Mutational dynamics and immune evasion in diffuse large B-cell lymphoma explored in a relapse-enriched patient series

Diffuse large B-cell lymphoma (DLBCL) is the most common non-Hodgkin lymphoma, representing ∼30% of all cases in western countries. It is an aggressive and molecularly heterogeneous form of lymphoma, with >18 different subtypes classified by the World Health Organization (WHO).^1,2  Primary refractory disease or relapse DLBCL (rrDLBCL), in which the disease progresses after an apparent effective treatment or clinical response to standard of care (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone [R-CHOP]), has long been recognized as a major contributor to the cancer’s mortality. The 3-year progression-free survival (PFS) in rrDLBCL cases is only 14%.³  Moreover, with the introduction of personalized and targeted therapies as viable therapeutic options, the appearance of multiple resistance mechanisms represents a major identified challenge. Thus, understanding the common features of drug resistance in DLBCL will not only contribute to our understanding of the leading cause of death in lymphoma, but may also assist in future drug design.

Next-generation-sequencing studies have refined the complexity of DLBCL subtypes and provided new categorizations of de novo DLBCL.^4-7  The genetics of rrDLBCL have been less studied, and only a limited number of studies have focused on tumor genetic evolution upon relapse by analyzing paired diagnostic and relapsed biopsies.^8-11  In recent years, the majority of sequencing studies in DLBCL have focused on providing a genetic “lymphopanel” to define the underpinnings of diversity and improve overall survival (OS) in this disease. We aimed to enhance DLBCL genomic profiling with a relapse-rich cohort of Nordic DLBCL cases, including 18 serially sampled cases, to contribute substantial information on progression and relapse.

A total of 83 cases and 102 biopsies (n [cases] = 83; b [biopsies] = 102) of DLBCL were included in this study. The whole-exome sequencing (WES) cohort consisted of 45 fresh-frozen biopsies from a Nordic cohort of 37 de novo DLBCL cases with matched normal DNA from peripheral blood, and 11 biopsies without matched normal from 5 patients. Patients were included from Oslo University Hospital (OUH) (n = 40; b = 54) and the Helsinki University Central Hospital Comprehensive Cancer Center (HUCHCCC) in Finland (n = 2; b = 4) (supplemental Tables 1-3; supplemental Figure 1).

A validation cohort using custom panel sequencing of 46 formalin-fixed paraffin-embedded biopsies was collected at OUH (n = 35) or at HUCHCCC (n = 11) (supplemental Tables 2 and 4; supplemental Figure 1), excluding cases with low DNA quality.

This study was approved by the Norwegian South Eastern Regional Ethics Committee (REK 2014/127). The study was performed according to the Declaration of Helsinki. All patients consented to written patient information.

Genomic DNA from fresh-frozen tissue and EDTA-treated blood was extracted using the Maxwell 16 DNA isolation automated magnetic bead instrument (Promega). Library preparation was carried out using SureSelectXT Human All Exon V5 (Agilent) per the manufacturer’s instructions at the OUH Genomics Core Facility, and whole exome capture and paired-end sequencing was performed using the Illumina HighSeq 2500 platform (average depth 289X, range 178-377X for tumor, 126X [range 88-173X] for normal; supplemental Figure 5). A benchmarked bioinformatics pipeline¹²  was used to process the sequencing reads and perform somatic variant calling (supplemental Materials and methods; supplemental Figure 4).

Copy-number aberration analysis was performed using FACETS¹³  version 0.5.14. Recurrent copy-number alterations by arm level and focal, <50% of the chromosome arm, peaks were analyzed by GISTIC2.0 (version 2.0.22). Regions of alteration with a q value of <0.25 were reported as significant. The WES cohort was also subjected to SNP6.0 analysis using the Affymetrix Genome Wide SNP6.0 Array platform (Gene Chip lot 4296976).^14,15 

POLYSOLVER software (v1; Broad Institute, Cambridge, MA) was used to infer the HLA class I alleles for each individual, and POLYSOLVER mutation calling and annotation was performed¹⁶  on WES cases with matched normal. Loss of heterozygosity (LOH) was determined by FACETS¹³  and POLYSOLVER.

