ColocML

ColocML quantifies co-localization between ion images from imaging mass spectrometry (imaging MS) to evaluate spatial correlations and benchmark co-localization measures.

Key Features:

• Machine Learning Integration: Leverages machine learning and deep neural networks to evaluate and improve assessment of spatial correlations between ion images.
• Gold Standard Dataset: Provides a gold standard of 2,210 ion-image pairs ranked for co-localization by 42 imaging MS experts from nine laboratories.
• Novel Co-localization Measures: Implements term frequency–inverse document frequency (TF-IDF) measures, median-thresholding followed by cosine score, and a semi-supervised deep learning Pi model, with the Pi model and cosine-after-median-thresholding achieving Spearman correlations of 0.797 and 0.794, respectively, against expert rankings.
• Comprehensive Evaluation: Evaluates existing co-localization measures and develops new metrics using expert-validated rankings to benchmark performance.

Scientific Applications:

• Large-scale METASPACE analysis: Applied to 10,273 molecules from 3,685 public METASPACE datasets to infer co-localization properties across extensive imaging MS data.
• Tissue molecular distribution studies: Enables quantitative assessment of molecular co-localization within tissue sections to support analysis of molecular interactions and distributions.

Methodology:

Uses pixel-based co-localization methods and expert ranking to create a gold standard, applies TF-IDF and median-thresholding with cosine scoring, and trains deep neural networks including a semi-supervised Pi model; Spearman correlation is used to compare measures to expert rankings.