songlab/multiz100way

Processed whole-genome alignment of 100 vertebrate species

For more information check out our paper and repository.

Source data:

• MSA was downloaded from http://hgdownload.soe.ucsc.edu/goldenPath/hg38/multiz100way/
• Human sequence was replaced with a newer reference: http://ftp.ensembl.org/pub/release-107/fasta/homo_sapiens/dna/Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.gz

Available MSAs:

• 89.zarr.zip contains human + 89 vertebrates (excluding 10 closest primates)
• 99.zarr.zip contains human + 99 vertebrates

Example usage:

from gpn.data import GenomeMSA

genome_msa = GenomeMSA(msa_path)
X = genome_msa.get_msa(chrom, start, end, strand="+", tokenize=False)

Coordinates:

• hg38 assembly
• chrom should be in ["1", "2", ..., "22", "X", "Y"]

Streaming (playing, few VEP queries)

• Faster setup (no need to download and unzip)
• Slower queries (depends on network connection)
• Multiple dataloader workers don't seem to work
• More CPU memory required to load: 10.41 GB
• Recommended if you just want to do a few queries, e.g. VEP for a couple thousand variants

• msa_path = "zip:///::https://huggingface.co/datasets/songlab/multiz100way/resolve/main/89.zarr.zip"
  

Local download (training, large-scale VEP)

• Requires downloading (34GB) and unzipping (currently quite slow, will try to improve)

  wget https://huggingface.co/datasets/songlab/multiz100way/resolve/main/89.zarr.zip
  7z x 89.zarr.zip -o89.zarr  # can still take 5 hours with 32 cores, will try to streamline this in the future
  

• Update: faster unzipping here, courtesy of lpigou
• Much faster to query
• Can have multiple dataloader workers
• Virtually no CPU memory required to load
• Recommended for training or VEP for millions of variants

• msa_path = "89.zarr"