Research map / Label-efficient and robust learning / Recognition and long-tail
GV-Rep: A Large-Scale Dataset for Genetic Variant Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GV-Rep, a ~7.5M-record genetic variant dataset aggregated from seven public databases plus a clinician-verified set, and benchmarks DNA foundation models (HyenaDNA, DNABERT2, NT, NT-V2) on pathogenicity, sQTL, gene-KO and indexing tasks.
Problem
Genetic variant interpretation lacks standardized, large-scale, context-rich datasets for deep learning, and existing benchmarks are binary, short-context and small, so it is unclear how well genomic foundation models represent variants.
Approach
Unifies heterogeneous GV effect databases (ClinVar, GTEx QTLs, Cell Passport, Score CRISPR screens, MAVEdb, GWAS, OLIDA) into a standardized (reference, alternative, annotation, label) format with variable-length sequence contexts, and evaluates frozen GFMs with a CNN head for classification, regression and embedding-based variant indexing.
Key results
- On ClinVar pathogenicity, DNABERT2 reached the highest AUROC at 73.87% ± 0.21, with Hyena, NT and NT-V2 at 65–69%.
- On sQTL classification, all models were near random (50.5%), with NT-V2 best at 54.10% AUROC.
- Finetuning raised indexing accuracy for Hyena from 0.428 to 0.662 and for DNABERT2 from 0.446 to 0.616; NT without finetuning was best at 0.558 among unfinetuned models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune Hyena with 1.3 million ClinVar GVs | HyenaDNA · Clinician-verified GV indexing | Variants Indexing Accuracy: 0.428 ± 0.09 -> 0.662 ± 0.16 | Helped |
| Fine-tune DNABERT2 with 1.3 million ClinVar GVs | DNABERT2 · Clinician-verified GV indexing | Variants Indexing Accuracy: 0.446 ± 0.06 -> 0.616 ± 0.06 | Helped |
| Fine-tune Nucleotide Transformer with 1.3 million ClinVar GVs | Nucleotide Transformer · Clinician-verified GV indexing | Variants Indexing Accuracy: 0.558 ± 0.10 -> 0.620 ± 0.10 | Helped |
| Fine-tune Nucleotide Transformer v2 with 1.3 million ClinVar GVs | Nucleotide Transformer v2 · Clinician-verified GV indexing | Variants Indexing Accuracy: 0.478 ± 0.06 -> 0.542 ± 0.05 | Helped |
| Use shorter sequence context | Fine-tuned genomic foundation models · ClinVar lung disease classification | AUC-ROC | Hurt |
| Use multi-species pretrained models instead of human-only models | DNABERT2 and Nucleotide Transformer v2 versus Hyena and NT-Human · ClinVar, sQTL, and Gene-KO tasks | Task performance | Helped |
| Use genomic foundation models for sQTL significance classification | Hyena, DNABERT2, NT-Human, and NT-V2 · sQTL significance classification | AUROC (%): 50.50± 1.07; 52.20 ± 2.62; 52.62 ±2.98; 51.33 ± 5.87; 54.10 ± 1.13 | No clear effect |
| Use genomic foundation models for Gene-KO fitness score prediction | Hyena, DNABERT2, NT-Human, and NT-V2 · Gene-KO fitness score prediction | MSE: 1.11± 0.09; 1.06 ± 0.11 | No clear effect |
Surprising findings
- In the Gene-KO regression task, three models converged to the same MSE (1.06), suggesting the added head learned the target distribution rather than meaningful variant representations.
- Nucleotide Transformer v2 was highly sensitive to context length in the lung-disease scaling experiment, while DNABERT2 and Hyena were more robust.
Limitations
- Clinical use requires human review; dataset is not validated for direct clinical decision-making.
- Lacks sensitive demographic attributes (ethnicity, sex) so fairness across populations cannot be evaluated.
- Frozen GFMs with a simple three-layer CNN head; sQTL and Gene-KO results near random suggest limited representation quality, and the clinician set is small (155 variants).
Experimental setup
Datasets: ClinVar Cell Passport Project Score GTEx eQTLs GTEx sQTLs GWAS Catalog MAVEDB OLIDA Clinician-verified GVs (CVGV)
Benchmarks: GV-Rep ClinVar pathogenicity GV-Rep sQTL significance GV-Rep Gene-KO fitness score CVGV variant indexing
Metrics: AUROC MSE top-10 label match accuracy
Baselines: Random HyenaDNA DNABERT2 Nucleotide Transformer Nucleotide Transformer v2
Builds on: ACMG-AMP guidelines Nucleotide Transformer DNABERT2 HyenaDNA BEND FAISS
Closest papers by meaning
- Training Flexible Models of Genetic Variant Effects from Functional Annotations using Accelerated Linear Algebra (ICML 2025 · similarity 0.86)
- DeepVar: An End-to-End Deep Learning Approach for Genomic Variant Recognition in Biomedical Literature (AAAI 2020 · similarity 0.85)
- Genomic Foundationless Models: Pretraining Does Not Promise Performance (ICLR 2026 · similarity 0.84)
- GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation (ICLR 2025 · similarity 0.83)
- Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks (ICML 2025 · similarity 0.82)
- WISER: Weak Supervision and Supervised Representation Learning to Improve Drug Response Prediction in Cancer (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.