WO2021173796A1

WO2021173796A1 - Methods of treating diffuse large b cell lymphoma based on particular genetic subtype

Info

Publication number: WO2021173796A1
Application number: PCT/US2021/019605
Authority: WO
Inventors: Louis M. Staudt; George W. WRIGHT; Da Wei Huang
Original assignee: The United States Of America, As Represented By The Secretary, Department Of Health And Human Services
Priority date: 2020-02-25
Filing date: 2021-02-25
Publication date: 2021-09-02

Abstract

Disclosed herein are methods of treating diffuse large B cell lymphoma based on particular genetic subtype.

Description

METHODS OF TREATING DIFFUSE LARGE B CELL LYMPHOMA BASED ON

PARTICULAR GENETIC SUBTYPE

CROSS REFERENCE TO RELATED APPLICATION

[0001] This patent application claims the benefit of U.S. Provisional Patent Application No. 62/981382, filed February 25, 2020, which is incorporated by reference in its entirety herein.

STATEMENT REGARDING

FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with Government support under project numbers ZIA BC 011006-05 and CT000200-29 by the National Institutes of Health, National Cancer Institute.

The Government has certain rights in the invention.

BACKGROUND OF THE INVENTION

[0003] Initial progress towards improved treatment of DLBCL subtypes came with the advent of gene expression profiling, which was used to define two prominent DLBCL subtypes, termed GCB and ABC, which comprise 80-85% of cases, with the remaining cases declared as “Unclassified.” This classification accounted for some of the heterogeneity in the clinical outcome following chemotherapy using the R-CHOP regimen. This so-call “cell of origin” (COO) methodology has proven useful in understanding the varied responses of DLBCL patients to targeted therapies such as ibrutinib, an inhibitor B cell receptor (BCR) signaling. Nonetheless, the COO distinction does not fully account for the heterogeneous responses and outcomes following either R-CHOP therapy or ibrutinib monotherapy. This is likely due to the fact that gene expression profiling provides a phenotypic description of cancers rather than a genetic description that encompasses tumor pathogenesis.

[0004] While recurrent genetic aberrations in individual genes have elucidated oncogenic mechanisms in DLBCL, progress towards improved treatment of these malignancies required the integration of data from genome and transcriptome sequencing as well as array-based comparative genomic hybridization to identify genes that were recurrently altered by mutations, translocations and/or copy number alterations. Mathematically distinct clustering methods were used to group DLBCL tumors into genetic subtypes that each shared a set of genomic aberrations in subtype signature genes. The utility of this genetic classification was evident by the association of the genetic subtypes with outcome following RCHOP therapy.

[0005] Clustering methods are limited by the necessity to place a tumor sample into no more than one subtype and by the fact that the subtype assignment of a particular tumor can vary if different tumors are included during the clustering process. Therefore, there is a need for additional methods of treating DLBCL, including methods based on the particular genetic subtype of DLBCL.

BRIEF SUMMARY OF THE INVENTION

[0006] In an embodiment, the invention provides a method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, without use of gene copy number, the method comprising: (a) obtaining from the subject a biopsy sample of the lymphoma; (b) (i) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1 A, or (ii) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB; (c) determining whether at least two weight values listed in the Table 1 A or IB are positive for each of the subtypes BN2, EZB, MCD, and ST2 based on the genetic features detected in (b); (d) calculating the probabilities that the sample belongs to (i) subtype Nl, and (ii) each subtype BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation

wherein P{ subtype] V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 1 A or IB based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate P(subtype|F), and wherein P(subtype|F) is calculated for each subtype using only V, weight values for that subtype;

(e) designating each subtype BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have /’(subtype) V) < 0.5; and (f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic: (1) if only one subtype has /’(subtype) V) > 0.9 with the other subtypes having /(subtype)F) < 0.9, the sample is classified as being the one subtype having /’(subtype) V) > 0.9; (2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype) V) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and (3) if there is more than one subtype having /’(subtype) V) > 0.9, the sample is classified as being each of the subtypes having /’(subtype) V) > 0.9.

[0007] In an embodiment, the invention provides a method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, with use of gene copy number, the method comprising: (a) obtaining from the subject a biopsy sample of the lymphoma; (b) (i) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2A, or (ii) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2B, wherein a copy number of each of the subtypes is a genetic feature detected; (c)determining whether at least two weight values listed in the Table 2 A or 2B are positive for each of the subtypes A53, BN2, EZB, MCD, and ST2 based on the genetic features detected in (b); (d) calculating the probabilities that the sample belongs to (i) subtype Nl, and (ii) each subtype A53, BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation e^xP(åi )

/(subtype \V) = — - l + exp(å Vi) wherein /(subtype) V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 2A or 2B based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate /(subtype) V), and wherein /(subtype) V) is calculated for each subtype using only V weight values for that subtype;

(e) designating each subtype A53, BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have /(subtype) V) < 0.5; and (f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic: (1) if only one subtype has /’(subtype) V) > 0.9 with the other subtypes having /’(subtype)F) < 0.9, the sample is classified as being the one subtype having /’(subtype) V) > 0.9; (2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype) V) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and (3) if there is more than one subtype having /’(subtype) V) > 0.9, the sample is classified as being each of the subtypes having /’(subtype) V) > 0.9.

[0008] Additional embodiments are as described herein.

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Fig. 1 A presents exemplary cancer subtype discovery and prediction using the LymphGen algorithm. Shown at left is the discovery of cancer subtypes starting with “seed” sets of cases using the GenClass algorithm. The LymphGen algorithm uses prevalences of genetic features to estimate the likelihood that a feature is associated with a subtype and combines these likelihoods to calculate a probability that a tumor belongs to a genetic subtype. An example is shown in which features associated with the EZB subtype are present (“1”) or absent (“0”) in a tumor, and the likelihoods that the tumor is EZB or non-EZB based on each feature. The final panel illustrates how LymphGen assigns a tumor using the subtype probabilities.

[0010] Fig. IB presents the frequency of cases with high probability (“Core”) or moderate probability (“Extended”) subtype assignments, genetically composite cases, and unassigned (Other) cases.

[0011] Fig. 1C presents the prevalence of various subtypes.

[0012] Fig. ID presents a diagram showing, top, prevalence of subtypes in DLBCL COO subgroups, bottom, prevalence of COO subgroups within each genetic subtype.

[0013] Figs. 2A-2F present tables of genetic features associated with the six LymphGen genetic subtypes of DLBCL: MCD (Fig. 2A), BN2 (Fig. 2B), EZB (Fig. 2C), ST2 (Fig. 2D),

A53 (Fig. 2E), and N1 (Fig. 2F). Log p-values are based on the difference in prevalence of each genetic feature in the indicated subtype versus all other samples. HL: Heterozygous loss; HD: Homozygous deletion; Gain: Single copy gain; Amp: amplification; Mut: Mutation; Trunc: Protein truncation mutation: Fusion: chromosomal translocation. [0014] Fig. 3 A presents a diagram showing genetic aberrations favoring immune escape in

MCD DLBCL.

[0015] Fig. 3B shows prevalence of CD79B and MFZ)55^L265P mutations in the indicated nodal and extranodal forms of DLBCL, shown according to the code at the top. The percent prevalence of tumors with the indicated genotypes in each of the indicated lymphoma types is shown according to the code.

[0016] Fig. 3C is a bar graph showing secondary extranodal involvement in genetic subtypes of nodal DLBCL. Shown are the prevalences of extranodal spread of tumors of the indicated genetic subtype to the CNS/vitreo-retina, testis, and breast.

[0017] Fig 3D is a bar graph showing prevalence of MCD-defming mutations in PCNSL and primary cutaneous lymphoma; Other NHL: Other non-Hodgkin lymphomas (see below).

[0018] Fig. 3E is a bar graph showing prevalence of BN2-defming mutations in the indicated types of marginal zone lymphoma (MZL) and in other NHLs.

[0019] Fig. 3F is a bar graph showing prevalence of EZB-defming mutations in FL, transformed FL and other NHLs.

[0020] Fig. 3G is a bar graph showing prevalence of ST2-defming mutations in NLPHL, THRLBCL and in other NHLs.

[0021] Fig. 4A is a diagram showing prevalence of DLBCL subtypes classified by LymphGen.

[0022] Figs. 4B-4D present diagrams showing prevalence of COO subgroups within genetic subtypes for the NCI cohort (Fig. 4B), the Harvard cohort (Fig. 4C), and the BCC cohort (Fig. 4D).

[0023] Figs. 4E and 4F present bar graphs showing prevalence of the indicated genetic features within the genetic subtypes defined in the Harvard (Fig. 4E) and BCC (Fig. 4F) cohorts in comparison with the NCI cohort.

[0024] Figs. 4G-4M present Kaplan-Meier plots of overall survival within the indicated DLBCL cohorts, either in all cases (NCI cohort, Fig. 4G), ABC cases (NCI cohort, Fig. 4H; Harvard cohort, Fig. 41; BCC cohort, Fig. 4J), and GCB cases (NCI cohort, Fig. 4K; Harvard cohort, Fig. 4L; BCC cohort, Fig. 4M). [0025] Fig. 4N is a bar graph showing hazard ratios (-log2 transformed) for the indicated comparisons between LymphGen subtypes in the indicated DLBCL cohorts. Error bars: SEM. Significance: ****: p<0.0001, ***: p<0.001, **: p<0.01, *: p<0.05.

[0026] Figs. 5A-5E present bar graphs showing gene expression signature expression in DLBCL subtypes for cancer processes (Fig. 5 A), B cell differentiation (Fig. 5B), B cell transcription factors (Fig. 5C), oncogenic signaling pathways (Fig. 5D), and tumor microenvironment (Fig. 5E). Shown is average log2 expression of signature genes in each subtype versus other DLBCL samples in the NCI cohort. Error bars: SEM.

[0027] Fig. 6A is a bar graph showing relative expression of the DHIT signature in the indicated subtypes within GCB DLBCL.

[0028] Fig. 6B is a graph showing prevalence of subtypes within DHIT+ GCB DLBCL cases.

[0029] Fig. 6C is a graph showing gene set enrichment analysis of the relationship of the DHIT and the GCB-4 and MYCUp- 4 signatures. Cases are ranked according to the T statistic for the comparison of DHIT+ and DHIT- EZB DLBCL. Kolmogorov-Schmirnov P-values are shown.

[0030] Fig. 6D is a graph showing correlation between the DHIT score and a linear model score derived using GCB-4 and MYCUp-4 signature averages. Each dot is an EZB case. A Pearson correlation p-value is shown.

[0031] Figs. 6E and 6F present Kaplan-Meier plots of survival for DHIT+ and DHIT- cases among EZB (Fig. 6E) and non-EZB (Fig. 6F) GCB cases.

[0032] Fig. 6G presents a table with genetic features that distinguish EZB-MYC+ (DHIT+) from EZB-MYC- (DHIT-) GCB DLBCL (top 2 panels), and features that distinguish EZB from other DLBCL cases that are shared by EZB-MYC+ and EZB-MYC-.

[0033] Fig. 6H is a bar graph showing prevalence of genetic features that distinguish EZB- MYC+ from EZB-MYC-in BL.

[0034] Fig. 7A is a schematic showing contribution of each genetic subtype to the indicated genetic aberrations in the BCR-dependent NF-AB pathway. [0035] Figs. 7B and 7C are bar graphs showing fraction DLBCL subtype cases with genetic alterations targeting the BCR-dependent NF-AB pathway (Fig. 7B) or negative regulators (Fig. 7C) of proximal BCR signaling.

[0036] Figs. 7D-7I present graphs showing fraction of cases expressing the IgV_H4-34 variable region or other IgV_H regions and fraction of cases expressing the indicated IgH constant regions for MCD (Fig. 7D), BN2 (Fig. 7E), A53 (Fig. 7F), ST2 (Fig. 7G), N1 (Fig. 7H), and EZB (Fig. 71).

[0037] Fig. 7J is a bar graph showing CRISPR-mediated knockout of BCR and NF-AB negative regulators promotes survival in models of MCD and BN2 DLBCL. Cas9+ cells expressing the indicated sgRNAs with GCB were cocultured with parental (GFP-) cells for the indicated times in ibrutinib. Increasing values indicate relative ibrutinib resistance of the sgRNA+ cells.

[0038] Figs. 7K-7S are bar graphs showing genome-wide CRISPR loss-of-function screens (BCR subunits, Fig. 7K; BCR-dependent NFKB activation, Fig. 71; IKB kinase, Fig. 7M; MYD88 pathway, Fig. 7N; PI3 kinase, Fig. 70; IRF4/SPIB, Fig. 7P; JAK/STAT, Fig. 7Q;

PRC2 complex, Fig. 7R; BCL2 family, Fig. 7S). The indicated Cas9+ models of MCD, BN2, and EZB were transduced with a genome-wide sgRNA expression library, and the sgRNA abundance was quantified before and after 3 weeks in culture. Asterisk: targeted by approved or investigational drugs.

[0039] Fig. 7T presents immunoblots showing the effect of BCR knockdown on signaling in a BN2 model. Riva cells were transduced with the indicated small hairpin RNAs (shRNAs) and the effect on BCR signaling was assessed by immunoblotting.

[0040] Fig. 7U is a line graph showing the effect of ibrutinib on Riva xenograft growth. Following the tumor establishment, mice (n=5/group) were treated with the indicated ibrutinib doses or vehicle control.

[0041] Fig. 8A provides a summary of the relationship between DLBCL COO subgroups and genetic subtypes (left), as well as the genetic, phenotypic, clinical attributes, and treatment implications of the subtypes (right). Prevalences were estimated using the NCI cohort, adjusting for a population-based distribution of COO subgroups (see Methods). [0042] Fig. 8B diagrammatically presents models of selection for shared genetic features in DLBCL subtypes.

[0043] Fig. 8C diagrammatically presents models of the shared genetic attributes of a DLBCL genetic subtype and indolent NHLs.

[0044] Fig. 8D diagrammatically presents model EZB-MYC+ and EZB-MYC- evolution. [0045] Figs. 9A-9L present genetic features in genetically composite DLBCL tumors (MCD/BN2, Fig. 9A; MCD/EZB, Fig. 9B; MCD/ST2, Fig. 9C; MCD/A53, Fig. 9D; BN2/N1, Fig. 9E; BN2/EZB, Fig. 9F; BN2/ST2, Fig. 9G; BN2/A53, Fig. 9H; ST2/N1, Fig. 91; EZB/N1/ST2/A53, Fig. 9J; EZB/ST2/A53, Fig. 9K; and EZB/A53, Fig. 9L). Shown are genetic features in DLBCL tumors that have been classified into a more than one subtype by the LymphGen algorithm (n=33). The significance of each genetic feature in the LymphGen model is presented according to the scale.

[0046] Figs. 10A-10D present tables with genetic features that distinguish DLBCL subtypes (MCD and BN2, Fig. 10A; BN2, Nl, EZB, and ST2, Fig. 10B; ST2 and A53, Fig. IOC; A53, Fig. 10D). Shown are genetic features in DLBCL biopsy samples that have been classified into a single subtype (n=329).

[0047] Fig. 10E presents bar graphs showing expression of BCL2 mRNA in the indicated subsets of DLBCL. Each value is the difference between the average BCL2 expression in the subset of cases relative to the mean expression over all cases. Left panel displays expression in genetic subtypes and gene expression subgroups. Right panel displays the influence of BCR translocation or BCL2 gain/amplification on expression levels in the indicated subset of cases. [0048] Fig. 1 OF is a bar graph showing relative expression of CD274 mRNA, encoding PD- Ll, and PDCD1LG2 mRNA, encoding PD-L2, in MCD tumors with fusions of the respective genes versus other, fusion-negative MCD tumors.

[0049] Fig. 10G is a bar graph showing prevalence of the indicated MCD-defming mutations in primary testicular, primary breast and primary intravascular lymphomas.

[0050] Fig. 10H presents a table with clinical and analytic aspects of the three DLBCL cohorts used in the present study (Chapuy et al., 2018; Ennishi et al., 2019a; Schmitz et al., 2018). [0051] Figs. 10I-10Q are bar graphs showing sensitivity, specificity and precision (positive predictive value (PPV)) for LymphGen predictor models based on the indicated availability of dataset types: mutations/4-class copy number assignments/no BCL2/BCL6 fusions, Fig. 101; mutations/4-class copy number assignments/BCL6 fusions/no BCL2 fusions, Fig. 10J; mutations/4-class copy number assignments/BCL2 fusions/no BCL6 fusions, Fig. 10K; mutations/no copy number data/BCL2/BCL6 fusions, Fig. 10L; mutations/no amp/homdel (loss; gain; WT)/ BCL2/BCL6 fusions, Fig. 10M; mutations/amp/homdel only (no hetloss; homdel)/BCL2/BCL6 fusions, Fig. ION; mutations/no copy number data/BCL2/BCL6 fusions, Fig. 10O; mutations; no copy number data/BCL6 fusions/no BCL2 fusions, Fig. 10P; mutations; no copy number data/BCL2 fusions/no BCL6 fusions, Fig. 10Q. Each plot is based on data from the NCI cohort and is a comparison with the LymphGen model based on a full complement of genetic data (mutations, 4- class copy number assignments (amplification, gain, heterozygous loss, homozygous deletion), and fusions (BCL2, BCL6). Models lacking copy number data did not include A53 since this subtype is defined primarily by copy number alterations.

[0052] Fig. 1 OR is a bar graph showing the relationship between clusters defined in the Harvard cohort and LymphGen-predicted subtypes. All LymphGen subtypes were predominantly composed of cases belonging to a single cluster, with the exception of cases in the N1 subtype, which were not assigned to a cluster.

[0053] Figs. 11 A-l 1C are Kaplan-Meier plots of overall survival in the indicated DLBCL cohorts (NCI cohort, Fig. 11 A; Harvard cohort, Fig. 1 IB; BCC cohort, fig. 11C) according to COO gene expression subgroup.

[0054] Fig. 1 ID presents Kaplan-Meier plots of overall survival in the BN2 (left) and A53 (right) genetic subtypes according to COO gene expression subgroup within the NCI cohort. [0055] Fig. 1 IE is a Kaplan-Meier plot of overall survival of DHIT+ and DHIT- subsets of GCB DLBCL.

[0056] Fig. 11 F is a receiver operating characteristic (ROC) curve for the binary probabilistic classifier of EZB-M+ versus EZB-M-. [0057] Figs. 11G and 11H are line graphs showing the effect of low doses of ibrutinib (5 mg/kg/day) on growth of xenografts of two MCD models, TMD8 (Fig. 11G), and HBL1 (Fig. 11H).

[0058] Figure 12 presents a scheme of set of features ordered for a given gene in a hierarchical manner as described in the Example.

DETAILED DESCRIPTION OF THE INVENTION

[0059] Described herein are treatment methods that integrate determination of the probability that an individual patient’s tumor belongs to a particular genetic subtype and that allow for the possibility that the tumor may have acquired more than one genetic program during its evolution, such that treatment for the patient can then be more practically applied.

[0060] In an embodiment, the invention provides a method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, without use of gene copy number, the method comprising: (a) obtaining from the subject a biopsy sample of the lymphoma; (b) (i) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1 A, or (ii) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB; (c) determining whether at least two weight values listed in the Table 1 A or IB are positive for each of the subtypes BN2, EZB,

MCD, and ST2 based on the genetic features detected in (b); (d) calculating the probabilities that the sample belongs to (i) subtype Nl, and (ii) each subtype BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation

wherein P( subtype] V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 1 A or IB based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate P(subtype|F), and wherein P(subtype|F) is calculated for each subtype using only V, weight values for that subtype;

(e) designating each subtype BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have /’(subtype) V) < 0.5; and (f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic: (1) if only one subtype has /’(subtype) V) > 0.9 with the other subtypes having /’(subtype)F) < 0.9, the sample is classified as being the one subtype having /’(subtype) V) > 0.9; (2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype) V) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and (3) if there is more than one subtype having /’(subtype) V) > 0.9, the sample is classified as being each of the subtypes having /’(subtype) V) > 0.9.

