Eukaryotic Gene Prediction - Usage Guide

Overview

Predict protein-coding gene structures in eukaryotic genomes by matching the pipeline to the available evidence: BRAKER3 (RNA-seq + protein), BRAKER1 (RNA-seq), GALBA or BRAKER2 (protein-only), Funannotate (fungi), GeMoMa (homology projection from a close relative), or Helixer/Tiberius (deep-learning ab initio when no evidence exists). The decisions that matter most are upstream of the tool: soft-mask repeats first, decontaminate and check assembly contiguity before training, and pick the right OrthoDB clade partition. Output is gene models in GFF3/GTF with protein and CDS sequences.

Prerequisites

# BRAKER3 (wraps AUGUSTUS + GeneMark-ETP + TSEBRA) - prefer the official container for reproducibility
conda install -c bioconda braker3

# Protein-only and homology routes
conda install -c bioconda galba

# Fungal end-to-end pipeline
conda install -c bioconda funannotate

# Evidence prep + evaluation
conda install -c bioconda hisat2 star samtools busco

# Python parsing
pip install gffutils pandas

Requires a soft-masked genome assembly (run repeat-annotation first). Pin the OrthoDB partition version and the repeat library for reproducibility.

Quick Start

Tell your AI agent what you want to do: - "Predict genes in my eukaryotic genome using RNA-seq and protein evidence" - "Run BRAKER3 on my soft-masked plant genome and check the gene-model sanity metrics" - "I only have protein evidence from a close relative - which pipeline should I use?" - "Annotate my fungal genome with Funannotate including UTRs and isoforms"

Example Prompts

Evidence-Matched Prediction

"Run BRAKER3 on my soft-masked genome with these RNA-seq BAM files and the Vertebrata OrthoDB partition, and report mono-exonic fraction and protein-length distribution."

"I have no RNA-seq and only distant proteins - run BRAKER2 with a broad OrthoDB clade, or Tiberius ab initio since this is a vertebrate."

Protein-Only and Fungi

"I have proteomes from two close relatives - run GALBA on my assembly."

"Annotate my fungal genome with Funannotate, then run the update step to add UTRs and isoforms."

Diagnosing a Poor Annotation

"My annotation has 45,000 genes for a vertebrate - check whether this is haplotigs, unmasked repeats, or split models."

"Compare assembly-BUSCO to proteome-BUSCO to tell whether my predictor missed present genes or the assembly is incomplete."

What the Agent Will Do

Confirm the assembly is soft-masked (repeats in lowercase) and decontaminated, and check contiguity before training
Choose the pipeline from available evidence (BRAKER3 > BRAKER1/2 > GALBA/Funannotate/GeMoMa > DL ab initio)
Prepare evidence (splice-aware RNA-seq alignment; smallest OrthoDB partition containing the clade)
Run the predictor from its container with the correct species/clade parameters
Extract protein and CDS sequences
Run the triage panel: gene count vs relative, mono-exonic fraction, protein-length distribution, mRNA:gene ratio, BUSCO genome-vs-proteome
Flag failure signatures and recommend a PASA/Iso-Seq update step if isoforms/UTRs are needed

Tips

Soft-masking is essential - Hard-masking truncates real repeat-overlapping genes; soft-masking lets a real gene span a repeat. Filter the repeat library against a protein DB so NLR/zinc-finger families survive.
Training quality dominates - A finder trained on a contaminated/fragmented assembly is confidently wrong genome-wide and BUSCO-green. Decontaminate and check assembly BUSCO/N50 first.
OrthoDB clade - Pick the smallest partition that still contains the clade (Vertebrata for a fish, not all Metazoa). GALBA wants a few close-relative proteomes instead.
One isoform, no UTRs by default - If mRNA:gene == 1.00, the annotation is isoform/UTR-naive; alternative-splicing and 3'-tag scRNA-seq analyses against it are untrustworthy. Add a PASA/Iso-Seq update step.
Fungi -> Funannotate - It is the de facto fungal standard (tiny-intron-aware, bundled EVM + PASA update); BRAKER's --fungus flag exists but Funannotate is the smoother fungal path.
High BUSCO-Duplicated - >5-8% with no known WGD means purge_dups the assembly first, not "deduplicate genes" afterward. In a known polyploid, high D is real - confirm via synteny/Ks and keep.
Reproducibility - Run from official containers; the GFF3 alone is not reproducible without the masked genome, evidence, trained parameters, and versions.

genome-annotation/repeat-annotation - Soft-mask repeats before gene prediction
genome-annotation/functional-annotation - Add functional annotation to predicted genes
genome-annotation/annotation-qc - BUSCO genome-vs-proteome, OMArk, gene-set sanity
read-alignment/star-alignment - Splice-aware RNA-seq aligner for evidence
genome-assembly/assembly-qc - Verify assembly quality and purge haplotigs first