ITSxRust : ITS Region extraction with partial‐chain recovery and structured diagnostics for long‐read amplicon sequencing

Abstract Long‐read amplicon sequencing is now routine for fungal metabarcoding, and locating the internal transcribed spacer (ITS) subregions within reads has become a throughput bottleneck. Existing extractors were designed for shorter reads and for datasets in which near‐identical sequences are common, and neither assumption holds for Oxford Nanopore or PacBio HiFi amplicons. Extraction failures are also reported opaquely, leaving users unable to distinguish a primer‐design problem from a parameter one. We present ITSxRust, an ITS extractor written in Rust. It delimits regions with a four‐anchor chain across the SSU 3′ end, both 5.8S boundaries and the LSU 5′ start, and falls back to two‐anchor pairs when a read lacks one or two flanks, so that every extracted region retains two profile‐HMM boundaries rather than a read edge. It emits per‐read confidence labels, structured failure codes and a per‐sample QC summary, and provides Nanopore and HiFi parameter presets. We benchmarked it against ITSx and ITSxpress v2 on two Nanopore datasets amplified with different primer pairs. Extraction rates depended strongly on amplicon design. On an ITS1/ITS4 library, ITSxRust recovered 75.3% of reads as full ITS against 69.9% for ITSx, but 76.3% as ITS1 against 93.8%. On an NS5/LR6 library of 54 taxa the ITS1 difference fell to 0.5 percentage points and the full ITS advantage persisted at 2.4. The shortfall is a property of the primer pair: ITS1/ITS4 leaves a median 30 bp of SSU against a 45 bp profile model, so the anchor cannot be matched in full. Where ITSx reported a higher ITS1 yield, 98.3% of the extra reads were ones it had itself flagged as lacking an SSU boundary, retaining a median 30 bp of SSU. Runtime was within 1% of a bare nhmmer search on both datasets, and 6.2 to 9.0 times faster than ITSx. ITSxRust makes the cause of extraction failure visible and quantifiable, which lets users act on primer design rather than guess. It is available under the MIT licence, packaged for Bioconda and Docker, and integrated into nf‐core/ampliseq.