lagosproject/ALPHAGenome-Embeddings

ALPHAGenome hg38 Embeddings

Pre-computed ALPHAGenome DNA foundation model embeddings for the entire human genome (hg38 / GRCh38).

The human genome is divided into ~22,000 non-overlapping 131 KB bins. Each bin's DNA sequence is embedded into a 3,072-dimensional latent space using the ALPHAGenome foundation model.

Companion project

These embeddings power the ALPHAGenome UMAP Explorer — an interactive browser visualization of latent relationships between genomic regions:

→ [GitHub: DNAEmbeddings](https://github.com/lagosproject/DNAEmbeddings) → [Live Demo](https://lagosproject.github.io/DNAEmbeddings/)

Scripts to download this dataset and re-run the UMAP pipeline are included in that repo.

Dataset structure

data/
  chr1_embeddings.npy     # float32 array, shape (N, 3072)
  chr1_metadata.csv       # columns: chrom, start, end
  chr2_embeddings.npy
  chr2_metadata.csv
  ...                     # one pair per chromosome (chr1–22, chrX, chrY)

Each .npy file and its paired .csv share the same row order — row i in the embeddings corresponds to row i in the metadata.

Quick usage

from huggingface_hub import hf_hub_download
import numpy as np
import pandas as pd

# Download a single chromosome
emb_path  = hf_hub_download("lagosproject/ALPHAGenome-Embeddings", "data/chr1_embeddings.npy", repo_type="dataset")
meta_path = hf_hub_download("lagosproject/ALPHAGenome-Embeddings", "data/chr1_metadata.csv",    repo_type="dataset")

embeddings = np.load(emb_path)          # (1750, 3072)
metadata   = pd.read_csv(meta_path)     # chrom | start | end

Or download everything at once (≈312 MB):

from huggingface_hub import snapshot_download
snapshot_download("lagosproject/ALPHAGenome-Embeddings", repo_type="dataset", local_dir="res/")

Statistics

Data sources

License

CC BY 4.0 — free to use for research and educational purposes with attribution.