Plasma versus Tissue Tumor Mutational Burden as Biomarkers of Durvalumab plus Tremelimumab Response in Patients with Metastatic Colorectal Cancer in the CO.26 Trial

To date, immunotherapy has shown limited activity in microsatellite stable (MSS) metastatic colorectal cancer (mCRC; refs. 1, 2), despite single-agent or combination immune checkpoint inhibition showing significant activity in patients with microsatellite instability-high (MSI-H) mCRC (3–5). Tumor mutation burden (TMB) has been assessed across many tumor types as a potential predictive biomarker (6–8). Most studies suggest that high TMB is associated with improved outcomes following immunotherapy. The hypothesis is that higher TMB leads to more neoantigens for the immune system to recognize (9). Recently, the FDA granted a tissue-agnostic accelerated approval for the use of pembrolizumab in patients with TMB ≥10 mutations/megabase (Mb) measured using the tissue-based FoundationOne CDx companion diagnostic based on the KEYNOTE-158 study (10). However, KEYNOTE-158 did not include colorectal cancer (11). The lack of prospective data showing TMB as predictive of benefits from immune checkpoint inhibition in mCRC leaves oncologists and patients with an approved agent that has limited tissue-specific data in one of the most commonly diagnosed cancers. Additionally, TMB is not well standardized across sequencing platforms, although efforts such as the TMB Harmonization Project (12) and the development of computational pipelines including TMBur (13) are underway to improve this issue.

The Canadian Cancer Trials Group (CCTG) CO.26 trial was a phase II study that randomized 180 patients with treatment-refractory mCRC to receive either durvalumab (an IgG1 monoclonal antibody against PDL1) in combination with tremelimumab (an IgG2 monoclonal antibody against CTLA4) or best supportive care (BSC) in a 2:1 ratio. The trial did not select for microsatellite instability (MSI) status. However, of the 165 patients whose MSI status was determined from plasma, only two (1.2%) had MSI-H tumors. The primary endpoint of OS among patients with MSS tumors has previously been reported [HR, 0.66; 95% confidence interval (CI), 0.49 to 0.89; P = 0.02], and exploratory analysis suggested that TMB derived from circulating tumor DNA (ctDNA) in plasma seemed beneficial in patients with MSS tumors. Among 163 patients with evaluable plasma-derived TMB (pTMB) ≥28 mutations/Mb, HR for death was 0.34 (95% CI, 0.18–0.96; P = 0.004) with P interaction = 0.07 (14).

Given the lack of prospective data evaluating tissue-based TMB in MSS mCRC and the relevant preliminary findings in CO.26 based on pTMB, we aimed to compare plasma and archival tissue-derived TMB and assess if either provided utility as a predictive biomarker of immune checkpoint inhibition. Additionally, we sought to explore the influence of other factors such as prior therapy on pTMB. This study includes the updated and final follow-up data from the trial that has not been reported (14).

CO.26 was an unblinded phase II clinical trial that randomized 180 patients (Table 1, Supplementary Table S1) with treatment-refractory mCRC 2:1 to receive either durvalumab 1,500 mg intravenously every 4 weeks + tremelimumab 75 mg intravenously every 4 weeks for the initial four cycles + BSC or BSC alone after written and informed consent. The study received institutional review board approval and was carried out per the Declaration of Helsinki and International Ethical Guidelines for Biomedical Research Involving Human Subjects (NCT02870920). Full study details and protocols were previously published (14). This study represents the updated results of the CO.26 trial with a database lock date of June 1, 2020, and tissue/plasma-based correlative analyses of TMB.

The original trial was designed to have 80% power and a two-sided alpha of 0.10 to detect a 35% reduction in death. Overall survival (OS) and progression‐free survival (PFS) were analyzed according to the intention to treat. OS was defined as the time from randomization until death or date of the last follow-up with patients alive during the last follow-up being censored. PFS was defined as the time from randomization until progression (defined by RECIST, version 1.1) or death, with patients censored during the last follow-up if an event had not occurred (15). Linear regressions were performed in R using the stats package (v3.6.3), and ß coefficients (slope) are reported with associated P values.

