TF-DFE: Topology-Guided Multi-View Ensemble Learning for Evidence-Aware Somatic Variant Pathogenicity Classification in Cancer Genomics

Paper: Topology-Guided Multi-View Ensemble Learning for Evidence-Aware Somatic Variant Pathogenicity Classification in Cancer Genomics
Authors: Asif Ahamed, Md. Tanvir Hasan, Most. Alisa Tabassum, Ahammad Hossain*, Md. Kamruzzaman, Jayanta Das, A.H.M. Rahmatullah Imon
Corresponding Author: Ahammad Hossain — ahammadstatru@gmail.com

---

Overview

TF-DFE (Topo-Fractal Dynamic Fuzzy Ensemble) is a scalable, evidence-aware framework for classifying somatic single nucleotide variants (SNVs) as pathogenic drivers or benign passengers. The framework combines:

• Standard biological predictors - REVEL, CADD, SIFT, PolyPhen2, GERP++, phyloP, structural/UniProt features
• Fractal sequence features - Frequency Chaos Game Representation (FCGR) at k=3 and k=4 (320 dimensions)
• Topological features - Persistent homology via giotto-tda (6 dimensions)
• Dynamic ensemble selection - Enhanced KNORA-Eliminate strategy

Final dataset: 207,266 balanced somatic SNVs curated from dbNSFP v5.3a (GRCh37)
Performance: MCC = 0.8702 | AUROC = 0.9666 | AUPRC = 0.9768 | Accuracy = 93.48%

---

Repository Structure

TF-DFE/
├── README.md
├── requirements.txt
├── scripts/
│   ├── 01_somatic_variant_processor.py       # dbNSFP parsing, CIViC/COSMIC rescue, labeling
│   ├── 02_remove_missing_values.py           # Selective missing value removal
│   ├── 03_remove_duplicates.py               # Deduplication on chr/pos/ref/alt
│   ├── 04_structural_feature_engineering.py  # UniProt + AlphaFold2 + SASA features
│   ├── 05_remove_leakage_columns.py          # Leakage column removal
│   ├── 06_dataset_balancing.py               # Downsample to balanced 207,266 variants
│   ├── 07_eda.py                             # Exploratory data analysis + figures
│   └── 08_tf_dfe_model.py                    # Main TF-DFE model, evaluation, SHAP
└── data/
    └── Final_Dataset.csv                     # Balanced dataset (207,266 variants, 24 features)

---

Pipeline

Run the scripts in order. Each script outputs a CSV that is used as input to the next step.

dbNSFP5.3a_grch37.gz
        │
        ▼
01_somatic_variant_processor.py  →  somatic_variant_dbNSFP.csv  (19,613,687 variants)
        │
        ▼
02_remove_missing_values.py      →  somatic_variant_dbNSFP_Removes_missing_values.csv  (9,241,413)
        │
        ▼
05_remove_leakage_columns.py     →  somatic_variant_dbNSFP_Removes_leakage.csv
        │
        ▼
03_remove_duplicates.py          →  somatic_variant_dbNSFP_Removes_leakage_missing_values_Deduplication.csv  (6,461,982)
        │
        ▼
04_structural_feature_engineering.py  →  Final_Dataset_UniProt_Removes_leakage.csv  (adds UniProt/AlphaFold2 features)
        │
        ▼
06_dataset_balancing.py          →  data/Final_Dataset.csv  (207,266 balanced variants)
        │
        ▼
07_eda.py                        →  EDA_Results/  (figures and statistics)
        │
        ▼
08_tf_dfe_model.py               →  results_publication/  (model results, figures, SHAP)

---

External Data Requirements

These large/licensed files must be downloaded separately before running the pipeline.

File	Source	Used In
dbNSFP5.3a_grch37.gz	dbNSFP	Script 01
01-Jan-2026-VariantSummaries.tsv	CIViC Releases	Script 01
Cosmic_CancerGeneCensus_v102_GRCh37.tsv	COSMIC	Script 01
cmc_export.tsv	COSMIC Mutant Census	Script 01
oncokb_biomarker_drug_associations.tsv	OncoKB	Script 01
uniprot_sprot_human.dat	UniProt	Script 04
AlphaFold2 PDB files (AF-*-model_v*.pdb)	AlphaFold DB	Script 04

Note: Place all external files in the same directory as the scripts before running, or update the file paths inside each script accordingly.

---

Installation

git clone https://github.com/asifahamed11/TF-DFE.git
cd TF-DFE
pip install -r requirements.txt

Python 3.8 or higher is required.

---

Quick Start (Skip to Model Training)

If you only want to reproduce the model results using the pre-processed dataset:

# Dataset is already available in data/Final_Dataset.csv
# Update DATA_PATH in script 08 if needed, then run:
python scripts/08_tf_dfe_model.py

Results will be saved to results_publication/.

---

Dataset

The final balanced dataset (data/Final_Dataset.csv) contains 207,266 somatic SNVs (103,633 pathogenic / 103,633 benign) with the following features:

Feature Group	Features	Dimensions
Genomic coordinates	chr, pos, ref, alt	4
Pathogenicity scores	REVEL, CADD, SIFT, PolyPhen2, GERP++, phyloP	6
Cancer gene annotations	IS_CANCER_GENE, IS_TIER1, IS_ONCOGENE, IS_TSG	4
Structural features	SASA, RELATIVE_SASA, PLDDT_SCORE	3
UniProt functional sites	IS_IN_DOMAIN, IS_ACTIVE_SITE, IS_BINDING_SITE, IS_TRANSMEMBRANE, DISTANCE_TO_ACTIVE_SITE	5
Label	LABEL_PATHOGENIC (0=benign, 1=pathogenic)	1

FCGR (320-dim) and TDA (6-dim) features are computed on-the-fly inside 08_tf_dfe_model.py.

---

Reproducing Paper Figures

All figures in the manuscript are generated by 08_tf_dfe_model.py and saved to results_publication/ as both .png and .tiff.

EDA figures (Fig 4a, 4b, 4c, 5, 5b) are generated by 07_eda.py and saved to EDA_Results/.

---

License

This project is licensed under the MIT License. See LICENSE for details.

---

Contact

Ahammad Hossain (Corresponding Author)
Department of Computer Science & Engineering, Varendra University, Bangladesh
Email: ahammadstatru@gmail.com
ORCID: 0000-0001-9081-5905