GSAman: A desktop "magic brush" for perfecting gene structure annotations

A genome sequence is only half a genome. The other half is knowing where its genes sit and what shape they take — and that half is still finished largely by hand.

Sequence quality has improved far faster than annotation quality. Long-read assembly now closes most gaps, yet gene models still carry errors that only surface when someone tries to use them: genes that overlap, alternative splicing variants, isoforms expressed at barely detectable levels. The mistakes travel — into expression studies, cross-species comparisons, crop breeding decisions.

Human judgment remains the best fix. Yet the software for it has thinned to a single survivor: Apollo2, which runs on a web server, bringing deployment chores, dependence on network stability, and limited access to local files and computing power.

GSAman, from a team led by Rui Xia of South China Agricultural University, Ling-Ling Chen of Yazhou Bay National Laboratory and Chengjie Chen of the Chinese Academy of Tropical Agricultural Sciences, takes a different route. Built on IGV — the genome browser many biologists already use — it runs entirely on a laptop: download, install, start editing.

Users drag exons, adjust boundaries, add or remove transcripts and undo freely, while the tool flags in real time whether RNA-seq reads back up the model on screen. Where evidence is thin, built-in prediction and homology modules can rebuild a model from scratch, and companion plugins let a team split a genome, curate in parallel, then merge the pieces.

The case study is Minghui 63, an elite indica rice. The team refined 11,496 gene models and improved structures on every chromosome; BUSCO completeness — the standard measure of how much of a genome's expected gene content is captured correctly — rose from 86.93% to 99.63%, above the Nipponbare reference. The same workflow has since been applied to pineapple, pig, fish, shrimp and crab.

The authors are candid about scope: this is one genome, and curation still demands expertise and practice; improvement is scored by annotation benchmarks, not yet by downstream biology.

GSAman stands for Gene Structure Annotation Manual Curator; the cover art nods to Ma Liang's magic brush, the legend of a boy whose painted creatures came alive. The authors expect two converging trends — language models assisting curators, and a growing stock of expert-curated annotations training better automated predictors — to make hand-finishing a routine step rather than a specialist craft.