hiyata/hantavirus-genome-dataset

Hantavirus Genome Dataset

Dataset Summary

The Hantavirus Genome Dataset is a collection of Hantaviridae genomic sequences compiled from NCBI GenBank and annotated with taxonomic, clinical, epidemiological, and host metadata. This dataset is specific to the virus family Hantaviridae, and provides annotations to support straightforward model training, including segment identity, isolate linkage across the tripartite genome, clade classification, clinical syndrome association, reservoir host species, and reassortment indicators.

Hantaviruses are negative-sense, single-stranded RNA viruses with a tripartite segmented genome (L, M, S segments) transmitted primarily through contact with infected rodents. They are the causative agents of Hemorrhagic Fever with Renal Syndrome (HFRS) in the Old World and Hantavirus Pulmonary Syndrome (HPS) in the Americas, with a combined global burden of tens of thousands of cases annually. This dataset was assembled to support genomic studies of host specificity, zoonotic spillover risk, phylogenetics, and machine learning-based classification of Hantaviridae.

If you use this dataset, please cite:

@article{carbajo2026sequence,
  author  = {Carbajo, Alan L and Vensko, Taylor A and Pellett, Philip E},
  title   = {Sequence Based Virus Host Prediction: A Curated Dataset and
             Generalizable Framework for Training Artificial Intelligence
             to Identify Viruses of Humans},
  year    = {2026},
  url     = {https://doi.org/10.1093/ve/veag009}
}

All genomic sequences were sourced from NCBI GenBank/NCBI Virus. Please also acknowledge NCBI in any work that uses this dataset.

---

Key Features

• Segment records — every row is one segment (L, M, or S), reflecting how Hantaviridae genomes are deposited in public databases
• Isolate linkage — an isolate_id field groups all three segments from the same physical isolate, enabling reconstruction of complete tripartite genomes
• Clade assignment — Old World vs. New World classification and strain-level clade labels for all well-characterised viruses
• Clinical syndrome — HFRS, HPS, mild (nephropathia epidemica), non-human, or unknown for every record
• Reservoir host — specific reservoir species with confidence classification (primary vs. incidental)
• Reassortment flagging — isolates where L, M, and S segments cluster into different clades are flagged as reassortment_suspected
• Sequence quality — per-segment completeness scores and per-protein coding completeness fractions
• Three-tier host labeling — human vs. non-human labels assigned via rule-based matching and AI-assisted annotation following the methodology of Carbajo et al. (2026)

---

Dataset Structure

Each row represents one genomic segment record. Records from the same physical isolate share an isolate_id and can be grouped to reconstruct full genomes or study reassortment.

Data Fields

Core Sequence and Taxonomy

Field	Type	Description
sequence	string	Nucleotide sequence (ACGT only, ambiguous characters removed)
accession	string	GenBank accession number including version
genus	string	Viral genus
virus_name	string	Full virus name as recorded in GenBank
strain_name	string	Strain or isolate identifier
isolation_date	string	Collection or isolation date
location	string	Raw geographic location from GenBank
standardized_location	string	Country-level standardized location
isolation_source	string	Source tissue or host from which the virus was isolated

Host Information

Field	Type	Description
host	string	Primary host classification: "human" or "non-human"
standardized_host	string	Standardized scientific name of the host organism
host_category	string	Broad host category (e.g., Mammal)
zoonotic	bool	Whether the virus is known to have zoonotic potential
lab_culture	bool	Whether the sequence was derived from a laboratory culture
wastewater_sewage	bool	Whether isolated from environmental wastewater
gemini_annotated	bool	Whether Google Gemini AI was used for host annotation (Tier 3)

Segment Information

Field	Type	Description
segment	string	Genome segment: "L", "M", "S", or "Unknown"
isolate_id	string	Normalised identifier linking L, M, and S segments from the same isolate
has_all_segments	bool	True if all three segments are present in the dataset for this isolate
matched_accessions	string	JSON object mapping segment labels to accessions: {"L": "AB123.1", "M": "AB124.1", "S": "AB125.1"}
segment_completeness	float	Fraction of expected full-length segment present (0.0–1.0)
is_complete_segment	bool	True if segment_completeness ≥ 0.90
coding_completeness	string	JSON object of per-protein completeness fractions (e.g., {"RdRp": 0.98} for L; {"Gn": 1.0, "Gc": 0.95} for M; {"N": 1.0} for S)

