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.
3. The resulting thermodynamic Laplacian representation produces a platform-invariant feature space — 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 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.

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.

Live Inference

A zero-install interactive inference interface is hosted on Hugging Face Spaces:

→ https://huggingface.co/spaces/Sharon-codes/Chara

Upload a patient-by-gene CSV (rows = samples, columns = HGNC symbols). The app aligns your cohort to the frozen 4,337-gene signature, computes risk scores, renders Kaplan–Meier–style survival curves, and returns a downloadable report — all in seconds.

Python Package

Installation

pip install chara-survival

Requires Python ≥ 3.10. The frozen model artefact (chara_model_4337.pkl) must be placed in your working directory or downloaded from the Releases page.

Programmatic Inference

import pandas as pd
from chara import CharaModel

# Load the frozen model
model = CharaModel.load("chara_model_4337.pkl")

# expression_matrix: patients × HGNC gene symbols
expression_matrix = pd.read_csv("patient_expression.csv", index_col=0)

# Align, scale, and infer
risk_scores, x_scaled, aligned_df, scaler, alpha_index = model.predict(expression_matrix)

print(f"Processed {len(expression_matrix)} patients")
print(f"Risk score range: [{risk_scores.min():.4f}, {risk_scores.max():.4f}]")

Survival Curve Extraction

import numpy as np

# Retrieve full survival functions for all patients
curves, times = model.survival_curves(x_scaled, alpha_index)

# Interpolate to clinical horizons
horizons = np.array([365.0, 1095.0, 1825.0])  # 1, 3, 5 years
survival = np.array([
    np.interp(horizons, times, row, left=1.0, right=row[-1])
    for row in curves
])
# survival[:, 0]  →  1-year survival probability per patient
# survival[:, 1]  →  3-year survival probability per patient
# survival[:, 2]  →  5-year survival probability per patient

Thermodynamic Graph Utilities

The chara package also exposes the graph primitives used during training:

from chara.graph import laplacian_from_edges, heat_kernel, exponential_chara_laplacian

# Construct a graph Laplacian from an edge list (e.g., STRING interactions)
L = laplacian_from_edges(edge_df, node_list, source="protein1", target="protein2", weight="score")

# Standard heat kernel (diffusion on Laplacian spectrum)
K = heat_kernel(L, diffusion_time=0.1)

# Chara exponential operator — weights edges by MD variance
L_chara = exponential_chara_laplacian(adjacency, edge_variance, tau=0.5)

API Reference

CharaModel

scale_external_expression

from chara import scale_external_expression

x, aligned, scaler = scale_external_expression(expression_df, feature_list)

Aligns an external expression matrix to a target feature list (zero-imputing missing genes, averaging duplicate symbols) and applies StandardScaler.

Input Specification

Repository Architecture

chara-survival/
├── chara/
│   ├── __init__.py          # Public API: CharaModel, scale_external_expression
│   ├── model.py             # Frozen CoxNet wrapper with strict feature alignment
│   ├── graph.py             # Thermodynamic Laplacian and heat-kernel operators
│   └── preprocessing.py    # Cross-platform StandardScaler pipeline
│
├── scripts/                 # Full research pipeline (01 → 19)
│   ├── 01_fetch_tcga.py     # TCGA-LUAD expression + survival data
│   ├── 02_generate_string.py
│   ├── 03_generate_chara.py # Thermodynamic Laplacian construction
│   ├── 04_chara_ood_validation.py
│   ├── 05_adversarial_poisoning.py
│   ├── 06_dirichlet_energy.py
│   ├── 07_biological_gsea.py
│   ├── 09_zeroshot_external_validation.py
│   ├── 11_clinical_frontier_metrics.py
│   ├── benchmark_frontiers.py
│   └── ...
│
├── assets/                  # Logos and figures
├── app.py                   # Gradio inference interface
├── chara_model_4337.pkl     # Frozen model artefact (4,337-gene signature)
├── requirements.txt
├── setup.py
├── pyproject.toml
└── LICENSE

Reproducibility

The complete research pipeline is contained in scripts/. Execution order follows the numeric prefix (01 → 19). MARTINI 3 MD trajectories require GROMACS ≥ 2023.3; all downstream graph construction and survival modelling is pure Python.

Key intermediate artefacts required to reproduce the frozen model from scratch:

Citation

If you use Chara in published research, please cite this repository until a formal preprint is available:

Sharon Melhi (2026). Chara Survival: Thermodynamic Graph Laplacian Survival Inference
for Out-of-Distribution Transcriptomic Oncology.
GitHub: https://github.com/Sharon-codes/Chara

People

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.