I have been reading deconvolution method to separate RNA-Seq mixture into different immune cell. It seems that, like this and this discussion metioned, expression count from RNA-Seq data is not indenpende (non-linearly correlated) and this would result in poor accuracy of separation.

However, there are many tools, like TIMER or CIBERSORT, available for such task. Is there a widely recognized way to handle such problem?

By the way, I came across this article stating that TPM value from salmon and kalisto may yield more linear result. Would this solve the non-linearity problem?