Introduction

Colorectal carcinoma (CRC) remains one of the most prevalent malignancies worldwide1. Despite advancements in multidisciplinary treatments, the prognosis for patients with locally advanced or metastatic CRC remains poor, largely due to high rates of distant metastasis, chemotherapy resistance, and relapse, which contribute to unsatisfactory long-term survival outcomes2,3. As a result, the identification of reliable prognostic factors is essential for optimizing treatment strategies and improving patient outcomes, particularly given the considerable variation in postoperative results and long-term prognosis among individuals4.

The human oral microbiota is a complex ecosystem consisting of over 770 species-level taxa, each adapted to different niches within the oral cavity 5. Evidence suggests that specific oral-derived species, including Porphyromonas gingivalis, Fusobacterium nucleatum, Streptococcus anginosus, Peptostreptococcus stomatis, and Prevotella intermedia, are associated with the occurrence of CRC6,7,8,9,10. These oral pathogens may influence CRC development by disrupting the gut microbiota via the oral-to-gut microbe translocation, or by activating the host immune inflammation11,12. For instance, the well-known pathogen P. gingivalis, upon ectopic colonization in the gut, can recruit myeloid immune cells, activate the NLRP3 inflammasome, and enhance the expression of inflammatory cytokines such as IL-1β, thereby promoting the onset and progression of colorectal cancer (CRC)13. While the role of oral-derived microbes in CRC development is identified, the mechanisms by which alterations in the oral microbiota are associated with CRC prognosis remain largely unknown.

The oral microbiota has been implicated in the progression of various malignancies, and certain species have been associated with cancer outcomes14. For example, Mohamed, et al. reported that a higher salivary carriage of the genus Candida was associated with a poor prognosis, while Malassezia was enriched in patients with favorable outcomes in oral squamous cell carcinoma15. Additionally, Du et al. found that reduced oral microbiota diversity was correlated with higher mortality in nasopharyngeal carcinoma, particularly among elderly patients16. A recent study observed that intra-tumoral infection of oral-derived F. nucleatum is associated with poorer disease-free survival and overall survival in patients with stage III CRC17. These findings underscore the critical role of the oral microbiota in influencing cancer prognosis, and further investigation of the prognostic function of oral microbiota in CRC is warranted.

In this study, we investigated the salivary microbiota of 312 CRC patients using 16S rRNA gene full-length sequencing. We identified microbial signatures associated with CRC outcomes and developed a microbial risk score (MRS) to predict CRC progression. Furthermore, we integrated the MRS with key clinical factors to construct a multi-factorial prognostic model with enhanced predictive performance. While these findings require validation in multi-center cohorts, our study pioneers the application of salivary microbiota profiling as a non-invasive prognostic tool in CRC.

Results

Characteristics and microbial community structure among patients with different clinical outcomes

In this study, we collected saliva samples from 312 CRC patients who were scheduled for surgical tumor resection. The median follow-up period for patients with CRC progression was 21.5 months, whereas for survival was 25.4 months, during which 59 experienced disease progression and 25 died (Fig. 1). Demographic and clinical characteristics of the participants are presented in Table 1. There were no statistically significant differences in basic demographic characteristics, such as age, sex, marital status, body mass index (BMI), family history of tumors, and alcohol consumption history, between non-progression and progression groups, and between survival and death groups. Patients with progression or with death outcomes were more likely diagnosed at an advanced clinical stage, exhibited higher rates of lymphatic invasion, perineural invasion, and lymphatic metastasis when compared to good outcome individuals (Table 1).

Fig. 1: The workflow of the study.
figure 1

Step 1: Saliva samples were collected from 312 CRC patients scheduled for surgical tumor resection. The median follow-up time was 21.5 months for patients with disease progression and 25.4 months for survival. Salivary microbiota profiling was performed using 16S rRNA full-length sequencing. Step 2: Species-level identification was conducted with inclusion criteria of prevalence ≥10% and relative abundance ≥0.01%, resulting in 98 species. A 1:1 cross-validation strategy was applied by randomly splitting the cohort into discovery (n = 156) and test (n = 156) sets, with 1000 subsampling iterations. Prognosis-related microbes were identified using univariate and multivariate Cox regression analyses, highlighting Campylobacter gracilis, Neisseria oralis, and Treponema medium as significant biomarkers. Step 3: A microbial risk score (MRS) was constructed based on the identified prognostic bacterial species. Clinical prognostic factors were also evaluated, and a comprehensive model integrating microbiota biomarkers (three species) and clinical factors (three variables) was developed. The performance of the comprehensive model was validated and compared with models based solely on the clinical factors, demonstrating superior prognostic accuracy.

