Colorectal cancer (CRC) is the third most prevalent cancer, which ranks the second in terms of mortality globally (Arnold et al., 2017; Cai and Liu, 2021). Although the incidence and mortality rates of CRC are decreasing in recent years, CRC is still one of the most life-threatening cancers and advanced CRC remains an incurable disease at metastatic stages, 5-year survival rate of 13% compared to 90% when diagnosed at an initial stage (Bray et al., 2018; Montalban-Arques and Scharl, 2019). The trend in younger patients, along with the continued burden, highlight the need for early detection of CRC that alleviates the incidence of metastatic CRC (Cheng et al., 2020). In particular, the clinical trials of patients are intimately associated with the CRC stages, thus there is an urgent requirement for accurately diagnosing the stages of CRC.

Over the past decades, the Tumour, Nodes, and Metastasis (TNM) staging system has contributed the cornerstone for the management of CRC patients. However, some problems have occurred with the TNM system, such as increasing complexity of CRC, poor clinical evidence, tumor deposits, and un-reproducibility between different observers (Quirke et al., 2007), which in turn confuse the accuracy of identified CRC stage and subsequent therapy. It is now widely recognized that the microbes contribute indispensable roles in host health and gastrointestinal tumor progression (Marchesi et al., 2016; Xiong, 2018; Guo et al., 2022). In accordance, intensive studies have shown dysbiosis (shift in gut commensal microbiota toward opportunistic pathogens) in the gut microbiota in CRC patients compared with healthy controls. Going forward, CRC-associated taxa have been identified for distinguishing healthy from colorectal adenomas, and CRC individuals (Shah et al., 2018; Wu et al., 2021; Coker et al., 2022). However, few studies have explored the dynamics of microbial communities over CRC progression. As a result, it is uncertain whether microbial taxa are indicative of each CRC stage, rather than just stratify case from control cohorts. However, this information is fundamental to establish CRC stage-dependent clinical trials.

Accumulating evidence depicts that the gut microbiota is an important etiological element in CRC initiation, progression, and metastasis (Kong et al., 2019; Cheng et al., 2020; Mizutani et al., 2020; Li et al., 2022). By this logic, identification of bacteria involved in CRC progression could provide new targets for CRC diagnosis and prevention (Montalban-Arques and Scharl, 2019; Fong et al., 2020). Indeed, case studies have proposed that the occurrence of CRC is attributed to the enrichment of Fusobacterium nucleatum (encoding FadA) (Flanagan et al., 2014), Bacteroides fragilis and Escherichia coli hosting polyketide synthase (pks) islands (Feng et al., 2015), Clostridium symbiosum (Xie et al., 2017), or Parvimonas micra (Löwenmark et al., 2020), respectively. It should be noted that the identification of “driver” taxa is generally based on their significant enrichment in CRC patients compared to healthy controls, which ignores the CRC stage (Flanagan et al., 2014; Löwenmark et al., 2020). In particular, a “driver” taxon is attributed to strong biotic interactions, rather than its sheer abundances, though this does not rule out that some “driver” taxa are numerically abundant (Dai et al., 2018). Recently, NetShift approach has been developed to quantify major changes in associations of each constituent taxon between healthy and diseased networks, in which the importance of a single species between health states can be calculated based on its topological features (Kuntal et al., 2019). By this logic, taxa that increase in their importance in the network of CRC patients could be the “driver” taxa, moving beyond enriched abundance. Furthermore, according to the “driver-passenger” model, the CRC driver taxa could be superseded by “passenger” bacteria that are better adapted to the conditions in and around carcinoma cells, thereby outcompeting the initial driver species (Feng et al., 2015; Bridges et al., 2019). For these reasons, a systematic analysis of CRC-associated bacteria along CRC progression is required from an ecological prospective, rather than case-control study.

In spite of a growing body of evidence with regard to the dysbiosis in gut microbiota in CRC patients, data on the association between in situ tumor microbial dynamics over CRC progression is lacking. The gut microbiotas are significantly varied as over human lifetime (Falony et al., 2016; Greenhalgh et al., 2016; Ghosh et al., 2020), thus a key challenge is to distinguish the alterations in microbial assembly over CRC progression from these as human aged. Additionally, faecal microbiota only partially mirrors mucosal microbiota in CRC, with low correlations between paired faecal and mucosal samples (Flemer et al., 2017). In this regard, the deployment of faecal microbes for mirroring tumor microbiota could bias the identification of CRC-associated bacteria. To overcome above obstacles, we explored the dynamics of bacterial communities in tumor tissues along the four CRC stages, and benign polyps as controls. We attempted to address the following concerns: (1) exploring the dynamics of bacterial communities over CRC progression, (2) identifying CRC-associated bacteria based on biotic interactions, (3) screening biomarkers for diagnosing each CRC stage, irrespective of host age.

