Introduction

Colorectal cancer (CRC) is the second leading cause of cancer mortality worldwide, contributing to an estimate of 850,000 deaths and ~1.8 million new cases in 20181,2. The majority of CRC cases are sporadic, with well-established lifestyle risk factors attributed to obesity, alcohol consumption and a diet enriched with red or processed meat3. The gut microbiome is an integral part of human health, and act as an important interface mediating the interactions between environmental cues, host biology, and CRC4,5. Research on CRC gut microbiome has consistently underlined the abundances of certain marker bacteria, among which Fusobacterium nucleatum has been most widely reported and intensively studied6,7,8,9,10.

The Gram-negative rod-shaped F. nucleatum is a common anaerobic member of the human oral microbiome, and it is currently composed of four subspecies (nucleatum, vincentii, animalis, and polymorphum)11. Mechanistic studies have demonstrated that F. nucleatum possesses several virulence factors, most notably FadA and Fap2, which enable the bacteria to potentiate colonic tumourigenesis. The adhesin FadA binds to E-cadherin in CRC cells and activates the β-catenin-dependent oncogenic pathways12, while the lectin Fap2 further facilitates F. nucleatum invasion into CRC cells by specifically binding to the tumour-enriched carbohydrate Gal-GalNAc13. Such interaction triggers the secretions of the pro-inflammatory (IL-8) and pro-metastatic (CXCL-1) cytokines, creating a tumour environment conditioned for accelerated growth and migratory tendency14. Recent studies have further highlighted that the bacteria could induce DNA damage in oral and colorectal cancerous cells15,16. As a result, enrichment of F. nucleatum in CRC microbiomes has been associated with more severe prognosis and poorer overall survival, particularly in a subset of patients with mesenchymal tumours17,18,19. Preclinical research demonstrated that F. nucleatum elimination by antibiotics reduced colorectal tumour proliferation in mice20. These evidences strongly support for the utilization of F. nucleatum as a target for CRC diagnosis and therapy, but current translational potential is hampered by the lack of insights into F. nucleatum diversity and its genomic characteristics in CRC patients.

The majority of microbiome studies were conducted in high-income countries, and such data are sparse regarding populations in developing settings, where host factors, diet and lifestyle could greatly influence the gut microbiome composition and function. Vietnam has an increasing ageing population adopting a more ‘Westernized’ diet and sedentary lifestyle21, where CRC incidence is predicted to climb and rank as among the top three cancers by 202522. Therefore, CRC microbiome studies in Vietnam are necessary to establish the basis for the implementation of microbiome-oriented strategies for CRC prevention, diagnosis, prognosis and therapy. We set out to investigate the microbiome signatures of Vietnamese CRC patients, by applying 16S-rRNA gene profiling on the saliva and gut tissues collected from patients with CRC and non-cancerous colorectal polyps. Additionally, different from prior studies, we used anaerobic culturing and whole genome sequencing (WGS) to study the genomic diversity of Fusobacterium colonising these CRC patients, allowing an in-depth and high-resolution examination of these bacterial populations.

Results

Gut mucosal, but not salivary, microbiomes differ significantly between CRC and controls

We enrolled 43 CRC patients (cases) and 25 patients with colorectal polyps (controls) between December 2018 and January 2020. 16S rRNA microbiome profiling was performed for all the saliva and gut tissue samples collected from the participants, including those originating from the diseased (CRC tumour or polyps) and the adjacent normal sites. Since the majority of bacterial biomass in gut tissues originates from the mucosa, the terms mucosal and tissue-associated microbiomes were used interchangeably. To limit the scope of this study, we selected participants with tumours/polyps detected in the distal colon or rectum. The patients’ demographic and clinical data were summarized in Table 1, which showed that there were no significant differences between the two groups. All polyps showed not more than low-grade dysplasia (i.e. non-cancerous), demonstrating the validity of our control group. No CRC patients have received chemo- or radiotherapy before surgery. Microbiome profiling identified 865 filtered amplicon sequence variants (ASVs – a marker for distinct taxonomic classification) among 66 saliva samples, with a median library size of 36,250 paired-end reads [IQR: 31,827–50,317]. Due to their lower microbial biomass, the library size of gut mucosal microbiomes was smaller (median: 17,711 [IQR: 9037–30,135]), with 1073 filtered ASVs detected across 129 tissue samples (seven removed). Initial quality check showed that the salivary and gut mucosal microbiomes were well separated on Bray-Curtis ordination (Supplementary Fig. 1A), and the sequenced mock community’s composition matched the manufacturer’s description (Supplementary Fig. 1B). Assessment of the rarefaction curves showed that both sequenced salivary and gut tissue samples attained sufficient sampling depth to recover the respective microbiome diversity (Supplementary Fig. 2).

