A review of the ranking's arithmetic found four things wrong, all of which made the score look better informed than it was. Measurements below are from this repo, not estimates. **Components now abstain instead of inventing a number.** A run without a VEP cache returns no allele frequencies, and rarity_score(None) read that as "absent from gnomAD, therefore maximally rare" and awarded every variant a free 0.25. jobs.has_frequencies / has_effect_scores record what the run actually produced, absent components are dropped from the weighted mean, and the remaining weights are renormalised so the score keeps its meaning. The UI shows "not looked up" rather than a bar, and the funnel stops calling a step "rare" when nothing was filtered. **Allele frequency is no longer a model feature.** It dominated: the same missense variant scored 0.887 at AF 0 and 0.0003 at AF 0.01. That double- counted, because the ranking already scores frequency explicitly, putting ~45% of every rank on one measurement; and it was circular, because ACMG assigns ClinVar's benign labels using frequency (BA1/BS1). Retraining without it moves missense AUROC from 0.872 to 0.500 — exactly random. The old figure was allele frequency, not variant-effect knowledge. The model therefore abstains unless CADD or AlphaMissense is present, since otherwise it only restates the consequence class. **Phenotype matching is weighted by information content** and HPO annotations are propagated up the ontology. Counting terms alike let "global developmental delay" (IC 0.93) count as much as "dilated left subclavian artery" (IC 7.88). **A real bug in the propagation, found by checking it.** The ancestor walk read a pre-order DFS backwards, which on a DAG lets a term resolve before one of its parents and inherit that parent alone instead of its lineage. It dropped 399 terms out of the phenotype branch, Camptodactyly and Chiari malformation among them. Now a true post-order, tested against a reference transitive closure. The ontology arithmetic moved to rarelens_ml.hpo so it is covered by tests, and rarelens_ml.benchmark measures the whole thing: across 10,178 published cases the causal gene ranks first 45.9-81.0% of the time against 5,269 genes, versus 0.02% for chance. docs/data.md reports that with its contamination (HPO's annotations come from these same case reports), and includes the measurement showing information-content weighting earns its place while propagation does not - kept anyway, for a reason the docs argue rather than assume.
242 lines
8.8 KiB
Python
242 lines
8.8 KiB
Python
import logging
|
|
import uuid
|
|
from dataclasses import asdict
|
|
from datetime import UTC, datetime
|
|
|
|
from fastapi import APIRouter, HTTPException, Query, status
|
|
from sqlalchemy import func, select
|
|
from sqlalchemy.exc import IntegrityError
|
|
from sqlalchemy.orm import selectinload
|
|
|
|
from app.config import settings
|
|
from app.db import SessionDep
|
|
from app.models import (
|
|
Case,
|
|
CasePhenotype,
|
|
DecisionState,
|
|
Job,
|
|
JobStatus,
|
|
Variant,
|
|
VariantDecision,
|
|
)
|
|
from app.schemas import (
|
|
CandidateOut,
|
|
CandidatePage,
|
|
CaseCreate,
|
|
CaseOut,
|
|
EvidenceOut,
|
|
FunnelOut,
|
|
JobOut,
|
|
ProvenanceOut,
|
|
ReportOut,
|
|
ScoreOut,
|
|
)
|
|
from app.services import candidates as case_view
|
|
from app.services import events, triage
|
|
from app.services.scoring import score_job
|
|
|
|
logger = logging.getLogger(__name__)
|
|
router = APIRouter()
|
|
|
|
REPORT_TOP = 5
|
|
|
|
|
|
async def _case_or_404(session: SessionDep, case_id: uuid.UUID) -> Case:
|
|
case = await case_view.get_case(session, case_id)
|
|
if case is None:
|
|
raise HTTPException(404, "case not found")
|
|
return case
|
|
|
|
|
|
async def _shortlisted_counts(session: SessionDep, case_ids: list[uuid.UUID]) -> dict[uuid.UUID, int]:
|
|
if not case_ids:
|
|
return {}
|
|
rows = await session.execute(
|
|
select(Job.case_id, func.count(VariantDecision.id))
