7. Outputs

With --output-prefix testing-4_<chr>, each reconcile pass writes a set of files sharing that stem. (For a --scatter run, the merged outputs use --out-prefix, defaulting to --output-prefix, plus a manifest.)

Files

File	Contents
<prefix>.gff3	Full reconciled annotation, all genes, all tiers, isoform-aware (1-based GFF3).
<prefix>.tier1.gff3	Tier 1 only.
<prefix>.tier2.gff3	Cumulative: Tier 1 + Tier 2.
<prefix>.tier3.gff3	Cumulative: Tier 1 + Tier 2 + Tier 3.
<prefix>.report.tsv	Per-gene report, one row per gene (columns below).
<prefix>.id_map.json	Helixer-locus → HFG ID map (keeps IDs stable across reruns).
<prefix>.report.json / .report.html / .run_stats.json	Machine- and human-readable run summaries.
<out-prefix>.manifest.json	(chunked runs only) plan + chunk prefixes + id_map.

Note

The tier GFF3s are cumulative: .tier2.gff3 contains Tiers 1–2, .tier3.gff3 contains Tiers 1–3. Pick the file at the confidence floor you want, or slice the full annotation later with utils filter --max-tier.

Reading the tiers

Tiers rank evidence strength, Tier 1 is what you trust most. The assignment is driven by the model's ORF and its homology support:

Tier	Meaning
Tier 1	Coherent CDS + protein-homology support (DIAMOND hit or positive BLAST score). The evidence-backed, highest-confidence set.
Tier 2	Coherent CDS but no homology (CDS-only), a good ORF without a protein match.
Tier 3	Helixer-only backstop, no projected CDS, but expressed / low-coverage RNA-seq support, a rescued silent gene.
Tier 4	Silent backstop with no CDS and no expression; also where TE-reclassified models are demoted.

Tip

Reconciliation preserves recall: silent Helixer genes with a valid ORF are kept rather than dropped. That is what Tiers 3–4 are for, filter by tier to trade completeness for precision, instead of trusting a flat annotation.

Reading the biotypes

Biotype is driven by the model's own ORF, not by whether evidence happens to corroborate it:

• protein_coding, the model carries a complete/coherent ORF. A good-ORF gene is coding regardless of expression or homology; absence of corroboration lowers confidence within coding (flagged PUTATIVE_CODING) rather than making it non-coding.
• lncRNA / ncRNA, no coherent ORF; never placed in Tier 1/2.
• transposable_element, only when a TE annotation (--te-annotation) is supplied and the model's TE overlap crosses the threshold; demoted to Tier 4.

The per-gene report

<prefix>.report.tsv has one row per gene with these columns:

gene_id, seqid, start, end, strand, tier, origin, biotype,
primary_transcript_id, num_isoforms, num_as_events, has_cds, protein_id,
max_tpm, junction_support, helixer_support, combined_score, aed, flags

Useful columns to scan:

Column	Tells you
tier, biotype	Confidence and coding class (above).
origin	Where the model came from (Mikado-reconciled vs HELIXER_ONLY backstop).
num_isoforms, num_as_events	Isoform multiplicity and alternative-splicing events recovered.
max_tpm, junction_support	RNA-seq expression and splice-junction backing.
helixer_support, combined_score, aed	Predictor confidence, the Mikado score, and annotation edit distance.
protein_id, has_cds	Homology accession (if any) and whether a CDS was emitted.
flags	QC flags, e.g. PUTATIVE_CODING, TE_OVERLAP, TIER_n.

Inspect and summarise

# Before/after table vs the raw Helixer input
helixforge stats --helixer helixer_output/Zea-mays-subsp-mays-B73_helixer.gff3 \
  --helixforge testing-4_chr1.gff3

# Annotation summary (gene/transcript counts, biotypes, CDS lengths)
helixforge utils summarize --gff3 testing-4_chr1.gff3

# Keep only Tier 1, protein_coding
helixforge utils filter --gff3 testing-4_chr1.gff3 --out tier1.gff3 --max-tier 1

See stats, utils summarize, and utils filter.

Next: the preflight you should run before all of this →