Table 1 Baseline characteristics of patients recruited in this study.
Full size table

Ordination by principal coordinate analysis (PCoA), based on phylogenetic-assisted isometric log-ratio (PhIRL) transformed value, showed that the salivary microbiomes of CRC and controls completely overlapped (Fig. 1a). Only active smoking within the last two years, but not CRC status, was significantly associated with the salivary microbiome structure (RDA, p-value = 0.033). Likewise, only two ASVs belonging to the genera Leptotrichia and Solobacterium were consistently identified as significantly more abundant in the CRC’s salivary microbiome (log2 fold change of 2.25 and 1.82 respectively, adjusted p-values < 0.05) (Supplementary Fig. 3A; Supplementary Table 1). These point to the high structural similarity in the salivary microbiome between the two groups. By contrast, the gut mucosal microbiomes differ significantly based on CRC status (Fig. 1b). CRC and diabetes significantly contributed to the variance in the gut microbiome (RDA, p-value < 0.05). Gut mucosa collected within a participant (tumour and non-tumour for CRC, biopsy and polyp for control) shared more similarity in their microbiomes than those of the same sample type between participants (Fig. 1c), resembling findings from previous research8. We also conducted these analyses using the weighted Unifrac and Bray-Curtis distances, which produced similar interpretations (Supplementary Fig. 4). Additionally, we performed unsupervised clustering on gut mucosal microbiomes, which showed the presence of two robust community state types (CSTs) supported by a mean accuracy of 90.67% (assessed by 50 iterations of nested cross-validation) in a random forest classification. This algorithm also identified that several ‘balances’ contributed significantly in separating the two CSTs (Supplementary Fig. 5). CST1 was generally more enriched in Gammaproteobacteria (mostly Escherichia) while CST2 had higher abundance of Actinobacteria (mainly Collinsella) and Lachnospiraceae (Fig. 1d). The two CSTs were similar in library size (p-value = 0.15, t-test), but different in CRC status (p-value = 0.002, Fisher-exact test), with the majority of control samples (72%) belonging to CST1. Samples from the same patients mostly shared the same CST membership (90.3%, n = 56/62 patients with paired microbiomes). These findings suggest that while CRC status mainly explained the dissimilarity observed in the gut mucosal microbiomes, their overall configurations were determined by the dominant presence of Gammaproteobacteria (Escherichia), possibly driven by an unknown or stochastic factor.

Fig. 1: The salivary and gut mucosal microbiomes of colorectal cancer patients.
figure 1

Principal coordinate analyses (PCoA), conducted on phylogenetic-assisted isometric log-ratio (PhILR) transformed data, of (a) 66 salivary microbiomes, and (b) 129 gut mucosal microbiomes, with different CRC groups and sample types denoted by different colours (see Keys; biopsies and polyps collected from controls, nontumours and tumours collected from cases). c Boxplot showing the distribution of pairwise beta-diversity, calculated on PhILR transformed values, observed in each gut microbiome category. Bold central lines denote the median, the upper whisker extends from the 75th percentile to the highest value within the 1.5*interquartile range (IQR) of the hinge, the lower whisker extends from the 25th percentile to the lowest value within 1.5*IQR of the hinge. Data points beyond the end of the whiskers are outliers. Asterisk markings represent statistically significant differences between groups, as calculated by posthoc Tukey test (p-values ranging from >0.01 to ≤0.05 (*); from >1e−5 to ≤0.01 (**); ≤1e−5 (***)). d Heatmap displaying the proportional abundances of 24 most abundant genera (prevalence ≥15%, mean relative abundance ≥1%, and accounting for ~85% of the gut mucosal microbiome composition), with headers showing the samples’ community state type (CST): CST1 (light gray), CST2 (dark gray), and the corresponding sample type: biopsy (light blue), polyp (dark blue), nontumour (pink), tumour (dark red). Genera were coloured according to their classifications at Phylum level (see Keys). Genera in black box represent ones with probable origin from the oral cavity. The contributions of 24 genera listed here were summarized in Supplementary Table 3. Source data are provided as a Source Data file.

Full size image

Enrichment of oral bacteria in the tumour gut tissues

We applied differential abundance analysis to rigorously detect bacteria enriched in the CRC tumours, by comparing results from different approaches, including ANCOMBC, DESeq2 and corncob (see Methods)23,24,25. Our analyses revealed that ASVs classified as bacteria of putative oral origin (Gemella, Peptostreptococcus, F. nucleatum, Leptotrichia, Selenomonas sputigena, and Campylobacter rectus) were overabundant in the tumour microbiomes, compared to control biopsies (Figs. 1d and 2b) (log2 foldchange of 0.84 to 4.17, adjusted p-values < 0.05; Supplementary Table 1). Within the CRC patients, tumours also showed an elevated presence of the aforementioned oral bacteria, alongside Hungatella, Lachnoclostridium, and Osillibacter, when compared to adjacent non-tumour tissues, albeit with less pronounced fold change (log2 foldchange of 0.87 to 1.95, adjusted p-values < 0.05, Fig. 2a). These increases were coupled with the reduction in abundances of commensal anaerobes in the tumour tissues, such as Blautia, Parabacteroides, Dorea, and Collinsella (log2 foldchange of −1.06 to −1.34, adjusted p-values < 0.05). Collectively, CRC-associative taxa (n = 11, log2 foldchange > 0, Fig. 2a) accounted for a mean cumulative relative abundance of 12% across 43 tumour microbiomes, with prevalence exceeding 90% (n = 39/43). When comparing between different cancer stages, the increased abundance of one taxon (Leptotrichia ASV-13, log2 fold-change 2.63, adjusted p-values < 0.05) was consistently associated with tumours of advanced stages (III-IV), compared to stage II (Fig. 2c). Results from DESeq2 alone additionally showed that F. nucleatum was also enriched in advanced CRC stages (adjusted p-value < 0.05). ASVs confidently assigned as F. nucleatum (n = 14) and Leptotrichia spp. (n = 16) were present at mean relative abundance of 4.6% (prevalence = 26/43) and 6.3% (prevalence = 22/43) in tumour microbiomes, respectively (Supplementary Table 2). We performed similar analysis within the control group and showed that only one ASV (Faecalibacterium) was consistently depleted in polyps compared to paired biopsies (log2 fold-change = −0.99, adjusted p-value < 0.05). However, when compared to CRC samples, Fusobacterium mortiferum, Tyzzerella, and Sutterella were significantly enriched in the control gut microbiomes (Fig. 2b, log2 foldchange of −2.6 to −5.27, adjusted p-value < 0.05; Supplementary Table 1).

