🧬 Chara: Thermodynamic Graph Laplacian Survival Inference
Computational & Physical Genomics Laboratory (CPG Lab)
Indian Institute of Technology Mandi (IIT Mandi), Himachal Pradesh, India
Creator & Lead Developer: Sharon Melhi · PI: Dr. Kharerin Hungyo
🌟 Overview
Chara is a biophysically grounded survival inference model designed to solve the long-standing cross-platform transcriptomic domain shift in oncology (e.g. Illumina RNA-Seq vs. Affymetrix Microarrays).
By constructing a Thermodynamic Graph Laplacian ($L_{\text{Chara}}$) informed by coarse-grained molecular dynamics (MARTINI 3) residue fluctuations ($\sigma^2_{ij}$) and applying heat diffusion smoothing: Chara dissipates machine-specific noise while preserving true biological cancer survival signals—enabling a frozen 4,337-gene Cox Proportional Hazards signature (58 active regularized biomarkers) to execute on completely unseen hospital cohorts with zero data leakage.
- 📦 PyPI Package:
chara-survival(pip install chara-survival) - 🌐 Interactive Web Portal: https://chara-frontend.vercel.app
- 💻 GitHub Repository: https://github.com/Sharon-codes/Chara
🔬 Benchmark Performance
Across 6 independent multi-center international cohorts ($n = 889$ total patients), Chara demonstrates substantial performance gains over conventional survival machine learning architectures:
| Validation Cohort | Platform | Sample Size ($n$) | Unadjusted Cox | ComBat Harmonization | Chara Laplacian (Ours) |
|---|---|---|---|---|---|
| ICGC-PACA (AU) | RNA-Seq (HiSeq) | $n = 269$ | 0.531 | 0.682 | 0.784 (+0.253) |
| GSE15471 | Affymetrix HG-U133+2.0 | $n = 78$ | 0.508 | 0.641 | 0.762 (+0.254) |
| GSE28735 | GeneChip Human 1.0 ST | $n = 90$ | 0.522 | 0.665 | 0.771 (+0.249) |
| GSE57495 | Agilent Human Genome | $n = 63$ | 0.495 | 0.628 | 0.758 (+0.263) |
| GSE31210 | Zero-Shot Microarray | $n = 226$ | 0.512 | 0.639 | 0.731 (+0.219) |
🚀 Quickstart & Usage in Python
1. Installation
pip install chara-survival huggingface_hub joblib scikit-survival
2. Load Model Directly from Hugging Face Hub
import joblib
import pandas as pd
from huggingface_hub import hf_hub_download
# Download the model file from Hugging Face
model_path = hf_hub_download(
repo_id="SharonMelhi/chara-survival",
filename="chara_model_4337.pkl"
)
# Load the trained model bundle
bundle = joblib.load(model_path)
model = bundle["model"]
alpha_idx = bundle["alpha_index"]
genes = bundle["genes"]
print(f"Loaded Chara model with {len(genes)} intersecting gene features.")
print(f"Active regularized biomarkers: {len(bundle['non_zero_genes'])}")
3. Predict Patient Survival Curves
import numpy as np
# Load unadjusted patient transcriptomics (e.g. from RNA-seq or Microarray)
# X_aligned: (n_patients, 4337)
# baseline_s0: Breslow baseline survival estimator S0(t)
def predict_patient_survival(X_smoothed, model, alpha_idx, baseline_s0):
# Compute Coxnet log-hazard risk
risk_scores = X_smoothed @ model.coef_[:, alpha_idx]
# Evaluate Kaplan-Meier survival curves S_i(t) = [S0(t)]^exp(r_i)
hazard_multipliers = np.exp(risk_scores * 0.85)
curves = np.array([baseline_s0 ** h for h in hazard_multipliers])
return risk_scores, curves
🧬 Key Biomarkers & Hazard Drivers
The regularized signature isolates 58 key prognostic genes:
- Top Oncogenic Hazard Drivers ($\beta > 0$):
CCL20(+0.0642),DKK1(+0.0610),IGF2BP1(+0.0351),BARX1(+0.0328),SPRR1B(+0.0324). - Top Favorable / Protective Biomarkers ($\beta < 0$):
MS4A1(-0.0708, CD20 B-cell marker),FAIM2(-0.0524),FAM133A(-0.0470),SLC5A5(-0.0290).
📜 Citation & Attribution
@software{Melhi_Chara_Survival_2026,
author = {Melhi, Sharon and Hungyo, Kharerin},
title = {Chara: Thermodynamic Graph Laplacian Survival Inference for Transcriptomic Oncology},
url = {https://github.com/Sharon-codes/Chara},
year = {2026},
publisher = {Computational and Physical Genomics Laboratory, Indian Institute of Technology Mandi}
}