Speaker
Description
Metabolite discovery can provide insights into disease mechanisms and help to identify potential biomarkers that contribute to the development of new treatments. We present a self-supervised deep learning approach for metabolite discovery. Molecular intensity distributions obtained via MALDI-MSI (matrix-assisted laser desorption/ionization mass spectrometry imaging) are compared with histological tissue coloration patterns in breast cancer samples from mice. While most deep learning studies in pathology focus on image classification, our work addresses the less common task of image similarity, linking MSI-derived molecular maps with visual features in stained tissue sections. Biomarker discovery is achieved by identifying specific ions or m/z-values overrepresented in tumor regions and exploring how these molecular markers vary across disease stages, including primary tumors, recurrences, and lung metastases derived from the same breast cancer cell line.
Each tumor sample comprises one reference histological image and approximately 900 MALDI-MSI ion-intensity images. HER2 staining serves as the spatial reference for aligning molecular and histological data. However, alignment is complicated by artifacts introduced during tissue preparation such as shearing, tearing or folding. To address these challenges, we adapted the self-supervised method Bootstrap Your Own Latent (BYOL) for comparing MSI-derived molecular distributions with stained tissue sections. This self-supervised setup eliminates the need for manual labeling of hundreds of MSI images per tumor, enabling efficient representation learning directly from raw data. Our methodological contribution lies in designing domain-specific augmentations that improve robustness to structural distortions and typical MALDI noise while preserving biologically relevant spatial information.
We systematically evaluated three configurations: (i) a pretrained ResNet without additional training; (ii) the standard BYOL model with default augmentations; and (iii) customized BYOL variants incorporating subsets of MSI-specific augmentations. Performance was assessed using receiver operating characteristic (ROC) curves on validation and test sets comprising nine tumors unseen during training, which were independently labelled by experts to define a ground truth. The customized augmentation strategy achieved area-under-curve (AUC) values between 0.8 and 0.95 on test data.
This capability enables large-scale comparison of unlabeled MSI datasets with histological references, is generalizable to a wide range of staining types, and paves the way for identifying molecules associated with specific pathological features.
85717608808