Introduction

The majority of patients with pancreatic ductal adenocarcinoma (PDAC) present with unresectable or metastatic disease at the time of diagnosis1. The incidence of PDAC continues to rise yearly for unknown reasons and is particularly notable in younger women2. Hereditary risk factors have been well described and in the advanced disease setting, germline alterations in BRCA1/2 and PALB2 or tumors deficient in homologous recombination repair (HRD), offer an opportunity for maintenance PARP inhibitors3. Notably, resistance to this precision approach has been associated with aneuploidy and the basal-like subtype4. Mismatch repair deficiency is present in approximately 1% of patients with advanced PDAC and the use of immune checkpoint inhibitors (ICIs) has provided heterogenous response rates for a biomarker that is considered tumor -agnostic5. Causative non-genetic risk factors have variable hazard ratios which are relatively low, including pre-existing diabetes, chronic pancreatitis, BMI and a smoking history1. However, our understanding how these factors may influence disease onset or impact subtype heterogeneity is limited.

The genomic landscape of PDAC has been extensively characterised and is dominated by KRAS mutations in over 90% with inactivation of the tumour suppressors, CDKN2A, TP53 and SMAD4 in 40–80%6,7,8. The accumulation of mutations in PDAC can present signatures particularly single base substitution signatures (SBS) highlighting the aetiology of disease including SBS3 denoting the presence of homologous recombination repair deficiency (HRD)9. Dominant SBS in this cohort included age-related SBS1 and 5 and SBS8 (unknown aetiology)10,11. Structural variant (SV) patterns are further diverse with approximately 20% of PDAC classified as unstable by Wadell et al describing PDAC with generally >200 SVs12, a cohort often synonymous with HRD-PDAC. Despite an improved understanding of tumour molecular biology, the treatment paradigm for PDAC has changed minimally in a decade. Modified (m) FOLFIRINOX represents the standard of care in the adjuvant setting for those suitable to receive it13. In advanced disease, mFOLFIRINOX, NALIRIFOX and gemcitabine nab/paclitaxel (GnP) are options, however median overall survival remains less than one year14,15,16. PDAC with wild-type KRAS (KRASWT) are enriched in fusions and BRAF alterations, considered targetable and may have distinct immunophenotypes and improved outcomes17. The clinical development of KRAS inhibitors targeting G12C has reinvigorated efforts across mutant KRAS alleles which may shape future management of PDAC18. Notably in NSCLC the KRASG12C mutant is associated with smoking. Although G12D is the most prevalent KRAS mutant in PDAC, there has been no association with specific risk factors. The COMPASS trial is a prospective observational study seeking to establish biomarkers in advanced PDAC through in-depth profiling prior to commencing chemotherapy. The primary objective was the feasibility of achieving a whole genome within 8 weeks of biopsy in over 80% of patients and this has been previously reported in the initial 62 patients19. Pre-defined secondary objectives of the trial included the determination of overall response rate (ORR) and overall survival (OS) according to chemotherapy regimen (physician choice) and evaluation of RNA based classifiers. Exploratory objectives included the identification of molecular predictors of chemotherapy and potential biomarkers of immunotherapy response, integrating epidemiological data. We have previously reported the prognostic impact of the basal-like and classical subtype identifiable by GATA6 expression20. Systemic inflammatory scores including the Gustave Roussy Immune score (GRIm-S) have also demonstrated prognostic value in our cohort21. In addition we have demonstrated hENT1 expression as a potential biomarker of gemcitabine response and the HRD subtype as predictor of platinum response in a collaborative study22.

Herein we report final, mature survival and integrated epidemiological and molecular data for the complete cohort of 268 patients enrolled in the COMPASS trial. In this dataset we describe the genomic landscape of patients with advanced PDAC including locally advanced unresectable PDAC and associate molecular features with outcomes and epidemiological risk factors. We demonstrate the continued prognostic impact of basal-like and classical subtyping. We highlight the enrichment of major imbalances in mutant KRAS in basal-like PDAC which has a distinct tumour microenvironment. In addition, patients with evidence of systemic inflammation at baseline represent a unique poor prognostic cohort with few CD8+ cells evident.

