Targeted Transcriptome and KRAS Mutation Analysis Improve the Diagnostic Performance of EUS-FNA Biopsies in Pancreatic Cancer

Pancreatic ductal adenocarcinoma (PDAC) is the 7th leading cause of cancer-related death worldwide (1). Most patients present with locally advanced or metastatic disease, with fewer than 20% presenting with lesions amenable to potentially curative surgery (2). The mainstay of treatment for patients with locally advanced and metastatic disease is chemotherapy, and the most commonly used regimens of gemcitabine/nab-paclitaxel and FOLFIRINOX may prolong survival for up to 11 months (3, 4). However, despite incremental improvements in recent years, the prognosis of PDAC remains dire with a 5-year survival rate of just 9% (5).

Multiple factors contribute to the dismal prognosis associated with PDAC. Many patients present with nonspecific symptoms, and current imaging modalities and biomarkers such as carbohydrate antigen 19.9 (CA19.9) may not detect early stage disease or clearly differentiate PDAC from other causes of solid pancreatic masses, which may include other malignancies [e.g., pancreatic neuroendocrine tumors (pNET), lymphomas, and metastases] and benign conditions (e.g., autoimmune pancreatitis and pseudotumoral lesions; ref. 6). Other factors contributing to poor outcomes in PDAC include the propensity for early distant metastasis, a complex tumor microenvironment characterized by dense stromal desmoplasia and immune dysregulation, and inherent resistance to standard treatments such as chemotherapy (7).

Screening programs have shown some benefit in applying early imaging or targeted molecular screening in high-risk populations although observed benefits remain limited largely to those with high familial risk, comprising only a small minority of all patients with PDAC (8). There have been conflicting reports on the benefits of reducing the time between symptom onset and PDAC diagnosis. For instance, some studies show poorer prognosis and higher risk of unanticipated metastasis with an increasing interval from symptom onset to diagnosis (9–12), while others demonstrate no significant prognostic impact from diagnostic delay (13, 14). However, improving the accuracy of current diagnostic procedures and reducing the time from first presentation to diagnosis and treatment remains appealing.

Endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) is considered the gold standard diagnostic method for the biopsy of suspicious pancreatic masses, and is a widely available procedure with low procedural morbidity and mortality (15). A recent meta-analysis reported pooled sensitivity for the diagnosis of PDAC of 85%, with high specificity of 98% (16). However, approximately 15% of patients fail to achieve a tissue diagnosis with their first attempt at biopsy and may require further diagnostic procedures. EUS-FNA sensitivity has also been reported to be lower in the setting of chronic pancreatitis, an established risk factor for PDAC (17). The accuracy of EUS-FNA as a diagnostic tool in PDAC can be improved with increasing operator experience, technical aspects of needle selection and rinsing, increasing the number of needle passes, and optimizing processing of the biopsy specimen (18).

In the era of precision medicine, adequate tissue samples for molecular testing are increasingly in demand, adding to the need for high-quality diagnostic biopsies to allow for genomic assessment. The genomic changes contributing to the development of PDAC have been well characterized. Activating mutations in the KRAS gene can be identified in approximately 90% of patients, and represent critical early events in the development and maintenance of cancers arising from premalignant pancreatic intraepithelial neoplasia lesions (19, 20). KRAS mutations are typically accompanied by the progressive loss of a number of tumor suppression genes, most commonly CDKN2A, TP53, and SMAD4, although significant genomic heterogeneity is observed (21). Metastasis has been demonstrated to be a relatively late phenomena in the genetic evolution of PDAC, occurring up to a decade or more after initial malignant changes appear (22). This observation raises hope that improved screening and diagnostic methods could reduce PDAC mortality by detecting cancers earlier in the natural disease course, when curative therapy remains possible.

Genomic characterization studies have provided valuable insights into the mutational landscape of PDAC, but have largely relied on sourcing tumor tissue from formalin-fixed, paraffin-embedded (FFPE) specimens, most commonly from surgically resected tissues, and rarely utilize EUS-FNA biopsies which may have lower tumor cellularity (23–27). As most patients have unresectable disease at diagnosis, EUS-FNA biopsies represent an appealing source of genetic material in PDAC (2). However, their use in this setting has previously been limited because of concerns regarding test sensitivity, tissue yield, and sample purity (16, 28).

Here, we describe the utility of EUS-FNA as a source of tissue for genetic material to integrate targeted genomic and global transcriptome profiling in all stages of PDAC. Using this approach, we report the identification of a genetic signature which differentiates PDAC from other malignant and benign causes of solid pancreatic masses, thus providing improved diagnostic accuracy over current strategies.