Driver genes were identified with MutSig2CV (Broad Institute)^4,17  and OncodriveFM and OncodriveCLUST programs¹⁸  (version 2.4.1). Genes with a q value of <0.1 were selected as harboring potential driver mutations.¹⁷ 

Clonal structure was inferred with sciClone (version 1.1)¹⁹  with loss-of-heterozygosity information from SNP6.0 array inferred by ASCAT. We used ClonEvol (https://github.com/hdng/clonevol) and Fishplot²⁰  (R version 3.4.1) to build tumor phylogenies. Targeted cancer drugs were identified through the Open Targets Platform (downloaded 09.19),²¹  with additional filtering collected from OncoTree (http://oncotree.mskcc.org/#/home; phenotypes available upon request). Variant deleterious effect was inferred through a combined annotation-dependent depletion (CADD)²²  phred score of over 10.

Immunohistochemical staining (IHC) for class I antigens was performed applying antibodies against HLA-A (EP1395Y; Abcam) and β2-microglobulin (B2M) (2M2; Lifespan BioSciences). Details on IHC staining are provided in supplemental Materials and methods.

GraphPad Prism (Version 8.01) was used to perform an unpaired 2-tailed Student t test with Welch correction, contingency χ², and the paired Student t test. R (version 3.4.1) was used for linear regression on time to relapse.

To address the underlying genomics of acquired and/or intrinsic resistance in DLBCL, WES was performed on 42 de novo DLBCL cases consisting of 56 biopsies; 13 cases had available serial biopsies and 34% of all biopsies were taken at relapse (supplemental Figures 1-4; supplemental Tables 1-3). Custom targeted sequencing, of 139 genes (supplemental Table 5), was performed on a second cohort of formalin-fixed paraffin-embedded samples representing 41 cases of which 23% were relapse biopsies (supplemental Figures 1 and 3; supplemental Tables 2, 4, and 5).

Given the unique enrichment of relapsed DLBCL in our cohort, we categorized biopsies as diagnostic DLBCL (from patients who did not experience relapse), diagnostic rrDLBCL, and relapse biopsies from rrDLBCL (Figure 1A), and analyzed multiple case sets stratified by disease subtype, survival, and relapse status. This approach allowed a comprehensive investigation into the heterogeneous subtypes included in our analysis (ABC, GCB, PMBCL, T-cell/histiocyte-rich; and more, supplemental Tables 1-4) and exploration based on resistance and progression. Mutation signatures and overall mutation burden in this Nordic cohort were similar to previously published western and Chinese cohorts (median of 125 coding mutations per case; supplemental Figures 5-7).^4,8,10,23-27  Using 2 different approaches for cancer driver gene prediction, we identified 118 potential driver genes including previously reported genes as well as multiple genes involved in antigen presentation (supplemental Tables 6 and 7). Four previously unidentified DLBCL driver genes were discovered: TMEM199, TBC1D24, IQSEC3, and SYNGR1 (supplemental Tables 6 and 7). Comparison of driver genes found in diagnostic biopsies from cases with a durable treatment response and cases that ultimately went on to relapse revealed an increased number of mutations in genes involved in antigen presentation, including HLA-A, HLA-B, CD74, and LILRA4 in the latter set of cases (Figure 1B; supplemental Figure 8; supplemental Table 6), implicating immune escape as a notable feature of DLBCL-intrinsic resistance.

In contrast to the initial high sensitivity of DLBCL to standard-of-care treatment, patients who experience relapse often have astonishingly aggressive and treatment-refractory disease. We observed that of the coding variants appearing at relapse, 8 recurrently mutated genes (ie, genes mutated in at least 17% of the tested relapse biopsies) were not mutated in any diagnostic pretreatment tumors: CACNA1B, FLRT1, KRAS, LRFN5, MPRIP, WNT8A, and NR1H2 (Figure 1B; supplemental Figure 12). However, all are represented in diagnostic biopsies in previously studied cohorts. Nevertheless, NR1H2 was reported as exclusively mutated in relapse biopsies by Morin et al, providing potential evidence for an association with resistance.⁸ NR1H2, a nuclear receptor for LXR-β, is ubiquitously expressed in tissues and is a key regulator of macrophage function, with a small molecule agonist currently in clinical trial.^28,29 

