CN: MF split normalised by per-probe reference BAF noise; genotype-based complete loss (CYTO-041)

The mosaic BAF split is now estimated over heterozygous calls against each probe's expected noise (the
reference heterozygous spread scaled to the sample), inverting the median of |N(+-d, sigma)|. A mosaic loss
lacking the heterozygous calls its PFB predicts gets MF 1 and is promoted (CYTO-020).

SIMURGARRAY_FUNCTIONAL_SPEC.md

@@ -2437,13 +2437,15 @@ about **algorithms or numerics**, the plan wins. Each such conflict found during
 
 - **CYTO-041** `§10.3 · v2` **MF refinements [OURS].** Two refinements of CYTO-018, from real-data validation
   (docs/validation/V2_cn.md, K-027):
-  - *noise-corrected split:* the median |BAF − 0.5| of unsplit heterozygotes is 0.6745·σ_b (σ_b = the sample's
+  - *noise-corrected split:* per-probe expected noise σ_i = the reference's heterozygous BAF spread of the probe ×
-    heterozygous BAF SD), so `d = √max(0, d_obs² − (0.6745·σ_b)²)`. Without this, noise alone gave MF ≈ 0.2 in
+    the sample's scale (median |BAF − 0.5| / spread over autosomal heterozygous calls, ÷ 0.6745; the sample's
-    diploid regions and waves were called as mosaic losses.
+    heterozygous BAF SD without a reference spread). Over the event's heterozygous calls, m = median(|BAF − 0.5| /
-  - *complete loss:* a mosaic loss with ≥ 20 finite-BAF probes, ≥ 5 expected heterozygotes (Σ 2p(1−p) over PFB p),
+    σ_i); the split in σ units δ solves Φ(m − δ) − Φ(−m − δ) = ½ (δ = 0 when m ≤ 0.6745) and d = median(σ_i)·δ.
-    and fewer than 0.25 × expected probes with 0.1 < BAF < 0.9 has MF = 1. With the default
+    Without this, noise alone gave MF ≈ 0.1–0.3 and genomic waves were called as mosaic losses.
-    `--max-mosaic-fraction 1.0` it is therefore promoted (CYTO-020) to a constitutional DEL with `HIGHFRACTION`.
+  - *complete loss:* a mosaic loss whose called genotypes hold fewer than a quarter of the heterozygotes the PFB
-    This catches constitutional deletions whose LRR is compressed towards the mosaic states.
+    predicts (Σ 2p(1−p) ≥ 3) has MF = 1. With the default `--max-mosaic-fraction 1.0` it is therefore promoted
+    (CYTO-020) to a constitutional DEL with `HIGHFRACTION`. This catches constitutional deletions whose LRR is
+    compressed towards the mosaic states.
 
 ---
 

src/simurg/cn/events.py

@@ -20,8 +20,19 @@ F64 = npt.NDArray[np.float64]
 I64 = npt.NDArray[np.int64]
 GAP_FACTOR = 150
 MIN_MF_PROBES = 20
-LOH_EXPECTED_HETS = 5.0  # CYTO-041: expected heterozygous probes needed to call a loss complete
+LOSS_EXPECTED_HETS = 3.0  # CYTO-041: expected heterozygous calls needed to call a loss complete
-LOH_HET_RATIO = 0.25  # CYTO-041: observed / expected heterozygotes below which a loss is complete
+LOSS_HET_RATIO = 0.25  # CYTO-041: observed / expected heterozygous calls below which a loss is complete
+
+
+def complete_loss(geno: npt.NDArray[np.uint8], pfb: F64) -> bool:
+    """CYTO-041 [OURS]: a loss is complete when its called genotypes lack the heterozygotes the PFB predicts
+    (Σ 2p(1 − p) ≥ 3 expected, fewer than a quarter observed); a mosaic loss keeps its heterozygotes."""
+    called = geno > 0
+    p = np.clip(np.where(np.isfinite(pfb), pfb, 0.5), 0.0, 1.0)[called]
+    expected = float((2 * p * (1 - p)).sum())
+    return expected >= LOSS_EXPECTED_HETS and float((geno[called] == 2).sum()) < LOSS_HET_RATIO * expected
+
+
 TYPES = ("del", "dup", "loh", "gainloh", "mdel", "mdup", "mloh")
 DEFAULTS = {"del": (10, 25_000), "dup": (10, 50_000), "loh": (500, 3_000_000), "gainloh": (300, 500_000),
             "mdel": (10, 25_000), "mdup": (10, 50_000), "mloh": (500, 3_000_000)}  # fmt: skip
@@ -117,21 +128,39 @@ def extract(chrom: str, path: npt.NDArray[np.int32], states: list[State], pos: I
     return out
 
