P7: differential-testing harness (simurg compare) and nightly run on the Pi
- compare/core: EXACT/NEAR/DISCREPANCY/UNKNOWN with the plan §16.3 tolerances (abs, float32 ULP), NaN = NaN, locus classes (assay, variant type, chromosome group, intensity-only); fields absent on both sides are not applicable - compare/records: GTC vs GTC (per-locus arrays, transforms per bin, controls, sample statistics; metadata listed, not counted) and VCF vs VCF (records keyed incl. ID; SNV/MNV/indel, merged, multi-allelic classes) - simurg compare A B (files or folders paired by name): summary.json, discrepancies.tsv.gz (never truncated), self-contained dashboard.html; --tolerance, --fail-on-discrepancy (exit 3) - scripts/nightly.py + install_nightly.sh (crontab 02:30): unit tests, GSA vs gtc2vcf (GTC + VCF), VCF regression vs golden, Omni2.5 subset vs vendor GTCs; first run PASS on every step - tests/unit/test_compare.py; docs/validation/P7.md, README
This commit is contained in:
kemal committed
1 parent
51747674ba
commit
3c33ae5cb1
10 files changed
+1268
-1
No files matched your search
Whitespace-only changes.
@@ -0,0 +1,223 @@
|
||||
"""Field-level comparison with tolerance categories (spec CLI-038; plan §16.3).
|
||||
|
||||
Every compared value falls in exactly one category:
|
||||
|
||||
* **EXACT** — bitwise equal (NaN equals NaN whatever its sign bit, K-014);
|
||||
* **NEAR** — within the field's tolerance;
|
||||
* **DISCREPANCY** — outside it (or NaN on one side only);
|
||||
* **UNKNOWN** — one side lacks the field or the record.
|
||||
|
||||
Counts are kept per locus class (assay, variant type, chromosome group, intensity-only), and every discrepancy is
|
||||
kept with its identifiers — nothing is hidden (plan §16.3).
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from collections.abc import Mapping, Sequence
|
||||
from dataclasses import dataclass
|
||||
from dataclasses import field as dc_field
|
||||
from typing import Any
|
||||
|
||||
import numpy as np
|
||||
import numpy.typing as npt
|
||||
|
||||
EXACT, NEAR, DISCREPANCY, UNKNOWN = 0, 1, 2, 3
|
||||
CATEGORY_NAMES = ("EXACT", "NEAR", "DISCREPANCY", "UNKNOWN")
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Tolerance:
|
||||
abs: float | None = None # absolute difference allowed
|
||||
ulp: int | None = None # float32 units in the last place allowed
|
||||
|
||||
@property
|
||||
def exact(self) -> bool:
|
||||
return self.abs is None and self.ulp is None
|
||||
|
||||
|
||||
EXACT_TOL = Tolerance()
|
||||
# plan §16.3 (bit-exact mode targets)
|
||||
GTC_TOLERANCES: dict[str, Tolerance] = {
|
||||
"raw_x": EXACT_TOL,
|
||||
"raw_y": EXACT_TOL,
|
||||
"controls_x": EXACT_TOL,
|
||||
"controls_y": EXACT_TOL,
|
||||
"genotypes": EXACT_TOL,
|
||||
"base_calls": EXACT_TOL,
|
||||
"normalization_transforms": Tolerance(ulp=2),
|
||||
"genotype_scores": Tolerance(abs=1e-5),
|
||||
"b_allele_freqs": Tolerance(abs=1e-5),
|
||||
"logr_ratios": Tolerance(abs=1e-5),
|
||||
"call_rate": Tolerance(abs=1e-6),
|
||||
"logr_dev": Tolerance(abs=1e-6),
|
||||
"gc10": Tolerance(abs=1e-6),
|
||||
"gc50": Tolerance(abs=1e-6),
|
||||
"gender": EXACT_TOL,
|
||||
"num_calls": EXACT_TOL,
|
||||
"num_no_calls": EXACT_TOL,
|
||||
"num_intensity_only": EXACT_TOL,
|
||||
"percentiles_x": EXACT_TOL,
|
||||
"percentiles_y": EXACT_TOL,
|
||||
}
|
||||
VCF_FLOAT_TOL = Tolerance(abs=1e-5) # GS, BAF, LRR, NORMX, NORMY, THETA, R
|
||||
VCF_EXACT_FIELDS = {"GT", "X", "Y"}
|
||||
|
||||
|
||||
def _float_ulps(a: npt.NDArray[np.float32], b: npt.NDArray[np.float32]) -> npt.NDArray[np.int64]:
|
||||
"""Distance in float32 ULPs (monotone integer mapping of the bit patterns)."""