Results

Baseline clinical characteristics of cohort (n = 268)

In total, 332 patients were screened for COMPASS. Screenfails occurred in 10 patients (3%) including 3 patients where a biopsy was felt unsafe, 25 patients (8%) withdrew from the study. Of 297 eligible patient who underwent a biopsy, 17 cases (5%) had an alternative diagnosis on biopsy and 12 (4%) had insufficient tumour to process for WGS (Fig. 1). Of the 268 patients included, median follow-up was 52 months. Median age was 64 years and 41% of patients were female. Patients were documented as having ECOG 0 performance status in 37% (n = 99). The majority of patients had metastatic disease (86%) with liver metastases present in 71% (n = 191). LAPC made up 14% of the cohort (n = 37). Of the 253 patients with available RNA-seq, 19% harboured the basal-like subtype. mFOLFIRINOX was the chosen regimen in 146 patients (55%) and 99 received GnP (36%). Two (1%) patients received cisplatin/gemcitabine and 5 (2%) gemcitabine only. Of the 252 patients who received chemotherapy, 8% (n = 21) were given an experimental agent on trial; survival in these patients was not different compared to those receiving chemotherapy alone (median 8.2 months vs 9.3 months, HR 1.18, 95% CI 0.75–1.84, p = 0.47). Planned treatment was not given in 16 (6%) who are categorised as having non-evaluable disease but included in the intention to treat analysis. Notably, those receiving GnP were significantly older than those who received mFOLFIRINOX (67 vs. 62 years, p < 0.001) (Supplementary Data 1).

Fig. 1: CONSORT diagram of patients screened and enrolled on the COMPASS trial.
figure 1

PDAC: Pancreatic ductal adenocarcinoma; WGS: Whole genomic sequencing; *none this refers to patients intended to receive combination chemotherapy post enrolment.

Full size image

Biopsies on COMPASS

The median number of cores per patient was 6 (range 1–10). Biopsies were taken of metastatic sites in 201 patients (75%) and of the primary pancreas mass in 67 (25%). The primary tumour was biopsied percutaneously in 30 patients (11%) when metastases were evident but not feasible for core biopsy. There was only one complication related to biopsy which was from a liver metastasis in a patient who had been on anticoagulation. The patient who had appropriately stopped anti-coagulants had a self-limiting bleed post biopsy. Median tumour cellularity of the 268 baseline and 15 progression biopsies was 80% (16–96) after laser capture microdissection. Sequencing depth of tumour and normal was on average 49X and 38X respectively.

Genomic landscape of cohort

KRAS mutations were present in 93% (n = 249) of patient samples (Table 1) and KRASG12D accounted for the majority (n = 114, 43%) (Fig. 2); KRASG12C occurred in 4/268 (1%). Of the 19 patients (7%) with KRASwt PDAC, activation of MAPK signalling occurred through small in-frame deletions in the β3-αC helix of BRAF gene (Class II) in 6 (32%) with a further three RAF altered cases identified (two fusions and BRAF p.600_601del). Four tumours harboured point mutations in ERBB2/3/4 and alternate drivers were not evident in 2 cases (Table 2). Median age at enrolment was 59 years in the KRASWT cohort compared to 65 years in patients with KRASm (p = 0.38) (Table 1). There was no difference in median CA19.9 at presentation between KRAS WT and KRASm PDAC. Major imbalances in mutant KRAS (KRASmaj) were present in 17% (n = 45), minor imbalances in 36% (n = 96) and KRASbal in 41% (n = 108). Loss of function mutations or deletions in additional driver genes were documented in TP53 (82%), CDKN2A (84%) and SMAD4 (47%) (Fig. 3). KRASWT PDAC were less likely to harbour inactivating mutations in TP53 (p = 0.0004) and SMAD4 (p = 0.03), however the frequency of biallelic loss of CDKN2A was similar across KRAS allelic states (p = 0.64) (Fig. 4A).

Fig. 2: KRAS mutational spectrum on the study.
figure 2

Frequency of specific KRAS mutations of tumours (n = 249) included together with KRAS WT PDAC proportion (n = 19).

Full size image
Fig. 3: Oncoprint depicting genomic characteristics of enrolled cases, including prevalence of the most common somatic mutations (top), COSMIC v2 mutational signatures, SNV, indel and SV patterns, tumour ploidy, KRAS allelic states.
figure 3

PDAC subtypes are annotated and include the Waddell classifier, Moffitt basal-like vs classical subtype, HRDetect scores and the Menghi score identifying the tandem duplicator phenotype. The T cell inflamed signature evaluated in the cohort is also included. SNV: single nucleotide variants; indel: insertion/deletion, SV: structural variation.

Full size image
Fig. 4: Molecular correlates of KRAS allelic states and overall survival.
figure 4

A Frequency of tumour suppressors according to KRAS allelic states (balanced n = 108, minor imbalance n = 96, major imbalance n = 45, WT n = 19).Two-sided Fisher’s exact test ***p = 8.5e-06 *p = 0.02. B Frequency of Moffitt basal-like (n = 49) vs classical subtype (n = 204) in KRAS allelic states. P-value derived from two-sided Fisher’s exact test. C The median overall survival according to chemotherapy received in the intention to treat population (n = 261, mFFX n = 146, GnP n = 99, None=16), 7 patients who received ‘other’ treatments are excluded. D Median overall survival and 95% CI according to specific KRAS mutation (wt n = 19, G12D n = 114, G12R n = 36, G12V n = 77, others n = 22). E Median overall survival and 95% CI according to KRAS allelic state (major n = 45, minor n = 96, balanced n = 108, wt n = 19).

