GIAE

Genome Interpretation & Annotation Engine — every prediction shows its reasoning.

GIAE is an explainability-first annotation engine. Where Bakta and Prokka return labels, GIAE returns labels with the full evidence stack, a numerical confidence score, the reasoning chain that produced it, and the alternatives it considered.

It's also faster and more accurate than Bakta on the genomes we've benchmarked.

Genome	GIAE F1	Bakta F1	Δ
phiX174	60.0 %	60.0 %	tied
λ phage	79.2 %	72.6 %	+6.6 %
T7	88.1 %	85.2 %	+2.9 %

Reproduce these numbers →

---

What's here

• Quickstart

  ---

  Five-minute path from pip install to your first annotated genome. CLI, Python library, and Docker walkthroughs.

• CLI reference

  ---

  Every command, every flag. interpret, parse, db, serve, worker, analyze, quick, info.

• Python API

  ---

  Use GIAE programmatically. Interpreter, Genome, Gene, Evidence, Interpretation, plugin protocol.

• REST API

  ---

  Full HTTP surface — auth, jobs, API keys, results. Built on FastAPI + Celery + Postgres + Redis.

• Architecture

  ---

  How GIAE actually works. Evidence flow, hypothesis generation, confidence scoring, conflict detection.

• Benchmarks

  ---

  Reproducible head-to-head against Bakta. Methodology, raw numbers, per-genome breakdowns.

• Deployment

  ---

  Run the full stack with docker compose. Env vars, scaling, ops.

• Extending GIAE

  ---

  Add a plugin — new evidence source, new analysis tool, new database.

---

Why explainability matters

Most annotation tools treat the output as the answer: a label, possibly a score, end of story. That works fine for well-characterised genes where the consensus is strong. It breaks down at the edges, which is where the interesting biology lives:

• Conflicting evidence — homology says X, motif says Y, domain says Z. Most tools silently pick one. GIAE flags the conflict.
• Weak signals — a single PROSITE hit isn't a function call. GIAE caps single-source-only hypotheses at 0.45 (LOW) so they don't masquerade as moderate.
• Dark matter — proteins with no sequence-detectable signal aren't failures, they're research priorities. GIAE ranks them.
• Curator vs. computational — a GenBank product annotation isn't the same kind of evidence as a BLAST hit. GIAE keeps them as separately typed evidence with different confidence weights.

The core idea: make uncertainty explicit, make reasoning auditable, let the user decide what to trust.

---

Architecture in 30 seconds

Genome ──► ORF prediction (pyrodigal + rescue + nested-finder)
        ──► Evidence extraction (PROSITE, Pfam, Diamond, UniProt, ESM-2)
        ──► Aggregation → Hypothesis generation → Confidence scoring
        ──► Functional annotation (COG / GO / normalised product)
        ──► Markdown · JSON · HTML · REST API

Every step keeps full provenance. The output isn't just "DNA polymerase III, HIGH", it's "DNA polymerase III, 0.87 HIGH, supported by Pfam PF00136 (e-value 2.1e-14) + UniProt P03722 + PROSITE PS51123, with no competing hypothesis above 0.50, COG category L (Replication), GO terms [GO:0003677, GO:0003887, GO:0006260]".

Read the full architecture doc →

---

Latest

GIAE 0.2.2 ships eight major capabilities at once: pyrodigal-backed ORF prediction, motif confidence tiering, Diamond BLASTP, validation suite, tRNA/rRNA detection, short-ORF rescue, phage-aware nested ORF detection, COG/GO functional annotation depth, and a full REST API server. See the changelog for the per-phase breakdown.

---

Citation

If GIAE contributes to your published work, please cite:

@software{giae_2026,
  author  = {{GIAE Contributors}},
  title   = {GIAE: Genome Interpretation and Annotation Engine},
  year    = {2026},
  version = {0.2.2},
  url     = {https://github.com/Ayo-Cyber/GIAE},
}

A formal application-note manuscript is in preparation.