jang1563/cbrn-physics-features

CBRN Physics Features

Pre-computed physics-informed distributional features for pathogen-agnostic biological threat detection in gene expression data.

Overview

This dataset contains per-sample and per-group features computed from the shape of gene expression distributions rather than the identity of individual genes. The four core features — Gini coefficient, Shannon entropy, normalized entropy, and Zipf exponent — are platform-agnostic: they require no gene annotation, no probe mapping, and no platform-specific normalization.

These features enable:

• Cross-platform infection detection (AUROC 0.720 across 5 microarray/RNA-seq platforms)
• Radiation biodosimetry from tissue gene expression (AUROC 0.793)
• Dose-response modeling with monotonic feature-dose relationships

Dataset Structure

File	Description	Rows	Key Columns
infection/per_sample_features.csv	Per-sample features across 10 GEO infection datasets	1,954	gini, shannon_entropy, normalized_entropy, zipf_gamma, dataset, condition
infection/cross_dataset_transfer.csv	Cross-dataset transfer AUROC (train on one platform, test on another)	51	train_dataset, test_dataset, auroc, model
infection/validation_transfer.csv	Development-to-validation transfer with bootstrap 95% CI	12	train_dataset, test_dataset, auroc, ci_low, ci_high
radiation/per_sample_features.csv	Per-sample features for brain, spleen, and skin tissue	82	tissue, radiation, group, gini, shannon_entropy
radiation/group_features.csv	Taylor k exponent with bootstrap CI per condition group	12	tissue, group, taylor_k, ci_low, ci_high
dose_response/per_sample_features.csv	Dose-response features across radiation doses (0-13 Gy)	82	tissue, dose_gy, gini, shannon_entropy

Total: 2,193 rows across 6 CSV files.

Features

Feature	Formula	Interpretation	Range
Gini coefficient	Lorenz curve area	Expression inequality (0 = uniform, 1 = one gene dominates)	[0, 1]
Shannon entropy	H = -sum(p log p)	Expression diversity (nats)	[0, log n]
Normalized entropy	H / log(n_expressed)	Scale-free diversity, comparable across platforms	[0, 1]
Zipf exponent (gamma)	OLS slope of log(expr) vs log(rank)	Power-law decay rate of ranked expression	[-3, 0]
Taylor k	OLS slope of log(Var) vs log(Mean)	Noise scaling exponent (requires replicates)	[1, 2]

Source Data

Features were computed from publicly available gene expression datasets:

• Infection: 10 GEO datasets (GSE161731, GSE157103, GSE63990, GSE60244, GSE72829, GSE236713, GSE100150, GSE111368, GSE40012, GSE73072) spanning RNA-seq, Affymetrix, Illumina, and Agilent platforms
• Radiation: 3 NASA GeneLab datasets (OSD-202 brain, OSD-211 spleen, OSD-237 skin)

Usage

import pandas as pd

# Load infection features
features = pd.read_csv("infection/per_sample_features.csv")

# Binary label: infected vs healthy
features["infected"] = (features["condition"] != "Healthy").astype(int)

# Train a simple classifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score

X = features[["gini", "shannon_entropy", "normalized_entropy", "zipf_gamma"]]
y = features["infected"]
scores = cross_val_score(LogisticRegression(), X, y, cv=5, scoring="roc_auc")
print(f"5-fold AUROC: {scores.mean():.3f} +/- {scores.std():.3f}")

Code

Full analysis pipeline, vendored feature computation library, and figure generation scripts:

github.com/jang1563/cbrn-physics-features

Citation

@software{kim2026physics,
  author = {Kim, Jangwoo},
  title = {Physics-Informed Distributional Features for CBRN Threat Detection},
  year = {2026},
  url = {https://github.com/jang1563/cbrn-physics-features},
  license = {MIT}
}