Argus-NET Dataset

This dataset contains the processed genomic compressed files used for training, validation and evaluation of Argus-NET, a hybrid CNN-Transformer deep learning framework for joint prediction of developmental (DEV; H3K27ac) and housekeeping (HK; H3K4me3) enhancer activity from multimodal genomic inputs.

Dataset composition:

train.tfrecord — Training set: chromosomes chr1, chr2, chr3, chr4, chr5, chr6, chr7, chr8, chr12, chr14, chr15, chr16, chr18, chr19, chr20, chrY (n = 777,869 examples)

val.tfrecord — Validation set: chromosomes chr11, chr22, chrX (n = 101,143 examples)

test.tfrecord — Held-out test set: chromosomes chr9, chr10, chr13, chr17, chr21 (n = 184,866 examples)

Input representation: Each example is a serialized format containing a 7-channel input tensor (1,000 bp × 7 channels) and two scalar regression targets:

Channel
Content

0–3
One-hot encoded DNA sequence (A, C, G, T)

4
ATAC-seq chromatin accessibility signal (quantile-normalized)

5
EP300 ChIP-seq co-activator binding signal

6
Functional regulatory element category (integer: PLS=1, pELS=2, dELS=3, CTCF-bound=4, DNase-H3K4me3=5)

Target
Content

dev
log₂(H3K27ac signal + 1)  developmental enhancer activity

hk
log₂(H3K4me3 signal + 1)  housekeeping enhancer activity

Genome assembly: GRCh38/hg38

Chromosome splitting strategy: Chromosomes were partitioned using a strict held-out scheme to prevent data leakage from spatial autocorrelation between neighboring genomic loci. No chromosome appears in more than one split. Chromosome indices were assigned by lexicographic sorting of chromosome names.

File format: compressed serialization format. Each record can be parsed using the following feature description:

python

feature_description = {
'x' : tf.io.FixedLenFeature([1000 * 7], tf.float32),
'dev': tf.io.FixedLenFeature([], tf.float32),
'hk' : tf.io.FixedLenFeature([], tf.float32),
'chr': tf.io.FixedLenFeature([], tf.int64),
}

Source data: Genomic sequences were extracted from the GRCh38 human reference genome. Regulatory element annotations, chromatin accessibility (ATAC-seq), histone modification (H3K27ac, H3K4me3), and co-activator binding (EP300 ChIP-seq) signals were derived from publicly available genome-wide chromatin profiling experiments across human cell types and tissues. Element functional categories were derived from an integrative registry of candidate cis-regulatory elements based on DNase-seq and ChIP-seq data.