Evolutionary and methodological considerations when interpreting gene presence–absence variation in pangenomes

While graph-based pangenomes have become a standard and interoperable foundation for comparisons across multiple reference genomes, integrating protein-coding gene annotations across pangenomes in a single 'pangene set' remains challenging, both because of methodological inconsistency and biological presence-absence variation (PAV). Here, we review and experimentally evaluate the root of genome annotation and pangene set inconsistency using two polyploid plant pangenomes: cotton and soybean, which were chosen because of their existing diverse high-quality genomic resources and the known importance of gene PAV in their respective breeding programs. We first demonstrate that building pangene sets across different genome resources is highly error prone: PAV calculated directly from the genome annotations hosted on public repositories recapitulates structure in annotation methods and not biological sequence differences. Re-annotation of all genomes with a single identical pipeline largely resolves the broadest stroke issues; however, substantial challenges remain, including a surprisingly common case where exactly identical sequences have different gene model structural annotations. Combined, these results clearly show that pangenome gene model annotations must be carefully integrated before any biological inference can be made regarding sequence evolution, gene copy-number, or PAV.