Furthermore, copy-number variation analysis revealed recurrent loss (either focal or arm level) of a region of chromosome 6p21 containing multiple antigen-presentation and antigen-processing genes (Figure 1B; supplemental Figure 9). Association of copy-number variations in 6p21 to clinical outcomes and immune-privileged site dissemination has been shown previously.^30,31  We detected focal losses of immune-processing genes LGMN and PSME3, in addition to genes involved in antigen presentation: B2M, PDIA3, and RFXAP (Figure 1B; supplemental Figure 9). These findings provide further support for the idea that immune cloaking may be important in DLBCL resistance.

We quantified variants in HLA-A and revealed an increased incidence in cases that ultimately relapsed (Figure 2; supplemental Table 8). Given the highly polymorphic nature of the MHC regions, we performed further in-depth alignments and investigations into the mutational prevalence, timing, impact on prognosis, and LOH. We revealed HLA-A mutagenesis and LOH as a plausible method of immune evasion in cases that went on to relapse (Figure 2A-C; supplemental Figure 10; χ² test; P < .02). Given that 2 of our cases were not treated with R-CHOP–like therapies, we retested this association without these 2 cases, and it remained significant (χ² test; P < .04). To determine whether these genomic changes in immune-related genes result in altered protein expression patterns, we performed IHC staining of HLA-A and B2M on 58 biopsies representing 55 cases of DLBCL pooled from all of our cohorts (Figure 2D-E). The majority of HLA-A losses were observed in the rrDLBCL diagnostic and relapse biopsies, but the difference did not reach statistical significance (Figure 2D; χ² test; P < .1). B2M expression was variable between the groups (supplemental Figure 11). Host immunogenetics, particularly HLA haplotypes and MHC class II variations, has emerged as a potential contributor to survival in DLBCL and follicular lymphoma patients.^32-34  HLA class I protein deficiencies have been reported in up to 43% of DLBCL cases.³⁵  We look forward to future investigations to unravel whether HLA-A mutational and protein status correlate with neoantigen profiles or the immune microenvironment of patients.

Thus far, our analyses of the 3 biopsy types have suggested that immune cloaking may be an intrinsic-resistance mechanism, as antigen-presentation alterations already exist at diagnosis and may predict response. Our relapse-rich cohort should provide significant information in the search for acquired relapse mechanisms and thereby potentially contribute to development of improved salvage therapies. Due to clinical practices, serial biopsies are rare, but we were able to assemble 18 serially sampled cases with corresponding detailed clinical annotations. Patient-wise, we compared mutations identified in the initial diagnostic biopsy to mutations identified in the relapse biopsy and integrated the molecular results with the clinical parameters including biopsy location (supplemental Figure 13). As time to relapse increased, we observed a strong trend of decreasing similarity between biopsies representing evolutionary divergence over time (Figure 3A-B). However, this observation was independent of lesion location. Leveraging the availability of matched blood and of copy-number information, we applied 2 clonality analysis tools (sciClone¹⁹  and ClonEvol²⁰ ) on 7 cases with serial biopsies available (Figure 3C; supplemental Figures 13 and 14). At each time point, all cases had clones that contained variants unique to the biopsy at ∼50% variant allelic fraction, thus suggesting 2 different dominant clones at diagnosis and relapse (Figure 3C; supplemental Figures 13 and 14), with a substantial overlap of variants between biopsies. This indicates branched evolution with late divergence of the tumors, and suggests that the dominant clone at diagnosis is eliminated by the given therapy and a unique clone, albeit derived from a common progenitor clone, dominates at relapse (Figure 3C). Founding clones frequently had alterations in antigen-presentation and -processing genes such as HLA genes and B2M. In 1 case, P5, multiple variants in HLA genes were gained throughout the oncogenic progression of the tumor. However, as in previous investigations, we were unable to find a recurrent mutation clearly responsible for resistance to R-CHOP–like therapies. This may suggest that resistance is intrinsic, or that multiple resistance mechanisms appear in each patient and are diverse among DLBCL cases.

