The study of the human microbiome serves several important purposes, encompassing a wide range of medical objectives. It can help identify the imbalances associated with various diseases, such as inflammatory bowel disease (IBD), diabetes, obesity, and cardiovascular diseases (Li et al., 2016; Jie et al., 2017; Haneishi et al., 2023). Studying the human microbiome enables the identification of microbial biomarkers for early diagnosis, prognosis, and disease monitoring (Boppana et al., 2024). It can also detect pathogenic microorganisms that may contribute to infections or chronic diseases (Dong et al., 2024). Additionally, microbiome profiles can be used to predict the risk of developing various diseases (He et al., 2024).

The development of sequencing platforms has revolutionized the study of microbial communities. The gold standard for studying the taxonomic composition of a bacterial community is the sequencing of the 16S rRNA gene (Woese and Fox, 1977). 16S rRNA gene is around 1,500 bp long and has 9 variable regions that collect the main evolutionary changes among microbial taxa (Stackebrandt and Goebel, 1994). Compared to whole genome sequencing (WGS), 16S rRNA sequencing is more cost-effective, making it accessible for large-scale studies and routine analysis. Also, the methodologies for 16S rRNA gene amplification, sequencing, and analysis are well-established, providing a robust framework for researchers (Bolyen et al., 2019).

There are extensive public databases (DBs) (e.g., SILVA, Greengenes, RDP) for reference, facilitating accurate taxonomic assignment (Wang et al., 2007; Quast et al., 2013; DeSantis et al., 2006; Cole et al., 2014). While large microbiome reference DBs offer numerous advantages, such as improved resolution and comprehensive taxonomic coverage, they also come with certain disadvantages. Large DBs require significant computational power and memory for searching and aligning sequences. The sheer volume of data in large reference DBs can lead to longer processing times for sequence alignment and classification (Baker, 2010). Large DBs often contain redundant sequences or highly similar entries, which can complicate classification and lead to ambiguities in taxonomic assignments. The likelihood of incorporating erroneous or misannotated sequences increases, which can reduce the accuracy of taxonomic classifications and potentially lead to false conclusions (Sczyrba et al., 2017). Thus, selecting an optimal reference DB is crucial for microbiome studies. An optimal reference DB ensures accurate identification and classification of microbial taxa, reducing the chances of misidentification or ambiguous results, which is essential for understanding the true composition of the microbiome (Monika Balvočiūtė et al., 2017).

The oral and gut microbiome are the two most commonly studied human microbiome. Studying the oral microbiome has several advantages over the gut microbiome. The oral microbiome typically has a lower microbial diversity compared to the gut microbiome (Human Microbiome Project C, 2012). Also, the oral microbiome has been extensively studied, resulting in well-characterized reference DBs such as Human Oral Microbiome Database (HOMD) specifically tailored for oral bacteria, which facilitates more accurate taxonomic assignment (Dewhirst et al., 2010).

For microbiome study, Illumina platform has been widely used. Illumina platforms can sequence millions of reads per run, making it suitable for large-scale studies. The cost of sequencing per base is relatively low, and it provides high accuracy with low error rates. However, typical sequencing read length is rather short (2 × 300 bps), which cannot cover the full-length of the 16S rRNA gene, which could lead to potential misclassification or ambiguous taxonomic assignment (Satam et al., 2023). Pacbio and Nanopore can provide long read sequences to overcome this limitation. Especially, Pacbio system can provide improved sequencing quality with the development of circular consensus sequencing (CCS) protocols which generates highly accurate long high-fidelity reads, also known as HiFi reads (Wenger et al., 2019). Callahan et al. demonstrated that Pacbio HiFi could offer a single-nucleotide resolution by DADA2 approach based on Amplicon Sequence Variant (ASV) classification (Callahan et al., 2019). Thus, we hypothesized that full-length 16S rRNA sequencing data produced by PacBio could be used to optimize reference database in human microbiome studies.