|
||||
|
||||
def key(x: npt.NDArray[np.float32]) -> npt.NDArray[np.int64]:
|
||||
i = x.astype(np.float32).view(np.int32).astype(np.int64)
|
||||
return np.where(i < 0, -(i & 0x7FFFFFFF), i)
|
||||
|
||||
return np.abs(key(a) - key(b))
|
||||
|
||||
|
||||
def categorize(a: npt.ArrayLike, b: npt.ArrayLike, tol: Tolerance) -> npt.NDArray[np.int8]:
|
||||
"""Element-wise category of two equally shaped arrays (numbers or strings)."""
|
||||
a, b = np.asarray(a), np.asarray(b)
|
||||
if a.shape != b.shape:
|
||||
raise ValueError(f"shape {a.shape} vs {b.shape}")
|
||||
out = np.full(a.shape, DISCREPANCY, np.int8)
|
||||
if a.dtype.kind in "fc" or b.dtype.kind in "fc":
|
||||
fa, fb = a.astype(np.float64), b.astype(np.float64)
|
||||
na, nb = np.isnan(fa), np.isnan(fb)
|
||||
same = (fa == fb) | (na & nb)
|
||||
if a.dtype == np.float32 and b.dtype == np.float32:
|
||||
same &= (a.view(np.uint32) == b.view(np.uint32)) | (na & nb) | ((fa == 0) & (fb == 0))
|
||||
out[same] = EXACT
|
||||
near = np.zeros(a.shape, np.bool_)
|
||||
valid = ~(na | nb) & ~same
|
||||
if tol.abs is not None:
|
||||
near |= valid & (np.abs(fa - fb) <= tol.abs)
|
||||
if tol.ulp is not None:
|
||||
near |= valid & (_float_ulps(a.astype(np.float32), b.astype(np.float32)) <= tol.ulp)
|
||||
out[near] = NEAR
|
||||
return out
|
||||
out[a == b] = EXACT
|
||||
return out
|
||||
|
||||
|
||||
@dataclass
|
||||
class FieldResult:
|
||||
"""Category counts of one field: overall and per locus class, plus the discrepancies themselves."""
|
||||
|
||||
field: str
|
||||
tolerance: str
|
||||
counts: dict[str, list[int]] = dc_field(default_factory=dict) # class -> [EXACT, NEAR, DISCREPANCY, UNKNOWN]
|
||||
discrepancies: list[dict[str, Any]] = dc_field(default_factory=list)
|
||||
|
||||
def add(self, cls: str, cats: npt.NDArray[np.int8]) -> None:
|
||||
c = self.counts.setdefault(cls, [0, 0, 0, 0])
|
||||
binc = np.bincount(cats.ravel().astype(np.int64), minlength=4)
|
||||
for k in range(4):
|
||||
c[k] += int(binc[k])
|
||||
|
||||
@property
|
||||
def total(self) -> list[int]:
|
||||
return self.counts.get("all", [0, 0, 0, 0])
|
||||
|
||||
|
||||
def tol_text(t: Tolerance) -> str:
|
||||
if t.exact:
|
||||
return "exact"
|
||||
parts = []
|
||||
if t.abs is not None:
|
||||
parts.append(f"abs<={t.abs:g}")
|
||||
if t.ulp is not None:
|
||||
parts.append(f"ulp<={t.ulp}")
|
||||
return ",".join(parts)
|
||||
|
||||
|
||||
def locus_classes(
|
||||
chroms: Sequence[str], snps: Sequence[str], assay_types: npt.ArrayLike, intensity_only: npt.ArrayLike
|
||||
) -> dict[str, npt.NDArray[np.bool_]]:
|
||||
"""Locus class masks (plan §16.3): assay, variant type, chromosome group, intensity-only."""
