Prior-guided factorization for reliable imputation of scRNA-seq data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding biological complexity.
• Technical dropout in scRNA-seq obscures true gene expression.
• Existing imputation methods struggle to differentiate technical noise from biological zeros.

Purpose of the Study:

• To develop a novel imputation framework, scZN, for scRNA-seq data.
• To accurately reconstruct cellular expression landscapes by addressing dropout.
• To improve downstream scRNA-seq analyses.

Main Methods:

• scZN models scRNA-seq data as a combination of a two-state transcription process and dropout.
• Imputation is formulated as nonnegative factorization of the raw count matrix.
• Prior knowledge and multiple regularizations guide learning and optimization.

Main Results:

• scZN accurately captures gene and cell expression distributions.
• The method effectively suppresses spurious gene activation.
• scZN outperforms existing state-of-the-art imputation methods across multiple datasets.
• Improved trajectory inference for stem cells and dentate gyrus data.
• Enhanced recovery of neuroinflammation pathways in Alzheimer's disease data.

Conclusions:

• scZN offers a unified framework for accurate and interpretable missing-value imputation in scRNA-seq.
• The framework enhances the reliability of scRNA-seq data for biological discovery.
• scZN significantly improves the performance of downstream analyses, including trajectory inference and disease pathway identification.