Recently, there have been several studies that simultaneously utilized PacBio and Illumina platform for microbiome study and compared their performance (Buetas et al., 2024; Souza et al., 2023; Katiraei et al., 2022). Especially, She et al. have performed microbiome analysis on 53 sites of 7 surface human organs using both Illumina V3-V4 short read sequencing and Pacbio 16S rRNA full-length sequencing (She et al., 2024). In this study, we tested if full-length 16S rRNA sequencing data produced by Pacbio could be used to serve as a reference DB and compared it with commonly used reference DB (e.g., HOMD) for coverage and classification performance against V3-V4 short read sequencing data. To validate the method, we applied the optimization method to gut microbiome data. Optimized reference DB was constructed with ASVs, and it was compared to SILVA and Greengene reference DB in taxonomy assignment and biomarker discovery against Illumina V3-V4 short read sequencing data.

Sequencing of the 16S rRNA gene is a widely accepted standard for analyzing the taxonomic composition of bacterial communities (Woese and Fox, 1977). Extensive public databases (e.g., SILVA, Greengenes, RDP) facilitate taxonomic assignment. Optimizing reference databases is crucial for human microbiome studies to ensure accurate identification and classification of microbial taxa, thereby reducing the chances of misidentification or ambiguous results (Monika Balvočiūtė et al., 2017). Recent advancements in PacBio technology can generate highly accurate, long high-fidelity reads, offering single-nucleotide resolution (Wenger et al., 2019; Callahan et al., 2019). In this study, we tested whether 16S full-length sequencing data produced by PacBio could be used to construct a reference database and evaluated its application using Illumina V3-V4 targeted short read sequencing data in human microbiome studies.

To evaluate whether PacBio long-read sequencing data could be used to construct a microbiome reference database, we used oral microbiome data for testing. The oral microbiome typically has lower microbial diversity compared to the gut microbiome (Human Microbiome Project C, 2012) and has been extensively studied, resulting in well-characterized reference DBs (18). First, we plotted the number of ASVs obtained from an individual to determine the minimum number of samples required to represent a population. For an individual oral sample, the average ASV count was 247, ranging from 34 to 440. Since combining and resequencing samples was not feasible, we randomly combined samples to simulate mixtures. When 4, 8, 16, and 32 samples were randomly combined, the number of ASVs detected increased gradually with the number of samples in the group (Figure 1B). To assess their efficiency in classifying Illumina data, we constructed a BLAST reference database using the PacBio ASVs obtained from various combinations. A stand-alone BLAST search was performed against Illumina data, and the results were compared against eHOMD to obtain discrete statistics. BLAST operates by aligning query sequences to a database of sequences, identifying regions of similarity using a heuristic algorithm to find high-scoring sequence alignments quickly. It produces a list of sequences in the database that are most similar to the query sequence, along with alignment scores and statistics, including identical nucleotide length and percentage (Altschul et al., 1990; Camacho et al., 2009).

When comparing eHOMD and PacBio_4 (4 samples mixed), which had a similar number of reference counts, the proportion of high identity (>97%) was significantly higher using eHOMD, while the overall positively blasted (>90%) percentage of reads was over 95% for both. Generally, sequence identity of 97, 95, and 90% or less for 16S rRNA genes is considered distinctive for species, genera, and family, respectively (Yarza et al., 2014; Tindall et al., 2010). The combination that showed comparable high identity performance to eHOMD was PacBio_16 (16 samples mixed). Thus, a minimum of 4 samples was sufficient to determine 95% of reads at the family level, while at least 16 samples were required to determine 95% of reads at the species level. Given that PacBio_32 included nearly 6,000 ASVs compared to eHOMD’s 1,032 sequences, we tested whether the eHOMD could be enhanced by adding ASVs from PacBio_32. Specifically, we filtered Illumina reads that showed less than 97% identity against eHOMD but higher than 97% identity against PacBio ASVs. We identified 130 ASVs, and the database created by combining eHOMD with these 130 ASVs (eHOMD+Pac) demonstrated the highest taxonomic assignment performance (Figure 1C). Taken together, with sufficient samples, PacBio full-length sequencing data can be utilized to construct a reference DB from a scratch for oral microbiome study.

To investigate any discrepancies in microbiome coverage between the PacBio DB and eHOMD, a phylogenetic tree was constructed using PacBio_32 OTU sequences combined with eHOMD. Phylogenetic analysis, which can be used for biological classification (de Queiroz and Gauthier, 1994) and predicting characteristics of clonal populations and unstudied species (Pearson et al., 2009), revealed that most of the trees included sequences from both databases, suggesting that both databases cover similar taxa (Figure 2).

