ColocML: machine learning quantifies co-localization between mass spectrometry images.

Area of Science:

• Biomedical imaging
• Computational biology
• Mass spectrometry

Background:

• Imaging mass spectrometry (imaging MS) is crucial for mapping molecular distributions in tissues.
• Computational methods analyzing spatial correlations (co-localization) are vital for imaging MS data interpretation.
• Current co-localization measures lack comprehensive evaluation, leading to inconsistent application and hindering progress.

Purpose of the Study:

• To address the lack of standardized evaluation for co-localization measures in imaging MS.
• To create a robust benchmark dataset and develop novel, high-performing co-localization algorithms.
• To provide a reliable tool for analyzing molecular spatial relationships in biological samples.

Main Methods:

• Developed ColocML, a machine learning framework for co-localization analysis.
• Created a gold standard dataset of 2210 ion image pairs ranked by 42 imaging MS experts.
• Evaluated existing methods and introduced novel measures using TF-IDF and deep neural networks.

Main Results:

• The semi-supervised deep learning Pi model and median thresholded cosine score achieved high correlation with expert rankings (Spearman 0.797 and 0.794).
• Demonstrated the utility of these measures by analyzing co-localization for over 10,000 molecules across thousands of public datasets.
• Identified top-performing co-localization methods for imaging MS data.

Conclusions:

• ColocML provides a validated, data-driven approach to co-localization analysis in imaging MS.
• The developed methods offer improved accuracy and reliability for spatial molecular correlation studies.
• This work establishes a benchmark for evaluating co-localization measures and facilitates future research in molecular imaging.