Full size image
Table 1 Demographic and baseline clinicopathologic characteristics of the of the cohort according to KRAS mutational status (n-268). Data are presented as the median (range) or number (%)
Full size table
Table 2 Proposed alternate drivers in KRAS WT cases included in the COMPASS trial (n = 19)
Full size table

Additional recurring somatic alterations were present in epigenetic genes including ARID1A (13%), KMT2C (8%), and KMD6A (7%), while loss of TGFBR2 (10%) and amplifications in MYC (10%) and GATA6(5%) were further documented. Deletions in the MTAP gene were common (n = 103, 38%%), associating with deletions in CDKN2A. Notably 26% of patients with CDKN2A homozygous deletions had intact MTAP (Fig. 3).

Mutational signatures in the cohort were dominated by age related substitution base signature 1 and 5 and SBS 8 of unknown aetiology similar to resected PDAC (Fig. 3). SBS3, association (attributed mutations >5%) with HRD-PDAC was evident in 15% (n = 40) of the cohort. This is slightly higher than the proportion of patients with HRDetect scores >0.7 (n = 34/261, 13%). We have previously reported HRDetect to be a superior biomarker of HRD-PDAC and platinum response9. In this updated analysis, HRDetecthi PDAC had an identifiable HRD genetic alteration in 53% (Supplementary Table 2).

Tumours in the COMPASS cohort were enriched for SBS17 (p = 0.05) and APOBEC mutagenesis (SBS2 and 3) (p = 0.05) compared to primaries (Supplementary Fig. 1A-B). SBS17 has been previously correlated with the prior use of 5-Fluorouracil23. We found no specific associated with epidemiological risk factors, however we confirmed that SBS17 correlated with tumour hypoxia p = 0.01 (Supplementary Fig. 1C).

The median number of SVs in the cohort was 109 with no difference according to KRAS allelic status (Fig. 3). Classifiers according to the Waddell subtyping documented unstable genotypes in 15%, stable in 12% locally rearranged genomes in 33% and scattered most commonly in 40% (Fig. 3). Diploid tumours were identified in 43% of samples, including 50% of primary tumours and 40% of metastatic sites. The basal-like subtype was found more commonly in KRASmaj (47%) compared to KRASmin (20%) or KRAS bal (9%) or KRASWT (11%), p < 0.001) (Supplementary Table 3, Fig. 4B).

Baseline Epidemiological Characteristics and subgroups

Smoking status and baseline BMI was not different for mutant specific alleles or KRAS allelic states (Table 1 and Supplementary Fig. 1D). In addition, specific smoking related mutational signatures (SBS4) were not commonly found in the genome. Pre-existing diabetes ( > 18 months prior to PDAC diagnosis) was documented in 2/19 (11%) of patients with KRASWT PDAC compared to 60/249 patients (24%) with KRASm (p = 0.26). Notably, when evaluating KRAS allelic states, 38% of patients with KRASmaj had evidence of pre-existing type II DM (p = 0.02, Supplementary Table 3) suggesting the importance of metabolic pathways in this cohort.

Genomic landscape and association with outcomes

Median OS in the ITT was 9.2 months (95% CI 8.0–10.2). In those receiving at least 1 cycle of mFOLFIRINOX, median OS was 10.6 months and in GnP 8.4 months. Median OS was 1.3 months in the 16 patients who did not receive treatment (Fig. 4C).

There were no statistically significant differences in median OS according to KRAS specific mutants although KRASG12R did have the longest median OS (12.1 months, 95% CI 6.7–15.5) and KRAS G12D the shortest OS (8.3 months, 95% CI 6.9–10.1) (Fig. 4D). Notably where tumours harboured major imbalances in mutant KRAS, median OS was 6.3 months (95% CI 4.0-8.3) compared to KRASmin (8.8 months, 95% CI 7.4–10.7), KRASbal (11.2 months, 95% CI 9.6–13.0) or KRASWT (8.5 months, 95% CI 7.0–15.0) log rank=0.001 (Fig. 4E). In addition, patients with KRASmaj had lower ORR (11%) to first line chemotherapy and had higher chance of progressive disease as best response (38%). In contrast, in tumours with balanced KRAS allelic dosage, the ORR was 32% and progression occurred in only 6% (Table 3).

