Table 2 Network variations used for technical assessment. NB: for all networks undefined correlations are set to 0
From: Exploiting single-cell expression to characterize co-expression replicability
Network | Method | Property tested |
|---|---|---|
UMI | • Spearman correlation of UMI data to make a network for each batch • Batch networks are rank standardized then aggregated | Do UMI expression estimates produce functional co-expression? |
CPM | • Spearman correlation of CPM normalized data to make a network for each batch • Batch networks are rank standardized then aggregated | What types of artifacts can sample standardization introduce? |
Batch-affected | • Spearman correlation across all samples using UMI data • Rank standardization | What impact does co-variation across batches have? |
Binary expression | • All non-zero values are set to 1 • Spearman correlation to make a network for each batch • Batch networks are rank standardized then aggregated | How informative is gene representation? |
Combat | • UMI data is log2 transformed then Combat is run for each celltype (ChC and Pv) • Spearman correlation to make a network for each cell type • Aggregate is made from the addition of rank-standardized ChC and Pv networks | Do methods for removing batch effects alter co-expression? |
Removal of unwanted variation (RUV) | • UMI data is log2 transformed then RUV is run for each cell type (ChC and Pv) using ERCC spike-ins as control genes • Spearman correlation to make a network for each cell type • Aggregate is made from the addition of rank-standardized ChC and Pv networks | What are the combined influence of batch correction and ERCC-based normalization? |
UMI excluding zeroes | • All zeroes are set to NA • Networks are made for each batch using pairwise Spearman correlation • Batch networks are rank standardized then aggregated | How does removing zeroes alter network topology and performance? |