Precomputed genomic / conservation / protein scores (CADD, phyloP, phastCons, ESM-1b, NT, GPN-MSA, HyenaDNA)
Biophysical protein features (BLOSUM62, Grantham, hydrophobicity, charge, volume, context)
Gene-disease association evidence from OpenTargets clinical trials

Models included

Model	Dataset	Features	AUROC	F1	Size
`xgb_precomputed.json`	songlab/clinvar	8 precomputed scores	0.982	0.941	~391 KB
`rf_protein.pkl`	Rain021217/clinvar-pathogenicity	20 protein features	0.888	0.509	~19 MB
`lr_protein.pkl`	Rain021217/clinvar-pathogenicity	20 protein features	0.880	0.356	~1 KB

Quick start

python

import pickle
import numpy as np
import xgboost as xgb

# Load XGBoost model (best AUROC)
model = xgb.XGBClassifier()
model.load_model("xgb_precomputed.json")

# Load preprocessing artifacts
with open("xgb_precomputed_imp.pkl", "rb") as f:
    imputer = pickle.load(f)
with open("xgb_precomputed_scaler.pkl", "rb") as f:
    scaler = pickle.load(f)

# Example input: [GPN-MSA, CADD, phyloP-100v, phyloP-241m, phastCons-100v, ESM-1b, NT, HyenaDNA]
# (obtain from VEP / precomputed tracks / songlab/clinvar)
sample = np.array([[1.5, 0.3, 0.7, -1.2, 0.1, -2.5, -0.8, 0.0]])
sample_imp = imputer.transform(sample)
sample_scaled = scaler.transform(sample_imp)
proba = model.predict_proba(sample_scaled)[0, 1]
print(f"Pathogenic probability: {proba:.3f}")

Batch inference

python inference.py --model xgb --input my_variants.csv --output scored_variants.csv

Gradio demo

python app.py

Data provenance

songlab/clinvar — Precomputed missense variant scores from Cheng et al. (2023) Science (AlphaMissense), Notin et al. (2022) (Tranception), Frazer et al. (2021) Nature (ESM-1b), GPN-MSA, Nucleotide Transformer, HyenaDNA, CADD, phyloP, phastCons.
Rain021217/clinvar-pathogenicity-prediction-dataset — Curated protein-level biophysical features derived from ClinVar missense variants with UniProt/VEP annotations.
songlab/clinvar_vs_benign — ClinVar variant_summary filtered to missense variants with review_status and consequence annotations.
opentargets/clinical_evidence — 2.7M clinical evidence records linking genes to diseases via clinical trials.
gonzalobenegas/hgmd-omim-gnomad — Variants annotated with OMIM/HGMD traits for gene-disease grounding.

Extended experiments

Experiment 4: Gene-disease association scoring

Built from opentargets/clinical_evidence:

71,419 unique target-disease pairs scored
Evidence score formula: 0.4*(max_phase/4) + 0.4*min(n_trials/10,1) + 0.2*min(n_datasources/5,1)
Ranges 0.04–1.00 (higher = stronger clinical evidence)

Experiment 5: Combined variant-to-therapy scoring

therapy_score = 0.5*pathogenicity_proba + 0.3*has_known_trait + 0.2*consequence_severity

Tier assignment:

Tier 1 (Actionable): pathogenicity ≥ 0.9 + known disease association
Tier 2 (High Risk): pathogenicity ≥ 0.7
Tier 3 (Moderate): pathogenicity ≥ 0.5
Tier 4 (Likely Benign): pathogenicity < 0.5

Files in this repository

File	Description
`xgb_precomputed.json`	Best model: XGBoost on precomputed scores (AUROC 0.982)
`rf_protein.pkl` / `lr_protein.pkl`	Fallback models on protein features (no precompute needed)
`inference.py`	Production inference script (`--model xgb` or `rf`)
`app.py`	Gradio demo (single variant + batch CSV scoring)
`gene_disease_evidence.csv`	71,419 gene-disease pairs with evidence scores
`plots/*.png`	ROC, PR, confusion matrix, feature importance, distributions, precision@k
`metrics_summary.json` / `extended_metrics.json`	Full experiment results
`research_memo.md`	Literature review + integration recommendation

Evaluation plots

See plots/ for all 6 evaluation figures.

Training details

Hardware: CPU-only (free-tier Hugging Face sandbox, 2 vCPU / 16 GB RAM)
Framework: scikit-learn + XGBoost
Train/test split: 80/20 stratified (exp2), provided train/test splits (exp3)
Imputation: median
Scaling: StandardScaler
Class balancing: class_weight="balanced" for RF/LR; dataset is naturally ~55:45 for exp2
Total training time: < 5 minutes for all experiments
Artifact size: ~19 MB total

Limitations

Only missense variants are covered; frameshift, nonsense, splice are out of scope for the protein-feature models.
The XGBoost model requires precomputed external scores (CADD, ESM-1b, etc.). If these are unavailable, fall back to the Random Forest on protein features.
Models were trained on ClinVar 2023–2024 data; performance may drift on newer submissions.
No clinical phenotype integration (OMIM, HPO) — this is planned for future OmniBiMol backend integration.

Integration into OmniBiMol

Patient VCF
    │
    ▼
[VEP Annotation]
    │
    ├─► Precomputed scores available? ──► XGBoost (AUROC 0.982)
    │
    └─► Only protein sequence? ──► Random Forest (AUROC 0.888)
    │
    ▼
[Gene-Disease Lookup] ──► OpenTargets evidence score
    │
    ▼
[Tier Assignment] ──► Tier 1/2/3/4
    │
    ▼
[OmniBiMol Backend API] ──► Top-k + confidence + evidence
    │
    ▼
[UI Report] ──► Sortable table → drug repurposing → wet-lab handoff

Citation

bibtex

@misc{omnibimol_variant_priority,
  title = {OmniBiMol Variant Priority Pipeline},
  author = {OmniBiMol Team},
  year = {2025},
  howpublished = {\url{https://huggingface.co/omshrivastava/omnibimol-variant-priority}}
}

License

MIT