Table 3 Overall response rates and overall survival of the cohort according to KRAS allelic status
Full size table

Inactivating mutations in CDKN2A, TP53 and SMAD4 did not influence OS. (Supplementary Fig. 1E–G). Although co-occurring alterations trended toward survival impact only the presence of GATA6 amplifications or mutations in PIK3CA associated with significant OS improvements on adjusted analyses (Supplementary Fig. 2). Mutational processes somewhat influenced outcomes. SBS3, which largely overlapped with HRDetect, was prognostic (Fig. 5A and B) while the presence of SBS17 in >5% mutations trended to inferior OS: median 7.4 (95% CI 6.1–10.5) vs 9.5 months (95% CI 8.7–10.6)p = 0.14 (Fig. 5C). APOBEC signatures had no influence on OS (Fig. 5D) and no impact on outcomes to platinum-based chemotherapy (Supplementary Fig. 3A). HRDetecthi continued to demonstrate a predictive role for mFOLFIRINOX (pi = 0.003) (Fig. 5B). The Waddell classifier documented superior outcomes in unstable PDAC (n = 41, p = 0.02) Of these, 54% (n = 22) had HRDetecthi PDAC which largely drove the OS benefit (Supplementary Fig. 3B).

Fig. 5: Genomic predictors and cox regression analyses associated with overall survival in advanced PDAC (n = 268).
figure 5

A Median overall survival and 95% CI according to presence of SBS3 (mutations >5% n = 40 vs <= 5% n = 228). B Overall survival according to HRDetect score and stratified by type of chemotherapy received (mFFX n = 144, GnP n = 96). C Median overall survival and 95% CI according to presence of SBS17 (mutations >5% n = 44 vs <= 5% n = 224). D Median overall survival and 95% CI according to presence of APOBEC SBS2 and SBS13 (mutations >5% n = 186 vs <= 5% n = 82). E Cox proportional hazards analysis of relevant variables, all variables were prespecified in the trial with the exception of the GRIm-S which was annotated retrospectively (n = 268). Error bars represent lower and upper 95% confidence interval. Chemo (mFFX p = 5.10e-20, GnP p = 3.84e-18); HRDetect high p = 2.10e-04; Moffitt basal-like p = 3.18e-05; GRIM high p = 5.50e-05; ECOG1 p = 1.32e-03. GnP= Gemcitabine/nab-paclitaxel; HRDetect hi is a score >0.7; GRIM= Gustave Roussy Immune Score.

Full size image

In a multivariable analysis, both patient and molecular factors influenced survival. The receipt of chemotherapy (p < 0.001), ECOG performance status 0 (p = 0.001) and HRDetecthi PDAC(p = 0.001) were associated with improved OS. A high GRIm-S (p < 0.001) and basal-like PDAC conferred inferior OS (Fig. 5E). Basal-like tumours demonstrated particularly poor OS when treated with mFFX (median OS 6.5 months, 95% CI 4.6–10). Median overall survival when treated with GnP was 7.6 months, (95% CI 6.3–10.2) noting physician choice of chemotherapy in this study (Supplementary Fig. 3C). We have previously documented the prognostic impact of systemic inflammatory scores, which did not associate with specific genotypes or transcriptomic subtypes but a high GRIm-S was more likely in males21. In this analysis we demonstrate the GRIm-S as highly prognostic on Cox proportional hazard analysis (p < 0.001). This cohort did not correlate with the basal-like subtype however did have a larger sum of tumour diameters at baseline (median 125 mm) compared to patients with a GRIm-Slo score (median 72.5 mm) p < 0.001. The ORR of these prognostic groups stratified by chemotherapy are shown in Supplementary Fig. 3D demonstrating the impact of HRD as a biomarker of mFFX response.

Locally advanced PDAC

Of 268 patients in the study, 37 (14%) had LAPC. Baseline epidemiological variables were similar between those with locally advanced and metastatic disease, with no differences between sex, age, smoking status, or history of diabetes. Patients with LAPC had a lower BMI (median 22.2 vs 24.4, p = 0.0013) and lower baseline CA19-9 (median 488.5 vs 2551, p = 0.0028) than metastatic cases and none of the LAPC cases had a high GRIm-S (Fig. 6A–C). Driver alternations were similar with no differences in SMAD4 loss between the two groups; however, major imbalances in mutant KRAS were absent in the LAPC group. Furthermore, the burden of SNVs, indels and SVs was significantly lower in LAPC and all LAPC cases demonstrated a classical RNA subtype (p = 0.0007) (Fig. 6D–G) In a univariable analysis median OS was 12.5 and 8.35 months for LAPC and metastatic cases, respectively (HR 1.5, 95% CI 1.071–2.193, p = 0.0187) (Supplementary Fig. 3E).

