Integration of Deep Learning Annotations with Functional Genomics Improves Identification of Causal Alzheimer's Disease Variants

Code and data for reproducing figures from:

Lakhani et al. Integration of Deep Learning Annotations with Functional Genomics Improves Identification of Causal Alzheimer's Disease Variants https://www.medrxiv.org/content/10.1101/2025.03.07.25323578v1

Data

All public data files are available on Zenodo:

Download all files

Using zenodo_get:

pip install zenodo_get
mkdir adfinemapping_public && cd adfinemapping_public
zenodo_get 10.5281/zenodo.19226023

Or using the Zenodo API directly:

pip install requests
python -c "
import requests, os
r = requests.get('https://zenodo.org/api/records/19226023')
for f in r.json()['files']:
    print(f'Downloading {f[\"key\"]}...')
    resp = requests.get(f['links']['self'], stream=True)
    with open(f['key'], 'wb') as out:
        for chunk in resp.iter_content(8192):
            out.write(chunk)
"

The dataset is ~4.5 GB across 32 files. All files are downloaded into a single flat directory. See zenodo_description.txt for a complete file-by-file description with column definitions.

Reproducing Figures

The public notebook process_focal_analysis_public.Rmd reproduces all main and supplementary figures from the downloaded data.

Requirements

R (>= 4.0) with the following packages:

install.packages(c("tidyverse", "data.table", "arrow", "patchwork",
                    "ggsci", "ggrepel", "scales", "PRROC"))

# Bioconductor packages
if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install(c("GenomicRanges", "IRanges", "ComplexUpset"))

Running

1. Download the Zenodo data to a local directory
2. Open process_focal_analysis_public.Rmd
3. Set DATA_DIR to the path where you downloaded the files
4. Knit or run all chunks

Repository Structure

File	Description
process_focal_analysis_public.Rmd	Public notebook — reproduces all figures from Zenodo data

Data Overview

LDSC Annotation Results (4 files)

Stratified LD Score Regression heritability enrichment across seven annotation types (Baseline, Glass Epigenomics, Roadmap, DeepSea, Enformer Tensorflow, Enformer TF/Glass, ChromBPNet) and three GWAS (Kunkle 2019, Wightman 2021, Bellenguez 2022). Includes multivariate LDSC with jointly-estimated tau*.

Fine-Mapping Results (4 files)

SuSiE/PolyFun results for Bellenguez 2022 under four annotation models: SuSiE (flat prior), Baseline, Baseline+Omics, Baseline+Omics+DL. Per-variant PIPs, credible sets, and effect sizes.

PRS Summaries (4 files)

Polygenic risk score AUC, odds ratios, and prevalence by decile across five ancestries (EUR, AFR, AMR, EAS, SAS). Aggregate statistics only (no individual-level data).

Cell-Type Annotations (9 files)

ATAC-seq peaks (4 cell types), ABC enhancer-gene predictions (4 cell types), and PLAC-seq chromatin interactions. Brain cell types: microglia, oligodendrocyte, astrocyte, neuron.

caQTL Predictions (8 files)

ChromBPNet and Enformer predicted allelic effects on chromatin accessibility for fine-mapped microglia variants under different centering strategies. Includes motifBreakR TF motif disruption analysis.

Other (3 files)

Bellenguez GWAS summary statistics (hg38, parquet) and finemapped variant annotation tables (FAVOR).

Citation

If you use this data or code, please cite:

Lakhani et al. Integration of Deep Learning Annotations with Functional Genomics
Improves Identification of Causal Alzheimer's Disease Variants. medRxiv (2025).
https://doi.org/10.1101/2025.03.07.25323578