Thread by @DPortik

What is the best method for taxonomic profiling in long-read shotgun #metagenomic datasets? We put several to the test in a new study, using @PacBio #HiFi and @nanopore data!!! Awesome collaboration with @ctitusbrown and @saltyscientist: biorxiv.org/content/10.110…. 🧵1/8

Top performers are long read methods including @bugseq and MEGAN-LR (using DIAMOND to NCBI nr), which had very high precision and solid recall (with no filtering necessary!). MMseqs2 and MetaMaps required some filtering to reduce false positives, but they also performed well. 2/8

Using methods designed for short reads (Kraken2, Bracken, Centrifuge) produced LOTS of false positives. This required heavy filtering to improve precision (simultaneously dropping recall), but they also produced inaccurate abundance estimates. 3/8

These suboptimal results for the short-read profiling methods occurred with both @illumina and long-read datasets for a given mock community, indicating the limitation is probably with the methods themselves and not the read type. 4/8

When @PacBio or @nanopore data were analyzed with ANY long-read profiler, the results were better than @illumina data plus a short-read profiler. There are clear advantages of using long-read shotgun metagenomics for taxonomic profiling (in addition to assembly and function). 5/8

Within long-read datasets, including shorter reads (< 2kb) has a negative effect on taxonomic profiling. These shorter reads are in the gray zone: too long for short-read profilers but not long enough for long-read profilers. Remove these shorter reads before analysis! 6/8

Long read accuracy still matters. Accuracy directly impacts the success of methods relying on protein predictions or exact kmer matches, and these methods performed better with @PacBio #HiFi data (median Q40) versus @nanopore data (including the new "Q20" chemistry). 7/8

Lastly, this study was intended to be highly transparent. All sequencing datasets are publicly available (NCBI, ENA). The outputs from every profiling method, along with Jupyter notebooks to reproduce summary analyses, are freely available on OSF: osf.io/bqtdu/.