Genome and Transcriptome Biomarkers of Response to Immune Checkpoint Inhibitors in Advanced Solid Tumors

Immune checkpoint inhibitors (ICI) have revolutionized the treatment of solid tumors with dramatic and durable responses in cancers with previously limited effective treatment options (1, 2). Overall response rates to ICIs are often improved compared with standard-of-care cytotoxic chemotherapies but remain relatively low in unselected populations; for example, reported overall response rates are 13.3% in head and neck squamous cell carcinomas (HNSCC) treated with nivolumab (3), and 5.3% in previously treated triple-negative breast cancers (4). Response rates in selected populations are higher, with objective response rates of 40% to pembrolizumab in mismatch repair deficient (dMMR) colorectal cancers (5) and 41% in patients with non–small cell lung cancer (NSCLC) with more than 1% tumor cell programmed death-ligand 1 (PD-L1) staining (6). dMMR, PD-L1 expression, and more recently tumor mutation burden (TMB) are clinically approved biomarkers for patient selection (7, 8). However, many patients who are negative for these markers also respond to ICI therapy.

Additional features, including tumor immunogenicity and evidence of an immune response, may be predictive of ICI efficacy. TMB and prediction of neoantigens have been associated with clinical benefit from ICIs (9–11). The tumor microenvironment, particularly the presence and spatial arrangement of T cells, has also been shown to influence response, and has been explored in various tumor types including NSCLC and melanoma (10, 12). These features have yet to be surveyed in broader advanced cancer populations.

Whole genome and transcriptome analysis (WGTA) provides the opportunity to explore a wealth of genomic and expression markers that may predict response and overall survival (OS) with ICIs. The Personalized OncoGenomics Program (POG) at BC Cancer, in Vancouver, Canada, performs WGTA on patients with advanced and metastatic disease, often after one or more lines of therapy, and therefore represents an important heterogeneous patient population in which to examine predictors of ICI response (13, 14). Here we explored proposed biomarkers of ICI sensitivity in a pan-cancer cohort of patients with advanced or metastatic disease who received ICI therapy, incorporating WGTA and clinical follow-up.

This work was approved by and conducted under the University of British Columbia BC Cancer Research Ethics Board (H12-00137, H14-00681), in compliance with the Tri-Council Policy Statement and the FDA regulations (Belmont) and the Good Clinical Practice principles (Helsinki), and approved by the institutional review board. The POG is registered under clinical trial number NCT02155621. Patients with advanced or metastatic disease gave informed written consent and were enrolled into POG as described previously (13, 14). Ninety-eight patients within the program, biopsied between April 1, 2014 and August 31, 2018, had received ICIs. A REMARK (REporting recommendations for tumour MARKer prognostic studies; ref. 15) checklist has been included for this study.

Full treatment histories, response, and survival data for this cohort were collected retrospectively using the BC Cancer Pharmacy database and chart review. Follow-up was censored on March 1, 2019. Patients who received ICIs up to 14 days prior to biopsy were considered ICI treated at time of biopsy. Patients were considered to have a durable clinical benefit (DCB) if they remained on treatment for greater than 6 months without disease progression [stable disease (SD), partial (PR) or complete response (CR)]. Time to progression (TTP) was defined as the time from ICI initiation to the date of discontinuation due to progression. OS was defined as the time from ICI initiation to death.

Tumor specimens were collected using needle core biopsies, endobronchial ultrasound biopsies, or tissue resection. Solid tumor specimens were snap frozen, while liquid biopsies were spun down into a cell pellet and resuspended, followed by extraction of DNA and RNA as described in Pleasance and colleagues (14). Constitutional DNA representing normal cells was extracted from peripheral blood.

PCR-free genomic DNA libraries and poly-A selected RNA libraries were constructed using standard protocols as described previously (16, 17). Tumor genomes were sequenced to a target depth of 80× coverage and normal peripheral blood samples to 40× coverage with 125 or 150 bp reads on the Illumina HiSeq platform. Transcriptomes were sequenced targeting 150–200 million 75-bp end reads on Illumina HiSeq2500 or NextSeq500. Sequencing characteristics for each sample are reported in Supplementary Table S1.