Pancreatic tissue samples were sourced from the Victorian Pancreatic Cancer Biobank (VPCB, HREC/15/MonH/117), which currently stores biospecimens from seven major tertiary centers in Victoria, Australia. For EUS-FNA biopsies, patients provide written informed consent for an additional needle pass at the time of their standard-of-care biopsy, which is snap frozen and stored in the VPCB. All biopsies for this study were obtained using either a 20-guage or a 22-guage ProCore Fine Needle Biopsy needle for both the standard-of-care cytology and the additional biopsy for this study. Relevant clinical data were extracted from retrospective review of medical records, and stored in a deidentified manner. Clinical diagnosis was confirmed by review of pathology reports and consensus clinical opinion from tertiary centre multidisciplinary meeting records (typically attended by surgeons, medical and radiation oncologists, radiologists, and pathologists). This study was performed in accordance with the principles of the Declaration of Helsinki, after approval by the Monash Health Human Research Ethics Committee (HREC/15/MonH/117 and LNR/17/MonH/335). Informed, written signed consent was obtained from all patients prior to initiating study procedures.

Tests for diagnostic accuracy were assessed by constructing 2 × 2 contingency tables using GraphPad Prism v8.0, with true positive cytology defined as those with confirmed malignancy or suspicious cytology. Cytology reported as “scant atypical cells” without a definite diagnosis, “indeterminate,” and “nondiagnostic” were considered false-negative results for patients with PDAC. Fisher exact test was used to calculate P values and statistical significance, and the Wilson–Brown method was used to calculate 95% confidence intervals (CI). For descriptive statistics of the yield of RNA and DNA, data are presented as the mean ± SEM.

RNA and DNA were simultaneously extracted from snap-frozen pancreatic biopsies following the manufacturer's protocol (Qiagen AllPrep DNA/RNA Universal Kit). Quantity of gDNA and RNA was assessed using the Nanodrop spectrophotometer (Thermo Fisher Scientific) and Qubit Fluorometer (Thermo Fisher Scientific), and quality assessed using Bioanalyser and TapeStation systems (Agilent). KRAS testing was performed on gDNA using the KRAS XL StripAssay (ViennaLab Diagnostics GmBH). For RNA sequencing (RNA-seq), we applied a minimum RNA integrity number (RIN) threshold of 4, although samples with RIN <4 were considered if there was at least 100 ng RNA available and clear 18s/28s peaks visible.

For the test cohort, RNA was sequenced on the Ion Torrent Proton Sequencer using the Ion Ampliseq Transcriptome Human Gene Expression Kit for library preparation, with amplified samples ligated, purified, and quantified by qPCR before being pooled for sequencing. Sequence reads were aligned to GRCh38 using STAR aligner. Read counts were calculated using Htseq.

Read counts for each gene in each sample were count-per-million normalized, posterior count of 1 added, and log₂ transformed. Genes with arithmetic average expression across all samples below the 90% percentile were removed. Genes were then ranked according to their ability to distinguish PDAC from normal pancreas and pancreatitis by constructing a ROC curve for each gene and calculating the AUC metric. A total of 20 genes with the highest AUC were selected. A total of 30 housekeeping genes were selected by ranking genes based on their coefficient of variation in the RNA-seq data across all samples and selecting the genes with the lowest coefficient of variation.

Raw affymetrix oligonucleotide microarray data (datasets E-MEXP-1121/E-MEXP-950, GSE15471, and GSE28735) were processed using the oligo package for R (29) using updated probeset definitions from Custom CDF v23 downloaded from http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/23.0.0/ensg.asp (30) and RMA-normalized as implemented in the oligo package (29). Raw Illumina oligonucleotide microarray data (datasets GSE101462, GSE101448) were vsn-normalized using the vsn package for R (31).

For gene expression derived from RNA-seq or microarray cohorts, gene expression scoring was calculated using the single-sample gene set enrichment analysis (ssGSEA) method, as implemented by the GSVA package on R (32, 33).

The diagnostic signature was assessed in the validation cohort using a custom-designed NanoString Custom CodeSet. RNA (50 ng) from each sample was added to a Master Mix containing the Hybridization Buffer and Reporter CodeSet, then underwent hybridization at 65°C for 16 hours before a ramp down to 4°C. Samples were immediately made up to 35 μL using RNAse free water, loaded into nCounter Sprint Cartridges and run using the SPRINT profiler (NanoString). Gene expression normalization was performed by dividing the expression of each signature gene by the geometric mean of the 30 housekeeping genes for each sample. Gene expression values were then z-transformed using the mean and SD of each gene in the NanoString cohort. To calculate the summarized gene expression score for each sample, z-transformed gene expression values were summarized into a single value for each sample using simple addition.

