Shout to extraordinary collaborators and co-authors on this, Karen Conneely, Cameron Russell, Mitchell Machiela, Marios Arvanitis, Janghee Woo.

For more info, I'm presenting this work on Thursday at #ASHG at 1:45pm, Room 210C

To make results easier to parse, we also release a summary statistics "portal" to view our results, including lots of QC metrics: somatic.emory.edu

We also release lots of code for this, which was all done on the DNA Nexus RAP in a cost effective way.

The IGH and IGL mutations in particular had really large effect sizes, so we then did a GWAS of a combined IGH + IGL phenotype, which resulted in a single hit: GRAMD1B. Remarkably, this is well characterized risk locus for CLL, suggesting that we are converging on CLL relevant biology.

We then quantified the variance explained on a liability scale of CH to 30 common aging-related diseases. CH explains far more 'liability' scale variance for hematologic malignancies than other classes (as one would expect), though chronic kidney disease appears high here as well.

New hits are generally indeed "weaker" than most classic CH drivers, suggesting that these are just underneath the "tip of the iceberg". We also replicate the telomere attrition mechanism reported here www.nature.com/articles/s41..., finding that the phenomenon is broader than previously observed

The strongest hits are indeed strongly enriched for canonical CH genes, but we also find non-coding mutations at FGF1, UGT2B7, DGKB, the TERT promoter, chr17 centromere, and immunoglobulin loci:

You can find PRSFNN code here: github.com/weinstockj/PRS . It takes in GWAS summary statistics + LD reference panel + annotations, and we compute the posterior using variational inference to make it fast.

A pleasure to build this with @aprilkim.bsky.social and @alexisbattle.bsky.social

We observed similar non-linear effects with AlphaMissense predictions, where low impact coding variants were prioritized, but highly pathogenic variants were not prioritized (presumably because these are so rare in real GWAS).

More compellingly, it also learns non-linear effects wrt to chromatin accessibility - variants in cCREs present in 10-50 cell types were more highly prioritized than variants in cCRE that are present in numerous (> 50) cell types, suggesting a preference for cell/tissue specificity.

In our annotation curation, we included lots of scATAC, cCREs from ENCODE, conservation from Zoonomia, pathogenicity from AlphaMissense, among others. Generally - PRSFNN "learns" that low-frequency SNPs in accessible chromatin are likely to have larger effect sizes (maybe not that surprising).

We connected the SNP annotations to the parameters of the prior distribution on the weights in a novel way with a neural network, so we're calling it Polygenic Risk Scores with Functional Neural Network (PRSFNN). We were excited to see that PRSFNN does well in benchmarks (at least in our hands).

Congratulations!