The Qiime2 computational ecosystem, a suite of analysis toolkits integrating methods from various sources, is commonly used to process and analyze microbiome sequencing data 16. Starting from raw sequencing reads produced by sequencing machines (either Illumina or PacBio), the first step of the computational pipeline is to group reads with similar sequences together. There are two algorithms to achieve this: denoising and clustering, which generate amplicon sequence variants (ASVs) and operational taxonomic units (OTUs), respectively. Denoising algorithms, such as DADA2 17, monitor the error rate introduced during the sequencing process and recognize reads originating from the exact same sequence, making them highly reproducible and precise. However, in some cases, ASVs may be too granular for comparing sequences across different samples. Therefore, studies have used clustering algorithms like Vsearch 18 to group similar ASVs into OTUs at an arbitrary cutoff, such as 97% for species-level resolution. For each OTU, one representative sequence is chosen and compared against known databases such as SILVA 19 and Greengenes 20, to assign taxonomy (i.e., species identification) to the corresponding OTU. The number of reads is then counted for each OTU in each sample to construct a feature table, which is the core data structure for subsequent bioinformatics analyses.

The general purpose of microbiome studies is to comprehensively describe collections of diverse microbes. Therefore, methodologies in ecological studies, which considers complex ecosystems of diverse organisms, are widely used 21. Using the feature table (produced by denoising or clustering algorithms) as a foundation, analyses of microbiome commonly fall into four categories of numerical methods: alpha diversity, beta diversity, differential abundance, and co-occurrence analysis.

Alpha diversity measures how diverse the microbes are contained in each sample, which can be viewed as “intra-sample” diversity. From each sample, quantitative values for richness (e.g. observed features), evenness (e.g. Simpson index 22) or both (e.g. Shannon entropy 23) can be generated. These values are then aggregated by groups to be tested statistically. For pairwise comparison between groups, Mann-Whitney U test is advised, as alpha diversity values are not normally distributed (hence not using t-test) 24. A healthy state of microbiome is commonly characterized by sufficient alpha diversity, composed of symbiotic, commensal microbes that outcompetes incoming pathogens 25. On the contrary, a disease state of the microbiome - i.e. dysbiosis-harbors more pathogenic bacteria and an imbalanced microbial community 26. Furthermore, under certain suboptimal health conditions, e.g. inflammatory bowel disease, the intrinsic alpha diversity could decrease, exhibiting a more fragile ecological system of microbial community 27.

Beta diversity measures the difference between a given pair of samples, i.e. “inter-sample” diversity. Various metrics are calculated to quantify these differences, including Euclidean distance, which calculates the straight-line distance between points in a multi-dimensional space; Bray-Curtis dissimilarity, which quantifies compositional dissimilarity based on relative abundances 28; and UniFrac, a phylogenetic metric that incorporates evolutionary relationships between observed organisms 29. The resulting distance matrix can be visualized on a 2D plane (i.e. plotted on a paper) via dimensionality reduction techniques, such as principal coordinates analysis (PCoA) 30, non-metric multidimensional scaling (NMDS) 31, t-distributed stochastic neighbor embedding (t-SNE) 32, or uniform manifold approximation and projection (UMAP) 33. To date, PCoA is still the most widely used visualization technique due to the fact that it’s simple, linear, and non-stochastic, hence suitable for moderate sample sizes. Nonetheless, the rapid increase in sequencing capability begins to result in large numbers of samples (possibly up to tens of thousands) in a single study 34. In such cases, mathematically sophisticated techniques like t-SNE and UMAP are required to resolve subtle patterns happening at different levels of biological phenomena. Together, these methods enable researchers to visualize and interpret complex microbial community relationships across different samples in an intuitive manner.

Both alpha and beta diversity consider the collective state of microbial communities, providing a holistic, non-granular overview of the diversity and differences of samples without focusing on specific taxa of microbes. Nonetheless, the precision and depth in high-throughput sequencing allows researchers to inspect the abundance of specific microbes with regard to disease conditions. By referencing sequence databases of known microbes 19, 20, enrichment and depletion analysis can be performed on a per-taxon basis, to identify key pathogens related to disease states.

One of the most widely used methods to reveal microbes that are differentially abundant across different patient groups is Linear discriminant analysis Effect Size (LEfSe) analysis 35. LEfSe combines linear discriminant analysis (LDA) with non-parametric statistical testing to identify features (e.g., microbial species) that are both statistically significant and biologically relevant by estimating the effect size (LDA score), which represents the magnitude of the difference in abundance between groups. This approach allows researchers to pinpoint specific microbial features that may play crucial roles in disease or health states. Microbial features reported by LEfSe can be further validated statistically using methods like the Mann-Whitney U test 24 for pairwise comparisons and the Kruskal-Wallis test 36 for multi-class (more than two classes) comparisons, respectively.

