Self-supervised clustering of mass spectrometry imaging data using contrastive learning.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Analytical Chemistry

Background:

• Mass spectrometry imaging (MSI) enables label-free molecular mapping of biological tissues.
• Identifying co-localized molecules is vital for understanding biochemical pathways.
• Manual annotation of large MSI datasets is infeasible, and small datasets hinder deep learning model training.

Purpose of the Study:

• To develop a self-supervised clustering approach for accurate molecular colocalization analysis in MSI data.
• To overcome the limitations of manual annotation and small dataset sizes in MSI analysis.
• To enable autonomous and high-throughput identification of co-localized molecules.

Main Methods:

• A self-supervised clustering method based on contrastive learning was employed.
• A deep convolutional neural network (CNN) was trained on MSI data without manual annotations.
• The CNN learned high-level spatial features from ion images for classification based on molecular colocalization.
• Self-labeling was used to fine-tune the CNN encoder and linear classifier.

Main Results:

• Contrastive learning effectively generated well-resolved clusters of ion image representations.
• The developed approach demonstrated excellent performance in clustering MSI data.
• The method enables autonomous and high-throughput identification of co-localized species.
• The approach successfully learned spatial features for molecular colocalization.

Conclusions:

• This self-supervised contrastive learning method offers an effective solution for MSI data clustering and molecular colocalization.
• The approach facilitates autonomous and high-throughput analysis, expanding spatial lipidomics, metabolomics, and proteomics.
• This method addresses key challenges in MSI data analysis, improving biochemical pathway understanding.