Sequences were aligned to human genome version hg19, and single-nucleotide variants (SNV) and insertions and deletions (indels) were called as described in Jones and colleagues (17) and Grewal and colleagues (16). Structural variants (SV) were identified followed by merging and annotating using MAVIS (v2.1.1; ref. 18), and filtered for high-quality events as described in Pleasance and colleagues (14). Variants were annotated to genes using SNPEff (v3.2) with the Ensembl database (v69; ref. 19).

Copy-number alterations were identified from read depth ratios in tumor and normal whole-genome sequencing (WGS) data using CNAseq (v0.0.6; ref. 20), and regions of loss of heterozygosity determined from allelic ratios using APOLLOH (v0.1.1; ref. 21). Estimated tumor content for each sample was obtained on the basis of the best fit of predicted allelic ratios and read depth ratios to observed data, followed by manual review, also taking into consideration the estimated tumor content from pathology review of hematoxylin and eosin–stained tumor sections.

Exomic mutation burden was computed as the total of all SNVs and indels that fall within protein coding sequence (specifically, annotated by SNPEff as HIGH, MODERATE, or LOW impact and with an entry in the “aachange” field), divided by the total coding exome size of 35,600,628 bases included in the gene annotations, based on one transcript per gene for Ensembl 69 gene models. Microsatellite instability–high (MSI-H) samples were identified as such using MSIsensor (v0.2; ref. 22).

For analysis of resistance alterations, frameshift and nonsense mutations were defined as truncating alterations, and missense variants as substitutions. Mutations used for this analysis were manually reviewed for validity in the Integrated Genomics Viewer (23). Biallelic loss indicated deletions predicted to affect all copies of the gene. Differences in alteration frequencies between responders (DCB) and nonresponders [no clinical benefit (NCB)] were calculated using a χ² test.

Reads were aligned and processed using Jaguar (v2.0.3; ref. 24) as described in Jones and colleagues (17). Gene level RPKM (reads per kilobase per million mapped reads) were calculated on the basis of Ensembl gene models (v69).

Immune cell type deconvolution was performed with the CIBERSORT R package (v1.04; ref. 25), using the LM22 cell subtype signature in absolute mode, with 1,000 permutations and no quantile normalization (Supplementary Table S3). Biopsies from hematopoietic malignancies, lymphomas, and thymic cancers, and organs of the hematopoietic system such as lymph nodes, were omitted from the immune CIBERSORT analysis due to the confounding presence of many immune cells unrelated to an antitumor immune response.

Using the LM22 cell weight matrix from CIBERSORT, we identified the genes that most effectively discriminate between M1 and M2 macrophages. Specifically, we calculated the fold change between the weight value for M1 and for M2 for each gene in the matrix, and the fold change for M1/M2 and the other 21 cell types. We selected the 10 genes with the highest weight for M1 and M2 (and not the other cell types) that also have the largest discrimination (fold change) between M1 and M2. The M1-M2 macrophage score for each sample was derived by calculating the mean expression (RPKM) of these 10 genes (CXCL11, IDO1, CCL19, CXCL9, PLA1A, LAMP3, CCR7, APOL6, CXCL10, TNIP3). Biopsies from hematopoietic malignancies and the hematopoietic system, such as lymph nodes, were also excluded from the M1-M2 macrophage analysis.

PD-L1 clinical staining tests were performed with the PD-L1 IHC 22C3 pharmDx test using the Dako autostainer.

Kaplan–Meier survival analysis was performed for TTP and OS using the R packages survival (v2.42.3) and survminer (v0.4.2). Differences in nonparametric survival functions were assessed across groups using log-rank tests. Cox proportional hazards models were performed using the R packages survival (v2.42.3) and forest model (v0.5.0). Log-rank tests were used to calculate P values, and in the case of ties, the Efron approximation was used. Linear regression models were used to account for histology in the DCB versus NCB analyses for mutation burden.

In the case of multivariate analyses using histology, only tumor types with more than one sample were included.

Thresholds for CD274 expression (80th percentile; 4.77 RPKM), CD8⁺ T cell (median; 0.054), and M1-M2 macrophage score (median; 8.51 RPKM) “high” groups were determined through examination of the naïve cohort distributions (Supplementary Fig. S1A) for each marker, and ensuring that sample sizes in each group were appropriate for combining markers later on. Ten mutations/Mb was used as a threshold for TMB as this has been previously reported as an appropriate threshold for hypermutation (26). High structural variant burden was defined as ≥ 75th percentile (187.25) for the cohort.

