Exploration of Tumor Biopsy Gene Signatures to Understand the Role of the Tumor Microenvironment in Outcomes to Lisocabtagene Maraleucel

The tumor microenvironment (TME) is an important predictor of outcomes in B-cell non-Hodgkin lymphoma (NHL). TME and tumor-associated macrophage stromal gene signatures have been associated with adverse outcomes to standard anthracycline and rituximab-based chemoimmunotherapy in patients with diffuse large B-cell lymphoma (DLBCL; ref. 1). The TME is also highly variable between different lymphoma subtypes and appears to impact prognosis and outcomes to novel immuno-oncology agents and targeted inhibitors (2–7). Pathogenesis of DLBCL involves evasion of immune recognition by T and natural killer cells as well as activation of immunosuppressive mechanisms [e.g., programmed cell death ligand 1 (PD-L1) pathway]. Multiple lymphoma studies have explored the role of the TME in response to checkpoint blockade (5, 8, 9). Specifically, the lymphoma TME has been described as exhibiting a spectrum between a more inflammatory environment such as in Hodgkin lymphoma, and a less inflammatory environment facilitating escape from immune surveillance observed in follicular lymphoma (FL) and DLBCL. On the basis of the relative responsiveness of these lymphomas to checkpoint inhibition, a more inflammatory TME has been associated with responsiveness to checkpoint inhibition (5).

How the TME affects CD19-directed chimeric antigen receptor (CAR) T-cell therapy is less well understood. Preclinical in vitro and in vivo evidence indicates that immunosuppressive macrophages can inhibit CAR T-cell function (10). Translational data from clinical B-cell NHL studies has shown that an inflammatory pretreatment TME with elevated CD3⁺ T cells correlated with response to CAR T-cell therapy (11–13).

Lisocabtagene maraleucel (liso-cel) is a CD19-directed, autologous 4-1BB co-stimulated CAR T-cell product administered at equal target doses of CD8⁺ and CD4⁺ CAR⁺ T cells (14). The TRANSCEND NHL 001 (TRANSCEND; NCT02631044) study enrolled patients with relapsed/refractory (R/R) large B-cell lymphoma (LBCL) who had progressed after two or more prior treatments. Eligible patients had DLBCL not otherwise specified (NOS), high-grade B-cell lymphoma with rearrangements of MYC and BCL2 and/or BCL6 (HGL), primary mediastinal B-cell lymphoma (PMBCL), LBCL transformed from indolent B-cell NHL, and grade 3B FL. Patients were heavily pretreated with a median of three prior lines of therapy, and two-thirds of patients were considered chemotherapy-refractory. The complete response (CR) rate after liso-cel treatment was 53%, with a low observed incidence of severe cytokine release syndrome (2%) and neurologic events (10%; ref. 15). Despite these encouraging results in a very high-risk patient population, and a probability of continued response at 2 years of 49.5% [95% confidence interval (CI), 41.4–57.0; ref. 16], not all patients achieve a CR or a durable response to liso-cel. Therefore, it is important to better understand what factors may influence response and durability.

Here, we examine tumor biopsies with associated long-term follow-up from a cohort of liso-cel–treated patients with LBCL to understand how the pretreatment TME may influence response. In addition, we identify changes in the posttreatment TME that are associated with the pretreatment TME and response.

The efficacy outcomes of TRANSCEND have been reported previously (15). The primary efficacy endpoint was objective response rate, defined as the proportion of patients who achieved a best overall response (BOR) of CR or partial response (PR) based on assessment by the independent review committee (IRC) per Lugano 2014 criteria (17). Protocol-specified efficacy assessments occurred at 1, 3, 6, 9, and 12 months after liso-cel infusion, respectively, and then every 6 months thereafter until study completion. For this study, main analyses are based on the assessments of CR and progressive disease (PD) per IRC to focus on patient subgroups with the largest differences in response. For the analyses based on assessment time points of 1, 3, 6, 9, or 12 months, a patient is classified as CR if the assessment was CR at that time point, and PD if a patient had progressed at any time up to that time point. Most analyses compared patients who achieved and maintained a CR (N = 27) at 3 months (month-3 CR) to those with PD (N = 32) by 3 months (month-3 PD).

Needle biopsy samples were obtained from patients with R/R LBCL who were enrolled in the TRANSCEND study (Table 1). The first biopsy was obtained before lymphodepleting chemotherapy, which preceded liso-cel infusion [day < 0 pretreatment (PRE)] by 2 to 7 days. The PRE biopsy was required per protocol unless an archival biopsy with no intervening treatments was available. In many cases, the PRE biopsy was not available for RNA sequencing (RNA-seq) analysis because the quality or quantity of the sample was not sufficient or because the biopsy material was needed for protocol-required disease confirmation.

An optional biopsy was obtained 11 days (±4 days) post-infusion [day 11 posttreatment (D11)], which was near the anticipated day of maximal liso-cel expansion. The D11 biopsy site was not necessarily the same location as the PRE biopsy site, as the site of both biopsies was at investigator discretion. Because the D11 biopsy depended on presence of accessible tumor, it could have led to a sample collection bias, so we compared the baseline characteristics of the patients with a D11 biopsy with those who had a pretreatment sample (Table 1).

