Tiara: Deep learning-based classification system for eukaryotic sequences

Abstract

Motivation: With a large number of metagenomic datasets becoming available, the eukaryotic metagenomics emerged as a new challenge. The proper classification of eukaryotic nuclear and organellar genomes is an essential step towards the better understanding of eukaryotic diversity. Results: We developed Tiara, a deep-learning-based approach for identification of eukaryotic sequences in the metagenomic data sets. Its two-step classification process enables the classification of nuclear and organellar eukaryotic fractions and subsequently divides organellar sequences to plastidial and mitochondrial. Using test dataset, we have shown that Tiara performs similarly to EukRep for prokaryotes classification and outperformed it for eukaryotes classification with lower calculation time. Tiara is also the only available tool correctly classifying organellar sequences. Availability and implementation: Tiara is implemented in python 3.8, available at https://github.com/ibe-uw/tiara and tested on Unix-based systems. It is released under an open-source MIT license and documentation is available at https://ibe-uw.github.io/tiara. Version 1.0.1 of Tiara has been used for all benchmarks.

Eukfinder: a pipeline to retrieve microbial eukaryote genomes from metagenomic sequencing data

“…Due to differences in input data for the methods examined, Eukfinder_short, and Refmapping (7) were used to examine short Illumina reads, while Eukfinder_long, EukRep (8) and Tiara (10) were used to examine assembled contigs of these reads. To ensure the comparisons were fair, we first identified the parameters for each program that maximized genome completeness and contiguity.…”

“…We compared the recovered genomes using EukRep with all three stringency cut-off modes and one analysis under lenient mode with the '-tie' set to prokaryotic. Tiara (10) utilizes previously trained neural networks to sort metagenomic reads into multiple categories, representing the taxonomic affiliation of each read within a given dataset. Tiara has two main parameters which can be adjusted in its workflow; the k-mer size, which adjusts the size of the DNA substrings used to compute sequence frequency, and the probability threshold, which disregards results with a probability score lower than that.…”

“…We examined Eukfinder (short and long) along with three other methods, Refmapping (7), EukRep (8) and Tiara (10), for their ability to recover the Blastocystis genome from the mock communities and human gut metagenome samples. QUAST v5.0.2 (34) was used to assess nuclear genome completeness measured by total genome length and fraction of Blastocystis genome recovered.…”

“…A recent study applied EukRep to 1198 metagenomic datasets and in total 14 novel eukaryotic genomes were recovered with estimated median completeness of 91% (9). Tiara, which uses a deep learning approach (a subfamily of machine-learning techniques), was introduced to identify eukaryotic and organellar genomes within metagenomic datasets (10). These studies demonstrate that it is possible to reconstruct microbial eukaryotic genomes from metagenomics datasets.…”

The high-throughput gene prediction of more than 1,700 eukaryote genomes using the software package EukMetaSanity

“…Taxonomy prediction using the MMseqs2 taxonomy subprogram for the MetaEukderived proteins identified 44% as belonging to the Phylum Ciliophora and the MAG had a relatively small proportion (0.33%) of interspersed repeat elements (n = 322). A cursory analysis using Tiara (Karlicki et al, 2021) revealed that the 13.5 Mbp MAG in question consisted of DNA sequences whose origins were 28.8% eukaryotic, 28.3% bacterial, 10.2% archaeal, 7.3% prokaryotic, and 25.4% unknown. When protein prediction was performed using Prodigal v2.6.3 (-p meta) (Hyatt et al, 2012), a tool for prokaryotic gene prediction, the number of recovered putative coding sequences increased from 1,943 to 9,020, suggesting that this particular MAG represented a binning error that combined genomic content across Domains.…”