Currently, pathological evaluation of stage I/II colon cancer, following the Union Internationale Contre Le Cancer (UICC) guidelines, is insufficient to identify patients that would benefit from adjuvant treatment. In our study, we analyzed tissue samples from 276 patients with colon cancer utilizing mass spectrometry imaging. Two distinct approaches are herein presented for data processing and analysis. In one approach, four different machine learning algorithms were applied to predict the tendency to develop metastasis, which yielded accuracies over 90% for three of the models. In the other approach, 1007 m/z features were evaluated with regards to their prognostic capabilities, yielding two m/z features as promising prognostic markers. One feature was identified as a fragment from collagen (collagen 3A1), hinting that a higher collagen content within the tumor is associated with poorer outcomes. Identification of proteins that reflect changes in the tumor and its microenvironment could give a very much-needed prediction of a patient’s prognosis, and subsequently assist in the choice of a more adequate treatment.

Knowing the precise location of analytes in the tissue has the potential to provide information about the organs’ function and predict its behavior. It is especially powerful when used in diagnosis and prognosis prediction of pathologies, such as cancer. Spatial proteomics, in particular mass spectrometry imaging, together with machine learning approaches, has been proven to be a very helpful tool in answering some histopathology conundrums. To gain accurate information about the tissue, there is a need to build robust classification models. We have investigated the impact of histological annotation on the classification accuracy of different tumor tissues. Intrinsic tissue heterogeneity directly impacts the efficacy of the annotations, having a more pronounced effect on more heterogeneous tissues, as pancreatic ductal adenocarcinoma, where the impact is over 20% in accuracy. On the other hand, in more homogeneous samples, such as kidney tumors, histological annotations have a slenderer impact on the classification accuracy.