The needle biopsy samples were formalin-fixed, paraffin-embedded (FFPE) material and were shipped to Charles River Laboratories or to Juno Therapeutics, a Bristol-Myers Squibb Company. Slides were cut from the FFPE blocks and were examined by a pathologist to determine and record the amount of tumor present, which was roughly 5 to 20 mm² and 20% to 100% tumor cell content, with most samples >60%.

In the TRANSCEND study (15), 269 patients with LBCL were infused with liso-cel and followed for a median of 12 months. Seventy-eight PRE samples and 73 D11 samples yielded quality RNA-seq data. To determine whether the RNA-seq study population was reflective of the full study population, we compared the response rates and progression-free survival (PFS) in the full study population with the patients in the PRE or D11 sets.

Institutional review boards at participating institutions (listed in Supplementary Appendix) approved the study protocol and amendments. The clinical study was done in accordance with the Declaration of Helsinki, International Conference on Harmonization Good Clinical Practice guidelines, and applicable regulatory requirements. All patients provided written informed consent for the use of biopsy material for analyses.

At Q² Solutions/EA Genomics, RNA was extracted from slides or curls from FFPE biopsy material without macrodissection. After extraction, TruSeq RNA Exome kits (Illumina) were used to yield barcoded libraries and sequencing was performed using Illumina HiSeq 2500 instruments. The target sequencing depth was 25 million paired-end reads for the first two batches and 50 million for the remaining four batches, with a sequencing length of 50 bp.

The quality of the reads in the fastq files was evaluated using FastQC (version 0.11.5). Illumina adapters were trimmed using cutadapt (version 1.15) and the resulting fastq files were used for downstream analyses. Sequence alignment was performed using STAR software (version 2.5.2b) using a two-pass mode and human genome reference hg38 (GENCODE GRCh38.p5 and the annotation from GENCODE Release 24). Counts were obtained using the quantMode GeneCounts option. Furthermore, the CAR sequence of liso-cel was joined to the genome so that the CAR sequence could be quantified. Alignments were evaluated using Picard tools (version 2.18) and samples with low-quality sequencing metrics were removed from analysis.

To assess how CAR T cells impacted the TME, PRE and D11 biopsies were compared. The final data set with passing quality RNA-seq data had 78 PRE samples and 73 D11 samples, with matching PRE and D11 samples for 30 patients.

The samples were collected and processed across six batches over 2.5 years. Although the key methodology was the same (sequencing vendor, library preparation kit, and Illumina sequencing technology), some procedures changed over time, including the total number of reads sequenced, the switch to a dual DNA/RNA extraction protocol, and the inclusion of samples with lower tumor fractions on the slides (Supplementary Table S1). Several batch correction approaches were explored by adding covariates to the DESeq2 (18) differential expression analysis (Supplementary Fig. S1) for the following key contrasts in this study: (A) CR versus PD at 3 months in the PRE samples and (B) D11 versus PRE. In both of those contrasts, using no batch correction yielded the highest number of differentially expressed genes. Adding a categorical covariate for six batches reduced the number of differential genes by ≥50% in the D11 versus PRE contrast. Batch 5 was all PRE samples and Batch 6 was mostly D11, so when they were treated as separate batches, the number of differential genes fell dramatically. Combining Batch 5 and Batch 6 preserved many of the D11 versus PRE differences. Several other covariates in the differential expression model were tested (e.g., percentage of tumor on slide, DV₂₀₀, batches), and the resulting fold-changes highly overlapped with the simple batch model (Supplementary Fig. S2). Given the above, all further analysis was performed using five batches (Batches 1–4 with 5 and 6 combined) as a covariate.

For sample level analyses, such as single sample gene set enrichment analysis (ssGSEA) scores and gene plot figures, we used the expression values adjusted by vst (DESeq2 R package; ref. 18), followed by removeBatchEffect (limma R package; ref. 19) for the five batch groups using the STAR counts as inputs. When performing differential expression analysis between two groups of samples, we used DESeq2 with the raw gene-level counts from the STAR output, with the five batch groups as a covariate. Supplementary Figure S3 shows the comparison of the fold-changes between response groups of the PRE samples.

For gene set enrichment analysis (GSEA; ref. 20), we used the clusterProfiler R package (version 3.14), which uses the R package fGSEA (21) internally and used 1,000 permutations to estimate the significance. With clusterProfiler, we searched the MSigDB (20) H (Hallmark; ref. 22) and C2 (Curated; ref. 20) gene sets as provided by the msigdbr R package (version 7.0.1). We used the GSEA module in ArrayStudio (version 10; Qiagen) to complement the MSigDB-based approach. The EZH2 gene sets, NUYTTEN_EZH2_TARGETS (UP and DN; ref. 23), were identified in the MSigDB C2 and ArrayStudio analysis. Additional EZH2 gene sets were generated from a reanalysis of the Affymetrix microarray data for GEO Series GSE49284 (24), which included three lymphoma cell lines (WSU-DLCL2, KARPAS-422, and SU-DHL-6) that had been treated with and were sensitive to the EZH2 inhibitor EPZ-6438. The three cell lines were analyzed to identify the genes most downregulated or upregulated by EZH2 inhibition. The three individual lists were then merged and named the Knutson_EZH2_inhib (up and dn). Another pair of gene sets identified in the top MSigDB C2 results were the SHIPP_DLBCL_vs_FOLLICULAR_LYMPHOMA (UP and DN; ref. 25).

