Study design and sample description

This study utilized a total of 90 tissues (52 PDAC tumours, 38 matched para-tumour tissues) and 546 plasma samples (198 PDAC, 25 CP and 323 healthy controls) to sequentially develop a PDACatch assay (Fig. 1, Additional file 2: Fig. S1, Tables 1 and 2). PDAC samples were collected from 232 PDAC patients (18 PDAC patients provided both tissue and plasma samples). Among 223 PDAC patients with known stage information, 43/119/38/23 cases were Stage I/II/III/IV (Additional file 2: Fig. S2A). In Phase I, we discovered de novo PDAC-specific markers by analysing the genomic DNA methylation profiles of PDAC tumours, normal tissues and plasma samples using the RRBS method [19]. In Phase II, markers were tested in additional tissue and plasma samples for their PDAC-discriminating accuracy. The most informative markers were selected to develop a targeted sequencing assay, PDACatch. In Phase III, PDAC classifiers were built and validated in 199 plasma samples to separate PDAC patients from healthy individuals or CP patients. Furthermore, we conducted a single-blinded test of the PDACatch classifier using an independent cohort of PDAC plasma and healthy controls. In Phase IV, we compared PDACatch and CA19-9’s performances in classifying PDAC plasma and explored an integrated classifier to further improve the accuracy. Note that a small number of tissue and plasma samples were shared in multiple phases of this study (see details in the relevant “Results” sections).

Table 1 Demographic and clinicopathological features of the study cohortsTable 2 Demographic and clinicopathological features of the study cohorts in Phase IVDiscovery of PDAC-specific methylation markers in tissue and plasma

We searched predefined methylation haplotype blocks (MHBs) for de novo PDAC-specific DNA methylation markers by first profiling the methylation patterns of 46 PDAC tumours, 28 para-tumour tissues and 143 plasma samples (20 PDAC, 123 healthy) using RRBS (Table 1). Multiple MHB-specific metrics (see the “Methylation haplotype measurements” section for details) were used to quantify methylation levels to identify MHBs containing PDAC-specific methylation haplotypes as markers via PDAC tissue group vs. para-tumour tissue group (T2T), PDAC tissue vs. healthy plasma (T2P) and PDAC plasma vs. healthy plasma (P2P) comparisons (Fig. 2A). The first set of 76 markers was yielded by intersecting the 1870 T2P markers and 700 T2T markers (Wilcoxon rank sum test, Benjamini and Hochberg FDR <0.05). The second set comprised 42 T2T markers located within −1500 to +1000 bp of the transcription start sites of 819 genes exhibiting aberrant methylation changes in PDAC tissues or plasma [21,22,23,24]. The third set of 53 markers was selected from P2P markers via a model-based cross-validation marker selection process using an AUC value of over 0.75 as the cut-off for qualified markers. In total, 171 de novo markers were compiled for downstream analysis.

Fig. 2

A PDAC-specific markers were discovered by T2T, T2P and P2P comparisons separately and then intersected and combined, as depicted in the figure. B Unsupervised hierarchical clustering of PDAC and para-tumour tissues based on their methylation measurements of the 750 assembled markers, which were ordered along the Y axis of the heatmap. C The receiver operating characteristic (ROC) curve of an SVM model built by using the 200 most discriminatory markers to cross-validate Phase II plasma samples

Gene set enrichment analysis (GSEA) of these de novo MHBs revealed that multiple cancer-related pathways or biological processes were enriched in their associated genes (FDR < 0.05, Fig. 3A, B and Additional file 3: Table S1) [33], including 7 MsigDB hallmark pathways known to be dysregulated in PDAC (Fig. 3B) [23, 34]. Furthermore, the PDAC marker genes previously published in the literature were also highly enriched in our top marker-associated genes (hypergeometric test, Additional file 4: Table S2) [21,22,23,24]. These results strongly supported that the de novo markers we selected are involved in PDAC carcinogenesis.

