cn_validate_truth: sensitivity by carrier frequency (rare < 5 % is the primary metric, K-027)

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01814GKVzZVL3FuF3DeiZXhv

scripts/v2/cn_tune_report.py

@@ -55,6 +55,8 @@ def main() -> int:
     val = json.loads((run / "val" / "summary.json").read_text())
     for t, bins in val["sensitivity"].items():
         print(f"  sensitivity {t}: " + ", ".join(f"{b} {x['found']}/{x['truth']}" for b, x in sorted(bins.items())))
+    for t, bins in val.get("sensitivity_by_carrier_frequency", {}).items():
+        print(f"  by frequency {t}: " + ", ".join(f"{b} {x['found']}/{x['truth']}" for b, x in sorted(bins.items())))
     return 0
 
 

scripts/v2/cn_validate_truth.py

@@ -6,7 +6,9 @@ structural-variant callset (1KGP_3202 GATK-SV/svtools freeze V3) and between rep
 
 ``truth.tsv``: ``chrom start end svtype sample`` for non-reference DEL/DUP genotypes of the 120 individuals
 (``bcftools query``; chr prefix removed). Matching: reciprocal overlap ≥ 50 % and the same class (DEL ↔ losses
-incl. mosaic, DUP ↔ gains). Only truth events with ≥ 10 array probes are counted as detectable.
+incl. mosaic, DUP ↔ gains). Only truth events with ≥ 10 array probes are counted as detectable. Sensitivity is also
+split by the event's carrier frequency among the individuals in ``truth.tsv`` (identical chrom/start/end/type):
+common events are normalised away by a median-LRR batch reference (K-027), so ``rare`` (< 5 %) is the primary metric.
 """
 
 from __future__ import annotations
@@ -22,6 +24,7 @@ import numpy as np
 
 from simurg.formats.bpm import read_bpm
 
+FREQ = [(0.0, 0.05, "rare<5%"), (0.05, 0.2, "5-20%"), (0.2, 1.01, "common>=20%")]
 BINS = [(25_000, 50_000), (50_000, 100_000), (100_000, 500_000), (500_000, 1_000_000), (1_000_000, 10**9)]
 CLASS = {"DEL": "DEL", "MDEL": "DEL", "DUP": "DUP", "MDUP": "DUP"}
 
@@ -70,15 +73,23 @@ def main() -> int:
         ign = set(a.ignore_filter) | {"PASS"}
         calls[s["sample_id"]] = [e for e in s["events"] if e["quality"] >= a.min_qual and set(e["filters"]) <= ign]
     truth: dict[str, list[tuple[str, int, int, str]]] = defaultdict(list)
+    carriers: dict[tuple[str, int, int, str], set[str]] = defaultdict(set)
     with a.truth.open() as fh:
         for line in fh:
             c, s_, e_, t, ind = line.rstrip("\n").split("\t")
             c = c[3:] if c.startswith("chr") else c
+            carriers[(c, int(s_), int(e_), t)].add(ind)
             if t in ("DEL", "DUP") and int(e_) - int(s_) + 1 >= 25_000:
                 truth[ind].append((c, int(s_), int(e_), t))
+    n_ind = len({i for v in carriers.values() for i in v}) or 1
+
+    def freq_class(key: tuple[str, int, int, str]) -> str:
+        f = len(carriers[key]) / n_ind
+        return next(name for lo, hi, name in FREQ if lo <= f < hi)
 
     # sensitivity per class and size bin
     sens: dict[str, dict[str, list[int]]] = defaultdict(lambda: defaultdict(lambda: [0, 0]))
+    sens_freq: dict[str, dict[str, list[int]]] = defaultdict(lambda: defaultdict(lambda: [0, 0]))
     prec = defaultdict(lambda: [0, 0])
     per_sample = []
     for sid, evs in calls.items():
@@ -93,6 +104,9 @@ def main() -> int:
             cell = sens[t][bin_of(e_ - s_ + 1)]
             cell[0] += ok
             cell[1] += 1
+            fc = sens_freq[t][freq_class((c, s_, e_, t))]
+            fc[0] += ok
+            fc[1] += 1
             hit_t += ok
         hit_c = 0
         cnv_calls = [e for e in evs if e["call"] in CLASS]
@@ -149,6 +163,10 @@ def main() -> int:
             t: {b: {"found": v[0], "truth": v[1], "rate": rate(v)} for b, v in sorted(d.items())}
             for t, d in sens.items()
         },
+        "sensitivity_by_carrier_frequency": {
+            t: {b: {"found": v[0], "truth": v[1], "rate": rate(v)} for b, v in sorted(d.items())}
+            for t, d in sens_freq.items()
+        },
         "precision": {t: {"in_truth": v[0], "calls": v[1], "rate": rate(v)} for t, v in prec.items()},
         "replicate_reproducibility": {
             k: {"matched": v[0], "comparisons": v[1], "rate": rate(v)} for k, v in rep.items()