HantaBERT/Orthohantavirus-Genome-Atlas

HantaBERT Data Pipeline

This repository is responsible for the entire process of collecting, cleaning, and standardizing Orthohantavirus genomic data for the HantaBERT project. The pipeline automates data extraction from NCBI GenBank to produce a ready-to-use dataset for machine learning.

Key Features

• Extraction Automation: Uses Biopython to fetch thousands of RNA sequences (S, M, L) and related metadata in batches from the NCBI database.
• Multi-task Labeling: Standardizes host, species, and geography labels from unstructured raw data.
• Geocoding: Integrates with the Nominatim API to convert country names into geographical coordinates.
• Quality Control: Filters sequences based on minimum length and metadata completeness.

Data Structure

The final output is final_hantavirus_dataset.csv, which includes:

• accession_id: Unique NCBI ID.
• host_label: Host classification (Human, Rodent, Others).
• geo_label_broad: Regional classification (Americas, Europe, Asia).
• sequence: Pure RNA nucleotide sequence.
• lokasi_geografis_koordinat: Lat/Lon points for map visualization.

Preparation & Installation

Use Python 3.12+

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

How to Run the Pipeline

Run the scripts sequentially according to the data dependency flow:

1. python src/01_fetch_ncbi.py: Fetches raw data.
2. python src/02_clean_labels.py: Cleans data and generates labels.
3. python src/03_geocoding.py: Determines location coordinates.

Open notebooks/eda_hantavirus.ipynb to view the data distribution analysis.

Dataset Card

Configurations

The Hub exposes three configurations matching the pipeline stages:

Config	File	Rows	Description
default	data/processed/final_hantavirus_dataset.csv	9,846	Fully cleaned, labeled, and geocoded data (recommended).
raw	data/raw/raw_hantavirus_ncbi.csv	9,950	Unprocessed records as fetched from NCBI GenBank.
interim	data/interim/interim_hantavirus.csv	9,846	Cleaned and labeled, prior to geocoding.

Usage

from datasets import load_dataset

# Recommended processed split
ds = load_dataset("<namespace>/<dataset-name>")  # config "default"

# Or load a specific pipeline stage
raw = load_dataset("<namespace>/<dataset-name>", "raw")
interim = load_dataset("<namespace>/<dataset-name>", "interim")

Data Fields

• accession_id (string): Unique NCBI GenBank accession ID.
• species_label (string): Orthohantavirus species (e.g. Orthohantavirus andesense).
• raw_host (string): Original host annotation from the source record (e.g. Homo sapiens).
• lokasi_geografis_name (string): Raw strain/location string from the source record.
• segment_type (string): RNA genome segment — S, M, or L.
• sequence_length (int64): Nucleotide length of the sequence.
• sequence (string): Nucleotide sequence (A/C/G/T).
• host_label (string, default/interim only): Standardized host class — Human, Rodent, Others, Unknown.
• geo_label_broad (string, default/interim only): Broad region — Americas, Europe, Asia, Others, Unknown.
• lokasi_geografis_koordinat (string, default only): lat, lon coordinates from Nominatim geocoding (empty when unresolved).

Label Distribution (default, n=9,846)

• host_label: Rodent 6,451 · Human 1,495 · Others 1,321 · Unknown 579
• geo_label_broad: Others 9,505 · Europe 255 · Unknown 51 · Americas 22 · Asia 13
• segment_type: S 4,175 · L 2,931 · M 2,729 (plus a few unnormalized variants)
• coordinates resolved: 2,535 / 9,846 rows

Source & Curation

Sequences and metadata are fetched from NCBI GenBank via Biopython, then standardized (host/species/geography labels), quality-filtered (minimum length and metadata completeness), and geocoded with the Nominatim API. See src/ for the reproducible pipeline.

Considerations

• Labels are machine-generated heuristics from unstructured source metadata and may contain noise.
• Classes are imbalanced (notably geo_label_broad, dominated by Others); account for this when training.
• The underlying records originate from NCBI GenBank and are subject to GenBank's terms of use.

Licensing

This dataset and pipeline are released under the Apache-2.0 license.