Archival FFPE tumor tissue and leukocytes from peripheral blood were utilized for tumor/matched reference sequencing. The Allprep DNA FFPE kit was used to purify genomic DNA per the manufacturer’s instructions from tissue, and QIAamp DNA Minikit was utilized for DNA extraction from blood-derived leukocytes. Sequencing libraries were created using the Agilent SureSelectXT method with subsequent hybridization and capture with SureSelect Human All Exon v6 baits (Agilent Technologies). Quantified, normalized, and pooled libraries were then multiplexed and sequenced using an Illumina platform with a sequencing protocol of 100-bp paired-end sequencing, with a read depth necessary to reach an average coverage before deduplication of 100× for reference or 400× for tumor samples.

Variant calling and TMB determination were assessed based on established guidelines (12, 16). Briefly, somatic variants were called on tumor-reference pairs through a combination of Manta (RRID:SCR_022997; v1.5.0) and Strelka v2.9.10, using default parameters and genome build GRCh37/hg19 (17, 18). VCF files produced by Strelka were converted to MAF format using vcf2maf v1.6.18 with default parameters, which included variant annotation by VEP v93 and flagging/removal of common variants using ExAC (ExAC_nonTCGA.r0.3.1.sites.vep.vcf.gz, downloaded on June 19, 2020; ref. 19). Variants were then intersected with consensus exon regions (GRCh37.p13; GCF_000001405.25, downloaded on June 19, 2020) to exclude those located outside exons and filtered using established criteria (12), such as VAF ≥ 0.05, tumor depth ≥ 25, and alternate allele count ≥ 3. We only included nonsynonymous mutations that were missense, nonsense, and in-frame/frameshift; overlapping/redundant variants were excluded. The 32.102474 Mb was used for the TMB denominator (12). Variants were considered subclonal if they occurred at a relative variant allele frequency (rVAF) of <10%. rVAF was defined as the allele frequency of a variant divided by the maximum detected allele frequency of a somatic variant. Variants occurring at <5% allele frequency were excluded in tissue samples because they were derived from FFPE and at risk of deamination artifact. MSI status was called via Mantis v1.0.4 using 2,539 sites obtained from the Mantis project repository and recommended parameters in which MSI/MMR is not available by standard-of-care immunohistochemistry or polymerase chain reaction before trial enrollment (20). Cases were classified as TMB high by tissue if TMB was ≥10 mutations/Mb.

Blood samples for ctDNA were collected in Streck tubes before study treatment and plasma was isolated. DNA extraction and next-generation sequencing were subsequently performed at Guardant Health using the GuardantOMNI next-generation sequencing 2.15 Mb, 500 gene panel to detect SNV, indels, fusions, copy-number amplifications, MSI-high status, and pTMB (21). pTMB was reported as mutations/Mb by the GuardantOMNI algorithm that includes all somatic synonymous and nonsynonymous SNV and indels, excluding germline, clonal hematopoiesis of indeterminate potential (CHIP), driver, and resistance variants with statistical adjustment for sample-specific tumor shedding of ctDNA and coverage (22). The TMB and MSI assessment algorithms have previously been described (23, 24). Variants occurring at a relative allele frequency <10% were considered subclonal. The rVAF of each variant was defined as the allele frequency of that variant divided by the maximum detected allele frequency of a variant in that sample and was only considered for SNV. Cases were classified as TMB high by plasma if TMB was ≥28 mutations/Mb based on our prior work and lack of other available cut points suggestive of the appropriate threshold (14).

Mutations called in tissue and plasma in genes from the GuardantOMNI panel were compared for each patient. One mutation was shown per gene and patient on the oncoprint in Fig. 1C and instances in which multiple alterations in a gene occurred per patient were colored based on the following hierarchy: alteration is present in tissue and plasma, absent in tissue, missing in plasma, and present in baseline, and tissue is unknown.

For minimal P value analyses, hazard ratios and P values were calculated using the survival R package (v3.1-8). Minimum and maximum TMB thresholds considered were chosen so that each group for comparison would have at least three patients. For clonal plasma TMB, this was 2.0 and 14.7 mutations/Mb, and for tissue TMB, it was 2.7 and 36.1 mutations/Mb, and for subclonal TMB, it was between 1.0 and 46.0 mutations/Mb. In total, 21 patients had no subclonal mutations detected and were counted as having a subclonal TMB of 0 mutations/Mb.