Fig. 2: Bacterial taxa significantly abundant among the examined classes.
figure 2

Taxa, or amplicon sequence variants (ASVs), were determined as significant and visualized in a and b if they were detected in at least two of the three tested approaches (ANCOMBC, DESeq2, corncob; adjusted p-value ≤ 0.05). a Log2 fold change of ASVs that differ between paired tumour and non-tumour mucosal microbiomes from case participants, using the full model ‘Patient + sample type’ (n = 86). b Log2 fold change of ASVs that differ between tumour and biopsy (control) mucosal microbiomes (n = 67). Log2 fold change was derived from ANCOMBC test output, and taxa of oral origin were coloured in pink. c Relative abundance of ASVs assigned as Fusobacterium mortiferum (n = 14), Fusobacterium nucleatum (n = 14), Leptotrichia (n = 16), and Collinsella (n = 14) in the tumour and nontumour mucosal microbiomes, stratified by cancer stages (III-IV vs. II). The error bar represents standard error of the mean relative abundance in each bar plot. Source data are provided as a Source Data file.

Full size image

To investigate bacterial co-occurrence and their potential interactions, we next constructed a correlation network of gut microbiomes from CRC patients (n = 86) (Fig. 3)26,27. Two oral bacteria clusters emerged from this network, one consisting of several Streptococcus and Veillonella taxa, and another composed mostly of aforementioned tumour-associated ASVs (Leptotrichia, Selenomonas, F. nucleatum, Streptococcus, Granulicatella, Gemella, Peptostreptococcus, and Parvimonas). The latter cluster exhibited positive correlation with E. coli (cor = 0.488, p-value = 9.2e−07), and antagonism toward the gut commensal Blautia (cor = −0.466, p-value = 2.97e−06). Besides, other tumour-associated ASVs such as Hungatella, Lachnoclostridium, and C. rectus were clustered alongside Negativibacillus and Eggerthella, which showed strong negative correlations with anaerobic gut commensals Dorea, Bacteroides, and Faecalibacterium. These findings highlight the potential competition between tumour-associated taxa and common gut commensal anaerobes. Other Fusobacterium species, F. mortiferum and F. varium were not linked to the oral clusters, showing that they were mainly gut inhabitants. Comparison with the network constructed from salivary microbiomes revealed that the same tumour-associated ASVs (F. nucleatum, Gemella, Selenomonas) formed similar clusters as observed in the CRC gut microbiomes (Supplementary Fig. 6).

Fig. 3: Correlation network of colorectal cancer gut mucosal microbiomes.
figure 3

The network was constructed from 117 most representative ASVs sampled from 86 mucosal microbiomes, outlining significant interactions detected by both CCLasso (p value ≤ 0.01 and absolute correlation strength > 0.37) and SpiecEasi. Positive and negative interactions were coloured as red and blue lines respectively, with line weight proportional to correlation strength. The ASVs (nodes) were coloured based on taxonomic family (see Legend), with sizes proportional to their relative abundances. The light green shaded area entails ASVs identified as members of the human oral microbiome (comparison with expanded Human Oral Microbiome Database); the blue shaded area covers ASVs identified as gut anaerobic commensals (Lachnospirales, Bacteroidales, Bifidobacteriales, Oscillospirales); and gray shaded area covers other tumour-associated ASVs (as identified in Fig. 2a). Source data are provided as a Source Data file.

Full size image

Diverse Fusobacterium colonizes CRC patients

Since F. nucleatum was more enriched in the tumour microbiomes and previously demonstrated to promote tumourigenesis, we next studied the population structure of Fusobacterium recovered from CRC patients. Six patients with a Fusobacterium relative abundance at the tumour site exceeding 10% (except for patient 18) and covered different cancer stages were selected for Fusobacterium isolation. In total, we isolated 56 presumptive Fusobacterium organisms, as identified by the matrix-assisted laser desorption/ionization time of flight mass spectrometer (MALDI-TOF), from the oral, nontumour and tumour samples of these patients (Table 2). Fifty-three successfully sequenced genomes belong to F. nucleatum (n = 38) and F. periodonticum (n = 15) species complexes, of which phylogenetic reconstruction was performed separately. Across the two phylogenies, we identified 14 phylogenetic clusters (PCs; each with 2–6 isolates exhibiting negligible genetic differences) and 12 singletons originating from this study’s collection, which were collectively named as PCs herein (17 F. nucleatum, 7 F. periodonticum, and 2 F. hwasookii). Each PC represents a single Fusobacterium strain, with multiple colonies picked from the same patient (Table 2). Core-genome phylogeny of F. nucleatum showed that tumour-associated isolates were detected in all four subspecies (animalis, vincentii, nucleatum, polymorphum) (Fig. 4a). In the F. periodonticum phylogeny, tumour-associated isolates (2 PCs isolated from P18, P40) formed a distinct cluster that is phylogenetically separated from the available references (Fig. 4b). These isolates all showed ~91% average nucleotide identity (ANI) to the closest F. periodonticum references, suggesting that they constitute a novel subspecies of this species complex, denoted herein as novel F. periodonticum (novelFperi). Likewise, one gut PC (H16_Fa) shared 93% ANI to the closest F. nucleatum references and were phylogenetically distant from the remaining F. nucleatum isolates, potentially indicative of a novel F. nucleatum subspecies.

