Score Matching for Differential Abundance Testing of Compositional High-Throughput Sequencing Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Modeling

Background:

• Sparse compositional data, common in high-throughput sequencing, presents challenges due to excess zeros.
• Existing models struggle to incorporate covariate information and handle correlated features effectively.
• Power interaction models offer a flexible framework for compositional data but require extension for covariates.

Purpose of the Study:

• To extend a-b power interaction models to include covariate information for heterogeneous populations.
• To develop a novel differential abundance (DA) testing scheme, cosmoDA, to address correlated features and reduce false positives.
• To evaluate cosmoDA's performance in estimating feature interactions and transformations using simulated and real-world biological data.

Main Methods:

• Extension of a-b power interaction models to incorporate covariate data.
• Development of cosmoDA, a DA testing scheme using penalized generalized score matching.
• Application of cosmoDA to simulated benchmarks, single-cell RNA-Seq, and microbial amplicon data.

Main Results:

• cosmoDA accurately estimates feature interactions in heterogeneous populations.
• The method significantly reduces the false discovery rate when testing differential abundance of correlated features.
• cosmoDA effectively estimates data-adaptive transformations and assesses the impact of zero-handling strategies.

Conclusions:

• cosmoDA provides a robust framework for differential abundance analysis in sparse compositional data with covariates.
• The method improves accuracy and reduces false positives, particularly in the presence of complex feature correlations.
• cosmoDA facilitates better understanding of biological data from single-cell and amplicon sequencing.