Feces samples (n = 439) were collected from two separate cohorts: cohort I from 2018 to 2020, The First Affiliated Hospital of Soochow University, 318 subjects, consisting of 206 colorectal cancer patients (average age of 62.2 ± 8.6 years old; with 125 males, 81 females) and 112 normal controls (56.2 ± 12.8 years old; 63 males and 49 females); and cohort II from 2018 to 2020, The First Affiliated Hospital of Soochow University, 121 subjects, consisting of 67 patients with colorectal cancer (mean age of 67.2 ± 9.3 years old; 43 males, 24 females) and 54 normal control groups (56.7 ± 13.5 years old; 30 males, 24 females) ( Supplementary Table S1 ). Fecal samples were collected from individuals exhibiting symptoms such as altered bowel habits, rectal bleeding, abdominal pain, or anemia and asymptomatic individuals who were screened for colonoscopy (Yu et al., 2017). When the intestinal flora is restored to the baseline level, the samples should be collected before or 1 month after the colonoscopy. The exclusion criteria were as follows: i) using antibiotics over the past 3 months, ii) taking a vegetarian diet, iii) conducting invasive medical interventions over the past 3 months, and iv) having a history of any cancer or inflammatory bowel disease (Liang et al., 2017). Participants were required to collect fecal samples in standardized containers at home and to store the samples immediately in their home refrigerator at -20°C. The frozen samples were then transported in an insulated polystyrene foam container to the hospital and immediately stored at -80°C until further analysis. The patient was diagnosed by colonoscopy and by a histopathologic examination of any biopsy. Informed consent was obtained from all subjects. The study was approved by the Institutional Review Board (IRB) of the First Affiliated Hospital of Soochow University (2021112).

Fecal samples were thawed on ice and DNA extraction was performed using the QIAamp DNA Stool Mini Kit according to the instructions of the manufacturer (Qiagen). Extracts were then treated with DNase-free RNase to eliminate RNA contamination. DNA quality and quantity were determined using a NanoDrop2000 spectrophotometer (Thermo Fisher Scientific). Primers and probes used in this study are shown in Supplementary Table S2 .

qPCR amplification is performed in a 20-μl reaction system of TaqMan Universal Master Mix II (Applied Biological Systems) containing 0.3 μmol/L of 96-hole reaction sheets and 0.2 μmol/L of reaction sheet seals. The thermal cycler parameters of an ABI PRISM 7900HT sequence detection system were 95°C 10 min × 45 cycles. Each experiment included positive/reference and negative controls (water as a template). Three measurements were repeated for each sample. The qPCR data were analyzed using sequence detection software (Applied Biological Systems) (Liang et al., 2017). Data analysis was performed according to the ΔCq method, using Cq_target − Cq_control and relative abundance = Power (2, −ΔCq). The nucleotide sequences of primers and probes are shown in Supplementary Table S2 .

Each participant received a Medline iFOB (Medline Industries Inc., Northfield, IL, USA, Lot nom. 768L11) collection tube. Medline iFOB is a qualitative FIT product exempted by the Clinical Laboratory Improvement Act (CLIA). The lower limit of hemoglobin detection set by the manufacturer was 50 g/g. The feces were exposed to ambient temperature for no more than 48 h between emptying and treatment. Each sample was processed in accordance with the instructions of the manufacturer, including verification of activated internal controls. The results of each test were interpreted by two members of the research team (Malagon et al., 2020; Knapp et al., 2021).

