Best Practices for Handling NA Values in Predicted Gene Expression Matrix

[Yuyu614614]

Hello,

I am using the Predict.py script to generate a gene expression matrix for a cohort. However, the resulting matrix contains some NA values.

Based on my understanding of the code, these NAs likely occur when SNPs required by a gene's prediction model are completely missing from the genotype data for all samples in my cohort, making it impossible to calculate the expression value for that gene.

My central question is: ​​What is the recommended best practice for handling these NA values in the predicted expression matrix for downstream analyses (e.g., association studies)?​​

Specifically, I am weighing two common approaches and would appreciate guidance:

​​Imputation with 0 (or another value):​​ Replacing NA with zero, implicitly assuming that the missing prediction is equivalent to no expression or a baseline level.

​​Deletion:​​ Removing the entire column (gene) that contains any NA values. This is simple but results in loss of data.

[Yuyu614614]

Follow-up Question:​​

I find that, if a single SNP genotype is missing (NA) in the input data, the prediction model for that individual's gene expression will return an NA for that gene.

Does this mean that the genotype data used for predicting gene expression ​​must be fully imputed​​ beforehand, with absolutely no missingness allowed? Is complete, gap-free genotyping a strict prerequisite for these models to generate a valid prediction for each sample?