Nhoodie/hgt-real-mutation-pairs

---

ℹ️ Dataset Note

This is the primary ground-truth dataset for HGT mutation prediction. It contains real biological mutation pairs with authentic HGT signatures (Ti/Tv ratio ~0.77, dominated by C>T and T>C transitions consistent with viral polymerase error patterns). This dataset is not affected by the ICI-DC degeneracy issues identified in synthetic datasets.

---

HGT Real Mutation Pairs

Real mutation pairs across 8 taxonomic domains for cross-domain horizontal gene transfer (HGT) mutation prediction.

Dataset Details

Split	Pairs	Mean mutation rate	Mean distance
train	3,317	~4.2%	~6 nt
test	826	~1.2%	2.2 nt

Domains

Domain	Sequences	Rationale
Archaea	1,260	Thermophile/halophile HGT
Bacteria	630	Primary HGT domain
Lichenes_HGT	210	Inter-kingdom HGT (fungi↔algae)
Metazoa	1,680	Microbiome HGT
Satellite_Viruses	84	Subviral agents
Viridiplantae	420	Limited bacterial→plant HGT
Viroids	294	Small RNA pathogens
Viruses	420	Phage transduction

Format

Each line is a JSON object:

{"parent": "ATGCGT...", "child": "ATGCGA...", "domain": "Bacteria", "distance": 3, "source_ids": ["AB123456"]}

Pipeline

1. Sequences fetched from NCBI with HGT-weighted sampling across 8 domains
2. Stop-codon-aware greedy chunking (~500bp)
3. Domain-aware nearest-neighbor pair building (C Levenshtein)
4. Leakage-free train/test split via connected-component clustering

Related Datasets

• Nhoodie/hgt-bootstrap-v1-synthetic — 8,112 synthetic pairs (ICI-DC v1)
• Nhoodie/hgt-bootstrap-v2-synthetic — 256K synthetic pairs (ICI-DC v2)