Data for Snyder and colleagues (27) was downloaded from Supplementary Data and https://github.com/hammerlab/multi-omic-urothelial-anti-pdl1 on April 28, 2020. High mutation burden was defined as ≥ 10/Mb. Mutation burden in this cohort was excluded from the multivariate Cox proportional hazards models as only one patient was above this threshold. Thresholds for high score were defined as: CD8 score ≥ 0.054 as the median was 0 (0.054 was the same value as the median in our cohort), M1-M2 score ≥ median (228.6), and CD274 expression ≥ 80th percentile (202.6). DCB was defined as a positive response (PR, CR, SD) without progression for at least 6 months. Data for Riaz and colleagues (28) were downloaded from the Supplementary Data and https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE91061 on April 28, 2020. High mutation burden was defined as ≥ 200 coding mutations, which corresponds to approximately 10/Mb. Clinical benefit was defined as a positive response (PR, CR, SD) only, as no progression data was available. Thresholds for high score were defined as: CD8 score ≥ median (0.075), M1-M2 score ≥ median (17.05), and CD274 expression ≥ 80th percentile (6.73). Patients with nonevaluable responses were not included in clinical benefit analyses. Raw data for Liu and colleagues (29) were downloaded from https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000452.v3.p1#restricted-access-section on May 25, 2020. SNV calling and gene expression profiling were performed as described above for our cohort. High mutation burden was defined as ≥ 10/Mb. Thresholds for high score were defined as: CD8 score ≥ 0.054, M1-M2 score ≥ median (8.17), and CD274 expression ≥ 80th percentile (1.51). DCB was defined as a positive response (PR, CR, SD) without progression for at least 6 months. Of these datasets, patients were only included if they were naïve for immunotherapy prior to the study and both transcriptome and mutation data were available, consistent with our study. CIBERSORT and M1-M2 scores were evaluated as described above on available expression data. Where relevant, lymph associated samples were excluded.

Cytolytic (CYT) scores were calculated using expression transcripts per million (TPM) as described in Rooney and colleagues, (30). T-cell exclusion signature (TIDE) scores and IFNγ signatures were obtained using normalized expression RPKMs with the TIDE web application http://tide.dfci.harvard.edu/ (31–33). TPMs were calculated as described in Pleasance and colleagues (14).

Sequence data, both WGS and RNAseq, have been deposited at the European Genome–phenome Archive (EGA, http://www.ebi.ac.uk/ega/) as part of the study EGAS00001001159 with accession numbers as listed in Supplementary Table S1. Clinical data, marker status, sequencing statistics, and CIBERSORT immune scores are presented in Supplementary Tables S1–S3. Complete somatic mutation variant call format (VCF) files, RPKM expression matrices, and tables of structural variants and coding mutations described in this study are available from https://www.bcgsc.ca/downloads/immunoPOG/. Data on coding mutations, copy changes, and expression from tumor samples in the POG program are also accessible from https://www.personalizedoncogenomics.org/cbioportal/. All other data supporting the findings of this study are available from the corresponding author on reasonable request.

We characterized tumor samples from 98 patients (Table 1; Supplementary Table S1), representing a diversity of histologies, biopsy sites, and treatment histories, who had received ICIs either following (naïve cohort, n = 82) or prior to biopsy (treated cohort, n = 16). Twenty different tumor types were profiled, with the majority of cases being lung (n = 26, 26%) and breast cancer (n = 13, 13%), or cutaneous melanoma (n = 11, 11%). The median age of advanced disease diagnosis was 57 (25–86 years) and 55% of patients were female. Biopsies were taken at a median of 10 months following advanced diagnosis (0 days–15 years), and the median follow-up time postbiopsy was 15 months (0–54 months). Patients in the naïve cohort received an ICI after the POG biopsy, typically as a third-line therapy (median, range: 1–10), and those in the treated cohort had received an ICI typically as a first-line therapy (median, range: 1–4). The naïve cohort was the primary focus of this study for examining genomic biomarkers predictive of response.