Full size image
Table 1 Patient demographic and clinical characteristics
Full size table

We investigated the oral microbiota diversity and composition differences in several clinical factors. There were no significant differences in oral microbiota diversity with differing prognostic outcomes. Shannon diversity tended to increase with the advanced TNM stage and lower tumor histological grade. We did not observe a significant difference in sex, lymphatic metastasis status, and perineural invasion status (Supplementary Fig. 1).

The oral bacterial species are significantly associated with the prognosis of CRC patients

To determine whether the oral bacteria were associated with CRC prognosis, we conducted a two-step identification. Firstly, the univariate Cox regression model was applied to 98 oral microbial species that appeared in at least 10% of the total samples with a relative abundance of at least 0.01%, to preliminarily screen the potential species. Twelve species were found to be substantially associated with CRC progression, and five species showed substantial associations with CRC survival (Supplementary Table 1 and Supplementary Table 2). Subsequently, those significant species were incorporated into a multivariate Cox regression model, adjusting for age, sex, tumor stage, and neoadjuvant chemoradiotherapy, to assess the independent effect of these bacteria on the prognosis of CRC patients. Finally, Campylobacter gracilis and Neisseria oralis were identified as the indicators associated with an increased risk of CRC progression, while Treponema medium was correlated with a reduced risk of progression in the total population cohort (Hazard ratio, HR = 2.63 for N. oralis, 95% confidence interval (CI) = 1.50–4.63, false discovery rate (FDR) = 0.08; HR = 2.27 for C. gracilis, 95% CI = 1.34–3.86, FDR = 0.09; HR = 0.41 for T. medium, 95% CI = 0.23–0.73, FDR = 0.09; Supplementary Table 1, Fig. 2A). No oral bacteria were found to be independently associated with CRC survival (Supplementary Table 2).

Fig. 2: Identification of prognostic bacterial biomarkers associated with PFS in colorectal cancer patients.
figure 2

A Kaplan–Meier curve illustrating PFS stratified by the detection status of identified oral bacterial species. P values were calculated using the log-rank test. Numbers below each graph indicate the number of patients at risk at different time points. ND not detected, Det detected, PFS progression-free survival. B Volcano plot showing the hazard ratio (HR) and statistical significance of candidate prognostic bacterial species. The x axis represents log2-transformed HR, while the y axis shows −log10-transformed false discovery rates (FDR). Circle size indicates the frequency of significance in 1000 cross-validation tests using univariate Cox regression, and color intensity (gray to red) represents the frequency of significance in multivariate Cox regression analysis. The multivariate Cox regression model was adjusted for age, sex, tumor stage, and neoadjuvant chemoradiotherapy.

Full size image

To identify robust oral microbial biomarkers, Monte Carlo simulations were utilized for cross-validation across these 98 oral microbial species. The univariate and multivariate Cox regression model was conducted within both training and validation datasets by random sampling 1000 times. Consequently, T. medium, C. gracilis, and N. oralis exhibited the most significant statistical association in both the training and validation datasets, confirming that these three oral bacteria are robust prognostic biomarkers for CRC (Fig. 2B).

