Meta-analysis of tumor- and T cell-intrinsic mechanisms of sensitization to checkpoint inhibition

Abstract

Checkpoint inhibitors (CPIs) augment adaptive immunity. Systematic pan-tumor analyses may reveal the relative importance of tumor-cell-intrinsic and microenvironmental features underpinning CPI sensitization. Here, we collated whole-exome and transcriptomic data for >1,000 CPI-treated patients across seven tumor types, utilizing standardized bioinformatics workflows and clinical outcome criteria to validate multivariable predictors of CPI sensitization. Clonal tumor mutation burden (TMB) was the strongest predictor of CPI response, followed by total TMB and CXCL9 expression. Subclonal TMB, somatic copy alteration burden, and histocompatibility leukocyte antigen (HLA) evolutionary divergence failed to attain pan-cancer significance. Dinucleotide variants were identified as a source of immunogenic epitopes associated with radical amino acid substitutions and enhanced peptide hydrophobicity/immunogenicity. Copy-number analysis revealed two additional determinants of CPI outcome supported by prior functional evidence: 9q34 (TRAF2) loss associated with response and CCND1 amplification associated with resistance. Finally, single-cell RNA sequencing (RNA-seq) of clonal neoantigen-reactive CD8 tumor-infiltrating lymphocytes (TILs), combined with bulk RNA-seq analysis of CPI-responding tumors, identified CCR5 and CXCL13 as T-cell-intrinsic markers of CPI sensitivity.

Keywords: CXCL9; biomarkers; checkpoint inhibitors; clonal TMB; immunogenicity; immunotherapy; meta-analysis; mutation; neoantigen.

Conflict of interest statement

Declaration of interests K.L. has a patent on indel burden and CPI response pending and outside of the submitted work, speaker fees from Roche tissue diagnostics, research funding from CRUK TDL/Ono/LifeArc alliance, and a consulting role with Monopteros Therapeutics. S.T. has received speaking fees from Roche, AstraZeneca, Novartis, and Ipsen. S.T. has the following patents filed: Indel mutations as a therapeutic target and predictive biomarker PCTGB2018/051892 and PCTGB2018/051893 and Clear Cell Renal Cell Carcinoma Biomarkers P113326GB. S.Q. reports personal fees and employment with Achilles Therapeutics (where he is CSO) outside of the submitted work. J.L.R. consults for Achilles Therapeutics. N.M. has received consultancy fees and has stock options in Achilles Therapeutics. N.M. holds European patents relating to targeting neoantigens (PCT/EP2016/ 059401), identifying patient response to immune checkpoint blockade (PCT/ EP2016/071471), determining HLA LOH (PCT/GB2018/052004), and predicting survival rates of patients with cancer (PCT/GB2020/050221). C.A. receives research salary from AstraZeneca and is an AstraZeneca Fellow and acting study physician on the MERMAID-1 study. C.A. holds pending patents in methods to detect tumor recurrence (PCT/GB2017/053289). C.A. and C.S. declare patent PCT/US2017/028013 for methods to detect lung cancer. C.A. has received speaker fees from Novartis, Roche Diagnostics, Bristol Myers Squibb, and AstraZeneca and was an advisory board member for AstraZeneca. M.D.H. has stock and other ownership interests in Shattuck Labs, Immunai, and Arcus Biosciences; reports honoraria from AstraZeneca and Bristol-Myers Squibb; has a consulting or advisory role with Bristol-Myers Squibb, Merck, Genentech/Roche, AstraZeneca, Nektar, Syndax, Mirati Therapeutics, Shattuck Labs, Immunai, Blueprint Medicines, Achilles Therapeutics, and Arcus Biosciences; receives research funding from Bristol-Myers Squibb (Inst); has patents, royalties, and other intellectual property (a patent has been filed by Memorial Sloan Kettering [PCT/US2015/062208] for the use of TMB for prediction of immunotherapy efficacy, which is licensed to Personal Genome Diagnostics); and receives travel and accommodation expense reimbursement from AstraZeneca, Bristol-Myers Squibb, and Eli Lilly. J.L. reports personal fees from Eisai, GlaxoSmithKline, Kymab, Roche/Genentech, Secarna, Pierre Fabre, and EUSA Pharma and grants and personal fees from Bristol-Myers Squibb, Merck Sharp & Dohme, Pfizer, and Novartis outside of the submitted work. C. Swanton acknowledges grant support from Pfizer, AstraZeneca, Bristol-Myers Squibb, Roche-Ventana, Boehringer-Ingelheim, Archer Dx (collaboration in minimal residual disease sequencing technologies), and Ono Pharmaceutical; is an AstraZeneca advisory board member and chief investigator for the MeRmaiD1 clinical trial; has consulted for Pfizer, Novartis, GlaxoSmithKline, MSD, Bristol-Myers Squibb, Celgene, Amgen, AstraZeneca, Illumina, Genentech, Roche-Ventana, GRAIL, Medicxi, Bicycle Therapeutics, and the Sarah Cannon Research Institute; has stock options in Apogen Biotechnologies, Epic Bioscience, and GRAIL; and has stock options and is co-founder of Achilles Therapeutics. C.S. holds European patents relating to assay technology to detect tumor recurrence (PCT/GB2017/053289), targeting neoantigens (PCT/EP2016/059401), identifying patient response to immune checkpoint blockade (PCT/EP2016/071471), determining HLA LOH (PCT/GB2018/052004), predicting survival rates of patients with cancer (PCT/GB2020/050221), and identifying patients who respond to cancer treatment (PCT/GB2018/051912), as well as a US patent relating to detecting tumor mutations (PCT/US2017/28013) and both a European and US patent related to identifying insertion/deletion mutation targets (PCT/GB2018/051892). S.R.H. is co-founder of Tetramershop and PokeAcell. D.B. reports personal fees from NanoString, outside this work, and he has a patent PCT/GB2020/050221 issued on methods for cancer prognostication.

