Machine learning-featured Secretogranin V is a circulating diagnostic biomarker for pancreatic adenocarcinomas associated with adipopenia

Abstract

Background: Pancreatic cancer is one of the most fatal malignancies of the gastrointestinal cancer, with a challenging early diagnosis due to lack of distinctive symptoms and specific biomarkers. The exact etiology of pancreatic cancer is unknown, making the development of reliable biomarkers difficult. The accumulation of patient-derived omics data along with technological advances in artificial intelligence is giving way to a new era in the discovery of suitable biomarkers.

Methods: We performed machine learning (ML)-based modeling using four independent transcriptomic datasets, including GSE16515, GSE62165, GSE71729, and the pancreatic adenocarcinoma (PAC) dataset of the Cancer Genome Atlas. To find candidates for circulating biomarkers, we exported expression profiles of 1,703 genes encoding secretory proteins. Integrating three transcriptomic datasets into either a training or test set, ML-based modeling distinguishing PAC from normal was carried out. Another ML-model classifying long-lived and short-lived patients with PAC was also built to select prognosis-associated features. Finally, circulating level of SCG5 in the plasma was determined from the independent cohort (non-tumor = 25 and pancreatic cancer = 25). We also investigated the impact of SCG5 on adipocyte biology using recombinant protein.

Results: Three distinctive ML-classifiers selected 29-, 64- and 18-featured genes, recognizing the only common gene, SCG5. As per the prediction of ML-models, the SCG5 transcripts was significantly reduced in PAC and decreased further with the progression of the tumor, indicating its potential as a diagnostic as well as prognostic marker for PAC. External validation of SCG5 using plasma samples from patients with PAC confirmed that SCG5 was reduced significantly in patients with PAC when compared to controls. Interestingly, plasma SCG5 levels were correlated with the body mass index and age of donors, implying pancreas-originated SCG5 could regulate energy metabolism systemically. Additionally, analyses using publicly available Genotype-Tissue Expression datasets, including adipose tissue histology and pancreatic SCG5 expression, further validated the association between pancreatic SCG5 expression and the size of subcutaneous adipocytes in humans. However, we could not observe any definite effect of rSCG5 on the cultured adipocyte, in 2D in vitro culture.

Conclusion: Circulating SCG5, which may be associated with adipopenia, is a promising diagnostic biomarker for PAC.

Keywords: adipopenia; biomarker; cachexia; diagnosis; machine learning; pancreatic adenocarcinoma; pancreatic cancer; prognosis.

Figure 1

Schematic workflow to identify diagnostic and prognostic markers for pancreatic adenocarcinoma. (A) Workflow to identify prospective diagnostic markers from three independent transcriptomes. (B) Workflow to predict a putative prognostic marker from the transcriptome of pancreatic adenocarcinoma (PAC) at The Cancer Genome Atlas (TCGA).

Figure 2

Performance of constructed random-forest (RF) models. (A–D) Plots summarizing the area under receiver operating characteristic curve (AUROC) (A), accuracy (B), F1 score (C), and kappa value (D) of RF models that classifies normal pancreas and pancreatic adenocarcinoma. (E, F) Plots highlighting the AUROC (E), accuracy (F), F1 score (G), and kappa value (H) of RF models that characterize the prognosis of PAC. (I, J) Tables summarizing the performance (I) and selected features (J) of XGBoost modeling, (K) Venn diagram showing SCG5 as the only common feature in two distinct ML models.

Figure 3

The expression profiles and clinicopathological signatures of Secretogranin V (SCG5). (A–C) Box plots presenting the levels of SCG5 transcript in NOP and PAC. (D) Box plot showing SCG5 expression at each stage of PAC. The boxplots (A–D) present the 25% quartile, the median, and 75% quartile. Student’s t- test (A–C) and one-way ANOVA (D) determined the statistical significance. ***p < 0.001; **p < 0.01; *p < 0.05. (E–G) Kaplan–Meier curves for overall survival of patients with PAC from TCGA cohort. (H) The forest plot showing the hazard ratio and 95% confidence intervals associated with the SCG5 level.

Figure 4

Plasma level of SCG5 distinguishes patients with PAC from those without PAC. (A) Plasma level of SCG5 in patients with PAC and those without PAC. (B, C) Plasma level in (B) male and (C) female participants. The boxplots (A–C) present the 25% quartile, the median, and 75% quartile. Student’s t- test computed the statistical significance. ****p < 0.0001, ***p < 0.001; **p < 0.01. (D) Receiver operating characteristic curve proposing optimal cut-off value (106.27 ng/ml) distinguishing patients with PAC from those without PAC. (E–H) Scatter plots estimating correlations between plasma SCG5 and BMI (E) or age (F–H). (I–O) Scatter plots showing correlations between plasma SCG5 and BMI (E) or age in females (I–L) and males (M–P).

Figure 5

SCG5 expression in human pancreatic tissues is positively associated with the size of subcutaneous adipocytes. (A) Bar plot showing pancreatic expression of SCG5 in the human Genotype-Tissue Expression (GTEx) portal. (B) Hematoxylin and eosin-stained images of human subcutaneous adipose tissues from the GTEx portal, (C, D) Plots displaying the (C) size and (D) distribution of the cross-sectional area of adipocytes from H&E-stained histological images. The boxplots (C) present the 25% quartile, the median, and 75% quartile. Student’s t- test (C) assessed the statistical significance. ****p < 0.0001.

Figure 6

Unbiased transcriptomic analysis shows that pancreatic SCG5 expression is associated with the adipopenia phenotype. (A) Bubble plot summarizing the results of gene set enrichment analysis (GSEA) dissecting adipose transcriptomic profiles from donors of pancreatic SCG5-high or -low groups. (B) Representative enrichment plots of cachexia and adipopenia-related gene sets. (C–F) Gene networks comparing correlations among top genes of each indicated gene set, human phenotype ontology (HP) Lipodystrophy (C), HP Cachexia (D), Gene Ontology biological process (GOBP) Mitochondrial respiratory chain complex assembly (E), and GOBP Brown fat cell differentiation (F).