jang1563/evo2-spaceflight-vep

Evo2 Zero-Shot VEP Scores for Spaceflight Radiation-Response Genes

Pre-computed zero-shot variant effect prediction scores from the Evo2 genomic foundation model (7B parameters) for 215,001 variants across 10 spaceflight radiation-response genes.

Code: github.com/jang1563/evo2-spaceflight-vep

Dataset Description

Each row is a single variant (SNV or indel) scored by Evo2 using an 8,192 bp context window with reverse-complement averaging.

Columns

Column	Type	Description
variant_key	string	chrom:pos:ref>alt identifier
gene	string	Gene symbol (BRCA1, TP53, CHEK2, DNMT3A, TERT, ATM, NFE2L2, CLOCK, MSTN, RAD51)
chrom	string	Chromosome (GRCh38)
pos	int	0-based genomic position
ref	string	Reference allele
alt	string	Alternate allele
region_type	string	Functional region: coding, intronic, promoter, utr5, utr3, intergenic, or radiation_del_mh1--radiation_del_mh4 (microhomology-flanked deletion regions)
delta	float	Evo2 delta score: score(alt) - score(ref). More negative = more damaging.
ref_score	float	Evo2 mean log-likelihood for reference sequence
alt_score	float	Evo2 mean log-likelihood for alternate sequence
clinvar_class	string	ClinVar classification (P/LP, B/LB, VUS, Conflicting_classifications_of_pathogenicity, etc.) or empty
clinvar_stars	int	ClinVar review star level (0--4) or empty

Gene Summary

Gene	Variants	ClinVar AUROC	Key Role
ATM	52,791	0.995	DSB sensor for radiation damage
BRCA1	36,901	0.988	DNA double-strand break repair
DNMT3A	19,693	0.996	DNA methyltransferase; astronaut CH
CLOCK	18,369	--	Circadian rhythm disruption in space
TERT	17,443	0.909	Telomere biology; NASA Twins Study
CHEK2	16,098	1.000	DNA damage checkpoint kinase
NFE2L2	15,230	--	Antioxidant response to radiation
TP53	14,240	0.989	Tumor suppressor; astronaut CH
MSTN	12,383	--	Muscle wasting in microgravity
RAD51	11,853	--	Homologous recombination repair

Usage

from datasets import load_dataset
import pandas as pd

# Load dataset
ds = load_dataset("jang1563/evo2-spaceflight-vep", split="train")
df = ds.to_pandas()

# Filter to a specific gene
atm = df[df["gene"] == "ATM"]

# Most damaging variants (most negative delta)
damaging = df.nsmallest(100, "delta")

# ClinVar pathogenic variants
pathogenic = df[df["clinvar_class"] == "P/LP"]

# Radiation-type mutations (C>A / G>T, characteristic of 8-oxoguanine)
radiation = df[
    ((df["ref"] == "C") & (df["alt"] == "A")) |
    ((df["ref"] == "G") & (df["alt"] == "T"))
]

Scoring Method

score(S) = (1/N) * SUM log P(s_{t+1} | s_1, ..., s_t)
delta    = score(S_alt) - score(S_ref)

• Model: Evo2 7B (evo2_7b, bfloat16)
• Window: 8,192 bp (optimal via ablation study)
• Reverse-complement averaging enabled
• Reference genome: GRCh38

Validation

• Mean AUROC 0.980 across 6 ClinVar-validatable genes (vs CADD 0.997, AlphaMissense 0.969)
• Evo2 achieves highest AUROC for CHEK2 (0.9996) and TP53 (0.989)
• DMS calibration: Spearman |rho| = 0.23--0.52 across 4 control genes (TP53 rho is negative due to LOF assay directionality)
• Non-coding: rho = -0.267 (p = 3.8e-14) vs TERT MPRA
• Radiation-type mutations more damaging across all 10 genes (all p < 0.005)

Intended Use

These scores provide computational supporting evidence (PP3/BP4 level under ACMG/AMP guidelines) for variant prioritization. They are NOT standalone diagnostic classifications. Use in conjunction with clinical evaluation, functional data, and other tools.

Limitations

• Scores reflect evolutionary sequence constraint, not direct functional measurement
• Non-coding predictions are exploratory for genes without MPRA validation
• 7B model only (40B requires >48 GB VRAM)
• ClinVar has known European ancestry overrepresentation

Citation

@software{kim2026evo2vep,
  title={Evo2 Zero-Shot Variant Effect Prediction for Spaceflight Genes},
  author={Kim, JangKeun and Mason, Christopher E.},
  year={2026},
  url={https://github.com/jang1563/evo2-spaceflight-vep}
}

References

• Nguyen et al. (2026). Sequence modeling and design from molecular to genome scale with Evo 2. Nature. doi:10.1038/s41586-026-10176-5
• Rutter et al. (2024). Protective alleles for spaceflight. Nature Communications 15:6158. doi:10.1038/s41467-024-50532-5

Acknowledgments

• Evo2 by ARC Institute
• MaveDB for DMS data
• Mason Lab, Weill Cornell Medicine
• Computing resources: Cayuga HPC, Cornell University