IIT Mandi
CHARA
Computational and Physical Genomics Laboratory · Indian Institute of Technology Mandi

Thermodynamic Graph Laplacian Manifold Alignment for Survival Inference

Chara resolves cross-platform transcriptomic domain shift by computing structural heat diffusion kernels across patient gene expression manifolds—enabling a frozen 4,337-gene Cox Proportional Hazards signature to execute on unseen clinical cohorts with zero data leakage.

Frozen Signature
4,337 Genes

Thermodynamic STRING Laplacian manifold features

Concordance Gain
+0.268 C-Index

Peak OOD gain over DeepSurv & Random Survival Forest

Test Integrity
Zero Leakage

Out-of-sample frozen cohort clinical validation

Inference Model
58 Biomarkers

Active regularized Coxnet hazard coefficients

BIOPHYSICAL LANDSCAPE

The Challenge of Cross-Platform Transcriptomics

IIT Mandi Research Focus

1. Sequencing Platform Discrepancies

High-throughput transcriptomics has revolutionized cancer prognostics. However, expression measurements vary wildly between Illumina RNA-Seq (HiSeq/NovaSeq), Affymetrix GeneChip Microarrays, and Agilent platforms. Differences in cDNA amplification, read depth, and probe hybridization introduce systematic non-biological shifts.

When a Cox Proportional Hazards model trained on TCGA Illumina data evaluates a local hospital's Affymetrix cohort, hazard predictions breakdown completely, dropping performance to random guessing.

2. Why Traditional Normalization Fails Clinical Testing

Algorithms like ComBat or Quantile Normalization rely on pooling training and testing samples together to adjust batch means. In clinical diagnostic practice, patient samples are received individually or in small batches.

Pooling test samples with training data leaks future patient statistics into the diagnostic model—violating out-of-sample machine learning rules and inflating performance metrics unrealistically.

Chara Thermodynamic Data Lifecycle

STAGE 1

Cohort Input CSV

Unadjusted patient transcriptomic counts (HGNC gene symbols).

STAGE 2

Graph Laplacian

Computes Gaussian affinity matrix W and degree matrix D.

STAGE 3

Heat Kernel Diffusion

Spectral smoothing kernel Ht = exp(-tL) dissipates platform noise.

STAGE 4

Survival Projection

Evaluates risk scores on frozen 4,337-gene Cox PH weights.

COMMUNITY RELEASES chara-survival v0.1.9 🤗 Hugging Face Hub

Official PyPI & Hugging Face Model Hub

Load the frozen Chara model weights directly in Python or install the library: pip install --upgrade chara-survival.

🤗 Hugging Face Hub PyPI Release
EMPIRICAL PERFORMANCE & SCIENTIFIC BENCHMARKING

Comprehensive Cross-Cohort Benchmarks

Evaluating Concordance Index (C-Index), 5-Year Time-Dependent AUC, and Brier Error Calibration across 6 international cohorts and multiple physical sequencing platforms.

What Datasets Were Used & Why Benchmarking Matters

1. The 6 Multi-Center Cohorts

We evaluated Chara on diverse discovery and out-of-distribution test cohorts:

  • • TCGA-PAAD: RNA-Seq (HiSeq), n=177 (Training)
  • • ICGC-PACA (AU): RNA-Seq, n=269 (Held-out)
  • • GSE15471: Affymetrix HG-U133 Plus 2.0, n=78
  • • GSE28735: GeneChip Human Gene 1.0, n=90
  • • GSE57495: Agilent Human Genome 4x44K, n=63
  • • GSE31210: Zero-Shot Lung Microarray, n=226

2. Why Platform Shift Fails ML

Microarrays measure fluorescence from optical probe hybridization with probe saturation limits, whereas RNA-Seq counts cDNA sequencing reads.

Without thermodynamic manifold alignment, conventional machine learning models (DeepSurv, Random Survival Forests) experience domain collapse, dropping from ~0.78 C-Index on RNA-Seq to random 0.50 on microarrays.