KRAS mutations in PDAC tissues were verified in duplicate by digital droplet PCR (ddPCR) with the inclusion of positive, negative, and no-template controls, following manufacturer protocols (KRAS G12/G13 or Q61H Screening Kit, Bio-Rad). Droplets (15,000–20,000 per well) generated using the Q200X droplet generator were transferred to a 96-well PCR plate, heat-sealed, and subjected to thermocycling in a C1000 touch thermal cycler (Bio-Rad) under the following cycling conditions: 95°C for 10 minutes, 40 cycles at 94°C for 30 seconds, and subsequently 55°C for 1 minute, then followed by an enzyme deactivation step through incubation for 10 minutes at 98°C. Amplified droplets were detected using a QX200 droplet reader (Bio-Rad Laboratories) with two fluorescent detectors (FAM and HEX. The determination of the number of mutation copies, ratio, and fractional abundance of the samples was adjusted by the Quanta-Soft software (Bio-Rad Laboratories) to fit a Poisson distribution model with a 95% confidence interval (CI). A minimum of three positive droplets across the two wells was required for a positive result for detection of rare events. The ratio was calculated as the number of copies per microliter of mutant allele, divided by copies per microliter of wild-type allele. The fractional abundance of mutant allele was measures by dividing the number of copies per microliter of mutant allele by the total copies per microliter of wild-type allele plus mutant allele.

We initially sought to determine the diagnostic accuracy of EUS-FNA by performing a retrospective chart review on 308 sequential pancreatic biopsies, where tissue was donated to the VPCB between 2016 and 2019. Biopsies were included whether EUS-FNA was performed for a solid pancreatic mass and/or imaging appearances suspicious for malignancy. Biopsies were excluded whether there was only a surgical biopsy, predominantly cystic lesion, ampullary or bile duct pathology, or inadequate clinicopathologic information for accurate classification (Fig. 1).

The 308 pancreatic biopsies included 218 with a final clinical diagnosis of PDAC, 51 with benign pathology, 21 with pNETs, and 18 with other tumors, including 12 metastases from other primary sites (three renal cell carcinomas, two lung cancers, two gastric cancers, and 1 each of melanoma, colorectal, ovarian, breast, and prostate cancer), 1 pancreatic sarcoma, 3 lymphomas and 2 patients with solid pseudopapillary tumors. The classification of a clinical diagnosis of PDAC was based on clinician consensus, and included patients who required multiple biopsies to establish a diagnosis, as well as some patients who never had a confirmed tissue diagnosis but were considered to have sufficient clinical features to warrant clinical diagnosis and subsequent treatment for PDAC.

Cytology confirmed a diagnosis of PDAC in 137 biopsies, and was suspicious or highly suspicious in a further 29. A total of 11 biopsies were reported as atypical without definite evidence of malignancy, and 41 were either nondiagnostic or inadequate. There was one false-positive result for PDAC, with initial cytology reported as adenocarcinoma, but at later review including clinical history and extensive IHC staining, the diagnosis was changed to pNET. Therefore, the sensitivity of EUS-FNA cytology in diagnosing solid pancreatic masses in our cohort was 78.9% (95% CI, 73.5%–83.5%) with a specificity of 98.1% (95% CI, 89.9%–99.9%), positive predictive value of 99.5% (95% CI, 97.3%–100.0%), negative predictive value of 48.6% (95% CI, 39.2–55.0%; Supplementary Table S1). In our series, 18.2% of patients underwent more than one procedure to establish a diagnosis (range, 1–3).

CA 19.9 is a widely used, well validated biomarker for PDAC, although inadequate sensitivity and specificity limit its utility in the diagnostic setting (34). We examined the diagnostic accuracy of CA 19.9 in our cohort when these data were available (n = 166 patients). As expected, CA 19.9 displayed significant variability across patient samples (serum levels ranging from <1 to >640,000 kU/L) and was a poor diagnostic biomarker for PDAC, with sensitivity of 71.2% (95% CI, 63.4%–80%%) and specificity of 63.6% (95% CI, 43.0%–80.3%).