[0061] In an embodiment, the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1 A.

[0062] In an embodiment, the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample all of the genetic features of the subtype N1 listed in Table 1A.

[0063] In an embodiment, the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB.

[0064] In an embodiment, the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample all of the genetic features of the subtype N1 listed in Table IB.

[0065] In an embodiment, the invention provides a method of treating a human subject with diffuse large B cell lymphoma (DLBCL), the method comprising: (A) classifying a DLBCL according to any of the above; and (B) treating the subject based on the one or more classifications of (A).

[0066] In an embodiment, the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-kB pathway is ibrutinib, acalabrutinib, or ONO/GS-4059. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib. [0067] In an embodiment, the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

[0068] In an embodiment, the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-kB pathway is ibrutinib, acalabrutinib, or ONO/GS-4059. In certain embodiments, the B cell receptor- dependent inhibitor of the NF-KB pathway is ibrutinib.

[0069] In an embodiment, the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1.

[0070] In an embodiment, the DLBCL is classified as ST2 or EZB, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl. In certain embodiments, the inhibitor of PI3 kinase is copanlisib, idelalisib, duvelisib, BYL-719, BKM-120 or GDC-0980. In certain embodiments, the inhibitor of MTORCl is sirolimus (rapamycin), temsirolimus, everolimus, OSI-027, AZD-8055, AZD-2014, or BEZ-235.

[0071] In an embodiment, the invention provides a method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, with use of gene copy number, the method comprising: (a) obtaining from the subject a biopsy sample of the lymphoma; (b) (i) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype Nl listed in Table 2A, or (ii) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype Nl listed in Table 2B, wherein a copy number of each of the subtypes is a genetic feature detected; (c)determining whether at least two weight values listed in the Table 2 A or 2B are positive for each of the subtypes A53, BN2, EZB, MCD, and ST2 based on the genetic features detected in (b); (d) calculating the probabilities that the sample belongs to (i) subtype Nl, and (ii) each subtype A53, BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation

wherein /’(subtype) V) is the probability that the sample is of a subtype, wherein each V, is a weight value listed in the Table 2A or 2B based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate /’(subtype) V), and wherein /’(subtype) V) is calculated for each subtype using only V weight values for that subtype;

(e) designating each subtype A53, BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have /’(subtype) V) < 0.5; and (f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic: (1) if only one subtype has /’(subtype) V) > 0.9 with the other subtypes having /’(subtype)F) < 0.9, the sample is classified as being the one subtype having /’(subtype) V) > 0.9; (2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype) V) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and (3) if there is more than one subtype having /’(subtype) V) > 0.9, the sample is classified as being each of the subtypes having /’(subtype) V) > 0.9.

[0072] In an embodiment, the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2A.

[0073] In an embodiment, the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2A.

[0074] In an embodiment, the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2B.

[0075] In an embodiment, the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2B.

[0076] In an embodiment, the invention provides a method of treating a human subject with diffuse large B cell lymphoma (DLBCL), the method comprising: (A) classifying a DLBCL according to any of the above; and (B) treating the subject based on the one or more classifications of (A). [0077] In an embodiment, the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib, acalabrutinib, or ONO/GS-4059. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib.

[0078] In an embodiment, the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

[0079] In an embodiment, the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib, acalabrutinib, or ONO/GS-4059. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib.

[0080] In an embodiment, the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1.

[0081] In an embodiment, the DLBCL is classified as ST2 or EZB, or EZB-MYC+ or EZB- MYC-, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl. In certain embodiments, the inhibitor of PI3 kinase is copanlisib, idelalisib, duvelisib, BYL-719, BKM-120 or GDC-0980. In certain embodiments, the inhibitor of MTORCl is sirolimus (rapamycin), temsirolimus, everolimus, OSI-027, AZD-8055, AZD-2014, or BEZ-235.

[0082] In an embodiment, the DLBCL is classified as A53, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib, acalabrutinib, or ONO/GS-4059. In certain embodiments, the inhibitor of the B cell receptor-dependent NF-KB pathway is ibrutinib.

TABLE 1A

TABLE IB

TABLE 2A

TABLE 2B

[0083] The below table includes possible drug targets for the various subtypes and non exclusive examples of drugs for those targets.

TABLE 3

[0084] The terms “lymphoma” and “lymphoid malignancy” as used herein refer to malignant tumors derived from lymphocytes and lymphoblasts. The phrase may refer to a broad lymphoma class (e.g., DLBCL, FL, MCL, etc.) or to a subgroup falling within a broad lymphoma class (e.g., GCB DLBCL, ABC DLBCL). Examples of lymphomas include, but are not limited to, follicular lymphoma (FL), Burkitt lymphoma (BL), mantle cell lymphoma (MCL), follicular hyperplasia (FH), small cell lymphocytic lymphoma (SLL), mucosa-associated lymphoid tissue lymphoma (MALT), splenic lymphoma, multiple myeloma, lymphoplasmacytic lymphoma, post-transplant lymphoproliferative disorder (PTLD), lymphoblastic lymphoma, nodal marginal zone lymphoma (NMZ), germinal center B cell-like diffuse large B cell lymphoma (GCB), activated B cell-like diffuse large B cell lymphoma (ABC) and primary mediastinal B cell lymphoma (PMBL), primary central nervous system lymphoma (PCNSL) and Waldenstrom’s macroglobulinemia (WM).

[0085] The terms “treat,” “inhibit,” and “prevent” as well as words stemming therefrom, as used herein, do not necessarily imply 100% or complete treatment, inhibition, or prevention. Rather, there are varying degrees of treatment, inhibition, or prevention of which one of ordinary skill in the art recognizes as having a potential benefit or therapeutic effect. In this respect, the inventive methods can provide any amount of any level of treatment. Furthermore, the treatment or prevention provided by the inventive method can include treatment or prevention of one or more conditions or symptoms of the condition, e.g., cancer, being treated or prevented. For example, treatment or prevention can include promoting the regression of a tumor. Also, for purposes herein, “prevention” can encompass delaying the onset of the condition, or a symptom or condition thereof.

[0086] An “effective amount” or “an amount effective to treat” refers to a dose that is adequate to prevent or treat cancer in an individual. Amounts effective for a therapeutic or prophylactic use will depend on, for example, the stage and severity of the disease or disorder being treated, the age, weight, and general state of health of the patient, and the judgment of the prescribing physician. The size of the dose will also be determined by the active selected, method of administration, timing and frequency of administration, the existence, nature, and extent of any adverse side-effects that might accompany the administration of a particular active, and the desired physiological effect. It will be appreciated by one of skill in the art that various diseases or disorders could require prolonged treatment involving multiple administrations, perhaps using the inhibitors in each or various rounds of administration. [0087] The following are certain aspects of the invention.

[0088] 1. A method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, without use of gene copy number, the method comprising:

(a) obtaining from the subject a biopsy sample of the lymphoma;

(b) (i) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample one or more genetic features of the subtype Nl listed in Table 1A, or

(ii) detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype Nl listed in Table IB;

(c) determining whether at least two weight values listed in the Table 1 A or IB are positive for each of the subtypes BN2, EZB, MCD, and ST2 based on the genetic features detected in (b);

(d) calculating the probabilities that the sample belongs to

(i) subtype Nl, and

(ii) each subtype BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation

wherein P( subtype] V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 1 A or IB based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate P(subtype|E), and wherein P{ subtype] V) is calculated for each subtype using only V, weight values for that subtype;

(e) designating each subtype BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have P(subtype|E) < 0.5; and

(f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic: (1) if only one subtype has /’(subtype) V) > 0.9 with the other subtypes having /’(subtype) V) < 0.9, the sample is classified as being the one subtype having /’(subtype) V) > 0.9;

(2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype)F) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and

(3) if there is more than one subtype having /’(subtype) V) > 0.9, the sample is classified as being each of the subtypes having /’(subtype) V) > 0.9.

[0089] 2. The method of aspect 1, wherein the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1A.

[0090] 3. The method of aspect 2, wherein the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample all of the genetic features of the subtype N1 listed in Table

IA.

[0091] 4. The method of aspect 1, wherein the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB.

[0092] 5. The method of aspect 4, wherein the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample all of the genetic features of the subtype N1 listed in Table

IB.

[0093] 6. A method of treating a human subject with diffuse large B cell lymphoma

(DLBCL), the method comprising:

(A) classifying a DLBCL according to any one of aspects 1-5; and

(B) treating the subject based on the one or more classifications of (A). [0094] 7. The method of aspect 6, wherein the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

[0095] 8. The method of aspect 6, wherein the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

[0096] 9. The method of aspect 6, wherein the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

[0097] 10. The method of aspect 6, wherein the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1 or an immune checkpoint inhibitor.

[0098] 11. The method of aspect 6, wherein the DLBCL is classified as ST2, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl. [0099] 12. A method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, with use of gene copy number, the method comprising:

(a) obtaining from the subject a biopsy sample of the lymphoma;

(b) (i) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype Nl listed in Table 2A, or

(ii) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype Nl listed in Table 2B, wherein a copy number of each of the subtypes is a genetic feature detected;

(c) determining whether at least two weight values listed in the Table 2A or 2B are positive for each of the subtypes A53, BN2, EZB, MCD, and ST2 based on the genetic features detected in (b);

(d) calculating the probabilities that the sample belongs to

(i) subtype Nl, and

(ii) each subtype A53, BN2, EZB, MCD, and ST2 determined in (c) to have at least two positive weight values using the equation e^xp(åi )

/(subtype |E) =

1 + exp(å_j ) wherein /( subtype] V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 2A or 2B based on the genetic features detected in (b) for that subtype, wherein i is the number of weight values used to calculate /(subtype) E), and wherein /( subtype] V) is calculated for each subtype using only V, weight values for that subtype;

(e) designating each subtype A53, BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (c) to have /( subtype] V) < 0.5; and

(f) classifying the sample with the probabilities calculated in (d) and any probabilities designated in (e) using the logic:

(1) if only one subtype has /(subtype) E) > 0.9 with the other subtypes having /(subtype) E) < 0.9, the sample is classified as being the one subtype having /(subtype) E) > 0.9;

(2) if only one subtype has /(subtype) E) > 0.5 with the other subtypes having /(subtype)E) < 0.5, the sample is classified as being the one subtype having /(subtype) E) > 0.5; and

[0100] 13. The method of aspect 12, wherein the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2A.

[0101] 14. The method of aspect 13, wherein the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2A.

[0102] 15. The method of aspect 12, wherein the detection of (b) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2B.

[0103] 16. The method of aspect 15, wherein the detection of (b) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2B.

[0104] 17. A method of treating a human subject with diffuse large B cell lymphoma

(DLBCL), the method comprising:

(A) classifying a DLBCL according to any one of aspects 12-16; and

(B) treating the subject based on the one or more classifications of (A).

[0105] 18. The method of aspect 17, wherein the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

[0106] 19. The method of aspect 17, wherein the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

[0107] 20. The method of aspect 17, wherein the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

[0108] 21. The method of aspect 17, wherein the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1 or an immune checkpoint inhibitor.

[0109] 22. The method of aspect 17, wherein the DLBCL is classified as ST2, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl. [0110] 23. The method of aspect 17, wherein the DLBCL is classified as A53, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

[0111] It shall be noted that the preceding are merely examples of embodiments. Other exemplary embodiments are apparent from the entirety of the description herein. It will also be understood by one of ordinary skill in the art that each of these embodiments may be used in various combinations with the other embodiments provided herein. [0112] The following examples further illustrate the invention but, of course, should not be construed as in any way limiting its scope.

EXAMPLE

[0113] The LymphGen algorithm was created to provide a probabilistic classification of a tumor from an individual patient into a genetic subtype. A genetic subtype was defined as a group of tumors that are enriched for genetic aberrations in a set of subtype predictor genes. These subtype predictor genes are identified by considering each possible combination of genetic aberrations (i.e. mutations, copy number alterations, or fusions) as a separate genetic “feature”, and scoring a tumor as positive if one or more of a feature’s genetic aberrations is observed. LymphGen uses the presence or absence of each subtype predictor feature to provide a probability that a tumor belongs to the subtype.

[0114] To implement LymphGen in DLBCL, what was first needed was to define the genetic subtypes to which a tumor could be assigned. For this subtype discovery effort, the GenClass algorithm (Schmitz et al., 2018) was used, which begins with a set of “seed” tumor subsets and iteratively assorts tumors into and out of these seeds, converging on a classification that maximizes the genetic distinctiveness of the resulting subtypes (Figure 1 A). First, seeds representing the four previously identified genetic subtypes were chosen: MCD (including ATYD88^L265F and CD79B mutations), BN2 (including BCL6 translocations and NOTCH2 mutations), N1 (including NOTCH 1 mutations), and EZB (including EZH2 mutations and BCL2 translocations). Within the remaining cases in the cohort (hereafter termed the “NCI cohort”), TP53 was the most frequently mutated gene (25.2%) that was not also significantly enriched in one of the previous subtypes. TP53 inactivation has been previously associated with aneuploidy in DLBCL (Bea et al., 2004; Chapuy et al., 2018; Monti et al., 2012). In the NCI cohort, tumors with a homozygous TP 53 deletion (5.9%) or the combination of a heterozygous TP 53 deletion and a TP 53 mutation (8.7%) had the most aneuploidy, as assessed by the number of gains and losses of chromosomal segments. A seed class was formed from these ZP53-altered cases, which was termed “A53” (aneuploid with TP53 inactivation). Mutations in TET2 , P2RY8 , and SGK1 were also recurrently mutated among the genetically unassigned cases (10.1%, 6.9%, and 6.9% of cases, respectively), with the majority (54%) of SGK1 mutations truncating the protein. TET2 mutations were significantly associated with truncating SGK1 mutations (P=0.001), and with P2RY8 mutations (P=0.033), leading us to create a second new seed class based on these features, which was termed “ST2” ( SGK1 and TET2 mutated). Using the 6 seed classes defined above, the GenClass algorithm assigned 54% of the cases.

[0115] LymphGen next develops a separate Bayesian predictor model for each GenClass subtype, which determines the probability that a tumor belongs to the subtype based on its genetic features (Figure 1 A). The algorithm defines subtype predictor features that distinguish the subtype from all other cases (P<0.001, Fisher’s exact test, prevalence > 0.2), and uses the prevalence of the feature in the subtype and its prevalence in other cases to estimate the likelihood that a tumor with that feature belongs to the subtype. These likelihood estimates are then used in Bayes formula to calculate the probability that an individual tumor belongs to a subtype based on its constellation of genetic features. Thus, for each DLBCL tumor, LymphGen calculates six probabilities, one for each GenClass-defmed subtype. Tumors with subtype probabilities of >90% or 50%-90% were defined a “core” or “extended” subtype members, respectively. Some tumors were core members of more than one subtype and were termed “genetically composite” (Figures 9A-9L). In the NCI cohort, the LymphGen algorithm identified 47.6% core cases, 9.8% extended cases, and 5.7% genetically composite cases (Figure IB). Altogether 329 (63.1%) of the 574 cases in the NCI cohort were classified, which is substantially greater than the 46.6% classified previously (Schmitz et ah, 2018) (Figures IB, 1C). The inability of LymphGen to classify the remaining cases stemmed from three issues: some tumors had a small number of features characteristic of one or more subtype, some had unique features that were not recurrent in DLBCL, and others had very few genetic features altogether.

[0116] In the resulting genetic taxonomy, each of the DLBCL gene expression subgroups was comprised of multiple genetic subtypes, with ABC tumors enriched for MCD, GCB tumors enriched for EZB and ST2, and Unclassified tumors enriched for BN2 (Figure ID, Figure 7A). Conversely, some genetic subtypes were largely comprised of tumors belonging to the same gene expression subgroup (MCD, Nl, EZB), while others were comprised of different gene expression subgroups, with BN2, A53 and ST2 being the most phenotypically diverse. Genetic attributes of DLBCL subtypes

[0117] A set of subtype-distinguishing genetic features that were significantly associated with a subtype (P<0.01) was selected and was identified in >10% of that subtype (Figures 2A- 2F, 10A-10D). Many subtype- defining mutations are likely due to AID-dependent somatic hypermutation (Schmitz et al., 2018), which produced truncating mutations in subtype-specific tumor suppressor genes (e.g. PRDMl , ETV6 , TOX, HLA-A , HLA-B , HLA-C in MCD, TNFRSF14 in EZB and NFKBIA in ST2).

[0118] A FYD88^L265F and CD79B mutations, the genetic hallmark of MCD, cooperatively activate NF-AB via the My-T-BCR supercomplex involving MYD88, TLR9 and the BCR (Phelan et al., 2018). Deletions of CDKN2A , encoding the cell cycle inhibitor pl6, likely accelerates proliferation in MCD tumors, and their viability is sustained by BCL2, which is upregulated epigenetically and by copy number gain/amplification (Figure 10E). MCD tumors evade immune surveillance in at least 72.5% of cases by acquiring homozygous deletions, truncating mutations or translocations resulting in: 1) Reduced antigen presentation due to inactivation of MHC class I or TAPI, a transporter that loads peptides onto MHC class I; 2) Decreased T cell activation due to gene fusions that elevate expression of CD274 and PDCD1LG2 , encoding PD-L1 and PD-L2, respectively; 3) Diminished NK activation due to CD58 inactivation (Challa-Malladi et al., 2011) (Figures 2A-2F, 3B).

[0119] BN2 is characterized by mutations that activate NOTCH2 and inactivate SPEN, a NOTCH antagonist, in 50% of tumors, 72% of which also have a BCL6 translocation. Mutations targeting components of the BCR-dependent NF-AB pathway (PRKCB, BCL10, TNFAIP3, TNIPl) are also prominent BN2 features, suggesting these tumors rely on BCR signaling for survival (see below). Interactions with NK cells and T cells are potentially compromised in BN2 by CD70 deletions. CCND3 mutations (Schmitz et al., 2012) foster vigorous proliferation in BN2.

[0120] Epigenetic dysregulation is a defining attribute of the EZB, involving inactivation of several epigenetic regulators (KMT2D, CREBBP, EP300, ARIDIA, IRF8, MEF2B, EBF1) and activation of EZH2, thereby altering germinal center (GC) B-cell differentiation (Mlynarczyk et al., 2019; Pasqualucci and Dalla-Favera, 2018). GC B-cell migration and signaling is altered by inactivation of the S1PR2/GNA13 pathway in EZB (Muppidi et al., 2014). PI3 kinase signaling in EZB is promoted by inactivating mutations and deletions of PTEN and MIR17HG amplification, encoding microRNAs that decrease PTEN expression. Recurrent PEL amplification may deregulate EZB metabolism and growth, as in normal GC B-cells (Heise et al., 2014). MHC class II expression and function in EZB is compromised by EZH2 activation (Ennishi et al., 2019b), CIITA inactivation (Steidl et al., 2011), and inactivation of HLA-DMB, which facilitates peptide loading onto MHC class II. This may perturb interactions between lymphoma cells and T follicular helper cells (TFH), as can TNFRSF14 inactivation (Mintz et al., 2019). STAT6 mutations may modulate the ability of T_FH-derived IL-4 to promote plasma cell differentiation (Weinstein et al., 2016).