We added several other lymphoma-related gene sets to our analysis that have been used to characterize lymphoma biology or patient subsets: molecular high-grade signature (26), stromal-1 and -2 (1), activated B-cell versus germinal center B-cell (27), and double-hit positive versus negative (28). We added several immune-related solid tumor gene sets identified in recent publications: IFNγ, Thorsson/Institute for Systems Biology scores (29), and a melanoma T-cell exclusion signature (30).

With these gene sets, we calculated single-sample scores using ssGSEA (with the R package GSVA version 1.42.0). We compared two groups of samples using a Wilcoxon rank sum test (R 3.6.1; refs. 31, 32).

To further explore the biologic differences between DLBCL and FL, we used a proprietary data set, plus NHL data from the International Cancer Genome Consortium (ICGC; https://dcc.icgc.org/releases/current/Projects/MALY-DE) and the Cancer Genome Characterization Initiative (CGCI; https://ocg.cancer.gov/programs/cgci/data-matrix). The proprietary data set was generated from commercially sourced FFPE samples (Avaden Biosciences): 70 DLBCL and 61 FL samples with RNA-seq performed using the same methods as applied to the liso-cel study samples. For the public data sets, ICGC had 40 DLBCL and 41 FL samples, while CGCI had 105 DLBCL and 13 FL samples.

We performed differential expression analysis with the commercially sourced DLBCL and FL samples and selected the top 200 genes for FL and 200 for DLBCL (Supplementary Table S2). These two gene lists were used to calculate ssGSEA scores for each sample and named FL_vs_DLBCL (genes higher in FL than DLBCL) and DLBCL_vs_FL (genes higher in DLBCL than FL). The two ssGSEA scores were combined to form a composite score (FL_like_score or FL-like score) by subtracting the DLBCL_vs_FL score from the FL_vs_DLBCL score.

In a prior study from the sponsor (33), a DLBCL TME classifier was constructed to identify a DLBCL high-immune infiltration phenotype, referred to herein as the DLBCL TME classifier. We applied the RNA-seq version of the classifier algorithm to label samples as TME-classifier positive or TME-classifier negative; otherwise, we used the classifier confidence score as a continuous score for characterization. We also used the RNA-seq data to construct a score and perform classification for the cell of origin (COO; ref. 34) and double-hit gene expression status (28) using the methods described in those publications.

We used Pearson correlations to find genes in PRE samples that were correlated (r > 0.8) to several TME-related genes (CD3E, CD163, COL1A1), and iteratively added genes correlated to any of the new genes converging to 1,051 genes. We then performed community detection based on label propagation (R/igraph; ref. 35) and selected the T-cell community, which had 42 genes. We used R/igraph for the network visualization and colored the nodes using the log₂-fold change of either month-3 CR versus PD in the PRE samples or CD8⁺ versus CD4⁺ log₂-fold change from a dataset of purified immune cells (GSE60424; ref. 36).

We used the internal DLBCL and FL data sets to train an FL classifier using a lasso model (cv.glmnet in R; ref. 37) to predict whether an NHL sample has FL histology (Supplementary Fig. S4). The classifier was then tested on the CGCI and ICGC public data sets and yielded an area under the curve of the receiver operating characteristic curve of 0.93.

Core FL and DLBCL samples to use in the training were determined by removing outliers beyond the 0.9 × interquartile range of both the FL-like and DLBCL-like ssGSEA scores to train the lasso classifier on the most confident samples. In the core training set, after removing the 0.9 × interquartile range outliers, there were 57 DLBCL and 53 FL samples. A lasso model (using cv.glmnet in R; ref. 37) with 10-fold cross-validation (family = “binomial”, standardize = T) was trained on the core samples using the 5,000 genes with the highest standard deviation. This resulted in 30 nonzero coefficients at the minimum lambda. The ssGSEA-based FL-versus-DLBCL score was concordant with the lasso FL classifier for >85% of the PRE samples in this dataset (Supplementary Fig. S5).

We used logistic regression and random forest models to compare importance of transformed FL (tFL) status and the FL-like gene signature with month-3 CR as the outcome. We used all 78 patients who had PRE tumor biopsy RNA-seq data and defined the binary outcome as a 1 if the patient was in CR at month 3 and used a 0 if the patient had any other outcome. When tFL was in the model, it was encoded as 1 if the patient was tFL and 0 otherwise. We show the different logistic regression models in Supplementary Table S3. The input signature scores for the random forest were reduced to a set of features with a Pearson correlation of <0.75. If scores in the starting set had a correlation or anticorrelation >0.75, then one of the scores was manually selected for input into the model (Supplementary Fig. S6A). The variable importance was determined and reported in Supplementary Fig. S6B using the mean decrease in the Gini coefficient where higher values are more important.

Kaplan–Meier curves were used to visualize PFS differences between groups. Hazard ratios and Cox proportional hazard statistics were calculated using either a continuous covariate or discretized categorical variable. For the Kaplan–Meier curves using the FL_like_score score and the Risueño TME score, an optimized threshold between the two groups of patients was greedily determined by finding the most significant split by Cox proportional hazard likelihood ratio test.

Three panels of five antibodies each were used to characterize the cellular content of the patient needle biopsy samples using multiplex immunofluorescence (mIF) with a PerkinElmer Vectra instrument as described previously (38, 39). Visiopharm was used to calculate the cell marker densities based on the hematoxylin and eosin–stained nuclei and marker intensities. Each panel was applied on a different slide; cell densities from each slide were analyzed independently (Supplementary Tables S4–S6).

