Binned multinomial logistic regression for integrative cell-type annotation.

Area of Science:

• Genomics
• Computational Biology
• Biostatistics

Background:

• Cell type annotation is crucial for single-cell genomics analysis but is currently time-intensive and subjective.
• Existing automated methods struggle to unify models across datasets with inconsistent cell type label resolutions.

Purpose of the Study:

• To develop a novel statistical estimator for cell type probability modeling that integrates multiple single-cell genomics datasets with varying label resolutions.
• To provide a unified and automated approach for cell type annotation.

Main Methods:

• Proposed a new multinomial logistic regression estimator.
• Solved a nonconvex optimization problem using a blockwise proximal gradient descent algorithm.
• Applied the method to ten single-cell RNA-seq datasets.

Main Results:

• The proposed method accurately estimates cell type probabilities across diverse scenarios, outperforming existing approaches in simulation studies.
• Demonstrated utility in predicting fine-resolution cell type labels on unlabeled data and refining coarse-resolution annotations.
• Generated novel scientific insights through differential gene expression analysis comparing peripheral blood before and after interferon treatment.

Conclusions:

• The developed method offers a robust and accurate solution for cell type annotation in single-cell genomics.
• Enables the integration of multi-dataset information with varying label granularities for improved biological insights.
• Provides an R package for accessibility and reproducibility.