Skip to content

Enformer

Resource Location

/labs/SysMedBio/Q/Resources/Genomics/Enformer

/Volumes/lab/SysMedBio/Resources/Genomics/Enformer

Q:\SysMedBio\Resources\Genomics\Enformer

Enformer is a deep-learning architecture developed by DeepMind and Calico Life Sciences designed to predict gene expression and chromatin states directly from DNA sequence.

Scientific Context

A major challenge in regulatory genomics is that genetic variants affecting gene expression often reside tens or hundreds of kilobases away from the promoters of the genes they regulate (distal regulatory elements, like enhancers). Traditional convolutional neural network models have limited receptive fields, making it difficult to capture these long-range regulatory interactions.

Enformer addresses this limitation by: 1. Transformer-Based Architecture: Utilizing self-attention layers to integrate information across genomic sequences over a receptive field of up to 100 kb on either side of a target locus. 2. Epigenomic and Expression Predictions: Directly predicting multiple epigenomic and gene-expression features (such as CAGE transcription rates, DNase-seq chromatin accessibility, and ChIP-seq histone modifications) across hundreds of human and mouse cell lines and tissues. 3. Variant Annotation: Quantifying how non-coding variants alter transcription levels and chromatin states, allowing researchers to predict the functional consequences of mutations without running the resource-heavy model locally.

The precomputed variant impact files in this directory are optimized for integration with the Ensembl Variant Effect Predictor (VEP) via the Enformer VEP plugin (Enformer.pm). The datasets provide two primary variant metrics: - SAD (SNP Activity Difference): The raw difference in Enformer prediction between the alternate allele and reference allele. - SAR (SNP Activity Ratio): The log2-transformed fold-change difference between the alternate and reference allele predictions.

Main Publication

  • Title: Effective gene expression prediction from sequence by integrating long-range interactions
  • Authors: Žiga Avsec, Vikram Agarwal, Daniel Visentin, Joseph R. Ledsam, Agnieszka Grabska-Barwinska, Kyle R. Taylor, Yannis Assael, John Jumper, Pushmeet Kohli, and David R. Kelley
  • Journal: Nature Methods (2021, Vol. 18, Pages 1196–1203)
  • DOI: 10.1038/s41592-021-01252-x

Available Files

The following files containing pre-computed Enformer regulatory impact scores for common genetic variants are available:

  • enformer_grch37.vcf.gz: Pre-computed Enformer scores mapped to the GRCh37/hg19 human assembly.
  • enformer_grch37.vcf.gz.tbi: Tabix index file enabling fast coordinate queries for the GRCh37 dataset.
  • enformer_grch38.vcf.gz: Pre-computed Enformer scores mapped to the GRCh38/hg38 human assembly.
  • enformer_grch38.vcf.gz.tbi: Tabix index file enabling fast coordinate queries for the GRCh38 dataset.