The first panel had antibodies for the EGFR (to identify the CAR T cells that were engineered to have EGFR protein expressed), CD19, CD20, CD4, and CD8. The second panel had antibodies for CD73, FOXP3, IDO1, CD163, and PD-L1 and was originally intended to look for immune-suppressive pathways. The third panel had antibodies for EGFR, KI67, GZMB, CD3, and programmed cell death protein 1 (PD-1) and was intended to characterize the state of both the endogenous and CAR T cells.

For this study, regions of high background were manually excluded from analysis. Next, cells were segmented and phenotyped on the basis of individual algorithms to implement models that were qualitatively assessed, per fluorophore and per serial slide. These algorithms, implemented in Visiopharm, included preprocessing steps to reduce background noise and compute textural features (e.g., polynomial blob filters) to accentuate nuclei, membrane, or the entire cell where appropriate. After initial segmentation of cellular objects, a series of post-processing steps apply morphologic operations (e.g., erosion, dilation, splitting) and filtering on the basis of a combination of spatial properties (e.g., area, shape), preprocessed signal intensities, and textural information. Thus, the final cell marker positivity was not based simply on thresholding of raw signal intensity, but rather a series of filters and processing to maximize cellular signal relative to background noise. Where chromogenic immunohistochemistry was available, that was used as a gold standard to guide algorithm development for the mIF quantitation. That said, there were some samples where the signal-to-noise ratio was low for certain markers and the false negative rate high for those individual cases. Such cases were noted and if the signal-to-noise ratio was deemed too low, they were excluded from analysis.

The following combinations of markers were also used to calculate cell densities in cells/mm²: KI-67⁺CD3⁻ [non–T-cell proliferation (we could not directly measure B-cell proliferation, as KI-67 and CD19 were not on the same panel)], CD3⁺PD-1⁺ (activated or exhausted T cells), CD3⁺GZMB⁺ (cytolytic T cells), and CD163⁺PD-L1⁺ or CD163⁺IDO1⁺ (activated macrophages). Differential analysis between groups of samples was performed using these markers or marker combinations with a Wilcoxon rank-sum test.

Bristol-Myers Squibb policy on data sharing may be found at https://www.bms.com/researchers-and-partners/independent-research/data-sharing-request-process.html. Data requests can be made using the link above. RNA-seq data for 14 of the 121 patients cannot be shared due to the informed consent agreement.

To identify gene expression differences in the PRE samples between responders and nonresponders, several definitions of response were explored. To focus on the differences between patient subgroups with the largest differences in response, we performed differential testing between PRE samples of patients who achieved a CR versus those with PD at defined time point response assessments of 1, 3, 6, 9, and 12 months. We additionally grouped patients by their BOR by comparing CR versus all other responses or CR/PR versus all other responses. Across those comparison groups, the highest number of differentially expressed genes in the PRE biopsies were found between patients who were in a CR at 3 months (month-3 CR) versus those who had PD by 3 months (month-3 PD; Fig. 1A). Although the number of differential genes from the comparisons varied, we found significant correlation between them (Supplementary Fig. S3). Using a ±1.5-fold change cutoff and a 0.05 FDR cutoff, 98 genes showed higher expression in PRE samples of patients with a month-3 PD (N = 32) and 205 genes had higher expression in PRE samples of patients with a month-3 CR (N = 27; Fig. 1B).

The full ranked list of genes from the differential test between month-3 CR and PD in the PRE samples was used for GSEA, which used gene sets from the Hallmark, lymphoma publications, and solid tumor signatures (Fig. 1C; Supplementary Table S7). The top normalized enrichment score for month-3 PD was from a gene set with higher expression in biopsies from patients with DLBCL versus patients with FL (DLBCL_vs_FL). Another gene set, Knutson_EZH2_inhib_dn, included genes that were downregulated by an EZH2 inhibitor applied to three DLBCL cell lines (GSE49284; ref. 24). Other gene sets included a variety of Hallmark gene sets related to various transcription regulators (e.g., MYC, E2F, MTORC1) and other proliferative and cell-cycle pathways (e.g., G2M and oxidative phosphorylation). For example, the EZH2 inhibitor and Jerby-Arnon T-cell exclusion (30) ssGSEA scores calculated for each sample were significantly higher in patients with month-3 PD (Fig. 1D).

Several gene sets in the PRE samples were associated with month-3 CR (Fig. 1C). The top gene set in patients achieving a month-3 CR was a gene set derived from an independent set of FL and DLBCL samples and defined by genes with higher expression in FL (FL_vs_DLBCL). Therefore, the patients with a month-3 CR displayed a more FL-like tumor phenotype (Fig. 2A). Figure 2B shows the PFS curves when patients were divided into two groups based on the 25% of patients with the highest FL-like scores (19 FL-like vs. 59 others; log-rank P = 0.042) or with the optimal FL-like score threshold that resulted in the smallest Cox proportional hazards P value (10 FL-like vs. 68 others; log-rank P = 0.014). A gene expression–based lasso classifier was generated to label samples as FL-like, which gave similar results (Supplementary Figs. S4 and S5). Several LBCL biopsies with a high FL-like score had tFL; however, the FL-like score was also high in several DLBCL NOS and HGL samples, and about half of the tFL samples had a low FL-like score (Fig. 2C). Further, in samples from patients with tFL, the FL-like score was significantly higher in patients achieving month-3 CR compared with those with PD (P = 0.012); whereas, in patients with DLBCL NOS or HGL, the FL-like score only trended higher in patients with month-3 CR compared with those with PD (P = 0.17; Fig. 2C). In addition, the FL-like score had a stronger association with PFS in the tFL patients than the combined group of DLBCL NOS or HGL patients (Supplementary Fig. S6C). There was a trend toward more tFL patients with a high FL-like score (44%) than DLBCL NOS/HGL patients with a high FL-like score (19%, P = 0.066; Supplementary Fig. S6C). Finally, to directly compare tFL status and the FL-like gene expression score in predicting month-3 CR as the outcome, we performed several logistic regressions and random forest models (Supplementary Fig. S6A and S6B). We found that the FL-like score was more significant and had more importance in predicting month-3 CR than the tFL status (Supplementary Table S3).

