Summary
A bioRxiv preprint describes Isocall, a deterministic method for jointly identifying known and novel transcripts across large PacBio long-read RNA-sequencing datasets. The authors report a joint call across 206 samples in 25 minutes after alignment and profiling.
A bioRxiv preprint posted on September 13, 2026 describes Isocall, a computational method for identifying RNA transcripts jointly across many long-read sequencing samples. The authors report that the method can process a 206-sample dataset containing 3.5 billion raw reads, with the joint transcript-calling step taking 25 minutes, using 1.3 GB of peak memory and eight threads after alignment and profile generation.
The work addresses a scaling problem in transcript discovery. Long-read RNA sequencing can read full transcript structures, including the way coding and non-coding segments are joined. This makes it possible to distinguish different transcript forms, or isoforms, produced from the same gene. As sequencing throughput grows, however, analysing hundreds of samples together creates a larger computational task than the small datasets targeted by many existing transcript-identification approaches.
How Isocall processes long-read RNA data
Isocall takes aligned, full-length non-concatemer reads and converts them into compact transcript profiles for each sample. It then merges those profiles and jointly identifies known and novel transcripts supported by reads in the analysed dataset.
The method is deterministic, meaning its calling procedure is designed to produce consistent results from the same inputs and settings. Its filtering can be adjusted at two levels: presets provide broad control, while individual parameters allow finer control over factors such as relative abundance and internal priming. These settings determine which candidate transcript structures are retained.
Joint calling is useful because transcript discovery is performed across the dataset rather than independently on every sample. That creates a common set of transcript structures that can be examined across a cohort, while still allowing the method to identify transcripts supported by reads in the data being analysed.
Accuracy and performance reported in the preprint
In accuracy benchmarks using WTC11 samples with SIRV spike-in controls, the authors report high precision. At the default settings, Isocall produced between zero and two false-positive transcripts per sample across the three SIRV mixes.
The authors then tested scalability with 206 samples from the Human Pangenome Reference Consortium, totalling 3.5 billion raw reads. After parallelised pbmm2 alignment and Isocall profile generation, the joint call across the complete dataset took 25 minutes, reached 1.3 GB of peak memory use and used eight threads. The reported timing applies to the call step, rather than necessarily to every stage from raw reads onward.
A separate analysis used Genome in a Bottle samples with matched SNP genotypes. Genetic variants at splice sites provide a way to check whether transcript calls reflect allele-specific changes in RNA processing. Isocall recovered 337 polymorphic splice sites in this analysis. One example was a de novo donor site in BTN3A1, associated with a complete isoform switch on the mutant allele.
The evidence is a computational methods evaluation in a preprint, using benchmark controls and human sequencing datasets. The reported results show how Isocall performed on the tested data and settings; broader use will depend on evaluation across additional datasets and workflows. The competing-interest statement notes that several authors are PacBio employees and shareholders, while one author receives research support from Illumina, PacBio and Oxford Nanopore.