RNA-seq only generic genes #76
Description
When we compared the correlation between gene percentiles generated by SOPHIE versus the manually curated dataset here, we noticed that there was a group of genes that SOPHIE identified as generic but were not found to be generic using the manually curated dataset.
In this case, SOPHIE was trained on recount2 (RNA-seq) dataset while the manually curated dataset was using array platform.
See #75 for details:
- Overall, it appears that lower expression values in the template (which correspond to RNA-seq only generic genes) are changed more than genes with higher values
- I believe this is likely due to the VAE compression -- possibly the ReLU activation function and/or gaussian constraint in the loss function.
- Why does this compression not affect genes with higher expression values (RNA-seq/array generic genes) as much? This is probably because these values are closer to the mean expression of the compendium. The compression is probably affecting genes on the outliers of the distribution more.
Why is this compression not seen in the array data?
- In the array data, the gene expression for the RNA-seq only genes vs RNA-seq/array genes were similar in the array training compendium.
- Overall the variance in array expression is lower compared to RNA-seq so there isnt' as much compression needed
So genes with low gene expression in the real experiment are getting a boost/increase after going through VAE (simulated experiment) which allows them to be detected as DE.
Possible solutions to consider:
- We don't want lowly expressed genes to get artificially detected as frequently DE.
- Would requiring varying parameters for the activation function and weighting for KL term in the loss function. This will need to be addressed in the future, probably not by this manuscript.
- Rescale RNA-seq data in some way