The immune and stromal gene sets also showed higher expression in PRE samples from the month-3 CR patient group. The gene sets representing the Lenz stromal-1 and -2 signatures (1) and the Immunosign 21 signature (11, 12) were enriched in patients with month-3 CR (Fig. 1A and C). Some of the strongest individual genes with higher expression in PRE samples of patients in month-3 CR include KLRB1, CD40LG, ICOS, CD28, and CCL21. Consistent with the PD sets, one of the CR-associated gene sets was upregulated by EZH2 inhibition in DLBCL cell lines (24). Taken together, these results indicate the PRE tumor biopsies of patients with a month-3 CR following liso-cel therapy have a higher expression of immune and stromal-related genes than those with PD.

We constructed a co-expression network of T-cell–related genes to explore T-cell biology in the PRE samples (Fig. 3). For each gene in the network, we compared the association with CR to the CD4⁺/CD8⁺ expression ratios of the purified T-cell data gene set (36). Half of the network was more CD4⁺-biased and had a stronger association with month-3 CR (P = 0.049). However, these analyses show that expression of CD4⁺- and CD8⁺-biased genes in the PRE biopsies were both associated with month-3 CR.

Gene expression analysis of D11 post-infusion samples showed that expression of T-cell– and macrophage-related genes were higher in patients achieving month-3 CR compared with patients with month-3 PD. CD3D and the liso-cel CAR construct (Fig. 4A and B) plus macrophage-associated genes (e.g., CD14) were also generally higher in D11 samples of patients with month-3 CR and increased more from PRE to D11 in patients with month-3 CR (Fig. 4C). Other T-cell– and myeloid-expressed genes show a similar pattern to CD3D and CD14 expression, respectively. B-cell genes (e.g., CD19) displayed a larger decrease at D11 in patients achieving month-3 CR (Fig. 4D).

Upon further exploration of the gene sets associated with month-3 CR, we found that pretreatment samples with high expression of cell cycle–related genes or genes positively regulated by EZH2 tended to have lower infiltrating T-cell levels. To inform this relationship with T cells, we found that the CD3E gene was anticorrelated to the EZH2 gene (r = –0.32), a panel of 5 proliferation-associated genes (r = −0.36), and the EZH2i ssGSEA score (r = −0.43; Supplementary Fig. S7A). Similarly, the EZH2i ssGSEA score had significant association with month-3 PD (P = 0.014), as did the EZH2 gene (P = 0.021), and the 5 proliferation genes (P = 0.032; Supplementary Fig. S7B). In addition, we verified the relationship between EZH2 and T-cell genes by replicating it in two additional public data sets with DLBCL and FL samples (Spearman ρ of −0.62 and −0.49 for ICGC and CGCI, respectively; Supplementary Fig. S7C).

To identify genes in the PRE samples that were associated with D11 T-cell or liso-cel CAR expression, we investigated correlations between PRE and D11 for 30 patient-matched samples. We first established a high correlation (Spearman ρ = 0.59) between the T-cell gene score and the liso-cel CAR in the D11 samples (Supplementary Fig. S8A). Next, using paired samples from patients with both a PRE and D11 sample, we calculated that CD3E in the PRE samples has a correlation (0.52) with the T-cell genes in the D11 samples, but a lower correlation (0.32) with the liso-cel CAR gene. Conversely, PRE levels of EZH2 had a stronger anticorrelation with D11 liso-cel CAR (−0.22) than with D11 T cells (−0.04), and PRE levels of CD14 had a positive correlation with D11 T cells (0.39), which was not as strong as the T-cell correlation (Supplementary Fig. S8B).

To characterize the PRE genes most associated with D11 T-cell and liso-cel infiltration, we sorted the PRE genes by average correlation with D11 T-cell and liso-cel CAR genes from the 30 patient-matched samples and then performed a GSEA (Supplementary Fig. S8C). Overall, the gene set results are similar to those found when comparing month-3 CR versus month-3 PD PRE samples. For example, the DLBCL_vs_FL and Knutson_EZH2_inhib_dn gene sets were both enriched in the negatively correlated direction, consistent with those same gene sets enriched in the month-3 PD samples. However, the gene sets with the highest positive enrichment were the Lenz stromal-1 set, two EZH2i upregulated sets, the Hallmark IFNγ response set, and the FL-like set. The top gene sets were similar to the month-3 CR gene sets, although the Hallmark IFN sets here were not significantly enriched in the month-3 CR analysis.