 
-def _mf_from_baf(call: str, baf: F64, baf_sd: float = 0.0, pfb: F64 | None = None) -> float | None:
+def _phi(x: float) -> float:
-    """CYTO-018 with the CYTO-041 refinements [OURS]: the split is corrected for BAF noise (the median |BAF − 0.5| of
+    return 0.5 * (1.0 + math.erf(x / math.sqrt(2.0)))
-    unsplit heterozygotes is 0.6745·σ_b), and a loss whose expected heterozygotes are absent is complete (MF 1)."""
+
-    fin = np.isfinite(baf)
+
-    mid = fin & (baf > 0.1) & (baf < 0.9)
+def _split_from_median(m: float) -> float:
-    b = baf[mid]
+    """δ ≥ 0 with median |N(±δ, 1)| = m (bisection; 0 when m ≤ 0.6745)."""
-    if call == "MDEL" and pfb is not None and fin.sum() >= MIN_MF_PROBES:
+    if _phi(m) - _phi(-m) <= 0.5:
-        p = np.clip(np.where(np.isfinite(pfb), pfb, 0.5), 0.0, 1.0)[fin]
+        return 0.0
-        expected = float((2 * p * (1 - p)).sum())
+    lo, hi = 0.0, m + 10.0
-        if expected >= LOH_EXPECTED_HETS and len(b) < LOH_HET_RATIO * expected:
+    for _ in range(60):
-            return 1.0
+        mid = 0.5 * (lo + hi)
+        if _phi(m - mid) - _phi(-m - mid) > 0.5:
+            lo = mid
+        else:
+            hi = mid
+    return 0.5 * (lo + hi)
+
+
+def _mf_from_baf(call: str, baf: F64, sigma: F64 | float = 0.0) -> float | None:
+    """CYTO-018 with the CYTO-041 noise correction [OURS]. ``baf``: heterozygous-call BAF of the event; ``sigma``:
+    each probe's expected BAF noise (reference heterozygous spread × the sample's scale). With m = median(|BAF − 0.5|
+    / σ), the split δ (in σ units) solves Φ(m − δ) − Φ(−m − δ) = ½ (the median of |N(±δ, 1)|; δ = 0 when
+    m ≤ 0.6745, i.e. no more spread than noise), and d = median(σ)·δ."""
+    sig = np.broadcast_to(np.asarray(sigma, dtype=np.float64), baf.shape)
+    ok = np.isfinite(baf) & (baf > 0.1) & (baf < 0.9) & np.isfinite(sig)
+    b, sg = baf[ok], sig[ok]
     if len(b) < MIN_MF_PROBES:
         return None
-    d_obs = float(np.median(np.abs(b - 0.5)))
+    if float(np.median(sg)) > 0:
-    d = math.sqrt(max(0.0, d_obs**2 - (0.6745 * baf_sd) ** 2))
+        sg = np.maximum(sg, 1e-3)
+        d = float(np.median(sg)) * _split_from_median(float(np.median(np.abs(b - 0.5) / sg)))
+    else:
+        d = float(np.median(np.abs(b - 0.5)))
     if call == "MDEL":
         f = 2 * d / (0.5 + d)
     elif call == "MDUP":
@@ -151,7 +180,7 @@ def _mf_from_lrr(call: str, lrr: F64, gamma: float) -> float | None:
 
 def annotate_stats(
     ev: Event, pos: I64, lrr: F64, baf: F64, geno: npt.NDArray[np.uint8], post: F64, states: list[State], m: float,
-    gamma: float, haploid: bool, baf_sd: float = 0.0, pfb: F64 | None = None,
+    gamma: float, haploid: bool, sigma: F64 | None = None, pfb: F64 | None = None,
 ) -> None:  # fmt: skip
     """CYTO-010/012/018/027 statistics in place (``lrr`` unsmoothed)."""
     sl = slice(ev.i0, ev.i1)
@@ -170,7 +199,11 @@ def annotate_stats(
     p = float(post[sl][:, cls].sum(axis=1).mean()) if cls else 0.0
     ev.quality = 60.0 if p >= 1.0 else float(min(60, max(0, round(-10 * math.log10(max(1e-300, 1 - p))))))
     if ev.mosaic:
-        mf = None if haploid else _mf_from_baf(ev.call, baf[sl], baf_sd, None if pfb is None else pfb[sl])
+        if not haploid and ev.call == "MDEL" and pfb is not None and complete_loss(g, pfb[sl]):
+            ev.mosaic_fraction = 1.0
+            return
+        hb = np.where(g == 2, baf[sl], np.nan) if sigma is not None else baf[sl]
+        mf = None if haploid else _mf_from_baf(ev.call, hb, 0.0 if sigma is None else sigma[sl])
         ev.mosaic_fraction = mf if mf is not None else _mf_from_lrr(ev.call, x, gamma)
 
 

src/simurg/pipeline/cn.py

@@ -288,6 +288,26 @@ def _resolve_sex(si: SampleInput, chroms: npt.NDArray[np.str_], lrr: F64, geno:
     return "U", "unresolved", ["sex_unresolved"]
 