|
||||
at = np.asarray(assay_types)
|
||||
io = np.asarray(intensity_only, dtype=np.bool_)
|
||||
|
||||
def group(c: str) -> str:
|
||||
c = c[3:] if c.lower().startswith("chr") else c
|
||||
if c in ("0", ""):
|
||||
return "unmapped"
|
||||
if c == "XY":
|
||||
return "PAR"
|
||||
if c in ("X", "Y"):
|
||||
return c
|
||||
if c in ("MT", "M"):
|
||||
return "MT"
|
||||
return "autosome" if c.isdigit() else "other"
|
||||
|
||||
groups = np.array([group(c) for c in chroms])
|
||||
indel = np.array([("D" in s or "I" in s) for s in snps])
|
||||
out = {"all": np.ones(len(groups), np.bool_), "assay=InfII": at == 0, "assay=InfI": at != 0,
|
||||
"type=SNV": ~indel, "type=indel": indel, "intensity_only": io} # fmt: skip
|
||||
for gname in ("autosome", "X", "Y", "PAR", "MT", "unmapped", "other"):
|
||||
m = groups == gname
|
||||
if m.any():
|
||||
out[f"chrom={gname}"] = m
|
||||
return out
|
||||
|
||||
|
||||
def compare_locus_field(
|
||||
name: str,
|
||||
a: npt.ArrayLike | None,
|
||||
b: npt.ArrayLike | None,
|
||||
tol: Tolerance,
|
||||
classes: Mapping[str, npt.NDArray[np.bool_]],
|
||||
ids: Sequence[str],
|
||||
context: Mapping[str, str],
|
||||
) -> FieldResult:
|
||||
"""Per-locus arrays (first axis = locus). Missing or empty on one side → UNKNOWN for every locus."""
|
||||
res = FieldResult(name, tol_text(tol))
|
||||
n = len(classes["all"])
|
||||
a_ok = a is not None and len(np.asarray(a)) == n
|
||||
b_ok = b is not None and len(np.asarray(b)) == n
|
||||
if not (a_ok and b_ok):
|
||||
cats = np.full(n, UNKNOWN, np.int8)
|
||||
else:
|
||||
el = categorize(np.asarray(a), np.asarray(b), tol)
|
||||
cats = el.max(axis=tuple(range(1, el.ndim))) if el.ndim > 1 else el # worst element of the locus
|
||||
for i in np.flatnonzero(cats == DISCREPANCY).tolist():
|
||||
res.discrepancies.append(
|
||||
{**context, "field": name, "locus": ids[i], "a": _fmt(np.asarray(a)[i]), "b": _fmt(np.asarray(b)[i])}
|
||||
)
|
||||
for cls, mask in classes.items():
|
||||
res.add(cls, cats[mask])
|
||||
return res
|
||||
|
||||
|
||||
def compare_scalar(name: str, a: Any, b: Any, tol: Tolerance, context: Mapping[str, str]) -> FieldResult:
|
||||
res = FieldResult(name, tol_text(tol))
|
||||
if a is None or b is None:
|
||||
cat = np.array([UNKNOWN], np.int8)
|
||||
else:
|
||||
aa, bb = np.asarray([a]), np.asarray([b])
|
||||
if aa.dtype.kind == "f" or bb.dtype.kind == "f":
|
||||
aa, bb = aa.astype(np.float64), bb.astype(np.float64)
|
||||
cat = categorize(aa, bb, tol)
|
||||
res.add("sample", cat)
|
||||
if cat[0] == DISCREPANCY:
|
||||
res.discrepancies.append({**context, "field": name, "locus": "", "a": _fmt(a), "b": _fmt(b)})
|
||||
return res
|
||||
|
||||
|
||||
def _fmt(v: Any) -> str:
|
||||
arr = np.asarray(v)
|
||||
if arr.ndim == 0:
|
||||
x = arr.item()
|
||||
return repr(x) if isinstance(x, float) else str(x)
|
||||
return "[" + ",".join(_fmt(x) for x in arr.ravel().tolist()) + "]"
|
||||
|
||||
|
||||
def merge_results(results: Sequence[FieldResult]) -> dict[str, FieldResult]:
|
||||
"""Sum results of the same field (e.g. over samples)."""