Continuous scores or categorical classifiers were calculated for other DLBCL- and lymphoma-related molecular features and are summarized in Supplementary Fig. S6D and S6E. The Reddy COO score (34) and intrinsic apoptosis (40) scores were not associated with month-3 CR or PFS (Supplementary Fig. S6F). Notably, although only 11% of patients in this cohort were classified as HGL based on the presence of gene rearrangements in MYC and either BCL2, BCL6, or both, 40% of patients had positive double-hit gene expression (ref. 28; Supplementary Fig. S6G), and those with positive double-hit gene expression had significantly shorter PFS (log-rank P = 0.0044; Supplementary Fig. S6H). The recently described DLBCL TME classifier (33) did not show a different PFS when using the published threshold (Supplementary Fig. S6I), but if a new threshold was used that optimally split the patients into TME+ and TME−, then the patients with a TME+ phenotype had a significantly longer PFS (log-rank P = 0.011), and the continuous probability score suggested that patients with a low TME score were enriched for nonresponders (Supplementary Fig. S6J and S6K). Although trending, neither the Immunosign 21 signature nor the death receptor signature (40) were significant in the month-3 CR versus PD comparison despite previous reports (11, 12).

The PRE (N = 58) and D11 (N = 53) tumor biopsies were examined by mIF to further examine the immune cell subsets (38, 39). On average, for the samples with matched RNA-seq and mIF data, the correlation between gene expression and cell density was moderate to high (Supplementary Fig. S9A), as measured by Spearman ρ. The T-cell genes (CD3D, PDCD1, and CD8A) were highly correlated (0.62–0.73) with their respective T-cell marker densities. Macrophage-related markers (CD163, PD-L1, and IDO1) had a similar range of correlations (0.66–0.79). However, B-cell markers CD19 and CD20 had lower correlations: 0.39 and 0.58, respectively.

Examination of the measured cell densities in the PRE samples demonstrated that the T-cell association trends with CR were highest for PD-1⁺ T cells using the month-1 assessment, while PD-1⁺ T cells with later time point assessments and general CD3⁺ cells (i.e., irrespective of the PD-1⁺ marker) did not trend with CR (Fig. 5A; Supplementary Fig. S9B). Conversely, a subset of CD163⁺ macrophages, those that were IDO1⁺ or PD-L1⁺, and proliferating tumor cells (measured as KI67⁺ and CD3^– cells) appeared to be higher in the PRE samples of patients with PD (Fig. 5A; Supplementary Fig. S9C and S9D).

In D11 biopsies, the EGFR marker was used to distinguish CAR T cells from endogenous T cells, with a correlation of 0.691 (Spearman ρ) between the liso-cel CAR gene expression in the RNA-seq data and the EGFR⁺ cell density in the mIF data (Supplementary Fig. S9A). Both endogenous (EGFR⁻) and CAR (EGFR⁺) CD4⁺ and CD8⁺ cells were higher in D11 samples of patients achieving month-3 CR versus patients with PD (Fig. 5B). Furthermore, the CAR T cells comprised on average 5% of the T-cell density in the D11 samples (Fig. 5B).

Although previous studies have described the influence of tumor and TME biology on CAR T-cell therapy outcomes, this is the largest study using both RNA-seq and mIF to examine the role of the pretreatment and posttreatment TME in patients with LBCL with long-term follow-up. On the basis of patient and disease characteristics of the PRE and D11 patient groups, the groups are generally similar to one another. The patient and disease characteristics of both groups are also similar to those of the overall study population as previously reported (15), except for numerically higher median C-reactive protein and lactate dehydrogenase in the PRE group and numerically higher proportion of the sum of the product of perpendicular diameters (SPD) ≥50 cm² in both the PRE and D11 groups (Table 1). The C-reactive protein is considered a measure of inflammatory state and lactate dehydrogenase and SPD are considered markers of tumor burden. The CR rate in patients with PRE and D11 RNA-seq samples was 47% and 59%, respectively, and, therefore, within the 95% CI of the CR rate reported for the overall population [CR rate of 53% (95% CI, 46.8–59.4; ref. 15]. Because the D11 biopsy requires accessible tumor, there is a concern that the patients with a D11 biopsy could have higher tumor burden compared with the PRE group. However, we found this not to be the case, as the D11 patient group had lower tumor burden (numerically lower baseline SPD and lactate dehydrogenase) and better outcomes (higher CR rate) than the PRE group. In addition, the D11 patient group did not exhibit a different PFS from the full cohort (log-rank P = 0.84).

Some lymphoma subtypes (e.g., classical Hodgkin lymphoma and T-cell histiocyte-rich LBCL) have a paucity of lymphoma cells amidst a dominant polyclonal microenvironment and appear to have a better response to immune checkpoint blockade than DLBCL (5). In a study with 25 patients, Rossi and colleagues (12) observed elevated Immunosign 21 gene signature and Immunoscore (measure of intratumoral densities of CD3⁺ and CD8⁺ T cells; ref. 41) were associated with a CR to axicabtagene ciloleucel therapy. In a study of 10 patients that employed a different CD19 CAR T-cell therapy (JWCAR029; ref. 13), increased expression of immune-related chemokine receptors, adhesion molecules, and activated T cells were observed in patients with a CR versus PR. Our RNA-seq and mIF data indicate pretreatment levels of T cells are associated with better response, which is consistent with the results of prior studies.