Currently there has been increased interest in the adenoma-specific markers that detect early CRC, partly due to the recognition that the bacterial communities are distinct along the control-adenoma-carcinoma sequence (Mizutani et al., 2020; Wu et al., 2021). By contrast, few studies have explored the bacterial communities over CRC progression, resulting in the uncertainty whether biomarkers could diagnose the four stages of CRC, rather than case-control. Additionally, as microbes are implicated in colorectal carcinogenesis, development and treatment outcome (Jin et al., 2019; Montalban-Arques and Scharl, 2019; Mizutani et al., 2020; Ting et al., 2022), understanding the dynamics in CRC development is a necessary initial step to developing a more complete understanding of both the ecology and etiology.

Increasingly evidence has shown that faecal bacterial communities are distinguishable from individuals with CRC or adenomas to controls (Wang et al., 2012; Flemer et al., 2018; Wu et al., 2021). In contrast, there is still lack of data about the bacterial profiles over CRC progression, especially these at cancerous tissue. Considering the functional importance of microbes in promoting CRC tumorigenesis (Cheng et al., 2020; Mizutani et al., 2020), and given that faecal microbiota only partially reflects mucosal counterpart in CRC (Flemer et al., 2017), this ignorance could bias the identification of CRC associated taxa. Trying to comprehensively collect the tumors comprising the four CRC stages, our results depicted that there were distinct segregations in bacterial community structure between polyps and tumors, and along the four CRC stages ( Figure 1 , Table S2 ). These patterns suggest that, despite the extensive physical variances among individuals (e.g., age, gender, and CEA level, Table S1 and Figure S1 ), each stage exerts sufficiently unique conditions to assemble communities that are consistent in structure according to BP or CRC stage.

As the stability of the microbiota could be affected by host disease (Xiong, 2018; Aho et al., 2019), we compared with the stability of bacterial communities over CRC progression. CRC initiation (T1 stage) sharply disrupted the stability of bacterial community compared with BPs ( Figure 2 ). There are several possible explanations for this pattern. First, the bloom of pathogenic taxa outcompete resident commensals, as supported by increased abundances of known pathogens in T1 tumors, such as Fusobacterium nucleatum, Bacteroides fragilis, and Escherichia fergusonii ( Figure S6 ). Second, pathogen invasion could attenuate host filtering on the colonization of external taxa, leading to a dominance of stochastic processes (e.g., random birth and death) governing community assembly (Mallon et al., 2015; Adair and Douglas, 2017). Accordingly, there was a heterogeneous bacterial community (higher AVD value) in T1 tumors ( Figure 2 ). However, as disease severity increased, inflammatory microenvironment could exert the dominant role of homogeneous selection, resulting in a convergent, disease-like microbial community (Subramanian et al., 2014; Xiong et al., 2017). Consistent with this assertion, the AVD values linearly decreased along CRC progression ( Figure 2 ), whereas phylogenetic diversity exhibited an opposing trend ( Figure S4B ). Given that microbes are implicated in the outcome of CRC therapy (Montalban-Arques and Scharl, 2019; Mizutani et al., 2020; Ting et al., 2022), the convergence in tumor bacterial community may partially explain why advanced CRC is difficult to be curative.

It has proposed a stratification of individuals into three distinct enterotypes (Arumugam et al., 2011), while others supported a concept of stratification based on bacterial abundance gradients (Flemer et al., 2017). We detected a stratification of individuals between distinct enterotypes and abundance gradients, as supported by distinct structures of bacterial communities when integrated the abundance of ASVs, e.g., abundance gradients ( Figure 1 , Table S2 ). However, we also found rapid replacement of ASVs over CRC progression, because there were linearly deceased numbers and low proportions of overlapped ASVs between BP and advanced CRC, as well as shuffling between adjacent two stages, e.g., distinct enterotypes ( Figure 3A ). This is apparent in the SPEC-OCCU plots ( Figure 3B ), where the majority of the ASVs exhibited low occupancy, indicating that few of them are consistently detected among individuals. In accordance, the tumor communities harbored significantly lower homogeneity among individuals in CRC T1, T2 and T3 stages, compared with these in BP ( Figure 2 ). These findings indicate that the tumor environments are less selective for a stable core community, resulting in heterogeneity in bacterial communities among patients. Considering that microbes determine the therapeutic efficiency of CRC, the heterogeneity in bacterial communities may guide the design of personalized medicine (Kong et al., 2019; Shi et al., 2020; Ting et al., 2022). Although this study was not designed to evaluate treatment outcome, clearly, there is a pressing need for longitudinal study exploring the associations between the tumor microbiota and treatment response in CRC patients.

