Isocall enables scalable transcript identification from long-read RNA-sequencing data

Abstract

Long-read RNA sequencing directly resolves the full structures of RNA transcripts. Advances in throughput now enable the generation of deeply sequenced cohorts of hundreds of samples, making joint transcript discovery across large datasets possible. However, existing transcript identification methods were designed for small datasets, which limits their applicability at this scale. Here, we present Isocall, a scalable and deterministic computational method for jointly calling transcripts from multiple PacBio long-read RNA sequencing samples. Isocall converts aligned full-length non-concatemer reads into compact per-sample transcript profiles, merges these profiles across samples, and jointly identifies known and novel transcripts supported by reads in the analyzed dataset. Filtering is tunable as presets provide coarse control and individual parameters, including relative abundance and internal priming thresholds, provide fine control. Isocall demonstrated high precision in our accuracy benchmarks, including the 69 transcript WTC11 SIRV spike-in controls, for which Isocall reported 0-2 false-positive transcripts per sample across the three SIRV mixes at default settings. To demonstrate scalability, we applied Isocall to 206 samples, totalling 3.5 billion raw reads, from the Human Pangenome Reference Consortium. After parallelized pbmm2 alignment and Isocall profile, the call step performed joint calling across the entire dataset in 25 minutes, using 1.3 GB of peak memory and 8 threads. Finally, in Genome in a Bottle samples with matched SNP genotypes, splice-site polymorphisms provide an additional measure of call accuracy: Isocall recovered 337 polymorphic splice sites, including a de novo donor site in BTN3A1 that corresponds to a complete isoform switch on the mutant allele.