The enormous microbial diversity prevalent in natural ecosystems represents a rich resource for discovery of hitherto unknown microbes and the novel genes/proteins they encompass. Estimates reveal that 99% of these microbes cannot be easily cultured in the laboratory [1]. The rapidly growing field of metagenomics directly investigates this microbial diversity by obtaining and sequencing the entire genomic content present in any given environmental sample. Since environmental samples contain hundreds of microbes, the sequencing step typically generates millions of sequenced fragments originating from the genomes of various microbes. Using computational methods, these sequences are subsequently analysed to identify new organisms, genes, proteins, and metabolic pathways.

An important aspect in metagenomic analysis is the identification of the source organism of each fragment (read/contig). This process, called binning, helps in identifying and enumerating the various taxa present in a given metagenomic sample. Binning approaches can be broadly classified into two categories, namely, composition-based and similarity-based. Composition-based approaches cluster and classify sequences based on their compositional characteristics, such as GC percent, codon usage, oligo-nucleotide frequency distribution, etc. [2, 3]. On the other hand, similarity-based approaches identify sequence(s) in reference databases, which display significant similarity with the read sequence [4, 5].

In similarity-based approaches, sequences in the database having significant similarity to the input reads are referred to as hits. While a read with hit(s) originating from a single organism is assigned to the organism corresponding to the hit(s), a read with hits originating from multiple organisms is assigned to the L owest C ommon A ncestor (LCA) of the taxa which correspond to these organisms. The specificity of assignment using the LCA approach is thus dependent on the spatial distribution of hits in the phylogenetic tree.

To improve the specificity of assignment of reads, similarity-based methods like MEGAN use bit-score of an alignment to identify a subset of hits which are significant, and subsequently assign the read to the LCA of this subset [4]. However, the MEGAN approach was seen to have a significantly high false positive rate and low specificity of assignments, especially in scenarios where reads originated from new organisms [5]. These limitations have been addressed by SOrt-ITEMS algorithm, which adopts a two-phase binning approach [5]. During the first phase, additional alignment parameters (identities, positives and the bit-score) are used by SOrt-ITEMS to ascertain the quality of the hits obtained for a read. Reads having hits with high quality alignments are allowed to be assigned at specific taxonomic levels such as species, genus or family. In contrast, the assignment of reads which generate alignments of lower quality is restricted to relatively higher taxonomic levels. Thus, the better the quality of the alignment, the more specific is the taxonomic level of assignment, and vice-versa. Once this level is identified, SOrt-ITEMS employs the 'orthology step' in the second phase to finally assign the read to the LCA of the taxa which correspond to the subset of hits that share a true orthologous relationship with the read sequence. While it is seen that the first phase of SOrt-ITEMS helps in significantly reducing the number of false positive assignments, the second phase ensures the specificity of assignments i.e. assignment of reads at relatively lower taxonomic levels (family, genus or species).

In spite of having significantly higher binning specificity and accuracy, the SOrt-ITEMS approach has the following two limitations. Firstly, since SOrt-ITEMS involves performing a reciprocal BLAST search, an additional time is spent (for every read) in this 'orthology step'. Summing this additional time for all reads in a typical metagenomic data set (having approximately 1-10 million sequences) results in a significant increase in the total computation time. Secondly, about 7-10% of input sequences still get misclassified by SOrt-ITEMS. This misclassification rate is significant since this will result in several thousands of sequences being classified wrongly. For example, if one analyzes the Sargasso sea metagenome data set consisting of greater than 7 million sequences [6], approximately 300,000 to 700,000 sequences are likely to be assigned to incorrect taxa. These wrong assignments will definitely impact the accuracy of several downstream analyses.

One way of reducing the number of misclassified sequences is by increasing the threshold of the 'minimum bin-size' parameter (i.e the minimum number of sequences to be assigned to a taxon for a bin to be created for that respective taxon). In this process, assignments to isolated taxa (i.e. taxa to which the number of sequences assigned is less than the minimum bin-size threshold) are discarded. Increasing the value of this threshold parameter, will thus reduce the number of false positive assignments. However this strategy will result in an increased number of unassigned sequences.

In this paper, we propose a new approach termed as DiScRIBinATE (Di stance Sc ore R atio for I mproved Bin ning A nd T axonomic E stimation) which attempts to address the limitations associated with SOrt-ITEMS. To maintain the accuracy of assignments, DiScRIBinATE retains the steps followed in the first phase of SOrt-ITEMS i.e. for finding an appropriate taxonomic level where the assignment of the read is to be restricted.