Construction of the oral MRSs as accurate predictors of CRC prognosis

To evaluate the performance of oral microbial biomarkers in predicting CRC prognosis, we developed every combination of oral MRS by integrating one, two, or all of the three above-mentioned prognostic bacteria. Using a Cox regression model, we applied these combinations of MRS to 1000 test datasets. Among all combinations, the MRS including three specific oral bacteria demonstrated a significant association with CRC prognosis, exhibiting the highest C-index (C-index = 0.68, 95% CI = 0.61–0.76). (Fig. 3A)

Fig. 3: The performance of microbial risk scores, clinical model, and comprehensive model.
figure 3

A Comparison of the predictive performance (C-index) of all combinations of microbial risk scores derived from the three identified prognostic species in the test datasets using the Cox regression model. B Kaplan–Meier curves for PFS stratified by microbial risk scores (MRS) into low (n = 38), moderate (n = 150), and high-risk groups (n = 124). Numbers below the graph indicate patients at risk at different time points. The P value is calculated by the log-rank test. C Kaplan–Meier curves for overall survival stratified by the MRS. The P value is calculated by the log-rank test. D Comparison of the predictive performance (C-index) of all combinations of clinical factors (perineural invasion, lymphatic metastasis, and tumor stage) in the test datasets using the Cox regression model. E Performance comparison of the clinical model, MRS model, and comprehensive model (integrating both clinical and microbial factors) in the test datasets using the Cox regression model. The P value for the comparison between the clinical model and the comprehensive model was calculated using a Z score test. The bars represent the 95% confidence intervals calculated by the bootstrap method. F Performance comparison of the clinical model, MRS model, and comprehensive model in the test datasets using the random survival forest (Rsf) method. The P value for the comparison between the clinical model and the comprehensive model was calculated using a Z score test. The bars represent the 95% confidence intervals calculated by bootstrap method. T. mediumTreponema medium,N. oralisNeisseria oralis, C. gralisCampylobacter gralis, Mod moderate, Rsf random survival forest.

Full size image

To better differentiate patient prognostic risks, we reclassified the MRS into three categories: 0 for MRS Low, 1 for MRS Moderate, and 2–3 for MRS-High risk. This stratification effectively distinguished progression risks among patients in addition to overall survival (Fig. 3B, C). Patients in the MRS-Low group, only 4 out of 38 patients exhibited progression (10.5%), compared to 20 out of 150 patients in the MRS-Moderate group (13.3%), and 35 out of 124 patients in the MRS-High group (28.2%) experienced disease progression post-surgery.

We also identified the clinical factors associated with the progression of CRC. Consistent with previous researches, tumor stage, perineural invasion, lymphatic metastasis were significantly associated with the progression of CRC in log-rank test, and were identified as clinical prognostic biomarkers (P < 0.05, Supplementary Fig. 2). Factors such as age, marital status, cigarette smoking, alcohol consumption history, histologic grade and lymphatic invasion did not show the significant association with cancer progression (Supplementary Fig. 2). We incrementally incorporated these three clinical factors into the model, and the one containing all three demonstrated a higher C-index compared to other clinical models in the 1000 test datasets (C-index: 0.78, 95% CI = 0.71 – 0.83) (Fig. 3D).

Furthermore, we integrated these clinical factors and oral MRSs to construct a combined model. We randomly partitioned the entire dataset into training and test datasets (1:1). In the training datasets, the C-index for the combined model achieved a C-index of 0.84 (95% CI: 0.75 ~ 0.91), and in the test datasets, the C-index for the combined model was 0.77 (95% CI: 0.66 ~ 0.86). The combined model exhibited a higher C-index than the clinical model 0.77 vs. 0.67, P = 2.33 × 10−5, with similar results observed in the train datasets and total population datasets 0.84 vs. 0.76, Ptraining datasets = 0.02, 0.81 vs. 0.72, Ptotal population = 3.25 × 10−5 (Fig. 3E, Table 2). To further assess the predictive performance of our oral MRS for CRC progression, we constructed a clinical model, an MRS model, and a combined model based on a random forest survival model. Consistently, the combined model outperformed the clinical model in both test and total population datasets, suggesting the stable predictive power of the MRS (Fig. 3F, Table 2).