|
||||
out: dict[str, FieldResult] = {}
|
||||
for r in results:
|
||||
m = out.setdefault(r.field, FieldResult(r.field, r.tolerance))
|
||||
for cls, c in r.counts.items():
|
||||
cur = m.counts.setdefault(cls, [0, 0, 0, 0])
|
||||
for k in range(4):
|
||||
cur[k] += c[k]
|
||||
m.discrepancies.extend(r.discrepancies)
|
||||
return out
|
||||
@@ -0,0 +1,223 @@
|
||||
"""GTC-vs-GTC and VCF-vs-VCF comparison on parsed data (spec CLI-038; plan §16.3)."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
from collections.abc import Mapping, Sequence
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Any
|
||||
|
||||
import numpy as np
|
||||
import numpy.typing as npt
|
||||
from numpy.lib import recfunctions as rfn
|
||||
|
||||
from simurg.compare.core import (
|
||||
DISCREPANCY,
|
||||
EXACT,
|
||||
GTC_TOLERANCES,
|
||||
UNKNOWN,
|
||||
VCF_EXACT_FIELDS,
|
||||
VCF_FLOAT_TOL,
|
||||
FieldResult,
|
||||
Tolerance,
|
||||
categorize,
|
||||
compare_locus_field,
|
||||
compare_scalar,
|
||||
tol_text,
|
||||
)
|
||||
|
||||
GTC_LOCUS_FIELDS = ("raw_x", "raw_y", "genotypes", "base_calls", "genotype_scores", "b_allele_freqs", "logr_ratios")
|
||||
GTC_SCALARS = ("call_rate", "logr_dev", "gc10", "gc50", "gender", "num_calls", "num_no_calls", "num_intensity_only")
|
||||
GTC_METADATA = ("sample_name", "sample_plate", "sample_well", "cluster_file", "snp_manifest", "imaging_date",
|
||||
"autocall_date", "autocall_version", "slide_identifier") # fmt: skip
|
||||
|
||||
|
||||
@dataclass
|
||||
class Comparison:
|
||||
"""Results of one pair (one sample)."""
|
||||
|
||||
label: str
|
||||
kind: str # "gtc" | "vcf"
|
||||
fields: list[FieldResult] = field(default_factory=list)
|
||||
metadata: dict[str, list[str]] = field(default_factory=dict) # field -> [a, b] when they differ
|
||||
notes: list[str] = field(default_factory=list)
|
||||
|
||||
|
||||
def compare_gtc(
|
||||
a: Any,
|
||||
b: Any,
|
||||
label: str,
|
||||
classes: Mapping[str, npt.NDArray[np.bool_]],
|
||||
names: Sequence[str],
|
||||
tolerances: Mapping[str, Tolerance] = GTC_TOLERANCES,
|
||||
) -> Comparison:
|
||||
"""Two parsed GTCs of the same manifest."""
