bioscore

Biomedical scoring toolkit — reproducibility, data quality, and model readiness metrics for computational biology.

---

Why this exists

Most computational biology teams have no automated quality gates. A researcher finishes a notebook, shares it — and nobody can reproduce it. A data scientist trains a model — and it silently fails on edge cases. A team deploys to production — and there's no bias audit.

bioscore closes these gaps with three one-command checks that plug into any workflow.

Install

pip install bioscore

Requires Python 3.9+. No external dependencies for core functions.

Quick Start

from bioscore import reproducibility, data_quality, model_readiness

# 1. Check if your notebook is reproducible
reproducibility("analysis.ipynb")
# → {"score": 0.65, "issues": ["missing seed", "no version pinning"], "level": "partial"}

# 2. Assess dataset quality before training
data_quality("dataset.csv", domain="oncology")
# → {"completeness": 0.8, "consistency": 0.9, "overall": 0.85}

# 3. Verify model is ready for production
model_readiness("model.pkl")
# → {"score": 0.72, "ready": false, "gaps": ["no validation split", "no bias audit"]}

---

Target Audience & Daily Use

🧬 Computational Biology Researcher

Their morning: Opens Jupyter, runs yesterday's analysis on new data. Shares notebook with labmates. Submits paper.

The problem: Six months later, nobody — including themselves — can reproduce the results. Random seeds weren't set. Package versions weren't pinned. The data source was a colleague's Dropbox link that's now dead.

How bioscore helps:

from bioscore import reproducibility

result = reproducibility("my_analysis.ipynb")
if result["level"] != "full":
    print("Fix before sharing:", result["issues"])

They run this before sharing any notebook. It catches missing seeds, unpinned versions, undocumented data sources. The level field (full / partial / minimal) gives a quick pass/fail.

Install: pip install bioscore in their notebook environment (conda, venv, or Colab).

---

📊 Data Scientist in Pharma/Biotech

Their morning: Pulls clinical trial data. Checks for missing values. Trains a survival model. Sends to review.

The problem: Datasets have silent gaps — 30% missing in one column, inconsistent row counts, domain-specific quality rules nobody checks automatically.

How bioscore helps:

from bioscore import data_quality

result = data_quality("clinical_data.csv", domain="oncology")
if result["overall"] < 0.7:
    print(f"Quality too low ({result['overall']}), fix before training")

They run this as the first cell in every analysis notebook. Domain-aware checks (oncology, agriculture, general) apply different quality thresholds. Prevents garbage-in-garbage-out silently.

Install: pip install bioscore in their data science environment.

---

🚀 ML Engineer / MLOps

Their morning: Reviews model PR. Checks metrics. Approves deployment to staging. Monitors production.

The problem: Models reach production without validation splits, bias audits, or input schemas. Issues surface only in production — expensive and risky.

How bioscore helps:

from bioscore import model_readiness

result = model_readiness("model_v2.pkl")
if not result["ready"]:
    print("Block deployment:", result["gaps"])

They add this to CI/CD pipeline as a deployment gate. If ready is false, the pipeline blocks. Gaps like "no validation split" or "no bias audit" are surfaced as actionable items.

Install: Add bioscore to requirements.txt or pyproject.toml in the ML pipeline project.

---

API Reference

reproducibility(source: str) -> dict

Evaluates a notebook or script for reproducibility best practices.

Checks: random seed, package version pinning, data source documentation, output preservation, environment specification, comments, docstrings, logging.

Returns: {"score": float, "issues": list[str], "level": "full"|"partial"|"minimal"}

data_quality(source: str, domain: str = "general") -> dict

Assesses a CSV dataset for completeness and consistency.

Domains: "general", "oncology", "agriculture" — each applies domain-specific quality weights.

Returns: {"completeness": float, "consistency": float, "overall": float}

model_readiness(source: str) -> dict

Evaluates a pickled ML model artifact for production readiness.

Checks: validation split, bias audit, performance metrics, version tag, input schema, error handling, documentation, test coverage.

Returns: {"score": float, "ready": bool, "gaps": list[str]}

---

Innovation

bioscore is the first lightweight, zero-dependency Python toolkit that unifies three critical pre-deployment checks for computational biology:

1. Reproducibility scoring — not just linting, but a weighted score with actionable issues
2. Domain-aware data quality — oncology and agriculture have different quality standards than general data
3. Model readiness gate — a binary pass/fail with specific gaps, designed for CI/CD integration

No other package combines all three. Most teams cobble together custom scripts. bioscore makes it pip install bioscore and one function call.

---

License

MIT © K-RnD Lab