Table 2 The C-index values of the clinical model, the MRS model, and the comprehensive model for predicting progression-free survival probability in the train set, test set, and the overall population
Full size table

Correlation between oral MRSs and predictive functional pathways

To investigate the functional roles of oral microbiota among patients with varying oral microbial risks, we used the PICRUSt2 tool to explore the potential metabolic functions of oral microbiota based on 16S rRNA gene sequences. Among the 344 identified KEGG pathways, we excluded those that appeared in fewer than 30% of patients and with an average relative abundance of below 1%, ultimately incorporating 282 metabolic pathways. Patients were categorized into two groups: MRS-low and MRS-moderate/high groups. Sixteen pathways exhibited significant differences between these two groups, with five pathways significantly increased in the MRS-moderate/high group. These pathways are predominantly associated with the promotion of cancer cell proliferation, specifically the Super-pathway of Polyamine Biosynthesis II and Polyamine Biosynthesis, which are linked to polyamine synthesis-an enhancement of cellular proliferation typically observed in cancer cells exhibiting active polyamine metabolism (Fig. 4B). Among the 11 metabolic pathways significantly enriched in the MRS-low group, the most difference in mean proportions was observed in the N-acetylglucosamine (GlcNAc) and N-Acetylgalactosamine pathways (Fig. 4B). This pathway is associated with the synthesis of GlcNAc and GalNAc, both of which can mitigate inflammatory responses, regulate glycosylation, and inhibit cancer cell proliferation. Subsequently, we analyzed the correlation between the differential metabolic pathways and identified oral microbial indicators. We found that species associated with an increased risk of CRC progression exhibited similar functional characteristics but differed markedly from those associated with a reduced risk of CRC progression. For instance, C. gracilis and N. oralis demonstrated a positive correlation with the increased KEGG pathways in the MRS-moderate/high group and a negative correlation with the depleted pathways. In contrast, the bacterium T. medium, which was enriched in the MRS-low group, displayed an opposite trend (Fig. 4A).

Fig. 4: The significantly different functional pathways between MRS-Low and MRS-moderate/high patients and their associations with identified prognostic species.
figure 4

A Heatmap of Spearman correlation between prognostic species and differential pathways. The KEGG pathways were predicted using PICRUSt2 to infer the functional shifts in the microbiota of MRS-Low and MRS-moderate/high patients. The strength of the color depicted Spearman’s correlation coefficients (negative correlation, blue; positive correlation, red). (*P < 0.05). B The differential analysis of the oral microbial community’s functions between MRS-moderate/high and MRS-low groups, using STAMP software. The analysis focused on KEGG pathways with an average relative abundance exceeding 1%, and the P value is corrected using the Bonferroni method. The difference in mean proportion for pathways showing significant differences in abundance was shown. The 95% confidence intervals and statistical significance (P value corrected) were indicated as well.

Full size image

Discussion

In this study, we systematically investigated the relationship between oral microbiota and CRC prognosis using full-length 16S rRNA sequencing data from saliva. We identified three prognostic biomarkers for CRC: C. gracilis, N. oralis, and T. medium. Additionally, we developed a prognostic prediction model based on the oral microbiota, which effectively predicts the prognosis of CRC patients. Given the high recurrence rates of CRC post-surgery, it is essential to identify suitable biomarkers for risk assessment18,19.

Although many studies have explored the microbiota-based prognosis biomarkers for CRC, the majority of previous research has predominantly concentrated on gut microbiota, which is usually tested by a fecal sample20,21,22. The collection and analysis of saliva samples are simpler, more cost-effective, and demonstrate better compliance than fecal samples8,23,24. Prognostic models based on fecal microbiota identified the enterotypical Prevotella and three microbial biomarkers for predicting clinical outcomes of CRC, achieving a C-index of 0.6925. Microbial signatures in tumor tissues have also demonstrated potential to refine prognostication in CRC patients26. Evidence of oral microbiota as prognostic indicators for CRC remains limited. In this study, we identified three oral bacteria associated with CRC prognosis based on follow-up data from patients, and constructed an oral MRS for predicting CRC prognosis. This approach enables the qualitative detection of the target species through a simple and straightforward PCR method, potentially providing advantages for monitoring the postoperative progression risk for CRC patients.

