Label Expression Value Problem

[oooo111]

Hello, author. During the calculation of the label, some cells have most of their gene expression values as 0. For example, when calculating onek1 and the B cells in GTEx, after taking the average expression across all individuals, only a few thousand genes have non-zero expression values. How do you handle these cases? Do you process them by directly sorting based on rank? If so, this might lead to different rankings for the same 0 values, resulting in different y values for those identical 0 values. Thank you!

[Yichao-Zhou]

Yes, we directly converted the zeros into rank-based values. But for most cell types and datasets I tested, the majority of genes have non-zero expressions after average across people. We discussed these settings for a long time, and we didn't find other options that can significantly solve the problem of sparsity. In one cell type/one model, there will not be different rankings for the same 0 values.

[Yichao-Zhou]

In the scPrediXcan settings, even though ctPred may give the same predictions for zero-expression genes, it is not necessarily mean that l-ctPred will give the same predictions for different individuals (the prediction accuracy may not be as high as those highly expressed genes, statistically). In the scPrediXcan/TWAS framework, it is the differences of individual genetic-regulated gene expression predictions with different genotypes lead to the potential findings of candidate causal genes.