[0121] ST2 is named for its enrichment in SGK1 and TET2 mutations. TET2 is an epigenetic regulator that catalyzes the hydroxylation of methylated cytosines in DNA. ST2 tumors acquire TET2 truncating mutations suggesting a tumor suppressor function, as in mouse GC B-cell lymphomagenesis (Dominguez et al., 2018). Inactivating SGK1 mutations also suggest a tumor suppressor function, possibly by modulating PI3 kinase signaling (Di Cristofano, 2017). JAK/STAT signaling is promoted in ST2 by inactivation of SOCS1, a JAK signaling inhibitor (Linossi and Nicholson, 2015), inactivation of DUSP2, a phosphatase for STAT3 (Lu et al., 2015), and by known STAT3 -activating mutations (Y640F, D661Y) (Crescenzo et al., 2015). Inactivating mutations in ST2 targeting P2RY8, a 7-transmembrane receptor, and its signaling mediator GNA13, prevent responses to S-geranylgeranyl-l-glutathione, which spatially confines normal GC B-cells, and inhibit ART activity (Lu et al., 2019). Finally, ST2 tumors foster NF-AB signaling by inactivating NFKBIA , encoding the NF-AB inhibitor lABer(Baeuerle and Baltimore, 1988).

[0122] A53 is characterized by TP53 mutations and deletions, long known to play a role in

DLBCL pathogenesis and prognosis (Monti et al., 2012; Young et al., 2007). A53 tumors also acquire homozygous deletions and mutations targeting 53BP1 ( TP53BP1 ), a DNA damage sensor that prevents aneuploidy (Celeste et al., 2002), consistent with the recurrent gains and losses of chromosome arms in A53 (Figures 2A-2F). Some A53 abnormalities have been associated with ABC DLBCL, including deletion of 6q, harboring the tumor suppressors TNFAIP3 and PRDM1, gain/amplification of 3q (Lenz et al., 2008b), focal amplification of NFKBIZ (Nogai et al., 2013), encoding an NF-AB co-activator, amplification of CNPY3, encoding a TLR9 trafficking regulator (Phelan et al., 2018), and BCL2 amplification (Figure 10G). Additional focal deletions target the tumor suppressors p73, a p53 family member, and ING1, a component of several epigenetic regulator complexes (Tallen and Riabowol, 2014). Finally, A53 tumors frequently delete or mutationally inactivate bΐ -microglobulin ( B2M ), providing a mechanism of escape from immune surveillance (Challa-Malladi et al., 2011).

[0123] N1 is characterized by gain-of-function NOTCH1 mutations, similar to those chronic lymphocytic leukemia and mantle cell lymphoma. These tumors additionally acquire mutations targeting B-cell differentiation regulators (ID3, BCOR) and IAB kinase b ( IKBKB ), including the V203I isoform that constitutively activates NF-AB (Cardinez et al., 2018).

Relationship of DLBCL genetic subtypes to other lymphoid malignancies [0124] While nodal DLBCLs typically present clinically in lymph nodes and other primary immune tissues, primary extranodal lymphomas present as tumors involving various non lymphoid organs. Primary extranodal lymphomas frequently acquire MFDSS^L265P and/or CD79B mutations (Figure 3B), as well as other MCD-defming mutations (Figure 3D). These lymphomas often arise in the central nervous system (CNS), ocular vitreo-retina, and testis, all considered “immune- privileged” sites because they tolerate allografts and permit only selective access by immune cells (Shechter et al., 2013). MCD, which arises in lymph nodes, spreads secondarily to extranodal sites in 30% of cases, and 46% of these occur at four sites that can give rise to primary extranodal lymphomas - the CNS, vitreo-retina, testis and breast - whereas other DLBCL subtypes spread to these sites significantly less often (P=5.0E-5; Figure 3C).

[0125] Genetic aberrations of several DLBCL subtypes reveal potential pathogenetic relationships with more indolent lymphomas. Mutations characteristic of BN2 link this subtype to marginal zone lymphomas (MZLs) (Figure 3E), befitting the essential role of NOTCH2 in the differentiation of follicular B-cells to marginal zone B-cells (Saito et al., 2003). BCL6 translocations, which characterizes BN2, are rare (2%) in indolent MZLs but common (50%) in MZLs that have transformed into aggressive large cell variants (Flossbach et al., 2011; Ye et al., 2008). Follicular lymphoma (FL), a frequently indolent disease, shares many genetic lesions with EZB, as does transformed FL that can histologically resemble DLBCL (Figure 3F). The genetic signature of ST2 betrayed an intriguing similarity with two histologically distinct lymphomas, nodular lymphocyte- predominant Hodgkin lymphoma (NLPHL) and T cell histiocyte-rich large B-cell lymphoma (THRLBCL) (Figure 3G). NLPHL is an indolent Hodgkin lymphoma variant that retains expression of GC B-cell genes, unlike other Hodgkin lymphomas, and can transform into an aggressive large cell form (Timens et ah, 1986). Morphological similarities between some cases of NLPHL and THRLBCL led pathologists to suspect a link between these entities, which was supported by their shared mutations in SOCS1 , DUSP2, SGK1 , and JUNB , which are all ST2-defming genes.

Validation of the LymphGen classification

[0126] To evaluate the reproducibility of the LymphGen algorithm in identifying genetic subtypes of DLBCL, it was used to assign tumors from two validation cohorts to genetic subtypes (Figures 10H-10R). The first (n=304) was used previously to identify DLBCL subtypes (denoted “Harvard”) and was analyzed by DNA sequencing for mutations, copy number in selected genomic regions, and BCL2/BCL6 rearrangements (Chapuy et ak, 2018). A second cohort (n=332) was used previously to identify signatures of poor prognosis in DLBCL (denoted “BCC”) and was analyzed here for mutations in 82 lymphoma-associated genes by targeted resequencing and for whole genome copy number aberrations (Ennishi et ak, 2019a).

[0127] Because each of the DLBCL cohorts had different data types available, LymphGen was designed to function using various combinations of mutational data (whole exome or gene panel resequencing), copy number data (regional or whole genome) and rearrangement data for BCL2 and BCL6. LymphGen has robust performance with varying genetic inputs (Figures 11 A- 1 ID).

[0128] To compare the LymphGen-assigned subtypes between the cohorts, the cohorts were first normalized to be equivalent to a population-based cohort with respect to overall COO composition (Scott et ak, 2015), given the known COO relationships of the genetic subtypes (Chapuy et ak, 2018; Schmitz et ak, 2018). In these normalized cohorts, the prevalences of each genetic subtype were roughly comparable (Figure 4A), and as were their COO compositions (Figures 4B-4D). Moreover, in the Harvard cohort, each LymphGen subtype was drawn predominantly from a single genetic “cluster”, as defined previously (Chapuy et ak, 2018), with a 75% overall agreement between the analytic methodologies (Figures 1 IE and 1 IF). [0129] The genetic features associated with each subtype in the three cohorts were generally comparable in prevalence (Figures 4E and 4F). To evaluate this similarity statistically, LymphGen subtypes were iteratively created using sets of mutational features in which one subtype-determining feature was omitted, and used the prevalence of the omitted feature within the resulting subtype to estimate the significance of its subtype association. These feature statistics were combined for a given subtype into a joint probability assessing the genetic similarity of that subtype among the cohorts (see Methods). Significant similarity was observed in the genetic compositions of MCD, BN2, EZB and ST2 (p<6.3E-7) while N1 could not be evaluated by this method due to the statistical dominance of NOTCH1 mutations. For A53, which is defined primarily by copy number alterations and not mutations, the association of TP53 mutation and/or deletion with aneuploidy was evaluated, as measured by the number of copy number alterations in each case, revealing significant genetic similarity between the A53 cases in each cohort (p<4.4E-9).

[0130] Next evaluated was the survival of patients in LymphGen subtypes in each cohort. The overall survival characteristics of the cohorts were distinct, as judged by COO associations, most likely reflecting accrual bias. Nonetheless, the genetic subtypes defined in each cohort had similar associations with overall survival, as judged by the Kaplan-Meyer method (Figures 4G- 4M), or by hazard ratios (Figure 4N). Within each cohort, MCD had an inferior survival, especially when compared with ST2 and BN2. Among ABC cases in each cohort, BN2 was favorable, especially when compared to MCD and A53. Among GCB cases, EZB had an inferior survival compared with ST2. Given these consistent survival trends, data were used from all three cohorts to estimate joint hazard ratios (Figure 4N). In this model, the survival of MCD was inferior compared with ST2, BN2 and all non-MCD patients (p<0.001), the survival of BN2 was favorable compared to MCD, A53 and to non-BN2 patients within ABC (p<0.01), and the survival of EZB was inferior compared to ST2 within GCB (P=0.032).

[0131] While the genetic subtypes clearly subdivided the outcomes within the ABC and GCB gene expression subgroups, the reverse was also true. Within BN2 and A53, the COO subgroups had significantly disparate survival characteristics, demonstrating that tumor genotype and phenotype should both be considered when attempting to understand the response to therapy. Phenotypic distinctions between genetic subtypes

[0132] Gene expression signatures offer glimpses into tumor phenotypes that are differentially manifested in DLBCL genetic subtypes (Schmitz et al., 2018). Signatures previously associated with MCD, BN2, Nl, and EZB were similarly associated with their corresponding LymphGen versions (Figures 5A-5E, 7A). Of potential therapeutic relevance, MCD and BN2 were characterized by signatures of BCR-dependent NF-AB activation, while EZB and ST2 were characterized by a PI3 kinase signaling signature. ST2 also expressed genes characteristic of GC B cells, JAK2 signaling, and glycolytic pathway activation as well as the Stromal-1 signature, which has been associated with favorable survival in DLBCL (Lenz et al., 2008a). A53 expressed p53 target genes at low levels but were unexpectedly devoid of multiple immune cell subpopulations.

[0133] Two gene expression signatures were recently defined, termed MHG (Sha et al., 2019) and DHIT (Ennishi et al., 2019a), that identify DLBCL patients with an inferior prognosis beyond the survival distinctions associated with COO. Since these two signatures were highly correlated in the NCI cohort (P=1.5E-14), DHIT was a focus, which was defined in GCB tumors with both BCL2 and MYC rearrangements and identifies GCB patients with an inferior prognosis (Ennishi et al., 2019a).

[0134] Using the NCI cohort, the relationship of DHIT to other gene expression signatures and genetic features was investigated. In GCB cases, EZB had significantly higher DHIT scores than other subtypes (P=0.002; Figure 6A). Among 30 DHIT+ GCB cases, the majority (70%) were either EZB or genetically composite cases with features of both EZB and A53 (Figure 6B). [0135] The gene expression signatures most correlated with DHIT were two that distinguish BL from DLBCL: the MHG signature (r=0.56) and GCB-4 (r=0.54), defined as the subset of GC B-cell signature genes that are expressed more highly in BL than in DLBCL (Dave et al., 2006). Signatures of intermediate zone and dark zone GC B-cells (GCB-9 and GCB-10, respectively; (Milpied et al., 2018)) were also correlated (r=0.44 and r=0.40, respectively), suggesting that DHIT reflects dynamic changes in GC B-cell differentiation. DHIT was also correlated with signatures reflecting MYC activity, notably MYCUp-4 (r=0.46), consisting of genes that are induced by MYC and are direct MYC binding targets (Zeller et al., 2006). DHIT also correlated (r=0.48) with a signature of adverse survival in DLBCL that includes MYC and its target genes GNL3 and NPM3 (Prolif-6) (Rosenwald et al., 2002).

[0136] It was hypothesized that DHIT is a composite signature that reflects both GC B-cell differentiation and MYC activity, and used GCB-4 and My cUp-4 to represent these phenotypes, given their association with DHIT by Gene Set Enrichment Analysis (p=2.6E-3 and 1.0E-4, respectively; Figure 6C). A linear model combining these signatures was significantly correlated with DHIT (P=2.2E-16; Figure 6D) and accounted for 60.2% of DHIT variance among EZB cases within GCB.

[0137] Within GCB, the survival of the DHIT- subset was better than DHIT+ (P=0.0111), in part reflecting the enrichment of DHIT- cases with ST2, BN2 or A53, subtypes with a favorable survival in GCB. Within the EZB subset of GCB, the survival of the DHIT+ subset was significantly worse than DHIT- (P=0.036), which was not true in the non-EZB subset (Figure 6E and 6F).

[0138] The association of genetic features with the DHIT+ and DHIT- subsets of EZB GCB cases was next explored (Figure 6F). The majority of EZB-defming genetic features were observed comparably in these two subsets with the exception of GNA13 mutations, which were more prevalent among DHIT+ than DHIT- cases (P=0.025). In keeping with a role for MYC in DHIT+ cases, MYC rearrangements, amplifications and mutations were significantly more common in DHIT+ than DHIT-cases (P=0.0079) and were also recurrent in non-EZB DHIT+ GCB cases. Mutations or double deletions of TP53 were more than twice as prevalent in DHIT+ than DHIT- cases (P=0.05), while the tumor suppressor DDX3X was mutated in a third of DHIT+ tumors but never in DHIT- tumors (P=2.5E-4). FOXOl, a transcription factor that is inactivated by PI3 kinase signaling, was targeted by mutations more than 3 times as often in DHIT+ than DHIT- cases.

[0139] Genetic features that were significantly more common in DHIT-cases included mutations targeting the NF-AB regulators A20 ( TNFAIP3 ) and CARD 11, as well as deletions of the tumor suppressor TP73. Given the association oiMYC abnormalities with DHIT+ EZB cases, hereafter DHIT+ EZB will be herefered to as “EZB-MYC+” and DHIT- EZB as “EZB-MYC-”. The genetic data were used to create a probabilistic model of EZB-MYC+ versus EZB-MYC- that could distinguish these subtypes effectively (permutation p-value=0.004, see Methods). [0140] Genes that were preferentially mutated in EZB-MYC+ cases were also frequently mutated in BL whereas those that were preferentially mutated in EZB-MYC- cases were not (Figure 6H). However, some genetic aberrations that define BL, such as ID3 and TCF3 mutations (Schmitz et al., 2012), were not observed in EZB-MYC+. Thus, the genetic program adopted by EZB-MYC+ is shared by BL, but these lymphomas are genetically distinct.

Functional genomics of DLBCL genetic subtypes

[0141] Next considered was whether the DLBCL genetic subtypes might offer insights into the response to targeted therapy. First investigated was BCR signaling since this pathway is altered by multiple genetic aberrations in DLBCL (Schmitz et al., 2018). Figures 7B and 7C depict the recurrent genetic lesions targeting BCR-dependent NF-AB activation and the relative number of cases in each genetic subtype with these aberrations. Each genetic subtype acquired potentially inactivating mutations and deletions in known negative regulators of proximal BCR signaling, suggesting that BCR signaling has a pervasive influence on lymphomagenesis (Figures 7A-7C). The genetic subtypes differed significantly in genetic aberrations that modify proximal BCR signaling, with mutations targeting the CD79B subunit of the BCR confined to the MCD, BN2 and A53 subtypes. By contrast, mutations targeting CD79A were enriched in EZB, in keeping with the distinct contributions of these subunits to BCR signaling and endocytic recycling (Busman-Sahay et al., 2013). MCD tumors were enriched in the M YD 88^L265P mutation, a hallmark of tumors in which the My-T-BCR supercomplex activates NF-AB at an endolysosomal location (Phelan et al., 2018). By contrast BN2 tumors acquire mutations that target the CBM signaling adapter complex ( PRKCB , BCL10 , and TRAF6 mutations) or impair two negative regulators of I/cB kinase, A20 ( TNFAIP3 ) and TNIP1. In aggregate, the BN2 and MCD subtypes had the highest prevalence of genetic lesions altering BCR signaling components (-63%), while the N1 and EZB subtypes had the lowest prevalence (Figure 7B).

[0142] The survival of ABC cell lines relies on engagement of autoreactive BCRs by self antigens, whereas GCB models rely on a “toncogenic” form of BCR signaling that is antigen- independent (Havranek et al., 2017; Young et al., 2015). One known autoreactive immunoglobulin (Ig) heavy chain variable (V_H) region, V_H4-34, binds to N-acetyl-lactosamine on self-glycoproteins and is enriched among ABC tumors (Young et al., 2015). The expressed IgV_H regions were reassembled using RNA-seq data from the NCI cohort, and observed that V_H4-34 was the dominant IgV_H region in MCD, BN2 and A53, providing evidence that these subtypes may rely upon self-antigen- dependent chronic active BCR signaling (Figure 7D-7I). Consistent with this hypothesis, these subtypes most often utilized IgM BCRs, which in normal B cells promote proliferation while IgGBCRs promote plasmacytic differentiation (Dogan et al., 2009).

[0143] To functionally evaluate the BCR pathway in DLBCL, cell lines that bear genetic hallmarks of the genetic subtypes were used. First investigated was whether the negative regulators of proximal BCR signaling that are genetically inactivated in DLBCL have an effect on chronic active BCR signaling in DLBCL models. The relative ability of cells to survive in the presence of sub-maximal concentrations of the BTK inhibitor ibrutinib was assayed as an effective proxy for BCR signaling strength, as described previously (Wilson et al., 2015b). In Cas9-expressing models of MCD and BN2, various BCR negative regulators were knocked out by expressing short guide RNAs (sgRNAs) together with GFP, and quantified the relative numbers of live, GFP+/sgRNA+ cells in the presence of ibrutinib compared with DMSO-treated cultures. Knockdown of each BCR negative regulator promoted survival in ibrutinib, as did knockdown of the NF-AB negative regulators A20 ( TNFAIP3 ) and TNIPl, whereas a control sgRNA had no effect (Figure 7J).

[0144] To investigate essential pathways in the genetic subtypes, whole genome loss- of- function CRISPR screens in Cas9-expressing models of MCD (TMD8, HBL1, OCI-LylO), BN2 (Riva), and EZB (OCI-Lyl, SUDHL4, WSU-DLCL2) was performed, as described (Phelan et al., 2018). For each gene targeted by the sgRNA library, a CRISPR screen score was calculated (see Methods), with negative scores indicating an essential gene. Based on this metric, MCD and BN2 models strongly depended on the BCR subunits CD79A and CD79B, with OCI-Lyl displaying a lesser, but evident, dependence (Figures 7K-7S). Downstream of the BCR, the signaling proteins that engage the NF-AB pathway by activating IAB kinase were selectively essential in the MCD and BN2 models, as was IAB kinase itself. Of particular note, the target of the drug ibrutinib, BTK, was essential in the MCD and BN2 models but not in the EZB models (Figures 7K-7S). Previous studies of ibrutinib in DLBCL have focused on the intense addiction of MCD models to BCR signaling (Davis et al., 2010; Lionakis et al., 2017; Phelan et al., 2018; Wilson et al., 2015b), but the contribution of BCR signaling in BN2 was not anticipated. Constitutive BCR signaling in the BN2 model was confirmed by knockdown of IgM or CD79A, which decreased phosphorylation of Src-family kinases, SYK and BTK (Figure 7T).

Accordingly, the growth of Riva xenografts was strongly suppressed by low doses of ibrutinib (Figure 7U), whereas similar treatment of MCD xenografts only modestly suppressed growth (Figures llG and 11H).

[0145] The whole genome CRISPR data was further used to predict the dependency of MCD, BN2 and EZB on other signaling and regulatory pathways that can by targeted by clinically available drugs (Figures 7K-7S; see Discussion). Two interacting transcription factors - IRF4 and SPIB - that regulate genes characteristic of ABC DLBCL were selectively essential in the MCD and BN2 but not EZB models (Yang et al., 2012). MCD models selectively depended on the IL-10 receptor a and b subunits, JAK1 and STAT3, consistent with autocrine IL-10 signaling in this subtype (Lam et al., 2008; Rui et al., 2016; Van Den Neste et al., 2018). The PRC2 chromatin repressor complex - including EZH2, SUZ12 and EED - was especially essential for the EZB models, all of which had gain-of-function EZH2 mutations. The PI3 kinase pathway, which can be activated by the BCR by different mechanisms in ABC and GCB models (Young et al., 2019), was essential in models of all three subtypes. BCL2 was also essential in all three genetic subtypes, whereas BCL-XL (BCL2L1) was selectively required in the MCD and BN2 models.

