Classification of 10 cancer types using imaging mass spectrometry

Aim: Typing of cancer subtypes requires histological, immunohistological and in some cases molecular pathological analyses. To support the pathologist in decision-making we have developed a mass spectrometry-based tumor typer using imaging mass spectrometry (IMS). Ten tumor types were included in the model including adenocarcinomas of the breast, liver, pancreas, colon and lung, squamous-cell carcinoma of the lung, urothelial carcinoma, lymphoma as well as pancreatic and gastrointestinal neuroendocrine tumors. Methods: Formalin-fixed paraffin-embedded (FFPE) tissue microarray sections (5 μm) were subjected to deparaffination, heat-induced epitope retrieval, in-situ trypsin digestion and deposition of a matrix solution for IMS analysis. MS data were acquired at a spatial resolution of 100 µm using an Autoflex Speed TOF/TOF system. Tissue sections were stained by Hematoxylin and Eosin after IMS and cancer regions were annotated. Spectra of datasets each including a specific tumor were used to train three different classification models using R statistical software. An independent subset of spectra was used to validate all algorithms. Immunohistochemistry was performed for validation of results. Results: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM) and Random Forest (RF) were used for the classification analysis. The cohort was separated into a training set (n=217) and a validation set (n=105). The models ranked the contribution of each tumor profiling towards the optimal separation of spectra derived from different tissues with an overall accuracy of 96%. In the validation set, tumors were correctly classified with an accuracy of 98% (LDA), 94% (SVM), and 93% (RF). The specificity in the recognition of the various tumor types was between 90 and 100%. Conclusion: For the first time IMS was applied to classify a large set of malignancies on FFPE tissue. The advantages of high accuracy, speed, saving materials and cost for consumables compared to IHC, make this method a great potential for routine diagnostics for supporting tumor classification.