Given that the gut microbiome is an important etiological element in the initiation and progression of CRC (Montalban-Arques and Scharl, 2019; Cheng et al., 2020; Mizutani et al., 2020; Guo et al., 2022), sufficient and accurate identification of CRC associated taxa could facilitate the targets for diagnosis and therapy. We identified the well known promoters in colorectal carcinogenesis, Fusobacterium nucleatum (Rubinstein et al., 2013), based on its sharply increased abundance in T1 stage tumors ( Figure 5A ). Similarly, three well known CRC-associated taxa, F. nucleatum, Bacteroides fragilis, and Campylobacter concisus (Guo et al., 2022), were enriched in the advanced T4 stage tumors ( Figure 5B ). Thus, there is a lack of consistency in the bacterial taxa associated with CRC progression, in accordance with studies conducted previously (Feng et al., 2015; Mizutani et al., 2020). This pattern supports the so-called “driver-passenger” model, which proposes that different bacteria sequentially implicate in CRC tumor initiation and progression. The “driver” bacteria are replaced by “passenger” bacteria that are better adapted to the conditions in and around cancerous cells (Tjalsma et al., 2012; Tilg et al., 2018). Of note, other bloomed ASVs have not been reported to be associated with CRC, instead, a few ASVs could be potential probiotics, such as Corynebacterium vitaeruminis (Colombo et al., 2017), and Streptococcus alactolyticus (Zhang et al., 2021). In this regard, it is cautious to identify CRC-associated taxa by increased abundance in the tumors. Going forward, we identified 13 “driver” bacterial ASVs by the NetShift model, moving beyond differential abundance, of which 6 ASVs were significantly enriched and the most abundant in T1 tumors ( Figure 4 ). Among the 6 ASVs, ASV1839 Clostridium saudiense is recently identified as an opportunistic pathogen with the potential to cause hepatocellular carcinoma, which translocates from the gastrointestinal system to biliary system (Yoon et al., 2022). Unexpectedly, ASV1414 Bifidobacterium stercoris and ASV2467 Parabacteroides distasonis were significantly enriched in T1 tumors ( Figure 4B ). B. stercoris is able to produce acetic acid and lactic acid, and promote antitumor immunity (Sivan et al., 2015), while the later could attenuate toll-like receptor 4 signaling and thus blocks colon tumor formation (Koh et al., 2018). One possible explanation for this counterintuitive pattern is that the hosts release specific chemicals that favor the recruitment of beneficial microbes or of antagonists able to suppress the growth of pathogens, according to the “cry-for-help” hypothesis (Rolfe et al., 2019), although further study is needed to validate its applicability in human diseases. Conversely, the butyrate-producing AVS1014 Faecalibacterium prausnitzii is an anti-inflammatory commensal bacterium (Miquel et al., 2013), thus its reduction is expected to be associated with CRC. Consistent with the above, the ratio of Fusobacterium nucleatum to Faecalibacterium prausnitzii has been identified as a valuable biomarker for screening early CRC (Guo et al., 2018). The beneficial ASVs decreased rapidly along advanced CRC ( Figures 4 , S5 ), which raises the intriguing possibility that tumors could recruit beneficial taxa to antagonize CRC-associated pathogens at the initial stage. Collectively, the taxa that trigger protumorigenic environments could be attributed to the enrichment of pathogenic strains, and also the depletion of probiotic species.