Table 2 Summary of Fusobacterium isolates recovered from this study.
Full size table
Fig. 4: Maximum likelihood phylogenies of Fusobacterium isolates from this study.
figure 4

a F. nucleatum species phylogeny constructed from the alignment of 516 core genes (89,900 variant sites; N = 57), using F. hwasookii clade as an outgroup. b F. periodonticum species phylogeny constructed from alignment of 863 core genes (106,738 variant sites, N = 26). Red circles at internal nodes denote bootstrap values ≥80. The associated metadata on the right describe the patient ID and clinical origin of isolates (where reference genomes were left blank), and the genomic presence of several virulence factors (fap2, fadA, fadA2, radD, fadA3, cbpF). Light blue shaded area covers isolates identified as novel F. periodonticum subspecies. The scale bars denote the estimated number of substitutions.

Full size image

The Fusobacterium population within each individual patient was diverse (2–7 PCs). Several Fusobacterium species/subspecies were detected in each patient’s saliva, sometimes with more than one PCs of the same subspecies (P18, P46) (Table 2). Likewise, we observed similar diversity in gut-associated isolates, with more than one PCs detected in three patients (P10, P16, P18). Most patients did not share the same Fusobacterium subspecies recovered from both oral- and gut-associated isolates, except for P16 (polymorphum). However, phylogenetic evidence confirmed that the two niches harboured distinct populations, which were ~16,955 SNPs apart (Fig. 4a). Particularly, oral Fusobacterium isolates from P18 (n = 9) belonged to six different PCs (mostly F. periodonticum and F. hwasookii), while 6/7 gut isolates were of a single novelFperi clone. By contrast, Fusobacterium from tumour and nontumour sites were frequently clustered in the same PC (n = 4; in P10, P16, P18 and P40), indicating that the same bacterial clones have colonised and spread beyond the tumour microenvironment. We used the mapping approach to confidently inspect the intraclonal variations within these PCs, and showed that they shared minimal genetic differences in the core genome (1–2 SNVs). These values fall in range with the variation observed in five other gut PCs (with either tumour or nontumour isolates; 0–5 SNVs) and five other oral PCs (1–10 SNVs).

Variation in Fusobacterium virulence gene content

We next sought to examine the presence of several Fusobacterium virulence factors, of which pathogenicity has been proven in experimental studies, including genes encoding adhesin (fadA, cbpF), lectin (fap2), and bacterial co-aggregation factor (radD)12,13,28,29. RadD is an autotransporter facilitating Fusobacterium’s interspecies interaction in polymicrobial biofilms29, while CbpF inhibits CD4+ T-cell response through CEACAM1 binding and activation30. Genomic screening showed that fap2 was present and intact in the majority of genomes from both species (49/53), with disruptive mutations occurring in some isolates, such as the tumour-associated F. nucleatum animalis in P46 (Fig. 4a). We also detected fadA in all isolates (except S18-65), with all F. periodonticum variants one amino acid shorter (codon A22) than the canonical FadA found in F. nucleatum (129 aa). The other elements showed variable presence among the examined genomes. For example, cbpF was present in all F. nucleatum nucleatum, F. nucleatum vincentii, and novelFperi, while radD was co-localised with fadA2/radA (a 122 aa fadA homolog) in 28 isolates. Another fadA homolog (fadA3) with unknown function was prevalent in both two Fusobacterium species. Phylogenies of FadA and CbpF showed that the two tree topologies were largely in agreement with those inferred from the core genomes, suggesting the absence of horizontal gene transfer (Supplementary Fig. 7). By contrast, the clustering pattern observed in the Fap2 phylogeny was concordant to subspecies classification for F. nucleatum nucleatum, F. nucleatum vincentii, and F. periodonticum, but was admixed for F. nucleatum polymorphum, F. hwasookii and F. nucleatum animalis (Supplementary Fig. 8A). fap2 encodes a very large protein of variable length (median of 3938 aa [range: 3436–4669]), and the protein length showed some correlation with its phylogenetic clustering, with variants >4200 aa (n = 6) all belonging to a monophyly composed of F. hwasookii and F. nucleatum polymorphum. Similarly, the RadD phylogeny did not concur with those inferred from the core genomes, and its length variation (median 3526 aa [range: 3461–3602]) also showed association with the tree topology (Supplementary Fig. 8B). radD was ~800 bp downstream of fadA2, which is flanked by an IS150 transposase on the F. nucleatum 23726 reference genome. This could explain the mobilization mechanism of radD-fadA2 across the Fusobacterium phylogeny. These data indicate that the autotransporter encoding genes fap2 and radD may have undergone frequent horizontal gene transfer or recombination in the F. nucleatum species complex.