Previous epidemiological studies have established a significant association between oral diseases and CRC. A cohort study of over 700,000 individuals reported a significant correlation between clinically assessed periodontitis and CRC (HR = 1.13, 95% CI: 1.03–1.24)27. Similar conclusions were drawn in another study on oral health, which found that women with moderate to severe periodontal disease exhibited a modest increased risk of CRC (HR = 1.22, 95% CI: 0.91–1.63)28. These findings suggest that poor oral health may, to a certain extent, facilitate the onset and progression of CRC. Subsequently, numerous studies have begun to focus on the role of oral bacteria in the occurrence of CRC. Zhang S et al.29 have evaluated the association between oral microbiota dysbiosis and CRC, and constructed a diagnostic model based on oral microbial markers, which can effectively distinguish between adenoma (AUC = 0.94) or CRC (AUC = 0.83) and healthy controls. Another study also found the value of oral microbiota as biomarkers for the detection of CRC, and a classification model containing oral microbiota can effectively distinguish CRC patients (AUC = 0.90), polyps (AUC = 0.89), and healthy controls8. These findings not only illuminate the connection between oral health and the risk of CRC but also underscore the potential of the oral microbiota as a diagnostic biomarker for CRC. Therefore, leveraging the oral microbiota as a biomarker for prognostic evaluation of CRC may represent an avenue for future research and clinical practice19,30,31.

Emerging evidence suggests that oral microbiota may contribute to the development of CRC through several mechanisms. First, the oral-gut microbial translocation axis facilitates direct pathogen infiltration. The oral cavity acts as a reservoir of gut microorganisms, daily swallowing introduces substantial bacterial loads to the gastrointestinal tract. Previous studies have reported an increased oral-gut transmission in CRC patients32. Additionally, animal models demonstrated that buccal F. nucleatum can migrate to the CRC locus and impair the therapeutic efficacy and prognosis of radiotherapy33. Second, certain oral pathogens can enter the bloodstream. Bacteremia caused by oral microbes, such as Peptostreptococcus and Gemella, has been linked to an increased risk of CRC34. Furthermore, oral dysbiosis modulates gut immunoinflammatory responses via the gut-mucosal immune interconnection. Microbial dysbiosis in the oral cavity can exacerbate inflammation in the gut by regulating host immunity, specifically through the introduction of colitogenic pathobionts and pathogenic Th17 cells into the gut35. These multilayered interactions underscore the importance of oral microbiota balance in the development of CRC.

Our study identified N. oralis and C. gracilis as predictors of poor CRC prognosis. In a prospective cohort of 793 individuals, elevated oral abundances of N. oralis (OR = 1.42, 95% CI = 1.01–2.00) and Campylobacter spp. (OR = 1.58, 95% CI = 1.12–2.24) were significantly associated with CRC risk36. C. gracilis demonstrates oral-gut translocation capacity, with gut colonization evidenced in inflammatory bowel disease and Crohn’s disease biopsies37,38. C. gracilis is also known as a pathobiont of dental diseases and periodontitis39,40. Periodontitis-driven systemic inflammation may contribute to CRC development (e.g., Th17 cell activation) by disrupting gut barrier integrity and amplifying microbial dysbiosis35. N. oralis typically appears in the oral cavity. One plausible mechanism is that N. oralis can convert ethanol into acetaldehyde in the oral cavity41. Acetaldehyde, a well-established carcinogen, is known for its substantial toxicity and mutagenicity39. These findings elucidate the mechanistic roles of these oral pathobionts in mediating unfavorable CRC prognosis through oral-gut axis translocation, systemic inflammation, and carcinogen production.

Our findings indicate that T. medium is associated with a reduced risk of CRC progression. T. medium, a member of the genus Treponema, is distinct from the notorious Treponema denticola, and research on T. medium remains limited. Although this bacterium has been identified in gingival plaques, there is still no consensus regarding its pathogenicity42. Our findings indicate that T. medium is associated with a reduced risk of CRC progression. The mechanisms by which T. medium may contribute to reducing CRC progression risk warrant further exploration in future studies.