Fig. 3

A, B Top 10 GO biological processes (A) and the 7 MSigDB hallmark cancer-related pathways (B) identified by GSEA that were enriched in the genes associated with the 171 de novo PDAC-specific markers. C–E Significantly enriched functional categories of genes associated with the 56 markers in the final PDACatch classifier, as identified by GO analysis. C Biological processes (top 20). D Cellular component. E Molecular function

Marker optimization and PDACatch assay development

In Phase II, we further reduced the number of PDAC markers to minimize model overfitting caused by the imbalance of hundreds of features and limited samples. To this end, all the selected markers were integrated with the PanSeer assay [32], and their separation power was validated. PanSeer is a targeted methylation sequencing assay that is highly sensitive for detecting early-stage cancer signals in blood. It is also readily customized for different sets of targets, making it versatile for investigating different types of cancers.

A combination of the PDAC de novo markers with the PanSeer markers formed a starting set of 750 markers for further testing (Fig. 2B). We filtered them by using them to discriminate PDAC tumours from para-tumour tissues (N = 27 and 17, respectively). Among them, 21 PDAC and 7 para-tumour tissues were previously used in Phase I. They were reused in Phase II to confirm that RRBS-discovered markers can also be consistently detected by targeted methylation sequencing. The top 200 most discriminating markers (p < 0.05, Wilcoxon rank sum test) were selected and preliminarily filtered based on their ability to classify PDAC and healthy plasma (N = 29 and 55, respectively) via cross-validation (Fig. 2C) and the distribution of their methylation haplotype measurements in these samples (chi-square test, p < 0.01). The 185 most significant markers were chosen to develop the final version of the PDACatch assay to detect the PDAC marker signature in blood.

Model building and evaluation of the PDACatch classifier for early PDAC detection

We then sought to develop a PDAC early detection classifier to separate PDAC plasma from healthy controls by the PDACatch assay. To this end, 94 PDAC and 80 healthy samples, which were age- and sex-matched, were randomly split into a training set and a validation set at a 2:1 ratio (Fig. 1 and Table 1). The training set included 19 PDAC plasma samples that were previously tested in Phase II and had sufficient remaining cfDNA. This was to increase the size of the training set to improve the trained classifier’s robustness; however, no samples were reused in validation to prevent biasing the validation results. Using training samples, the 56 most discriminatory markers for PDAC were identified by 10-fold cross-validation incremental feature selection (Additional file 2: Fig. S3 and Additional file 5: Table S3) to build an SVM-based classifier with a high AUC of 0.93 in the training set (sensitivity = 71%, specificity = 91%) (Fig. 4A). The PDACatch classifier was then validated in the left-out validation set and achieved a similar AUC of 0.91 (sensitivity = 84%, specificity = 89%) using the same cut-off as in the training set, demonstrating its consistency and robustness (Fig. 4A, B). Covariant analysis also showed that the PDACatch classifier was independent of age, sex, tumour location and size (Fig. 4C–F).

Fig. 4

Performance of the PDACatch classifier in differentiating PDAC from healthy plasma and covariate analysis of the PDACatch classifier. A ROC curve of the PDACatch classifier distinguishing PDAC and healthy plasma in the training and validation cohorts. B The PDACatch classifier scores across different types of samples. In Panel B, samples were labelled with cohorts (Training/Validation), pathological types (H: Healthy; CP: chronic pancreatitis) and stages (I, II, I-IIA and IIB-IV). During covariate analysis, PDACatch scores of PDAC and healthy plasma samples were grouped by sex (C), age (D), tumour size (E) and tumour location (F). In E, brackets and parentheses indicate inclusion and exclusion of size, respectively. Wilcoxon rank sum test: ns, not significant (0.05 < p ≤ 1.0); *: 0.01 < p ≤ 0.5; **: 0.001e−03 < p ≤ 0.01; ***: 0.0001 < p ≤ 0.001; ****: p ≤ 0.0001