Tumors with a high TMB and dMMR have been associated with a good response to ICIs in previously untreated tumors of a single type (5, 34). We explored whether TMB was predictive in our naïve cohort comprised of mixed tumor types and prior exposure to other cancer therapies. We calculated TMB as mutations per Mb both within coding regions (exomic TMB; Fig. 1A) and across the entire genome (genomic TMB). Measuring TMB in coding regions could reflect mutations with direct potential to elicit an immune response, and is more consistent with previous studies that use whole-exome sequencing or panels such as FoundationOne CDx (35, 36). However, we observed that the two approaches produced strongly correlated TMB values (Fig. 1B; r = 0.95, P < 2.2 × 10⁻¹⁶, Spearman). Exomic TMB ranged from 0.70 to 85 mutations/Mb (median: 4.6) in the naïve cohort, with cutaneous melanomas and lung cancers having some of the highest TMBs (Fig. 1A). Only three samples in this cohort were dMMR, one colorectal and two lung cancers, all of which had TMB > 30 mutations/Mb.

Patients that demonstrated a DCB (see Materials and Methods) did not have a significantly higher TMB overall than those with NCB (Fig. 1C; median 12.7 vs. 4.3, P = 0.16 and Fig. 1E; median 18.4 vs. 5.8, P = 0.053), though there was a significant difference after correction for histology [P = 0.035 (exome) and P = 0.029 (genome)]. A high TMB (≥ 10 mutations/Mb; Materials and Methods; Supplementary Fig. S1A) was strongly predictive of longer TTP for both exome (Fig. 1D; median 179 days vs. 79, P = 0.007) and genome-derived counts (Fig. 1F; median 156 days vs. 79.5, P = 0.0027). As TMB can vary substantially by tumor type, we additionally noted that this observation remained true after accounting for histology [exome; HR = 0.36 (0.16–0.83), P = 0.02, genome; HR = 0.34 (0.15–0.74), P = 0.007]. TMB was less predictive of OS, and using the same TMB threshold, exome-derived counts (Fig. 1D; median 442 days vs. 240, P = 0.039) were more effective than genomic (Fig. 1F; median 396 vs. 265, P = 0.17). This also remained constant after correction for tumor type [exome; HR = 0.41 (0.18–0.92), P = 0.03, genome; HR = 0.54 (0.26–1.12), P = 0.1]. Overall, there was a stronger stratification of patients based on TMB for TTP than OS, which is likely influenced by higher TMBs being a poor prognostic marker (37, 38).

The burden of SVs in the genome was also evaluated as a potential prognostic marker for ICIs, but we could not identify a consistent trend with response (Supplementary Figs. S1B and S1C). Despite some difference in OS between patients with high (≥75th percentile) and low SV burden, we did not see the same trend with TTP on ICI therapy, and we did not observe a difference in SV burden between patients with and without clinical benefit. This may indicate some OS benefit associated with higher structural variant burden that is independent of treatment with ICIs.

The composition of the tumor microenvironment, in particular, the presence and spatial arrangement of immune cells, has been reported to influence patient response to ICIs (12). As these investigations have generally been conducted using IHC, we sought to investigate the predictive power of immune cell gene expression profiles derived from WGTA (Materials and Methods). We observed a trend of higher T-cell expression scores in samples biopsied from lymph nodes and in lymphoid cancers (Supplementary Fig. S1D, P = 0.063, Wilcoxon rank sum). As presence of immune cells in these tissues was expected, and may not specifically indicate detection of an antitumor immune response, we excluded these (n = 17 naïve samples of mixed histologies; Supplementary Table S1) from immune cell expression analyses. Immune cell expression profiles varied by tumor type (Fig. 2A), with highest variability found in CD8⁺ T cell, M0 macrophage, and neutrophil scores [standard deviation (StdDev) 0.19, 0.16, and 0.16, respectively]. Cholangiocarcinomas displayed the highest overall levels of immune cell expression, which may be in part a result of the lower tumor content of these sample biopsies (Supplementary Fig. S1E). Cholangiocarcinoma, skin cutaneous melanomas (SKCM), and HNSCCs displayed the highest levels of expression for CD8⁺ T cells.