Discussion

[0146] The extreme genetic and phenotypic heterogeneity of DLBCL presents a challenge to the development of precision therapies. Here is provided a genetic framework in which to understand therapeutic responses in subsets of DLBCL tumors defined by shared pathogenesis. The LymphGen classification unifies two recent genetic profiling studies (Chapuy et al., 2018; Schmitz et al., 2018) and was also evident in the independent BCC cohort. This classification breaks DLBCL into seven genetic subtypes that differ with respect to oncogenic pathway engagement, gene expression phenotype, tumor microenvironment, survival rates following R- CHOP immunochemotherapy, and potential therapeutic targets (Figure 8A). As such, this taxonomy provides a framework for understanding the biologic diversity that is encompassed within the pathological diagnosis of DLBCL and will likely shed light on the heterogenous responses of DLBCL to cytotoxic and molecularly targeted therapies.

[0147] The investigation provides insight into the mechanisms by which a DLBCL genetic subtype acquires a shared genetic program (Figure 8B). In one model, the epigenetic nature of the cell-of- origin for a subtype necessitates certain oncogenic events that endow the B-cell precursor with the hallmarks of cancer, including unlimited proliferation and survival. The differential acquisition of genetic aberrations by GCB and ABC DLBCL, which originate from GC and post-GC cells, respectively, may fit this model (Lenz et ah, 2008b; Schmitz et ah, 2018). Alternatively, a precursor B-cell could randomly acquire a “founder” genetic lesion, the nature of which dictates the subsequent selection of secondary genetic lesions. For example, overexpression of Myc kills normal cells unless they also have lesions that prevent cell death, such as a BCL2 translocation (Evan et ak, 1992). The probabilistic approach raises a third, hybrid possibility. A substantial subset of DLBCL tumors (5.7%) had a high probability of belonging to more than one genetic subtype. This suggests a model in which one genetic program is initially acquired by a tumor and then a second is subsequently acquired because it confers an additional favorable phenotypic attribute (Figure 8B).

[0148] The DLBCL genetic subtypes have intriguing similarities to more indolent lymphoma types: BN2 resembles MZLs, EZB resembles FL, and ST2 resembles both NLPHL and THRLBCL. Three models could account for these genetic relationships (Figure 8C). A “direct evolution” model suggests that some DLBCL patients have a concurrent but undiagnosed low- grade malignancy that acquires additional genetic lesions, transforming it into an aggressive DLBCL subtype. Consistent with this model, pathologists recognize histologically “composite lymphomas” that have, at diagnosis of DLBCL, evidence of a concurrent low-grade lymphoma in the same biopsy (Kuppers et ak, 2014). For example, in composite lymphomas with both marginal zone and large cell components, the large cells frequently acquire BCL6 translocations, as is typical of BN2 (Flossbach et ak, 2011). A “branched evolution” model posits the existence of a pre-malignant B-cell clone that can become an indolent lymphoma or a DLBCL, depending on the nature of additional genetic alterations it acquires. In some cases of transformed FL, for example, the transformed lymphoma shares some genetic features with the antecedent FL, while each lymphoma type has genetic attributes not shared by the other (Green et ak, 2013). A final “convergent evolution” model suggests that indolent lymphomas and DLBCL subtypes can separately select the same genetic program to acquire a particular oncogenic phenotype while differing in other attributes. Future genetic studies of composite and transformed lymphomas may shed light on these evolutionary models.

[0149] Sequential tumor evolution is perhaps best exemplified by EZB-MYC+ and EZB- MYC-, which emerged from the analysis of the DHIT signature (Ennishi et al., 2019a). The poor survival associated previously with DHIT+ lymphoma is not only due to the adverse survival of EZB-MYC+ cases but also to the enrichment of the DHIT- subset with ST2, BN2 and A53, which all have excellent outcomes among GCB cases. DHIT+ EZB cases were enriched for aberrations targeting MYC and four other genes, all of which are also frequently mutated in BL. These tumors also expressed genes that are characteristic of normal GC B-cells and are more highly expressed in BL than DLBCL. These genetic and phenotypic associations suggest that EZB-MYC+ should be considered a seventh genetic subtype of DLBCL that arises from EZB- MYC- tumors with the acquisition of these genetic lesions (Figures 8A, 8D). EZB-MYC+ cases expressed genes that are both bound and transactivated by MYC in B-cells (Zeller et al., 2006), consistent with additional cryptic abnormalities that deregulate MYC, as described (Hilton et al., 2019), or enhance its function. Since most EZB-MYC+ tumors had a BCL2 fusion, this subtype may account in large measure for the adverse survival of “double hit” lymphomas.

[0150] Another intriguing genetic relationship links the MCD subtype to primary extranodal lymphomas, including those involving the CNS, vitreo-retina and testis, which are all sites of immunologic privilege. Mutations in MCD-defming genes are characteristic of primary skin, breast, uterus, adrenal and intravascular lymphomas, suggesting that these tissues may confer “relative” immune privilege by only permitting entry to certain immune subpopulations (Shechter et al., 2013). Notably, nodal MCD often secondarily involves these immune-privileged sites, perhaps allowing them to evade immunologic surveillance. A majority (72.5%) of MCDs tumors acquired lesions that inactivated MHC class I antigen presentation or activation of T and NK cells, with some acquiring multiple lesions that inhibit immune responsiveness. Moreover, some PCNSL tumors genetically abrogate immune responsiveness despite arising in an immune- privileged site (Figure 3D) (Chapuy et al., 2016). These observations suggest a quantitative, not qualitative, model for immune evasion by MCD-like aggressive lymphomas. [0151] The combined genetic, phenotypic, clinical and functional data show that the DLBCL genetic subtypes respond differentially to standard immuno-chemotherapy and also may facilitate precision the development of targeted therapies (Figure 8A). Genetic lesions targeting the BCR-dependent

[0152] NF-AB pathway were most frequent in the MCD, BN2, and A53 subtypes, as was the expression of the BCRs using the autoreactive V_H4-34 region, suggesting that these subtypes rely on this pathway may be sensitive to BTK inhibitors. Indeed, tumors with MFDSS^L265P and CD79B mutations have been associated with a high rate of response to ibrutinib (>80%) in relapsed DLBCL and in PCNSL (Grommes et al., 2017; Lionakis et al., 2017; Wilson et al., 2015b). The apparent addiction of MCD tumors to BCR signaling is associated with the My-T-BCR supramolecular complex, which is present in MCD and PCNSL (Phelan et al., 2018). Among the genetic subtypes, BN2 had the highest prevalence of lesions affecting the BCR-dependent NF- AB pathway. Moreover, a BN2 model relied on BCR signaling to activate BTK, and was highly sensitive to ibrutinib. These considerations support the clinical evaluation of ibrutinib in BN2 cases, as does the high rate of response to ibrutinib in MZL (Noy et al., 2017), a genetic cousin of BN2.

[0153] The PI3 kinase pathway was essential in MCD, BN2 and EZB models, likely for different mechanistic reasons. In MCD cells, the My-T-BCR supercomplex coordinates NF-AB signaling on endolysosomal membranes in close proximity to the mTORCl complex, which likely contributing to the sensitivity of this subtype to PI3 kinase inhibition. By contrast, the BN2 model Riva does not form the My-T-BCR complex (Phelan et al., 2018), implying that it engages PI3 kinase as a consequence of conventional BCR signaling at the plasma membrane. The genetic and gene expression analysis implicated PI3 kinase signaling in EZB and potentially also in ST2. These subtypes most likely activate PI3 kinase as a consequence of a “toncogenic” BCR signaling Over one quarter of EZB tumors delete PTEN and/or amplify MIRI7HG which encodes microRNAs that downregulate PTEN expression. ST2, which had the highest PI3 kinase signature expression, often inactivates SGK1 genetically, perhaps promoting PI3 kinase- dependent activation of AKT and susceptibility to both PI3 kinase and mTORCl inhibitors. [0154] Other molecular targets in MCD and BN2 include the master regulatory transcription factors IRF4 and SPIB, which heterodimerize and together direct much of the ABC gene expression profile (Yang et al., 2012). IRF4 and SPIB are both downregulated in expression by lenalidomide, a drug that has shown promise in combination with other agents in DLBCL, such as ibrutinib (Goy et al., 2019; Wilson et al., 2015a; Yang et al., 2012). The activity of I/cB kinase, which is required in MCD and BN2 models, can be attenuated by treatment with BET inhibitors targeting BRD4 (Ceribelli et al., 2014). JAK1 is activated by autocrine secretion of IL- 10 in MCD, as evidenced by high expression of a JAK1 gene expression signature, and dependency of MCD models on the IL-10 receptor, JAK1 and STAT3. Selective JAK1 inhibitors are being developed in lymphoma and one, INCB040093, has shown activity in combination with a PI3 kinase ^inhibitor in non-GCB DLBCL (Phillips et al., 2018). The MYD88^L265P isoform, which is present in many MCD tumors, spontaneously coordinates a signaling complex involving IRAKI and IRAK4 (Ngo et al., 2011), both of which are essential in this subtype, supporting the evaluation of IRAK4 inhibitors in MCD, especially in combination with a BTK inhibitor (Kelly et al., 2015). By contrast, EZB models bearing an EZH2 mutation were preferentially sensitive to knockdown of components of the PRC2 co-repressor complex and thus may respond preferentially to EZH2 inhibitors such as tazemetostat. BCL2 was required in MCD, BN2 and EZB models while BCL-XL was required in MCD and EZB, suggesting that agents such as venetoclax or navitoclax may provide benefit and may be synergistic with BTK inhibitors (Mathews Griner et al., 2014).

[0155] Given the above evidence that R-CHOP chemotherapy and targeted therapies may be differentially active in particular genetic subtypes, the LymphGen algorithm should be a useful tool in DLBCL clinical trials that would extend the utility of COO assays. It is contemplated that the LymphGen classification will find initial utility in the retrospective analysis of clinical trials in DLBCL. Given the genetic complexity of DLBCL, it is challenging to identify and statistically verify the association of individual genetic alterations with clinical outcome given the problem of multiple hypothesis testing. This problem is mitigated by the fact that there are only 6 LymphGen DLBCL subtypes. Because the LymphGen subtypes differentially acquire mutations in particular signaling and regulatory pathways (e.g. BCR-dependent NF-AB pathway in BN2 and MCD, PI3 kinase pathway in EZB) and have distinct microenvironmental compositions, it is anticipated that LymphGen subtypes may differ in their response to therapies targeting oncogenic signaling pathways as well as immunotherapies. Ultimately, if a LymphGen subtype is enriched for therapeutic responses, it could be used as a selection criterion for an expansion cohort in a subsequent clinical trial.

Methods

[0156] The goal was to design an algorithm that would calculate the probability of a given DLBCL sample belonging to one of 6 defined genetic subtypes, and to assign the sample to a subtype(s) based on these probabilities. Because genome analysis of patient samples is not always comprehensive, LymphGen was designed to use as input any combination of mutational data, copy number data, and BCL2/BCL6 rearrangement data, allowing for any platform besides mutational data to be omitted. Mutational data can be derived from whole exome/genome sequencing or from targeted panel resequencing. Copy number data can be binned to 4 classes (amplification, gain, heterozygous deletion, homozygous deletion) or can be binned into just 2 classes (increased or decreased). For analyses in which copy number data are not available, LymphGen operates in a 5-subtype mode, omitting A53 since it is defined predominantly by copy number abnormalities. Data were used from the NCI cohort to model the performance of LymphGen given various types of input data and calculated the sensitivity, specificity and precision (positive predictive value) for the subtype assignments compared with the assignments using all data types optimally (Figures 10I-10Q). Models lacking BCL2 rearrangements suffered in predicting EZB and models lacking BCL6 rearrangements data suffered in predicting BN2. A lack of copy number data primarily affected prediction of EZB, MCD and ST2. Nonetheless, models constructed only from mutational data performed acceptably, with sensitivity above 81%, specificity above 98%, and precision above 79%.

Revision of Genclass procedure

[0157] Much of the hierarchical modeling methodology used by LymphGen, particularly as it relates to the definition of features, relies on the methods defined in the statistical supplement of ref. (Schmitz et ak, 2018).

[0158] As a first step towards developing a classifier, an expanded version of the previous Genclass iterative prediction method was used. This expansion added two new classes (A53 and ST2) and incorporated synonymous mutations into the predictor. Other than the modifications listed below, the Genclass algorithm was implemented as previously described (Schmitz et al., 2018).

1. With the addition of ST2 and A53, the number of possible classifications (including “Other”) was expanded from 5 to 7.

2. The final Genclass classification was used from the Schmitz paper as a starting seed for the BN2, EZB, MCD, and N1 groups.

3. Samples previously classified as “Other” and with SGK1 truncations, P2RY8 mutations, or TET2 mutations were set to ST2 in the initial seed.

4. Samples previously classified as “Other” and not part of the ST2 core that had either a) both a TP53 mutation and a single-copy TP53 loss, or b) a homozygous TP53 deletion were set to A53 in the initial seed.

5. To account for the fact that synonymous and non-coding mutations may be useful in identifying the presence of somatic hypermutation, a “Synon” feature is considered for each gene. These features, in additional to all of the mutations that affect protein coding, include all synonymous mutations within 4kb of the transcription start site, whether in the coding region or in the 5’ UTR.

6. When identifying features associated with the A53 subgroup, focal single-copy losses were included as potential copy-number features, even when not combined with mutations.

7. When identifying features associated with the A53 subgroup, “GAIN” features — consisting of samples for which the gene was covered by a segment of 30MB or less, which indicated a copy- number increase of one or more copies — were included as potential copy-number features. These were distinct from the Amplification (AMP) features, which required an increase of at least two copies. Also, combinations of gains with mutations or truncations were considered as potentially associated with A53.

8. When identifying features associated with the A53 subgroup, features indicating gains, amplifications, heterozygous deletions, or homozygous deletions of chromosome arms were identified as those samples that had at least 80% of a chromosomal arm having a given copy- number change. Whole chromosome features were identified as those samples which had the same copy -number feature for both arms of a chromosome.

[0159] As before, combination features which combine mutation and copy-number change features are used, provided these sub-features each include at least four samples, with at least one-half of the samples of the resulting combination having the associated mutation and one- fourth of the samples of the resulting combination having the copy-number change.

[0160] With the new seeds and the revised feature set, the Genclass algorithm was run on the Schmitz data set, resulting in 31 samples classified as A53, 93 samples classified as BN2, 73 samples classified as EZB, 74 samples classified as MCD, 19 samples classified as Nl, 20 samples classified as ST2, and 264 samples classified as Other. This revised Genclass classification was used as the starting point for the new LymphGen classifier.

LymphGen Methodology

[0161] The new LymphGen classifier includes several improvements. First, while previously only a single feature was allowed to be included for each gene, the new modeling allows for multiple features for a gene to be included in a hierarchical fashion with different weights. So, for example, both truncating and non-truncating mutations may be suggestive of a particular class, but it may be that truncating mutations are more predictive and so are given more weight. Second, unlike Genclass, the LymphGen predictor is probabilistic, which allows us to report the confidence of the prediction and allows a sample to share characteristics of multiple classes. [0162] The LymphGen algorithm creates separate naive Bayes predictors for each of the six primary classes (BN2, EZB, MCD, Nl, ST2, A53), as has been done for genetic predictors of COO subgroups (Scherer et al., 2016). Each predictor will have its own set of features and its own weights given to those features. The set of features considered for possible association with a class are the same as those used in the Genclass prediction, with the exception that, under certain circumstances detailed below, LOSS features are allowed to be associated with non-A53 classes.

Measures of feature significance

[0163] In the prediction algorithm, there are two measures of significance for the relationship between a given class C and feature F. [0164] Consider the following 2X2 table, where the entries in each cell represent the number of samples that do or do not have a given feature and were or were not classified as a given class according to the revised Genclass prediction described above.

TABLE 4

The first measure of significance used is “Statistical Significance,” defined to be the Fisher exact p-value associated with the above 2x2 table. The second measure used is “Effect Size,” as defined in terms of the log odds ratio:

It was found that using the log odds ratio itself was too sensitive in the case of low-frequency features, so the significance was shrunk by subtracting its standard error.

So that the final measure of Effect Size is given by

ES(C,F ) = OR(C, F) - SE(C, F)

This value is undefined for cases in which one of the cells of the 2x2 matrix is equal to zero. This was handled by setting all 0 cells to be equal to ¼, which is the value that maximizes the Effect Size as defined above.

Gene list selection

[0165] Separate gene lists were defined for each class according to the following rules:

1. Genes are included in the model of a given class in the order of the Statistical Significance of their most statistically significant feature.

2. If a copy-number feature of a gene/arm is included in the model, all copy-number features within 15MB are excluded from further consideration.

3. Only those genes with at least one feature that was found in at least 20% of the class and had a statistical significance (p<0.001) were considered.

Feature selection within a gene

[0166] The set of features for the LymphGen model separately considered mutations that either included or excluded subclonal events, and either included or excluded synonymous mutations. These subclonal and synonymous mutations generally made up a small fraction of the mutations in a given gene; so, although their inclusion or exclusion may improve model performance, there were insufficient examples to accurately estimate weights for the different mutation types. It therefore made sense to select one of the MUTATION, Synon, SubMUTATION or SubSynon features without further division. So, for each gene/class combination, the one that had the strongest statistical association with the subtype to use as the “mutation” feature was selected. Similarly, the strongest statistical association was chosen from among TRUNC and SubTRUNC to represent the “TRUNC” feature for that gene/subtype combination. This same methodology was applied to the combination features as well; so that, for example, only one of “AMP TRUNC” or “AMP SubTRUNC” would be chosen.

[0167] Within the copy-number features, it was simplest and most biologically believable to assume that either increases in copy number or decreases in copy number for a given gene or arm will be associated with a given subtype, but not both. Therefore, for each gene/class combination, the most statistically significant (non-combination) copy-number feature was identified. If this feature indicated an increase in copy number, then AMPs (and GAINs for the class being A53) along with their associated combination features were retained, while any features or combinations representing a loss of copy number for that gene were eliminated. If the most significant copy-number feature indicated a loss of copy number, then the reverse is true. [0168] Given that the changes in copy number occurred over segments that often contained multiple genes, it is not possible to distinguish computationally which one of several adjacent genes was responsible for an observed effect, while on the basis of known biology the effective gene is clearly identifiable. In the initial run of the algorithm, there were several instances of such a confusion occurring, and the incorrect gene was chosen. To prevent this, copy-number features were excluded from any other gene that was within 1MB of any of the following genes: CDKN2A, NOTCH2, REL, SPIB, USP7.

[0169] BCL2 and BCL6 fusions were also included as separate features, and if found to be significant (p<0.001) would be used as the sole feature to represent their respective genes.

Hierarchical feature selection within a gene

[0170] The previous Genclass prediction was restricted so that only the most significant feature would represent each gene, and that each gene would only be associated with the modeling of a single class. In this new version, the possibilities were expanded so that different gene features could be included in the same gene model and influence that model to different degrees. It may be, for example, that truncations are more indicative of a class. To this end, the set of features were ordered for a given gene in the hierarchical manner indicated in Fig. 12, where each level is a subset of the level above it, so that TRUNCs are a subset of MUTATIONS and AMPs are a subset of GAINs.

[0171] As stated above, GAIN and LOSS features are only included in the A53 model. For a given gene/class combination, features were selected in line with the following rules:

1. Only features that are individually statistically significant (p<0.05) are selected.

2. If both mutation features and copy-number features are included, then combination features are excluded.

3. Level 1 features should be separated from the Level 2 features (e.g., truncations being considered distinct from non-truncating mutations) if: a. Both the number of samples in the class that were in the Level 2 feature but were not in the Level 1 feature, and the number of samples that were in the Level 1 feature but not the Level 2 feature, were at least 3. b. Even excluding those samples that had the Level 1 feature, the Level 2 feature still had an association with the class that was statistically significant atp<0.05. c. The Effect Size for the Level 2 feature is larger than the Effect Size for the Level 1 feature. (Biologically, more disruptive change should be more predictive of subtype).