Other oral microbes, such as F. nucleatum43, P. intermedia44, Rothia45, and P. gingivalis11, have also been identified as the higher abundance microbes enriched in the oral cavity of CRC patients. The well-known oral-derived microbes, F. nucleatum, have also been found in the feces and tumor tissues in CRC patients, increasing the risk of CRC by inducing inflammation, disrupting the gut microbiota, and exacerbating metabolic disorders. Its critical virulence factor, RadD, facilitates the attachment to CRC cells and plays a vital role in promoting tumorigenesis46. Moreover, F. nucleatum can effectively bind to host cells through its virulence factor, adhesin FadA. This interaction enhances bacterial adhesion, disrupts the intestinal barrier47,48. Besides, F.nucleatum can induce immune suppressive effect by promoting M2 polarization of macrophages via the TLR4/IL-6/PSTAT3/c-MYC signaling pathway49. These pathogen–host interactions are involved in the carcinogenesis of CRC. In this study, we did not observe the significant associations between these microbes and CRC progression, which can partly be explained by the fact that the molecular events driving carcinogenesis and the biological processes driving disease progression (e.g., metastasis, therapeutic resistance) frequently involve distinct biological mechanisms. Our current prognostic analyses are confined to CRC patient cohorts where oral microbial profiles inherently exhibit baseline homogeneity due to their shared carcinogenesis-associated signatures. Consequently, bacterial taxa already enriched in CRC populations (e.g., F.nucleatum) may demonstrate attenuated prognostic discriminative power.

The oral microbiota plays an important role in the biosynthesis and metabolic degradation of amino acids and carbohydrates50. Our study revealed that those bacteria, enriched in high oral microbial risk patients, are positively correlating with metabolic pathways, which are crucial for cellular proliferation, gene expression, and growth51. These bacteria had opposite functions to those associated with a reduced risk of CRC progression. Specifically, we identified N. oralis and C. gracilis, positively correlating with metabolic pathways involved in the synthesis of arginine and proline, phospholipids, and polyamines. Polyamines, such as putrescine, spermidine, and spermine, play an important role in regulating the cell cycle and apoptosis52. Therefore, oral microbiota may facilitate cancer growth and metastasis by increasing the biosynthesis of polyamines, providing cancer cells with essential polyamines. In contrast, the analysis of metabolic pathways revealed that T. medium was positively correlated with several metabolic routes, including beta-alanine metabolism, L-histidine biosynthesis and degradation, serine and glycine metabolism, and the non-mevalonate pathway. Interestingly, serine and glycine metabolism, as well as the non-mevalonate pathway, are closely associated with the synthesis of the antioxidant glutathione, while the beta-alanine pathway is considered a precursor for the synthesis of carnosine, which functions as an intracellular antioxidant and pH buffer. Oxidative stress is closely associated with cancer development, resulting in cellular damage and DNA mutations51.

In this study, several limitations should be considered. Firstly, this study primarily utilized a single-center design and lacks an external validation cohort, which restricts the generalizability of our predictive model, necessitating future validation through multi-center studies with independent cohorts. Nevertheless, we conducted a series of sensitivity analyses to ensure the robustness of our results, including randomization of the sample division and different model-constructed approaches. Additionally, the relatively small number of progression events (n = 59) may affect the statistical power of our findings. Future prospective investigations incorporating multi-center designs, expanded sample sizes, and extended follow-up periods are required to reliably identify CRC prognostic biomarkers. Finally, while our analysis suggests potential microbial metabolic pathway involvement, the proposed associations between identified microbes, microbial functional metabolic pathways, and clinical outcomes require experimental validation through in vitro and in vivo studies to elucidate the mechanisms underlying the influence of oral microbiota on CRC progression.

Besides, we acknowledge several methodological constraints of this study. First, while implementing enhanced mechanical lysis, quantitative assessment of microbial lysis efficiency was not conducted, leaving potential Gram-positive bacterial under-lysis unverified. Second, the inherent primer bias of 16S rRNA sequencing may skew microbial composition, which has the potential to be biased towards amplifying specific bacteria. As a high percentage of host genomic DNA is present in oral specimens, 16S rRNA sequencing is widely used in oral microbiota research due to its cost-effectiveness, while metagenomic sequencing data are necessary for more precise information on taxonomic composition. Finally, inferred metabolic pathways through PICRUSt2 disregard strain-level heterogeneity. Future studies should integrate shotgun metagenomics with metabolomic profiling to establish host-microbiota interactions.

In summary, our study suggests potential associations between N. oralis, C. gracilis, and T. medium with the risk of postoperative progression in CRC. By integrating clinical and microbial factors, we achieved improved predictive accuracy for progression risk compared to clinical factors alone. These findings collectively provide preliminary evidence for exploring oral microbiota-based biomarkers for CRC progression, the implementation of oral microbial indicators in CRC clinical surveillance required future confirmatory studies.

