Multiresolution categorical regression for interpretable cell-type annotation.

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Categorical response regression models often feature multi-resolution structures where categories can be grouped into coarser levels.
• Understanding how predictors influence probabilities at different category resolutions is crucial for accurate modeling.

Purpose of the Study:

• To propose a unified, data-driven method for fitting high-dimensional multinomial logistic regression models with multi-resolution response categories.
• To enable the identification of predictors that are relevant at specific category resolutions (coarse vs. fine) or irrelevant.

Main Methods:

• Developed a novel method for fitting multinomial logistic regression models that explicitly accounts for the multi-resolution structure of response categories.
• Proposed a scalable algorithm applicable to both overlapping and non-overlapping coarse category definitions.
• The method is designed to leverage the inherent multi-resolution structure for improved statistical properties.

Main Results:

• The proposed method successfully distinguishes predictors based on their impact on coarse versus fine categories.
• Statistical analysis demonstrates that the method effectively utilizes the multi-resolution structure, outperforming existing estimators.
• Application to cell-type annotation using gene expression data yielded novel biological insights.

Conclusions:

• The developed method provides a powerful tool for analyzing complex categorical data with inherent hierarchical structures.
• It offers a data-driven approach to identify predictor relevance at varying levels of category granularity.
• The findings have significant implications for improving cell-type annotation methodologies in biological research.