Machine learning for cell classification and neighborhood analysis in glioma tissue

Area of Science:

• Computational Biology
• Pathology
• Machine Learning

Background:

• Accurate cell classification is essential for analyzing tissue heterogeneity and cell organization in multiplexed, spatially resolved single-cell studies.
• Variations in sample handling and staining can lead to inconsistent marker profiles for identical cell types, complicating classification.
• Downstream analyses like cell counting, interaction quantification, and disease-specific niche identification rely heavily on precise cell classification.

Purpose of the Study:

• To develop and validate a robust, automated methodology for cell classification using multiplexed immunofluorescence data from low-grade gliomas.
• To improve the accuracy and reproducibility of cell classification by employing machine learning architectures resilient to staining and scanning variations.
• To leverage accurate cell classification for the unsupervised identification of specialized tissue niches within tumor microenvironments.

Main Methods:

• Utilized two distinct machine learning architectures for cell classification on multiplexed immunofluorescence data.
• Incorporated illumination-invariant features to enhance robustness against variations in sample handling and scanning artifacts.
• Developed a confidence measure for automated cell classification decisions, allowing for the exclusion of low-confidence predictions.
• Applied an unsupervised learning approach based on graph neural networks to identify cell niches using the classified cell data.

Main Results:

• Achieved an initial cell classification accuracy of 83.1% without requiring sample standardization.
• By excluding low-confidence classifications using the developed measure, the accuracy was improved to 94.5%.
• Successfully re-detected known specialized tissue niches in previously published datasets using the unsupervised niche identification approach.
• Demonstrated that the proposed cell classification method can lead to refined niche definitions potentially relevant for glioma sub-groups.

Conclusions:

• The presented automated machine learning methodology offers a robust solution for cell classification in complex multiplexed immunofluorescence data.
• The confidence measure significantly enhances classification accuracy, making the approach more reliable for downstream biological interpretation.
• The integration of cell classification with graph neural network-based niche identification provides a powerful framework for exploring tumor microenvironment heterogeneity in gliomas.