Differentiating Gastric Cancers from Acid Peptic Diseases through Integrative Targeted Proteomics and Machine Learning Approaches

Abstract

Gastric cancers (GCs) are often diagnosed in advanced stages owing to nonspecific early symptoms resembling Acid Peptic Diseases (APDs). Despite recent efforts, a simple, liquid biopsy-based multiprotein panel prediagnostic assay capable of differentiating GCs from APDs is lacking. Mass spectrometry (MS)-based targeted proteomics methods, including Multiple Reaction Monitoring (MRM), are utilized as the method of choice to develop Laboratory Developed Tests (LDTs) that revolutionize GC early diagnosis and screening. In this study, a 22-min MS-MRM LDT was developed and tested to quantify a serum protein panel in 135 serum samples from treatment-naive cases of GCs, APDs, and healthy individuals. Notably, a novel Deep Neural Network (DNN)-based pattern recognition scoring architecture, integrated with a model explainability tool (SHAP), was developed to score and categorize GCs. The MRM-MS assay produced minimal carryover and matrix effects, with adequate limits of detection/quantification. Quantities of SAA1 and IGFBP2, as determined through ELISA, demonstrated similar sensitivity compared to the LDT. Importantly, the DNN-based scoring architecture efficiently differentiated GCs from the rest of the samples (AUROC = 0.95), with average precision marking >0.90 and minimal bias in protein expression affecting model performance. This LDT can serve as a prediagnostic screening method to distinguish GCs from APDs, guiding clinicians and patients in proceeding with a confirmatory diagnosis.