VariFuse

Somatic variant pathogenicity classifier that fuses ESM-2 protein embeddings with tabular biological features.

Two models are trained on the same feature set, CV split, and threshold, so the comparison is fair:

• LightGBM over biological, structural, and ESM-derived features
• Cross-Attention fusion net that cross-attends scalar features against ESM-2 residue tokens (seed-averaged ensemble, EMA weights)

Features

• Predictors: REVEL, CADD, SIFT, PolyPhen2, GERP++, phyloP
• Structural (AlphaFold): SASA, relative SASA, pLDDT, domain/site flags
• Protein LM: ESM-2 (esm2_t33_650M_UR50D, 1280-d) embeddings + variant score
• Cancer annotations: cancer-gene / tier1 / oncogene / TSG flags

Leakage-prone columns are dropped before modeling. A leakage-audit config also ablates the meta-predictor scores to measure Type-1 circularity.

Pipeline

Scripts in src/ run in order. Each reads the previous output and writes to its own outputs/ subfolder.

#	Script	Stage
01	01_dbnsfp_processor.py	Parse/filter dbNSFP
02	02_remove_missing_values.py	Drop missing values
03	03_remove_duplicates.py	Deduplicate variants
04	04_feature_engineering.py	UniProt + AlphaFold features
05	05_remove_leakage.py	Drop leaking columns
06	06_clean_and_finalize.py	Final cleaning
07	07_dataset_balancing.py	Balance classes
08	08_prepare_esm_dataset.py	Map sequences, ref-AA check
09	09_prepare_external_esm_dataset.py	Build external ClinVar/DMS set
10	10_extract_esm_features.py	ESM-2 embeddings (internal + external)
11	11_train_and_evaluate.py	LightGBM vs Cross-Attention (gene-level CV, SHAP, McNemar)
12	12_external_validation.py	Gene-disjoint external validation
13	13_generate_figures.py	Figures

Stage 09 runs before 10 so both sets get embedded in one pass:

python src/10_extract_esm_features.py --dataset both

Setup

git clone https://github.com/asifahamed11/VariFuse.git
cd VariFuse
pip install -r requirements.txt
export VARIFUSE_DATA_DIR=/path/to/your/Datasets

ESM extraction uses fair-esm + torch. A CUDA GPU is recommended (FP16 auto). See data/README.md for the expected data layout.

Run

python src/09_prepare_external_esm_dataset.py
python src/10_extract_esm_features.py --dataset both
python src/11_train_and_evaluate.py
python src/12_external_validation.py
python src/13_generate_figures.py

Evaluation

5-fold gene-level StratifiedGroupKFold with a nested inner fold for early stopping. Thresholds maximize MCC subject to recall ≥ 0.90, falling back to max-F₂. Reports MCC, AUROC, AUPRC, Brier, precision/recall/F1, and a McNemar test between the two models. SHAP is computed for LightGBM.

Outputs

• outputs/10_model_evaluation/ , results.json, comparison_table.csv, oof_predictions.npz, shap_values_*.npz
• outputs/12_external_validation/ , external_validation.json, external_validation_table.csv, external_predictions.npz
• figures/ , fig01–fig12

License

MIT. See LICENSE.