GLASS: a graph learning algorithm for screening splice-aware alignments of long-read RNA-seq data

With the continuous development of long-read RNA-seq, obtaining reliable splice-aware alignments remains a major challenge in transcriptomic analysis. Here, we propose GLASS, a graph-learning-based, precision-oriented post-alignment filtering method for long-read RNA-seq data. GLASS constructs a Read–AS bipartite graph to model the relationships between reads and alternative-splicing events, and uses a BipartiteGCN model to identify reads with potentially unreliable or annotation-inconsistent splice patterns. Across multiple datasets and transcript assemblers, GLASS reduced annotation-inconsistent splice signals and improved conservative transcript reconstruction precision, with a modest sensitivity trade-off. Additional analyses using updated assemblers, alternative annotations, minisplice/2passtools-processed alignments and PacBio datasets further supported the robustness of this precision-oriented filtering trend. We also clarify that annotation-inconsistent splice patterns are not necessarily erroneous and may include true but unannotated isoforms. Therefore, GLASS should be viewed as a complementary filtering strategy for conservative transcript reconstruction.