enformer-data

Dataset Summary

This dataset contains the specific genomic intervals used for training, validating, and testing the Enformer model, a deep learning architecture for predicting functional genomic tracks from DNA sequence. The intervals are provided for both human and mouse genomes. As done in the publication, we modified the Basenji2 dataset by extending the input sequence to 196,608 bp from the original 131,072 bp using the hg38 reference genome.

• Source Publication: Avsec, Ž., et al. "Effective gene expression prediction from sequence by integrating long-range interactions." Nat Methods 18, 1196–1203 (2021).
• Genome Builds:
  • Human: hg38
  • Mouse: mm10

Repository Content

The repository includes two tab-separated values (TSV) files and two Jupyter notebooks:

1. human_intervals.tsv: 38,171 genomic regions (excluding header).
2. mouse_intervals.tsv: 33,521 genomic regions (excluding header).
3. data_human.ipynb: Code to create human_intervals.tsv.
4. data_mouse.ipynb: Code to create mouse_intervals.tsv.

Dataset Structure

Data Fields

Both files follow a standard genomic interval format:

Column	Type	Description
chrom	string	Chromosome identifier (e.g., chr18, chr4)
start	int	Start coordinate of the interval
end	int	End coordinate of the interval
split	string	Data partition assignment (train, test, or val)

Statistics

File	Number of Regions	Genome Build
human_intervals.tsv	38,171	hg38
mouse_intervals.tsv	33,521	mm10

Data Splits

Human (hg38)

Split	Count
train	34,021
valid	2,213
test	1,937

Mouse (mm10)

Split	Count
train	29,295
valid	2,209
test	2,017

Usage

from huggingface_hub import hf_hub_download
import pandas as pd

file_path = hf_hub_download(
    repo_id="Genentech/enformer-data",
    filename="human_intervals.tsv",
    repo_type="dataset"
)
df_human = pd.read_csv(file_path, sep='\t')

file_path = hf_hub_download(
    repo_id="Genentech/enformer-data",
    filename="mouse_intervals.tsv",
    repo_type="dataset"
)
df_mouse = pd.read_csv(file_path, sep='\t')