Chara Survival

Thermodynamic Graph Laplacian Survival Inference for Transcriptomic Oncology

A frozen 4,337-gene thermodynamic intersection signature that transfers across sequencing platforms — zero retraining required.

🌟 The Problem

Standard survival models — Cox proportional hazards, Random Survival Forests, DeepSurv — are trained on RNA-seq cohorts (typically TCGA) and collapse catastrophically when applied to microarray data (GEO). The cause is structural: uncorrected platform variance and feature distribution shift corrupt the learned risk landscape. The concordance index, already noisy at 0.50–0.55 on in-distribution data, degrades to random or sub-random when the platform changes.

This failure is not a modelling artefact. It is a physical problem — raw transcript counts do not encode the molecular interaction topology that determines biological function. Chara corrects this at the feature level, before training even begins.

💡 The Solution

Chara grounds gene expression features in thermodynamic graph Laplacians derived from MARTINI 3 coarse-grained molecular dynamics (MD) simulations. Specifically:

1. MARTINI 3 MD trajectories are run for key oncogenic protein systems (KRAS, CMYC/MAX, PTPN11, MUT-TP53) across biological replicates.
2. Exponential heat kernels are computed from the symmetrised graph Laplacian of the STRING protein–protein interaction network, weighted by MD-derived edge variances via the Chara exponential operator: $$W_{\text{Chara}}(i, j) = W_{\text{STRING}}(i, j) \cdot \exp\left(\tau \cdot Z(\sigma^2_{ij})\right)$$
3. The resulting thermodynamic Laplacian representation produces a platform-invariant feature space via spectral heat diffusion $H_t = \exp(-tL)$ — the spectral geometry of molecular interaction rather than raw transcript abundance.
4. A frozen CoxNet trained on TCGA-LUAD using these 4,337 thermodynamically-stabilised features (58 active regularized biomarkers) is applied directly to unseen cohorts without fine-tuning.

The result is a model that generalises across the RNA-seq ↔ microarray boundary as a matter of physical principle, not statistical luck.

🔬 Benchmark Performance

Evaluated zero-shot on GSE31210 (Affymetrix Human Genome U133 Plus 2.0, n = 226, completely held-out lung adenocarcinoma microarray cohort — never seen during training):

Chara achieves a +0.267 absolute improvement in OOD C-index over the next-best deep learning baseline, with monotonically improving time-horizon AUC — an unusual and clinically meaningful signature of robust calibration rather than threshold overfitting.

🚀 Quickstart: Python Package

Installation

pip install --upgrade chara-survival

1-Line Python Inference (Auto-Downloads from Hugging Face Hub)

import pandas as pd
import chara

# 1. Load the frozen pretrained model (auto-fetches from Hugging Face Hub if needed)
model = chara.load_model()

# 2. Ingest your patient cohort CSV (Rows: Samples, Columns: HGNC Gene Symbols)
df = pd.read_csv("patient_expression.csv", index_col=0)

# 3. Predict full Kaplan-Meier survival curves & patient hazard risk scores
curves, times, risk_scores = model.predict_survival_curves(df)

print(f"Evaluated {len(df)} patients across {len(times)} months.")
print(f"Mean Risk: {risk_scores.mean():.4f} | 5-Year Survival: {curves[:, 60].mean() * 100:.1f}%")

Load Directly from Hugging Face Hub

from huggingface_hub import hf_hub_download
import joblib

model_path = hf_hub_download(repo_id="SharonMelhi/chara-survival", filename="chara_model_4337.pkl")
bundle = joblib.load(model_path)
print(f"Loaded Chara bundle with {len(bundle['genes'])} features and {len(bundle['non_zero_genes'])} biomarkers.")

🌐 Interactive Web Portal & Comparison Sandbox

Try the live, zero-install interactive inference suite and multi-model benchmark sandbox:

👉 https://chara-frontend.vercel.app

🧬 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

@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}
}

👥 Authors & Lab

Dr. Kharerin Hungyo Principal Investigator Computational & Physical Genomics Lab Indian Institute of Technology Mandi kharerin@iitmandi.ac.in

Sharon Melhi Computational Biologist Creator, Chara Survival LinkedIn · Email

Special thanks to Khushi Mhamane for her constant encouragement, thoughtful discussions, and belief in this research from its earliest stages. — Sharon Melhi

📄 License

Released under the MIT License. © 2026 Sharon Melhi.