4. If the Level 1 and Level 2 features are not considered distinct, then the most statistically significant one is selected and the other excluded.

5. If only the copy number or mutation arm of the hierarchy has features selected according to the above criterion, and if the statistical significance of the most statistically significant combination feature is greater than the statistical significance of the highest-level feature in the remaining arm, then that combination feature replaces the highest-level feature in the arm, with any lower feature being considered as a distinct subset of the combination feature.

Example 1 : ETV6 in MCD

[0172] Considering the MCD subtype and the Synon feature produced the following 2x2 table, which has an Effect Size of 2.25 and a Statistical Significance, according to the Fisher’s exact test, of 3.25xl0^-16.

TABLE 5

This was more significant than the results of similar 2x2 tables based on MUTATION, SubMUTATION or SubSynon features, so the Synon feature was used to represent mutations going forward.

[0173] Since this model was not for the A53 subtype, the LOSS and GAIN features were removed from consideration. The HOMDEL feature was the next most significant copy-number feature. It produced the following 2x2 table (samples without copy-number data are excluded) with an Effect Size of 1.24 and a p-value of 0.0033. TABLE 6

[0174] Since there were both copy-number and mutation features that were significant with p<0.05, these features were treated separately rather than merged into a combination feature. [0175] The TRUNC feature resulted in the following 2x2 table, which had an Effect Size of 2.57 and a Statistical Significance of 2.2xlO^-11.

TABLE 7

[0176] This was more significant that the Sub TRUNC feature, so the Sub TRUNC feature was not used. Since it was significant with p<0.05, had a higher Effect Size than the Synon feature, and included 7 MCD samples (3 or more), the possibility of separating the truncations from the other mutations was considered. Excluding the truncations resulted in the following table, which had a p-value of 9.9xl0^-6:

TABLE 8

[0177] Since this significance is also less than 0.05, and there were 15 MCD samples (3 or more) with non- truncating mutations, it was confirmed that the TRUNC feature should be separated from the Synon feature. If this had not resulted in a significant p-value, then only the Synon feature would have been used, since it had a better statistical significance than the TRUNC feature. [0178] So, as a final result, the samples were divided into 4 groups according to aberrations of ETV6:

TABLE 9

Example 2: IRF4 in MCD

[0179] The most significant feature for MCD associated with the IRF4 gene was the combination feature, including SubSynon and LOSS. It can be represented by the following 2x2 table, which has an Effect Size of 0.77 and a Statistical Significance of 7.3xl0^-4:

TABLE 10

[0180] Since the p-value was less than 0.001, and the 20 MCD samples with this feature represented greater than 20% of the total set of MCD samples, the IRF4 gene was considered for inclusion in the model. This feature was more significant than any of the mutation features, and no copy- number feature was found that had a significance p-value less than 0.05. Therefore, this combination feature was chosen as the top of the hierarchy. However, the TRUNC feature produced the following 2x2 table with a Statistical Significance of 0.0491 (<0.05) and an Effect Size of 0.88: TABLE 11

[0181] So IRF4 TRUNC was included as a sub-feature of the IRF4 Synon-LOSS combination feature. Thus, the final result would be to divide the samples into three groups according to aberrations of IRF4:

TABLE 12

Single-class sample prediction

[0182] In this section is described how it was identified that the likelihood a sample is part of a particular class. The methodology is based on a categorical naive Bayes. According to naive Bayes, given a set of observations x = [x₁ ... , x_n] and a condition M an estimate of the probability of having that condition can be made as

where P₀ indicates a prior probability and M represents “not M.” If it is “naively” assumed that [x₁ ..., x_n] are independent, and further assume a flat prior, then this can be rewritten as

By defining

then this reduces to

[0183] Now suppose for a given feature, there is the following 2x2 table:

TABLE 13

The likelihood of having a feature can be empirically estimated as

So that

However, as before, not accounting for the degree of uncertainty in the empirical estimates resulted in an over-emphasis of features with few examples. The standard error of this value was estimated:

and in practice the following was used

If Ti_! or m₁ is equal to 0, then these are set to ¼, the value that maximizes the . The confidence that a sample is in class c is then calculated as

where the sum is over all genes associated with the class, and the 5)’ s are calculated based on the observed status for that sample in gene i If a sample matches more than one category (say for example, both a HOMDEL and a TRUNCATION for ETV6), the category associated with the largest 5) is used.

Example lb: ETV6 in MCD

[0184] Returning to Example 1 above, an S_ETV6 can be defined for a given sample depending on what (if any) abnormality that sample had in ETV6. The total number of MCD samples (N) is 74, and the total number of non-MCD samples (M) is 500. So, if sample j had a HOMDEL for ETV6, then for that sample the following is true:

If instead that sample had a truncating mutation in ETV6, then

Since both of these values are positive, having either indicates an increased likelihood of a sample being MCD. If a sample had both a HOMDEL and a truncating mutation, then the larger value (2.37) would be used. [0185] On the other hand, if the sample was wild type for ETV6, then 1 1 1

V E, TV6J ⁱ = -0.77 ,

6 74 ⁺ 465 500 which as a negative value indicates increased likelihood that the sample is not MCD.

Example 2b: IRF4 in MCD

[0186] Looking back to Example 2 in the previous section, if sample j had a loss of IRF4, then

0.54.

If the sample had wild-type IRF4, then

1 1 1

-0.262 .

74 ⁺ 432 500

Suppose (counterfactually) that IRF4 and ETV6 were the only genes associated with MCD. Then the confidence value for MCD would be calculated as

So, if a sample had a truncating mutation of ETV6, but was wild type for IRF4, the confidence value would be

0.892 .

Alternatively, for a sample that was wild type for ETV6 but had a loss of IRF4, it would have a confidence value of MCD equal to exp(— .77 + 0.54)

P M, CD.j = 0.442 .

1 + exp(— 0.77 + 0.54) Combining models to generate final sample call

[0187] Following the methods described above will result in each sample having confidence values between 0 and 1 for each of the 6 classes. If a sample had a confidence value between 0.5 and 0.9 in one class and less than 0.5 in all the 5 remaining classes, then the sample is called “adjacent” to that first class. If a sample has a confidence value of greater than 0.9 for one class and less than 0.9 for the 5 other remaining classes, then it is called “core” for that class. If a sample had a confidence value of greater than 0.9 in multiple classes, then it is called a “composite” sample that has qualities of all classes for which it had a confidence value greater than 0.9. For example, a sample may be called “composite EZB/A53”. If a sample had no class with a confidence level greater than 0.5, or had multiple classes with a confidence level between 0.5 and 0.9 but no classes with a confidence level greater than 0.9, then it was called “Other”. Note that in the majority of samples, either no class had a confidence level greater than 0.5, or there would be a single class with a confidence level greater than 0.9. So, the side cases discussed above were relatively rare.

Application of LymphGen to imperfect data

[0188] The data on which the LymphGen algorithm was trained included whole-exome data for all genes, complete copy-number data, and information regarding the fusion status of BCL6 and BCL2. It further included high-coverage Haloplex data, which allowed for the identification of subclonal events. It is recognized that not all of the features indicated in the model may be available. For example, they may only have a limited gene panel, lack information on fusions, or lack copy-number information. Alternatively, they may have copy- number information but lack the ability to distinguish single-copy gains from amplifications, or perhaps they only detect high- level amplifications.

[0189] If it were attempted to use the model as originally defined on such samples, their confidence values would be penalized for not having features that were not tested for. Therefore, when presented with a new data set, the model is customized to match the available data on that sample. To do this, the gene and hierarchical feature selection of the LymphGen algorithm (as defined above) are altered to exclude any feature that was not available on the data set. So for example, features would be excluded from all genes that were not available on the tested set’s gene panel: if copy- number data were unavailable, all copy -number features would be excluded; if BCL2 fusion information were unavailable, then that feature would be excluded (but this may result in the inclusion of BCL2 mutations if that information was available), etc. This reduction of the set of allowable features only extends to the final development of the naive Bayes predictor. The initial Genclass results that were used as the basis to train the model remain unchanged.

[0190] The following are additional restrictions on class prediction in the case of incomplete data:

1. Since the most prominent attribute of A53 is extensive copy-number changes, if a sample lacks copy-number information, the A53 subclass is excluded from consideration and no A53 confidence is calculated.

2. Since the prediction of the N1 subtype relies exclusively on mutations of NOTCH1, if information regarding NOTCH1 mutations is not available, then the N1 subclass is excluded from consideration and no N1 confidence is calculated.

3. With the exception of the prediction of N1 (which as stated relies solely on the N1 gene), in order for a sample to be predicted as a particular class (or as a composite including that class), that sample must include predictive features from at least two genes that were part of the predictive model.

Evaluation of model performance on subclasses of features

[0191] Although when using the methods described above, it is possible to predict samples using only a subset of features, this will likely result in reduced predictive accuracy. The degree of degradation depends on which features are missing. It may be that the missing features are of little importance and there is little change in the prediction, or it may be that crucial information is not available and the classification of a given subtype is heavily compromised.

[0192] It is therefore important to evaluate how much the loss of features affects predictor performance by assuming that the class prediction of the training-set samples on the complete set of features represents the gold standard, and then compare the results to a prediction of the training-set samples with a model based on the subset of features. Since, depending on which features are missing, the prediction of some subtypes may degrade more than others, the performance on each class separately is evaluated. The set of samples predicted are compared by the full model as being of the given class (such as BN2) or a composite including that class (such as BN2/EZB) with the set of samples predicted as that class (including composites) under the subset model, and report standard accuracy statistics such as sensitivity, specificity, and precision.

Prediction on validation sets

[0193] The BCCA cohort (Ennishi et al., 2019a) did not have whole-exome data available, but instead had sequencing data on a select gene panel. All mutation, truncation, and composite features for genes not included in the gene panel were excluded when training the model for this data set.

[0194] Copy-number calls were generated from the Affymetrix SNP6.0 array through the use of the PennCNV and OncoSNP algorithms. It was observed that this produced significantly fewer homozygous deletions than were observed on the training set, suggesting that this method had reduced sensitivity in distinguishing homozygous deletions from heterozygous losses. Therefore, all HOMDEL features and their associated composite features were removed when training the predictor for this data.

[0195] The Harvard cohort (Chapuy et al., 2018) included full-exome data on all samples; however, it was not clear what sort of mutation blacklist or gene annotation was used to develop their list of mutations. Since they provided a large set of samples, and the features used in the model were generally prevalent, mutation and truncation features (along with their composites) were included only for those genes for which there was at least one mutation found in the Harvard data. The copy- number data for the Harvard samples was generated from the exome data, which resulted in 65 regions of suspected copy-number change that applied to all samples. Further, only the direction of copy-number change was indicated for these samples, with no indication of their magnitude. It was decided, therefore, to exclude all copy-number and composite features from the predictors of every class except A53. For A53, GAIN and LOSS features were only included based on genes or arms that were included among the 65 regions, and the inclusion of the most significant feature for each region regardless of size were only allowed. Model verification via gene cross-validation

[0196] Given that there was no gold standard against which to compare the prediction results on the validation set, an alternative way was needed to demonstrate that the classes found on the training data were also found on the validation cohorts. To do this, it was considered that (with the exception of Nl) the classes were defined by sets of features that frequently co-occurred. It was theoretically possible that the classes identified did not in fact represent distinct biology and only represented coincidental co-occurrence of various features; but if that were the case, it would not be expected to see a similar co-occurrence on the independent validation cohorts. [0197] For the BN2, EZB, MCD and STD subtypes, co-occurrence was tested for by looking at the relationship between the presence or absence of features for a gene and the model score with that gene excluded. If a class was defined based on a purely coincidentally co-occurring set of features, then on an independent cohort there would be no relationship between the features for a gene and the model score based on all of the other predictive features. If instead the relationship was not coincidental, then it would be expected that those samples that had predictive features for that gene would tend to have additional features associated with that class, and so have a higher predictor score than those without that feature.

[0198] Consider the example of ETV6 within MCD. Calculate the predictive score for MCD, excluding the ETV6 features, and calculate their ranks.

The score for ETV6 is then calculated as the sum of ranks for those samples with an ETV6 feature.

UETV6 j has

ETV6 feature [0199] If it is assumed that each sample was equally likely to have an ETV6 feature, this would reduce to the standard Mann-Whitney U test. However, recognize that the features are not uniformly distributed among the samples. Some samples have more reported mutations of all types than other samples, either due to genomic instability or differences between samples in assay sensitivity. This will result in increased co-occurrence of all features (whether predictive or not) beyond what one would expect by chance. Therefore, it needs to be demonstrated that the relationship between ETV6 (in this example) and the MCD score is greater than the relationship between the score and randomly occurring features. To do so, generate random sets of patients 5₁ ...,5_{10 000}, each with size equal to the number of samples with ETV6, and with the probability of a sample being selected weighted according to the number of mutations reported for that patient. Then calculate a U score for each of these selections.

Then report the following Monte Carlo p-value for the association between the gene and the subtype, where / is the indicator function.

A global p-value for a given class can be calculated by using Fisher’s method to combine the p- values of the individual predicative genes associated with that class.

The A53 subtype was characterized by an overall increase in copy-number changes associated with TP53 alterations. To test the validity of this subtype, a standard Mann-Whitney U test was used to check whether the total number of copy-number changes was higher in samples with TP53 alterations than in those without such alterations.

Genetic prediction of EZB-MYC+ andEZB-MYC-

[0200] After observing the connection between the EZB subtype and the DHIT+ class, a genetically-based classification of the EZB-MYC+ and EZB-MYC- groups was created. This was done in a similar manner to the development of the other binary naive Bayes submodels of the LymphGen algorithm but with a few modifications. First, the prediction was only done within the samples that were classified as EZB by the LymphGen algorithm, with the gene expression classification of DHIT cases being used as the starting point. Second, given the much smaller set of training data, the stringency of the feature selection criteria was reduced, requiring statistical significance p<0.01 which previously was p<0.001. This resulted in a predictor that included three composite features.

[0201] Since the Harvard data lacked classification calls for the DHIT subtype, and the BCCA samples lacked the complete genomic data required to make the genomic prediction, evaluation of the finding on a completely independent data set was not possible. Instead, a permutation test was used to demonstrate that the genomic predictor of EZB-MYC+ had greater association with the DHIT classification than one would expect by chance, even taking into account the possibility of overfitting. To this end, the DHIT class labels were randomly permuted among the EZB samples and repeated the prediction algorithm, and then tested the agreement between the permuted class labels and the predictor score, according to a Wilcoxon test. This was repeated 1,000 times. In only three of those times was the agreement between the permuted class labels and predictor score more significant than what was observed for the unpermuted DHIT labels, resulting in a permutation p- value of 0.004.

Additional statistical tests and methods

[0202] Digital gene expression values, copy number, and mutation calls were generated as previously described (Chapuy et ah, 2018; Ennishi et ah, 2019a; Schmitz et ah, 2018). The DHIT score was generated from the digital gene expression as previously described (Ennishi et al., 2019a). All remaining signature averages were calculated as the mean of the normalized, log₂- transformed digital gene expression values for all genes in the signature. These values were then linearly transformed, such that across all samples the resulting normalized averages had a median of zero and an interquartile range of 1.35 (the interquartile range of a standard normal distribution). For Figure 6D, a multivariate regression of the GCB4 and MYCUp-4 signature averages was fit to the DHIT score for all EZB cases (including composite EZB cases). The results of this fit for each EZB sample was plotted against the observed DHIT score for that sample. The reported p-value was based on an F-test for the total significance of the model.

[0203] Within the BCCA cohort, disease-specific survival was used as the endpoint, while overall survival was used within the training and Harvard cohorts. The hazard ratios and confidence intervals for survival differences between subtypes indicated in Figure 4N were calculated according to the Cox proportional hazard model, based on all samples of the types being compared within the specified cohort. Models, including a single binary variable indicating class, were used to generate the within-study estimates, with p-values generated according to a log-rank test. For the combined estimates, a multivariate model was used, which included, in addition to the binary class variable, a categorical co-variate indicating the cohort. For these models, the p-value was generated from a score test.

[0204] All reported p-values are two-sided. The statistical significances for differential prevalence of features between classes were calculated with a Fisher’s exact test.

Estimation of DLBCL Genetic Subtype Prevalence

[0205] Since the NCI DLBCL was deliberated enriched for ABC and Unclassified cases, it was estimated what the prevalence of these subtypes would be in a population-based cohort of DLBCL cases. Using the published prevalence of COO subgroups within a population-based cohort (Scott et al., 2015), the prevalences of each genetic subtype within the NCI cohort were adjusted based on the percentages of the ABC, GCB and Unclassified subgroups within each genetic subtype. The assumption behind this normalization is that the relationship between COO subgroup and DLBCL genetic subtype is relatively invariant, as observed to be the case in the three cohorts analyzed in the present study (Figures 4B-4C).

RNA-seq analysis

[0206] The gene expression signature database was first described in ref. (Shaffer et al.,

2006) and are available at api. gdc. cancer .gov/ data / cf7cd89e-da75-45fe-a4d5-e89e491f45d6. Digital gene expression was calculated as described (Schmitz et al., 2018). Expressed immunoglobulin heavy chain genes were assembled from the RNA-seq data as described (Bolotin et al., 2017). Prevalence of genetic alterations in other non-Hodgkin lymphomas

[0207] Prevalence of lymphoma cases with mutations in genes that characterize particular genetic subtypes of DLBCL were derived from ref. (Schmitz et al., 2018). Prevalence of mutations in other types of non-Hodgkin lymphoma (NHL) were calculated using published datasets derived whole exome, whole genome, whole transcriptome, or targeted resequencing. As a negative control, the prevalence of cases with mutations were determined from whole exome sequencing data in a non- Hodgkin lymphoma cohort consisting of BL, FL, chronic lymphocytic leukemia and mantle cell lymphoma. References for the prevalence calculations in Figure 3B, Figure 3D-G and Figure 6H are as follows:

Non-Hodgkin lymphomas:

BL (Bouska et al., 2017a; Love et al., 2012; Richter et al., 2012)

Chronic lymphocytic leukemia (Amin et al., 2016; Landau et al., 2017; Landau et al., 2015; Ljungstrom et al., 2016; Puente et al., 2015; Quesada et al., 2011; Wang et al., 2011)

FL (Bouska et al., 2017b; Green et al., 2013; Green et al., 2015; Hellmuth et al., 2018; Krysiak et al., 2017; Okosun et al., 2016; Pasqualucci et al., 2014; Tsukamoto et al., 2017; Zamo et al., 2018) Mantle cell lymphoma (Agarwal et al., 2019; Bea et al., 2013; Wu et al., 2016; Zhang et al., 2014)

Primary extranodal lymphomas:

Primary CNS lymphoma (Braggio et al., 2015; Bruno et al., 2014; Chapuy et al., 2016; Fontanilles et al., 2017; Fukumura et al., 2016; Hattori et al., 2019; Hattori et al., 2017; Hickmann et al., 2019; Nakamura et al., 2016; Vater et al., 2015; Zhou et al., 2018b)

Primary cutaneous lymphoma (Ducharme et al., 2019; Mareschal et al., 2017; Zhou et al., 2018a) Primary vitreoretinal lymphoma (Yonese et al., 2019)

Primary testicular lymphoma (Chapuy et al., 2016; Kraan et al., 2014)

Primary breast lymphoma (Cao et al., 2017; Franco et al., 2017) Primary intravascular lymphoma (Schrader et al., 2018; Shin et al., 2019; Suehara et al., 2018) Primary uterine lymphoma (Cao et al., 2017)

Marginal zone lymphoma: (Clipson et al., 2015; Ganapathi et al., 2016; Hyeon et al., 2018; Johansson et al., 2016; Kiel et al., 2012; Martinez et al., 2014; Parry et al., 2013; Parry et al., 2015; Pillonel et al., 2018; Rossi et al., 2012; Spina et al., 2016)

FL and transformed FL: (Bouska et al., 2017b; Green et al., 2013; Green et al., 2015; Hellmuth et al., 2018; Krysiak et al., 2017; Okosun et al., 2014; Okosun et al., 2016; Pasqualucci et al., 2014; Tsukamoto et al., 2017; Zamo et al., 2018)

Nodular lymphocyte-predominant Hodgkin lymphoma (NLPHL) and T-cell histiocyte-rich large B cell lymphoma (THRLBCL): (Hartmann et al., 2016; Schuhmacher et al., 2019)

Functional genomics

Cell culture

[0208] HBL-1, TMD8, RIVA (also known as RI-1), OCI-Ly8 and BJAB cell lines were grown at 37°C in the presence of 5% C02 and maintained in RPMI (GIBCO) supplemented with 10% Tetracycline- tested fetal bovine serum (Atlanta Biologicals) and 1% pen/strep and 1% L- glutamine (Invitrogen). All cell lines were regularly tested for mycoplasma using the MycoAlert Mycoplasma Detection Kit (Lonza) and cell line identity was confirmed by DNA fingerprinting examining 16 regions of copy number variants (Jonathan Keats, personal communication).

