Builds a ranked, downloadable menu of candidate flanking SNP markers around a pathogenic variant, for PGT-M linkage and karyomapping. Syngamy then reads SNP arrays after an experiment and reports which parental genome is present, on which chromosomes, and across which regions within them. Progenitor builds a SNP array for a parent nobody genotyped, out of the haploid cells that parent produced. It works out which of the two parents it built from the Y those cells carry, writes the array out as a file usable anywhere an array goes, and calls the parental origin of every input against it.

Developed by & for the Egli Lab at Columbia University Irving Medical Center.

Research use only. Candidate markers require validation and per-family phasing in a qualified genetics laboratory. Not a clinical diagnostic.

Live at originmarker.app, and in parallel at ezrakruger.cc/originmarker.

Genotype files are read in the browser. There is no endpoint that receives one.

Running it

docker compose up -d --build

Then http://localhost:8091. No API keys are required; optional ones are listed in deploy/README-deploy.md and the app degrades gracefully without them.

Without Docker:

python -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/uvicorn app.main:app --reload
cd web && npm install && npm run dev

The Syngamy analysis also runs standalone, with more than the page exposes:

python -m originmarker.origin --father sperm.txt --samples embryo1.txt embryo2.txt

Automation

scripts/checks.sh	every self-check, each from the directory it needs. A root-level loop reports two of them as failing for no reason
scripts/release-check.sh	the version the build reports, the newest changelog entry and the citation table must agree. Add --tag to require the tag too
scripts/deploy.sh	copy, rebuild and verify, with the health check as a gate rather than a printout
cli/examples.check.ts	the bundled examples still produce the rates their own text claims

CI runs the first, second and fourth on every push and pull request, which is what makes Dependabot's grouped minor and patch updates mergeable on a green check.

Tests

export PANELBUILDER_CACHE=tests/fixtures PANELBUILDER_CACHE_TTL=0
.venv/bin/python -m pytest tests/
cd web && npm run build && for f in src/*.check.ts; do node "$f"; done

Offline, against recorded API responses in tests/fixtures/. npm run build is tsc -b && vite build, and the -b is the typecheck: bare tsc has no inputs here and exits 0 regardless.

Layout

originmarker/	the library: panel building, array reading, the HMM, the report writer. Every module carries a self-check, run as python -m originmarker.<module>
app/	FastAPI over that library. app/main.py is the only entry point
web/	the SPA. Syngamy and Progenitor run entirely in the browser; nothing is uploaded
data/	23MB of bundled deCODE recombination maps, read from disk and never re-downloaded
tests/	offline, against recorded API responses in tests/fixtures/
tools/	one-off analysis scripts, kept because their numbers are cited in the docs
audit/	the accuracy audit and its record
docs/	validation write-ups per major version
deploy/	the deployment runbook

Documentation

Method, scope, data sources and limitations: Documentation, Syngamy and Progenitor.

Version history and the bugs each release fixed: CHANGELOG.md.

Accuracy audit, 27 public arrays with bench-established answers: audit/.

Command line

om exposes the same modules the browser runs, with every constant as a flag. The web tool infers them all and offers no knobs on purpose; this does the opposite.

node --experimental-strip-types cli/om.ts constants

om stage <array>	material, dropout, marker floor
om link <parent> <sample>...	child, duplicate, haploid product, or unrelated
om origin <parent> <sample>	which parent's copy is missing, per region
om cohort <dir> --ref <array>	the same for every confirmed child under a directory
om census <dir>	haploid products per donor group
om reconstruct <product>...	a parent's genotypes from that parent's haploid cells
om enrich <regions.tsv>	positional enrichment against a marker-matched null
om constants	every tunable, its value, and why it is that value

Omitting every flag is exactly the configuration the web tool runs and the audits measured. Moving one is printed in the output, since a number produced under a changed constant is not comparable with the validation figures.

License

Apache 2.0.