Can genetic score based on predicted gene expressions offer better prediction than polygenic scores?

Short answer: no (at least not with the existing eQTL datasets, which are based on bulk-tissue RNAseq)

A neat exploration by @ollie_pain et al https://twitter.com/biorxiv_genomic/status/1334309617430622209

In rheumatoid arthritis (RA) combining PRS and transcriptome score offered slight improvement, but RA is an outlier here, with 70% of its heritability originating from HLA locus. https://twitter.com/doctorveera/status/1329848753499594752?s=20

A common assumption of such studies is that the risk variants mechanism goes through gene expression and constructing scores using gene expression, which is more proximal to disease, will offer better prediction.

But often the scenario often tend to be like this , where gene expression is a peripheral correlate and is more distal from the disease. As a result gene expression based scores perform poorer than SNP based scores

I'll give you one example from my PhD project. I conducted TWAS for ADHD and ADHD symptoms and compared the genetic correlation between ADHD and ADHD symptoms with transcriptomic correlation (based on imputed gene expressions)

The genetic correlation between ADHD and ADHD symptoms was 1, but the transcriptomic correlation was zero. I still keep wondering why.
This is a slide from my PhD defence which @cathrynlewis might remember; she was one my opponents :)

I used to be an eQTL/TWAS fanatic, but my enthusiasm has decreased over years as accumulating evidence suggests that existing eQTLs (particularly for brain tissues) are simply a drop of true signal mixed in a big bowl of noise.

Cell type specific and developmental stage specific eQTLs might solve these issues in the future. Many people are doing great work in single cell eQTLs and I am looking forward to how these studies will change our understanding of the role eQTLs in human diseases and traits.