|
|
.join(Variant, Variant.job_id == Job.id)
|
|
.join(VariantDecision, VariantDecision.variant_id == Variant.id)
|
|
.where(Job.case_id.in_(case_ids), VariantDecision.state == DecisionState.shortlisted)
|
|
.group_by(Job.case_id)
|
|
)
|
|
return dict(rows.all()) # type: ignore[arg-type]
|
|
|
|
|
|
async def _latest_jobs(session: SessionDep, case_ids: list[uuid.UUID]) -> dict[uuid.UUID, Job]:
|
|
if not case_ids:
|
|
return {}
|
|
jobs = await session.scalars(
|
|
select(Job).where(Job.case_id.in_(case_ids)).order_by(Job.created_at.desc())
|
|
)
|
|
latest: dict[uuid.UUID, Job] = {}
|
|
for job in jobs: # ordered newest first, so the first one wins
|
|
latest.setdefault(job.case_id, job)
|
|
return latest
|
|
|
|
|
|
def _as_case_out(case: Case, job: Job | None, shortlisted: int) -> CaseOut:
|
|
out = CaseOut.model_validate(case)
|
|
out.latest_job = JobOut.model_validate(job) if job is not None else None
|
|
out.shortlisted = shortlisted
|
|
return out
|
|
|
|
|
|
@router.get("", response_model=list[CaseOut])
|
|
async def list_cases(session: SessionDep):
|
|
cases = (
|
|
await session.scalars(
|
|
select(Case).order_by(Case.created_at.desc()).options(selectinload(Case.phenotypes))
|
|
)
|
|
).all()
|
|
ids = [c.id for c in cases]
|
|
jobs, counts = await _latest_jobs(session, ids), await _shortlisted_counts(session, ids)
|
|
return [_as_case_out(c, jobs.get(c.id), counts.get(c.id, 0)) for c in cases]
|
|
|
|
|
|
@router.post("", response_model=CaseOut, status_code=status.HTTP_201_CREATED)
|
|
async def create_case(payload: CaseCreate, session: SessionDep):
|
|
case = Case(
|
|
name=payload.name,
|
|
vcf_uri=payload.vcf_uri,
|
|
assembly=payload.assembly,
|
|
phenotypes=[CasePhenotype(hpo_id=p.hpo_id, label=p.label) for p in payload.phenotypes],
|
|
)
|
|
session.add(case)
|
|
try:
|
|
await session.commit()
|
|
except IntegrityError: # cases.name is unique
|
|
await session.rollback()
|
|
raise HTTPException(409, f"a case named {payload.name!r} already exists") from None
|
|
await session.refresh(case, attribute_names=["phenotypes"])
|
|
return _as_case_out(case, None, 0)
|
|
|
|
|
|
@router.get("/{case_id}", response_model=CaseOut)
|
|
async def get_case(case_id: uuid.UUID, session: SessionDep):
|
|
case = await _case_or_404(session, case_id)
|
|
counts = await _shortlisted_counts(session, [case_id])
|
|
return _as_case_out(case, await case_view.latest_job(session, case_id), counts.get(case_id, 0))
|
|
|
|
|
|
@router.get("/{case_id}/jobs", response_model=list[JobOut])
|
|
async def list_jobs(case_id: uuid.UUID, session: SessionDep):
|
|
result = await session.scalars(
|
|
select(Job).where(Job.case_id == case_id).order_by(Job.created_at.desc())
|
|
)
|
|
return result.all()
|
|
|
|
|
|
@router.post("/{case_id}/annotate", response_model=JobOut, status_code=status.HTTP_202_ACCEPTED)
|
|
async def annotate(case_id: uuid.UUID, session: SessionDep):
|
|
case = await _case_or_404(session, case_id)
|
|
# Commit `running` before launching: a local run that dies instantly is marked failed by its
|
|
# watcher, and a later status write here would overwrite that.
|
|
job = Job(case_id=case.id, status=JobStatus.running)
|
|
session.add(job)
|
|
await session.commit()
|
|
try:
|
|
job.workflow_ref = await events.launch(job.id, case.vcf_uri, case.assembly)
|
|
except events.LaunchError as e:
|
|
job.status = JobStatus.failed
|
|
job.log = str(e)
|
|
job.finished_at = datetime.now(UTC)
|
|
await session.commit()
|
|
await session.refresh(job)
|
|
return job
|
|
|
|
|
|
@router.post("/{case_id}/score", response_model=ScoreOut)
|
|
async def score(case_id: uuid.UUID, session: SessionDep):
|
|
await _case_or_404(session, case_id)
|
|
job = await case_view.latest_job(session, case_id, status=JobStatus.succeeded)
|
|
if job is None:
|
|
raise HTTPException(409, "no finished annotation to score")
|
|
try:
|
|
scored, version = await score_job(job.id, session)
|
|
except Exception as e:
