PAVR

PAVR: Perturbation-Aware Variant Representations from Genomic Foundation Models

Variant effect prediction asks whether a small DNA change can perturb molecular function and alter phenotype. Genomic foundation models (GFMs) provide strong sequence representations, but effects span local allelic disruption, proximal regulatory response, and distal functional consequences. Collapsing these into one global embedding can wash out variant-specific signal and entangle distinct effect scales.

PAVR (Perturbation-Aware Variant Representation) repurposes frozen GFMs by comparing token-level hidden states of paired reference and alternative sequences. It builds multi-view ref–alt token fields and uses an explicit dual-branch design with soft complementary receptive fields:

• Allele branch — local allelic perturbation at the edited bases
• Context branch — context-mediated / regulatory response

The two scales stay disentangled before fusion, improving predictive performance and mechanistic interpretability over conventional global differential embeddings (validated on ClinVar, sQTL, and Gene-KO across six GFMs).

This repository is a self-contained demo of the paper-default PAVR (no ablation switches). Evo2 weights and the GFM encoding pipeline are not included; features are assumed cached offline.

---

Layout

PAVR/
├── script/
│   ├── train.py           # ClinVar 4-class trainer
│   ├── common/            # layers + task losses
│   └── pavr/
│       └── model.py       # PAVR adapter (paper default)
├── data/
│   ├── PAVR.png                     # method overview figure
│   ├── clinvar_pathogenicity.csv
│   ├── dataset_summary.json
│   └── evo2_7b_window_gastric_1k.pt   # ~2 GB frozen Evo2-7B cache
├── result/window_gastric_1k_Evo2_7B/  # 5-seed demo (AUROC ≈ 0.862)
├── environment.yml
├── environment.full.yml
├── requirements.txt
└── README.md

---

Demo data

Test data (including the cached feature tensors) is hosted on Hugging Face:
https://huggingface.co/datasets/xw97/PAVR_test_data

File	Content
data/clinvar_pathogenicity.csv	2000 gastric ClinVar variants, 1 kb windows
data/evo2_7b_window_gastric_1k.pt	Cached tensors: h_ref, h_alt, variant_mask, padding_mask

Splits: train 1400 / val 300 / test 300.

The .pt cache is ~2 GB and is gitignored by default. Download it from the Hugging Face dataset above, or use Git LFS / a release asset.

---

Reported result

result/window_gastric_1k_Evo2_7B/ (PAVR + frozen Evo2-7B, 5 seeds):

Metric	Mean ± std
Macro AUROC	0.862 ± 0.018
Macro F1	0.519 ± 0.044

---

Environment

conda env create -f environment.yml
conda activate pavr

environment.full.yml is a full dump of the authors’ Genomic env (optional).

---

Train

cd PAVR
export PYTHONPATH=script

python script/train.py \
  --features data/evo2_7b_window_gastric_1k.pt \
  --csv data/clinvar_pathogenicity.csv \
  --output-dir result/rerun_window_gastric_1k \
  --seeds 0 1 2 3 4 \
  --device cuda

PYTHONPATH=script python -c "from pavr import PAVRClassifier, PAVRConfig; print('ok')"

---

License & contact

No license file is included in this repository. Please confirm licensing with the project owner before reuse or redistribution.