Naught all zeros in sequence count data are the same.

Area of Science:

• Genomics
• Bioinformatics
• Statistical Modeling

Background:

• Genomic studies generate multivariate count data, frequently with excess zeros.
• These zeros can introduce artifacts in statistical analyses, necessitating specialized modeling approaches.
• Existing zero-handling models may yield divergent results when identifying differentially expressed sequences.

Purpose of the Study:

• To evaluate the performance of different zero-handling models on gene-expression and microbiome datasets.
• To develop a conceptual framework for understanding zero-generating processes in sequence count data.
• To provide guidelines for selecting appropriate models for sparse sequence count data analysis.

Main Methods:

• Application of various zero-handling models to real-world gene-expression and microbiome data.
• Development of a four-process conceptual framework for zero generation.
• Simulation studies to assess model behavior under different zero-generating conditions.

Main Results:

• Different zero-handling models showed substantial disagreement in identifying differentially expressed sequences.
• Simple count models proved adequate across various zero-generating processes, even when unknown.
• Zero-inflation models were found to be suitable only under specific, less common biological and experimental conditions.

Conclusions:

• The choice of zero-handling model significantly impacts the identification of differentially expressed sequences.
• Simple count models offer a robust and often sufficient approach for analyzing sparse sequence count data.
• Guidelines are proposed for selecting optimal models, emphasizing the limitations of zero-inflation in many contexts.