However, to ensure the specificity of assignments, DiScRIBinATE circumvents the time consuming 'orthology' step of SOrt-ITEMS by utilizing a quicker 'alternative approach' based on the ratio of bit-score and distance information obtained from the hits corresponding to a read. We demonstrate that the proposed 'alternative approach' significantly decreases the time taken for binning by half. Besides, a novel reclassification strategy incorporated in DiScRIBinATE reduces the misclassification rate to around 3 - 7%. This misclassification rate is around 1.5 - 3 times lower than that by SOrt-ITEMS and 3 - 30 times lower as compared to MEGAN. We also demonstrate that adopting this novel strategy does not increase the percentage of unassigned reads.

Evaluation of similarity-based binning approaches (using simulated metagenomic data sets) have indicated that a significant percentage of reads get misclassified or are categorized as unassigned [4, 5]. This is expected since a majority of reads in metagenomic data sets originate from new or partially characterized genomes, the sequences of which are absent in existing reference databases. Consequently, a majority of metagenomic sequences generate poor alignments with sequences in reference databases. The SOrt-ITEMS algorithm could accurately assign reads having poor alignment quality [5]. The SOrt-ITEMS algorithm ensured accuracy by adopting an elaborate work-flow for judging alignment quality before taxonomic assignment. In spite of this elaborate work-flow, validation studies indicated that a sizeable fraction of reads (7-10%) still get misclassified. This is due to the fact that some reads generate stray, yet significant hits (i.e with good alignment quality), with incorrect (and generally isolated) taxa present in reference databases. It is due to this reason that even an algorithm like SOrt-ITEMS (which takes in account the alignment quality before taxonomic assignment) assigns such reads to incorrect taxa. The latter reads (assigned to incorrect taxa) are manifested in the output as isolated assignments. One way of reducing the number of misclassified sequences is by discarding such isolated assignments. However, adopting this strategy will results in increased number of sequences categorized as unassigned.

In the current study, a novel reclassification strategy has been devised that attempts to reclassify assignments to isolated taxa instead of simply discarding them. The premise of this strategy is the following. The SOrt-ITEMS algorithm [5] revealed that over-binning (incorrect assignments to taxa that are related to the source taxon only at a higher taxonomic level) was one of the primary causes of misclassification.

Hence, it is likely that most of the assignments to isolated taxa could also be a result of over-binning. Therefore to improve binning accuracy, the method presented in this paper attempts to re-classify assignments to isolated taxa at progressively higher taxonomic levels, rather than simply discarding them. The results obtained in this study have revealed that adopting this strategy significantly reduces the percentage of wrong assignments without increasing the percentage of unassigned reads.

The method described in this paper also incorporates a quicker 'alternative approach' (as compared to the time consuming orthology approach used by SOrt-ITEMS) based on the ratio of bit-score and distance information. The use of such a ratio (for improving specificity) is based on the following premise. Bit-score is one of the parameters in the BLAST output which reflects the overall quality of an alignment. Higher bit-scores indicate higher similarity between the query and the hit sequences. Furthermore, higher similarity also indicates that the organism corresponding to the query and hit sequences are taxonomically close to each other (i.e. the phylogenetic distance between them is low). Hence a ratio of bit-score/distance is directly indicative of the taxonomic relatedness.

DiScRIBinATE uses the above principle to improve the specificity of assignments. For each query sequence, DiScRIBinATE identifies a set of candidate ancestors (from amongst the set of significant hits) and assigns the query to a candidate ancestor having the highest bit-score/distance ratio. Using this approach ensures that the query sequence is assigned to a taxon that is taxonomically closest to it (thereby improving specificity). The specificity of assignments using this alternate approach is seen to be comparable to those obtained using the orthology based approach. Figure 5 illustrates an example which demonstrates the process by which this alternative approach helps in improving the specificity of assignments. It is interesting to note that a similar approach had been used earlier for the identification of horizontal gene transfer in metagenomic samples [9].

Results indicate that, as compared to the orthology based approach, this alternative 'bit-score/distance approach', coupled with the reclassification step reduces the binning time by half and also significantly reduces the percentage of wrong assignments. As demonstrated in the validation results, the scale of reduction is as high as 30 times.

The three taxonomic binning methods described in this study use the output of a BLAST search as input. However, it is important to note that the time spent on the taxonomic assignment from BLAST results is really a tiny proportion as compared to the time spent on BLAST search. Besides developing efficient methods that can rapidly and accurately derive taxonomic inferences from BLAST outputs, it is also necessary to develop approaches that can reduce the time spent on performing the actual BLAST search. Our current research focus is on development of novel methods that reduce the computational time associated with typical BLAST searches.

A significant reduction in binning time, coupled with a superior assignment accuracy (as compared to existing binning methods), indicates the immense applicability of the proposed algorithm in rapidly mapping the taxonomic diversity of large metagenomic samples with high accuracy and specificity.