Copyright © 2021 Elsevier Inc. All rights reserved.

Figures

Graphical abstract

Figure 1

Design of the meta-analysis study Input studies to the meta-analysis (Figure 2) results (top) and validation cohorts for the multivariable predictive modeling (Figure 3) (bottom).

Figure S1

Supplementary meta-analysis data, related to Figure 2 Panel A shows the correlation in biomarker effect sizes for radiological response and overall survival clinical endpoints (Spearman's correlation). Panel B shows response rate by number of NMD-escape mutations for all available samples. Panel C shows results from previously published histology specific biomarkers, or metrics that could not be calculated in > 75% of the cohort samples.

Figure 2

The biomarker landscape of CPI response (A) Previously published biomarkers are shown as rows and individual cohorts within the CPI1000+ cohort as columns. The heatmap indicates the effect size of each biomarker in each cohort, measured as the log2 odds ratio (OR) for response “CR/PR” versus no response “SD/PD/NE” derived from logistic regression. Blue denotes association with response, red association with no response. Drug class and cohort sizes are annotated, and the right-hand forest plot shows the overall effect size and significance of each biomarker in meta-analysis across all studies, based on effect sizes and standard errors from each individual cohort. p values are shown from meta-analysis (random effects, on account of the different tumor types), with the first set of p-values including all samples (p-meta all cohorts) and last set (p-meta validation cohorts) including validation cohorts only (i.e., when a biomarker was originally discovered in a cohort, this cohort was excluded from the meta-analysis). For clarity of plotting, outlier OR values were capped between OR = 0.1 and OR = 10 (all outlier values were nonsignificant results skewed by rare event counts, and raw (uncapped) values were still used in the meta-analysis). (B) The CPI1000+ cohort broken into cancer/drug subgroups for combinations with two or more independent cohorts. OR effect sizes are shown on the y axis, and biomarkers that are either significant in the pan-cancer 2A analysis or within an individual subgroup are shown. Colors are arbitrary and are used only to distinguish the groups. (C) Correlation between biomarkers that are measured on a continuous scale. (D) Proportion of variance explained for each category of biomarker, for each study, calculated using logistic regression pseudo-R2.

Figure S2

(A) shows significant histology or drug-specific biomarker interactions identified in the CPI1000+ cohort (using histology*biomarker and drug*biomarker interaction terms in logistic regression), and (B) shows dinucleotide variant associations with CPI response, related to Figures 2 and 4

Figure 3

A multivariable predictor of CPI response outperforms TMB (A) Feature importance scores from XGBoost for the multivariable model, corresponding to 1,000 Monte Carlo sampling rounds. (B) Design, samples included, and features utilized in the final model training. (C) The top five feature importance scores from the final pan-cancer model. (D) ROC curves and AUC values for the multivariable predictor benchmarked to TMB, as a univariable comparator, in the three independent test cohorts (not used in any of the model training steps). p values report the significance of improved performance for the multivariable versus TMB model using DeLong’s test.