Survival analyses (Kaplan-Meier and Cox proportional hazards) were performed using the R packages survival (v3.1-8) and survminer (v0.4.7). Cytotoxic agents used to examine the association with therapy were oxaliplatin, fluorouracil, irinotecan, capecitabine, floxuridine, and cisplatin. The length of time on cytotoxic agents was defined as the total exposure time to chemotherapy agents (cumulative over different treatment lines) for all therapies received before study enrollment in CO.26. Timing was grouped into ≥3, ≥1, and <1 years, as this was previously reported to correlate with TMB (25), and the groups were close to the 25th and 75th percentiles in our cohort. The two patients with no listed cytotoxic agents before enrollment in CO.26 were excluded from this analysis, as the group is too small to compare. When examining the impact of DNA repair mutations on TMB, we included any patients with DNA repair alterations in plasma in the following genes: ATM, BRCA2, BRCA1, POLE, POLQ, POLD1, MLH1, MSH6, MSH2, POLH, and PMS2. All multivariable models included sex, disease side, ECOG status, liver metastases, and age.

Somatic mutations detected in plasma by GuardantOMNI were classified as putative CHIP mutations if the mutations occurred in genes or as specific variants commonly annotated as CHIP or heme-derived in literature or detected in the sequencing of healthy reference donors (23). Variants detected at an allele frequency ≥2% were considered CHIP variants for this study.

Data used for this analysis were part of a clinical trial and are housed at the CCTG. An investigator who wishes to analyze data from this study must make a formal request to the CCTG using the specific form available at https://www.ctg.queensu.ca/docs/public/policies/TMG-FRM-0199-Application-for-Data-Sharing-V2-2017Feb23.docx to ensure that data analysis maintains confidentiality of patient information and data-sharing stipulations from the patient consent forms and the participating jurisdictions. The application must be submitted to datasharing@ctg.queensu.ca. CCTG will review requests according to the institution’s data-sharing and access policy available at https://www.ctg.queensu.ca/docs/public/policies/DataSharingandAccessPolicy.pdf.

Among the 180 patients enrolled in CO.26 (Supplementary Fig. S1A; Supplementary Table S1), 168 had plasma samples available from baseline (Fig. 1A; Supplementary Fig. S1A), of whom 98% (n = 165) were evaluable for plasma TMB (pTMB). Patients were randomized 2:1, 70% (n = 118) received durvalumab and tremelimumab (D + T), and 30% (n = 50) received BSC. Additionally, 109 patients had WES available from archival tissue with corresponding tissue TMB, 108 of whom had evaluable pTMB (Fig. 1A). Tissue samples were taken an average of 3.4 years before the baseline plasma sample (median range, 0.4–13.0 years). Patients had a median age of 65 (39–87) years, and one-third of patients (33%) were female (Table 1; ref. 14). Similar baseline characteristics were observed between the original full trial cohort, those with assessable pTMB, and those with additional archival tissue (Table 1).

Among the 165 patients with assessable baseline pTMB, two (1%) had microsatellite instability detected in plasma. In the paired tissue exomes, both seemed to be microsatellite stable (MSS, Materials and Methods); however, neither had a clinical-grade assay available to confirm MSI status. The two samples with detected MSI-H in plasma had a 10-fold higher pTMB than plasma samples without MSI detected [median, 160.9 mutations/Mb (values: 74.69, 247.08] vs. 16.3 mutations/Mb, Fig. 1B], yet a much smaller (1.4-fold) difference between the TMB in the corresponding tissue samples [median, 8.8 mutations/Mb (values: 5.64, 11.87] vs. 6.4 mutations/Mb). Among 33 patients who had available clinical MSI testing, 100% were MSS by WES, and their baseline plasma samples were also determined to be MSS. MSI-H is already an FDA-approved predictive biomarker in mCRC, so all subsequent outcome analyses exclude the two patients with MSI-H detected in plasma.

