Work

A Verified Gene-Model Dataset for Multi-Host Flavor-Trait Engineering: 28 Pinned Reference Genes and 7,878 Host Models Across 87 Species, with an Independent Second-Method Check v1

Year: 2026
Type: preprint
Abstract: Designing a flavor-trait edit in a crop requires knowing, for that specific host genome, which gene carries the target activity — not a reference gene from a model species. We describe the verified gene-model dataset that underlies the Scoville Splice design packages: 28 reference genes across eight trait modules, each pinned to a curated UniProt entry and re-checked by name, and 7,878 host gene models built across 87 host species (81 genome assemblies). Each host model passes fixed identity, coverage and completeness gates and a reciprocal-best-hit test against the curated references. As an independent check, a second method — nearest-neighbour comparison of protein-language-model (ESM2) embeddings against 169 references and decoys — agreed with the reciprocal-best-hit assignment for 88.1% of models, with a further 5.6% matching a different member of the same protein family. We report the acceptance rules, the catalytic-site conservation of the models, and the annotation pitfalls that required manual handling: merged RefSeq models, RefSeq exception records, draft-assembly gaps, near-identical gene copies, and inferred coding termini. All work is computational. § 1 Introduction A design package that tells a customer to knock out or over-express a gene is only as good as its identification of that gene in the customer's host. A reference enzyme from Arabidopsis or Capsicum is a starting point; the sequence that must actually be edited is the host's own ortholog, which may differ in identity, copy number and annotation quality. Early design drafts in this programme leaned on reference sequences and placeholder constructs; the dataset described here replaces that with host-specific gene models, each resolved from the host's own genome and each carrying an explicit record of how confident the assignment is. Two commitments shape the dataset. First, every reference gene is pinned to a curated database entry and re-fetched and name-checked, so that a module always resolves to the intended enzyme. Second, every host model is accepted only against fixed, stated gates, and anything resolved by inference — an extended start codon, a close-relative stand-in, a gene with near-identical copies — is flagged in the model rather than hidden. This paper documents the dataset so that the design packages built on it can be audited independently.
Author: P.I. Ford
Language: English
Open Access status: green
OpenAlex ID: W7236203639
Cites: 5
Cited by:
Related to: 10