Mycofier: a new machine learning-based classifier for fungal ITS sequences

Fungi represent an essential functional component of Earth’s biodiversity, not only because of their roles as decomposers, mutualists and pathogens, but also because they are the second most speciose eukaryotic kingdom [1, 2]. Several rRNA genes have been used to explore their diversity and used as method for their identification; these include the small ribosomal subunit (SSU), the large ribosomal subunit (LSU) and the internal transcribed spacer (ITS) [3]. The ITS region has a higher PCR amplification success rate compared with other phylogenetic markers such as RPB1, SSU and LSU; also, it has a species discrimination power throughout the entire fungal kingdom and a defined barcode gap. Given these advantages, Schoch et al. [4] proposed ITS as the standard barcode for fungi. ITS includes the ITS1 and ITS2 regions, separated by the 5.8S gene in the nuclear rDNA repeat unit [5]. The entire ITS region has commonly been sequenced with traditional Sanger approaches with a typical amplicon that ranges between 450 and 700 bp. Either the ITS1 or the ITS2 regions have been targeted in recent high-throughput sequencing studies because the entire ITS region is still too long for illumina sequencing, the predominant method today [6–8]. In particular, the ITS1 region has been used recently for fungal phylogeny, taxonomic placement and for environmental surveys [9–11]. ITS1 is a hypervariable region that allows species identification and subgeneric phylogenetic inference [12–14]. However, pairwise alignments are less effective and show a comparative lower efficiency than alignment free methods when it comes to taxonomic assignment of sequences that show a high sequence divergence between their members such as ITS1.

Machine learning-based algorithms have been used as a response to the problems in computational biology such as classification of biological data. Among these tools, the Naïve Bayesian classification method is simple yet can be extremely efficient. This type of classifier is based on the Bayes theorem and “Naïve” refers to the assumption that data attributes are independent from each other. Even when the independency of data attributes is violated, the Bayesian method can still be optimal [15]. A Naïve Bayes classifier assigns an object to a class based on the probability the object has according to its features. In bioinformatics, the Naïve Bayesian classification method has been reported to perform well on problems similar to the classification of sequence data, such as the Ribosomal Database Project (RDP) Classifier [16].

The aim of this work was to develop a machine learning-based classifier for classifying fungal ITS1 sequences according to the NCBI taxonomy at the genus level. Here we explored the use of Naïve Bayes with different features and parameters in order to develop a classifier for ecological studies using high-throughput data.

Accurate sequence classification is a crucial factor to assess fungal community diversity and ecological studies. At present, high-throughput sequencing technologies produce millions of sequences such as ITS1 with no bioinformatics tool to correctly classify such sequences. Supervised machine learning approaches have been very accurate in several bioinformatic prediction methods. This work describes the construction of a fungal ITS1 database based on criteria that would minimize incorrect assignment of taxa, and its use in the development of a machine learning-based classifier.

Besides the fact that reliable reference databases and taxonomies are critical to assign sequence reads to their right phylogenetic affiliation, the development of supervised machine learning classifiers needs curated data for the consequent construction of training and exploration data sets. In the absence of a curated fungal ITS1 database, the first goal was the construction of a curated database of fungal ITS1 sequences. The primary sequence data source (ITS1 sequences from GenBank that recovers Refseq, CBS and nr database sequences) had to be filtered out due to the presence of low-quality sequence data and inaccurate taxonomic information deposited.

The taxonomic and clustering filters were the steps where more sequences were discarded, as shown in Fig. 1. In public databases there are too many fungal ITS1 sequences lacking taxonomic annotation at the genus level and also many environmental sequences without a good annotation [17]. Redundancy was reduced in the clustering step, removing all identical sequences and also the ones that were subsequences of other longer sequences. Nilsson et al. [18] reported that interspecific variability varies throughout the different fungal species and there is not a unifying stringent upper limit for defining all fungal species. In fact, we applied several clustering parameters in order to determine the percentage of identity that would best define natural groups in the fungal kingdom. Our result indicates that there is no sweet spot to determine this, as variation between taxa is as large as variation within taxa (data not shown). In addition to this, clustering even at high percent identity (99 %) led to a dataset with insufficient coverage for many fungal taxa. However, we selected 99 % to avoid loosing taxonomic resolution, while still reducing sequence redundancy.

The reduction of the dataset affected classification accuracy since the number of data per class in learning algorithms is low and can lead to over fitting and suboptimal performances [19]. Indeed, this phenomenon was shown in the classification accuracies of the different models constructed according to the number of sequences allowed for each fungal genus.

Naïve Bayes classifiers have been developed to classify other sequences commonly used for bacterial and fungal barcoding but that cannot be aligned accurately, such as the RDP classifier [16, 20]. The RDP classifier is a Naïve bayes classifier, fast and effective for bacterial and archaeal organisms based on the 16S ribosomal RNA gene. Recently, this classifier has been adapted for the identification of fungal sequences using two markers, the 28S rRNA subunit and the ITS region. For the last one, UNITE and the Warcup ITS training set (sequences retrieved from the UNITE + INSD datasets) are used [21]. Deshpande et al. reported the same problem as we had in classification of fungal ITS sequences due to taxa coverage since several orders could not be represented in the Warcup ITS training set [21].

Mycofier is a classifier that was built to accurately classify fungal ITS1 sequences, choosing the best features based on a curated database of only ITS1 sequences. Indeed Mycofier was developed to specifically classify fungi based on ITS1 sequences.

Although the BLAST approach provided higher classification accuracy, our classifier does not require a pairwise sequence alignment step, which improves speed and lowers computational demands. The accuracy of our classifier is above 87 %. This is an initial effort to develop a machine learning-based classifier for large sequence data sets of hyper variable nature like the ITS1. The ITS1 database and consequently Mycofier are based on the availability of high quality sequences. Development of a classifier with more coverage will be accomplished with the inclusion of underrepresented taxa in the future.

This study reports the Mycofier tool for the classification of fungal ITS1 sequences. This classifier includes a novel and curated training data set built with a set of sequences from specialized and curated databases. Our training set can still be improved by including good-quality, curated sequence data to improve coverage. Therefore, the classification tool coupled to the use of this database provides accurate identification of ITS1 fungal sequences obtained from NGS technologies. The features used for building the Mycofier classifier make it advantageous over BLAST because it does not require and it is not limited to a pairwise comparison between two sequences. Given its probabilistic nature, Mycofier also captures sequence diversity within the model, this is something that is not available in BLAST searches. In addition, our classifier represents an alternative to phylogenetic placement methods such as pplacer [22].