Genes associated with the 56 markers of this classifier were annotated and a number of cancer-related genes or gene families were identified, including HOX family [35] and TBX family [35, 36] members (Additional file 5: Table S3). Several have been proposed for the detection of PDAC or other cancers in blood, including BCAN, IKZF1, TBX15, BNC1 and SHOX2 [17, 21, 32, 37, 38]. Gene Ontology (GO) analyses showed significantly enriched molecular function categories for DNA binding or transcription factor activity (Fig. 3E). It may be worth exploring whether these transcription factors have regulatory roles in PDAC carcinogenesis.

While testing CP samples, we found that the PDACatch classifier showed limited accuracy in stratifying CP from PDAC. We then rebuilt an SVM-based classifier to separate PDAC from CP plasma, which achieved an AUC of 0.85 for samples in the validation set (Additional file 2: Fig. S4A). This PDAC-CP classifier exhibited a consistent accuracy for PDAC across all stages (Additional file 2: Fig. S4B) with no significant covariate differences (Additional file 2: Fig. S4C-F). Although the limited CP samples likely have reduced performance during validation, the results did suggest potentially great feasibility in differentiating PDAC from CP using ctDNA methylation as markers to reduce misdiagnosis due to the lack of discriminatory symptoms.

Comparison of the PDACatch classifier with serum CA19-9 levels

As mentioned earlier, serum CA19-9 is commonly used as a blood marker to stratify PDAC risk. Thus, it is necessary to compare the performance of the PDACatch classifier with CA19-9 in all samples with available test results for CA19-9 to assess PDACatch’s clinical utility and significance.

We compared the classification accuracy by PDACatch and CA19-9 on all 92 PDAC and 37 healthy cases with known CA19-9 levels from the training and validation samples of Phase III (Fig. 1 and Table 2) and found that on balance, PDACatch was modestly more accurate than CA19-9, as demonstrated by the fact that PDACatch has a higher, or at least an equal, AUC score than CA19-9 for PDAC of each stage (Fig. 5A and Table 3). Importantly, PDACatch was more sensitive in detecting Stage I (sensitivity = 80 and 68% for PDACatch and CA19-9, respectively, Additional file 2: Fig. S5) or early-stage (I-IIa) PDAC plasma than CA19-9 (sensitivity = 76 and 70% for PDACatch and CA19-9, respectively). Note that in this comparison, PDACatch and CA19-9 had the same specificity of 89%. These results indicate that PDACatch may be more advantageous in detecting early PDAC cases than CA19-9.

Fig. 5

Independent test of the PDACatch classifier and its comparison and integration with CA19-9. A ROC curves of CA19-9, PDACatch and the combined classifier in differentiating PDAC and healthy controls in the training and validation sets. B Comparison of the sensitivities of CA19-9, PDACatch and the combinatorial classifier in classifying PDAC of different stages. The specificity was fixed at 89%. Error bars: 95% CI of sensitivity. C ROC curve of the PDACatch classifier on the independent test samples. D Comparison of the sensitivity of PDACatch and CA19-9 in detecting CA19-9-negative PDAC cases in the study cohorts. Note that the specificity was of the entire cohort by the classifier. E Predicted probability scores for noncancerous cases (n = 37) and Stage I (n = 25), Stage II (n = 44), Stage III (n = 18) and Stage IV (n =5). The same samples were also tested for CA19-9 levels. Orange dots show the CA19-9-positive cases (>37 U/ml), and blue dots show CA19-9-negative cases (≤37 U/ml)

Table 3 Performance of PDACatch, CA 19-9 and the combined model to classify all cases of the training and validation sets that have CA19-9 levels (bootstrapped 1000 repetitions at 95% CIs)

