songlab/hg38-variant-consequences

VEP consequences with CRE annotations for all possible SNVs in the human genome hg38

Pre-computed VEP consequences for all possible single nucleotide variants (SNVs) in the human genome (GRCh38/hg38), with cis-regulatory element (CRE) annotations from ENCODE SCREEN.

Code: https://github.com/gonzalobenegas/hg38-variant-consequences

Dataset Details

• VEP container: docker://ensemblorg/ensembl-vep:release_109.1
• VEP flags: --most_severe --distance 1000
• CRE source: ENCODE SCREEN

CRE Annotation

For non-exonic variants (intergenic, intronic, upstream/downstream gene variants), the consequence_cre column contains the CRE class based on overlap with cis-regulatory elements:

CRE Class	Description
PLS	Promoter-like signature
pELS	Proximal enhancer-like signature
dELS	Distal enhancer-like signature
CA-H3K4me3	Chromatin accessible + H3K4me3
CA-CTCF	Chromatin accessible + CTCF
CA-TF	Chromatin accessible + TF
CA	Chromatin accessible only
TF	TF binding only

Variants within 500bp of a CRE (but not overlapping the core) are annotated with the _flank suffix (e.g., PLS_flank).

Priority order: core annotations override flank annotations; within each category, earlier classes in the table take precedence.

Schema

Column	Type	Description
chrom	String	Chromosome (1-22, X, Y)
pos	UInt32	Position (1-based)
ref	String	Reference allele
alt	String	Alternate allele
consequence	String	Most severe VEP consequence (original)
consequence_cre	String	Consequence with CRE class for non-exonic variants

For exonic variants, consequence and consequence_cre are identical. For non-exonic variants overlapping CREs, consequence_cre contains the CRE class while consequence retains the original VEP annotation.

Files are sorted by (chrom, pos, ref, alt).

Download

Using Hugging Face CLI

pip install -U huggingface_hub
hf download songlab/hg38-variant-consequences --repo-type dataset

This downloads to ./hg38-variant-consequences/.

Using wget

wget https://huggingface.co/datasets/songlab/hg38-variant-consequences/resolve/main/{1..22}.parquet
wget https://huggingface.co/datasets/songlab/hg38-variant-consequences/resolve/main/{X,Y}.parquet

Usage with Polars

The recommended approach for joining variants with consequences is a per-chromosome loop with streaming. This avoids memory issues when working with the large annotation files.

import polars as pl

VARIANTS_PATH = "variants.parquet"
CONSEQUENCES_DIR = "hg38-variant-consequences"
CHROMS = [str(i) for i in range(1, 23)] + ["X", "Y"]

Example 1: Small variants file (load into memory)

variants = pl.read_parquet(VARIANTS_PATH)

results = []
for chrom in CHROMS:
    chrom_variants = variants.filter(pl.col("chrom") == chrom)
    consequences_lf = pl.scan_parquet(f"{CONSEQUENCES_DIR}/{chrom}.parquet")

    joined = chrom_variants.lazy().join(
        consequences_lf,
        on=["chrom", "pos", "ref", "alt"],
        how="left",
        maintain_order="left",
    ).collect(engine="streaming")

    results.append(joined)

final = pl.concat(results)

Example 2: Large variants file (stream from disk)

For variants files with millions of rows, keep both sides lazy:

results = []
for chrom in CHROMS:
    consequences_lf = pl.scan_parquet(f"{CONSEQUENCES_DIR}/{chrom}.parquet")

    joined = pl.scan_parquet(VARIANTS_PATH).filter(
        pl.col("chrom") == chrom
    ).join(
        consequences_lf,
        on=["chrom", "pos", "ref", "alt"],
        how="left",
        maintain_order="left",
    ).collect(engine="streaming")

    results.append(joined)

final = pl.concat(results)