Given the strong similarities between diagnostic and relapse biopsies and the observed evolutionary divergence with time, we sought to understand whether relapse sampling advances our understanding of treatment options enough to warrant rebiopsies. We assessed the evolutionary timing of mutations in druggable targets. There were a median of 23 mutations per biopsy in genes targetable by known antineoplastic drugs as listed in the Open Targets Platform.²¹  To incorporate the functional impact of each variant on driving cancer, we analyzed all druggable variants using CADD²²  (Figure 3D-E). Private variants in druggable genes were found more frequently in relapse biopsies than in initial diagnostic biopsies (Figure 3D; P < .0221). Although our results indicate that relapse sampling is important for guiding treatment, obtaining relapse biopsies can present ethical and practical issues for clinicians. Therefore, sampling of the peripheral blood for circulating tumor DNA or circulating tumor cells may provide an alternative investigational route for relapse sampling.^36-38 

Two of the previous studies on serially sampled rrDLBCL cases revealed a novel increase in mutations associated with the NF-κB pathway and that immune surveillance was disrupted by indels and copy number variations in CD58 and B2M.²³  In contrast, in our paired samples, these alterations were seen in both the diagnostic and relapse biopsies. Our results suggest that immune-surveillance targeting also occurs at diagnosis, and should be taken into consideration when making clinical decisions on the use of the growing arsenal of immune therapies. Theoretically, these patients may be able to develop resistance to checkpoint blockade and neoantigen therapy, but may be sensitive to engineered immune therapies such as chimeric antigen receptor–T-cell therapies and NK-cell therapies. CD19-engineered chimeric antigen receptor T cells have shown long-term durable response rates in up to 58% in rrDLBCL.^39-43  More broadly, our findings yield insight into the development of resistance in DLBCL, and highlight antigen processing and presentation as key targets for genomic aberrations in rrDLBCL cases. Moreover, our findings provide a clear investigational route into the effect of HLA-A mutations on relapse status in future clinical correlative studies.

The authors thank Merete Hektoen for training and support in DNA sample preparation, and Jinchang Sun and Anna Lid from the Genomics Core Facility at the Oslo University Institute for Cancer Research for their contributions.

This work was supported by the Norwegian Research Council (grant nos. 221580 and 218241).

J.F.W., S.N., D.V., and L.B.A. are associated members, L.A.M.-Z., E.H., R.A.L., and H.H. are members of the leadership group, and O.M. is the head of the Norwegian Cancer Genomics Consortium.

Contribution: J.F.W., J.H.M., E.B.S., and H.H. conceptualized and designed the study; J.F.W., G.T., B.B., V.H., S. Leppä, A.P., and H.H. collected and organized samples; J.F.W. analyzed data and drafted the article; J.F.W., S.N., C.B.S., M.S.L., L.B.A., D.V., E.H., and O.C.L. performed bioinformatics analyses; B.J., A.S., and R.A.L. analyzed and provided supportive design for HLA allele inference; O.C.L. and M.S.L. provided statistical analysis support; Y.N.B., S. Leppä, and H.H. collected clinical data; L.A.M.-Z. and S. Lorenz provided sequencing technologies and expertise, as well as generation of the sequencing data; K.B. provided pathology support and IHC evaluation; J.F.W., S.N., C.B.S., M.S.L., E.B.S., and H.H. provided data interpretation; O.M., J.H.M., S.N., D.V., E.H., O.C.L., M.S.L., Y.N.B., C.B.S., E.B.S., and H.H. provided critical revision of the article; and all authors read and approved the final manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Jillian F. Wise, University of Oslo, Ullernchausseen 70, Montebello, 0310 Oslo, Norway; e-mail: jillian.wise@medisin.uio.no; and Harald Holte, Oslo University Hospital, Institute for Cancer Research, PB 4953 Nydalen, 0424 Oslo, Norway; e-mail: hhe@ous-hf.no.