🧪 LabBridge

Turn a messy plate-reader / assay export into a clean, standardized, QC-checked, ML-ready dataset — with an explainable column mapping, deterministic quality checks, automatic scientific validation, and a full audit trail.

▶️ 2-min walkthrough  ·  🔗 Try it live (no install)  ·  💻 Deterministic pipeline + Streamlit UI · 32 passing tests

---

The problem it solves

Every lab-software onboarding starts with the same manual chore: raw instrument files arrive with inconsistent headers, mixed units, odd date formats, missing wells, overflow reads, and replicate disagreements. Someone cleans that by hand, and the cleaned data is trusted on faith.

LabBridge automates that first mile and shows its work — so the cleaned data can be audited, and so the scientific validity of the data (not just its tidiness) is checked before anyone models on it.

What it does

1. Detect layout — long vs grid plate formats.
2. Map columns — exact / synonym / fuzzy / content-inference, each with a confidence score and a green / yellow / red review bucket. Ambiguous columns (e.g. Anlyst → operator, ~78%) are surfaced for a human to confirm — you stay in control of every column.
3. Normalize — µM/mM → nM, decimal commas, embedded units, mixed date formats → ISO 8601, well coordinates (a01 → A1), control-type spellings.
4. QC flag — missing / overflow / negative / invalid-well / duplicate-well, robust MAD outliers, replicate-CV disagreement, control sanity.
5. QC Intelligence — standard-curve fit, replicate reliability, anomaly detection, and Levey-Jennings control monitoring with Westgard rules, ending in a fitness verdict: clean data is necessary, but only validated data is safe to model on.
6. Audit + summarize — nothing changes silently; every transformation and flag is logged with before/after.

No database, no auth, no LLM. The engine is deterministic and testable; the Streamlit UI is a thin 6-step wizard that only calls the pipeline.

Run it locally (2 commands)

pip install -r requirements.txt
streamlit run labbridge/app.py

Then click Load sample plate export in the UI — a 96-well plate with realistic, deliberately-injected data issues.

Run the tests

pytest labbridge/tests -q          # 32 passing

Headless pipeline / CLI

python -m labbridge.pipeline labbridge/data/messy_plate_long.csv --out out/

Layout

labbridge/
├── core/
│   ├── schema.py      # canonical schema + column synonyms + unit tables
│   ├── parsers.py     # file read + layout detection
│   ├── mapping.py     # column -> canonical mapping with confidence scores
│   ├── normalize.py   # units / decimals / dates / wells / control types
│   ├── qc.py          # validation + robust statistical QC flags
│   ├── qc_intel.py    # standard curve, replicates, control monitoring, verdict
│   └── audit.py       # transformation log + data-quality summary
├── data/              # deterministic sample plate + injected-error log
├── tests/             # pytest suite (mapping / normalize / qc / pipeline)
├── pipeline.py        # headless end-to-end engine + CLI
└── app.py             # Streamlit 6-step wizard over the pipeline

Deliberate scope decisions

• Plate edge-effect detection is deferred. With few, heterogeneous samples per plate, outer-vs-interior comparison isn't statistically reliable; flagging it would be false confidence. Proper detection needs a uniformity plate or row/column detrending.
• No LLM. Mapping is deterministic and explainable. An LLM fallback for genuinely ambiguous headers is a later, optional, human-confirmed add-on.

License

MIT — see LICENSE.