The establishment of gut microbiota—host health relationship and modeling algorithms facilitates the identification of bio-indicators diagnosing disease severity, which is a key goal of microbiome research. Multiple studies have extensively shown human age to be a strong covariate of the gut microbiota (Falony et al., 2016; Greenhalgh et al., 2016), thus host age may overshadow changes in commensals associated with CRC. For this reason, we teased apart the effect of ageing-related and CRC-related changes in the bacterial community. To achieve this, we identified age-discriminatory ASVs that featured the age of BP controls. The top nine age-discriminatory ASVs contributed a high consistency (r = 0.924, P < 0.001) between diagnosed and chronologic ages ( Figure S6 ). After removal of the nine age-discriminatory ASVs, we further identified CRC stage-discriminatory taxa for diagnosing BPs and the four CRC stages, with an overall 87.4% accuracy. It is worthy to emphasize that this optimization substantially improves the performance of our diagnosis model compared with the model neglecting age effect (71.6% accuracy, Table S5 ). As a result, the diagnosis accuracy was not affected by host age ( Figure S8 ). Our diagnosis model could add the designation of CRC stage-dependent clinical trials, especially the patients with poor pathological feature and un-reproducibility between two doctors (Quirke et al., 2007). Similarly and more importantly, no CRC patients were falsely diagnosed as BPs ( Table 1 ), illustrating that there was no false-negative in diagnosing CRC when ignored the CRC stage. In this regard, our diagnosis model is reliable to warn the cancerization of CRC with 100% accuracy, which in turn alerts patients to further treatment. One might argue that the low sampling sizes at the T1 (N = 9) and T4 (N = 6) stages ( Table S1 ), which could weaken statistics power. Here we collected unique and comprehensive tumors (four CRC stages), instead of fecal samples. We want to point out the CRC tend to be inspected at later stages, while T4 patients generally refuse surgery. Consequently, we collected limited tumors from T1 and T4 patients. However, the highly diagnosed accuracy of patients at the two stages indicated that our diagnosis model was not strongly affected by the sampling sizes ( Table 1 ). Despite high heterogeneity of bacterial communities among patients at each CRC stage ( Figure 1 ), we are surprised that, given the high number of ASVs (N = 3,601) identified across the 95 individuals, only 15 ASVs (accounting for 0.4%) accurately discriminated human BPs and CRC stages. This small subset affords unique experimental opportunities that could be prioritized based on their feature importances to the diagnosis model ( Figure 5 ). For example, it has been shown that levels of Parvimonas micra in faecal samples from CRC patients are significantly higher compared to healthy individuals (Löwenmark et al., 2020), as also observed in the present study ( Figures 5A, B ). Thus, P. micra is proposed as non-invasive biomarkers for CRC (Löwenmark et al., 2020). Similarly, a metagenomic analysis of faecal microbiome reveals that P. micra and Solobacterium moreii are positively associated with CRC (Yu et al., 2017). Besides confirming known associations of ASV3221 P. micra, ASV2618 S. moreii, ASV3540 Fusobacterium nucleatum and ASV3132 Peptostreptococcus stomatis with CRC, we found significant associations with several species, including ASV640 Campylobacter hominis, ASV740 Collinsella aerofaciens among others ( Figure 5 ). Clearly, the pressing steps would be to isolate representatives of these CRC stage-discriminatory ASVs and to explore their roles in CRC carcinogenesis.

This is the few attempts to explore CRC-associated taxa and underlying mechanisms in CRC progression from an ecological perspective. Tumor microenvironments are less selective for stable core community, leading to heterogeneity in bacterial communities among patients, as supported by higher AVD, lower occupancy and specificity. However, tumors could recruit beneficial taxa to antagonize CRC-associated pathogens at the initial stage, in accordance with the “cry-for-help” pattern that has not been recognized before. Based on the “driver-passenger” distribution of CRC-associated ASVs, CRC initiation could be attributed to the enrichment of pathogenic strains, and also the depletion of probiotic species. By distinguishing age- from CRC stage-associated taxa, the diagnosis model for accurately diagnoses (an overall 87.4% accuracy) BPs and the four CRC stages, especially without false-negative for CRC patients. The diagnosis model could add CRC stage stratification, especially the patients with poor pathological feature and un-reproducibility between doctors. Additional works will be imperative to eventually validate them in much larger cohorts before clinical deployment.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/ Supplementary Material .

The studies involving human participants were reviewed and approved by the ethics committees of Hwa Mei Hospital, University of Chinese Academy of Science (No. PJ-NBEY-KY-2020-042-01). The patients/participants provided their written informed consent to participate in this study.

XD, JL and JX conceived and designed the research. PC, JL, and HS conducted the experiments. JX contributed analytical tools. PC, JL, and HS analyzed the data. PC wrote the manuscript with help from JX, JL and XD. All authors contributed to the article and approved the submitted version.