Are we throwing away good data? Evaluation of chimera detection algorithms on long-read amplicons reveals high false-positive rates across algorithms

Long-read amplicon sequencing has enabled us to return to full-length DNA barcodes, which benefit from the higher taxonomic resolution in metabarcoding-based biodiversity studies. However, chimeric sequences (artificial constructs formed when incomplete amplicons fuse during polymerase chain reaction (PCR)) remain challenging, potentially skewing diversity estimates and ecological inferences. Here, we benchmark three <i>de novo</i> chimera detection algorithms, <i>uchime_denovo</i>, <i>removeBimeraDenovo</i>, and <i>chimeras_denovo</i>, on simulated and empirical eukaryotic full-ITS (rRNA ITS1-5.8S-ITS2) datasets to evaluate their precision, sensitivity, and effects on the final OTUs composition/community structure. Upon simulated data, <i>uchime_denovo</i> achieved the highest precision even with default settings, whereas other algorithms displayed high false-positive chimera rates without setting adjustments. Similarly, the tests upon empirical data showed that <i>uchime_denovo</i> had lower false positive rates, whereas about half of the sequences in the putative chimeric batch were false positives when using <i>chimeras_denovo</i> and <i>removeBimeraDenovo</i>. We found that most of the false-negative chimeras contained multiple 5.8S regions, indicating PacBio library preparation artifacts rather than PCR artifacts. However, OTU-level comparisons indicated that overall richness and community-ordination patterns remain largely consistent across different chimera-filtering approaches with or without accounting for false positives and negatives.