We also observed that the median CD8⁺ T-cell score in responders (DCB) was over double that of nonresponders, though the overall difference was not significant (Fig. 2B; median 0.13 vs. 0.05, P = 0.11). Higher CD8⁺ T-cell scores (≥ median; Supplementary Fig. S1A) were predictive of a longer TTP (Fig. 2C; 102 days vs. 70, P = 0.0059) and OS (Fig. 2C; 370 days vs. 160, P = 0.0014). In contrast to TMB, high CD8⁺ scores had a more profound influence on OS than TTP, highlighting the underlying positive prognostic value of CD8⁺ T cells (39). Evaluating expression signatures from other immune cells revealed that only CD8⁺ T-cell scores were predictive for OS (P = 0.03, Holm–Bonferroni correction), consistent with the central role of CD8⁺ T cells in eliminating tumor cells (40).

M1 macrophages are known to have antitumor characteristics (41). Conversely, M2 macrophages are thought to have opposing effects and repolarization of these to an M1 phenotype may improve therapy response (41, 42). As individual M1 and M2 macrophage expression scores were not predictive, and the ratio of M1/M2 macrophages has been described as having prognostic value in cancer (43), we developed a score for M1-M2 macrophages expression (Materials and Methods), and investigated its utility in predicting response to ICIs. We discovered that the difference in M1-M2 score between patients with a DCB and NCB was more pronounced than for the CD8⁺ T-cell score (Fig. 2D; median 21.2 vs. 7.6, P = 0.018). Similar to the CD8⁺ score, a high M1-M2 macrophage score (≥ median; Supplementary Fig. S1A) was associated with longer TTP (Fig. 2E; median 102 vs. 65, P = 0.041) and OS (median 373 days vs. 146, P = 0.0012).

In addition to CD8⁺ T-cell scores, and M1-M2 macrophage expression, we were also able to validate previously published immunologic biomarkers of response in our cohort (Supplementary Fig. S2A—S2L) including a CYT score (30) and IFNγ expression signature (31) which both correlated well with our CD8⁺ score (R = 0.8 and 0.74, respectively, Spearman). The previously published TIDE (33) had an inverse relationship (R = −0.5, Spearman) with our CD8⁺ score, and was not predictive in our pan-cancer cohort, unlike in the melanoma tumors in which this was originally developed.

Expression of measured by IHC, is used as a clinical biomarker to select patients that may benefit from ICIs (8), but there is evidence that PD-L1–low or –negative patients can also respond to therapy (44, 45). We therefore sought to investigate the utility of RNA expression of the PD-L1 gene (CD274) as a biomarker of response to ICIs in this cohort. Expression of CD274 was highest in a lymphoma and a thymic carcinoma (Fig. 2F), consistent with its natural expression in the immune cells composing these tumors. Some LUNG, cholangiocarcinoma, and SKCM also exhibited high expression of CD274. Comparison with PD-L1 IHC staining available for a subset (n = 11) of NSCLCs showed that samples with the highest PD-L1 staining in tumor cells (>50%) consistently had the highest levels of CD274 expression (Supplementary Fig. S1F; P = 0.047, Kruskal–Wallis), suggesting that CD274 expression may be a useful surrogate for IHC staining. Furthermore, consistent with a recent study examining PD-L1 status (46), we observed that CD274 expression was not correlated with TMB (R = 0.051, P = 0.65, Spearman).

Median CD274 expression was not significantly higher in responders (Fig. 2G; median 8.9 RPKM vs. 1.8, P = 0.12), though the data may suggest a trend. In addition, we noted that many responders had little to no CD274 expression, and that nonresponders often also had high expression. Interestingly, the highest expressing responders encompassed a variety of tumor types (one melanoma, one thymic, one colorectal, one NSCLC, and one sarcomatoid carcinoma of the lung) and not just NSCLCs for which the biomarker is currently validated. Patients with high CD274 expression (≥80th percentile, 4.8 RPKM) had a longer TTP (median 104 days vs. 79, P = 0.032), but had no OS benefit (Fig. 2H; median, 340 days vs. 271, P = 0.87). This suggests that, while high CD274 expression can be observed across tumor types, it may not be the most informative biomarker for response in a pan-cancer setting.

