vthawfeek/mtdna-fm-pathogenicity

mtDNA-FM Pathogenicity Adapter (LoRA r=4)

LoRA adapter for pathogenic variant prediction (binary) on top of vthawfeek/mtdna-foundation-model.

Trained on ClinVar pathogenic mtDNA variants (positive class) vs gnomAD common variants AF > 0.01 (negative class).

Usage

from mtdna_fm.model.model import MtDNAForVariantPathogenicity, MtDNAModel
from peft import PeftModel

base = MtDNAModel.from_pretrained("vthawfeek/mtdna-foundation-model")
model = MtDNAForVariantPathogenicity(base)
model = PeftModel.from_pretrained(model, "vthawfeek/mtdna-fm-pathogenicity")
model.eval()

LoRA Configuration

• r = 4, lora_alpha = 8
• target_modules: query, key, value, dense
• lora_dropout = 0.1
• pos_weight = 2.5 in BCE loss (class imbalance correction)

Task

Binary pathogenicity prediction. Classifier uses the hidden state at the variant-position token (not the CLS token) — pathogenicity is a local property.

Limitations

Training data is limited (~2,000 ClinVar pathogenic + ~5,000 gnomAD negatives). Variants in underrepresented functional regions (tRNA, rRNA) may have lower accuracy.