Discussion

Our study revealed the composition of microbiome perturbations at the tumours of Vietnamese patients with CRC and non-cancerous colorectal polyps. Tumour-enriched taxa include mostly bacteria of putative oral origin, such as F. nucleatum, Leptotrichia, Gemella, C. rectus, and Selenomonas, which agrees with findings from previous studies profiling either gut mucosal or faecal microbiomes in different CRC populations8,9,10,31. Compared to these studies, some CRC-indicative taxa (Parvimonas, Solobacterium, Porphyromonas) were not included in our findings because we applied a conservative approach in reporting differential abundance testing. Indeed, ASVs assigned to Parvimonas and Porphyromonas only showed significant enrichment in tumour microbiomes in either DESeq2 or ANCOMBC test. Thus, these slight differences likely stem from technical rather than biological reasons, highlighting that the proliferation of oral bacteria at the gut mucosa could be a universal signature of CRC microbiomes. We found that several of these oral taxa shared identical ASVs between the oral and gut niches, pointing to a probable oral origin of tumour-associated taxa. Our analysis found that these bacteria also display a co-occurrence pattern in the tumour microbiome, which agrees with the frequent presence of polymicrobial biofilms composed of oral taxa (F. nucleatum, Peptostreptococcus, Gemella) in colorectal tumours32. Among the oral bacteria, F. nucleatum stands out for its ability to form “bridging” interactions with other bacteria via the presence of several adhesins11. F. nucleatum was recently reported to secrete FadA with amyloid properties, which confers acid tolerance and provides a scaffold for biofilm formation33. In addition, our analyses pointed to the significant presence of Leptotrichia in tumour microbiomes, especially in advanced cancer. This association, however, has only been noted in few studies31,34. This may be due to the differences in sampling location, as tumours excised from the distal colon (as performed for all cases in our study) were reported to harbour a higher abundance of Leptotrichia, compared to those originating from the proximal colon34. Regarding the oral microbiome, we reported that only two taxa (Leptotrichia and Solobacterium) were significantly enriched in the saliva of CRC patients. Previous research has revealed that several bacteria (Parvimonas, Haemophilus, Prevotella, Neisseria) were significantly depleted in the CRC’s oral microbiome, compared to healthy controls35. This discrepancy could stem from different oral sampling methods (saliva vs. cheek swab), study population (Asian vs. European), control populations (polyp vs. healthy), and analytical tools employed.

Asides from oral taxa, Hungatella overabundance was the most significant signature of CRC microbiome in our dataset. This falls in line with results from a recent metagenomic meta-analysis, showing that Hungatella hathewayi’s specific choline trimethylamine-lyase gene (cutC) was significantly enriched in the faecal microbiomes of CRC patients10. Moreover, colonic H. hathewayi could induce hypermethylation in prominent tumour suppressor genes, thus silencing their functions and promoting intestinal epithelial cell proliferation36. On the other hand, we found that F. mortiferum was the most significantly enriched taxon in the polyp control group. F. mortiferum was known as a hallmark for dysbiosis in infectious diarrhoea37, and recent studies have also reported the abundance of F. mortiferum in patients with colorectal polyps38,39. Furthermore, this species was shown to be present in the gut microbiomes of ~60% of a cohort in Southern China, albeit in very low abundance (~0.5%)40. Unlike other Fusobacterium species, F. mortiferum was devoid of distinctive virulence factors such as adhesins FadA and Fap241. The association between F. mortiferum and colorectal polyps will need to be further addressed in future studies.

Despite the increasing importance of F. nucleatum in the pathogenesis of CRC and other invasive diseases11, genomic characterisation of these bacteria from patient populations is currently limited due to technical difficulties in Fusobacterium isolation. Here, we applied targeted culturomics approach, which combines anaerobic culturing, high-throughput identification by MALDI-TOF and WGS, to study the Fusobacterium population in high resolution and help uncover novel bacteria42. Indeed, we discovered novel subspecies of both F. nucleatum and F. periodonticum from culturing the gut tissues, showing that the microbiomes in non-Western settings offer untapped diversity. Using metagenomic assemblies from Chinese faecal microbiomes, Yeoh and colleagues have proposed several new Fusobacterium species (based on 95% ANI cutoff)41. Our WGS approach provided more accurate and complete realization of the bacterial genomes, which contributes to the global representation of Fusobacterium diversity (with 26 non-duplicate assemblies added). Furthermore, our approach allows for delineation of bacteria from tumour and non-tumour sites, which is inaccessible by faecal metagenomes. Nevertheless, targeted culturomics generally has low sensitivity, and bacterial recovery is subjected to factors such as storage time and condition. Therefore, our approach could not capture the high diversity of Fusobacterium in the oral niche43, which likely explains the absence of close genetic relatedness between oral and gut Fusobacterium isolates. Previous research deploying WGS has demonstrated that oral and tumour-originated F. nucleatum shared little genetic divergence (0–183 SNVs), supporting the notion that oncogenic Fusobacterium arise from the patient’s oral microbiome44. Similarly, using arbitrarily primed PCR, Komiya and colleagues showed that identical F. nucleatum strains were isolated from 6/14 paired gut tissue-saliva samples45, but WGS was not conducted to verify the exact genetic differences. The populations of Fusobacterium colonising the oral cavity and gut were heterogeneous within some individuals, even at the subspecies level, which mirrors the diversity observed previously for gut commensals such as Bifidobacterium46. Chronic infections with Helicobacter pylori at the stomach, which increases the risk of gastric cancer, usually result in extensive clonal propagations detected by WGS within each patient, though isolates were collected in a single timepoint47. This prolonged colonization scenario contrasts with our observations in three CRC patients (P10, P16, P18), in which two to three Fusobacterium strains (with minimal intraclonal variation) were present at the tumour and non-tumour gut tissues. Given that CRC could take years to develop, we hypothesize that the Fusobacterium population at tumour sites might fluctuate in response to the frequent seedings from the highly diverse oral source. Additionally, identical Fusobacterium strains have been retrieved from the colonic tumours and liver metastasis of the same patient, suggesting the metastatic potential of tumour-borne Fusobacterium20. Future longitudinal study design is necessary to investigate the Fusobacterium population dynamic within CRC patients, including those from different geographical regions.