The most commonly mutated genes across tissue samples were APC (63% of patients, Fig. 1C), TP53 (55% of patients), and KRAS (51%), consistent with prior reports (26). Mutations in these genes were often corroborated in the baseline plasma samples (71% of APC mutations in WES were also detected in plasma; 84% and 93% of TP53 and KRAS mutations detected by WES were also detected in plasma). Mutations in DNA repair genes were also common in the tissue samples in this cohort (e.g., ATM and BRCA2 mutations in 14% and 12%), and these were less often corroborated in the corresponding plasma samples (11% and 6% of ATM and BRCA2 mutations).

pTMB was significantly higher than tissue TMB [median, 15.3 mutations/Mb (IQR, 9.5–26.2) vs. 6.5 mutations/Mb (3.9–12.0), P = 8.7 × 10⁻⁰⁶; Fig. 1D, paired Wilcoxon rank-sum], and the two were not correlated (r = 0.13, P = 0.20 Spearman, Fig. 1E). When considering pTMB from clonal mutations only [>10% relative VAF (rVAF); Materials and Methods], there was less difference between plasma and tissue TMB [median, 5.8 mutations/Mb (IQR, 3.8–8.6) vs. 6.5 mutations/Mb (3.9–12.0), P = 0.021, Fig. 1D] and a weak correlation between the two (r = 0.25, P = 0.010, Fig. 1F, Spearman). We did not observe a correlation between total pTMB and the time (years) between the tissue and the plasma sample (Spearman R = 0.13, P = 0.20) nor between the fold change between the tissue and the plasma and the time between the two samples (R = −0.13, P = 0.19). There was a weak positive correlation between the time between the samples and clonal pTMB (R = 0.34, P = 0.00034; subclonal R = 0.02, P = 0.84); however, no trend was noted when looking at the fold change in TMB (clonal R = 0.0051, P = 0.96; subclonal R = −0.11, P = 0.27).

Results reported initially from the trial found that durvalumab + tremelimumab improved median OS by 2.50 months compared with BSC [HR, 0.72 (90% CI, 0.54–0.97); P = 0.07; ref. 14]. The final improvement in OS for D + T compared with BSC was 2.99 months after database lock [median, 6.60 vs. 3.61 months; HR, 0.68 (95% CI, 0.49–0.93); P = 0.017; Fig. 2A]. As with the original report that showed a minimal difference in PFS of 0.1 months, the final PFS for the trial was not different between D + T and BSC with a difference in median follow-up of 0.0 months [median, 1.84 vs. 1.84 months; HR, 0.94 (95% CI, 0.68–1.29); P = 0.70; Supplementary Fig. S2A].

Analysis conducted in our original report (minimal P value approach) suggested that the best predictive threshold for dichotomizing patients based on pTMB was 28 mutations/Mb (14). With updated follow-up and inclusion of other variables known to influence the outcome in a multivariable Cox model, we continued to find that durvalumab + tremelimumab was associated with improved OS in patients with a pTMB ≥28 mutations/Mb versus BSC alone (HR, 0.34; 95% CI, 0.13–0.85; P = 0.02; interaction P value = 0.070; Fig. 2B and C), as well as a higher disease control rate [complete response, partial response, stable disease; 23.8% (95% CI, 5.59–42.03) vs. 0% (95% CI, 0.0–0.0); P = 0.068, Fisher’s exact test; Fig. 2D]. A difference in OS between treatment arms was not observed for patients with a pTMB <28 mutations/Mb [HR, 0.93 (0.60–1.43); P = 0.73; Supplementary Fig. S2B, multivariable Cox model].

Dichotomizing patients based on the FDA-approved predictive threshold of 10 mutations/Mb in tissue did not provide value as a predictive biomarker for durvalumab + tremelimumab vs BSC for OS [HR, 0.71 (95% CI, 0.28–1.80); P = 0.47, interaction P value = 0.62; Fig. 2E and F] or disease control rate (19.0% vs. 10.0%; P = 1; Fig. 2G). In a multivariable Cox model, the only predictive features for survival in patients with a tissue TMB ≥10 mutations/Mb were the ECOG status and the presence of liver metastases (Fig. 2F). Interestingly, there was a trend toward improved OS on D + T in patients with a tissue TMB <10 mutations/Mb in a multivariable Cox model (HR, 0.61; 95% CI, 0.34–1.07; P = 0.083; Supplementary Fig. S2C).