Fig. 6: Baseline characteristics of LAPC compared to metastatic cases enrolled on COMPASS.
figure 6

Boxplots in Panel A, B, D-F depict 1st quartile, median and 3rd quartile +/− 1.5 interquartile range. P-values derived from two-sided Wilcoxon rank sum test. A BMI in locally advanced (stage III, n = 37) compared to stage IV cases. B CA19.9 in locally advanced (stage III, n = 37) compared to stage IV cases. C Frequency of GRIm-S low (0) or high (2) in locally advanced (stage III, n = 37) compared to stage IV cases. Molecular features of LAPC (n = 37) compared to met PDAC (n = 231) D SNV count in locally advanced vs stage IV cases, E indel count and F structural variant count and G frequency of Moffitt basal-like vs classic subtype.

Full size image

Progression biopsies

Fifteen patients had progression on chemotherapy biopsies (6%) of which 5 (33%) were basal-like. The median time from baseline to progression biopsy was 6 months (2-27) and the same lesion was biopsied in 10 (67%) cases. Progression biopsies harboured the highest median number of SNV mutations (p = 0.01) and insertion/deletions (p = 0.004) and SVs (p < 0.001) compared to baseline metastases or primaries in the COMPASS dataset. In addition SV burden was highest in progression biopsies (median 157) compared to baseline COMPASS biopsy (median 109) and resected primaries (median 95) p = 0.01 suggesting increasing genomic instability as PDAC progresses (Supplementary Fig. 4A–D).

Non-evaluable patient cohort

Patients who did not receive any chemotherapy (n = 16) or those who died without a scan (non-evaluable n = 29) are a cohort of patients with very short median survival (7.1 weeks). These patients had a larger median sum of tumour diameter (115 mm) at diagnosis compared to those patients who had at least one RECIST response documented (73 mm) p < 0.001. In this cohort of 45 cases, 36% had a GRIm-Shi vs 12% of other cases (p = 0.005), there was however no difference in the prevalence of basal-like cases (25 vs 18% p = 0.30). This suggests an additional cohort of patients with aggressive disease of different aetiology, or perhaps very late presentation resulting in rapid clinical decline.

Second line treatments and long-term survivors

Of the 249 patients who started chemotherapy, 47% of patients received a second line approach. A matched targeted therapeutic approach of patients enrolled on the COMPASS trial was undertaken in 10% (n = 25 of 249) (Supplementary table 4). Of these 5 patients had KRAS-WT PDAC including one with HER2 point mutation (p.L755S, OncoKB level 3B).

Additional genotypes were consistent with mismatch repair deficiency (n = 1) and homologous recombination deficiency (HRD) (n = 12, KRASm in 11/12). Other genomic alterations targeted, included KRASG12C mutation (n = 1), amplification in CDK4/6 (n = 3), ERBB3 point mutation (n = 1) BRAF p.486_491del (n = 1) and a BRAF fusion. There was no difference in median OS in patients receiving matched versus unmatched second-line approaches (median OS 14.7 versus 12.8 months, HR 0.79 95% CI 0.49–1.27, p = 0.32) (Supplementary Fig. 3F).

Of the 10 patients (4%) still alive at the end of the study, one patient lost to follow-up is excluded (median OS 55.5 months, range 19–86 months), one had LAPC and four were HRDetecthi (3 gBRCA2, 1 gBRCA1 one unknown.) An additional patient had a germline pathogenic variant in MSH6 with a tumour consistent with MSI-H PDAC on second line immunotherapy and one patient with KRASWT PDAC (BRAF in frame deletion) had yet to progress on first line mFOLFIRINOX.

Immunohistochemical analysis of CD8+ lymphocytes

Immunohistochemical analysis of CD8+ lymphocytes was available in 149 cases including 49 primaries and 100 metastases. The median number of positive cells per mm2 was 131 (4-3634) with large variability. There was a moderate correlation between CD8 expression and IHC (Spearman rho 0.4,p < 0.001) (Fig. 7A).

Fig. 7: Analyses of CD8 expression using immunohistochemistry (IHC, n = 149) or RNA-seq (n = 253) across cohorts included in COMPASS.
figure 7

Boxplots in Panel B–F depict 1st quartile, median and 3rd quartile +/− 1.5 interquartile range. P-values derived from two-sided statistical tests. A Spearman correlation of CD8 IHC vs RNA-seq expression. B Comparison of CD8 IHC expression in primary biopsy sites vs metastatic sites. C CD8 expression according to specific KRAS mutation Left) analysis by IHC Right) analysis by RNA-seq. D CD8 expression (RNA-seq) in HRDetect hi vs lo cases. E CD8 expression according to Moffitt subtype (basal-like vs classical) in liver metastases Left) analysis by IHC, Right) analysis by RNA-seq. F CD8 expression by IHC in liver metastases comparing GRIm-S lo vs hi.

