Gene-Miner: an evidence-integrating Nextflow pipeline for reference re-annotation and protein-coding gene discovery

The principal bottleneck in genomics has shifted from assembly to annotation. Reference catalogues reach near-complete gene sets only through billions of reads and years of curation, which most genomes lack. No reference annotation is final, so a reproducible pipeline that re-derives a protein-coding catalogue from modest public data serves both unannotated genomes and audits of existing ones. Gene-Miner, a portable Nextflow workflow, takes the union of RNA-seq transcript reconstruction, two ab initio finders (AUGUSTUS and the evidence-trained GeneMark-ETP) and protein homology, then removes contaminant- and transposable-element-derived models. On identical inputs it matched or exceeded BRAKER3 in conserved-ortholog completeness on all four genomes, while BRAKER3 reproduced reference exon structures more closely — a measure that ranks agreement with the reference rather than correctness. Gene-Miner instead recovered more of every reference annotation and proposed more loci the references lack: on rice, 95.1% BUSCO completeness (poales_odb10), 69% of RAP-DB genes recovered and 16,542 loci RAP-DB does not contain, from four public libraries (118 million read pairs, 2.7% of the reads behind the RefSeq annotation) and Swiss-Prot. Completeness was 92.0% on soybean, 98.8% on Drosophila melanogaster and 97.6% on Caenorhabditis elegans , with 66–88% of each reference annotation recovered. Each locus is released in the GFF3 with its evidence flags and one of three confidence tiers, set by how many of RNA-seq, homology and orthology support it. Gene-Miner extends and audits an existing annotation rather than reproducing it. Curated reference databases remain the authority, and a consensus pipeline such as BRAKER3 stays closer to a reference annotation than a detection-first union does. In 3 h per animal genome, 6 h for rice and 22 h for soybean on one 24-core workstation, it produces an isoform-aware catalogue near-complete in conserved orthologs and candidate model updates for expert review: structural revisions, novel loci and RNA-seq-supported isoforms, each with the evidence behind it.