The two well-described major virulence genes (fadA and fap2) were identified in the majority of Fusobacterium genomes, regardless of niche. This concurs with previous research reporting the high prevalence of fadA and fap2 in F. nucleatum and F. periodonticum metagenomic assemblies from a cohort in China41. These suggest that Fusobacterium with high virulence potential are prevalent in the human population, and the genetic presence of fadA and fap2 is not suitable for predicting the risk of Fusobacterium-related CRC. All gut-derived novelFperi isolates harboured the examined virulence genes (fadA, fap2, radD, and cbpF), which was more similar to F. nucleatum compared to F. periodionticum. Moreover, fap2 and radD showed variation in gene length and evidence of horizontal gene transfer, underlying the significance of dynamic evolutionary processes in shaping Fusobacterium’s virulence landscape. A recent study using Fusobacterium WGS has also reported that fap2 could be either missing or highly divergent in tumour-derived F. nucleatum, suggesting the mobile nature of fap248. Since Fap2 orchestrates F. nucleatum invasion into CRC tumour cells via specific binding to Gal-GalNAc, this ligand-receptor interaction was recently proposed as a target for clinical intervention in Fusobacterium-enriched CRC49. Interestingly, our genetic analysis predicted that fap2 was either missing or truncated in some gut-associated Fusobacterium isolates, which may indicate the complex lifestyle of Fusobacterium once colonising the gut environment.

Some limitations were notable in our study design. Due to ethical concerns, patients with colorectal polyps were selected as the control group, instead of healthy age-matched individuals. Our interpretations do not extend to cancer in the proximal colon, though previous reports have noted that proximal CRC tumours had a higher Fusobacterium abundance50. The sample size of cultured Fusobacterium isolates was moderate and did not include longitudinal sampling, so it was not possible to investigate the bacterial evolution in longer timeframe. Besides, our saliva sampling might not fully reflect the microbiome compositions at other oral sites, as well as to capture the whole diversity of Fusobacterium, which is more abundant in subgingival dental biofilms. Notwithstanding these shortcomings, our study reconfirmed the prominent role of oral anaerobic conglomerates in CRC microbiome in an understudied Asian population, and provided new insights into the genomic diversity of the oncobacterium Fusobacterium. The observed diversity in this organism should be taken into account when designing future diagnostic or therapeutic tools that target Fusobacterium.

Methods

Study design and sample collection

This prospective case-control study enrolled adult Vietnamese patients (≥18 years old) admitted at Binh Dan Hospital, a large surgical hospital in Ho Chi Minh City Vietnam, from December 2018 to January 2020. This study received ethical approval from the Ethics Committee of Binh Dan Hospital (690/BVBD-QD), and the study was performed in compliance with the Declaration of Helsinki. Written informed consent was obtained from all study participants. Cases were defined as patients diagnosed with left-sided colorectal cancer (distal colon and rectum) of stage II onward, who received colectomy treatment and underwent non-antibiotic pre-operative bowel preparation. Controls were patients diagnosed with colorectal polyps (single/scattered non-cancerous polyps at distal colon or rectum), who received polypectomy at the hospital.

Demographic and clinical information were collected from study participants at recruitment. Cancer stage classification was based on the TNM Staging system51. A saliva sample (~3 mL) was collected within three hours pre-operation from each study participant (by spitting into a sterile container). For cases, the mucosa epithelia at the tumour and adjacent non-tumour (2–10 cm away from the tumour) sites were collected aseptically from the excised colon. For controls, we collected colorectal polyps and 2–3 biopsies of non-polyp mucosal epithelium (~50 mg) during colonoscopy. All clinical samples were stored on ice and transported back to the laboratory within 4 h, then were stored in -80 °C until further experiments.

16S rRNA gene sequencing

Microbiome profiling was performed on recruited 43 cases and 25 controls. Total DNA was extracted from whole biopsies and polyps (due to their small size), whole tumour (mucosa plus tumour tissue), and nontumour tissues (n = 136) using the FastDNA spin kit for soil (MP Biomedicals, USA), with bead-beating step on Precellys 24 homogenizer (Bertin Instruments, France). Though our approach targets the whole tissue and not just the mucosa, the bacterial biomass in the mucosa still comprise the majority of tissue-associated microbiome. Thus, the terms mucosal and tissue-associated microbiomes were used interchangeably. DNA from the saliva samples (n = 67, one missing) was extracted using the ReliaPrep Blood gDNA Miniprep (Promega, USA). For microbiome profiling, all samples underwent primary PCR amplification (30 cycles) using the conventional V4 primers (515F-806R) and KAPA HiFi Hot Start DNA polymerase (KAPA Biosystems, USA), and secondary PCR was performed to add dual-indexes (IDT, USA) to each sample, following procedures optimized in a published protocol52,53. Additionally, we applied the same procedures to a positive control (Zymo mock community, Zymo Research, USA) and six negative controls (two for each DNA extraction kits, and two no-template PCR amplifications). 16S rRNA sequencing was performed for all samples on one run of the Illumina MiSeq platform, to generate 250 bp paired-end reads.