The RNA-seq and mIF data allowed us to characterize the subtype and functional state of the TME immune cells. We found that PD-1⁺ T cells in the PRE samples are trending as more associated with CR than total T cells (Supplementary Fig. S9D). However, no significant correlation was observed between BOR to tisagenlecleucel and PD-1⁺ T cells (42). The T-cell correlation network analysis of the PRE samples shows that both CD4⁺ and CD8⁺ genes are associated with month-3 CR (Fig. 3), although CD4⁺ genes may be slightly more associated. In summary, the pretreatment state of T cells demonstrated that CD4⁺, CD8⁺, and PD-1⁺ T cells may be important elements of the TME to support a positive response to treatment. In the PRE samples, the greater number of differential genes between CR and PD was found when using the month-3 assessment, which could imply that the tumor biology plays a larger role in the midterm response, whereas initial response (∼1 month) and longer-term response (>12 months) may be more related to tumor burden, product composition, or longer-term CAR dysfunction or persistence.

D11 post-infusion samples showed increased immune infiltration (T cells and macrophages) in patients who developed a CR versus PD. In addition, in D11 biopsies, although infiltrating CAR T cells were associated with a CR, only approximately 5% of the T cells within the post-infusion biopsies were CAR T cells. This is similar to the percentage of CAR T cells reported with axicabtagene ciloleucel treatment in 17 patients (43). We and others (43) hypothesize that cytokines and other factors produced by CAR T cells invading the tumor result in an inflamed TME, enabling endogenous T cells and macrophages to aid in response.

We used the 30 patients with paired PRE and D11 samples to understand how pretreatment tumor biology may lead to D11 T-cell infiltration. As expected, there was a significant correlation between pretreatment levels of T-cell genes and post-infusion levels of CAR T cells and endogenous T-cell gene expression. However, the gene set in the PRE samples that was most correlated to the D11 T cells was the Lenz stromal-1 signature (Supplementary Fig. S8C). The stromal-1 signature is prognostically favorable in rituximab, cyclophosphamide, doxorubicin hydrochloride, vincristine sulfate, and prednisone treatment (1) and was also significantly enriched in this study in the PRE samples of patients achieving a month-3 CR (Fig. 1C). One aspect of stromal biology is exemplified by the higher expression of chemokines such as CCL21, a T-cell chemoattractant expressed by stromal cells that is related to T-cell trafficking in tumor-draining lymph nodes (44) but is also associated with migration of malignant lymphocytes (45, 46). In summary, a higher pretreatment level of T cells and stromal cells is correlated with increased immunity after liso-cel infusion and higher CR rates.

Liso-cel–treated patients with LBCL who had month-3 PD tended to have noninflamed tumors with higher levels of tumor cell proliferation. Using gene set enrichment, we found that EZH2 targets and transcription regulators (e.g., MYC, E2F, MTORC1, or other proliferative and cell-cycle pathways) were observed in patients with PD. We demonstrated that the proliferation and EZH2 target score in PRE biopsies were inversely correlated with T-cell gene expression in pretreatment and posttreatment biopsies in this study and other LBCL cohorts. Recent studies of lymphoma biopsies have shown DLBCL tumors have higher expression of proliferation genes than FL tumors (47). In addition, a double-hit gene expression signature (28) identified 40% of patients with a positive signature in this study, rather than 13% HGL as seen in the full study (15), and patients with this signature had shorter PFS. This is in contrast with the observation that the clinically determined HGL group had no significant difference in outcome in the full study. We also found that the melanoma tumor intrinsic gene set (30) was higher in DLBCL tumors where there were fewer T cells and where the patients had a worse month-3 response. In addition, the mIF data confirmed the RNA observation that PRE samples with higher tumor cell proliferation (i.e., non–T-cell KI-67⁺) at pretreatment had worse response rates.

The DLBCL TME gene expression classifier was developed and shown to enrich for responders to the cereblon E3 ligase modulator (CELMoD) agent avadomide (CC-122; ref. 33). DLBCL TME classifier–positive samples were previously shown to have more stromal, macrophage, and/or T-cell content than classifier-negative samples, which had more B-cell and proliferative gene expression. We found that patients with very low TME classifier probability scores were more likely to progress, consistent with both the higher proliferation and reduced levels of stromal cells observed in the TME classifier–negative patients. In addition, despite the impact of COO on other LBCL therapeutic outcomes, there is not an association of COO with outcome in this cohort.

There are several drugs which could potentially change the tumor into a more favorable state for liso-cel or CAR T-cell infiltration and response. In lymphoma, EZH2 activity-enhancing mutations leads to increased cell proliferation and reduced antigen presentation (5). CDK4/6 inhibitors have been shown in both posttreatment clinical samples and mouse models to increase tumor immunity (48). Immunomodulatory drugs (IMiD) and CELMoD agents have been shown to increase IFN pathway genes in lymphoma cell lines (49) and in tumors after patients with DLBCL receive avadomide treatment (50). These observations suggest future experiments with EZH2 or CDK4/6 inhibitors, IMiDs, or CELMoDs to explore the interplay between tumor biology and the TME.