Hantaviridae-Specific Annotations

Field	Type	Description
clade	string	Phylogenetic clade assignment (e.g., "Old World / Hantaan", "New World / Sin Nombre")
clinical_syndrome	string	Associated human disease: "HFRS", "HPS", "mild", "non-human", or "unknown"
pathogenicity_class	string	Human pathogenicity: "high", "moderate", "low", or "unknown"
reservoir_host	string	Primary reservoir species (e.g., "Peromyscus maniculatus")
reservoir_confidence	string	Confidence in reservoir assignment: "primary", "incidental", or "unknown"
collection_method	string	How the sample was collected: "trapped rodent", "human clinical", "cell culture", "environmental", or "unknown"
outbreak_associated	bool	Whether this strain has been associated with a documented human outbreak
reassortment_suspected	bool	True when segments of the same isolate assign to different clades

---

Annotation Methods

All Hantaviridae-specific annotations were assigned computationally using a curated knowledge base compiled from ICTV taxonomy, CDC outbreak records, and peer-reviewed literature. These are reference-based annotations, not predictions — a record receives an annotation only when its strain name, virus name, or metadata contains a substring matching a curated entry. Records that do not match any entry receive "Unknown" rather than a guess.

Clade Assignment

Clade labels were assigned by substring matching the virus name, strain name, and organism field against a curated dictionary of characterised Hantaviridae strains following ICTV nomenclature and published phylogenetic analyses. For example, any record whose metadata contains "sin nombre" is assigned "New World / Sin Nombre"; any record containing "puumala" is assigned "Old World / Puumala". Records whose metadata does not match any known strain receive "Unknown".

This approach assigns clades based on established strain identity rather than sequence-based phylogenetic placement. Researchers requiring phylogenetically confirmed clade assignments should perform independent tree-based analysis.

Clinical Syndrome

Clinical syndrome was derived deterministically from clade assignment, using the well-established association between Hantavirus species and disease presentation:

Clinical syndrome	Assignment basis
HFRS	Hantaan, Seoul, Dobrava-Belgrade — documented causes of hemorrhagic fever with renal syndrome
mild	Puumala, Saaremaa — causative agents of nephropathia epidemica, a mild HFRS variant
HPS	Sin Nombre, Andes, Bayou, Black Creek Canal, and related New World strains — causative agents of hantavirus pulmonary syndrome
non-human	Tula, Prospect Hill, Thottapalayam, bat-associated strains — no documented human disease
unknown	Strain not matched to any characterised species in the knowledge base

Pathogenicity Class

Pathogenicity was assigned directly from clinical syndrome, reflecting documented case fatality rates in the literature:

Pathogenicity class	Clinical syndrome	Approximate CFR
high	HFRS, HPS	5–40% depending on strain
moderate	mild	<1%
low	non-human	No documented human disease
unknown	unknown	Insufficient data

This is a coarse classification for filtering purposes. Researchers studying pathogenicity quantitatively should consult primary literature for strain-specific case fatality rates.

Reservoir Host

Reservoir host was assigned by substring matching against the host, isolation_source, organism, and description fields from the GenBank record, using a curated dictionary of known Hantaviridae reservoir associations. The confidence field reflects the nature of the association:

• primary — the assigned species is the well-established natural reservoir based on published ecological and serological studies (e.g., Peromyscus maniculatus for Sin Nombre virus)
• incidental — the host is a spillover host, not the natural reservoir (e.g., Homo sapiens)
• unknown — the metadata did not match any entry in the reservoir dictionary

Reservoir assignments are based on metadata present in the GenBank record. If a record does not specify a host or the host is ambiguous, the assignment will be "Unknown".

Outbreak Association

A record is flagged outbreak_associated = True if the virus name or strain name contains a substring matching a list of strains with documented human outbreaks compiled from CDC and WHO outbreak records. This flag indicates association with an outbreak-causing species, not that the specific sequenced isolate was itself sampled during an outbreak.

Reassortment Suspected

reassortment_suspected = True is set at the isolate level for any isolate where L, M, and S segments assign to different clades. This is a computational heuristic — discordant clade assignments can indicate genuine reassortment, but can also reflect incomplete annotation in GenBank records, novel divergent strains, or annotation errors. This flag surfaces candidates for downstream investigation and does not confirm reassortment events. Confirmation requires phylogenetic analysis with appropriate reference sequences.

Host Labeling (human vs. non-human)

Host labels were assigned using the three-tier system described in Carbajo et al. (2026):

1. Tier 1 — Regex matching of scientific names and common terms against the host qualifier and organism fields in the GenBank source feature
2. Tier 2 — Extended pattern matching across isolation source and description fields, including clinical terminology
3. Tier 3 — Google Gemini AI (gemini-1.5-flash) contextual inference for records unresolvable by Tiers 1 and 2; flagged with gemini_annotated = True

For Hantaviridae, the large majority of records resolve at Tier 1 or Tier 2 since most sequences are submitted with explicit rodent host metadata.