 
+def _baf_noise(
+    baf: npt.NDArray[np.float64], geno: npt.NDArray[np.uint8], auto: npt.NDArray[np.bool_],
+    ref_het_sd: npt.NDArray[np.float32] | None, baf_sd: float,
+) -> npt.NDArray[np.float64]:  # fmt: skip
+    """CYTO-041: expected heterozygous BAF noise per probe, the reference spread scaled to this sample
+    (median |BAF − 0.5| / spread over autosomal heterozygous calls = 0.6745 × scale); ``baf_sd`` without a reference
+    spread or where a probe has none."""
+    out = np.full(len(baf), baf_sd, dtype=np.float64)
+    if ref_het_sd is None:
+        return out
+    h = np.asarray(ref_het_sd, dtype=np.float64)
+    hc = auto & (geno == 2) & np.isfinite(baf) & np.isfinite(h) & (h > 0)
+    if hc.sum() < 100:
+        return out
+    scale = float(np.median(np.abs(baf[hc] - 0.5) / np.maximum(h[hc], 1e-3))) / 0.6745
+    ok = np.isfinite(h) & (h > 0)
+    out[ok] = np.maximum(h[ok], 1e-3) * scale
+    return out
+
+
 def call_sample(si: SampleInput, ctx: dict[str, Any], opts: CnCallOptions) -> dict[str, Any]:
     """One sample: preprocessing, per-region HMM, events, VCF + segments. Returns the summary row + events."""
     from simurg.cn import hmm
@@ -346,6 +366,7 @@ def call_sample(si: SampleInput, ctx: dict[str, Any], opts: CnCallOptions) -> di
     seg_sd = _robust_sd(lrr[auto], 0.05)
     b_het = baf[auto & (baf > 0.3) & (baf < 0.7)]
     baf_sd = _robust_sd(b_het, 0.02)
+    sigma = _baf_noise(baf, geno, auto, model.ref_het_sd, baf_sd)  # CYTO-041
     dip, hap = hmm.diploid_states(opts.mosaic_prior), hmm.haploid_states(opts.mosaic_prior)
     counts: dict[str, int] = {}
     events: list[Any] = []
@@ -392,7 +413,7 @@ def call_sample(si: SampleInput, ctx: dict[str, Any], opts: CnCallOptions) -> di
             post = hmm.forward_backward(e, sw, prior)
             evs = extract(chrom, path, states, p)
             for ev in evs:
-                annotate_stats(ev, p, x_raw, b, gt, post, states, m_sp, ctx["gamma"], haploid, baf_sd, pfb[idx])
+                annotate_stats(ev, p, x_raw, b, gt, post, states, m_sp, ctx["gamma"], haploid, sigma[idx], pfb[idx])
             evs = post_filter(evs, [x_raw[ev.i0 : ev.i1] for ev in evs], th, counts)
             for s_, e_ in _runs(path):
                 st = states[int(path[s_])]

tests/unit/test_cn.py

@@ -12,7 +12,7 @@ from typer.testing import CliRunner
 
 from simurg.cli.main import app
 from simurg.cn import hmm
-from simurg.cn.events import Event, Thresholds, _mf_from_baf, apply_thresholds, effective_size, extract
+from simurg.cn.events import Event, Thresholds, _mf_from_baf, apply_thresholds, complete_loss, effective_size, extract
 from simurg.cn.gcwave import gc_correct
 from simurg.cn.model import gc_fractions, population_bfreq
 from simurg.formats.gtc import Gtc, write_gtc
@@ -237,12 +237,15 @@ def test_mf_refinements() -> None:  # CYTO-018 / CYTO-041
     b = 0.5 + rng.normal(0, sd, 400)
     mf = _mf_from_baf("MDEL", b, sd)
     assert mf is not None and mf < 0.08
-    assert (_mf_from_baf("MDEL", b, 0.0) or 0) > 0.05  # uncorrected, the old estimate was biased upwards
+    assert (_mf_from_baf("MDEL", b) or 0) > 0.05  # uncorrected, the old estimate was biased upwards
     # a true 40 % mosaic loss is still recovered
     d = 0.4 / (2 * (2 - 0.4))
     b = 0.5 + rng.choice([-d, d], 400) + rng.normal(0, sd, 400)
     assert abs((_mf_from_baf("MDEL", b, sd) or 0) - 0.4) < 0.08
-    # no heterozygotes where ~50 are expected: a complete (constitutional) loss
+    # no heterozygous calls where ~12 are expected: a complete (constitutional) loss
-    b = rng.choice([0.0, 1.0], 100) + rng.normal(0, 0.01, 100)
+    hom = np.full(30, 1, dtype=np.uint8)
-    assert _mf_from_baf("MDEL", np.clip(b, 0, 1), sd, np.full(100, 0.3)) == 1.0
+    assert complete_loss(hom, np.full(30, 0.3))
-    assert _mf_from_baf("MDEL", np.clip(b, 0, 1), sd, np.full(100, 0.01)) is None  # too few expected
+    assert not complete_loss(hom, np.full(30, 0.02))  # too few expected
+    mixed = hom.copy()
+    mixed[:8] = 2
+    assert not complete_loss(mixed, np.full(30, 0.3))