Most current approach to metagenomic classification employ short next generation sequencing (NGS) reads that are present in metagenomic samples to identify unique genomic regions. NGS reads, however, might not be long enough to differentiate similar genomes. This suggests a potential for using longer reads to improve classification performance. Presently, longer reads tend to have a higher rate of sequencing errors. Thus, given the pros and cons, it remains unclear which types of reads is better for metagenomic classification. We compared two taxonomic classification protocols: a traditional assembly-free protocol and a novel assembly-based protocol. The novel assembly-based protocol consists of assembling short-reads into longer reads, which will be subsequently classified by a traditional taxonomic classifier. We discovered that most classifiers made fewer predictions with longer reads and that they achieved higher classification performance on synthetic metagenomic data. Generally, we observed a significant increase in precision, while having similar recall rates. On real data, we observed similar characteristics that suggest that the classifiers might have similar performance of higher precision with similar recall with longer reads. We have shown a noticeable difference in performance between assembly-based and assembly-free taxonomic classification. This finding strongly suggests that classifying species in metagenomic environments can be achieved with higher overall performance simply by assembling short reads. Further, it also suggests that long-read technologies might be better for species classification.

Quantification and identification of microbial genomes based on next-generation sequencing data is a challenging problem in metagenomics. Although current methods have mostly focused on analyzing bacteria whose genomes have been sequenced, such analyses are, however, complicated by the presence of unknown bacteria or bacteria whose genomes have not been sequence. We propose a method for detecting unknown bacteria in environmental samples. Our approach is unique in its utilization of short reads only from 16S rRNA genes, not from entire genomes. We show that short reads from 16S rRNA genes retain sufficient information for detecting unknown bacteria in oral microbial communities. In our experimentation with bacterial genomes from the Human Oral Microbiome Database, we found that this method made accurate and robust predictions at different read coverages and percentages of unknown bacteria. Advantages of this approach include not only a reduction in experimental and computational costs but also a potentially high accuracy across environmental samples due to the strong conservation of the 16S rRNA gene.

Efforts such as International HapMap Project and 1000 Genomes Project resulted in a catalog of millions of single nucleotides and insertion/deletion (INDEL) variants of the human population. Viewed as a reference of existing variants, this resource commonly serves as a gold standard for studying and developing methods to detect genetic variants. Our analysis revealed that this reference contained thousands of INDELs that were constructed in a biased manner. This bias occurred at the level of aligning short reads to reference genomes to detect variants. The bias is caused by the existence of many theoretically optimal alignments between the reference genome and reads containing alternative alleles at those INDEL locations. We examined several popular aligners and showed that these aligners could be divided into groups whose alignments yielded INDELs that agreed strongly or disagreed strongly with reported INDELs. This finding suggests that the agreement or disagreement between the aligners’ called INDEL and the reported INDEL is merely a result of the arbitrary selection of one of the optimal alignments. The existence of bias in INDEL calling might have a serious influence in downstream analyses. As such, our finding suggests that this phenomenon should be further addressed.

Background: Most current metagenomic classifiers and profilers employ short reads to classify, bin and profile microbial genomes that are present in metagenomic samples. Many of these methods adopt techniques that aim to identify unique genomic regions of genomes so as to differentiate them. Because of this, short-read lengths might be suboptimal. Longer read lengths might improve the performance of classification and profiling. However, longer reads produced by current technology tend to have a higher rate of sequencing errors, compared to short reads. It is not clear if the trade-off between longer length versus higher sequencing errors will increase or decrease classification and profiling performance.Results: We compared performance of popular metagenomic classifiers on short reads and longer reads, which are assembled from the same short reads. When using a number of popular assemblers to assemble long reads from the short reads, we discovered that most classifiers made fewer predictions with longer reads and that they achieved higher classification performance on synthetic metagenomic data. Specifically, across most classifiers, we observed a significant increase in precision, while recall remained the same, resulting in higher overall classification performance. On real metagenomic data, we observed a similar trend that classifiers made fewer predictions. This suggested that they might have the same performance characteristics of having higher precision while maintaining the same recall with longer reads.Conclusions: This finding has two main implications. First, it suggests that classifying species in metagenomic environments can be achieved with higher overall performance simply by assembling short reads. This suggested that they might have the same performance characteristics of having higher precision while maintaining the same recall as shorter reads. Second, this finding suggests that it might be a good idea to consider utilizing long-read technologies in species classification for metagenomic applications. Current long-read technologies tend to have higher sequencing errors and are more expensive compared to short-read technologies. The trade-offs between the pros and cons should be investigated.