|
|
# No registry, no model behind the alias, a model that will not load: all of these are
|
|
# the environment being unready, not a bug in the request. Name what is missing and keep
|
|
# the exception in the log, where it is useful, rather than in the UI, where it is noise.
|
|
logger.exception("scoring case %s failed", case_id)
|
|
raise HTTPException(
|
|
503,
|
|
f"no model available: {settings.model_name}@{settings.model_alias} "
|
|
f"via {settings.mlflow_tracking_uri}",
|
|
) from e
|
|
return ScoreOut(case_id=case_id, scored=scored, model_version=version)
|
|
|
|
|
|
@router.get("/{case_id}/candidates", response_model=CandidatePage)
|
|
async def list_candidates(
|
|
case_id: uuid.UUID,
|
|
session: SessionDep,
|
|
gene: str | None = None,
|
|
impact: str | None = Query(None, pattern="^(HIGH|MODERATE|LOW|MODIFIER)$"),
|
|
max_af: float | None = Query(None, ge=0, le=1),
|
|
state: str | None = Query(None, pattern="^(shortlisted|dismissed|undecided)$"),
|
|
limit: int = Query(50, ge=1, le=500),
|
|
offset: int = Query(0, ge=0),
|
|
) -> CandidatePage:
|
|
case = await _case_or_404(session, case_id)
|
|
view = await case_view.build(session, case)
|
|
|
|
items = view.candidates
|
|
if gene:
|
|
items = [c for c in items if (c.variant.gene or "").upper() == gene.upper()]
|
|
if impact:
|
|
items = [c for c in items if c.variant.impact == impact]
|
|
if max_af is not None:
|
|
items = [c for c in items if c.variant.gnomad_af is None or c.variant.gnomad_af <= max_af]
|
|
if state == "undecided":
|
|
items = [c for c in items if c.variant.decision is None]
|
|
elif state is not None:
|
|
items = [c for c in items if c.variant.decision and c.variant.decision.state.value == state]
|
|
|
|
labels = view.labels
|
|
return CandidatePage(
|
|
# The funnel describes the whole case, not the filtered view.
|
|
funnel=FunnelOut(**asdict(view.funnel)),
|
|
evidence=EvidenceOut(**asdict(view.evidence), missing=view.evidence.missing),
|
|
weights=triage.weights_in_use(view.evidence),
|
|
items=[CandidateOut.from_candidate(c, labels) for c in items[offset : offset + limit]],
|
|
total=len(items),
|
|
limit=limit,
|
|
offset=offset,
|
|
)
|
|
|
|
|
|
@router.get("/{case_id}/report", response_model=ReportOut)
|
|
async def report(case_id: uuid.UUID, session: SessionDep) -> ReportOut:
|
|
case = await _case_or_404(session, case_id)
|
|
view = await case_view.build(session, case)
|
|
labels = view.labels
|
|
|
|
def decided(state: DecisionState) -> list[CandidateOut]:
|
|
return [
|
|
CandidateOut.from_candidate(c, labels)
|
|
for c in view.candidates
|
|
if c.variant.decision is not None and c.variant.decision.state == state
|
|
]
|
|
|
|
shortlisted, dismissed = decided(DecisionState.shortlisted), decided(DecisionState.dismissed)
|
|
top = [
|
|
CandidateOut.from_candidate(c, labels)
|
|
for c in view.candidates
|
|
if c.variant.decision is None
|
|
][:REPORT_TOP]
|
|
prediction = next((c.variant.prediction for c in view.candidates if c.variant.prediction), None)
|
|
return ReportOut(
|
|
case=_as_case_out(case, view.job, len(shortlisted)),
|
|
funnel=FunnelOut(**asdict(view.funnel)),
|
|
generated_at=datetime.now(UTC),
|
|
provenance=ProvenanceOut(
|
|
job_id=view.job.id if view.job else None,
|
|
vep_version=view.job.vep_version if view.job else None,
|
|
finished_at=view.job.finished_at if view.job else None,
|
|
model_name=prediction.model_name if prediction else None,
|
|
model_version=prediction.model_version if prediction else None,
|
|
),
|
|
shortlisted=shortlisted,
|
|
dismissed=dismissed,
|
|
top=top,
|
|
)
|