Overcoming missing data in spatial metabolomics with machine learning imputation to accelerate downstream discovery.

Area of Science:

• Biomedical data analysis
• Computational biology
• Metabolomics

Background:

• Mass spectrometry imaging (MSI)-based spatial metabolomics generates large datasets with significant missing values.
• Current imputation strategies lack comprehensive evaluation for their impact on spatial analysis accuracy.

Purpose of the Study:

• To systematically evaluate eight imputation methods for spatial metabolomics data.
• To identify optimal imputation strategies balancing accuracy and spatial structure preservation.

Main Methods:

• Developed an evaluation framework assessing imputation accuracy and spatial cluster structure preservation.
• Tested methods on six diverse benchmark datasets (mouse brain/liver, human kidney/stomach, plant seeds) with simulated missing values.
• Compared existing methods with a novel graph convolutional network (GCN)-based approach.

Main Results:

• Random Forest (RF) demonstrated superior overall performance across both imputation accuracy and spatial cluster preservation.
• Graph Convolutional Network (GCN) ranked second, showing strong performance in both evaluation dimensions.
• The study provides a robust comparison of imputation techniques tailored for spatial metabolomics.

Conclusions:

• RF and GCN are recommended as highly effective imputation methods for spatial metabolomics.
• This benchmark study offers critical guidance for researchers in selecting appropriate imputation strategies to enhance data analysis and interpretation.