A unified framework for selecting and evaluating cell-type-specific gene co-expressions in single-cell data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene co-expression networks are crucial for understanding gene relationships within specific cell types.
• Current methods for inferring these networks from single-cell data often lack robust false positive control, potentially leading to inaccurate biological conclusions.
• Existing evaluation metrics may not accurately reflect method performance without proper control for false positives.

Purpose of the Study:

• To systematically compare p-values versus correlation strength for selecting gene pairs in single-cell co-expression network inference.
• To develop and validate a simulation method for generating empirical p-values for benchmarking co-expression estimation methods.
• To establish rigorous standards for evaluating and comparing gene co-expression network inference methods.

Main Methods:

• Systematic comparison of p-value and correlation strength criteria for selecting correlated gene pairs from single-cell data.
• Extension and validation of a simulation method to generate empirical p-values for co-expression estimation.
• Development of adjusted comparison strategies accounting for varying gene pair counts and expression-level biases.

Main Results:

• P-value-based selection of gene pairs is more robust than correlation strength for identifying meaningful relationships.
• The validated simulation method reliably generates empirical p-values for benchmarking.
• Fair comparison necessitates adjusting for the number of identified gene pairs and inherent biases in ground truth networks.

Conclusions:

• P-values provide a more reliable criterion for selecting gene pairs in single-cell co-expression network analysis.
• The developed simulation framework enhances the benchmarking of co-expression estimation methods.
• This study provides practical guidance for reliable gene pair selection and sets higher standards for evaluating network inference methods.