Zero-shot variant effect prediction across coding and non-coding regions of 10 spaceflight radiation-response genes using the Evo2 genomic foundation model (7B parameters). Pre-computed scores for 215,000+ variants are available on Hugging Face.
| Gene | Variants | Evo2 AUROC | AlphaMissense | CADD | REVEL |
|---|---|---|---|---|---|
| CHEK2 | 16,098 | 1.000 | -- | 0.999 | 0.832 |
| ATM | 52,791 | 0.995 | 0.926 | 0.999 | 0.937 |
| DNMT3A | 19,693 | 0.996 | 1.000 | 1.000 | 0.972 |
| TP53 | 14,240 | 0.9887 | 0.988 | 0.988 | 0.952 |
| BRCA1 | 36,901 | 0.988 | 0.960 | 0.995 | 0.748 |
| TERT | 17,443 | 0.909 | 0.969 | 0.999 | 0.996 |
| Mean | 0.980 | 0.969 | 0.997 | 0.906 |
Bold = highest AUROC among all tools for that gene. ClinVar P/LP vs B/LB (>=2-star review). 4 additional genes (CLOCK, NFE2L2, MSTN, RAD51) scored but lack sufficient ClinVar P/LP variants for AUROC.
Highlights:
- Evo2 achieves the highest AUROC of any tool for CHEK2 (0.9996) and TP53 (0.989)
- On ATM (52K variants), Evo2 outperforms AlphaMissense by +0.069 AUROC
- Radiation-type mutations (C>A via 8-oxoguanine) are more damaging than other mutations across all 10 genes (all p < 0.005; ATM p = 1.1e-26)
- Evo2 scores complement existing tools (Spearman rho 0.41--0.83 vs CADD), capturing distinct signal
- Non-coding validation: Spearman rho = -0.267 (p = 3.8e-14) against TERT promoter MPRA data
| Gene | Category | Spaceflight Relevance |
|---|---|---|
| BRCA1 | DMS control | DNA repair; risk allele in Rutter et al. 2024 |
| TP53 | DMS control | Most mutated gene in astronaut clonal hematopoiesis (7/14 astronauts) |
| CHEK2 | DMS control | DNA damage checkpoint kinase |
| DNMT3A | DMS control | 2nd most mutated in astronaut CH; found in Inspiration4 crew |
| TERT | Novel target | Telomere biology; 14.5% elongation in NASA Twins Study |
| ATM | Novel target | DSB sensor for galactic cosmic ray (GCR) radiation damage |
| NFE2L2 | Novel target | Master antioxidant TF; protective in 4 ISS mouse studies |
| CLOCK | Novel target | Circadian rhythm; 16 sunrises/day disruption on ISS |
| MSTN | Novel target | Muscle wasting protection in microgravity |
| RAD51 | Novel target | Core HR recombinase for strand invasion at DSBs |
- DMS Calibration -- Spearman correlation with deep mutational scanning fitness scores (4 control genes); Pejaver 2022 likelihood ratio threshold calibration
- ClinVar Validation -- Independent AUROC on ClinVar P/LP vs B/LB (>=2-star), stratified by review status, benchmarked against 5 tools
- Non-Coding Validation -- TERT promoter MPRA (Kircher 2019), ENCODE cCRE enrichment analysis
All variant scores are available on Hugging Face:
from datasets import load_dataset
ds = load_dataset("jang1563/evo2-spaceflight-vep")
# Filter to a specific gene
atm = ds.filter(lambda x: x["gene"] == "ATM")
# Get pathogenic variants
pathogenic = ds.filter(lambda x: x["clinvar_class"] == "P/LP")Each variant includes: chrom, pos, ref, alt, gene, region_type, delta (Evo2 score), clinvar_class, clinvar_stars.
scripts/
00_window_ablation.py # Window size ablation (4K--32K bp)
01_prepare_variants.py # Generate variant sequences (SNVs + indels)
02_score_variants.py # Evo2 scoring engine with checkpointing
03_calibrate_dms.py # DMS calibration + Pejaver LR thresholds
04_validate_clinvar.py # ClinVar validation + temporal split
05_score_noncoding.py # TERT MPRA + ENCODE enrichment
06_score_indels.py # Indel scoring (in-frame vs frameshift)
07_benchmark_tools.py # Multi-tool comparison (AM, CADD, REVEL, etc.)