We have previously established a method to optimize the processing of snap-frozen EUS-FNA biopsies to maximize yield of genetic material (35). When an EUS-FNA biopsy was available, we extracted DNA and RNA simultaneously and performed KRAS mutation testing. There was significant variability in the yield and quality of genomic material extracted from these unselected biopsies, particularly with regards to RNA. The average yield of RNA was 2,839 ± 341.6 ng, and DNA 2,427 ± 289.4 ng. The mean RIN was 3.4 ± 0.15, while the quality of DNA was generally higher, with a mean DIN of 7.1 ± 0.13. Only one of 175 DNA samples (0.006%) failed quality control testing for KRAS mutation analysis, due to a very low yield of genetic material from a paucicellular specimen. The KRAS XL StripAssay (ViennaLab Diagnostics GmBH) was selected as a commercially validated, relatively cost-effective method of testing with quick turnaround time, which can detect mutations in specimens comprising 1%–5% mutation positive cells.

KRAS mutation analysis was available for 174 PDAC samples and 23 benign tissues, and had diagnostic sensitivity for PDAC of 86.8% (95% CI, 80.1%–91.0%), and specificity of 95.7% (95% CI, 79.0%–99.8%). The single positive result in a benign biopsy was an equivocal KRAS G12V mutation, present at the very lower limit of detection of the assay. On subsequent follow up, a pancreatico-duodenectomy confirmed no invasive malignancy but focal areas of low-grade pancreatic intraepithelial neoplasia (PanIN-1B), malignant precursor lesions of which 10%–30% may harbor pathogenic KRAS mutations (36).

Notably, in the group of 51 patients with PDAC with only atypical features or nondiagnostic cytology, a pathologic KRAS mutation could be detected in 34 of 42 (81%) available samples, suggesting the presence of malignant cells within the biopsy despite negative cytology. We hypothesized that deeper genomic examination of these biopsies may reveal markers of PDAC even in low yield, cytologically uncertain clinical specimens.

To determine the gene expression profile of EUS-FNA biopsies, we performed whole human transcriptome sequencing on 96 patients with PDAC, and 38 non-PDAC controls. The clinical features of the 134 patients in this initial RNA-seq cohort are summarized in Supplementary Table S2. We used AmpliSeq, a whole transcriptome sequencing approach which provides targeted amplification of greater than 20,000 RNA targets using a single primer pool, and allows for differential gene expression profiling using small starting RNA quantities (10 ng) and requires fewer total sequencing reads when compared with other RNA-seq approaches (37).

The non-PDAC control biopsies included normal pancreatic tissues obtained at surgery (n = 14), as well as EUS-FNA biopsies from patients with autoimmune pancreatitis (n = 10), nonspecific pancreatitis (n = 6), and pNET (n = 8). The control EUS-FNA biopsies were selected to include causes of suspicious pancreatic inflammation or solid masses on imaging which would generally warrant an urgent biopsy to exclude malignancy.

In our cohort, EUS-FNA PDAC biopsies provided a feasible source of genetic material for molecular analysis, and displayed a distinct gene expression profile when compared with non-PDAC controls. We noted 2 patients initially diagnosed with pancreatitis (with benign cytology and no clinical diagnosis of malignancy) were found to have outlying gene expression profiles in our test set, clustering with PDAC samples (Fig. 2A). Pathologic KRAS G12D mutations were detected in both apparently benign biopsies. On retrospective chart review, both patients ultimately developed progressive symptoms after observation, and were ultimately diagnosed with PDAC at a later date.

The presence of pathogenic KRAS mutations in combination with a distinct gene expression profile in nondiagnostic biopsies suggests that transcriptomic profiling of patients with a clinical suspicion of malignancy may be more sensitive than standard cytology in distinguishing PDAC from other causes of pancreatic masses. However, the time and costs associated with performing whole transcriptome profiling mean that it is not currently feasible for routine clinical practice. Similarly, to meet our aims of identifying a pragmatic, clinically relevant and time-efficient test, we elected not to incorporate additional somatic mutation testing to our algorithm beyond the near-ubiquitous KRAS. We subsequently sought to develop a targeted gene signature which could be feasibly and rapidly performed on an initial diagnostic EUS-FNA biopsy.

