dnabert2-clinvar-pathogenicity

DNABERT-2 — ClinVar Variant Pathogenicity Classifier

Fine-tuned from zhihan1996/DNABERT-2-117M on ClinVar high-confidence (≥2 review stars) single nucleotide variants (SNVs) for binary pathogenicity classification.

Test-set metrics

Gene-level 80/10/10 split on GRCh38 SNVs (no gene appears in both train and test). Test set is 16.3% pathogenic — class-weighted loss used during training.

Cancer gene panel (BRCA1/2, TP53, MLH1/2, MSH2/6, ATM, CHEK2, PALB2, APC, PTEN, RAD51C): AUROC 0.819 · AUPRC 0.733 · F1 0.629

Input format

A 512 bp DNA string (characters A/C/G/T only), centered on the variant position with the ALT allele introduced at index 255 (0-based centre of the window). Extracted from the GRCh38 reference genome using a ±256 bp window.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tok = AutoTokenizer.from_pretrained("whitedevil0089devil/dnabert2-clinvar-pathogenicity", trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
    "whitedevil0089devil/dnabert2-clinvar-pathogenicity", trust_remote_code=True)
model.eval()

# sequence = 512 bp window from GRCh38, ALT allele injected at position 255
sequence = "ACGT" * 128   # replace with your real sequence

inputs = tok(sequence, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    logits = model(**inputs).logits

prob_pathogenic = torch.softmax(logits, dim=-1)[0, 1].item()
label = "Pathogenic" if prob_pathogenic >= 0.5 else "Benign"
print(f"Pathogenicity probability: {prob_pathogenic:.4f}  →  {label}")

Confidence tiers (used in the demo app)

Data and training

• Source: ClinVar variant_summary.txt (GRCh38, SNVs, ≥2 review stars, 2026)
• Labels: Pathogenic + Likely pathogenic → 1 · Benign + Likely benign → 0 · VUS dropped
• Split: by GeneSymbol (prevents gene-level leakage)
• Context: 512 bp window from GRCh38 reference, ALT allele injected
• Imbalance: class-weighted CrossEntropyLoss (16.3% pathogenic)
• Optimizer: AdamW · LR 3e-5 · warmup 6% · cosine decay · 5 epochs · batch 64
• Hardware: NVIDIA RTX PRO 6000 Blackwell (96 GB) · fp16

Limitations

• SNVs only — insertions, deletions, structural variants not supported
• Trained on germline ClinVar variants — not validated on somatic mutations
• Gene-level split means genes absent from ClinVar are unseen during training
• AUROC of 0.83 is good for ranking; the model should not be used as a sole clinical decision-making tool

Citation

@inproceedings{zhou2024dnabert2,
  title={DNABERT-2: Efficient Foundation Model and Benchmark for Multi-Species Genome},
  author={Zhou, Zhihan and Ji, Yanrong and Li, Weijian and Dutta, Pratik and
           Davuluri, Ramana and Liu, Han},
  booktitle={ICLR},
  year={2024}
}

ClinVar: Landrum et al., Nucleic Acids Research 2020.