDNA platforms will be achieved through better technical error suppression, for example, with duplex-adapters⁴⁷ and additional tools to distinguish true MRD signal from biological confounders, such as clonal hematopoiesis,⁴⁸ that represent limitations of our study. Although compelling, the efficacy and feasibility of dynamic ctDNA-guided decision-making warrant establishment in a prospective setting.

Altogether, our comprehensive analysis of ctDNA in the serial plasma samples of trial patients provides a roadmap for better understanding of high-risk B-cell lymphoma on the liquid biopsy. Our findings, both in terms of validation and novelty, are of high translational importance revealing previously unexplored molecular principles in the ctDNA that can be used in diagnostics, response evaluation, as prognostic tools, and in understanding lymphoma biology. We anticipate a growing interest in the platforms integrating both mutational and fragmentation features of the ctDNA in future LBCL studies, including ctDNA-based clinical trial designs and personalized treatment.

This article was supported by the Academy of Finland (grant 311171), the Biomedicum Helsinki Foundation, the Digital Precision Cancer Medicine Flagship iCAN, the Finnish Cancer Organizations, the Sigrid Juselius Foundation, the Southern Finland Regional Cancer Center FICAN South, Helsinki University Hospital, The University of Helsinki, the Ida Montin Foundation, the Orion Research Foundation, and the Paolo Foundation.

Contribution: L.M., A.A., and S.L. conceived the study; L.M. and S.L. wrote the manuscript; A.P., J.J., Y.N.B., M.-L.K.-L., P.B., M.B., M.J., I.F., L.T.G.M., Ø.F., S.J., H.H., and S.L. provided samples and clinical data; A.P. and M.A. established the cell-free DNA extraction; M.L. and P.E. performed sequencing; L.M., A.A., P.M., and E.P. analyzed the data; A.A. preprocessed the raw sequencing data; S.L. supervised the study and data analysis; and all authors read and approved the manuscript.

Conflict-of-interest disclosure: S.L.: Janssen-Cilag: Consultancy, research funding*; Roche: Consultancy, honoraria, research funding*; Takeda: Consultancy, honoraria, research funding*; Novartis: Consultancy, honoraria; Bayer: Research funding*; Celgene: Consultancy, research funding*; Gilead: Consultancy*. H.H.: Gilead: Consultancy, honoraria*; Takeda: Consultancy, honoraria*; Novartis: Consultancy, honoraria*; Genmab: Consultancy, honoraria*, Incyte: Consultancy, honoraria*; Roche: Consultancy, honoraria*. Ø.F.: Novartis: Consultancy, honoraria*. M.B.: Incyte, Mundipharma, Schain Research, Roche, Pfizer, Bristol-Myers Squibb, and Janssen-Cilag: Consultant, advisor*; Takeda: Research grant*. The remaining authors declare no competing financial interests.

*Not related to this study.

Correspondence: Sirpa Leppä, Department of Oncology, Helsinki University Central Hospital Cancer Center, P.O. Box 180, FI-00029 Helsinki, Finland; e-mail: sirpa.leppa@helsinki.fi.