It is noted, however, that there are hundreds of oral microbial species to be tested. This results in increased false-positive results, a problem known as the multiple testing problem 37. To address this, corrections for multiple comparisons are essential to control the false discovery rate (FDR). One of the earliest methods developed for this purpose is the Bonferroni correction 38, which adjusts the significance threshold by dividing it by the number of tests performed. While this method is straightforward, it is often overly stringent, limiting the statistical power to reveal any biologically relevant targets. In contrast, the Benjamini-Hochberg procedure is a more modern and widely used approach that controls the FDR, providing a balance between identifying true positives and controlling false positives 39. The Benjamini-Hochberg method is generally preferred in microbiome research for its ability to maintain statistical power while appropriately managing the risk of false discoveries. Through these methods, individual microbial taxa that are significantly associated with different conditions can be identified.

In a complex microbial community like the oral microbiome, various bacteria coexist either in a symbiotic relationship or exhibit antagonistic competition. These interactions can be reflected by the abundance of sequencing reads across multiple samples under different host physiological conditions. For instance, Streptococcus and Veillonella were shown to co-occur in the tongue microbiome 40, consistent with their metabolic interdependence demonstrated in vitro 41. Such co-occurrence analysis is generally based on Spearman’s correlation 42, as the relationships between microbes usually do not follow a linear pattern. Co-occurrence analysis produces a correlation matrix representing pairwise correlations between any given pair of microbes. Positive and negative values indicate co-occurring and mutually exclusive relationships, respectively. This correlation matrix can be further visualized in a microbial network to highlight clusters of co-occurring microbes potentially exhibiting symbiosis.

The relationship between oral microbiome and oral cancer has long been a subject of research, which has been addressed in great depth in the context of common oral diseases such periodontitis 43, 44, 45, 46, 47. Some studies used oral mucosal swabs which allow the collection of paired contralateral normal samples, while others use saliva for a more consistent sampling process.

The first 16S rDNA study on oral microbiome and oral cancer was by Schmidt et al., which found that oral cancer and precancerous samples had a significantly decreased abundance of phyla Firmicutes (genus Streptococcus) and Actinobacteria (genus Rothia) 48. By using the UniFrac distances based on only 12 taxa, the authors were able to separate oral cancer samples from the normal ones 48. A subsequent study on saliva microbiome showed that HNSCC patients exhibited significantly reduced alpha diversity, and higher levels of Streptococcus 49. These saliva-derived findings were opposite to many other studies based on mucosal swabs, which all indicated increased levels of alpha diversity, and higher prevalence of Fusobacterium, Prevotella, Porphyromonas, and Peptostreptococcus on OSCC sites compared to normal mucosa 50, 51, 52. On the contrary, Streptococcus was consistently shown to be depleted on OSCC lesion sites 50, 51, 52. Of note, most of the OSCC-associated taxa were periodontal pathogens, which were further demonstrated in a study using metagenomic shotgun sequencing 53. An additional saliva-based study also showed increased Fusobacterium and decreased Streptococcus in OSCC patients compared to oral leukoplakia and post-operative patients 54. Furthermore, a saliva study involving 248 subjects of all OSCC stages (I to IV) clearly showed increasing Fusobacterium and decreasing Streptococcus with cancer progression 55. Finally, a meta-analysis by Yu et al. integrating 18 studies involving 1056 participants concluded that OSCC patients are enriched in Fusobacteria (genus Fusobacterium) but depleted in Actinobacteria and Firmicutes (genus Streptococcus) 56.

To date, various oral cancer-related clinical factors have been examined with regard to the oral microbiome. Similarly, information in the oral microbiome was utilized as a prognostic biomarker during the treatment process of OSCC. An intriguing study related to HPV status indicated that the presence of Fusobacterium is linked to better survival outcomes in OSCC patients, particularly those without traditional risk factors 57. In addition, two studies have examined the effect of conventional treatment on the saliva microbiome, revealing a prolonged decrease in alpha diversity for months to years following surgery and chemotherapy 58, 59. Compositional differences in microbiome were reported between responders and non-responders of chemotherapy, suggesting the use of microbial signatures as prognostic indicators 59. Given the robust associations between microbiome composition and cancer prognosis, as well as the observed impact of treatment on microbial diversity, a non-invasive early detection test (The CancerDetect for Oral & Throat cancer™, or CDOT) utilizing salivary host and microbiome RNA signature was developed to offer high specificity (94%) and sensitivity (90%) for OSCC detection 60.