Datasets:

GustavoHCruz
/

DNA_coding_regions

	---
	license: cc-by-4.0
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data/train-*
	dataset_info:
	features:
	- name: accession
	dtype: string
	- name: organism
	dtype: string
	- name: sequence
	dtype: string
	- name: introns
	list:
	- name: after
	dtype: string
	- name: before
	dtype: string
	- name: end
	dtype: int64
	- name: gene
	dtype: string
	- name: sequence
	dtype: string
	- name: start
	dtype: int64
	- name: exons
	list:
	- name: after
	dtype: string
	- name: before
	dtype: string
	- name: end
	dtype: int64
	- name: gene
	dtype: string
	- name: sequence
	dtype: string
	- name: start
	dtype: int64
	- name: proteins
	list:
	- name: end
	dtype: int64
	- name: gene
	dtype: string
	- name: sequence
	dtype: string
	- name: start
	dtype: int64
	splits:
	- name: train
	num_bytes: 11536678696
	num_examples: 1677609
	download_size: 5448417115
	dataset_size: 11536678696
	task_categories:
	- text-classification
	- token-classification
	- translation
	tags:
	- Exons
	- Introns
	- Proteins
	- DNA
	pretty_name: DNA Coding Regions
	size_categories:
	- 1M<n<10M
	---

	# DNA Coding Regions Dataset

	This is a curated collection of genomic sequences extracted directly from NCBI GenBank, designed to support research in introns and exons classification, DNA-to-protein translation, gene structure analysis, and biological sequence modeling with deep learning architectures.

	---

	## Source and Extraction Pipeline

	All records were extracted from GenBank using [Biopython](https://biopython.org/).
	The dataset construction followed a reproducible data processing pipeline written in Python, which:

	- Downloads and parses GenBank records.
	- Extracts genomic DNA sequences, their associated exons, introns, and coding sequences (CDS).
	- Processes the `strand` orientation to produce normalized sequences.
	- Removes duplicate entries based on `(sequence, organism)` pairs.
	- Assembles each record into a structured JSONL format suitable for machine learning models.

	The GenBank search query used for data collection was:

	```
	"genomic DNA"[Filter]
	AND ("exon"[Feature Key] OR "intron"[Feature Key])
	AND "CDS"[Feature Key]
	AND ("3"[SLEN] : "16384"[SLEN])
	````

	You can find more information about the pipeline in the GitHub from the [DNA Coding Regions](https://github.com/GustavoHCruz/CodingDNATransformers) repository.

	---

	## Dataset Structure

	Each entry in the dataset corresponds to a unique DNA sequence, identified by its GenBank accession.
	The dataset is serialized in JSON Lines (`.jsonl`) format and can be loaded with the Hugging Face `datasets` library.

	### Example record

	```json
	{
	"accession": "NC_045512.2",
	"organism": "Homo sapiens",
	"sequence": "ATTAAAGGTTTATACCTTCCCAGGTAACAAACCAACCAACTTTCGAT...",
	"introns": [
	{
	"sequence": "TTGTAGACCAGTGCAGTA...",
	"start": 1450,
	"end": 1783,
	"gene": "ORF1ab",
	"before": "ATGCCDG",
	"after": "TAACAFG"
	}
	],
	"exons": [
	{
	"sequence": "ATGGACACAAGTCAGG...",
	"start": 1,
	"end": 1449,
	"gene": "ORF1ab",
	"before": null,
	"after": "GT"
	}
	],
	"proteins": [
	{
	"sequence": "MESLVPGFNEKTHVQLSLPVLQVRDVLVRGFGDSVEEVL...",
	"start": 1,
	"end": 4405,
	"gene": "ORF1ab"
	}
	]
	}
	````

	---

	## Field Descriptions

	\| Field \| Type \| Description \|
	\| ------------- \| ------ \| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- \|
	\| accession \| `str` \| GenBank accession number for the DNA sequence. \|
	\| organism \| `str` \| Name of the organism from which the sequence was derived. \|
	\| sequence \| `str` \| The full genomic DNA sequence (processed strand). \|
	\| introns \| `list` \| List of intronic regions associated with this DNA sequence. Each item contains: <ul><li>`sequence`: only the nucleotide sequence of the intron</li><li>`start`, `end`: coordinates relative to the DNA sequence</li><li>`gene`: gene name when annotated</li><li>`before`, `after`: short flanking sequences</li></ul> \|
	\| exons \| `list` \| List of exonic regions associated with this DNA sequence. Same structure as `introns`. \|
	\| proteins \| `list` \| List of coding sequences (CDS) translated to amino acid sequences, with: <ul><li>`sequence`: protein sequence</li><li>`start`, `end`: coordinates in the DNA sequence</li><li>`gene`: gene name</li></ul> \|

	---

	## Applications

	This dataset can be directly used for:

	* DNA to protein translation modeling
	* Exon and Introns classification
	* Splicing prediction
	* Genomic representation learning
	* Bioinformatics-focused LLM pretraining (DAPT)

	---

	## Loading Example

	```python
	from datasets import load_dataset

	dataset = load_dataset("gu-dudi/DNA_coding_regions")

	print(dataset)
	print(dataset["train"][0])
	```

	---

	## Dataset Metadata

	* Source: NCBI GenBank
	* Processed with: Biopython, Pandas, tqdm
	* Maintainer: [Gustavo Henrique Ferreira Cruz](https://huggingface.co/GustavoHCruz)
	* License: Open for research and educational use
	* Format: JSON Lines (UTF-8)

	---

	## Disclaimer on data completeness

	Not all genomic entries in this dataset contain every type of annotation (exons, introns, and proteins).

	While the GenBank records were filtered to include sequences annotated with "exon", "intron", and "CDS" feature keys, the underlying annotations in GenBank are not always deterministic or complete.

	Some sequences may include only exons or introns without corresponding protein-coding regions, or vice versa.

	This reflects the inherent variability and curation differences across submissions in the GenBank database.

	---


	### Citation

	If you use this dataset in your research, please cite:

	```bibtex
	@misc{gustavo_henrique_ferreira_cruz_2025,
	author = {Gustavo Henrique Ferreira Cruz},
	title = {DNA\_coding\_regions (Revision 16f4e3a)},
	year = 2025,
	url = {https://huggingface.co/datasets/GustavoHCruz/DNA_coding_regions},
	doi = {10.57967/hf/7238},
	publisher = {Hugging Face}
	}
	```
	---

	### Version and Integrity

	* Version: 1.0
	* Total entries: 1,677,609
	* Deduplication: duplicates removed based on `(sequence, organism)` pair
	* Strand normalization: handled during extraction

	---

	### Notes

	* Coordinates (`start`, `end`) are relative to the parent DNA sequence.
	* `sequence` inside each intron/exon corresponds only to that region (not the full DNA).
	* Protein sequences are already translated amino acid chains, not nucleotide fragments.

	---

	Developed as part of the Master’s research project on DNA sequence understanding and translation using deep learning models.