PRISM-models / README.md
NotSoSharpDressedMan's picture
README: concise, key fixes per cohort
f929b7e verified
|
Raw History Blame Contribute Delete
3.42 kB
---
pretty_name: PRISM protein imputation models
viewer: false
tags:
- genomics
- proteomics
- pqtl
- pwas
- twas
- uk-biobank
- fenland
- decode
---
# PRISM-models
Protein imputation models (cis + trans) for protein-wide association studies with GWAS summary
statistics. One file per protein, holding the SNP weights, the per-SNP association table and the LD matrix.
| path | cohort | models | version |
|---|---|---|---|
| `UKB-EUR/CisTrans/<protein>.RData` | UK Biobank Pharma Proteomics (Olink), European ancestry | 2,877 | 2026-10-09 |
| `Fenland/CisTrans/<gene>_<SeqId>.RData` | Fenland (SomaScan) | 3,785 | 2026-09-18 |
| `deCODE/CisTrans/<deCODE file>.RData` | deCODE (SomaScan) | 4,708 | 2026-10-09 |
## What this version fixes
**UKB-EUR (2026-10-09)**
- **Effect allele corrected.** For variants with MAF near 0.5, `A1` named the other allele of `BETA`. The
labels came from a batch-restricted frequency run, while the GWAS used the full sample. All trans loci
were retrained, and the inflated R² values (some above 1) are gone.
- **GRCh38 positions added** to `ss` as `POS.38`.
- **A trans locus that had silently failed to train** (PPP1R12A locus 2) is now included.
- Trans loci are capped at 3 Mb, and positions use a repaired GRCh38 rsID map.
**deCODE (2026-10-09)**
- **Strand flips corrected.** deCODE alleles are on the GRCh38 strand and the LD panel is on GRCh37, so
in reverse-oriented regions some SNPs got the wrong sign. Alleles are now complemented before alignment,
and the affected trans loci were retrained.
- Trans loci are capped at 3 Mb.
**Fenland (2026-09-18, first release)**
- Indels are removed, and multi-gene assays use the union of all their genes' cis windows, across
chromosomes. Trans loci are capped at 3 Mb.
## How the models were built
Cis = target gene ± 1 Mb; trans = Bonferroni-significant SNPs outside cis (MHC removed), clumped and
merged into loci. Each locus is trained on summary statistics as a SUMMIT ensemble (ElNet, LASSO, MCP,
MNet, SCAD over six LD-clumping levels, PUMAS subsampling), with LD from 49,341 UK Biobank EUR
individuals. The cis locus and all trans loci of a protein are then merged into one model.
- **UKB-EUR:** trans loci from our own GWAS on individual-level UKB-PPP data; cis from the BLISS UKB models.
- **Fenland:** trans and cis loci trained from the published Fenland summary statistics (N = 10,708).
- **deCODE:** trans loci from the published deCODE summary statistics; cis from the BLISS deCODE models.
## File contents
| object | content |
|---|---|
| `weight` | data.frame with one column, `weight`; rownames = SNP |
| `ss` | `SNP CHR BP BETA SE P N A1 A2 MAF CIS LOCI R2` (UKB-EUR also has `POS.38`) |
| `LD.ut`, `LD.snp` | upper triangle of the LD matrix, and its SNP names |
| `CHECKED`, `R2.PUMAS`, `stats`, `VALIDATION` | training diagnostics |
- `A1` is the effect allele of `BETA` and `weight`.
- `BP` is GRCh37 for UKB-EUR and Fenland, and GRCh38 for deCODE.
- `LOCI` is 0 for cis and ≥ 1 for a trans locus.
- `R2` is the locus's PUMAS R², so a model's total is `sum(unique(ss$R2))`.
## Loading (R)
Only the upper triangle of the LD matrix is stored; `load-model-slim.r` rebuilds the full matrix.
```r
source("load-model-slim.r")
m <- load_model_slim("UKB-EUR/CisTrans/A1BG.RData") # m$weight, m$ss, m$matrix.LD
```
`MANIFEST.md5` lists the md5 of every model file (`md5sum -c MANIFEST.md5`).