Genome-scale annotation

DeepGO-GSPA

Annotate a whole genome or metagenome and assess it per contig. Every protein is annotated with DeepGO-PlusPlus-Light (a multi-evidence ensemble); GSPA then computes genome-scale quality metrics and can optionally enforce taxon consistency, completeness and coherence over the Gene Ontology — recording full provenance for every call.

1 Input data

A nucleotide genome (optionally with a GFF3), or pre-called proteins.

Nucleotide FASTA, one contig per record (a chromosome, a MAG, or many contigs). If you omit GFF3, genes are called automatically for prokaryotes only.
If supplied, the CDS in it are translated and mapped to their contig instead of calling genes. For eukaryotes, viruses, or any assembly with introns/non-standard gene models, provide GFF3 or upload proteins directly.
— or —
Already have predicted proteins? Upload them instead of a genome; they are annotated on a single synthetic contig.

2 Organism

Inferred from the predicted functions by default — override only if you already know it.

Leave on "Infer" and the domain is read off the predictions: the high-confidence GO terms are tested against the NCBI taxon constraints, and the proteome is placed in the domain whose constraints they violate least (Asaad-style). The inferred domain is shown with its evidence and used for taxon-consistency enforcement. Override only if you already know the organism.
Relaxes quality thresholds for a fragmented / incomplete bin.

3 Genome-scale metrics

Function predictor is fixed to DeepGO-PlusPlus-Light (DIAMOND BLAST-KNN + STRING Net-KNN + a learned integrator).

Per contig is the right unit for a metagenome or multi-replicon assembly — a chromosome, a plasmid and a phage score separately, which pooling would hide.
4 Quality enforcement — optional, off by default

Beyond measuring quality, GSPA can repair the prediction set. Each action is consistency-gated and fully logged.

Run the SAT4J taxon-constraint pass: GO terms that cannot occur in the organism (NCBI taxon constraints) are acted on. Vendored from the genome-scale-pfp-adjust constraints.
Each missing essential function is promoted onto the protein with the strongest near-ancestor evidence and marked imputed — never fabricated when there is no evidence.
Fix obligate heteromeric-complex singletons and missing has_part partners. Advanced: runs the ELK reasoner, so it is slower.
Record how every annotation was assigned — the originating predictor and any enforcement action — and emit the enforcement-actions log.
Load an example into the form:
FASTA GFF3
FASTA GFF3
FASTA GFF3
FASTA GFF3
Examples only fill the fields; press Annotate genome when you are ready. Download links expose the exact input files so you can inspect or adapt them.

Runs asynchronously — you'll get a live result page. Each example is ~22 genes and finishes in about a minute.

How it works

  • Genes → proteinsCDS from your GFF3 are translated. If no GFF3 is supplied, genes are called for prokaryotic genomes only.
  • Function predictionEach protein is annotated with DeepGO-PlusPlus-Light against the Gene Ontology.
  • Infer the organismThe domain is read off the confident predictions via the GO taxon constraints (Asaad-style) — not assumed — and shown with its evidence.
  • Per-contig metricsCompleteness, coherence, consistency, information content — one row per contig.
  • Optional enforcementRemove taxon-impossible terms, impute essentials, repair coherence (SAT4J).
  • ProvenanceEvery assigned (or removed) function is traced to its source.

The examples

Four demo assemblies are bundled with matching GFF3 annotations. Loading an example sets the upload fields and organism domain but does not submit the job. Gene calling is deliberately limited to prokaryotes; for eukaryotes and viruses, use a GFF3 annotation or upload proteins directly. Distinguishing prokaryotic from eukaryotic sequence reliably from a raw FASTA alone is not a cheap browser-side check, so this form uses the organism selection and validates unsafe combinations.