Comparing the two thresholds (10 and 28 mutations/Mb in tissue and plasma), we found that 42% (n = 45/106) of patients would be discordant for high or low TMB between their tissue and plasma samples (Fig. 2H). Despite pTMB being higher than tissue TMB, 62% of discordant patients were categorized as tissue TMB high and pTMB low. Only three patients (3%) were considered to have high TMB in both samples.

We attempted to identify an optimal threshold for OS with tissue TMB using a minimal P value approach (Supplementary Fig. S2D–S2F) and found this to be <3.4 mutations/Mb [HR, 0.12 (95% CI, 0.020–0.75); P = 0.023, interaction P value = 0.0062; Supplementary Fig. S2E and S2F].

As clonal pTMB (>10% rVAF) was closer to tissue TMB, we explored the impact of clonal pTMB on the outcome. Using a minimal P value approach, we identified the optimal threshold for separating patients based on improved OS with D + T to be ≥10.6 mutations/Mb [HR, 0.20 (95% CI, 0.068–0.59); P = 0.0035, interaction P value = 0.036; Fig. 3A; Supplementary Fig. S3A and S3B]. This remained significant in a multivariable Cox model including sex, disease side, ECOG status, and age [HR, 0.10 (95% CI, 0.014–0.79); P = 0.029; Fig. 3B]. There was no difference in disease control rate, but between the two treatment arms [33.33% (95% CI, 9.48–57.19) in D + T vs. 14.29% (0.0–40.21); P = 0.62, in BSC; Fig. 3C]. Using the 10.6 mutations/Mb threshold in clonal pTMB and the FDA-approved tissue threshold of 10 mutations/Mb, a lower proportion of patients (30%, n = 31; Fig. 3D) were discordant between the groups compared to using the pTMB threshold of 28 mutations/Mb. A total of 6% of patients (n = 6) were high in both groups. Using the minimal P value approach, we found the optimal threshold for separating patients based on OS for subclonal pTMB (<10% rVAF) to be 25.9 mutations/Mb [HR, 0.32 (0.13–0.72); P = 0.011, interaction P value = 0.069; Supplementary Fig. S3C and S3D), which is close to the 28/Mb for total pTMB. When examined in a multivariable model, subclonal pTMB remained significant [HR = 0.20 (95% CI, 0.061–0.69); P = 0.010].

Given that TMB has been shown to increase in tissue samples as a result of prior cytotoxic therapy (25) and plasma samples for CO.26 were collected after multiple cytotoxic chemotherapy regimens, we aimed to examine the association between therapy, pTMB, and its impact on the outcome. Because all except two patients had an exposure to cytotoxic agents (Materials and Methods) before enrollment in CO.26, we explored the association between pTMB and cytotoxic agents using the total length of time on any cytotoxic agent. Additionally, we considered the use of EGFR antibodies and regorafenib because these are common therapies for mCRC and have been linked to the development of resistance alterations (25, 27). A longer time on cytotoxic agents was associated with higher pTMB (r = 0.28, P = 0.00031; Fig. 4A, Spearman), and this was specific to pTMB derived from subclonal mutations (<10% VAF, r = 0.26, P = 0.0 0093; clonal mutations r = 0.037, P = 0.64). We did not observe a difference in OS between D + T and BSC in patients on cytotoxic agents for the longest time [≥3 years; HR, 0.76 (95% CI, 0.31–1.86); P = 0.54, interaction P value = 0.91; Fig. 4B] and those on for the shortest [<1 year; HR, 0.72 (0.37–1.38); P = 0.32; Fig. 4C].