Clinically approved biomarkers for patient selection for ICI therapy, such as PD-L1 positivity or dMMR status, are currently considered in isolation (47). We investigated the potential for improved predictive value of combinations of biomarkers. Using TMB, CD8⁺ T-cell score, M1-M2 score, and CD274 expression, we performed multivariate Cox proportional hazards models, also incorporating tumor type to control for any underlying variation in response (see Materials and Methods). High CD8⁺ T-cell scores (≥ median; HR = 0.41, P = 0.03) and TMB (≥10 mutations/Mb; HR = 0.36, P = 0.05) were most predictive of TTP (Fig. 3A). M1-M2 scores were most beneficial for OS (≥ median; HR = 0.26, P = 9.4 × 10⁻⁰⁴) and a high TMB or CD8⁺ T-cell score also provided prognostic benefit (HR = 0.36, P = 0.03 and HR = 0.46, P = 0.06, respectively). Of the 33 samples with high CD8⁺ T-cell score, 75% also showed a high M1-M2 score; 12.5% of samples were high for only one of these. Despite this overlap, these results indicate that both biomarkers contribute to patient response. The presence of high CD274 expression did not provide any additional benefit in TTP [≥ 80th percentile; HR = 0.84 (0.32–2.22), P = 0.72] or OS [HR = 0.93 (0.35–2.47), P = 0.89], indicating that any benefit from CD274 expression can be explained in our cohort by positivity for another marker. We additionally examined the markers as continuous covariates but did not find the results to be as robust (overall model TTP; P = 0.04, OS; P = 0.6). It is possible that the effect of these markers on response and prognosis is only observed at higher levels, consistent with high TMB thresholds described previously (35, 36) and is not as discriminating at lower values.

Samples with a DCB had more positive markers than those with NCB (Fig. 3C; median 2 vs. 1, P = 0.0028), excluding CD274 due to its lack of predictive value. Interestingly, all patients with a DCB were positive for at least two markers, except for one patient with stomach adenocarcinoma whose biopsy sample was negative for all markers that we evaluated. Samples with more markers had significantly improved TTP [Fig. 3D; HR = 0.62 (0.47–0.81), P = 0.0006] and OS [Fig. 3D; HR = 0.56 (0.42–0.75), P = 0.0001].

In addition to our pan-cancer cohort, we also found that there may be predictive value in using the combination of TMB, CD8⁺ T-cell score, and M1-M2 score in independent, previously published cohorts from single tumor types (Supplementary Fig. S3; refs. 27, 28). While the number of samples in these cohorts is smaller, precluding statistical significance, we still observe a trend in the same direction toward the individual predictive value of CD8 score, TMB, and M1-M2 macrophage score, and that responders trend in the direction of having more than one marker. In another published melanoma cohort by Liu and colleagues (29), the authors report no trend with respect to TMB or T-cell score; our reanalysis of this melanoma cohort confirmed their findings, and additionally showed no trend with M1-M2 macrophage score. The authors attributed the lack of observed association partly to confounding rare tumor subtypes. This exception to the consistent trends found in other tumor datasets underscores the complexity of evaluating biomarkers in heterogeneous cohorts, and suggests that some rare tumor subtypes may be more amenable to alternative biomarkers, the elucidation of which will require further detailed study.

We examined alterations in genes in antigen presentation and JAK/STAT signaling pathways, as these have been reported as ICI resistance mechanisms (48, 49). Considering all types of somatic alterations in these genes, we observed no difference in the proportion of altered samples between DCB and NCB groups (Fig. 4A; 24% vs. 31%, P = 0.78, χ² test). Notably, however, all samples with biallelic alterations (events which affect all copies of the gene due to copy loss or loss of heterozygosity) were in the nonresponder group. This observation could not reach statistical significance (0% vs. 12%, P = 0.28) due to the small number of samples with these events (n = 8). In the 16 patients who progressed on ICIs before WGTA, biallelic losses were only present in patients who had previously exhibited a DCB, but the small number of patients (n = 8) prevented a statistical interpretation of this observation (Fig. 4B).