|
||||
res = Comparison(label, "gtc")
|
||||
ctx = {"pair": label}
|
||||
if a.num_snps != b.num_snps:
|
||||
res.notes.append(f"locus count differs: {a.num_snps} vs {b.num_snps}")
|
||||
return res
|
||||
for f in GTC_LOCUS_FIELDS:
|
||||
va, vb = getattr(a, f, None), getattr(b, f, None)
|
||||
if (va is None or len(va) == 0) and (vb is None or len(vb) == 0):
|
||||
continue # absent on both sides (e.g. BAF/LRR in GTC v3): not applicable
|
||||
res.fields.append(compare_locus_field(f, va, vb, tolerances[f], classes, names, ctx))
|
||||
# normalization transforms: one row per bin
|
||||
ta, tb = a.normalization_transforms, b.normalization_transforms
|
||||
fr = FieldResult("normalization_transforms", tol_text(tolerances["normalization_transforms"]))
|
||||
if len(ta) != len(tb) or not len(ta):
|
||||
fr.add("bins", np.full(max(len(ta), len(tb), 1), UNKNOWN, np.int8))
|
||||
else:
|
||||
ua = rfn.structured_to_unstructured(ta).astype(np.float32)
|
||||
ub = rfn.structured_to_unstructured(tb).astype(np.float32)
|
||||
cats = categorize(ua, ub, tolerances["normalization_transforms"]).max(axis=1)
|
||||
fr.add("bins", cats)
|
||||
for i in np.flatnonzero(cats == DISCREPANCY).tolist():
|
||||
fr.discrepancies.append({**ctx, "field": "normalization_transforms", "locus": f"bin{i}",
|
||||
"a": str(ua[i].tolist()), "b": str(ub[i].tolist())}) # fmt: skip
|
||||
res.fields.append(fr)
|
||||
# controls: one row per control bead
|
||||
for f in ("controls_x", "controls_y"):
|
||||
ca, cb = getattr(a, f), getattr(b, f)
|
||||
fr = FieldResult(f, tol_text(tolerances[f]))
|
||||
fr.add("controls", categorize(ca, cb, tolerances[f]) if len(ca) == len(cb) else np.array([UNKNOWN], np.int8))
|
||||
res.fields.append(fr)
|
||||
for f in GTC_SCALARS:
|
||||
res.fields.append(compare_scalar(f, _scalar(a, f), _scalar(b, f), tolerances[f], ctx))
|
||||
for f in ("percentiles_x", "percentiles_y"):
|
||||
pa, pb = getattr(a, f), getattr(b, f)
|
||||
fr = FieldResult(f, tol_text(tolerances[f]))
|
||||
fr.add("sample", categorize(pa, pb, tolerances[f]) if len(pa) == len(pb) else np.array([UNKNOWN], np.int8))
|
||||
res.fields.append(fr)
|
||||
for f in GTC_METADATA:
|
||||
va, vb = str(getattr(a, f, "")), str(getattr(b, f, ""))
|
||||
if va != vb:
|
||||
res.metadata[f] = [va, vb]
|
||||
if a.version != b.version:
|
||||
res.notes.append(f"GTC versions differ: {a.version} vs {b.version} (fields absent in one are UNKNOWN)")
|
||||
return res
|
||||
|
||||
|
||||
def _scalar(g: Any, f: str) -> Any:
|
||||
if f == "gender":
|
||||
return g.gender_char
|
||||
v = getattr(g, f, None)
|
||||
if isinstance(v, float) and math.isnan(v):
|
||||
return v
|
||||
return v
|
||||
|
||||
|
||||
# ------------------------------------------------------------------------------------------------ VCF
|
||||
@dataclass
|
||||
class VcfRecord:
|
||||
chrom: str
|
||||
pos: int
|
||||
vid: str
|
||||
ref: str
|
||||
alt: str
|
||||
fields: dict[str, str] # FORMAT key -> value of the compared sample
|
||||
|
||||
|
||||
def vcf_classes(r: VcfRecord) -> list[str]:
|
||||
c = r.chrom[3:] if r.chrom.lower().startswith("chr") else r.chrom
|
||||
grp = "autosome" if c.isdigit() else {"X": "X", "Y": "Y", "MT": "MT", "M": "MT"}.get(c, "other")
|
||||
alts = r.alt.split(",")
|
||||
if all(len(x) == len(r.ref) == 1 for x in alts):
|
||||
vtype = "SNV"
|
||||
elif all(len(x) == len(r.ref) for x in alts):
|
||||
vtype = "MNV"
|
||||
else:
|
||||
vtype = "indel"
|
||||
out = ["all", f"chrom={grp}", f"type={vtype}"]
|
||||
if "," in r.vid or ";" in r.vid:
|
||||
out.append("merged_site")
|
||||
if len(alts) > 1:
|
||||
out.append("multiallelic")
|
||||
return out
|
||||
|
||||
|
||||
def _gt_norm(gt: str) -> str:
|
||||
parts = gt.replace("|", "/").split("/")
|
||||
if any(p == "." for p in parts):
|
||||
return "." if len(parts) == 1 else "./."