Comparing patients who had received anti-EGFR antibodies before CO.26 and those who had not, we found that pTMB was higher in the treated group [median, 24.90 mutations/Mb (IQR, 15.32–38.30) vs. 13.41 mutations/Mb (7.90–21.07); P = 2.2 × 10⁻⁰⁷; Fig. 4D]. Similar to cytotoxic agents, we found that this was specific to subclonal pTMB [median, 18.20 (IQR, 7.66–30.64) vs. 5.74 (1.12–12.50), P = 9.7 × 10⁻⁰⁸; clonal, 5.75 (3.83–7.66) vs. 5.75 (3.83–8.62), P = 0.88; Fig. 4D]. As with prior clinical follow-up (27), we did not observe any difference in OS between BSC and D + T in patients who had received prior anti-EGFR therapy [HR, 0.61 (95% CI, 0.35–1.039); P = 0.069, interaction P value = 0.60; Fig. 4E] and those who had not [HR, 0.69 (0.43–1.11); P = 0.13; Fig. 4F]. When limiting the analysis to the group of patients who received prior anti-EGFR, the interaction between pTMB (≥28/Mb) and treatment arm did not persist [interaction HR, 0.64 (0.22–1.86), P interaction = 0.41; subclonal pTMB ± 25.9/Mb, interaction HR, 0.69 (0.23–2.13), P interaction = 0.52]. However, the presence of liver metastases was not prognostic in this subset of patients [univariable HR, 1.59 (0.84–3.010); P = 0.15], and the ECOG status was weakly prognostic [HR, 1.97 (1.013–3.85); P = 0.050]. Because both the covariates are known to be strongly associated with the outcome (14, 28), it suggests that we are underpowered to determine if there was a direct impact of EGFR-associated pTMB on the outcome. We did not observe a difference in pTMB between patients who had prior regorafenib exposure and those who had none [prior regorafenib (median, 21.5 mutations/Mb) vs. no regorafenib (16.3 mutations/Mb); P = 0.22; Supplementary Fig. S4].

We sought to explore whether other DNA repair mutations, except those linked to MSI, may be associated with a higher pTMB. We found that patients with any DNA repair alteration detected in plasma (genes in Fig. 1C) had a higher pTMB than those without [median, 22.68 (IQR: 16.76–37.35) vs. median 11.12 (6.70–17.24); P = 9.3 × 10⁻¹³; Fig. 4G]. Interestingly, unlike therapy-associated pTMB, a higher pTMB with DNA repair mutations in plasma was observed for both clonal [median, 5.75 (IQR, 3.83–9.10) vs. median 4.79 (2.87–6.70), P = 0.0098] and subclonal [median, 16.82 (IQR, 7.18–31.12) vs. median 4.79 (0.96–12.21), P = 2.8 × 10⁻⁰⁸] pTMB. Despite the association with a higher clonal pTMB, we did not observe a difference in OS between D + T and BSC in patients who had a DNA repair mutation [HR, 0.67 (95% CI, 0.41–1.10); P = 0.12, interaction P value = 0.98; Fig. 4H] and those without [HR, 0.60 (0.37–1.00); P = 0.049; Fig. 4I]. Limiting the analysis to the group of patients with DNA repair mutations, we found that the interaction between the treatment arm and pTMB (≥28/Mb) remained [interaction HR, 0.36 (0.13–1.01); P interaction = 0.052]. This was true for clonal [interaction HR, 0.27 (0.062–1.14); P interaction = 0.075] and subclonal [interaction HR, 0.37 (0.13–1.052); P-interaction = 0.062] pTMB. There were a few patients to assess in the group without DNA repair mutations (only one patient ≥28/Mb in the BSC arm).

To understand the impact of each factor independently (length of time on cytotoxic agents, prior anti-EGFR exposure, and presence of a mutation in a DNA repair gene), we performed a multivariable linear regression including these factors, as well as the time between the tissue and plasma samples. The association between each factor and higher total pTMB remained significant (EGFR antibodies β = 9.00, P = 0.00014; DNA repair mutation β = 13.04, P = 5.3 × 10⁻⁰⁸). The length of time on cytotoxic agents was associated with an increase in pTMB of 0.24 mutations/Mb per month on therapy (P = 0.0085). There was no positive association between pTMB and the time between the tissue and the plasma samples when included in the model (β = −0.090, P = 0.053). For clonal pTMB, there was a slight positive trend for increased pTMB with the time between samples (β = 0.060, P = 2.0 × 10⁻⁰⁴); however, our results remained consistent with a significant association between DNA repair mutations and higher clonal pTMB (β = 1.68, P = 0.028). For subclonal mutations, all our observations remained significant (anti-EGFR, β = 9.86, P = 6.23 × 10⁻⁰⁵; DNA repair, β = 11.41, P = 2.94 × 10⁻⁰⁶; length of time on cytotoxic agents, β = 0.28, P = 0.0027), and there was a negative trend for the time between samples (β = −0.15, P = 0.0027).

