Novel protein biomarkers for early detection of pancreatic ductal adenocarcinoma using longitudinal serum protein measurements and machine learning.

Abstract

e14536 Background: Pancreatic ductal adenocarcinoma (PDAC) is highly aggressive with one- and five-year average net survival rates of approximately 28% and 8%, respectively (NHS Digital 2023). The majority of PDACs are diagnosed in late stages, when surgical resection and cure are not possible. Without better screening tools, by 2030, PDAC is expected to become the second leading cause of cancer death (Rahib et al. 2014). Undoubtedly, there is an exceedingly critical need for an early detection tool of PDAC. Our study utilizes multiplexed sensitive targeted proteomics, a unique biobank with longitudinal serum samples, and machine learning to measure and evaluate the change in abundance of select serum proteins over time. With this, we identified a novel panel of serum biomarkers for possible early detection of PDAC. Methods: Liquid chromatography coupled to selected reaction monitoring (LC-SRM) based targeted proteomics was used to measure the abundances of 140 candidate biomarker proteins in longitudinal serum samples of 82 subjects with PDAC and 82 healthy control subjects selected from the Department of Defense Serum Repository (DODSR) between 1988 to 2020. A random forest machine learning approach was used to develop two classification models using the SRM data. One model (termed the “cross-sectional” model) was trained on data from samples that were drawn during the “diagnosis window” (defined as 100 days before to 100 days after the date of each case subject’s diagnosis). The other “longitudinal” model was trained using data precomputed to measure change in relative abundance from 1-4 years before diagnosis to the “diagnosis window.” Results: The cross-sectional and longitudinal models yielded high classification performance, achieving area under the curve (AUC) values of 0.89 and 0.91, respectively. This of utilizing serum measurements from four years to one year prior to diagnosis as a reference point. The feature selection algorithm consistently selected fifteen proteins for each type of model. Among the ten proteins identified and overlapping between the two models, Gelsolin, Cartilage oligomeric matrix protein, Tetranectin, TIMP metallopeptidase inhibitor 1, and Apolipoprotein A4 have been linked to PDAC in previous studies measuring biomarker levels in sera, plasma, or tissue. Conclusions: Our study identified additional novel protein biomarkers which, after validation, may be utilized with the other proteins in the classifiers for earlier detection of pancreatic adenocarcinoma. Our findings highlight the promise of detecting PDAC earlier by determining longitudinal protein abundance change in serum and the potential substantial changes to PDAC outcomes if successful in identifying PDAC at earlier stages.