CRISPR-Cas9 pooled screens for essential genes Cas9-clone generation

[0209] Cell lines were transduced multiple times with either pTO-Cas9-hygro or pCW-Cas9- bsr, selected and single cell cloned using the limiting dilution technique. Cell clones were tested for functional Cas9 cutting after transduction with sgRNAs targeting surface markers including CD20 or ICAMl. Clones were selected based upon loss of surface expression within the transduced population as measured by FACS 8-14 days after addition of doxycyline. sgRNA vector construction and cloning of individual sgRNAs

[0210] The pLKO-based sgRNA vector was purchased from Addgene (#52628). The puromycin gene was removed and replaced with a puroR-GFP fusion protein previously described(Ngo et al., 2006) using Gibson assembly. The resulting plasmid was digested with BfuAI and incubated with shrimp alkaline phosphatase before isolating the backbone. Complementary sgRNA sequences flanked by ACCG on the 5’ end, and CTTT on the 3’ of the reverse strand, were annealed, diluted, and ligated into the cut vector with T4 ligase (New England Biolabs) according to the manufacturer’s instructions. All transformations were performed in Stbl3 bacteria and grown at 30° C. sgRNA library construction

[0211] The genome-wide “Brunello” sgRNA library(Doench et al., 2016) was purchased from Addgene and transformed using electroporation in Stbl4 bacteria (Invitrogen). Transformations were grown at 30°C overnight on 24.5cm² bioassay plates maintaining at least 500X coverage. Colonies were scraped, spun and plasmid DNA was isolated using Qiagen Plasmid Maxi Kits (Qiagen).

Virus production and transduction of sgRNA library

[0212] Lentiviruses were produced in 293FT cells (Invitrogen) by cotransfecting sgRNA vectors with packaging vectors pPAX2 (Addgene #12260) and pMD2.g (Addgene #12259) in a 4:3:1 ratio using Trans-IT 293T (Mirus) according to manufactures instructions. Supernatants were harvested 48 and 72 hours later, filtered using 0.45 um HV Durapore membranes (Millipore) and incubated with Lenti-X concentrator (CloneTech). Virus was concentrated according to manufacturer’s instructions, aliquoted and frozen. For genome-wide screenings, virus titration was performed on target cell populations. Transduced cells were split and incubated with or without puromycin until untransduced control cells were dead. The transduction percentage was calculated as the ratio of viable cells in puromycin selected cell populations versus non-selected cell populations.

Pooled sgRNA screening

[0213] For genome-wide screens, 2 individual replicates were transduced using viral titers to attain an average of 500 copies of each sgRNA. Cultures were carried for the duration of the 21- day screen maintaining 500X coverage. Antibiotic selection was started 3-4 days after transduction and carried out until untransduced control cells were dead. An aliquot of 50 million cells was pelleted and frozen for a day 0 time point and doxy cy cline was added to the culture media at 500 ng/mL final concentration. Transduced cells passaged every two days with media containing doxy cy cline. At day 21 after doxy cy cline induction cells were pelleted and frozen. DNA was isolated from frozen cell pellets using the QIAmp DNA Blood Maxi kits (Qiagen). sgRNA library sequencing preparation and sequence extraction

[0214] Sequencing libraries were prepared as described (Phelan et al., 2018) using a 2 round nested PCR approach to isolate the sgRNA sequence from genomic DNA in the first round and to add Illumina sequencing adapters in the second round. Products were amplified using ExTaq (Takara) in 18 cycles for both rounds of amplification. Amplicons were size selected using eGels (Invitrogen) and libraries were quantitated using the Kapa quantification kit for Illumina Platforms according to the manufacturer’s instructions (Kapa Biosystems) or by Qubit quantification (Thermo Fisher Scientific). Single end 75 bp read sequencing was performed on a NextSeq500 system using the Illumina NextSeq 500High Output v2 kit to achieve an average sequencing coverage of 250X. Libraries were multiplexed using indexes compatible with the Illumina TrueSeq HT kit.

[0215] Oligonucleotides contained eight base pair indices and a variable length adapter (8-15 bps) to prevent monotemplate stretches during sequencing. sgRNA sequences were extracted from the sequencing reads using a custom script and aligned to the sgRNA library sequences using the Bowtie 2.2.9 algorithm. (Langmead et al., 2009) Raw read counts were normalized to 40xl0⁶per sample and increased by 1 before calculating a CRISPR screen score as follows: All sgRNAs with an average normalized read count below 50 at day 0 were removed due to low coverage. The average log2 fold change was computed between day 21 and day 0 for each replicate and then a Z-score was calculated from the average log2 fold change per gene.

Analysis of tumor suppressor genes

[0216] Cas9-expressing TMD8, Riva, and HBL1 cells were infected with lentiviruses co expressing an sgRNA and a puromycin resistance-GFP fusion gene. 3 days after infection cells were selected with puromycin and Cas9 expression was induced by addition of doxycycline. After 7 days, selected cells were washed with media and mixed at a ratio of 1 : 10 with uninfected cells of the respective cell lines. Cell mixtures were treated with Ibrutinib (Selleckchem) or DMSO (Sigma) for a period of 21 days. Based on the overall viability of the cultures, ibrutinib doses were increased in the first 10 days of treatment to a final concentration of 2.5 ng/mL (TMD8) and 10 ng/mL (HBL1 and Riva) and kept constant thereafter. Relative growth of cells co-expressing GFP and sgRNAs was monitored by FACS. Cell proliferation and survival of these cells was determined proportional to day 0 of the drug treatment. Growth differences of ibrutinib treated cells were normalized to DMSO treated control cells.

Mouse xenograft experiments

[0217] Murine xenograft models of the MCD genetic subtype were established using the TMD8 and HBL1 cell lines, and a model of the BN2 subtype was established using the Riva cell lines. Tumors were established by subcutaneous injection of lxKFRiva, TMD8 or HBL1 cells into the right flank of female non-obese diabetic/severe combined immunodeficient/common gamma chain deficient (NSG) mice (Jackson Laboratory). Tumor growth was monitored by measuring tumor size in two orthogonal dimensions. Tumor volume was calculated by using the formula ½(long dimension)(short dimension)². Eleven days after injection of the tumor cells, the average tumor volume reached 200 mm³ and drug therapy was started. The Riva tumor bearing NSG mice were divided into three groups of 5 mice each, with comparable tumor burden between groups as evaluated by tumor volume. Ibrutinib (MedChemExpress) was dissolved in 50% DMSO and given by intraperitoneal (i.p.) injection daily at 3 or 6 mg/kg/day for 12 days. Control mice received the same amount of 50% DMSO by i.p. injection. Mice with TMD8 or HBL1 xenografts were divided into 2 groups of 5 mice, with the treatment group receiving ibrutinib at 5 mg/kg/day for 12 days and the control group receiving 50% DMSO. Tumor volume was monitored during this time. At day 12 after initiation of the therapy, all of the mice were euthanized. All animal experiments were approved by the National Cancer Institute Animal Care and Use Committee (NCI ACUC) and were performed in accordance with NCI ACUC guidelines.

[0218] References

Agarwal, R., Chan, Y. C, Tam, C. S., Hunter, T., Vassiliadis, D., Teh, C. E., Thijssen, R., Yeh, P., Wong,

S. Q., Ftouni, S., et al. (2019). Dynamic molecular monitoring reveals that SWI-SNF mutations mediate resistance to ibrutinib plus venetoclax in mantle cell lymphoma. Nat Med 25, 119-129.

Alizadeh, A. A., Eisen, M. B., Davis, R. E., Ma, C., Lossos, I. S., Rosenwald, A., Boldrick, J. C., Sabet, H.,

Tran, T., Yu, X., et al. (2000). Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403, 503-511.

Amin, N. A., Seymour, E., Saiya-Cork, K., Parkin, B., Shedden, K., andMalek, S. N. (2016). A Quantitative Analysis of Subclonal and Clonal Gene Mutations before and after Therapy in Chronic Lymphocytic Leukemia. Clin Cancer Res 22, 4525-4535.

Baeuerle, P. A., and Baltimore, D. (1988). I kappa B: a specific inhibitor of the NF-kappa B transcription factor. Science 242, 540-546.

Bea, S., Colomo, L., Lopez-Guillermo, A., Salaverria, L, Puig, X., Pinyol, M., Rives, S., Montserrat, E., and Campo, E. (2004). Clinicopathologic significance and prognostic value of chromosomal imbalances in diffuse large B-cell lymphomas. J Clin Oncol 22, 3498-3506.

Bea, S., Valdes-Mas, R., Navarro, A., Salaverria, L, Martin-Garcia, D., Jares, P., Gine, E., Pinyol, M.,

Royo, C., Nadeu, F., et al. (2013). Landscape of somatic mutations and clonal evolution in mantle cell lymphoma. Proc Natl Acad Sci U S A 110, 18250-18255.

Bolotin, D. A., Poslavsky, S., Davydov, A. N., Frenkel, F. E., Fanchi, L., Zolotareva, O. L, Hemmers, S., Putintseva, E. V., Obraztsova, A. S., Shugay, M., et al. (2017). Antigen receptor repertoire profiling from RNA-seq data. Nat Biotechnol 35, 908-911.

Bouska, A., Bi, C., Lone, W., Zhang, W., Kedwaii, A., Heavican, T., Lachel, C. M., Yu, J., Ferro, R.,

Eldorghamy, N., et al. (2017a). Adult high-grade B-cell lymphoma with Burkitt lymphoma signature: genomic features and potential therapeutic targets. Blood 130, 1819-1831.

Bouska, A., Zhang, W., Gong, Q., Iqbal, J., Scuto, A., Vose, J., Ludvigsen, M., Fu, K., Weisenburger, D.

D., Greiner, T. C., et al. (2017b). Combined copy number and mutation analysis identifies oncogenic pathways associated with transformation of follicular lymphoma. Leukemia 31, 83-91.

Braggio, E., Van Wier, S., Ojha, J., McPhail, E., Asmann, Y. W., Egan, J., da Silva, J. A., Schiff, D.,

Lopes, M. B., Decker, P. A., et al. (2015). Genome-Wide Analysis Uncovers Novel Recurrent Alterations in Primary Central Nervous System Lymphomas. Clin Cancer Res 21, 3986-3994.

Bruno, A., Boisselier, B., Labreche, K., Marie, Y., Polivka, M., Jouvet, A., Adam, C., Figarella-B ranger, D., Miquel, C., Eimer, S., et al. (2014). Mutational analysis of primary central nervous system lymphoma. Oncotarget 5, 5065-5075.

Busman-Sahay, K., Drake, L., Sitaram, A., Marks, M., and Drake, J. R. (2013). Cis and Trans Regulatory Mechanisms Control AP2 -Mediated B Cell Receptor Endocytosis via Select Tyrosine- Based Motifs. PLoS One 8, e54938.

Cao, X. X., Li, J., Cai, H., Zhang, W., Duan, M. H., and Zhou, D. B. (2017). Patients with primary breast and primary female genital tract diffuse large B cell lymphoma have a high frequency of MYD88 and CD79B mutations. AnnHematol 96, 1867-1871.

Cardinez, C., Miraghazadeh, B., Tanita, K., da Silva, E., Hoshino, A., Okada, S., Chand, R., Asano, T., Tsumura, M., Yoshida, K., et al. (2018). Gain-of-function IKBKB mutation causes human combined immune deficiency. J Exp Med 215, 2715-2724.

Celeste, A., Petersen, S., Romanienko, P. J., Femandez-Capetillo, O., Chen, H. T., Sedelnikova, O. A., Reina-San-Martin, B., Coppola, V., Meffre, E., Difilippantonio, M. J., et al. (2002). Genomic instability in mice lacking histone H2AX. Science 296, 922-927.

Ceribelli, M., Kelly, P. N., Shaffer, A. L., Wright, G. W., Xiao, W., Yang, Y., Mathews Griner, L. A., Guha, R., Shinn, P., Keller, J. M., et al. (2014). Blockade of oncogenic IkappaB kinase activity in diffuse large B-cell lymphoma by bromodomain and extraterminal domain protein inhibitors. Proc Natl Acad Sci U S A 111, 11365-11370.

Challa-Malladi, M., Lieu, Y. K., Califano, O., Holmes, A. B., Bhagat, G., Murty, V. V., Dominguez-Sola, D., Pasqualucci, L., and Dalla-Favera, R. (2011). Combined genetic inactivation of beta2- Microglobulin and CD58 reveals frequent escape from immune recognition in diffuse large B cell lymphoma. Cancer Cell 20, 728-740.

Chapuy, B., Roemer, M. G., Stewart, C., Tan, Y., Abo, R. P., Zhang, L., Dunford, A. J., Meredith, D. M., Thomer, A. R., Jordanova, E. S., et al. (2016). Targetable genetic features of primary testicular and primary central nervous system lymphomas. Blood 127, 869-881.

Chapuy, B., Stewart, C., Dunford, A. J., Kim, J., Kamburov, A., Redd, R. A., Lawrence, M. S., Roemer, M. G. M., Li, A. J., Ziepert, M., et al. (2018). Molecular subtypes of diffuse large B cell lymphoma are associated with distinct pathogenic mechanisms and outcomes. Nat Med 24, 679-690.

Clipson, A., Wang, M., de Leval, L., Ashton-Key, M., Wotherspoon, A., Vassiliou, G., Bolli, N., Grove, C., Moody, S., Escudero-Ibarz, L., et al. (2015). KLF2 mutation is the most frequent somatic change in splenic marginal zone lymphoma and identifies a subset with distinct genotype. Leukemia 29, 1177-1185.

Crescenzo, R., Abate, F., Lasorsa, E., Tabbo, F., Gaudiano, M., Chiesa, N., Di Giacomo, F.,

Spaccarotella, E., Barbarossa, L., Ercole, E., et al. (2015). Convergent mutations and kinase fusions lead to oncogenic STAT3 activation in anaplastic large cell lymphoma. Cancer Cell 27, 516-532.

Dave, S. S., Fu, K., Wright, G. W., Lam, L. T., Kluin, P., Boerma, E. I, Greiner, T. C., Weisenburger, D. D., Rosenwald, A., Ott, G., et al. (2006). Molecular diagnosis of Burkitt's lymphoma. N Engl J Med 354, 2431-2442.

Davis, R. E., Brown, K. D., Siebenlist, U., and Staudt, L. M. (2001). Constitutive nuclear factor kappa B activity is required for survival of activated B Cell-like diffuse large B cell lymphoma cells. J Exp Med 194, 1861-1874.

Davis, R. E., Ngo, V. N., Lenz, G., Tolar, P., Young, R. M., Romesser, P. B., Kohlhammer, H., Lamy, L., Zhao, H., Yang, Y., et al. (2010). Chronic active B -cell-receptor signalling in diffuse large B-cell lymphoma. Nature 463, 88-92.

Di Cristofano, A. (2017). SGK1: The Dark Side of PI3K Signaling. CurrTop Dev Biol 123, 49-71. Doench, J. G., Fusi, N., Sullender, M., Hegde, M., Vaimberg, E. W., Donovan, K. F., Smith, L, Tothova, Z., Wilen, C., Orchard, R., et al. (2016). Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9. Nat Biotechnol 34, 184-191.

Dogan, L, Bertocci, B., Vilmont, V., Delbos, F., Megret, J., Storck, S., Reynaud, C. A., and Weill, J. C. (2009). Multiple layers of B cell memory with different effector functions. Nat Immunol 10, 1292- 1299.

Dominguez, P. M., Ghamlouch, H., Rosikiewicz, W., Kumar, P., Beguelin, W., Fontan, L., Rivas, M. A., Pawlikowska, P., Armand, M., Mouly, E., et al. (2018). TET2 Deficiency Causes Germinal Center Hyperplasia, Impairs Plasma Cell Differentiation, and Promotes B-cell Lymphomagenesis. Cancer Discov 8, 1632-1653.

Ducharme, O., Beylot-Barry, M., Pham-Ledard, A., Bohers, E., Viailly, P. J., Bandres, T., Faur, N., Frison, E., Vergier, B., Jardin, F., et al. (2019). Mutations of the B-cell receptor pathway confer chemoresistance in primary cutaneous diffuse large B-cell lymphoma leg-type. J Invest Dermatol.

Ennishi, D., Jiang, A., Boyle, M., Collinge, B., Grande, B. M., Ben-Neriah, S., Rushton, C., Tang, J., Thomas, N., Slack, G. W., et al. (2019a). Double-Hit Gene Expression Signature Defines a Distinct Subgroup of Germinal Center B-Cell-Like Diffuse Large B-Cell Lymphoma. J Clin Oncol 37, 190-201. Ennishi, D., Takata, K., Beguelin, W., Duns, G., Mottok, A., Farinha, P., Bashashati, A., Saberi, S.,

Boyle, M., Meissner, B., et al. (2019b). Molecular and Genetic Characterization of MHC Deficiency Identifies EZH2 as Therapeutic Target for Enhancing Immune Recognition. Cancer Discov 9, 546- 563.

Evan, G. L, Wyllie, A. H., Gilbert, C. S., Littlewood, T. D., Land, H., Brooks, M., Waters, C. M., Penn, L. Z., and Hancock, D. C. (1992). Induction of apoptosis in fibroblasts by c-myc protein. Cell 69, 119- 128.

Flossbach, L., Antoneag, E., Buck, M., Siebert, R., Mattfeldt, T., Moller, P., and Barth, T. F. (2011). BCL6 gene rearrangement and protein expression are associated with large cell presentation of extranodal marginal zone B-cell lymphoma of mucosa-associated lymphoid tissue. Int J Cancer 129, 70-77.

Fontanilles, M, Marguet, F., Bohers, E., Viailly, P. J., Dubois, S., Bertrand, P., Camus, V., Mareschal, S., Ruminy, P., Maingonnat, C., et al. (2017). Non-invasive detection of somatic mutations using next-generation sequencing in primary central nervous system lymphoma. Oncotarget 8, 48157- 48168.

Franco, F., Gonzalez-Rincon, J., Lavemia, J., Garcia, J. F., Martin, P., Bellas, C., Piris, M. A., Pedrosa,

L., Miramon, J., Gomez-Codina, J., et al. (2017). Mutational profile of primary breast diffuse large Bcell lymphoma. Oncotarget 8, 102888-102897.

Fukumura, K., Kawazu, M., Kojima, S., Ueno, T., Sai, E., Soda, M., Ueda, H., Yasuda, T., Yamaguchi, H., Lee, J., et al. (2016). Genomic characterization of primary central nervous system lymphoma.

Acta Neuropathol 131, 865-875.

