IsoPrism predicts isoform usage from short-read transcriptomes through structure-guided competitive allocation
Year: 2026
Type: article
Abstract: Short-read (SR) single-cell and spatial transcriptomics characterize gene expression but provide limited resolution of within-gene isoform usage. Here we present IsoPrism, a structure-guided competitive allocation framework that predicts isoform proportions from gene-level expression. IsoPrism combines individual expression and local expression context with transcript structural annotations, using shared iterative updates to preserve non-negative, gene-normalized proportions. Joint supervision from individual long-read (LR) counts and group-level isoform usage enables prediction in samples without matched LR measurements. Across five transfer scenarios derived from three paired transcriptomic resources, IsoPrism achieved the highest mean dominant isoform accuracy, compositional similarity, and cluster-level allocation correlation among eight evaluated methods. Ablation analyses supported complementary contributions from structural queries, expression context, iterative allocation, and dual-resolution supervision. Applied to a developing mouse cortex atlas, IsoPrism identified stage- and cell-type-associated isoform-usage patterns, including variation in the predicted relative usage of two Pkm transcripts. IsoPrism enables isoform-usage analysis of SR transcriptomes by transferring information from paired LR references, extending the analytical value of existing single-cell and spatial gene-expression atlases.
Source: Briefings in Bioinformatics
DOI: 10.1093/bib/bbag621
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $3,885
Award: G5313112709
OpenAlex ID: W7225092830