Additionally, because 5~10% of PDAC cases do not have elevated CA19-9 levels due to genetic background [39], we specifically evaluated whether PDACatch can accurately detect PDAC cases considered to be negative in the CA19-9 test (termed CA19-9-negative cases, defined as CA19-9 levels lower than 37 U/ml, N = 21) from healthy controls (N = 33). Indeed, it correctly identified 13 out of 21 CA19-9-negative PDAC cases in the training and validation cohorts (sensitivity =54, 75 and 62% for the training, validation and combined cohorts, respectively) at a specificity of 91% (Fig. 5D and Additional file 2: Fig. S4A). Taken together, PDACatch not only outperformed CA19-9 in detecting early-stage PDAC patients but also accurately identified CA19-9-negative PDAC cases.

Lastly, we explored integrating CA19-9 with the PDACatch classifier to potentially maximize the model accuracy. To this end, we used cases from the training and validation sets for PDACatch that had known CA19-9 levels for the combinatorial model’s training (23 healthy, 62 PDAC) and validation (14 healthy, 31 PDAC), respectively. The combinatory classifier was trained by logistic regression and achieved an AUC of 0.93 (sensitivity = 82%, specificity = 87%); in validation, it achieved an AUC score of 0.96 (sensitivity = 94%, specificity = 93%), which was higher than either parental classifier (0.87 and 0.90 for CA19-9 and PDACatch, respectively) in classifying the same validation samples (Fig. 5, Table 3).

Because the combinatorial classifier had consistent performances in both training and validation cohorts, we further compared the combinatorial classifier’s performances with CA19-9 in classifying all the cases of these 2 cohorts. Indeed, the combinatorial classifier had an AUC of 0.94, higher than CA19-9 (AUC = 0.89, Fig. 5A–E and Table 3). Additionally, it was more sensitive than CA19-9 when classifying Stage I (sensitivity = 92 and 68% for the combinatorial classifier and CA19-9, respectively, p < 0.05, McNemar’s test, Additional file 2: Fig. S5) or early-stage PDAC plasma (I-IIa) (sensitivity = 88 and 70% for the combinatorial classifier and CA19-9, respectively, p < 0.05, McNemar’s test, Additional file 2: Fig. S6). These results suggest that early detection of PDAC may be improved by combining PDACatch and CA19-9.

Independent test of the PDACatch classifier to distinguish PDAC and healthy plasma samples

To independently verify the PDACatch classifier’s utility in classifying PDAC plasma, we conducted a single-blinded classification on a cohort of preoperative PDAC (N = 74) plasma samples and healthy controls (N = 65, Fig. 1 and Table 1) obtained from ProteoGenex, a biobank in the USA. The PDACatch assay was performed on these samples, and the same classifier and cut-off were applied to label these samples as PDAC or normal.

The results showed that for the blind-test cohort, PDACatch achieved an AUC of 0.91 (sensitivity = 82%, specificity = 88%, Fig. 5C) in classifying PDAC cases, reaching a degree of accuracy that was essentially identical to that of the validation cohort (AUC = 0.91, sensitivity = 84%, specificity = 89%), further confirming its robustness and consistency. Stagewise, PDACatch detected early-stage PDAC (I-IIa) at a sensitivity of 80% and advanced-stage PDAC (IIb and above) at 83%, both of which were also consistent with the results of the validation cohort.

Importantly, PDACatch correctly identified all 7 CA19-9-negative PDAC samples in this cohort, achieving a sensitivity of 100% (Fig. 5D). While the number of such cases was relatively small in this cohort (22 of the 74 PDAC samples had serum CA19-9 levels measured), combined with the results of the same test on CA19-9-negative cases of the training and validation cohorts, it nonetheless demonstrated the PDACatch classifier’s consistent accuracy at detecting CA19-9-negative PDAC cases. Taken together, we found that the PDACatch classifier performed consistently in classifying PDAC plasma in the independent blind-test cohort as it did in the training and validation cohorts, confirming its robustness and utility for the noninvasive detection of PDAC in blood.