One patient with NSCLC is particularly remarkable; she presented (ages 50) with stage IV squamous cell lung cancer with bilateral lung masses, mediastinal adenopathy, and liver metastases. She had a 10-pack-year smoking history but had quit 20 years prior to this diagnosis. This was her fourth cancer diagnosed within 10 years, having had surgery for a T2 breast cancer, T3N1 colorectal cancer, and a stage IA endometrial adenocarcinoma. Next-generation sequencing of germline DNA was uninformative; however, subsequent clinical testing identified constitutive MLH1 promoter hypermethylation in the patient. For her NSCLC she received palliative radiation to her lung mass and initiated first-line pembrolizumab 4 weeks later. The patient underwent a biopsy for WGTA both before (liver biopsy) and after (supraclavicular node) pembrolizumab. On the basis of WGTA this patient was a strong candidate for receiving ICIs (dMMR, high TMB, CD8⁺ T cell and M1-M2 scores, and high CD274 expression; Fig. 4C). Although her liver and lung nodules did respond to pembrolizumab, she had progression in her supraclavicular fossa and bone after 104 days on therapy, therefore, she did not have a DCB. The second biopsy of the growing supraclavicular mass was taken 135 days after treatment initiation which revealed biallelic frameshift mutations in both JAK1 and HLA-A, neither of which were found in the other biopsy (liver) before treatment (Fig. 4C). This may represent an acquired resistance mechanism to ICIs, or presence of a resistant subclone which was selected for by ICI treatment, which highlights the importance of profiling metastatic lesions.

As the complexity of cancer care increases with technological advances and emergence of new therapeutic options, identification of the optimal patient population for any given intervention is crucial. Current molecular testing to align patients to immunotherapies are usually panel- or IHC-based approaches focused on individual biomarkers, for example, MMR status and PD-L1 expression. However, it is becoming clear that individual biomarkers are neither necessary nor sufficient to select patients most likely to benefit. WGTA has the advantage of producing a wealth of information that would otherwise require multiple independent tests, which is particularly valuable given our findings that multiple markers are predictive of ICI response. Our accrual of WGTA data for a pan-cancer cohort of ICI-treated patients will permit examination of more unusual responders, and can be used retrospectively (see Data Availability) to identify and evaluate new markers of response and resistance.

Our results lend support to development of TMB as a tumor agnostic clinical biomarker, confirming previous findings (11, 34), and the recent tumor-agnostic FDA approval of pembrolizumab in patients with a TMB >10 mutations/Mb using the FoundationOne CDx assay (7). Our results also demonstrate a strong relationship between exome- and genome-derived TMB counts with little difference in predictive value, providing further support for the applicability of WGTA in clinical testing.

Our findings emphasize the fundamental predictive value of immune-related markers for ICI response. Consistent observations across tumor types, cohorts, and different measures of T-cell activity [CD8 score, IFNγ score (31), and cytolytic score (30)] suggest a clear role for expression-based signatures of T-cell activity in a pan-cancer context. Our macrophage score, reflecting the expression of M1 and M2 macrophage markers, has not been previously described and evaluation both in our cohort and in published datasets suggests this may warrant further investigation as another independently predictive expression-based immune-related marker of response.

In contrast, we found that CD274 expression was not predictive in a tumor agnostic context, despite good concordance with PD-L1 IHC, a clinical biomarker evaluated in specific tumor types including NSCLCs. Even within this population, there are many PD-L1 diagnostic tests available with varying thresholds paired with specific drugs (50), suggesting that the clinical benefit from PD-L1 may be tumor or drug specific.

ICI resistance is reported across multiple cancer types with patient-based and in vivo studies suggesting roles for antigen presentation and JAK/STAT signaling pathways (48, 49, 51). Loss-of-function mutations in JAK1 similar to those found here have been previously associated with resistance to pembrolizumab in patients with melanoma (51), supporting similar mechanisms of resistance across tumor types. Our finding that biallelic mutations and copy losses in these pathways were limited to patients with NCB, or patients post-ICI treatment, suggests that complete loss of gene function may be key to resistance. This has parallels to the patterns of alterations of tumor suppressor genes during cancer development (52). We did not observe recurrent alterations, suggesting that resistance may be driven by loss of a pathway, rather than a single gene. Future work in larger patient cohorts will expand on the clinical relevance of these and other potential resistance alterations.

