glio-sign
v0.1.0
Published
GLIO-SIGN — the only npm/JavaScript-native mutational-signature engine: glioma-specialised, governed, multi-modal. Opportunity-corrected SBS96 catalogs, Poisson exposure refit, de-novo extraction, presence testing, CN-48/SV-32/ID-83/RNA-SBS-192 modalities
Maintainers
Readme
glio-sign
The only npm/JavaScript-native mutational-signature engine. Glioma-specialised,
governed, multi-modal. Every other tool in the field (SigProfiler, MuSiCal, sigminer,
deconstructSigs, MutationalPatterns, sigfit, PASA, ...) is Python or R; GLIO-SIGN installs
with npm install and runs in Node, a browser, or a serverless function with zero Python
required for its core, and an optional real Python numeric sidecar for heavy de-novo work.
Research use only (RUO). Not for diagnostic or therapeutic use. GLIO-SIGN emits governed, abstaining research objects; it never names or recommends a clinical action.
Why it exists
Mutational-signature analysis is dominated by excellent Python/R tools, but none ships for the JavaScript ecosystem, none carries a governance/abstention discipline, and none unifies the glioma-specific process modalities (RNA-SBS, CSF-MRD monitoring, mechanism anchors, therapy contraindication) that glioma research and tumour boards actually need. GLIO-SIGN does all three.
On the shared, benchmarkable tasks it is competitive with the state of the art (pinned-Docker,
reproducible — see docs/BENCHMARK.md):
- De-novo: matches SigProfilerExtractor (0.9995 vs 0.9997) at ~17x the speed.
- Assignment: beats MuSiCal 30/30 cohorts on glioma panels (sign-test p = 1.9e-9).
- Honest limitation: MuSiCal's minimum-volume NMF still wins de-novo separation of highly collinear signatures — a certified open gap, stated plainly, not hidden.
Full landscape and per-tool verdicts: docs/COMPETITIVE.md.
Install
npm install glio-sign
# or: pnpm add glio-sign # or: yarn add glio-signSelf-contained: one ESM + one CJS bundle, self-contained TypeScript types, zero runtime dependencies. Node >= 20.
Quick start (pure TypeScript, no Python)
import {
GLIOMA_FAMILIES, // clock | TMZ | MMR | polymerase | ROS | other | ...
familyOf, // COSMIC signature name -> glioma process family
MODULE_ID, // "glio-sign"
MODULE_VERSION,
} from 'glio-sign';
familyOf('SBS11'); // -> "TMZ" (temozolomide)
familyOf('SBS6'); // -> "MMR" (mismatch-repair deficiency)
GLIOMA_FAMILIES; // the fixed family taxonomyGoverned report (abstains rather than fabricates)
import { buildSignatureReport } from 'glio-sign';
// buildSignatureReport(inputs, deps) returns a SignatureReport that has passed the
// 4-layer guardEmission (prohibited-language + hard-stops + gates + final-invariant),
// carries provenance dataClass on every artifact, and ABSTAINS below the detectability
// floor instead of guessing. See docs and types for the input/dep contract.Numeric sidecar (optional, high fidelity)
Heavy de-novo extraction and the full Poisson/KL numeric kernels run in an optional Python FastAPI sidecar; the TS core and the sidecar agree to 1e-6. The typed client:
import { GlioSignClient } from 'glio-sign';
const client = new GlioSignClient({ baseUrl: 'http://localhost:8000' });
// client.refit(...), client.extract(...), etc. — each returns a provenance-wrapped envelope.Boot the sidecar (needs the Python package / Docker image):
npx glio-mmb-sign serve --port 8000
npx glio-mmb-sign doctor # environment + readiness check
npx glio-mmb-sign routes # list the governed endpointsWhat is inside
| Layer | What | |---|---| | Catalogs | SBS-96, RNA-SBS-192, ID-83, CN-48, SV-32, DBS-78, mtDNA (COSMIC v3.6) | | Core | opportunity correction, Poisson-KL refit, de-novo NMF/mvNMF, presence test, over-identification, negative-control calibration | | Glioma modules | mechanism anchors (C1), common-vs-rare (C2), radiotherapy imprint (F1), lineage-conditional prognosis (F2), therapy contraindication (F3), CSF/TISF MRD monitor (D1), clustered/kataegis (A3), single-cell (B1) | | Governance | 4-layer guardEmission, abstention, dataClass provenance, hash-chained audit |
Deployment
npm is one of several supported form factors — see docs/DEPLOYMENT.md
for the npm library, the PyPI wheel (to rival pip install musical directly), the Docker
image, the REST/OpenAPI service, and the GitHub Action, with a per-use-case guide.
License
Proprietary, research use only. See LICENSE. Bundled reference dictionaries carry
their own upstream licenses; gated datasets are not redistributed.
