DriverSub-SVM

DriverSub-SVM is a machine learning framework for cancer subtype classification.
It integrates the heterogeneity of personalized driver genes with the commonalities of global driver genes, enabling the identification of subtype-specific biomarkers.
This comprehensive approach enhances the accuracy and interpretability of cancer classification.

---

1. Environment Requirements

• Python 3.9
• Required libraries:
  • numpy
  • pandas
  • scikit-learn
  • scipy
  • tqdm
  • optuna

---

2. Input Data

We provide processed data for three cancer types: BRCA, THCA, and STAD.

BRCA data clarification: The PAM50 subtype dataset contains 1218 breast cancer samples, while the multi-omics data (mutation, CNV, expression) includes 1101 samples. After intersecting these two sources, 1089 samples with both subtype labels and complete multi-omics features were retained for model training and evaluation. This number is consistent with Table 1 in the manuscript, the specific dimensions of the other two cancer datasets are provided in the manuscript.

Files:

• InfluenceGraph_Sig_Fls_inter_vstCNV
  A protein functional interaction network constructed by integrating SignaLink3 and Reactome.
  Gene pairs with concurrent interactions in both networks are retained.

• MutMatrix_Sig_Fls_inter_vstCNV
  Combines somatic mutation and CNV data.
  A gene is labeled as "aberrant" (value = 1) if it has a mutation or CNV value of 1 or -1 in a sample; otherwise, it is 0.

• OutMatrix_Sig_Fls_inter_vstCNV
  VST-normalized transcriptomic expression values are converted into z-scores.
  Genes with absolute z-score ≥ 2δ are marked as abnormally expressed (1), otherwise 0.

• Subtype Labels:

  Cancer Type	Subtype Definition
  BRCA	PAM50: LumA, LumB, Her2, Basal, Normal
  THCA	Histological types: Classical, Follicular, Tall cell, Other
  STAD	Expression-based: GS, CIN, EBV, MSI

---

3. Source Code Description

File Name	Function
Data.py	Loads input datasets.
Random_walk.py	Implements a random walk to assess interactions between mutated and dysregulated genes.
personalized.py	Applies the BPR algorithm to personalize gene ranking based on mutation profiles.
global_driver.py	Uses the Condorcet method to identify global driver genes across samples.
model_training_cross.py	Trains an OAO-MSVM classifier for subtype prediction and performs evaluation.

---

4. How to Run

Run the following command from the Code directory: python model_training_cross.py

This script will:

Perform k-fold cross-validation

BRCA: 10-fold (due to larger sample size)

THCA & STAD: 5-fold (due to smaller sample sizes)

Output average evaluation metrics (accuracy, precision, recall, F1-score)

Generate SHAP values and KEGG pathway input data for downstream interpretability analysis.

5.Notes

This project supports reproducibility. All datasets and preprocessing steps used for BRCA, THCA, and STAD are included or described in this repository. See the /data/ and /Code/ directories for more details.

For questions or issues, please refer to the supplementary materials or contact the authors.