Methods

Participants and sample collection

From December 2018 to April 2021, we recruited 361 patients diagnosed with stage I to IV colorectal adenocarcinoma at Sun Yat-sen University Cancer Center. All patients underwent elective surgical resection, and saliva samples were collected preoperatively. Forty-nine patients were excluded as follows: (1) 42 patients were excluded due to postoperative sampling; and (2) 7 patients were excluded because their sequencing data did not meet quality control standards. Ultimately, 312 patients with colorectal adenocarcinoma were included in this analysis. Institutional review board approval was obtained from the Human Ethics Committee of Sun Yat-sen University Cancer Center (The approved number: B2024-542-01), and informed consent was obtained from all patients participating. The study data have been collected in accordance with the Declaration of Helsinki. This study followed the Standards for Reporting of Diagnostic Accuracy (STARD) reporting guideline (Supplementary Table 3).

All patients were collected the clinical information, including sex, age, marital status, body mass index (BMI), cigarette smoking history, alcohol consumption history, family history of tumors, neoadjuvant chemoradiotherapy, tumor stage, histologic grade, lymphatic invasion, perineural invasion, and lymphatic metastasis. During the study registration period, the saliva samples were collected from patients who were instructed to refrain from eating and drinking for at least 30 min before the collection process. The participants were guided to open the collection tubes and let saliva flow into them effortlessly, which was collected using sterile 50 ml centrifuge tubes. The collected saliva was temporarily stored in an ice box and transferred to 2 ml centrifuge tubes within two hours, and subsequently stored at −80 °C for long-term preservation.

DNA extraction and library construction

In this study, saliva samples from patients were extracted using the DNeasy PowerSoil Pro Kit (QIAGEN, Germany). Saliva specimens were subject to mechanical lysis through bead-beating processing (four cycles of 5-minute agitation at ~2700 rpm for 50 Hz) using the Vortex-Genie 2 system (Scientific Industries) with zirconium dioxide and yttrium oxide beads. The construction of the 16S rRNA amplicon library was described previously53,54. Specifically, DNA extracted from each salivary sample served as the template for amplifying the full-length region of the 16S rRNA gene. We amplified the full-length 16S rRNA gene using universal primers 27 F (5’-AGR GTT YGA TYM TGG CTC AG-3’) and 1492 R (5’-RGY TAC CTT GTT ACG ACT T-3’), which included 12-bp barcodes. Saliva-derived DNA was amplified using KAPA HiFi HotStart DNA Polymerase (KAPA Biosystems) for 27 cycles. The process involved denaturation at 95 °C for 30 s, annealing at 57 °C for 30 s, and extension at 72 °C for 1 min. We used Agencourt AMPure XP (Beckman Coulter) for fragment selection and purification of the PCR products. We prepared the SMRTbell libraries from the purified amplicons by ligating adapters and sequenced them using the PacBio Sequel platform (Pacific Biosciences). We obtained high-quality circular consensus sequence (CCS) reads from the raw PacBio sequencing data using SMRT Link software (v9.0.0, Pacific Biosciences) and assigned multiplexed libraries to each sample using Lima (v2.0.0) based on the barcodes.

Analysis of 16S rRNA full-length region sequencing data

We employed a customized DADA2 (v1.22.0) workflow for the PacBio full-length 16S sequencing data to conduct quality control, denoising, and identification of amplicon sequence variants (ASVs) from the demultiplexed CCS, ultimately yielding representative sequences and an abundance matrix derived from ASVs55. ASVs detected in fewer than five samples or with a total sequence count of fewer than ten were excluded. The “phyloseq” package in R (version 4.4.0) was employed to standardize the sequencing depth of each sample to 250056. Species annotation of ASVs was performed using the pre-trained SILVA database (version 138), categorizing species information across six taxonomic levels: phylum, class, order, family, genus, and species, while calculating the relative abundance of various taxa at different classification levels57. Alpha diversity index and beta diversity metrics were computed using the R package “vegan”58.

