Evo2 DNA foundation model + population-scale WGS → variant & haplotype scores that illuminate Alzheimer’s genetics across ADNI, HPRC & UK Biobank.
DNA foundation models (e.g. Evo2-7B) learn regulatory grammar directly from sequence.
Here we score every variant and full-length haplotype in the APOE region (chr19:44.84–44.92 Mb, GRCh38) and show:
| Figure | Highlight | Dataset |
|---|---|---|
| 2 | Δ-scores track AD GWAS peaks | IGAP / ADNI GWAS |
| 3 | Haplotype diversity & ancestry effects | Human Pangenome |
| 4 | Scores predict amyloid & cognition | ADNI WGS + PET |
| 5 | MRI, proteomic & PheWAS landscape | UK Biobank |
The full manuscript title is “Advancing Human Population Genomics with DNA Foundation Models: Insights from the APOE Locus.”
# 1️⃣ Clone + install R deps (renv for reproducibility)
git clone https://github.com/huthvincent/Advancing-Human-Population-Genomics-with-DNA-Foundation-Models.git
cd Advancing-Human-Population-Genomics-with-DNA-Foundation-Models
# 2️⃣ Point to data & Evo2 model weight directory
export APOE_DATA_DIR=/path/to/inputs
export EVO2_MODEL_DIR=/path/to/evo2-7b
# 3️⃣ Re-run Figure 2 (as example)
Rscript Figure2/0.EVO2_scoring_GWAS.R
Rscript Figure2/1.Fig2A.RTip: Each folder’s 0* script builds inputs (variant ↦ FASTA ↦ Δ-score). Subsequent scripts generate statistics & plots.
⸻
Path Role Figure2/ Variant-level scoring & AD GWAS overlay Figure3/ Pangenome haplotypes • ancestry correlations Figure4/ ADNI haplotypes → amyloid & cognition Figure5/ UKB scale associations (MRI, proteomics, PheWAS) utils/(optional) helper scripts for Evo2 API, plotting themes
⸻
| Figure | Folder | Main Outputs |
|---|---|---|
| 2 | Figure2/ |
Manhattan & scatter plots comparing Evo2 Δ-scores with AD GWAS signals |
| 3 | Figure3/ |
Diversity & ancestry panels of APOE haplotypes from the Human Pangenome |
| 4 | Figure4/ |
ADNI diagnosis, cognition and amyloid-PET associations with haplotype scores |
| 5 | Figure5/ |
MRI maps, proteomic volcano + enrichment, and a PheWAS Manhattan plot |
Run scripts in each folder sequentially (e.g.
0*→1*…) after configuring paths to Evo2 weights and datasets.
| Resource | What it’s for | Access / Link |
|---|---|---|
| Evo2-7B weights | Sequence Δ-score generation | https://github.com/ArcInstitute/evo2/tree/main/evo2 |
| GRCh38 ref (Chr 19 slice) | Base sequence for variant/haplotype editing | GRCh38 FASTA |
| ADNI WGS + phenotypes | Individual haplotypes & PET / cognition (Fig 4) | Apply via https://adni.loni.usc.edu |
| Human Pangenome assemblies (HPRC) | Diverse haplotypes (Fig 3) | Public: https://humanpangenome.org |
| UK Biobank genotypes, MRI, SWI, Olink, Phecodes | Large-scale associations (Fig 5) | UKB Research Platform |
| LDlinkR API (optional) | LD calculations for plots | https://ldlink.nci.nih.gov |
| R packages | Stats & plots (dplyr, ggplot2, robustbase, gprofiler2, cerebroViz, …) |
Auto-installed via renv::restore() |
Note Some datasets (ADNI, UKB) require prior access approval.
Ensure credentials and local paths are configured before running scripts.
⸻
MIT for code, CC-BY 4.0 for docs.
⸻
Rui Zhu rui.zhu.rz399@yale.edu
Xiaopu Zhou xiaopu.zhou@sickkids.ca