From 6d48d6137c15654c3289add19a8a0960130aa833 Mon Sep 17 00:00:00 2001
From: Daniel Nachun <daniel.nachun@gmail.com>
Date: Fri, 4 Sep 2026 11:09:46 -0700
Subject: [PATCH] fix variant normalization

---
 R/sumstatsQc.R                   | 17 ++++++++++++++---
 tests/testthat/test_sumstatsQc.R | 25 +++++++++++++++++++++++++
 2 files changed, 39 insertions(+), 3 deletions(-)

diff --git a/R/sumstatsQc.R b/R/sumstatsQc.R
index 1e907463..3ad17b83 100644
--- a/R/sumstatsQc.R
+++ b/R/sumstatsQc.R
@@ -4412,10 +4412,21 @@ krigingOutlierQc <- function(
     if (length(ids) == 0L) {
         return(ldSketch)
     }
-    keep <- is_in(
-        normalizeVariantId(.ldSketchMatchIds(ldSketch)),
-        normalizeVariantId(unique(ids))
+    # Match on the (chrom, pos, allele) tuple rather than the raw id string.
+    # The entry ids have just been canonicalized to chr:pos:A2:A1 by the QC
+    # above, while panel ids are passed through verbatim by
+    # `.repairVariantIds()` whenever their allele fields are already valid DNA
+    # -- so a panel keyed chr:pos:A1:A2 (which is how a PLINK .bim writes ids;
+    # a .pvar writes REF:ALT and is therefore already canonical) matches
+    # nothing under string equality and the panel is silently emptied. This is
+    # the matcher `.ldFromSketchMatch()` and every other LD lookup already use.
+    panelIds <- .ldSketchMatchIds(ldSketch)
+    matched <- matchVariants(
+        panelIds,
+        unique(ids),
+        removeStrandAmbiguous = FALSE
     )
+    keep <- is_in(seq_along(panelIds), matched$idxA)
     .ldSketchSubset(ldSketch, keep)
 }
 
diff --git a/tests/testthat/test_sumstatsQc.R b/tests/testthat/test_sumstatsQc.R
index f06345e1..51bb06ed 100644
--- a/tests/testthat/test_sumstatsQc.R
+++ b/tests/testthat/test_sumstatsQc.R
@@ -3324,6 +3324,31 @@ test_that("a tag-named panel entry survives the end-to-end sketch subset", {
     )
 })
 
+test_that("the sketch survives a panel keyed chr:pos:A1:A2 (PLINK .bim order)", {
+    # Pre-fix regression: harmonization canonicalizes the entry ids to
+    # chr:pos:A2:A1, while panel ids are passed through verbatim whenever their
+    # allele fields are already valid DNA. A panel keyed the other way round --
+    # which is how a PLINK .bim writes ids, whereas a .pvar writes REF:ALT and
+    # is already canonical -- matched nothing under string equality, so
+    # .subsetSketchToIds() silently emptied the sketch and RSS fine-mapping /
+    # TWAS weights aborted later with "not present in the LD sketch panel".
+    h <- .ssQ_makeHandleVid()
+    si <- getSnpInfo(h)
+    h@snpInfo$SNP <- paste0("chr1:", si$BP, ":", si$A1, ":", si$A2)
+    ss <- GwasSumStats(
+        study = "g1",
+        entry = list(.ssQ_makeEntryGr(
+            snp_ids = paste0("chr1:", c(100L, 200L, 300L, 400L), ":G:A")
+        )),
+        genome = "hg19",
+        ldSketch = h
+    )
+    out <- summaryStatsQc(ss, pipCutoffToSkip = 0, nCutoff = 0)
+    kept <- pecotmr:::.ldSketchMatchIds(getLdSketch(out))
+    expect_length(kept, 4L)
+    expect_setequal(pecotmr:::parseVariantId(kept)$pos, c(100L, 200L, 300L, 400L))
+})
+
 test_that("summaryStatsQc: slalom z-mismatch resolves sign-flipped variants against the panel", {
     # Pre-fix regression: a sign-flipped variant kept its input-orientation SNP,
     # which is absent from the panel, so .applyLdMismatchQcToEntry errored with