Myeloid cells can have suppressive functions against CAR T cells (10). Yan and colleagues showed that 3 patients with higher levels of macrophage genes had PRs compared with 5 patients with CRs who had lower levels of macrophages (13). However, in Rossi and colleagues, patients with higher pretreatment levels of macrophage genes also had higher levels of T-cell and immune genes in general and were more likely to respond to treatment (12). In our study, we found the pretreatment gene expression patterns for myeloid markers were not associated with CR or PD. However, results obtained through mIF indicated that higher PD-L1⁺ or IDO1⁺ macrophages may trend towards being higher in PRE samples of patients with PD (Fig. 5A; Supplementary Fig. S9). This data, along with the observation that macrophages were more prevalent in D11 samples of patients with month-3 CR versus PD, suggest macrophages have both pro- and antitumor activity depending on their activation state. Thus, depletion of myeloid cells may be detrimental, and it may be more productive to alter the functional state of macrophages in LBCL prior to therapy to improve response to CAR T cells.

The observation that patients with a high FL-like gene expression pattern had a better response to liso-cel provides additional clues for understanding at least one of the mechanisms of response. To further understand FL-like gene expression, a more detailed look at FL biology is warranted. FL is characterized by tumor-infiltrating lymphocytes, including regulatory T cells and tumor-associated macrophages, and the tumor cells depend heavily on the TME for proliferation and survival (4). We identified a subset of patients with LBCL who exhibited an elevated FL-like gene expression signature that was associated with favorable outcomes in response to liso-cel. Although a high FL-like score was trending to be more common in patients with tFL, and the latter had a better outcome in the full study (15), the FL-like score was more important than tFL status in several statistical models of predicting CR at month 3. This suggests that FL-like gene expression is not exclusive to tFL and that it could play an important role in understanding response. The current data set suggests that an “FL-like” TME in LBCL is associated with favorable response to liso-cel therapy and provides the basis for further investigation of the association of pretreatment gene expression and response to liso-cel.

Although the pretreatment gene expression data highlight some significant correlates for response, gaps remain in our full understanding of the role of the TME. Further analysis of the tumor biopsy signatures integrated with other patient factors, drug product features, or persistence of functional CAR T cells post-infusion may help inform a more complete view of the mechanisms of response and relapse. In addition, lymphoma tumor genetics, such as PD-L1 structural variants or other tumor genetic alterations (MYC, TP53, NFkB, NOTCH2, BCL2/6/10, FAS), have been shown to help clarify the mechanisms of tumor immune escape (5), and tumor genetic data is not available for these patients.

This analysis identified novel gene expression signatures associated with response to liso-cel and provided insights into how tumor biology and the TME may affect response to liso-cel and CAR T-cell therapies in LBCL in general. These data will generate hypotheses to explore additional combination therapy strategies and next-generation CAR T-cell therapies.

N. E. Olson, B. A. Fox, and K. J. Newhall report a patent for WO2021035194A1 pending to Bristol Myers Squibb. D.J. Reiss reports personal fees and nonfinancial support from Bristol Myers Squibb during the conduct of the study; personal fees from Bristol Myers Squibb outside the submitted work. Y. Kim reports other support from Bristol Myers Squibb during the conduct of the study. J.S. Abramson reports grants and personal fees from Bristol Myers Squibb during the conduct of the study; personal fees from Caribou Biosciences, Incyte, Takeda, Janssen, Genentech, Eli Lilly, Kite Pharma, AbbVie, AstraZeneca, Epizyme, Genmab, Century Therapeutics, Regeneron, MorphoSys, BeiGene, Ono Pharma, Kymera, Bluebird Bio, C4 Therapeutics, Allogene, Karyopharm; grants and personal fees from Mustang Bio; and personal fees from Novartis outside the submitted work. C. McCoy reports other support from Bristol Myers Squibb during the conduct of the study; other support from Amgen; and other support from Bristol Myers Squibb outside the submitted work. K.J. Newhall reports other support from Bristol Myers Squibb during the conduct of the study; other support from Cellectis, Inc. outside the submitted work. B.A. Fox reports other support from Bristol Myers Squibb during the conduct of the study. No disclosures were reported by the other authors.

N.E. Olson: Conceptualization, formal analysis, supervision, validation, writing–original draft, project administration, writing–review and editing. S.P. Ragan: Resources, data curation, formal analysis, validation, investigation, writing–original draft, writing–review and editing. D.J. Reiss: Conceptualization, formal analysis, writing–review and editing. J. Thorpe: Resources, methodology, project administration, writing–review and editing. Y. Kim: Formal analysis, methodology. J.S. Abramson: Conceptualization, data curation, writing–original draft, writing–review and editing. C. McCoy: Methodology, writing–review and editing. K.J. Newhall: Conceptualization, resources, supervision, writing–original draft, project administration, writing–review and editing. B.A. Fox: Conceptualization, formal analysis, supervision, validation, visualization, writing–original draft, project administration, writing–review and editing.

This study was funded by Bristol-Myers Squibb.

The authors thank the patients for participating in the study. We thank David Kuo and Matt Stokes for contributions to the gene expression analyses; Trevor Do, Suzana Cuoto, Chung-Wein Lee, Falon Gray, and Shradha Srinivasan for contributions to the mIF experiments; Gui Gao for contributions to clinical data analysis; and Jason Dubovsky for scientific input. All authors contributed to and approved the manuscript. Writing and editorial assistance were provided by Jeremy Henriques, PhD, and Lauren Connor, BS, ELS, of The Lockwood Group (Stamford, CT, USA), funded by Bristol Myers Squibb.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.