Interpreting MALDI imaging data for rare types of ampullary cancer using machine learning

Rare tumor diseases are difficult to diagnose and there is a lack of routine diagnostic procedures. Approaches must be found that allow comprehensive identification and evaluation of prognostic relevant target proteins or transcripts. Analyzing rare ampullary cancer, respectively, to their prognosis and predictive factors by machine learning (ML) based matrix-assisted laser desorption/ionization (MALDI) time-of-flight (TOF) imaging is a first step towards providing new solutions for diagnostics of those cancer samples. In this study, we investigated a cohort of ampullary adenocarcinomas, including intestinal, pancreatic and cases of unknown subtypes, to identify differences in the proteome. Human formalin-fixed paraffin-embedded (FFPE) tissues were pathologically assessed, immunohistological stained, MALDI Imaging detected, and ML-related analyzed. We enable MALDI imaging as a diagnostic complement for immunohistochemical analysis and provide a MALDI Imaging neural network for broad application in tumor diagnostics. Moreover, using tools from ML model explainability, we determined a small subset of influential m/z-values from the trained models. The transformation of locally established ML networks dependent on one proteomic application source to other similar application sources (without peak picking or other pre-processing) is the basis for future rare cancer patient data collection.