Sequencing of plasma samples also allows the detection of variants in hematopoietic stem cells associated with clonal expansion (CHIP). This phenomenon is associated with an increased risk of developing hematologic cancer and OS (29). A total of 41% of patients (N = 68) had at least one CHIP variant detected in their baseline plasma sample (Fig. 5A), a similar frequency to another cohort of mCRC reporting 36% with CHIP (30). CHIP mutations were most frequently found in DNMT3A (14% cohort, 34% of patients with CHIP mutations, n = 23 patients, Fig. 5A), followed by ASXL1, BCOR, and CREBBP (all in 4% of cohort, n = 6, Supplementary Fig. S5), similar to genes reported previously in mCRC (30). Patients with CHIP alterations were older than those without (median, 68 years vs. 64.5, P = 0.02, Fig. 5B, Wilcoxon rank-sum). We also noted that patients with CHIP alterations had a higher pTMB than those without it (median, 21.1 vs. 14.7; P = 0.0089, Fig. 5C, Wilcoxon rank-sum). However, after inclusion in a linear regression model with the length of time on cytotoxic agents, the use of anti-EGFR antibodies, and the presence of a DNA repair mutation in plasma, with age and sex, the association between the presence of CHIP and a higher pTMB was no longer statistically significant (b = 3.24, P = 0.15). There was no difference in the presence of CHIP between patients with different primary tumor locations (43% with CHIP for the rectum/left colon vs. 35% in the right colon, P = 0.39, Fisher’s exact test); however, we did observe a trend toward a higher likelihood of liver metastases in patients with CHIP (78% with mets vs. 65%, P = 0.09).

We did not observe any association between the presence of CHIP alterations and OS in patients who received BSC [HR, 1.12 (95% CI 0.63–2.031); P = 0.69; Fig. 5D] or D + T [HR, 1.11 (0.76–1.64); P = 0.59, interaction P value = 0.96; Fig. 5E].

We previously identified that a pTMB of 28 mutations/Mb may help select MSS mCRC patients likely to benefit from a combination of durvalumab and tremelimumab (14). In this study, we further explored the landscape of pTMB and found that this association was strong when considering either clonal or subclonal mutations. A higher threshold for the prognostic impact of TMB for immune checkpoint inhibitor (ICI) derived from plasma, specifically Guardant OMNI, used for this study, has been reported in non–small cell lung cancer (20 mutations/Mb; ref 22), suggesting that plasma may be a useful tool for reporting TMB in other tumor types, but the threshold may differ from what has been seen in other studies for tissue. In our study, we did not find that tissue TMB provided a useful biomarker.

The notion that TMB increases over time and that it can be exacerbated by successive cancer therapies, which directly damage DNA, or by applying a specific selective pressure suggests that repeat TMB testing may be used to monitor changes in TMB over time, which is facilitated by the use of ctDNA. The higher proportion of subclonal variants detected in plasma suggests there is a high likelihood of capturing ongoing and emerging mutation processes in plasma, including those induced by cancer therapies. We did not observe any association between prior therapy and the outcome, which may suggest that for MSS mCRC, the mutational mechanism is less important than the sheer number of potentially immunogenic mutations detectable in the plasma. Furthermore, our finding that clonal and subclonal pTMB are effective at selecting patients (at different thresholds) suggests that low-frequency passenger mutations obtained during therapy may still contribute to overall tumor immunogenicity.

There is limited literature surrounding CHIP and ICI response in solid tumors; however, a study reported a longer time on PD1/PDL1 agents for non–small cell lung cancer and melanoma patients with DNMT3A mutations (21 cycles; range, 10–40) compared with patients with wild-type DNMT3A (7 cycles, range, 1–13), although a small sample size meant that this did not reach statistical significance (31). This is possibly because CHIP mutations are more relevant in tumor types with a higher mutation burden, such as lung cancers, melanomas, and MSI-H CRC, which were underrepresented in this trial. However, prior work has suggested that the presence of CHIP may be a positive prognostic factor in mCRC and the presence of poorly functioning T cells in patients with CHIP may impact the immune response to checkpoint inhibition and requires further investigation (30).

