False signals induced by single-cell imputation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution gene expression analysis.
• scRNA-seq data analysis faces challenges with abundant zero values, indicating missing data or no expression.
• Existing imputation methods often rely on inherent data structures, limiting their information gain and validity.

Purpose of the Study:

• To evaluate the risk of false positives and irreproducible differential expression introduced by various imputation methods for scRNA-seq data.
• To compare the performance of six different imputation techniques on simulated and real scRNA-seq datasets.
• To assess the impact of imputation on the reproducibility of cell-type specific markers across different sequencing technologies.

Main Methods:

• Six imputation methods were applied to simulated and permuted real scRNA-seq datasets.
• Evaluated false positive gene-gene correlations and differentially expressed genes.
• Assessed marker reproducibility using matched 10X and Smart-seq2 data before and after imputation.

Main Results:

• Imputation methods varied significantly in their introduction of false positives; smoothing methods (MAGIC, knn-smooth, dca) generated numerous false positives.
• Model-based methods produced fewer false positives, contingent on cell-type diversity.
• All imputation methods reduced cell-type marker reproducibility, though large effect size markers were more robust.

Conclusions:

• Imputation in scRNA-seq analysis can create circularity, leading to false-positive findings and necessitating cautious interpretation of statistical tests.
• Filtering by effect size can mitigate, but not eliminate, imputation-induced biases.
• SAVER demonstrated the lowest risk of false or irreproducible results, making it the recommended imputation method.