Skip to content

Roadmap

GIAE is an evolving platform. The strategic direction is moving from sequence-based interpretation to multi-modal structural and functional reasoning. This page is a living document — we update it as work lands.

For the broader product vision (hosted SaaS, enterprise tier), see PRODUCT_STRATEGY.md.

✅ Recently shipped (0.2.2)

Phase What landed Lift
Phase 1 pyrodigal-backed ORF prediction Replaced naive ATG scanner
Phase 2 Motif confidence tiering + --mode flag Reduced false positives on motifs
Phase 3 Diamond BLASTP plugin ~10× faster than NCBI BLAST+
Phase 4A Validation suite (precision/recall/F1) Reproducible quality metrics
Phase 4B Aragorn (tRNA) + Barrnap (rRNA) detection Non-coding RNA coverage
Phase 5 Short-ORF rescue (RBS + codon usage gate) +1 TP each on λ and T7, 0 new FP
Phase 6 Functional annotation depth (COG / GO / normalised product) Closed Bakta's last advantage
Phase 7 Phage-aware nested ORF detection (--phage) T7 F1 86.2 % → 88.1 %, 0 new FP
Phase C REST API + Celery worker + Docker stack Production deployment surface
Phase D Version + report-label fixes Polish

The cumulative effect: GIAE matches or beats Bakta on three reference phage genomes (benchmarks →), and runs as a CLI, library, or REST service.

🎯 Next horizon (0.3.0)

Foldseek / AlphaFold structural homology

  • What: A FoldseekPlugin that emits STRUCTURAL_HOMOLOGY evidence by searching the AlphaFold Database (AFDB) for structural homologs.
  • Why: PhiX174 and other compact phages have proteins with no detectable sequence-based signal but conserved 3D folds. Sequence search is at its biological floor here; structural search isn't.
  • Status: Design ready. The STRUCTURAL_HOMOLOGY evidence type already exists. Plugin scaffold sketched in extending.md.
  • Hard part: AFDB is large (~250 GB downloaded, several TB uncompressed). Need a hosted index option for users without the storage.

Translational coupling detection

  • What: Detect ORFs whose start codon sits within a small window downstream of a stop codon in the same frame — the canonical signature of translationally coupled overlapping genes (PhiX174 A*, B, E, K all use this).
  • Why: Closes the PhiX174 60 % F1 ceiling that even Bakta can't beat with sequence-based methods.
  • Status: Diagnostic evidence collected (the missing genes consistently have RBS=0 at canonical positions). Algorithm sketched.

Bacterial-genome scaling

  • What: Validation suite on 4–6 Mb bacterial genomes (E. coli K-12, B. subtilis, M. tuberculosis). Memory profiling, parallel-worker benchmarks, possibly a streaming gene-iterator.
  • Why: Currently validated on phages only. Bacterial-scale is the real-world target.
  • Hard part: Per-gene API calls in online mode become a throughput bottleneck. Need batch UniProt requests or a local embedding-based pre-filter.

🌅 Mid-horizon (0.4.0)

Comparative-genomics mode

  • What: giae diff genome_a.gb genome_b.gb — show functional divergence between two interpretations side by side.
  • Why: Strain comparison, phage variant analysis, host adaptation studies all want this.
  • Status: Concept; data model already supports it.

Evidence-network visualisation

  • What: Web UI panel showing evidence flow per gene as a graph — each evidence type a node, conflicts highlighted, reasoning chain as edges.
  • Why: The reasoning chain is GIAE's central differentiator; visualising it makes it immediately obvious to non-bioinformatician consumers.
  • Status: Prototyping in the Next.js frontend.

Streaming output

  • What: Server-sent events on /api/v1/jobs/{id}/stream — emit Interpretation records as they complete, instead of waiting for the whole genome.
  • Why: Bacterial genomes are slow; streaming makes progress visible to the user.

🌌 Long-horizon (1.0.0)

Hosted SaaS

  • Frontend dashboard (already scaffolded in frontend/).
  • API key tier with rate limiting and metering.
  • "Dark matter" cross-genome database — index every unannotated protein across all uploaded genomes, continuously re-annotated by structural AI as new methods land.

See PRODUCT_STRATEGY.md for the full product vision.

Self-correcting annotations

  • What: When new structural data, UniProt entries, or HMM models appear, re-interpret old GIAE predictions and surface what changed.
  • Why: Annotations rot. Our reasoning chains make it possible to ask "would this prediction still hold today?" — that's a useful capability nobody else has.

How to influence the roadmap

  • Open a feature request with a clear use case → issue tracker
  • Send a PR — extension points are documented in extending.md
  • Bring a benchmark — a real genome we don't currently handle well is the most actionable signal we can get
  • Discuss it first for anything that would touch the engine core (discussions)

Last updated: 2026-05-09 (v0.2.2 release)