CDScompare: a frame-sensitive GFF comparison tool for improved detection of coding sequence discrepancies, with applications to wheat and rice annotations

The rapid expansion of high-quality plant genome assemblies, driven by advances in sequencing technologies and cost reductions, has led to a proliferation of genome assemblies. Moreover, for a single genome, multiple annotation sources often coexist: full-genome predictions from dedicated pipelines (sometimes split into high- and low-confidence sets), manually curated annotations focused on particular gene families, and independent annotations from public repositories such as NCBI. Accurately comparing these alternative annotations is essential to identify and resolve discrepancies, either by selecting the best gene model or, at minimum, by flagging conflicting gene structures at the same locus. Existing comparison tools can detect differences such as gene overlap, fusion or splitting, and some can also detect differences in coding/non-coding boundaries. However, when it comes to coding sequences, none of these tools take the reading frame into account despite its crucial role in interpreting protein-coding regions. This limits their ability to detect functionally relevant annotation conflicts, since the same nucleotide segment, when read in a different frame, yields a completely different protein.

We present CDScompare, a Python-based tool that performs frame-aware comparison of GFF annotation files with a focus on CDS regions. This is valuable not only for researchers involved in genome annotation, but also for the many scientists working with gene families of interest, who often face conflicting annotations across multiple sources. As case studies, we applied CDScompare to compare (i) two public annotations of the hexaploid bread wheat (Triticum aestivum), and (ii) two annotations of rice (Oryza sativa) cv. Azucena. CDScompare revealed many CDS conflicts overlooked by existing tools, including cases where overlapping or identical CDS boundaries result in completely different protein sequences due to reading frame differences. Despite the size of the wheat GFF input files (> 1 GB in total), CDScompare processed the comparison in under one minute on a standard laptop, thanks to several algorithmic optimizations.

CDScompare introduces a necessary layer of precision in structural annotation comparison by incorporating reading-frame awareness. It enables more informed decisions for annotation pipeline comparison and supports downstream analyses where functional annotation integrity is critical. CDScompare is open source and freely available.