3. Clinical Trial Zero Leakage

In prospective oncology trials, patients arrive sequentially. Methods that perform batch harmonization (like ComBat) require pooling future test samples with past data, leaking test statistics.

Chara solves this by applying frozen graph Laplacian smoothing individually to each incoming cohort without accessing training labels.

What is Concordance Index (C-Index)?

The C-Index measures a survival model's ability to correctly rank patient risk. It evaluates whether a patient who died earlier was correctly assigned a higher risk score than a patient who survived longer.

  • • 0.50 = Random guessing (equivalent to a coin flip).
  • • 0.70+ = Strong clinical utility for patient stratification.
  • • 0.784 (Chara Peak) = State-of-the-art predictive discrimination across independent cohorts.

What is Brier Error Score?

The Brier Score assesses the calibration accuracy of predicted survival probabilities over time. It calculates the mean squared difference between predicted survival probability and actual patient survival status.

  • • Lower is better (0.0 represents zero prediction error).
  • • Unadjusted models suffer from error drift (> 0.45) at 5 years.
  • • Chara maintains tight calibration error (< 0.18) across 5 years.

Concordance Index (C-Index) Comparison

Comparing Unadjusted Cox vs ComBat vs Chara Laplacian across 5 cohorts

5-Year Brier Error Score Drift

Calibration error over time (Lower = Better)

Detailed Cross-Platform Benchmark Table

Each cohort represents an independent international study on a different physical platform.

Validation Cohort Platform Sample Size (n) Unadjusted Cox ComBat Adjustment Chara Laplacian (Ours)
ICGC-PACA (AU) RNA-Seq (Illumina HiSeq) n = 269 0.531 0.682 0.784 (+0.253)
GSE15471 Affymetrix HG-U133 Plus 2.0 n = 78 0.508 0.641 0.762 (+0.254)
GSE28735 GeneChip Human Gene 1.0 ST n = 90 0.522 0.665 0.771 (+0.249)
GSE57495 Agilent Whole Human Genome n = 63 0.495 0.628 0.758 (+0.263)
CLINICAL APPLICATIONS & LITERATURE CITATIONS

Case Studies & Literature Papers

Real-world oncology applications demonstrating Chara in precision medicine, early cancer detection, and drug discovery.

Case Study 1 Pancreatic Adenocarcinoma

Pancreatic Ductal Adenocarcinoma (PDAC) Survival Stratification

Pancreatic ductal adenocarcinoma is notorious for high inter-patient heterogeneity and low 5-year survival rates (< 9%). Applying models trained on TCGA-PAAD directly to Australian hospital cohorts (ICGC-PACA) traditionally resulted in severe misclassification.

Chara Impact: By constructing a graph Laplacian manifold across 269 ICGC patient expression profiles, Chara successfully aligned the 4,337-gene signature, achieving a C-Index of 0.784 and clearly stratifying high-risk vs. low-risk survival arms.

Cancer Cell (2017) · TCGA Network Read Paper
Case Study 2 Cross-Platform Validation

Cross-Platform Microarray to RNA-Seq Validation

Translating legacy biomarker signatures developed on 1990s Affymetrix HG-U133 Microarray chips (e.g. GSE15471) onto modern Illumina NovaSeq RNA-seq platforms poses a major hurdle for pharmaceutical companies.

Chara Impact: Chara heat diffusion smoothing projects Affymetrix probe intensities onto the same thermodynamic manifold as RNA-seq transcript counts in a single mathematical step, enabling cross-generational clinical validation.

Hepatogastroenterology (2008) · Badea et al. Read Paper
Case Study 3 Precision Oncology

Early-Stage Cancer Detection & Precision Oncology

In early-stage oncological screening, subtle expression signals are easily masked by sequencing noise. Chara manifold projection isolates true structural variance from platform noise, aiding computational target discovery.