In summary, we have presented a unique WTGA dataset for pan-cancer ICI-treated tumors, and demonstrated that a combination of biomarkers including high TMB, CD8⁺ expression, and M1-M2 macrophage scores can predict both TTP and OS on ICIs. These data support a potential move away from PD-L1 testing as the clinical biomarker of choice for immunotherapy, and point toward the use of a multifaceted assessment of many tumor features, as is afforded by WGTA. Challenges in measuring and standardizing biomarkers for clinical use across tumor types, as illustrated by work to standardize TMB measures across panel platforms (53), demonstrate the importance of exploration in pan-cancer cohorts for hypothesis generation, and follow-up with prospective clinical trials. A pan-cancer approach based on WGTA data and integrating multiple biomarkers based on the data presented here has been implemented in a phase II clinical trial at BC Cancer (Vancouver, Canada), CAPTIV-8, which uses WGTA to select patients for treatment with the PD-L1 inhibitor atezolizumab (NCT04273061). This illustrates the potential of integrative genomics approaches to identify patients best suited for immunotherapies in current medical practice.

A. Pender reports other from Roche Canada during the conduct of the study; A. Pender also reports other from Guardant Health and BMS outside the submitted work. C.J. Grisdale reports grants from NIH during the conduct of the study. R.A. Moore reports grants from Genome Canada, TFRI, and BC Cancer Foundation during the conduct of the study. J.-M. Lavoie reports personal fees from Hoffmann-La Roche Limited outside the submitted work. S. Yip reports receiving (advisory boards) travel allowances from Amgen, AstraZeneca, Bayer, Norvatis, Pfizer, and Roche, none of which are associated with the submitted work. H. Lim reports other from Ipsen, Merck, BMS, Taiho, Eisai, and Roche outside the submitted work. D.J. Renouf reports grants and personal fees from Roche, as well as personal fees from Bayer, Celgene, Servier, Ipsen, Taiho, and AstraZeneca outside the submitted work. S.J.M. Jones reports grants from BC Cancer Foundation, Terry Fox Research Institute, Canada Foundation for Innovation, Canada Research Chairs, and BC Knowledge Development Fund during the conduct of the study. J. Laskin reports grants from BC Cancer Foundation and Roche Canada during the conduct of the study; grants from Roche Canada; and personal fees from Roche Canada, Pfizer, Eli Lilly, and AstraZeneca outside the submitted work. No disclosures were reported by the other authors.

A. Pender: Conceptualization, data curation, supervision, writing-original draft, writing-review and editing. E. Titmuss: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. E.D. Pleasance: Conceptualization, formal analysis, supervision, investigation, methodology, writing-original draft, writing-review and editing. K.Y. Fan: Data curation, formal analysis, writing-original draft, writing-review and editing. H. Pearson: Conceptualization, data curation. S.D. Brown: Investigation, writing-review and editing. C.J. Grisdale: Data curation, writing-review and editing. J.T. Topham: Investigation, writing-review and editing. Y. Shen: Resources, writing-review and editing. M. Bonakdar: Resources, writing-review and editing. G.A. Taylor: Resources, writing-review and editing. L.M. Williamson: Resources, supervision, writing-review and editing. K.L. Mungall: Resources, data curation, methodology, writing-review and editing. E. Chuah: Resources, methodology, writing-review and editing. A.J. Mungall: Resources, methodology, writing-review and editing. R.A. Moore: Resources, methodology, writing-review and editing. J.-M. Lavoie: Data curation, writing-review and editing. S. Yip: Writing-review and editing. H. Lim: Writing-review and editing. D.J. Renouf: Writing-review and editing. S. Sun: Writing-review and editing. R. Holt: Conceptualization, writing-review and editing. S.J.M. Jones: Supervision, funding acquisition, writing-review and editing. M.A. Marra: Supervision, funding acquisition, writing-review and editing. J. Laskin: Conceptualization, supervision, funding acquisition, writing-original draft, writing-review and editing.

The POG program, and subsequent studies, would not be possible without the participation of our patients and families, and support from the POG team. This work was supported by the British Columbia Cancer Foundation, Genome British Columbia (B20POG), Genome Canada and Genome BC (202SEQ, 212SEQ, 12002), Canada Foundation for Innovation (20070, 30981, 30198, 33408, 35444), and the BC Knowledge Development Fund.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.