Figure 4

Mutational processes associated with CPI response (A) Forest plot of each mutation signature and its association with CPI response, with odds ratio values shown on the y-axis, and p-values derived from meta-analysis (hence the results are not biased by mixing histology types). (B) Proportion of signature 7 (UV) mutations (left) and the number of dinucleotide variants (DNVs) per tumor (middle), split by histology type. The panel on the right shows the correlation between signature 7 proportion and DNV count. p value and correlation coefficient are from Spearman’s rank test. (C) Grid of substitutions from the reference amino acid (rows) to the mutated amino acid (columns). The heatmap is colored from low to high, based on the simple count of each observed ref > alt change in the cohort, shown on a log10 scale. The first grid (left) shows the data for SNVs, and the second grid (middle) shows data for DNVs. The first barplot (middle) then quantifies the number of unique changes observed for SNVs and DNVs, and the second barplot shows the proportion of amino acid changes resulting in a radical amino acid change (i.e., Grantham distance ≥ 100) compared to those resulting in a conservative change (Grantham distance 

Figure 5

Somatic copy-number alteration (SCNA) profile of CPI responders versus nonresponders (A) Frequency of somatic copy-number gain (top) and loss (bottom) across the genome for CPI responders (“CR/PR”) versus nonresponders (“SD/PD”). (B) Cytobands with significantly different loss or gain frequencies in responders versus nonresponders,with p-value derived from Fisher's exact test, and q values from FDR correction.. (C) Fine mapping of the 9q34 locus to identify the peak of differential loss frequency between groups. (D) TRAF2-loss percentage frequencies for cohorts with a significant difference between responders and nonresponders, with p-value derived from Fisher's exact test. (E) Probability of haploinsufficiency (pLI) scores from the gnomAD/ExAC consortium data (n = 125,748 germline human samples).

Figure S3

9q34 (TRAF2) analysis and immune evasion data, related to Figure 5 Panel A shows drug sensitivity screening data for two compounds, for TRAF2 heterozygously mutated versus TRAF2 wild-type cell lines. Panel B shows immune evasion analysis, measuring as the % of patients with an antigen presentation pathway defect between tumors with 9q34 wild-type (i.e., no loss) compared to 9q34 loss tumors. The left barplot includes either a somatic copy number loss, or a non-synonymous mutation, in an antigen presentation pathway gene. The right plot includes non-synonymous mutations only. Antigen presentation pathway genes were defined as per (Rosenthal et al., 2019), also see methods. Panel C shows the frequency of whole chromosome loss in TCGA for the set of cancer types included in the CPI1000+ study.

Figure S4

Cytobands with significantly different copy-number loss or gain frequencies in responders versus nonresponders, related to Figure 5 Analysis is split by 4 tumor/drug types.

Figure 6

Focal amplification and deletion profile of CPI responders versus nonresponders (A) CPI response rate (% “CR/PR”) in patients with focal amplification (defined as copy number ≥ 5) or homozygous deletion (copy number = 0) compared to wild-type (nonamplified/deleted) tumors. The analysis was conducted for all oncogenes/tumor suppressor genes with greater than 5% amplification/deletion frequency, and p-values were derived from Fisher's exact test. (B) Counts of CCND1 amplification by histology. (C) mRNA expression for CCND1 in responders versus nonresponders from the Mariathasan et al. urothelial cancer cohort, with p-value derived from Mann-Whitney U test. (D) Overall survival analysis in MSK-IMPACT urothelial cancer CPI-treated patients for CCND1-amplified versus wild-type tumor groups. (E) Overall survival analysis in MSK-IMPACT urothelial cancer non-CPI-treated patients for CCND1-amplified versus wild-type tumor groups. (F) Overall survival analysis in MSK-IMPACT pan-cancer CPI-treated patients for CCND1-amplified versus wild-type tumor groups. (G) Overall survival analysis in MSK-IMPACT pan-cancer non-CPI-treated patients for CCND1-amplified versus wild-type tumor groups.

Figure 7

CD8+ neoantigen-reactive single-cell RNA-seq and CPI1000+ cohort analysis (A) Single-cell RNA sequencing (RNA-seq) data from neoantigen multimer negative versus positive CD8+ TILs. The top plot shows the sorting of multimer positive versus negative T cells, and the bottom plot shows differential gene expression analysis between multimer-positive versus multimer-negative cells, with log2 fold change (FC) shown on the x axis and −log10 value on the y axis. Significant genes with > 2 FC upregulation (log2(FC) > 1) and false discovery rate (FDR) < 0.05 are shown blue. (B) The same FC upregulation values from (A) on the y axis and then overlaid on the x axis is upregulation scores from the CPI1000+ cohort (log2[FC] values for responders “PR/CR” versus nonresponders “SD/PD”). The panel only shows genes significantly upregulated in both experiments. (C) Patient-level mRNA data for the two most strongly unregulated genes (CXCL13 and CCR5) from (B) from the CPI1000+ cohort, with p-value derived from Mann-Whitney U test..

Figure S5

Clustering by common germline SNP panel to ensure no duplicate participants were recorded in the CPI1000+ cohort, related to STAR methods Columns are patients, rows are SNPs.

Figure S6

Purity, sequencing coverage, and choice of exome capture kits do not correlate with TMB scores in the CPI1000+ cohort, related to STAR methods

Figure S7

Purity vs TMB correlations by study, related to STAR methods