Full size image

Although there was no difference in CD8 counts according to site of biopsy, liver metastases trend to lower CD8 (n = 94) compared to primaries (n = 158) (p = 0.07, Fig. 7B). Across specific mutant KRAS alleles and allelic states the median number of CD8+ cells was similar. KRASQ61X had numerically higher median CD8 count although numbers were small (Fig. 7C). HRDetecthi cases documented similar CD8 counts compared with HRDetectlo (p = 0.85, Fig. 7D). This did not change when separating primary biopsies and liver biopsies. Basal-like liver metastases did however have trend towards a higher CD8 count compared to classical (p = 0.06, Fig. 7E) and notably liver metastases from patients with GRIm-Shi patients had fewer CD8 (p = 0.002, Fig. 7F). Data for full cohort (primaries and metastases) in Supplementary Fig. 5.

In line with these findings, an analysis of a gene signatures predictive of immune checkpoint inhibitor response showed that basal-like liver metastases appeared more inflamed than classical (Fig. 8A–E) while there was no difference in liver metastases from other subgroups. Diverse immune gene sets were further enriched in basal-like liver metastases compared to classical. Aside from a higher expression of CD8 T- cells, basal-like liver metastases were enriched in NK cells markers (KLRF1/KLRC1). In parallel, expression of markers presenting inhibition of antigen presenting cells and T cells were also enriched, (Fig. 8F, Supplementary Table 1).

Fig. 8: Immune signatures and prognostic subgroups in advanced PDAC.
figure 8

A–E T-cell Inflamed signature (RNA) in liver metastases according to A KRAS mutation (wt n = 19, G12D n = 114, G12R n = 36, G12Vn = 77, Q61 n = 16) (B) KRAS allelic states (wt n = 19, balanced n = 108, minor n = 96, major n = 45) C HRDetect hi (n = 227) vs lo (n = 34). D Moffitt subtype (basal-like n = 49, classical n = 204) E GRIm-S hi (n = 225) vs lo (n = 43) Boxplots depict 1st quartile, median and 3rd quartile +/− 1.5 interquartile range. P-values derived from two-sided statistical tests. F Expression of varying immune gene sets according to Moffitt subtype in liver metastases (basal-like n = 45, classical n = 122, see Supplementary Table 1 for genes).

Full size image

Discussion

Precision oncology in PDAC has been limited to HRD-PDAC and small numbers within the KRAS WT cohort with rare oncogenic fusions or mutations in the BRAF gene. As we begin to integrate KRAS inhibitors into clinical practice an understanding of the reliance on MAPK signaling and the influence of KRAS on the tumor microenvironment will remain crucial for combination strategies. In addition, it is likely that chemotherapy will continue to be required and biomarkers selecting for the regimens are currently lacking.

We consistently demonstrate the importance of the basal-like RNA subtype as a poor prognostic group. Although enriched in tumours with increased KRAS dosage, only the Moffitt signature was prognostic on Cox proportional hazard regression analysis. In addition, we demonstrate by IHC a trend to higher CD8 infiltration in basal-like liver metastases, along with transcriptional metrics of a more inflamed TME. Notably, in comparing immune gene set expression in liver metastases from basal-like and classical PDAC, we see a striking imbalance in basal-like tumours. Although enriched by RNA expression in CD8 and NK cell markers compared to classical liver metastases, we also see higher expression of co-inhibitory molecules of both antigen presenting cells and T cells, indicative of the mounting of an immunosuppressive response. A knowledge of the TME in distinct subsets will be important for future trial design as recent pre-clinical data suggest that KRAS inhibitors may have deeper responses in basal-like PDAC24 which is in line with our findings and the observation that KRAS inhibitors may require functional CD8 T cells to maximize response25.

Furthermore, patients with evidence of systemic inflammation, exhibit fewer CD8+ cells by IHC in liver metastases. The GRIm-S has been shown to predict response to immune checkpoint inhibitor. Our data suggest that this score associates with a ‘cold’ TME, and this cohort of patients have an inferior outcome as previously described21. Notably, higher systemic inflammation may be more prevalent in patients with a higher volume of disease at baseline and, when considering patients with non-evaluable disease, the sum of tumor diameter at baseline was the single variable that predicted rapid progression. This potential relationship between larger disease volume and tumor progression with a more immunosuppressive TME26,27 in the context of different molecular subtypes, will be further evaluated in the PASS-01 trial28.

Genomic data associating with improved outcomes included HRDetecthi PDAC. In this updated analysis, we observe HRDetecthi scores in 13%, with an unknown etiology in 46% of these. When considering unstable genomes or those with a structural variant burden >200, only those with high HRDetect scores had improved survival, underscoring heterogeneity within this subgroup. We did not see differences in CD8 infiltration or inflamed scores according to HRDetect and notably recent data suggest the importance of CTLA-4 inhibitors rather than PD-1 inhibitors in this cohort29.

