Imputing Single-Cell Protein Abundance in Multiplex Tissue Imaging.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning in Medicine

Background:

• Multiplex tissue imaging offers powerful single-cell spatial proteomics and transcriptomics for tissue characterization.
• Current limitations include restricted molecule detection, tissue loss, and protein probe failures, hindering utility.
• Addressing these challenges is crucial for advancing spatial biology and precision medicine.

Purpose of the Study:

• To demonstrate machine learning's capability in imputing single-cell protein abundance using multiplex tissue imaging data.
• To evaluate and compare different machine learning algorithms for imputation accuracy.
• To assess the impact of incorporating cellular spatial information on imputation performance.

Main Methods:

• Comparison of machine learning techniques: regularized linear regression, gradient-boosted regression trees, and deep learning autoencoders.
• Imputation of single-cell protein abundance using multiplex tissue imaging datasets from a breast cancer cohort.
• Integration of cellular spatial information to enhance imputation accuracy.

Main Results:

• Machine learning successfully imputed single-cell protein expression with a mean absolute error between 0.05-0.3 on a [0,1] scale.
• Inclusion of cellular spatial information significantly improved imputation performance.
• Imputed data was utilized to predict biopsy treatment status, demonstrating biological relevance.

Conclusions:

• Machine learning provides a feasible approach for imputing single-cell abundance levels for numerous proteins in multiplex imaging.
• Cellular spatial data integration substantially enhances the accuracy and utility of protein imputation.
• The imputed single-cell protein abundance data holds significant potential for biological applications, such as treatment response prediction.