Each gene was ordered in their ability to distinguish PDAC from non-PDAC controls and the top 20 genes within the top 20% of abundance were selected to create a diagnostic signature for PDAC for further analysis and validation (Fig. 2B and C). Interestingly, using this agnostic strategy, we identified a number of genes that have previously been associated with PDAC either from a potentially diagnostic or prognostic perspective, as well as a number of genes that have not previously been associated specifically with PDAC. The diagnostic performance of each individual gene was assessed (Fig. 2C) and combined to generate a 20-gene signature score (Fig. 2D and E). A ROC curve was generated to assess the diagnostic performance of the gene signature in our test cohort, and demonstrated an excellent predictive AUC of 98% (Fig. 2F). Although selecting a larger number of genes would have performed similarly (Supplementary Fig. S1), more genes were not selected because of the incremental cost of testing for the expression of more genes without contributing to additional predictive performance. Selecting a lower number of genes would have also performed similarly (Supplementary Fig. S1); however, we were concerned that picking fewer genes would create a gene expression score with a lower signal-to-noise ratio under the possible scenario that one of the highly ranked genes was false discovery and behave like random noise when validating in an independent cohort.

We next applied our diagnostic signature to five publicly available cohorts of patients containing both PDAC and non-malignant controls (either benign specimens or microdissected adjacent normal tissue): E-MEXP-1121/E-MEXP-950, GSE101462, GSE15471, GSE28735, and GSE101448 (Supplementary Table S3). We generated heat maps and ROC curves to assess the diagnostic performance of our gene signature in each cohort (Fig. 3). The predictive AUC in the respective external validation cohorts was 82%, 98%, 89%, 94%, and 96%, although we noted that the cohort with the lowest diagnostic signature performance (E-MEXP-1121/E-MEXP-950) did not include all of our diagnostic genes in their dataset.

In addition to the 20-gene signature, we also used our RNA-seq data to select a number of genes that were enriched in pancreatitis, pNETs and/or genes that were overexpressed in normal pancreas, including a number of genes with potential prognostic potential, but lacked statistical significance in our RNA-seq data and were added to our NanoString CodeSet for exploratory purposes (data not shown). The final selected genes were then used to create a custom NanoString CodeSet with 165 target mRNA and 35 housekeeping genes, including our signature genes, for testing in an independent patient cohort. Analyses of the wider utility of the full gene set are planned in ongoing studies. The NanoString system utilizes a simple workflow which accommodates input of relatively low RNA quality and quantity (25 ng) and allows complementary capture and reporter probes for all mRNA targets of interest to be mixed with RNA in a single hybridization reaction with no need for library preparation, with subsequent digital counting of color-tagged codes for each mRNA target (38). Results are available within 24 hours, an attractive feature if applied in the clinical setting as data could be used in real time alongside standard cytology to aid in the interpretation of biopsy results.

Using the NanoString Custom CodeSet, we next tested our diagnostic gene signature in an independent local cohort of a further 60 EUS-FNA patient biopsies. The validation cohort consisted of 24 patients with cytologically confirmed PDAC, 20 patients with indeterminate or nondiagnostic cytology, 10 patients with clinically and cytologically benign pancreatic disease, and 6 patients with cytologically confirmed pNETs (Supplementary Table S4). Notably, one of the patients with a final clinical diagnosis of pNET was initially erroneously diagnosed with adenocarcinoma on cytologic assessment. The diagnostic gene signature profile for this sample scored lowly, consistent with other pNETs, and lower than the PDAC biopsies. Of the 20 cytologically nondiagnostic samples, the final clinical diagnoses consisted of two further pNETs, one benign pancreatitis, and a further 17 PDACs. The mean RIN was 4.8 (range, 2.3–8.7) in the validation cohort.

The performance of the diagnostic gene signature in this cohort is shown in Fig. 4. Among the 20 genes, three were noted to perform poorly in the expanded cohort and we elected to refine our diagnostic panel to a 17-gene signature (Fig. 4A and B). The combined gene signature score was consistently low in non-PDAC samples, and high for the majority of samples with a final clinical diagnosis of PDAC (Fig. 4C). When considering the potential clinical utility of the signature, we were keen to optimize specificity to avoid false positives. We therefore chose a cut-off score of −1.5 to establish a definite diagnosis of PDAC, as the minimum score which did not include non-PDAC samples. Using our gene expression score at this cutoff has a specificity of 100% and a sensitivity of 70.7% (Table 1). With a clinical perspective in mind, we defined a suspicious clinical scenario in combination with either a positive KRAS mutation or a signature score above our specified cutoff as a positive diagnosis. Using these diagnostic criteria has a specificity of 100% and a sensitivity of 87.8% outperforms current standard cytology in test sensitivity and specificity (Table 1). This result is particularly encouraging when it is considered that we included a significant number of samples in which multiple biopsies were required to establish a diagnosis using standard cytology, while only a single biopsy was used to evaluate our gene signature and KRAS phenotype.