Microbiome data analysis

All data analyses were conducted in R (v4.1.1) and Rstudio using multiple packages, including ‘dada2’, ‘phyloseq’, ‘DESeq2’, ‘ANCOMBC’, ‘corncob’, ‘philr’, ‘ggplot2’, ‘vegan’, ‘SpiecEasi’ and others23,24,25,27,54,55,56,57. Generated sequence reads were analysed under the amplicon sequence variant framework (ASV) using DADA258,59. Chimeric sequences were detected and removed independently for each sample. Taxonomic assignment (up to the species level) was performed using the RDP Naïve Bayesian Classifier implemented in ‘dada2’ package, on the SILVA v138 train dataset60. Further filtering removed ASVs matching the following criteria (1) classified as ‘Mitochondria’ or ‘Archaea’, (2) unclassified at Kingdom or Phylum level, (3) identified as kitome or contamination from mock community (except Escherichia and Enterococcus ASVs), or (4) identified as low abundant singletons (abundance ≤ 10 counts and present in only one sample). This resulted in 2,461 ASVs detected across 203 samples (68 participants), totalling 5,250,754 sequences.

Saliva and gut mucosal microbiomes were then analysed separately. For saliva microbiomes, we removed singleton ASVs with abundances <79 sequences (third quartile threshold) and one sample with low sequencing depth. The filtered ASVs (n = 865) were aligned using PASTA61, and a maximum likelihood phylogeny was constructed under the GTR + G model using IQ-Tree (with 1000 rapid bootstrap)62. The resulting phylogeny was used to transform the ASV count matrix into isometric log-ratio (ILR) ‘balances’ (weighted log-ratio between two ASVs), using the “philr” package56,63. Ordination was performed using principle coordinate analysis (PCoA) on a calculated Euclidean distance matrix. To identify covariates which explain the salivary microbiome structures, we performed redundancy analysis on the ‘balance’ value matrix of 62 samples with complete metadata. We repeat the same analytical procedures on the gut mucosal microbiome data. Low-abundance singleton ASVs (<44 sequences – third quartile threshold) and seven samples with low sequencing depth (<1300 sequences each, as assessed by rarefaction curve) were removed, retaining 1073 ASVs across 129 samples for downstream analyses. We tested the association between covariates and the gut mucosal microbiome structures using redundancy analysis, performed on the ILR-transformed ‘balance’ values of 120 samples with complete metadata. The ILR-transformed values were used to calculate the beta-diversity, within and between participants. In addition, the gut mucosal microbiomes (n = 129) were clustered into community state types (CSTs) using the partition around medoid (pam) algorithm on the calculated ILR-transformed distance matrix, with the optimal number of CSTs (k = 2) determined by gap statistic and average silhouette width (asw)64. The random forest classification algorithm (10,000 trees) was then used to identify ‘balances’ differentiating the two CSTs, using the package ‘randomforestSRC’65. We further assessed the performance of this model using 50 iterations of nested cross-validation (five-fold cross-validations for both the outer and inner loops), as implemented in Python’s Sklearn library.

Evaluating differential abundances

In order to detect ASVs that showed significantly differential abundance between two examined groups, we utilized the compositional data analysis approach implemented in ANCOMBC23. In addition, the same comparisons were performed using DESeq2 and corncob to check for consistent results, as recommended in recent benchmark studies66,67. The comparisons include salivary microbiomes in cases (n = 43) and controls (n = 23); paired tumours (n = 43) against adjacent non-tumours (n = 43); paired polyps (n = 16) against non-polyp biopsies (n = 16); tumours (n = 43) against non-polyp biopsies (n = 24); tumours of cancer stage III-IV (n = 24) against stage II (n = 18). For paired comparison within cases and controls, the model design was set to “~Patient + sample_type” to increase statistical power68. Multiple hypothesis testing was corrected using Holm or Benjamini–Hochberg method, setting false discovery rate as 0.05. ANCOMBC and corncob approaches were carried out using default parameters. For DESeq2, library size corrections were estimated using ‘poscounts’ method. All comparisons were performed using likelihood ratio test, and ASVs with adjusted p-value < 0.05 (and base mean >20 for DESeq2) were considered significant hits. To minimize the number of false positives, ASVs which showed significant hits in at least two tested methods were considered differentially abundant and included in final interpretation. We performed BLAST for ASV sequences of interest against the expanded Human Oral Microbiome Database (HOMD; www.homd.org/), and species identification was assigned if the ASV showed >99% nucleotide similarity to that in the database.

Correlation network

We constructed a correlation network of gut mucosal microbiomes from colorectal cancer patients (n = 86), using 117 most representative ASVs, defined as ones with abundance of at least 10 sequences detected in at least 15 samples. This filtering resulted in a median sample retainment rate of 77% [70–85%]. The correlation network was constructed using CCLasso, with 250 bootstrap and three-fold cross validation26. Interactions with adjusted p values < 0.01 and absolute correlation strength >0.37 were considered significant hits. Additionally, a separate correlation network was inferred using SpiecEasi on the same dataset27. Both these methods have been demonstrated to produce robust performance in a recent benchmark study69. To avoid spurious hits, only significant interactions detected by both the CCLasso and SpiecEasi approaches were included in the final visualization. We applied the same procedures to construct correlation networks of microbiomes in saliva samples (n = 66, 115 ASVs) and controls’ gut mucosa (n = 43, 90 ASVs).

