Fluoro-forest: a random forest workflow for cell type annotation in high-dimensional immunofluorescence imaging with...

Area of Science:

• Computational biology
• Bioinformatics
• High-resolution imaging analysis

Background:

• Cyclic immunofluorescence (IF) enables detailed cell phenotyping and tissue organization analysis.
• Current workflows often use unsupervised clustering and cluster-level annotation, lacking statistical rigor and potentially leading to misclassification.
• Marker expression averages in existing methods lack statistical validation for cell type assignment.

Purpose of the Study:

• To develop an end-to-end pipeline for accurate cell type annotation in complex IF data.
• To improve upon existing methods by incorporating statistical evaluation and a semi-supervised approach.
• To provide a tool that enhances classification accuracy and interpretability in high-dimensional IF datasets.

Main Methods:

• Implementation of a semi-supervised, random forest-based pipeline for cell type prediction.
• Utilizing cluster-based sampling for efficient training data generation.
• Integrating downstream visualization for enhanced interpretability of cell annotations.

Main Results:

• The Fluoro-forest pipeline demonstrates higher accuracy in cell annotation compared to deep learning and probabilistic methods.
• Achieved superior classification with a training dataset comprising less than 5% of the total cells.
• The pipeline provides cell type probabilities and performance metrics for user assessment.

Conclusions:

• The proposed Fluoro-forest pipeline offers a statistically robust and accurate method for cell type annotation in cyclic IF.
• It significantly improves classification accuracy, even with limited training data, outperforming existing approaches.
• The tool is freely available, facilitating its adoption in complex IF data analysis workflows.