We noted that several of the outlier PDAC samples which displayed low expression of the diagnostic signature were also KRAS wild type. Given the high frequency of KRAS mutations in PDAC, we hypothesized that these outlier samples were likely predominantly comprised of benign or inflamed pancreatic tissue with low tumor cellularity.

To extract genetic material of optimal quality, we elected to utilize snap-frozen EUS-FNA biopsies which are processed in their entirety. The concurrent specimen which is sent to pathology can be assessed for cellularity, but may not accurately reflect the cellularity of the frozen sample, thereby making it difficult to clearly differentiate true negative signature expression results from those due to sampling error.

As KRAS is commonly expressed in PDAC cells, we hypothesized that measurement of the KRAS mutation allele fraction (MAF; mutant KRAS/wild-type KRAS) may provide a useful surrogate marker of tumor cellularity in our specimens, and allow us to assess whether sampling error may be responsible for the KRAS wild-type, low PDAC signature expressing specimens. From the 41 patients with PDAC included in the validation cohort, ddPCR analysis was performed on a representative set of 32 specimens where adequate tumor-derived DNA was available (Fig. 5A). As expected, MAF varied across the population but closely correlated with higher RNA signature expression (Fig. 5B and C), suggesting higher tumor cellularity in these samples and providing a rational explanation for lower signature expression in the outlier PDAC samples.

Cytologically nondiagnostic specimens with negative KRAS, low signature expression and low MAF could therefore be presumed to be samples with very low (or no) tumor cells and would represent the population who would require further diagnostic biopsies. In total, five samples were identified which failed to meet the minimum requirement of either definitive cytology, positive KRAS (by StripAssay testing or MAF >1%), or a signature score above the defined PDAC cutoff. Therefore, these samples (8.3%) would represent the patient cohort who are likely to require a further diagnostic biopsy, a significantly lower proportion than we observed in our retrospective review of real-world patients.

Establishing an adequate, timely tissue diagnosis of PDAC remains a challenge in the clinic, particularly in the era of precision medicine. Tissue availability for genomic profiling is increasingly desirable, with treatment guidelines now recommending germline and somatic testing as standard management for patients with PDAC (39).

Given that the standard of care for the diagnosis of PDAC in most patients is an EUS-FNA biopsy, it is critical that we can make a prompt and accurate clinical diagnosis using these tissues of variable yield and cellularity, to ensure patients can access genomic testing in a comprehensive and timely fashion. Only a minority of patients with PDAC present with resectable or borderline resectable disease, and a definitive diagnosis by EUS-FNA is vital in this population prior to initiation of treatment such as neoadjuvant chemotherapy. Identifying prognostic biomarkers is also of potential interest in this population, particularly to identify patients who may not benefit from either chemotherapy or surgical resection. We note that most patients in our study had locally advanced or metastatic disease, with early-stage disease underrepresented.

The diagnostic accuracy of EUS-FNA in our review is in keeping with similar published series, in which the sensitivity of EUS-FNA has been reported at 83%–93%, and diagnostic accuracy between 59% and 100% (40). A recent systematic review and meta-analysis of the diagnostic accuracy of EUS-FNA including over 1,800 patients reported several limitations to interpreting reported data, including a lack of reference standard, lack of consistent follow-up data to confirm a clinical diagnosis (therefore likely underestimating the rate of false negatives), and potential biases in patient selection and cytology interpretation (40). We also note that some studies reported any “atypical” results as positive biopsies; however, in our series we counted biopsies with focal atypia or atypical cells of uncertain significance as negative results, as clinically these would not be considered adequate for diagnosis in our local experience. It is therefore perhaps not surprising that our large “real-world” observational study reports slightly lower diagnostic sensitivity when compared with some studies included in this meta-analysis, several of which included as few as 12 patients (41–43).

In our series, we identified several patients with false-negative cytology results, and 1 patient with a false-positive result, in whom appropriate treatment was delayed. While cases such as these are uncommon, they highlight the weaknesses of standard cytology in assessment of EUS-FNA biopsies, and the potential dangers in erroneously treating patients based on an incorrect clinical diagnosis. These cases demonstrate potential diagnostic utility for early molecular profiling of suspected PDAC, and the opportunity to prevent missed or delayed diagnoses and subsequent delays in time to first treatment.

Current diagnostic biomarkers such as CA19.9 and IHC staining have limited utility in the setting of uncertain cytologic results. Other methods such as liquid biopsy to assess circulating tumor cells, circulating tumor DNA or exosomes are of increasing clinical interest. However, these methods currently appear to lack sensitivity for the diagnosis of PDAC, particularly at early stages, and are perhaps better poised to play a role in the monitoring of disease progression and treatment response (44). Circulating microRNA signatures have shown more promise as potential diagnostic biomarkers, but are hampered by a lack of standardized method of assessment for reproducibility, and difficulties in accurate quantification due to their small size (45).

