HelixForge

Helixer is a deep-learning ab initio gene finder that works across species without retraining, but it emits exactly one transcript per gene and carries no evidence support. HelixForge keeps that gene set as a stable anchor, reconciles it against RNA-seq assemblies, splice junctions, and protein alignments through Mikado, and writes a tiered, isoform-aware, ID-stable annotation.

The reconciliation principle

The predictor is the backbone. Evidence refines a model only where it demonstrably improves it, and nothing the predictor gets right is discarded.

RNA-seq and protein evidence are scored against each Helixer model; a model changes only when the evidence outscores the original. That single rule is what separates HelixForge from a replace-everything re-annotation.

Reconciles, does not replace

Evidence is scored against each Helixer model; a model changes only when the evidence outscores the original.

Adds the missing isoforms

Helixer is structurally capped at one transcript per gene. HelixForge reconstructs alternatives where transcript evidence supports them.

Uses the predictor's confidence

Helixer per-base confidence is injected as an external metric, targeting correction where the predictor is least certain.

Rescues silent genes

Predicted genes without RNA-seq support are preserved when they carry a valid ORF, so recall is not traded for precision.

Tiered output

Every gene is labeled by evidence tier, so downstream users filter by confidence rather than trusting a flat annotation.

Scales without retraining

No species-specific training, no manual curation. Scales across HPC via genome chunking.

How it works

The pipeline runs in four stages, prep (evidence preparation), Mikado (reconciliation engine), reconcile (model selection, backstop rescue, tiering), and output (validated GFF3 plus reports):

Helixer prediction  ─┐
  + per-base confidence
                     │
RNA-seq              ├──▶  Mikado-anchored reconciliation  ──▶  tiered, isoform-aware
  StringTie + junctions                                          evidence-backed GFF3
                     │     score before modifying
Protein homology     ┘     keep a change only if it outscores
  miniprot alignments      call isoforms where evidence supports
                           confidence-gated splice correction
                           rescue silent genes with valid ORFs

All CDS are stop-inclusive, following the Helixer and Ensembl convention. A gene whose CDS cannot be made structurally coherent is flagged and emitted without a CDS rather than repaired heuristically.

Start here

Install HelixForge pip + extras, the external-tool list, and the doctor preflight.

Maize (B73) tutorial A complete end-to-end walkthrough in the order data actually flows.

CLI reference Every command and subcommand, generated from the live CLI.

Benchmarks Maize B73 vs the NAM v5 reference annotation.