Ganapathi, K. A., Jobanputra, V., Iwamoto, F., Jain, P., Chen, J., Cascione, L., Nahum, O., Levy, B., Xie, Y., Khattar, P., et al. (2016). The genetic landscape of dural marginal zone lymphomas.

Oncotarget 7, 43052-43061.

Goy, A., Ramchandren, R., Ghosh, N., Munoz, J., Morgan, D. S., Dang, N. H., Knapp, M., Delioukina,

M., Kingsley, E., Ping, J., et al. (2019). Ibrutinib plus lenalidomide and rituximab has promising activity in relapsed/refractory non-germinal center B-cell-like DLBCL. Blood 134, 1024-1036.

Green, M. R., Gentles, A. J., Nair, R. V., Irish, J. M., Kihira, S., Liu, C. L., Kela, L, Hopmans, E. S., Myklebust, J. H., Ji, H., et al. (2013). Hierarchy in somatic mutations arising during genomic evolution and progression of follicular lymphoma. Blood 121, 1604-1611.

Green, M. R., Kihira, S., Liu, C. L., Nair, R. V., Salari, R., Gentles, A. J., Irish, J., Stehr, H., Vicente- Duenas, C., Romero-Camarero, L, et al. (2015). Mutations in early follicular lymphoma progenitors are associated with suppressed antigen presentation. Proc Natl Acad Sci U S A 112, El 116-1125. Grammes, C., Pastore, A., Palaskas, N., Tang, S. S., Campos, C., Schartz, D., Codega, P., Nichol, D., Clark, O., Hsieh, W. Y., et al. (2017). Ibrutinib Unmasks Critical Role of Bruton Tyrosine Kinase in Primary CNS Lymphoma. Cancer Discov 7, 1018-1029.

Hartmann, S., Schuhmacher, B., Rausch, T., Fuller, L., Doling, C., Weniger, M., Lollies, A., Weiser, C., Thumer, L., Rengstl, B., et al. (2016). Highly recurrent mutations of SGK1, DUSP2 and JUNB in nodular lymphocyte predominant Hodgkin lymphoma. Leukemia 30, 844-853.

Hattori, K., Sakata-Yanagimoto, M., Kusakabe, M., Nanmoku, T., Suehara, Y., Matsuoka, R., Noguchi, M., Yokoyama, Y., Kato, T., Kurita, N., et al. (2019). Genetic evidence implies that primary and relapsed tumors arise from common precursor cells in primary central nervous system lymphoma. Cancer Sci 110, 401-407.

Hattori, K., Sakata-Yanagimoto, M., Okoshi, Y., Goshima, Y., Yanagimoto, S., Nakamoto-Matsubara, R., Sato, T., Noguchi, M., Takano, S., Ishikawa, E., et al. (2017). MYD88 (L265P) mutation is associated with an unfavourable outcome of primary central nervous system lymphoma. Br J Haematol 177, 492-494.

Havranek, O., Xu, J., Kohrer, S., Wang, Z., Becker, L., Comer, J. M, Henderson, J., Ma, W., Man Chun Ma, J., Westin, J. R., et al. (2017). Tonic B-cell receptor signaling in diffuse large B-cell lymphoma.

Blood 130, 995-1006.

Heise, N., De Silva, N. S., Silva, K., Carette, A., Simonetti, G., Pasparakis, M., and Klein, U. (2014). Germinal center B cell maintenance and differentiation are controlled by distinct NF-kappaB transcription factor subunits. JExp Med 211, 2103-2118.

Hellmuth, J. C., Louissaint, A., Jr., Szczepanowski, M., Haebe, S., Pastore, A., Alig, S., Staiger, A. M., Hartmann, S., Kridel, R., Ducar, M. D., et al. (2018). Duodenal-type and nodal follicular lymphomas differ by their immune microenvironment rather than their mutation profiles. Blood 132, 1695- 1702.

Hickmann, A. K., Frick, M., Hadaschik, D., Battke, F., Bittl, M., Ganslandt, O., Biskup, S., and Docker,

D. (2019). Molecular tumor analysis and liquid biopsy: a feasibility investigation analyzing circulating tumor DNA in patients with central nervous system lymphomas. BMC Cancer 19, 192.

Hilton, L., Tang, J., Ben-Neriah, S., Alcaide, M., Jiang, A., Grande, B. M., Rushton, C., Boyle, M., Meissner, B., Scott, D., and Morin, R. D. (2019). The double hit signature identifies double-hit diffuse large B-cell lymphoma with genetic events cryptic to FISH. Blood in press.

Hodson, D. J., Shaffer, A. L., Xiao, W., Wright, G. W., Schmitz, R., Phelan, J. D., Yang, Y., Webster, D.

E., Rui, L., Kohlhammer, H., et al. (2016). Regulation of normal B-cell differentiation and malignant B-cell survival by OCT2. Proc Natl Acad Sci U S A 113, E2039-2046.

Hyeon, J., Lee, B., Shin, S. H., Yoo, H. Y., Kim, S. J., Kim, W. S., Park, W. Y., and Ko, Y. H. (2018). Targeted deep sequencing of gastric marginal zone lymphoma identified alterations of TRAF3 and TNFAIP3 that were mutually exclusive forMALTl rearrangement. Modem pathology : an official journal of the United States and Canadian Academy of Pathology, Inc 31, 1418-1428.

Johansson, P., Klein-Hitpass, L., Grabellus, F., Arnold, G., Klapper, W., Pfortner, R., Duhrsen, U., Eckstein, A., Durig, J., and Kuppers, R. (2016). Recurrent mutations in NF-kappaB pathway components, KMT2D, and NOTCHl/2 in ocular adnexal MALT-type marginal zone lymphomas. Oncotarget7, 62627-62639.

Kelly, P. N., Romero, D. L., Yang, Y., Shaffer, A. L., 3rd, Chaudhary, D., Robinson, S., Miao, W., Rui, L., Westlin, W. F., Kapeller, R., and Staudt, L. M. (2015). Selective interleukin-1 receptor-associated kinase 4 inhibitors for the treatment of autoimmune disorders and lymphoid malignancy. J Exp Med 212, 2189-2201.

Kiel, M. J., Velusamy, T., Betz, B. L., Zhao, L., Weigelin, H. G., Chiang, M. Y., Huebner-Chan, D. R., Bailey, N. G., Yang, D. T., Bhagat, G., et al. (2012). Whole-genome sequencing identifies recurrent somatic NOTCH2 mutations in splenic marginal zone lymphoma. J Exp Med 209, 1553-1565. Kraan, W., vanKeimpema, M., Horlings, H. M, Schilder-Tol, E. J., Oud, M. E., Noorduyn, L. A., Kluin, P. M., Kersten, M. J., Spaargaren, M., and Pals, S. T. (2014). High prevalence of oncogenic MYD88 and CD79B mutations in primary testicular diffuse large B-cell lymphoma. Leukemia 28, 719-720. Krysiak, K., Gomez, F., White, B. S., Matlock, M., Miller, C. A., Trani, L., Fronick, C. C., Fulton, R. S., Kreisel, F., Cashen, A. F., et al. (2017). Recurrent somatic mutations affecting B-cell receptor signaling pathway genes in follicular lymphoma. Blood 129, 473-483.

Kuppers, R., Duhrsen, U., and Hansmann, M. L. (2014). Pathogenesis, diagnosis, and treatment of composite lymphomas. Lancet Oncol 15, e435-446.

Lam, L. T., Wright, G., Davis, R. E., Lenz, G., Farinha, P., Dang, L., Chan, J. W., Rosenwald, A., Gascoyne, R. D., and Staudt, L. M. (2008). Cooperative signaling through the signal transducer and activator of transcription 3 and nuclear factor-{kappa}B pathways in subtypes of diffuse large B-cell lymphoma. Blood 111, 3701-3713.

Landau, D. A., Sun, C., Rosebrock, D., Herman, S. E. M., Fein, J., Sivina, M., Underbayev, C., Liu, D., Hoellenriegel, J., Ravichandran, S., et al. (2017). The evolutionary landscape of chronic lymphocytic leukemia treated with ibmtinib targeted therapy. Nat Commun 8, 2185.

Landau, D. A., Tausch, E., Taylor-Weiner, A. N., Stewart, C., Reiter, J. G., Bahlo, J., Kluth, S., Bozic, L, Lawrence, M., Bottcher, S., et al. (2015). Mutations driving CLL and their evolution in progression and relapse. Nature 526, 525-530.

Langmead, B., Trapnell, C., Pop, M., and Salzberg, S. L. (2009). Ultrafast and memory -efficient alignment of short DNA sequences to the human genome. Genome Biol 10, R25.

Lenz, G., Wright, G., Dave, S. S., Xiao, W., Powell, J., Zhao, H., Xu, W., Tan, B., Goldschmidt, N., Iqbal, J., et al. (2008a). Stromal gene signatures in large-B-cell lymphomas. N Engl J Med 359, 2313-2323.

Lenz, G., Wright, G. W., Emre, N. C., Kohlhammer, H., Dave, S. S., Davis, R. E., Carty, S., Lam, L. T., Shaffer, A. L., Xiao, W., et al. (2008b). Molecular subtypes of diffuse large B-cell lymphoma arise by distinct genetic pathways. Proc Natl Acad Sci U S A 105, 13520-13525.

Linossi, E. M., and Nicholson, S. E. (2015). Kinase inhibition, competitive binding and proteasomal degradation: resolving the molecular function of the suppressor of cytokine signaling (SOCS) proteins. Immunological reviews 266, 123-133.

Lionakis, M. S., Dunleavy, K., Roschewski, M., Widemann, B. C., Butman, J. A., Schmitz, R., Yang, Y., Cole, D. E., Melani, C., Higham, C. S., et al. (2017). Inhibition of B Cell Receptor Signaling by Ibmtinib in Primary CNS Lymphoma. Cancer Cell 31, 833-843 e835.

Ljungstrom, V., Cortese, D., Young, E., Pandzic, T., Mansouri, L., Plevova, K., Ntoufa, S., Baliakas, P., Clifford, R., Sutton, L. A., et al. (2016). Whole-exome sequencing in relapsing chronic lymphocytic leukemia: clinical impact of recurrent RPS15 mutations. Blood 127, 1007-1016.

Love, C., Sun, Z., Jima, D., Li, G., Zhang, J., Miles, R., Richards, K. L., Dunphy, C. H., Choi, W. W., Srivastava, G., et al. (2012). The genetic landscape of mutations in Burkitt lymphoma. Nat Genet 44, 1321-1325.

Lu, D., Liu, L., Ji, X., Gao, Y., Chen, X., Liu, Y., Liu, Y., Zhao, X., Li, Y., Li, Y., et al. (2015). The phosphatase DUSP2 controls the activity of the transcription activator STAT3 and regulates TH17 differentiation. Nat Immunol 16, 1263-1273.

Lu, E., Wolfreys, F. D., Muppidi, J. R., Xu, Y., and Cyster, J. G. (2019). S-Geranylgeranyl-L-glutathione is a ligand for human B cell-confinement receptor P2RY8. Nature 567, 244-248.

Mareschal, S., Pham-Ledard, A., Viailly, P. J., Dubois, S., Bertrand, P., Maingonnat, C., Fontanilles,

M., Bohers, E., Ruminy, P., Toumier, L, et al. (2017). Identification of Somatic Mutations in Primary Cutaneous Diffuse Large B-Cell Lymphoma, Leg Type by Massive Parallel Sequencing. J Invest Dermatol 137, 1984-1994.

Martinez, N., Almaraz, C., Vaque, J. P., Varela, L, Derdak, S., Beltran, S., Mollejo, M., Campos- Martin, Y., Agueda, L., Rinaldi, A., et al. (2014). Whole-exome sequencing in splenic marginal zone lymphoma reveals mutations in genes involved in marginal zone differentiation. Leukemia 28,

1334-1340.

Mathews Griner, L. A., Guha, R., Shinn, P., Young, R. M., Keller, J. M., Liu, D., Goldlust, I. S., Yasgar, A., McKnight, C., Boxer, M. B., et al. (2014). High-throughput combinatorial screening identifies drugs that cooperate with ibrutinib to kill activated B-cell-like diffuse large B-cell lymphoma cells.

Proc Natl Acad Sci U S A 111, 2349-2354.

Milpied, P., Cervera-Marzal, L, Mollichella, M. L., Tesson, B., Brisou, G., Traverse-Glehen, A., Salles,

G., Spinelli, L., and Nadel, B. (2018). Human germinal center transcriptional programs are desynchronized inB cell lymphoma. Nat Immunol 19, 1013-1024.

Mintz, M. A., Felce, J. H., Chou, M. Y., Mayya, V., Xu, Y., Shui, J. W., An, J., Li, Z., Marson, A., Okada, T., et al. (2019). The HVEM-BTLA Axis Restrains T Cell Help to Germinal Center B Cells and Functions as a Cell-Extrinsic Suppressor in Lymphomagenesis,. Immunity In press.

Mittal, S. K., and Roche, P. A. (2015). Suppression of antigen presentation by IL-10. Curr Opin Immunol 34, 22-27.

Mlynarczyk, C., Fontan, L., andMelnick, A. (2019). Germinal center-derived lymphomas: The darkest side of humoral immunity. Immunological reviews 288, 214-239.

Monti, S., Chapuy, B., Takeyama, K., Rodig, S. J., Hao, Y., Yeda, K. T., Inguilizian, H., Mermel, C., Currie, T., Dogan, A., et al. (2012). Integrative analysis reveals an outcome-associated and targetable pattern of p53 and cell cycle deregulation in diffuse large B cell lymphoma. Cancer Cell 22, 359-372.

Muppidi, J. R., Schmitz, R., Green, J. A., Xiao, W., Larsen, A. B., Braun, S. E., An, J., Xu, Y., Rosenwald, A., Ott, G., et al. (2014). Loss of signalling via Galphal3 in germinal centre B -cell-derived lymphoma. Nature 516, 254-258.

Nakamura, T., Tateishi, K., Niwa, T., Matsushita, Y., Tamura, K., Kinoshita, M., Tanaka, K., Fukushima, S., Takami, H., Arita, H., et al. (2016). Recurrent mutations of CD79B and MYD88 are the hallmark of primary central nervous system lymphomas. Neuropathology and applied neurobiology 42, 279-290.

Ngo, V. N., Davis, R. E., Lamy, L., Yu, X., Zhao, H., Lenz, G., Lam, L. T., Dave, S., Yang, L., Powell, L, and Staudt, L. M. (2006). A loss-of-function RNA interference screen for molecular targets in cancer. Nature 441, 106-110.

Ngo, V. N., Young, R. M., Schmitz, R., Jhavar, S., Xiao, W., Lim, K. H., Kohlhammer, H., Xu, W., Yang, Y., Zhao, H., et al. (2011). Oncogenically active MYD88 mutations in human lymphoma. Nature 470, 115-119.

Nogai, H., Wenzel, S. S., Hailfinger, S., Grau, M., Kaergel, E., Seitz, V., Wollert-Wulf, B., Pfeifer, M., Wolf, A., Frick, M., et al. (2013). IkappaB-zeta controls the constitutive NF-kappaB target gene network and survival of ABC DLBCL. Blood 122, 2242-2250.

Noy, A., de Vos, S., Thieblemont, C., Martin, P., Flowers, C. R., Morschhauser, F., Collins, G. P., Ma,

S., Coleman, M., Peles, S., et al. (2017). Targeting Bruton tyrosine kinase with ibrutinib in relapsed/refractory marginal zone lymphoma. Blood 129, 2224-2232.

Okosun, J., Bodor, C., Wang, J., Araf, S., Yang, C. Y., Pan, C., Boiler, S., Cittaro, D., Bozek, M., Iqbal,

S., et al. (2014). Integrated genomic analysis identifies recurrent mutations and evolution patterns driving the initiation and progression of follicular lymphoma. Nat Genet 46, 176-181.

Okosun, J., Wolfson, R. L., Wang, J., Araf, S., Wilkins, L., Castellano, B. M., Escudero-Ibarz, L., Al Seraihi, A. F., Richter, J., Bemhart, S. H., et al. (2016). Recurrent mTORCl -activating RRAGC mutations in follicular lymphoma. Nat Genet 48, 183-188.

Parry, M., Rose-Zerilli, M. J., Gibson, J., Ennis, S., Walewska, R., Forster, J., Parker, H., Davis, Z., Gardiner, A., Collins, A., et al. (2013). Whole exome sequencing identifies novel recurrently mutated genes in patients with splenic marginal zone lymphoma. PLoS One 8, e83244.

Parry, M., Rose-Zerilli, M. J., Ljungstrom, V., Gibson, J., Wang, J., Walewska, R., Parker, H., Parker,

A., Davis, Z., Gardiner, A., et al. (2015). Genetics and Prognostication in Splenic Marginal Zone Lymphoma: Revelations from Deep Sequencing. Clin Cancer Res 21, 4174-4183.

Pasqualucci, L., and Dalla-Favera, R. (2018). Genetics of diffuse large B-cell lymphoma. Blood 131, 2307-2319.

Pasqualucci, L., Khiabanian, H., Fangazio, M., Vasishtha, M., Messina, M., Holmes, A. B., Ouillette,

P., Trifonov, V., Rossi, D., Tabbo, F., et al. (2014). Genetics of follicular lymphoma transformation.

Cell reports 6, 130-140.

Phelan, J. D., Young, R. M., Webster, D. E., Roulland, S., Wright, G. W., Kasbekar, M., Shaffer, A. L., 3rd, Ceribelli, M., Wang, J. Q., Schmitz, R., et al. (2018). A multiprotein supercomplex controlling oncogenic signalling in lymphoma. Nature 560, 387-391.

Phillips, T. J., Forero-Torres, A., Sher, T., Diefenbach, C. S., Johnston, P., Talpaz, M., Pulini, J., Zhou, L., Scherle, P., Chen, X., and Barr, P. M. (2018). Phase 1 study of the PBKdelta inhibitor INCB040093 +/- JAK1 inhibitor itacitinib in relapsed/refractory B-cell lymphoma. Blood 132, 293-306.

Pillonel, V., Juskevicius, D., Ng, C. K. Y., Bodmer, A., Zettl, A., Jucker, D., Dimhofer, S., and Tzankov, A. (2018). High-throughput sequencing of nodal marginal zone lymphomas identifies recurrent BRAF mutations. Leukemia 32, 2412-2426.

Puente, X. S., Bea, S., Valdes-Mas, R., Villamor, N., Gutierrez-Abril, J., Martin-Subero, J. L, Munar,

M., Rubio-Perez, C., Jares, P., Aymerich, M., et al. (2015). Non-coding recurrent mutations in chronic lymphocytic leukaemia. Nature 526, 519-524.

Quesada, V., Conde, L., Villamor, N., Ordonez, G. R., Jares, P., Bassaganyas, L., Ramsay, A. J., Bea, S., Pinyol, M., Martinez-Trillos, A., et al. (2011). Exome sequencing identifies recurrent mutations of the splicing factor SF3B1 gene in chronic lymphocytic leukemia. Nat Genet 44, 47-52.

Richter, J., Schlesner, M., Hoffmann, S., Kreuz, M., Leich, E., Burkhardt, B., Rosolowski, M., Ammerpohl, O., Wagener, R., Bemhart, S. H., et al. (2012). Recurrent mutation of the ID3 gene in Burkitt lymphoma identified by integrated genome, exome and transcriptome sequencing. Nat Genet 44, 1316-1320.

Rosenwald, A., Wright, G., Chan, W. C., Connors, J. M., Campo, E., Fisher, R. L, Gascoyne, R. D., Muller-Hermelink, H. K., Smeland, E. B., Giltnane, J. M., et al. (2002). The use of molecular profiling to predict survival after chemotherapy for diffuse large-B-cell lymphoma. N Engl J Med 346, 1937- 1947.

Rossi, D., Trifonov, V., Fangazio, M., Bruscaggin, A., Rasi, S., Spina, V., Monti, S., Vaisitti, T., Amiga, F., Fama, R., et al. (2012). The coding genome of splenic marginal zone lymphoma: activation of NOTCH2 and other pathways regulating marginal zone development. J Exp Med 209, 1537-1551.