Specific mutant KRAS alleles did not associate with differences in median OS although those patients with KRASG12R had the longest OS in a univariable analysis. This finding has been reported recently in a large series by Yousef et al where patients with stage IV PDAC and KRASG12R had a median OS of 25 months30. KRAS WT PDAC is known to be prognostic in PDAC; 7% of patients in our trial lacked a KRAS mutation in keeping with the literature however OS was not significantly longer perhaps owing to small numbers receiving a targeted approach31,32. We do further note that high GRIm-S was evident in 6 (32%) KRASWT cases suggesting that KRAS WT PDAC are a heterogenous group. When evaluating KRAS allelic states, balanced number of mutant versus wildtype alleles demonstrated longest median OS. Although major imbalances in mutant KRAS were not prognostic on MVA, we do note a higher frequency of type II DM > 18 months in this cohort suggesting the importance of metabolic pathways and glucose metabolism in driving the dosage of KRAS.

Importantly, SMAD4 loss did not associate with locally advanced PDAC which represented 14% of our cohort. LAPC cases did not have major imbalances in KRAS or basal-like PDAC. Furthermore, all LAPC had low scores of systemic inflammation. These findings suggest selection for metastatic disease in basal-like PDAC and in the presence of inflammatory states.

Unsurprisingly, despite low numbers, we observe a higher burden of mutations/indels and structural variation in progression biopsies compared to baseline COMPASS biopsies and primary resections, highlighting progressive cell cycling and genomic instability in progression of PDAC. The advanced PDAC cohort was further enriched in signature 17 and APOPEC signatures underscoring the potential selection for specific mutational processes.

In this study, 47% of patients received second line therapy including 10% of the cohort who received a matched approach. This highlights a need to invest in biomarker driven strategies in the first line setting since less than half receive second-line treatment and given the unique subgroups highlighted in this study.

The COMPASS trial had a number of limitations. This was a non-randomized study, and chemotherapy was administered at the discretion of the investigator. Since drug dosing was not prescriptive, true progression free survival was not calculated. It is notable however that median OS for mFOLFIRINOX and GnP are similar to that of the PRODIGE2414 and MPACT15 trials, respectively. The PASS-01 trial has recently reported early results and will represent a cohort to establish biomarkers from a randomized approach28. Beyond the IHC based quantification of CD8+ cells in the entire tumoral area, RNA-based immunophenotyping on COMPASS is predominantly from tumor enriched material, representing immune cells in the immediate vicinity of malignant epithelia. It is notable that both levels of analyses yield similar results, suggesting a relatively robust association of T cell inflamed phenotype with basal-like liver metastases. Not all analyses were prespecified and thus considered hypothesis generating. The PASS-01 trial will evaluate formalin-fixed paraffin-embedded slides from cohorts using CODEX immunophenotyping, which will be additionally informative.

In conclusion, the COMPASS trial demonstrates the importance of patient, molecular and systemic factors in determining outcome in PDAC. KRAS G12R and low levels of RAS signaling may associate with improved outcomes. Importantly the basal-like subtype and patients with evidence of systemic inflammation represent cohorts with inferior OS selecting for metastatic disease. Distinct TMEs warrant specific strategies for these subgroups that could include targeting KRAS with immunotherapy strategies.

Methods

Patient population

The COMPASS trial (NCT02750657) was a prospective study designed to explore biomarkers of chemotherapy response and was approved by the University Health Network Research Ethics Board (REB:15-9596).

Patients with unresectable advanced PDAC were included. Patients had to have a lesion suitable for biopsy prior to commencing planned chemotherapy with mFOLFIRINOX or GnP as first line treatment for advanced disease. All patients were ECOG 0-1 and the choice of combination chemotherapy backbone was at the discretion of the treating medical oncologist. Dosing modification was as per institutional guidelines and was not prescriptive in this study. Patients were included after histological confirmation of PDAC and only if sufficient tissue for attempted WGS. Response to therapy was assessed using CT scans and measured using RECIST 1.1. All patients were enrolled from December 2015 until June 2020 with follow-up censored on October 1, 2023. Patients on this study were enrolled at the Princess Margaret Cancer Centre (Toronto, Ontario, Canada), McGill University Health Centre (MUHC, Montreal, Quebec, Canada), Kingston General Hospital (Kingston, Ontario, Canada), the Centre hospitalier de I’Universite de Montreal (CHUM) and Nova Scotia Health authority. Written informed consent was obtained from all patients prior to enrolment on the clinical trial. This included consent to publish variables presented.

