| # Gene identity and orthology contract | |
| VariantHound joins dog, human, and mouse evidence only through release-qualified stable identifiers. Gene symbols are useful labels, but they are mutable and are never accepted as the sole join key. | |
| ## Fail-closed release states | |
| `data/manifests/orthology.json` has two states: | |
| - `unmaterialized`: the policy is defined, but the Ensembl release, assemblies, timestamp, and checksum are deliberately null and the mapping arrays must be empty; | |
| - `pinned`: all release metadata and species annotation assemblies are present, the source artifact has a SHA-256 checksum, and at least one validated mapping exists. | |
| Ranking builds must refuse orthology evidence unless the manifest is `pinned`. This keeps a development placeholder or live API response from becoming scientific evidence by accident. | |
| The repository keeps the lightweight checked-in manifest `unmaterialized`. The manual Hugging Face workflow uses the immutable descriptor in `data/manifests/ensembl-compara-116.json` to generate and validate the full `pinned` manifest at build time. | |
| ## Identifier rules | |
| 1. Resolve Ensembl gene IDs first and NCBI Gene IDs second. | |
| 2. Preserve the species taxon on every identity. | |
| 3. Retain canonical symbols and aliases for display and search only. | |
| 4. Record retired, replaced, ambiguous, and unverified IDs instead of silently rewriting them. | |
| 5. Put records that cannot be resolved deterministically into the `unresolved` report and exclude them from scoring. | |
| The normalized record shape is defined by `schemas/gene-identity.schema.json`. | |
| ## Orthology rules | |
| The browser-facing mapping is dog-centered. Each record is one dog-to-human or dog-to-mouse edge and preserves: | |
| - the Ensembl homology identifier and source record identifier; | |
| - one-to-one, one-to-many, or many-to-many homology type; | |
| - Ensembl's high-confidence flag and normalized confidence when available; | |
| - an explicit `resolved` or `ambiguous` status; | |
| - warnings needed by downstream explanations. | |
| Ambiguous edges are retained and flagged so the ranker can expose or down-weight them; they are never collapsed by choosing the first symbol match. | |
| ## Assembly rules | |
| Gene identities record the annotation assembly supplied by the pinned Ensembl release. Coordinate-bearing canine data has a separate target of `UU_Cfam_GSD_1.0`. Any liftover must retain original and target coordinates, chain-file identity, and mapping status. Orthology does not itself imply that coordinates were lifted successfully. | |
| ## Materialization sequence | |
| 1. Pin one Ensembl release and download a complete dog-human-mouse Compara export. | |
| 2. Record the raw artifact checksum and the three annotation assemblies. | |
| 3. Normalize stable identifiers and create an unresolved-ID report. | |
| 4. Expand one-to-many and many-to-many relationships into explicit edges. | |
| 5. Validate the complete manifest, then change `release_status` to `pinned`. | |
| 6. Compare row counts, unresolved rates, and ID drift against the previous release before publication. | |
| ## Ensembl 116 materialization | |
| Ensembl's genome-specific homology dumps are not complete independently: each contains an arbitrary subset intended to reduce redundancy. VariantHound therefore verifies and unions the dog, human, and mouse `protein_default` files before filtering dog-human and dog-mouse edges. | |
| ```bash | |
| python3 pipelines/ingest_ensembl_orthology.py | |
| python3 pipelines/validate_data.py --orthology dist/ensembl/orthology.json | |
| ``` | |
| The release is accepted only when it produces exactly 41,504 unique edges with the pinned species and homology-type counts. The normalized data orients every edge dog-first, retains source homology IDs and metrics, and marks every non-one-to-one relationship ambiguous. Compara does not include canonical gene symbols in this bulk table, so symbols remain null rather than being guessed or used as join keys. | |
| This contract is intentionally in place before bulk ingestion because symbol drift and assembly changes are data-layer concerns, not evaluation-time cleanup. | |