One caveat to using a threshold derived from plasma is that not all ctDNA assays are created similarly. Our threshold of pTMB may not be directly applicable to pTMB derived from other assays. Understanding the nuances in pTMB and clonality will be important for selecting patients for ICI. Additionally, our CO.26 study is a combination PDL1 and CTLA4 trial, and there may be differences in applicable TMB thresholds with different antibodies. For example, early results from the phase Ia/Ib trial with the next-generation CTLA4 inhibitor botensilimab + balstilimab in MSS CRC have seen a 12-month OS rate of 61%, and 0/10 responders had a TMB ≥10 mutations/Mb (32). Similar results were found with the phase II basket PD1 and CTLA4 TAPUR study (33), in which they found insufficient evidence of clinical activity for patients with high TMB (≥9 mutations/Mb FoundationOne panel) MSS CRC. We attempted to find an ideal cutoff for tissue TMB in this cohort that came out far lower than the FDA-approved 10 mutations/Mb for pembrolizumab. It may be that studies so far (33) have been underpowered to properly evaluate tissue TMB. Additionally, it may be that only subsets of TMB, for example, clonal or most immunogenic, are applicable in MSS CRC as biomarkers of response.

In conclusion, pTMB (clonal and subclonal) may provide a TMB value predictive of response to a combination of PDL1 and CTLA4 inhibitors in MSS CRC. Further exploration using a validation cohort will be necessary to determine its use in clinical practice. Acquisition of plasma for sequencing is more feasible than tumor biopsies and allows for repeat evaluation of TMB during a patient’s treatment course when our therapeutic interventions may introduce vulnerabilities.

J.M. Loree reports nonfinancial support from AstraZeneca during the conduct of the study; personal fees from Amgen, Roche, Merck, and Pfizer; nonfinancial support from Saga Diagnostics; grants and nonfinancial support from Personalis Inc and Guardant Health; and grants and personal fees from Ipsen and Novartis outside the submitted work. H.F. Kennecke reports personal fees from Natera outside the submitted work. Y.S. Lee reports employment at AstraZeneca with stock or stock options. K. Quinn is an employee and stock holder of Guardant Health. D.J. Renouf reports grants and personal fees from Roche and Bayer and personal fees from Sanofi and Ipsen outside the submitted work. D.J. Jonker reports other support from CCTG during the conduct of the study. E.X. Chen reports personal fees from AstraZeneca during the conduct of the study. No disclosures were reported by the other authors.

J.M. Loree: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing. E. Titmuss: Formal analysis, visualization, writing–original draft, writing–review and editing. J.T. Topham: Formal analysis, visualization, writing–review and editing. H.F. Kennecke: Resources, writing–review and editing. H. Feilotter: Resources, writing–review and editing. S. Virk: Resources, writing–review and editing. Y.S. Lee: Resources, writing–review and editing. K. Banks: Resources, writing–review and editing. K. Quinn: Resources, writing–review and editing. A. Karsan: Methodology, writing–review and editing. D.J. Renouf: Resources, writing–review and editing. D.J. Jonker: Resources, funding acquisition, methodology, writing–review and editing. D. Tu: Formal analysis, methodology, writing–review and editing. C.J. O’Callaghan: Conceptualization, formal analysis, supervision, funding acquisition, investigation, methodology, writing–review and editing. E.X. Chen: Conceptualization, supervision, funding acquisition, investigation, writing–review and editing.

AstraZeneca provided durvalumab and tremelimumab as well as financially supported sequencing costs associated with correlative analyses. The CCTG is supported by the Canadian Cancer Society. J.M.L. and D.J.R. have received Michael Smith Health Professional Investigators awards, which helped support this study. Funds from the BC Cancer Foundation helped support this study. E.X.C. received honoraria from AstraZeneca and participated in clinical trials sponsored by AstraZeneca.