Given the high frequency of KRAS mutations in PDAC, the detection of pathogenic KRAS variants in a nondiagnostic biopsy with high clinical suspicion of PDAC is very likely to represent malignant disease. Several other groups have previously reported on methods of improving EUS-FNA sensitivity by adding KRAS mutation testing to standard cytology (46–50). A meta-analysis pooling over 900 patients from eight studies reported that the addition of KRAS mutation testing to standard cytology increased diagnostic sensitivity from 80.6% to 88.7%, and calculated that the repeat biopsy rate could be reduced from 12.5% to 6.8% if nondiagnostic biopsies were only repeated in the setting of a wild-type KRAS result (28).

In accordance with these findings, in our series, 34 of 42 (81%) cases of PDAC with nondiagnostic biopsies had detectable somatic KRAS mutations, compared with only 1 of 23 (4%) biopsies with benign pathology (in that case, the patient had an intraductal papillary mucinous neoplasm, which would itself warrant further investigation and management). We also identified ddPCR as a highly sensitive method of KRAS-mutant allele detection, which is commonly used for ctDNA analysis but may also prove a novel surrogate marker of cellular adequacy in frozen EUS-FNA biopsies.

Transcriptome profiling is of increasing interest in identifying clinical phenotypes in PDAC, and has been postulated as a method of identifying treatment sensitivity (26). There is limited published data on transcriptomic analysis of EUS-FNA derived RNA in PDAC, and RNA quality and yield remains challenging. A recent study of 74 biopsies from 37 patients using both EUS-FNA and 20G biopsies reported that only 28 (37.8%) biopsies met their specified quality thresholds of RIN>3 and 60 ng total RNA for transcriptomic profiling, and these were predominantly the larger bore biopsies which are not currently considered standard of care (51). Similarly, Rodriguez and colleagues used EUS-FNA derived RNA for transcriptomic profiling to distinguish between benign and malignant biopsies in 48 patients, but despite including only cases with definitive cytology (no “suspicious” or atypical biopsies), 12 of 45 samples (26.7%) were excluded due to inadequate quality (52). While the yield and quality of RNA from EUS-FNA in these studies were disappointing, they support our findings that EUS-FNA–derived RNA can be valuable in transcriptomic profiling. Our results are particularly encouraging when you consider that both of these previous studies were much more selective in their biopsy selection and did not include any nondiagnostic specimens.

Here, we demonstrate the utility of gene expression analysis in the initial diagnostic process. Our choice of the NanoString platform as a targeted gene expression approach offers clear advantages by allowing for sensitive results using low input and relatively lower quality RNA with rapid turnaround time, and could therefore feasibly be incorporated into clinical management without significant delay when compared with standard cytologic assessment.

Our retrospective cohort excluded patients with incomplete clinical records and cystic or non-pancreatic pathology, and therefore was likely biased toward patients with true positive PDAC. In contrast, our validation set was biased toward samples which had no clear diagnosis, and had required multiple biopsies to establish a diagnosis. We also chose to apply a strict cut-off score to optimize specificity at a level much higher than many standard diagnostic tests. The broader gene panel has been designed to also clearly differentiate pancreatitis and pNET samples, and therefore may have even wider clinical utility. Several genes included in our diagnostic signature have previously been observed to have potential prognostic significance, and we plan to further explore the predictive and prognostic potential of the broader gene panel as our clinical dataset matures.

To calculate the gene expression signature score using the NanoString platform, we elected to use a simple calculation formula that is not mathematically dependent on gene expression, maximizing its generalizability to future samples. While in our NanoString cohort, the parameters mean and SD for each gene were calculated to perform the z-transformation, these parameters can be reused in future testing without recalculation.

While we have focused here on the use of genomic profiling to improve diagnostic accuracy, there are also clear potential benefits in using EUS-FNA–derived molecular profiling as an aid to predict for therapeutic response. Despite significant advances in the characterization of PDAC at a genomic level, the implementation of precision medicine in PDAC has proved slow and challenging. Barriers to the successful implementation of targeted therapy in PDAC include the adequacy of specimens and feasibility of extracting genetic material from FFPE, unacceptable delays in accessing and processing tissue and subsequent molecular reports, and patient factors including frequent clinical decline prior to commencing therapy on clinical trials (53).