08_radiation_signatures.py # Radiation mutation functional impact
09_cross_species.py # Mouse ortholog scoring (GRCm39)
10_entropy_landscape.py # Positional entropy landscape
11_astronaut_variants.py # Published astronaut mutation scoring
12_make_figures.py # Figure generation
utils/ # Config, gene coordinates, parsers, benchmarking
slurm/ # SLURM job templates for HPC
data/
download_data.sh # Automated data download script
figures/ # Pre-rendered figures
results/ # Example result files
- Python >= 3.11
- CUDA-capable GPU with >= 48 GB VRAM (A40 or A100)
- PyTorch >= 2.5 with CUDA 12.1+
# Create environment
conda create -n evo2 python=3.11 -y
conda activate evo2
# Install Evo2
pip install evo2 torch flash-attn
# Install analysis dependencies
pip install -r requirements.txt# Set your project root
export EVO2_ROOT=/path/to/your/project
# 1. Download reference data and benchmarks
bash data/download_data.sh
# 2. Download DMS data from MaveDB
python scripts/download_mavedb.py
# 3. Prepare variants for a gene
python scripts/01_prepare_variants.py --gene BRCA1
# 4. Score variants (requires GPU)
python scripts/02_score_variants.py --gene BRCA1 --window-size 8192
# 5. Validate against ClinVar
python scripts/04_validate_clinvar.py --gene BRCA1
# 6. Run full benchmark
python scripts/07_benchmark_tools.py --gene all
# 7. Generate figures
python scripts/12_make_figures.py --fig allFor SLURM-based HPC environments, see scripts/slurm/template.sh.
score(S) = (1/N) * SUM log P(s_{t+1} | s_1, ..., s_t)
delta = score(S_alt) - score(S_ref)
- More negative delta = more damaging variant
- All scores use reverse-complement averaging for strand symmetry
- Window size: 8,192 bp (optimal via ablation; AUROC 0.992 across 3 genes x 4 sizes)
- Deterministic at bfloat16 precision
Click to expand all figures
| Figure | Description |
|---|---|
| Fig 2 | Window size ablation: AUROC vs window size, score stability |
| Fig 3 | DMS calibration: Evo2 vs DMS fitness (4 control genes) |
| Fig 4 | ClinVar validation: multi-tool AUROC comparison (6 genes) |
| Fig 5 | Per-gene constraint landscapes (all 10 genes) |
| Fig 6 | Non-coding: TERT MPRA correlation, ENCODE cCRE enrichment |
| Fig 7 | Radiation mutation impact across 10 genes |
| Fig 8 | Astronaut variant scoring |
| S1 | Tool orthogonality matrix |
| S3 | Frameshift vs in-frame indel analysis |
| Dataset | Source | Build |
|---|---|---|
| Reference genome | GRCh38 no-alt analysis set | hg38 |
| ClinVar | NCBI ClinVar VCF | hg38 |
| BRCA1 DMS | Findlay 2018 (MaveDB 00000097-0-2) |
-- |
| TP53 DMS | Giacomelli 2018 (MaveDB 00000068-b-1) |
-- |
| CHEK2 DMS | McCarthy-Leo 2024 (MaveDB 00001203-a-1) |
-- |
| DNMT3A DMS | Garcia et al. 2025 | -- |
| TERT MPRA | Kircher 2019 (MaveDB 00000031-b-1) |
-- |
| AlphaMissense | Cheng et al. 2023 | hg38 |
| CADD | v1.7, Rentzsch et al. 2021 | hg38 |
| REVEL | v1.3, Ioannidis et al. 2016 | hg38 |
| ENCODE cCREs | ENCODE Project | hg38 |
@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}
}Please also cite the Evo2 model:
@article{nguyen2026evo2,
title={Sequence modeling and design from molecular to genome scale with Evo 2},
author={Nguyen, Eric and others},
journal={Nature},
year={2026},
doi={10.1038/s41586-026-10176-5}
}- Evo2 -- Genomic foundation model (ARC Institute)
- Rutter et al. 2024 -- Protective alleles for spaceflight (Nature Communications)
- SOMA Atlas -- Space Omics and Medical Atlas (Nature)
- Brojakowska et al. 2022 -- Astronaut clonal hematopoiesis (Communications Biology)
MIT License. See LICENSE for details.