ADR-D19: F-RESEARCH targets

Status:

Accepted

Date:

2026-05-05

Phase:

F-RESEARCH

Context

F-RESEARCH risks becoming an open campaign. Without an explicit cap, research questions multiply and the phase loses focus. The plan needs a small, ordered set of ideas.

Decision

Three directed targets, in order:

  1. Lineage feature (1 week). Whether a candidate GO term is an ancestor or descendant of a term already known for the protein. Implementable as a single registry feature.

  2. GeOKG embeddings (1 week, conditional). Replace anc2vec features with multi-curvature hyperbolic + Euclidean GO embeddings (Bioinformatics 2025).

  3. Multi-K ensemble (1 week, optional). Combine K ∈ {5, 10, 20} instead of K=5 fixed.

Each idea produces an ExperimentRun. Wins integrate into the canonical pipeline; losses go to the insights appendix (D30).

Consequences

  • Phase capped at 2-3 weeks.

  • Ideas that did not make the list (PROTEA-DL, retrieval-neural, R-GCN over GO-DAG) deferred to F11 post-defense.

Resolution

Closed.