Fusobacterium isolation and whole genome sequencing

Fusobacterium isolation was performed on six selected case patients (P10, P16, P18, P28, P40, P46), whose Fusobacterium relative abundance in the tumour microbiome exceeded 0.5% as inferred by microbiome profiling. The respective samples (saliva, tumour, and non-tumour tissues) were subjected to anaerobic culturing in a Whitley A35 anaerobic workstation (Don Whitley Scientific, UK) supplied with 5% CO2, 10% H2, and 85% nitrogen gas, following an established Fusobacterium isolation procedures45. Briefly, gut tissues were thawed on ice, and ~100 mg tissues were aseptically excised and anaerobically homogenized, using sterile surgical blades, in phosphate buffer supplemented with L-cysteine HCl (500 mg/L), Tween 80 (500 mg/L), and 0.1% resazurin. The suspension (100 µL) was then plated onto the selective media (EG agar supplemented with L-cysteine HCl, 50 ml/L of defibrinated sheep blood, 7 mg/L of crystal violet, 5 mg/L of vancomycin, 30 mg/L of neomycin, and 25 mg/L of nalidixic acid; Sigma-Aldrich, Germany). Thawed saliva samples were plated directly on the selective media. Plates were incubated at 37 °C for 48–72 h, and colonies (up to 10) resembling that of Fusobacterium were picked from each plate and sub-cultured on new EG media to confirm purity and select for single colonies. The isolate’s taxonomic identities were queried using MALDI-TOF, and those characterised as Fusobacterium species were retained. A total of 56 Fusobacterium isolates were recovered and subjected to DNA extraction using the Wizard genomic extraction kit (Promega, USA). For each isolate, 1 ng DNA was used to prepare the sequencing library using the Nextera XT library preparation kit, following the manufacturer’s instruction. Normalized libraries were pooled and sequenced on an Illumina MiSeq platform to generate 250 bp paired-end reads.

Pangenome analysis, phylogenetic reconstruction and screening for virulence genes

FASTQC was used to check the sequencing quality of each read pair70, and Trimmomatic v0.36 was used to trim sequencing adapters and low-quality reads71. For each isolate, the trimmed read set was input into Unicycler v0.4.9 to construct the de novo assembly, using default parameters, and contigs of size over 500 bp were retained72. The assemblies were checked for traces of contamination using Checkm, and three assemblies were shown contaminated and discarded73. The resulting assemblies were of adequate quality, with median size of 2,125,169 bp [IQR: 2,067,843–2,168,429], median number of contigs of 133 [IQR: 86–173] and the median N50 of 35,535 bp [IQR: 21,685–51,953]. Prokka v1.13 was used to annotate the assemblies, using the well-annotated F. nucleatum 23726 (accessed via FusoPortal) as reference74. To provide preliminary taxonomic classification up to the subspecies level, FastANI was used to calculate the average nucleotide identity (ANI) between the individual assembly and a set of Fusobacterium references, with an ANI value ≥95% denoting a shared species/subspecies75. The pangenomes of 57 F. nucleatum/hwasookii isolates (38 sequenced herein plus 19 references) and 25 F. periodonticum isolates (15 sequenced herein plus 10 references) were constructed separately using panX76. The respective core genome from each species complex was aligned, with invariant sites removed, producing SNP alignments of 89,900 bp (F. nucleatum/hwasookii complex) and 106,738 bp (F. periodonticum complex). These were input into RAxML to construct maximum likelihood phylogenies, under the GTRGAMMA substitution model with 300 rapid bootstraps77. Using the pangenome analysis output, we screened for the presence of several known Fusobacterium virulence genes (fap2, fadA, radD, cbpF). The intact presence or synteny of each genetic element was checked manually by gene alignment (Seaview) or genome visualization (Artemis) tools78,79. Visualization of phylogenetic tree and associated metadata was performed using package ‘ggtree’80. Individual protein sets were aligned and inspected in Seaview, and phylogenies were constructed in RAxML, using the PROTGAMMAGTR model and 200 rapid bootstraps.

Intra-clonal variation examination

To investigate intra-clonal variation with high confidence, we examined single nucleotide variants (SNV) among isolates belonging to the same phylogenetic cluster (Fig. 4 and Table 2), using the mapping approach recommended previously81. For each phylogenetic cluster, trimmed fastq files from the isolates were concatenated and input into Unicycler to construct a pan-assembly, with contigs less than 500 bp removed. This pan-assembly was ordered against an appropriate Fusobacterium reference using ABACAS, creating a pseudogenome reference82. Trimmed paired-end reads from each isolate were mapped against this reference using a custom wrapper script. Briefly, mapping was conducted using BWA MEM algorithm and samtools v1.883,84, with duplicate reads removed using PICARD, followed by indel realignment by GATK85. SNVs were detected using the haplotype-based caller Freebayes86, and low quality SNVs were removed using bcftools if they met any of the following criteria: consensus quality <30, mapping quality <30, read depth <4, ratio of SNVs to reads at a position (AO/DP) < 85%, coverage on the forward or reverse strand <1. The bcftools ‘consensus’ command was used to generate a pseudosequence87, integrating the filtered SNVs and invariant sites, and masking the low mapping region (depth <4) and low-quality SNVs with ‘N’. The presence of high quality SNVs were validated by manual visualization of output bam files in Artemis, and SNV pertaining to recombination, transposons, plasmids, or repetitive elements were excluded from interpretation.