More recently, several studies have reported more success in the use of molecular profiling of PDAC to guide precision therapy (54–56). The Know Your Tumor study included some FFPE-derived FNA biopsies, although they did not report on the success rate of extracting adequate genetic material from these specimens (56), which are known to present challenges in terms of both yield and quality, as processing can lead to degradation and fragmentation of nucleic acids (57). The COMPASS study was able to demonstrate the feasibility of real-time comprehensive genomic profiling in 60 PDAC biopsies, with successful analysis of 98% of samples (55). However, they did not use standard-of-care biopsies, but required additional percutaneous core biopsies (predominately obtained from liver metastases) and specialized processing including laser capture microdissection, which is costly and not widely available (55). In contrast, our more pragmatic approach aims to maximize the use of current standard-of-care EUS-FNA diagnostic biopsies, to minimize additional interventions for patients and reduce the time from presentation to molecular analysis and reporting. Our diagnostic signature could feasibly be utilized to screen the adequacy of EUS-FNA biopsies for further analysis.

There are some limitations to our study. While we were able to demonstrate the diagnostic utility of a transcriptomic signature in our patient cohort and demonstrate the feasibility of molecular profiling using snap-frozen EUS FNA biopsies, it is clear that some biopsies remain inadequate, and the minimum cellularity required to process snap-frozen EUS-FNA biopsies for molecular profiling is difficult to define. However, our exploratory analysis suggests that our gene signature score or a KRAS MAF >1% as measured by ddPCR may be a good marker of adequate tumor cellularity in KRAS-mutant specimens.

We did not record patient reported outcome measures (PROM) as part of this study; however the potential emotional burden arising from delays in diagnosis, including the need for repeat procedures and prolonged time from first presentation to confirmation of malignancy and initiation of treatment should not be underestimated, and prospective studies could consider including PROMs to assess this further. The financial burden of genomic assessment has also not been addressed here, although it is possible that optimizing diagnostic procedures and processing of tumor material may balance the costs of molecular analysis, by reducing health care resource use and costs associated with repeat procedures.

Overall, our study emphasizes the value of routine tissue biobanking and demonstrates the utility of EUS-FNA biopsies as a source of material for clinically relevant genomic and transcriptomic analysis, which warrants further assessment of diagnostic, prognostic, and predictive utility in prospective clinical trials.

Our results confirm that EUS-FNA biopsies can be used to establish a molecular diagnosis of PDAC, even in cases with borderline or nondiagnostic standard cytology. Our validated diagnostic gene signature with KRAS mutation analysis significantly improves upon the diagnostic accuracy of current standard procedures, and could feasibly be implemented into clinical practice to reduce the need for repeat procedures. Transcriptomic analysis of PDAC at the time of diagnosis has widespread potential clinical utility beyond increasing current test accuracy, and future areas to explore could include use of this information as predictors for prognosis and therapy selection.

Further evaluation of the clinical utility of our signature is planned in a multi-centre, prospective clinical trial which will also evaluate the predictive value of tumor molecular profiling (ACTRN12620000762954). This study will further evaluate the diagnostic performance of our gene signature in a larger patient cohort, in combination with a clinically focused commercial gene panel which will be used to identify patients with potentially actionable molecular phenotypes, and has been designed to further demonstrate that genomic profiling of EUS-FNA biopsies can improve upon current diagnostic procedures and therapeutic selection in PDAC.

The datasets (including accession codes or web links to publicly available datasets) and analyses generated from this study are available upon reasonable request from the corresponding author.

J. Lundy, H. Gao, B.J. Jenkins, and D. Croagh report a patent for Methods of detecting and or diagnosing pancreatic cancer AU2021902314 pending. No disclosures were reported by the other authors.

J. Lundy: Conceptualization, data curation, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. H. Gao: Conceptualization, data curation, formal analysis, validation, investigation, methodology, writing–review and editing. W. Berry: Data curation, writing–review and editing. S. Masoumi-Moghaddam: Data curation, writing–review and editing. B.J. Jenkins: Conceptualization, resources, supervision, writing–review and editing. D. Croagh: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing–review and editing.

This research was supported by a Victorian Cancer Agency Collaborative Research Grant (ICOUGI18023).

The authors would like to acknowledge the contribution of Trevor Wilson and the team in the MHTP Medical Genomics Facility for their assistance with Next Generation Sequencing; Zdenka Prodanovic for her oversight of the VPCB; and all the clinicians, scientists, and patients who contributed their time and tissues to the VPCB to enable this research.

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.