|
||||
return "/".join(sorted(parts, key=int))
|
||||
|
||||
|
||||
def _cat_scalar(va: str, vb: str, tol: Tolerance) -> int:
|
||||
if va == vb:
|
||||
return EXACT
|
||||
x, y = _num(va), _num(vb)
|
||||
if math.isnan(x) and math.isnan(y):
|
||||
return EXACT
|
||||
if math.isnan(x) or math.isnan(y):
|
||||
return DISCREPANCY
|
||||
if x == y:
|
||||
return EXACT
|
||||
if tol.abs is not None and abs(x - y) <= tol.abs:
|
||||
return 1 # NEAR
|
||||
return DISCREPANCY
|
||||
|
||||
|
||||
def compare_vcf_records(
|
||||
a: Mapping[tuple[str, int, str, str, str], VcfRecord],
|
||||
b: Mapping[tuple[str, int, str, str, str], VcfRecord],
|
||||
label: str,
|
||||
float_tol: Tolerance = VCF_FLOAT_TOL,
|
||||
) -> Comparison:
|
||||
"""Records keyed by (CHROM, POS, REF, ALT, ID); FORMAT fields of one sample on each side."""
|
||||
res = Comparison(label, "vcf")
|
||||
ctx = {"pair": label}
|
||||
keys = list(a) + [k for k in b if k not in a]
|
||||
fields = sorted({f for r in a.values() for f in r.fields} | {f for r in b.values() for f in r.fields})
|
||||
counts: dict[str, dict[str, list[int]]] = {f: {} for f in ["record", *fields]}
|
||||
disc: dict[str, list[dict[str, Any]]] = {f: [] for f in ["record", *fields]}
|
||||
tols = {f: (Tolerance() if f in VCF_EXACT_FIELDS else float_tol) for f in fields}
|
||||
|
||||
def bump(f: str, classes: list[str], cat: int) -> None:
|
||||
per = counts[f]
|
||||
for cls in classes:
|
||||
c = per.get(cls)
|
||||
if c is None:
|
||||
c = per[cls] = [0, 0, 0, 0]
|
||||
c[cat] += 1
|
||||
|
||||
for k in keys:
|
||||
ra, rb = a.get(k), b.get(k)
|
||||
r = ra or rb
|
||||
assert r is not None
|
||||
classes = vcf_classes(r)
|
||||
where = f"{r.chrom}:{r.pos}:{r.ref}>{r.alt} {r.vid}"
|
||||
if ra is None or rb is None:
|
||||
bump("record", classes, UNKNOWN)
|
||||
disc["record"].append({**ctx, "field": "record", "locus": where,
|
||||
"a": "present" if ra else "absent", "b": "present" if rb else "absent"}) # fmt: skip
|
||||
continue
|
||||
bump("record", classes, EXACT)
|
||||
for f in fields:
|
||||
va, vb = ra.fields.get(f), rb.fields.get(f)
|
||||
if va is None and vb is None:
|
||||
continue # the field does not apply to this record on either side (e.g. BAF of merged sites)
|
||||
if va is None or vb is None:
|
||||
cat = UNKNOWN
|
||||
elif f == "GT":
|
||||
cat = EXACT if _gt_norm(va) == _gt_norm(vb) else DISCREPANCY
|
||||
else:
|
||||
cat = _cat_scalar(va, vb, tols[f])
|
||||
bump(f, classes, cat)
|
||||
if cat == DISCREPANCY:
|
||||
disc[f].append({**ctx, "field": f, "locus": where, "a": va, "b": vb})
|
||||
for f in ["record", *fields]:
|
||||
fr = FieldResult(
|
||||
f, "present in both" if f == "record" else ("exact" if f in VCF_EXACT_FIELDS else tol_text(float_tol))
|
||||
)
|
||||
fr.counts = counts[f]
|
||||
fr.discrepancies = disc[f]
|
||||
res.fields.append(fr)
|
||||
return res
|
||||
|
||||
|
||||
def _num(v: str) -> float:
|
||||
try:
|
||||
return float(v)
|
||||
except ValueError:
|
||||
return math.nan
|
||||
Reference in new issue
Block a user