Given the substantial size difference between the PacBio_32 ASVs and eHOMD databases, we aimed to optimize the PacBio ASVs. One method to optimize the database is by constructing a phylogenetic tree, trimming closely related branches, and retaining the representative taxa (Mikula, 2018). To find the optimal condition, terminal branches were trimmed at various thresholds. When these trimmed ASVs were used to BLAST Illumina sequencing data, a negative correlation was observed between the threshold and identity outcome. With a trimming threshold of 0.0005, the number of ASVs in the BLAST reference database was reduced by 50%, while the BLAST search performance remained similar to that of the PacBio_32 ASVs (Figure 3F). Taken together, this approach allows for efficient database optimization while maintaining high taxonomic assignment accuracy.

To evaluate whether PacBio ASVs could be applied to other less-studied microbiomes, we tested them against gut microbiome data. The gut microbiome, particularly in the small intestine, presents unique challenges. The microbial community composition in the small intestine differs from that in fecal or oral samples, often containing a higher proportion of fastidious and less well-characterized bacteria, which complicates taxonomic identification (Thadepalli et al., 1979; Villmones et al., 2022). Obtaining samples from the small intestine typically requires invasive procedures such as endoscopy or intubation, which are more complex, costly, and uncomfortable for patients compared to non-invasive fecal or oral sample collection (Booijink et al., 2007). Additionally, the small intestine has a lower microbial biomass compared to the colon, making it more difficult to obtain sufficient microbial DNA for analysis (Hayashi et al., 2005). We constructed optimized gut microbiome reference DB using gut PacBio ASVs.

A pre-trained Naive Bayes classifier was prepared using the Greengene DB, SILVA DB, and gut PacBio ASVs. Gut Illumina V3-V4 paired-reads microbiome data from the ileum, jejunum, and anus were taxonomically assigned by each classifier. Unlike BLAST, the Naive Bayes classifier assigns taxonomy to rRNA sequences by calculating the probability of the sequence belonging to a particular taxon. It is fast and efficient for classifying large numbers of sequences and provides taxonomic assignments with confidence scores, which depend on the quality and comprehensiveness of the training DB (Wang et al., 2007). At the genus level, the overall relative abundance showed similar proportions regardless of the DB used (Figure 5A). At the species level, classifiers trained with Greengene and SILVA DBs assigned more than 50% of the operational taxonomic units (OTUs) as Bacteroides, while the classifier trained with the PacBio DB distinguished most OTUs to specific species. Moreover, some species were only identified by the PacBio DB-trained classifier. Our results support that a well-curated, microbiome-specific DB can improve the reliability of 16S sequencing analyses and taxonomic annotations (Ritari et al., 2015; Sierra et al., 2020). Taken together, an improvement in species assignment was observed when using the PacBio DB across all four gut microbiome samples compared to the other two DBs.

One of the primary purposes of microbiome studies is to discover biomarkers for diseases (Hajjo et al., 2022). Biomarker discovery can provide a deeper understanding of disease mechanisms (Cani, 2018) and can be applied to disease prediction and treatment (Veziant et al., 2021; Marcos-Zambrano et al., 2021). We applied LEfSe to evaluate the biomarker discovery efficiency using classifiers trained with various reference DBs. The choice of reference DB significantly impacted the identification of significant taxa. The PacBio reference DB identified significantly more species compared to the other reference DBs. Although further validation is necessary, having more candidate species increases the likelihood of identifying important taxa.

In addition, recent advancements in the accuracy of sequencing long DNA reads using Nanopore technology, particularly in homopolymer regions, may present a new potential method for preparing microbiome reference DBs (Mantas Sereika et al., 2022).

In conclusion, full-length 16S rRNA sequencing data produced by PacBio can be used to construct an optimized microbiome reference database that demonstrates coverage and efficiency comparable to the well-established HOMD in oral microbiome studies. Applying these optimization methods to gut microbiome data indicated that this approach could be extended to other microbiomes, enhancing the accuracy of microbiome classification and improving biomarker discovery.