harari/human_grch38_segment_sample

Human GRCh38 Genome Segments

Dataset Description

This dataset contains 16,384 base pair segments from the human reference genome (GRCh38) prepared for Sparse Autoencoder (SAE) training with the Evo2 model. The segments are extracted using a sliding window approach with 75% overlap.

Dataset Details

• Total segments: 718,648
• Segment size: 16,384 base pairs
• Stride: 4,096 base pairs (75% overlap)
• Source genome: GRCh38.primary_assembly (GENCODE Release 41)
• Chromosomes: chr1-22, chrX, chrY, plus alternative/random contigs
• Filtering: Segments with >10% N-content removed

Data Structure

Each segment contains:

• chromosome: Chromosome identifier
• start: Start position (0-based)
• end: End position (exclusive)
• sequence: DNA sequence (16,384 bp)
• strand: Strand orientation (always "+")
• n_content: Fraction of N bases (ambiguous nucleotides)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("harari/human_grch38_segment")

# Access a segment
segment = dataset['train'][0]
print(f"Chromosome: {segment['chromosome']}")
print(f"Position: {segment['start']}-{segment['end']}")
print(f"Sequence length: {len(segment['sequence'])}")

Intended Use

This dataset is designed for:

• Training Sparse Autoencoders on genomic sequences
• Analyzing DNA sequence patterns with foundation models
• Genomic representation learning
• Evo2 model feature extraction and interpretation

Citation

If you use this dataset, please cite the original GRCh38 reference genome and GENCODE annotations:

@article{grch38,
  title={The reference genome sequence of the human genome},
  journal={Nature},
  year={2013}
}

@article{gencode,
  title={GENCODE: the reference human genome annotation},
  journal={Genome Research},
  year={2012}
}

Data Source

• Reference genome: GENCODE GRCh38.p13 Release 41
• URL: https://www.gencodegenes.org/human/release_41.html
• Processing: Sliding window chunking with N-content filtering

Preprocessing

1. Downloaded GRCh38.primary_assembly.genome.fa.gz
2. Extracted 16,384 bp segments with 4,096 bp stride
3. Filtered segments with >10% N-content
4. Added genomic coordinates and metadata

License

This dataset is released under CC-BY-4.0 license, consistent with the GENCODE data usage terms.