Statistical analysis

The demographic, socioeconomic, and clinical characteristics of patients were described and compared among different clinical outcomes using the Pearson χ2 test for categorical variables, and one-way ANOVA for continuous variables. Permutational multivariate analysis of variance was applied to evaluate the statistical significance of differences in microbial composition across various clinical factors59. The C-indexes between the clinical models and comprehensive models were compared by using the Z score test60. All statistical analyses and visualization procedures were performed using R software (version 4.4.0) with designated R packages.

Survival analysis

Progression-free survival (PFS) was defined as the period from surgical removal to the recurrence or progression of CRC, or death from any cause. Overall survival was defined as the period from the surgical period to death from any cause. To investigate potential prognostic factors for the CRC, we used Kaplan–Meier survival analysis and the Log-rank test to compare the relationships between various factors and prognosis, converting all variables into categorical forms, such as sex, marital status, BMI, family history of tumor, alcohol drinking history, cigarette smoking history, neoadjuvant chemoradiotherapy, tumor stage, histologic grade, lymphatic invasion, perineural invasion, lymphatic metastasis. Furthermore, we calculated the HRs and their corresponding 95% CIs for different factors influencing CRC prognosis using the Cox proportional hazards regression model. These analyses were performed using the “survival” and “survminer” R packages. Regarding the sample size for PFS analysis, the comparison between the number of progression events59 and the number of prognostic factors used in the multivariate Cox proportional hazards model indicates that the “minimum of 10 events per predictor” rule was satisfied61.

Identification of prognostic oral bacteria

A total of 291 bacterial species were observed, of which 98 were included in the prognosis analysis, which met criteria with a frequency exceeding 10% in the total sample and a relative abundance >0.01%. Patients were divided into two groups based on the median abundance of each species, and a univariate Cox regression model was performed. Species that were statistically significant were subsequently incorporated into a multivariate Cox regression model, adjusted for sex, age, tumor stage, and neoadjuvant chemoradiotherapy.

Next, we applied Monte Carlo simulation cross-validation for a robust selection, by randomly partitioning the entire datasets into discovery and validation sets (1:1) and performing 1000 times randomized resampling. In the discovery set, we performed univariate and multivariate Cox regression models to identify the species associated with prognosis (P < 0.05). Species with significance in the discovery stage were then included in the validation set for confirmation. Ultimately, three bacterial species were identified as prognostic biomarkers, demonstrating substantial effect sizes (HR > 1.5 or HR < 0.8) and significance (corrected FDR < 0.1).

Construction of MRSs and prognostic prediction models

Three validated oral bacterial species significantly associated with CRC prognosis were included to construct an MRS. These species were C. gracilis, N. oralis, both associated with poor prognosis, and T. medium, which was conversely associated with good prognosis. The presence of species indicating poor prognosis or the absence of species indicating good prognosis was scored as one point each. Each patient was assigned an MRS ranging from 0 to 3. To evaluate the performance of the combinations, 50% of the samples (156 out of 312) were randomly selected 1000 times without replacement to create a pool of test datasets, and the concordance index (C-index) across the test datasets was used as an indicator of model stability. For clear stratification, a three-layered risk score (MRS), which included the above three bacteria, was categorized by assigning scores 0 for MRS low, 1 for MRS moderate, and 2–3 for MRS high risk. Furthermore, to comprehensively predict the probability of PFS in CRC patients. We constructed a clinical model based on selected clinical factors, including tumor stage, lymphatic invasion, and perineural invasion, all of which were statistically significant in a univariate Log-rank test. A comprehensive model was constructed by integrating both clinical factors and MRS. We calculated the concordance index (C-index) to evaluate the performance of the models and compared the differences in predictive performance between models incorporating and excluding MRS.

Metagenomic PICRUSt analysis

We applied the “Phylogenetic Investigation of Communities by Reconstruction of Unobserved States” (PICRUSt2) tool to infer functional variations within microbial communities62. Subsequently, we utilized STAMP software to evaluate significant differential metabolic pathways between moderate/high and low MRS groups, applying the Bonferroni correction with adjusted p value < 0.05 considered significant. The correlation between oral microbes and the significantly related differential pathways was performed by Spearman’s rank correlation test.