---

Segment Biology and the Tripartite Genome

Hantaviruses have a tripartite negative-sense RNA genome. Each segment encodes distinct proteins critical to the viral replication cycle:

Segment	Length (~bp)	Encoded Proteins	Function
L	~6,530	RdRp	RNA-dependent RNA polymerase — genome replication and transcription
M	~3,616	Gn, Gc	Envelope glycoproteins — cell entry, neutralizing antibody targets
S	~1,696	N (± NSs)	Nucleocapsid protein — genome encapsidation, innate immune antagonism

Because GenBank submissions store each segment as an independent record, this dataset follows the same convention. To work with complete genomes, filter to rows where has_all_segments == True and group by isolate_id.

---

A Note on Geographic Coordinates

Geographic coordinates are not included in this dataset. NCBI GenBank records for Hantaviridae do not systematically include precise coordinates — location information is recorded as free-text country or region names in the geo_loc_name field, available here via location and standardized_location. Researchers requiring coordinates should geocode the standardized_location field using a gazetteer or cross-reference accession numbers with the original publications.

---

Clade Reference Table

Clade	Geography	Clinical Syndrome	Primary Reservoir
Old World / Hantaan	East Asia	HFRS (severe)	Apodemus agrarius
Old World / Seoul	Global	HFRS (moderate)	Rattus norvegicus
Old World / Dobrava-Belgrade	Europe	HFRS (severe)	Apodemus flavicollis
Old World / Puumala	Europe	mild (NE)	Myodes glareolus
Old World / Tula	Europe	non-human	Microtus spp.
Old World / Amur	East Asia	HFRS	Apodemus peninsulae
New World / Sin Nombre	North America	HPS (severe)	Peromyscus maniculatus
New World / Andes	South America	HPS (severe)	Oligoryzomys longicaudatus
New World / Bayou	North America	HPS	Oryzomys palustris
New World / Black Creek Canal	North America	HPS	Sigmodon hispidus
New World / Laguna Negra	South America	HPS	Calomys laucha
New World / Araraquara	South America	HPS	Necromys lasiurus
Bat-associated	Africa, Asia	non-human	Chiroptera spp.

---

Limitations and Biases

• Sampling bias — sequences from clinically significant strains and well-studied regions are overrepresented relative to poorly-characterised or environmentally-isolated viruses
• Segment imbalance — S segment records are the most abundant due to widespread use as a typing target; not all isolates have all three segments sequenced
• Annotation coverage — clade, clinical syndrome, reservoir, and pathogenicity fields are "Unknown" for novel or divergent strains not yet in the knowledge base
• Reassortment flag — a computational heuristic only; confirmation requires independent phylogenetic analysis
• No coordinates — location data reflects GenBank text fields; precise geographic coordinates are unavailable (see note above)
• Reference-based annotations — all Hantaviridae-specific fields are derived from strain name matching; a misnamed or ambiguously named GenBank record may receive an incorrect or "Unknown" annotation

---

Usage Examples

Load the dataset

from datasets import load_dataset

ds = load_dataset("hiyata/hantavirus-genome-dataset", split="train")
print(ds)
print(ds.column_names)

Filter by clade and clinical syndrome

hfrs = ds.filter(lambda x: x["clinical_syndrome"] == "HFRS")
hps  = ds.filter(lambda x: x["clinical_syndrome"] == "HPS")
print(f"HFRS records: {len(hfrs)}")
print(f"HPS records:  {len(hps)}")

Get complete isolates (all 3 segments present)

complete = ds.filter(lambda x: x["has_all_segments"] == True)
df = complete.to_pandas()

for isolate_id, group in df.groupby("isolate_id"):
    segs = group.set_index("segment")["sequence"].to_dict()
    # segs = {"L": "ATCG...", "M": "ATCG...", "S": "ATCG..."}

K-mer vectorization for machine learning

import numpy as np
from itertools import product

def kmer_freq(seq, k=4):
    kd    = {''.join(k): i for i, k in enumerate(product('ACGT', repeat=k))}
    freq  = np.zeros(4**k)
    total = len(seq) - k + 1
    for i in range(total):
        kmer = seq[i:i+k]
        if kmer in kd:
            freq[kd[kmer]] += 1
    return freq / total if total > 0 else freq

X = np.array([kmer_freq(row["sequence"]) for row in ds])
y = np.array([1 if row["host"] == "human" else 0 for row in ds])

---

License

MIT. Sequences sourced from NCBI GenBank are in the public domain under their respective submission terms.