seedbreaker-lora-v2

SeedBreaker — miRNA Seed Disruption Classifier

Predicts whether a rare germline SNV disrupts a microRNA binding site in a gene 3′UTR.

SeedBreaker is a LoRA-fine-tuned ESM-2 classifier trained to detect loss-of-binding mutations at miRNA seed positions (2–8) within AGO2 CLIP-confirmed 3′UTR binding sites. It is designed for rare germline variant prioritisation and site-set burden testing in cancer epidemiology cohorts.

🔗 GitHub: Cancer-epi-unit/SeedBreaker

---

Validation

Known pathogenic 3′UTR variants (ClinVar) are 4.4× enriched for HIGH-tier SeedBreaker calls versus common population variants from gnomAD v4.1:

OR = 4.44 (95% CI 2.10–9.40), p = 6.55 × 10⁻⁶ (Fisher's exact test, n=44 ClinVar vs n=969 gnomAD)

---

Model details

---

Training data

• 745,492 labelled examples from in-silico saturation mutagenesis
• Source: 278,425 AGO2 CLIP-confirmed binding sites (ENCORI) across 192 miRNAs and 11,140 genes (hg38)
• Labels assigned by Watson-Crick complementarity disruption at seed positions 2–8
• Chromosome-level train/val/test splits (test = chr8 + chr18, val = chr4)
• 3:1 disruptive:preserving class balance

---

Performance

Held-out test set — chr8 + chr18 (38,919 examples):

---

Usage

Recommended — via SeedBreaker pipeline

git clone https://github.com/Cancer-epi-unit/SeedBreaker
cd SeedBreaker

python3 score.py \
  --vcf    variants.vcf.gz \
  --sites  Pre_data/ENCORI/mirna_sites_encori_annotated_sorted.bed.gz \
  --genome Pre_data/Genome/GRCh38.primary_assembly.genome.fa \
  --seeds  Pre_data/miRBase/mirna_seeds_hsa.json \
  --model  c3114203/seedbreaker-lora-v2 \
  --out    results/scored_variants.tsv

Direct PEFT loading

from transformers import AutoConfig, AutoModelForSequenceClassification, AutoTokenizer
from peft import PeftModel
import torch

base_id  = "facebook/esm2_t33_650M_UR50D"
lora_id  = "c3114203/seedbreaker-lora-v2"

tokenizer = AutoTokenizer.from_pretrained(base_id)

cfg = AutoConfig.from_pretrained(base_id)
cfg.num_labels = 2
model = AutoModelForSequenceClassification.from_pretrained(base_id, config=cfg)
model = PeftModel.from_pretrained(model, lora_id)
model.eval()

# ref and alt are 50 bp windows centred on the variant
sequence = ref_50bp + "NNN" + alt_50bp
inputs   = tokenizer(sequence, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    prob_disruptive = torch.softmax(model(**inputs).logits, dim=-1)[0, 1].item()

print(f"Disruption probability: {prob_disruptive:.4f}")
# >= 0.8 combined score → HIGH tier

Output tiers

Combined score = 0.4 × physics_score + 0.6 × model_score, where physics_score reflects Watson-Crick disruption at the seed position.

---

Tissue-specific scoring

GTEx v11 co-expression weights can be applied post-scoring to downweight variants in binding sites that are not active in the tissue of interest:

python3 apply_tissue_weights.py \
  --scores results/scored_variants.tsv \
  --tissue breast \
  --out    results/scored_variants_breast.tsv

In breast tissue, GNAS emerges as the top hit — five ClinVar pathogenic variants disrupting binding sites for miR-138-5p, miR-18a/b-5p, and miR-150-5p, all achieving tissue_score ≥ 0.92.

---

Limitations

• Scores only variants within ENCORI AGO2 CLIP-confirmed sites (278,425 sites). Variants outside these sites are not scored.
• Trained on in-silico labels derived from Watson-Crick chemistry — not experimentally validated at the functional level.
• Does not model gain-of-function (new site creation).

---

Citation

Preprint forthcoming on bioRxiv.

Please also cite:

• ENCORI: Zhou K et al. (2026) Nature Methods
• ESM-2: Lin Z et al. (2023) Science 379:1123–1130
• PEFT: Mangrulkar et al. (2022) https://github.com/huggingface/peft
• miRBase v22: Kozomara A et al. (2019) Nucleic Acids Research 47:D155

---

Developed at the Oxford Cancer Epidemiology Unit, University of Oxford.