Epidemiological data were documented prospectively and included duration of diabetes, smoking history (Ever/Never) and BMI at first presentation. The Gustave Roussy Immune Score (GRIm-S) was calculated retrospectively in all patients21. This systemic inflammatory score combines neutrophil to lymphocyte ratio (NLR > 6 = 1 point) albumin ( < 35 = 1 point) and LDH ( > upper limit normal = 1 point) and dichotomises patients into GRIm-S high (hi) and low (lo). A score of ≥2 is designated GRIm-Shi. The COMPASS trial was approved by the Institutional Review Boards of the participating sites (University Health Network, Toronto, Ontario, Canada; MUHC Centre for Applied Ethics, Montreal, Quebec, Canada; and Queen’s University Health Sciences and Affiliated Teaching Hospitals Research Ethics Board, Kingston, Ontario, Canada); and was conducted in accordance with the Declaration of Helsinki.

Whole genome sequencing

Frozen biospecimens underwent laser capture microdissection (LCM) for tumor enrichment.

Sequenced reads were processed as previously described19,33,34. Briefly, sequencing reads were aligned to hg38 using bwa mem (version 0.7.17)51. Duplicate reads were marked with Picard version 2.21.4 (https://github.com/broadinstitute/picard). Germline variant calling was performed using GATK version 4.1.235, SNVs were identified as the intersection of calls by Strelka2 (version 2.9.10)36 and MuTect2 (version 4.1.2)37. Indels were identified as the overlap between two of five callers - Strelka236, MuTect237, manta (version 1.6.0)52, SVaBA (v134)37, and DELLY2 (version 0.8.1)37. Copy number segments, tumor cellularity and ploidy (number of sets of chromosomes) were called by CELLULOID33. Additional details are found in Supplementary Data 2. Somatic structural rearrangements were called as the consensus from two out of three variant callers - SVaBA, DELLY2, manta. Tumours with ploidy >2.3 were classified as polyploid. Recurrent somatic alterations were identified using dndscv38 and GISTIC39. Tumours with an SV burden> 200 are considered unstable. Substitution base signatures (SBS) are annotated as per Alexandrov40. HRDetect scores followed the same algorithm as previously reported by Davies et al.41 with a score of >0.7 deemed HRDetecthi. All germline variants detected were reviewed by a genetic counsellor (S.H.) and documented as pathogenic, likely pathogenic, variant of unknown significance, or benign. Where KRASm were present, allelic states were documented, major imbalances were defined as mutant to wild-type copy number ≥3, minor imbalances 1-2 copy differences and balanced. Molecular tumour board provided recommendations for the treating physician for second line therapy approaches and where implemented this was deemed a matched approach.

RNA sequencing

RNA-sequencing and analysis was performed at the Ontario Institute for Cancer Research as described previously10. Briefly, reads were aligned to the human reference genome (hg38) and transcriptome (Ensembl v.100) using STAR v2.7.4a42. Gene expression was calculated in transcripts of exon per million reads mapped using stringtie v2.0.643. Putative fusions were identified using STAR-Fusion v1.9.0 53. A modified Moffitt classification (basal-like vs classical) was applied to each sample with sufficient RNA for analysis as previously described44,45, where samples with high expression of Moffitt basal-like genes and absence of Moffitt classical genes were classified as basal-like and other samples as classical. Hypoxia gene expression signatures were calculated as the difference between median normalized expression of hypoxia high genes and hypoxia low genes46.

Immune signatures and IHC

Published immune signatures and gene sets were evaluated in the cohort including a T cell inflamed signature shown to predict responsiveness to immune-checkpoint inhibitors in other solid tumours47,48,49 (Supplementary Table 1). CD8 immunohistochemical analysis (IHC) was performed in a subset of cases using the CD8-4B11-L-CE antibody (NovoCastra, Leica Biosystems), at 1:25 dilution. Tumoral regions were reviewed and annotated by a pathologist (RG) and digital image analysis (QuPath) enumerated the number of positive cells per µm2 of tumor area. Samples were classified as CD8hi vs CD8lo using the median no./mm2. Spearman correlation evaluated IHC and RNA expression of CD8.

Statistical analysis

Baseline clinical characteristics were compared across cohorts using Wilcoxon rank sum or Kruskal-Wallis test. Categorical characteristics were compared using Fisher’s exact test. The Kaplan-Meier and Cox proportional hazards models were used to estimate overall survival (OS) selecting clinical and genomic variables with an established association with OS. OS was calculated from the date of enrolment on the COMPASS trial. Exploratory associations between driver genes, genomic features and OS were tested in univariable Cox regression models. For comparative purposes we have referenced an early-stage resectable cohort of PDAC which have undergone WGTS (n = 177) with additional annotated clinical data as previously published34. Statistical significance was set at p < 0.05. All statical analyses were performed in R version 3.6.2 53.

Statistics and reproducibility

No statistical tests were used to predetermine sample sizes. Patients were excluded only when there is insufficient tumour for sequencing. Exclusions have been recorded in the consort diagram. Patient treatment was not randomized and investigators were not blinded to study findings during treatment selection process.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.