Our updated study is published! We evaluated methods for taxonomic profiling and classification with long-read #metagenomic datasets. This included @PacBio and @nanopore data, and was a group effort with @saltyscientist and @ctitusbrown: bmcbioinformatics.biomedcentral.com/articles/10.11… 🧵1/8
Top performers include sourmash, @bugseq, and DIAMOND + MEGAN-LR, which displayed overall high precision and solid recall. We recommend trying all of these methods, and each has their own specific advantages. 2/8
Sourmash was magical with HiFi data, having the best precision/recall tradeoff with detection down to 0.001% relative abundance (other methods hit 0.05%). Sourmash works best with highly accurate reads - regardless of read length (!), including @PacBio and @illumina data 🤔. 3/8 
Methods designed for short reads (Kraken2, Bracken, Centrifuge) produced LOTS of false positives 😬. Filtering helped but severely dropped recall. The final balance between precision and recall was never better than results from suitable long-read methods. 4/8
When @PacBio or @nanopore data were analyzed with ANY suitable long-read method, the results were generally better than @illumina data plus a short-read method. There are clear advantages of using long reads to build taxonomic profiles 🙌. 5/8
Within long-read datasets, beware of shorter reads (< 2kb)! They negatively impact taxonomic classification. These shorter reads generally lack multiple genes, which are needed to boost long-read method performance. Remove these during library prep or before analysis! 6/8
Long read accuracy matters. Methods incorporating protein prediction or exact kmer matching simply performed better with @PacBio #HiFi data (median Q40) 🔥. 7/8
Finally, this was intended to be a fully open-access study. The publication (BMC Bioinformatics), sequencing datasets (NCBI, ENA), and analysis outputs (OSF: osf.io/bqtdu/) are all publicly available. 8/8
• • •
Missing some Tweet in this thread? You can try to
force a refresh
