Optimized cell type signatures revealed from single-cell data by combining principal feature analysis, mutual...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Machine learning is crucial for analyzing single-cell expression data, aiding in tasks like cell annotation and clustering.
• Identifying key genes that differentiate cell populations is challenging yet vital for biological insight.

Purpose of the Study:

• To develop a framework for objective and accurate identification of small gene sets that optimally separate distinct cell phenotypes.
• To enhance the interpretability of machine learning findings in single-cell analysis.

Main Methods:

• Utilized Principal Feature Analysis (PFA) for feature selection, reducing redundancy and identifying informative genes.
• Integrated a Seurat preprocessing tool and a PFA script, employing mutual information for balancing accuracy and gene set size.
• Included a validation component for evaluating gene selection information content using binary and multiclass classification.

Main Results:

• The framework successfully identified small gene subsets (around ten genes) with high information content for separating phenotypes across diverse single-cell datasets.
• Demonstrated explainability in unsupervised learning by revealing cell-type specific signatures.
• Overcame limitations in objectively identifying informative gene sets.

Conclusions:

• The developed framework provides an objective method for selecting minimal yet highly informative gene sets from single-cell data.
• This approach significantly improves the interpretability of cell phenotypes and machine learning outcomes.
• The provided code facilitates the application of this method in biological research.