Rui, L., Drennan, A. C., Ceribelli, M., Zhu, F., Wright, G. W., Huang, D. W., Xiao, W., Li, Y., Grindle, K. M., Lu, L., et al. (2016). Epigenetic gene regulation by Janus kinase 1 in diffuse large B-cell lymphoma. Proc Natl Acad Sci U S A 113, E7260-E7267.

Saito, T., Chiba, S., Ichikawa, M., Kunisato, A., Asai, T., Shimizu, K., Yamaguchi, T., Yamamoto, G., Seo, S., Kumano, K., et al. (2003). Notch2 is preferentially expressed in mature B cells and indispensable for marginal zone B lineage development. Immunity 18, 675-685.

Scherer, F., Kurtz, D. M., Newman, A. M., Stehr, H., Craig, A. F., Esfahani, M. S., Lovejoy, A. F., Chabon, J. J., Klass, D. M., Liu, C. L., et al. (2016). Distinct biological subtypes and patterns of genome evolution in lymphoma revealed by circulating tumor DNA. Science translational medicine 8, 364ral55.

Schmitz, R., Wright, G. W., Huang, D. W., Johnson, C. A., Phelan, J. D., Wang, J. Q., Roulland, S., Kasbekar, M., Young, R. M., Shaffer, A. L., et al. (2018). Genetics and Pathogenesis of Diffuse Large B-Cell Lymphoma. N Engl J Med 378, 1396-1407.

Schmitz, R., Young, R. M., Ceribelli, M., Jhavar, S., Xiao, W., Zhang, M., Wright, G., Shaffer, A. L., Hodson, D. J., Buras, E., et al. (2012). Burkitt lymphoma pathogenesis and therapeutic targets from structural and functional genomics. Nature 490, 116-120.

Schrader, A. M. R., Jansen, P. M, Willemze, R., Vermeer, M. H., Cleton-Jansen, A. M., Somers, S. F., Veelken, H., van Eijk, R., Kraan, W., Kersten, M. J., et al. (2018). High prevalence of MYD88 and CD79B mutations in intravascular large B-cell lymphoma. Blood 131, 2086-2089.

Schuhmacher, B., Bein, J., Rausch, T., Benes, V., Tousseyn, T., Vomanen, M., Ponzoni, M., Thumer,

L., Gascoyne, R., Steidl, C., et al. (2019). JUNB, DUSP2, SGK1, SOCS1 and CREBBP are frequently mutated in T-cell/histiocyte-rich large B-cell lymphoma. Haematologica 104, 330-337.

Scott, D. W., Mottok, A., Ennishi, D., Wright, G. W., Farinha, P., Ben-Neriah, S., Kridel, R., Barry, G.

S., Hother, C., Abrisqueta, P., et al. (2015). Prognostic Significance of Diffuse Large B-Cell Lymphoma Cell of Origin Determined by Digital Gene Expression in Formalin-Fixed Paraffin- Embedded Tissue Biopsies. J Clin Oncol 33, 2848-2856.

Sha, C., Barrans, S., Cucco, F., Bentley, M. A., Care, M. A., Cummin, T., Kennedy, H., Thompson, J. S., Uddin, R., Worrillow, L., et al. (2019). Molecular High-Grade B-Cell Lymphoma: Defining a Poor-Risk Group That Requires Different Approaches to Therapy. J Clin Oncol 37, 202-212.

Shaffer, A. L., Rosenwald, A., Hurt, E. M., Giltnane, J. M., Lam, L. T., Pickeral, O. K., and Staudt, L. M. (2001). Signatures of the immune response. Immunity 15, 375-385.

Shaffer, A. L., Wright, G., Yang, L., Powell, J., Ngo, V., Lamy, L., Lam, L. T., Davis, R. E., and Staudt, L.

M. (2006). A library of gene expression signatures to illuminate normal and pathological lymphoid biology. Immunological reviews 210, 67-85.

Shechter, R., London, A., and Schwartz, M. (2013). Orchestrated leukocyte recruitment to immuneprivileged sites: absolute barriers versus educational gates. Nat Rev Immunol 13, 206-218.

Shin, S. H., Kim, Y. J., Lee, D., Cho, D., Ko, Y. H., Cho, J., Park, W. Y., Park, D., Kim, S. J., and Kim, W. S. (2019). Analysis of circulating tumor DNA by targeted ultra-deep sequencing across various non- Hodgkin lymphoma subtypes. Leuk Lymphoma 60, 2237-2246.

Spina, V., Khiabanian, H., Messina, M., Monti, S., Cascione, L., Bruscaggin, A., Spaccarotella, E.,

Holmes, A. B., Arcaini, L., Lucioni, M., et al. (2016). The genetics of nodal marginal zone lymphoma.

Blood 128, 1362-1373.

Steidl, C., Shah, S. P., Woolcock, B. W., Rui, L., Kawahara, M., Farinha, P., Johnson, N. A., Zhao, Y., Telenius, A., Neriah, S. B., et al. (2011). MHC class II transactivator CIITA is a recurrent gene fusion partner in lymphoid cancers. Nature 471, 377-381.

Suehara, Y., Sakata-Yanagimoto, M., Hattori, K., Nanmoku, T., Itoh, T., Kaji, D., Yamamoto, G., Abe,

Y., Narita, K., Takeuchi, M., et al. (2018). Liquid biopsy for the identification of intravascular large Bcell lymphoma. Haematologica 103, e241-e244.

Tallen, G., andRiabowol, K. (2014). Keep-ING balance: tumor suppression by epigenetic regulation.

FEBS Lett 588, 2728-2742. Timens, W., Visser, L., and Poppema, S. (1986). Nodular lymphocyte predominance type of Hodgkin's disease is a germinal center lymphoma. Lab Invest 54, 457-461.

Tsukamoto, T., Nakano, M., Sato, R., Adachi, H., Kiyota, M., Kawata, E., Uoshima, N., Yasukawa, S., Chinen, Y., Mizutani, S., et al. (2017). High-risk follicular lymphomas harbour more somatic mutations including those in the AID-motif. Sci Rep 7, 14039.

Van Den Neste, E., Andre, M., Gastinne, T., Stamatoullas, A., Haioun, C., Belhabri, A., Reman, O., Casasnovas, O., Ghesquieres, H., Verhoef, G., et al. (2018). A phase II study of the oral JAK1/JAK2 inhibitor ruxolitinib in advanced relapsed/refractory Hodgkin lymphoma. Haematologica 103, 840- 848.

Vater, L, Montesinos-Rongen, M., Schlesner, M., Haake, A., Purschke, F., Spmte, R., Mettenmeyer,

N., Nazzal, L, Nagel, L, Gutwein, I, et al. (2015). The mutational pattern of primary lymphoma of the central nervous system determined by whole-exome sequencing. Leukemia 29, 677-685.

Wang, L., Lawrence, M. S., Wan, Y., Stojanov, P., Sougnez, C., Stevenson, K., Werner, L., Sivachenko, A., DeLuca, D. S., Zhang, L., et al. (2011). SF3B 1 and other novel cancer genes in chronic lymphocytic leukemia. N Engl J Med 365, 2497-2506.

Weinstein, J. S., Herman, E. L, Lainez, B., Licona-Limon, P., Esplugues, E., Flavell, R., and Craft, J. (2016). TFH cells progressively differentiate to regulate the germinal center response. Nat Immunol 17, 1197-1205.

Wilson, W. H., Popplewell, L. L., Phillips, T., Kimball, A. S., Chhabra, S., Ping, J., Neuenburg, J., Cavazos, N., Staudt, L. M., and de Vos, S. (2015a). Multicenter Phase lb Dose-Escalation Study of Ibrutinib and Lenalidomide Combined with Dose-Adjusted EPOCH-R in Patients with Relapsed/Refractory DLBCL. Blood 126, 1527-1527.

Wilson, W. H., Young, R. M., Schmitz, R., Yang, Y., Pittaluga, S., Wright, G., Lih, C. J., Williams, P. M., Shaffer, A. L., Gerecitano, J., et al. (2015b). Targeting B cell receptor signaling with ibrutinib in diffuse large B cell lymphoma. Nat Med 21, 922-926.

Wu, C., de Miranda, N. F., Chen, L., Wasik, A. M., Mansouri, L., Jurczak, W., Galazka, K., Dlugosz- Danecka, M., Machaczka, M., Zhang, H., et al. (2016). Genetic heterogeneity in primary and relapsed mantle cell lymphomas: Impact of recurrent CARDll mutations. Oncotarget 7, 38180- 38190.

Yang, Y., Shaffer, A. L., 3rd, Emre, N. C., Ceribelli, M., Zhang, M., Wright, G., Xiao, W., Powell, J., Platig, J., Kohlhammer, H., et al. (2012). Exploiting synthetic lethality for the therapy of ABC diffuse large B cell lymphoma. Cancer Cell 21, 723-737.

Ye, H., Remstein, E. D., Bacon, C. M., Nicholson, A. G., Dogan, A., and Du, M. Q. (2008).

Chromosomal translocations involving BCL6 in MALT lymphoma. Haematologica 93, 145-146.

Yonese, L, Takase, H., Yoshimori, M., Onozawa, E., Tsuzura, A., Miki, T., Mochizuki, M., Miura, O., and Arai, A. (2019). CD79B mutations in primary vitreoretinal lymphoma: Diagnostic and prognostic potential. European journal of haematology 102, 191-196.

Young, K. H., Weisenburger, D. D., Dave, B. J., Smith, L., Sanger, W., Iqbal, J., Campo, E., Delabie, J.,

Gascoyne, R. D., Ott, G., et al. (2007). Mutations in the DNA-binding codons of TP53, which are associated with decreased expression of TRAILreceptor-2, predict for poor survival in diffuse large B-cell lymphoma. Blood 110, 4396-4405.

Young, R. M., Phelan, J. D., Wilson, W. H., and Staudt, L. M. (2019). Pathogenic B-cell receptor signaling in lymphoid malignancies: New insights to improve treatment. Immunological reviews 291, 190-213.

Young, R. M., Wu, T., Schmitz, R., Dawood, M., Xiao, W., Phelan, J. D., Xu, W., Menard, L., Meffre,

E., Chan, W. C, et al. (2015). Survival of human lymphoma cells requires B-cell receptor engagement by self-antigens. Proc Natl Acad Sci U S A 112, 13447-13454.

Zamo, A., Pischimarov, J., Schlesner, M., Rosenstiel, P., Bomben, R., Horn, H., Grieb, T., Nedeva, T.,

Lopez, C., Haake, A., et al. (2018). Differences between BCL2-break positive and negative follicular lymphoma unraveled by whole-exome sequencing. Leukemia 32, 685-693.

Zeller, K. L, Zhao, X., Lee, C. W., Chiu, K. P., Yao, F., Yustein, J. T., Ooi, H. S., Orlov, Y. L., Shahab, A.,

Yong, H. C., et al. (2006). Global mapping of c-Myc binding sites and target gene networks in human B cells. Proc Natl Acad Sci U S A 103, 17834-17839.

Zhang, J., Jima, D., Moffitt, A. B., Liu, Q., Czader, M., Hsi, E. D., Fedoriw, Y., Dunphy, C. H., Richards,

K. L., Gill, J. L, et al. (2014). The genomic landscape of mantle cell lymphoma is related to the epigenetically determined chromatin state of normal B cells. Blood 123, 2988-2996.

Zhou, X. A., Louissaint, A., Jr., Wenzel, A., Yang, J., Martinez-Escala, M. E., Moy, A. P., Morgan, E. A.,

Paxton, C. N., Hong, B., Andersen, E. F., et al. (2018a). Genomic Analyses Identify Recurrent Alterations in Immune Evasion Genes in Diffuse Large B-Cell Lymphoma, Leg Type. J Invest Dermatol 138, 2365-2376.

Zhou, Y., Liu, W., Xu, Z., Zhu, H., Xiao, D., Su, W., Zeng, R., Feng, Y., Duan, Y., Zhou, J., and Zhong, M.

(2018b). Analysis of Genomic Alteration in Primary Central Nervous System Lymphoma and the Expression of Some Related Genes. Neoplasia 20, 1059-1069.

[0219] All references, including publications, patent applications, and patents, cited herein are hereby incorporated by reference to the same extent as if each reference were individually and specifically indicated to be incorporated by reference and were set forth in its entirety herein. [0220] The use of the terms “a” and “an” and “the” and “at least one” and similar referents in the context of describing the invention (especially in the context of the following claims) are to be construed to cover both the singular and the plural, unless otherwise indicated herein or clearly contradicted by context. The use of the term “at least one” followed by a list of one or more items (for example, “at least one of A and B”) is to be construed to mean one item selected from the listed items (A or B) or any combination of two or more of the listed items (A and B), unless otherwise indicated herein or clearly contradicted by context. The terms “comprising,” “having,” “including,” and “containing” are to be construed as open-ended terms (i.e., meaning “including, but not limited to,”) unless otherwise noted. Recitation of ranges of values herein are merely intended to serve as a shorthand method of referring individually to each separate value falling within the range, unless otherwise indicated herein, and each separate value is incorporated into the specification as if it were individually recited herein. All methods described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The use of any and all examples, or exemplary language (e.g., “such as”) provided herein, is intended merely to better illuminate the invention and does not pose a limitation on the scope of the invention unless otherwise claimed. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the invention.

[0221] Preferred embodiments of this invention are described herein, including the best mode known to the inventors for carrying out the invention. Variations of those preferred embodiments may become apparent to those of ordinary skill in the art upon reading the foregoing description. The inventors expect skilled artisans to employ such variations as appropriate, and the inventors intend for the invention to be practiced otherwise than as specifically described herein. Accordingly, this invention includes all modifications and equivalents of the subject matter recited in the claims appended hereto as permitted by applicable law. Moreover, any combination of the above-described elements in all possible variations thereof is encompassed by the invention unless otherwise indicated herein or otherwise clearly contradicted by context.

Claims

CLAIMS:

1. A method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, without use of gene copy number, the method comprising:

(a) (i) detecting in a biopsy sample of lymphoma two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1A, or

(ii) detecting in a biopsy sample of lymphoma two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB;

(b) determining whether at least two weight values listed in the Table 1 A or IB are positive for each of the subtypes BN2, EZB, MCD, and ST2 based on the genetic features detected in (a);

(c) calculating the probabilities that the sample belongs to

(i) subtype Nl, and

(ii) each subtype BN2, EZB, MCD, and ST2 determined in (b) to have at least two positive weight values using the equation

wherein P( subtype] V) is the probability that the sample is of a subtype, wherein each V is a weight value listed in the Table 1 A or IB based on the genetic features detected in (a) for that subtype, wherein i is the number of weight values used to calculate P(subtype|E), and wherein P{ subtype] V) is calculated for each subtype using only V, weight values for that subtype;

(d) designating each subtype BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (b) to have P(subtype|E) < 0.5; and (e) classifying the sample with the probabilities calculated in (c) and any probabilities designated in (d) using the logic:

(1) if only one subtype has /’(subtype)!⁷) > 0.9 with the other subtypes having P( subtype] V) < 0.9, the sample is classified as being the one subtype having P( subtype] V) > 0.9;

(2) if only one subtype has ./(subtype) J⁷) > 0.5 with the other subtypes having P{ subtype] V) < 0.5, the sample is classified as being the one subtype having P( subtype] V) > 0.5; and

(3) if there is more than one subtype having P( subtype] V) > 0.9, the sample is classified as being each of the subtypes having ./(subtype) J⁷) > 0.9.

2. The method of claim 1, wherein the detection of (a) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 1A.

3. The method of claim 2, wherein the detection of (a) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table 1 A and detecting in the sample all of the genetic features of the subtype N1 listed in Table 1A.

4. The method of claim 1, wherein the detection of (a) comprises detecting in the sample two or more genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample one or more genetic features of the subtype N1 listed in Table IB.

5. The method of claim 4, wherein the detection of (a) comprises detecting in the sample all of the genetic features of each of the subtypes BN2, EZB, MCD, and ST2 listed in Table IB and detecting in the sample all of the genetic features of the subtype N1 listed in Table IB.

6. A method of selecting a treatment for a human subject with diffuse large B cell lymphoma (DLBCL), the method comprising:

(A) classifying a DLBCL according to any one of claims 1-5; and

(B) selecting the treatment of the subject based on the one or more classifications of

(A).

7. The method of claim 6, wherein the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

8. The method of claim 6, wherein the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

9. The method of claim 6, wherein the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

10. The method of claim 6, wherein the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1 or an immune checkpoint inhibitor.

11. The method of claim 6, wherein the DLBCL is classified as ST2, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl.

12. A method of classifying a diffuse large B cell lymphoma (DLBCL) of a human subject, with use of gene copy number, the method comprising:

(a) (i) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype Nl listed in Table 2A, or (ii) detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2B, wherein a copy number of each of the subtypes is a genetic feature detected;

(b) determining whether at least two weight values listed in the Table 2A or 2B are positive for each of the subtypes A53, BN2, EZB, MCD, and ST2 based on the genetic features detected in (a);

(c) calculating the probabilities that the sample belongs to

(i) subtype Nl, and

(ii) each subtype A53, BN2, EZB, MCD, and ST2 determined in (b) to have at least two positive weight values using the equation

wherein P( subtype] V) is the probability that the sample is of a subtype, wherein each Vi is a weight value listed in the Table 2A or 2B based on the genetic features detected in (a) for that subtype, wherein i is the number of weight values used to calculate (subtype|F), and wherein P( subtype] V) is calculated for each subtype using only V, weight values for that subtype;

(d) designating each subtype A53, BN2, EZB, MCD, and ST2 determined not to have at least two positive weight values determined in (b) to have (subtype|F) < 0.5; and

(e) classifying the sample with the probabilities calculated in (c) and any probabilities designated in (d) using the logic:

(1) if only one subtype has P(subtype|J⁷) > 0.9 with the other subtypes having P{ subtype] V) < 0.9, the sample is classified as being the one subtype having P( subtype] V) > 0.9; (2) if only one subtype has /’(subtype) V) > 0.5 with the other subtypes having /’(subtype) V) < 0.5, the sample is classified as being the one subtype having /’(subtype) V) > 0.5; and

13. The method of claim 12, wherein the detection of (a) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2A.

14. The method of claim 13, wherein the detection of (a) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2A and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2A.

15. The method of claim 12, wherein the detection of (a) comprises detecting in the sample two or more genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample one or more genetic features of the subtype N1 listed in Table 2B.

16. The method of claim 15, wherein the detection of (a) comprises detecting in the sample all of the genetic features of each of the subtypes A53, BN2, EZB, MCD, and ST2 listed in Table 2B and detecting in the sample all of the genetic features of the subtype N1 listed in Table 2B.

17. A method of selecting a treatment for a human subject with diffuse large B cell lymphoma (DLBCL), the method comprising:

(A) classifying a DLBCL according to any one of claims 12-16; and (B) selecting the treatment of the subject based on the one or more classifications of

(A).

18. The method of claim 17, wherein the DLBCL is classified as BN2, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

19. The method of claim 17, wherein the DLBCL is classified as EZB, and the treatment is administration of an effective amount of tazemetostat or CPI-1205.

20. The method of claim 17, wherein the DLBCL is classified as MCD, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.

21. The method of claim 17, wherein the DLBCL is classified as Nl, and the treatment is administration of an effective amount of an inhibitor of NOTCH1 or an immune checkpoint inhibitor.

22. The method of claim 17, wherein the DLBCL is classified as ST2, and the treatment is administration of an effective amount of an inhibitor of PI3 kinase or mTORCl.

23. The method of claim 17, wherein the DLBCL is classified as A53, and the treatment is administration of an effective amount of an inhibitor of B cell receptor-dependent NF-KB signaling.