Chara Impact: Allows research labs to compute baseline hazard curves for early-stage surgical resection candidates, prioritizing patients for targeted adjuvant therapy regimens.

Nature (2009) · Stratton et al. Read Paper
Case Study 4 Pipeline Integration

PyTorch & Scanpy Pipeline Integration

Bioinformatics teams building deep survival autoencoders or single-cell pseudo-bulk models can use `chara-survival` as a zero-leakage pre-processing step inside PyTorch data loaders or AnnData objects.

Chara Impact: Eliminates the need for manual batch correction pipelines in Python, returning aligned matrices in under 3.2 seconds.

Genome Biology (2018) · Wolf et al. Read Paper
HIGH-PERFORMANCE COMPARATIVE SURVIVAL ENGINE

Interactive Inference & Model Comparison

Upload a cohort CSV file to evaluate real survival trajectories and compare Chara against baseline survival architectures side-by-side.

DATA FORMAT SPECIFICATION Cohort Transcriptomics Matrix

What Data to Upload & How to Structure Your CSV

1. Column & Row Structure

Column 1: Patient_ID (e.g. PATIENT_01, TCGA-LUAD-01, GSM12345).

Columns 2 to N: Standard HGNC Gene Symbols (e.g. CCL20, DKK1, MS4A1, FAIM2, TP53, KRAS, EGFR, etc.).

2. Expression Units

Chara accepts unadjusted expression counts across all standard measurement platforms:

• RNA-Seq: TPM, FPKM, RPKM, or log2(CPM+1).

• Microarray: Log2 RMA / MAS5 probe intensity values.

3. Zero Data Leakage

You do not need to perform manual batch correction before uploading.

Chara projects your expression matrix onto the frozen 4,337-gene thermodynamic manifold in real time without training label leakage.

Example CSV Structure (First 4 Columns Preview):
Patient_ID CCL20 DKK1 MS4A1 FAIM2 ... [remaining genes]
PATIENT_01 14.82 12.45 2.10 1.95 ...
PATIENT_02 3.12 2.80 16.40 14.90 ...

Cohort Input File

CSV Files Only

Drag & Drop Cohort CSV Here

Rows: Patient ID | Columns: HGNC Gene Symbols

or click to browse
SELECT ARCHITECTURE:
4 Modes
ACTIVE EVALUATION ARCHITECTURE:
Chara Thermodynamic Laplacian (Ours): Spectral heat diffusion operator Ht = e-tL on MD-weighted STRING manifold with 58 regularized Coxnet biomarkers. (C-Index: 0.784, Zero-leakage OOD Transfer)

No Active Cohort Data

Upload a transcriptomic CSV on the left to compute real thermodynamic survival trajectories.

LABORATORY AUTHORS & ACKNOWLEDGMENTS

Research Team & Acknowledgments

Computational and Physical Genomics Laboratory · IIT Mandi.

Dr. Kharerin Hungyo

Dr. Kharerin Hungyo

Principal Investigator · Assistant Professor, IIT Mandi

PhD from IIT Bombay with postdoctoral research fellowships at Penn State University and ENS de Lyon. Directs the Computational & Physical Genomics Lab studying biophysical chromatin organization, nucleosome positioning, and AI/ML survival architectures.

kharerin@iitmandi.ac.in
Sharon Melhi

Sharon Melhi

Research Intern · CPG Lab, IIT Mandi

Creator and developer of Chara Survival. Research Intern at the Indian Institute of Technology Mandi working under Dr. Kharerin Hungyo.

Research Interests:
Drug Discovery Computational Oncology Precision Medicine Early Cancer Detection Computational Biology
GitHub · LinkedIn · Email

Special Acknowledgment

Khushi Mhamane

My heartfelt thanks to Khushi Mhamane for her constant encouragement, thoughtful discussions, and belief in this research from its earliest stages.

— Sharon Melhi