This study used IBM SPSS 25.0 software (SPSS Inc., Chicago, IL, USA) for data statistical analysis and logical regression analysis. Classification data were presented by frequency (percentage), chi-square analysis, or Fisher exact probability method. Continuous data were presented as mean ± standard deviation, and group differences were analyzed using t-test or non-parametric Mann–Whitney U test. All tests were bilateral, and P <0.05 indicated significant statistical differences. In this study, the software R 3.4.3 (http://www.R-project.org, The R Foundation) was used to construct the logistic regression forecasting model, and charts were analyzed using the R software, where the receiver operating feature (ROC) curve was drawn using the Hiplot visual drawing website (https://hiplot.com.cn/). According to the actual flora abundance of the case and predicted situation of the model, the ROC curve was drawn. The area under the curve (AUC) and the differentiation size of the model were also evaluated. The larger the AUC value of the model, the better the prediction effect. Generally, AUC ≥0.8 indicates good differentiation and clinical significance, and AUC <0.6 indicates poor clinical application value. The best truncation value (cutoff value) was determined by the ROC curve and calculated for sensitivity, specificity, diagnostic accuracy, diagnostic ratio, positive prediction rate, positive result likelihood ratio, negative result likelihood ratio at the best cutoff value, and 95% confidence interval. The net gain of the prediction model is evaluated by the predictive model’s decision curve (decision curve analysis, DCA). DCA, as a simple way to evaluate clinical predictive models, diagnostic trials, and molecular markers, is able to better estimate the model for clinical value than AUC by integrating patient or decision-maker preferences into the analysis (Vickers and Elkin, 2006). After determining the best cutoff value of the predictive model, its nomogram was drawn, which can transform the complex regression equation into a visualized graph and make the results of the prediction model more readable, facilitating the evaluation of the patient.

The progression of enteritis-related bowel cancer and intestinal flora was analyzed through the gutMEGA database (http://gutmega.omicsbio.info/). Log₂(colorectal cancer/NOR) > 3 and Log₂(colorectal cancer/ADE) > 3 and P <0.05 were set. The associated abundance differences ( Figures 1A, B ) were obtained through screening studies. The Venn map showed that there were 13 species of intestinal flora in colorectal cancer and normal human that have significant differences in abundance and 9 species in colorectal cancer and adenoma, where the Pc, Gm, Pm, Cs, and Ps bacteria achieved statistical differences simultaneously in both groups of controls (Log₂Ratio > 3, P < 0.05, Figure 1C ).

In all five bacteria, Gm and Pm performed better in distinguishing colorectal cancer from healthy control than the other three target bacteria, and the relative abundance of Gm and Pm in colorectal cancer patients was significantly higher than that in healthy control groups (P < 0.001; Figure 2A ). The AUC of Gm in cohort I is 0.799 (95% confidence interval of 0.751–0.842; P < 0.001; Figure 2B ) and that of Pm in cohort I is 0.791 (95% confidence interval of 0.749–0.838; P < 0.001; Figure 2B ). At the best cutoff value which maximizes the sensitivity and specificity in ROC analysis, Gm had a sensitivity of 97.09% to 206 colorectal cancer patients and 112 healthy people, specificity of 61.61%, negative predictive value (NPV) of 92.00%, and positive predictive value (PPV) of 82.30%. The sensitivity of Pm to this cohort was 93.20%, specificity was 58.04%, NPV was 82.28%, and PPV was 80.33%. This was further validated in the independent cohort II of 67 colorectal cancer patients and 54 healthy controls. The relative abundance of Gm and Pm in colorectal cancer patients remained significantly higher than in the healthy control groups (P < 0.001; Figure 2C ). As a single factor which distinguishes colorectal cancer patients and the control groups, the AUC of the fecal biomarker candidate Gm was 0.823 (0.762–0.893, P < 0.001, Figure 2D ), and that of Pm was 0.774 (0.712–0.838, P < 0.001, Figure 2D ).

In line with the data after sequencing, the combined use of the bacterial markers tested showed better diagnostic performance than Gm or Pm alone, with AUC increasing to 0.8615 ( Figure 3B ). It was found that the simple linear combination of Gm, Pm, Pc, Ps, and Cs had higher AUC (0.862) than other combinations (two to four markers) or Gm (0.799) or Pm (0.791) alone. The relative abundance of the five bacteria in patients with colorectal cancer was significantly higher than that in healthy control groups (P < 0.001, Figure 3A ). Under the best cutoff value, the five bacteria (Gm, Pm, Pc, Ps, Cs) can distinguish patients with colorectal cancer from healthy controls, with a sensitivity of 67.96%, specificity of 89.29%, NPV of 60.24%, and PPV of 92.11%, which show better diagnostic performance than Gm or Pm alone ( Table 1 ). The pairing comparison of AUC showed that the colorectal cancer diagnostic ability of the five-bacteria group was better than that of the group Gm or Pm alone (P < 0.05).

The diagnostic performance of the five bacteria was further demonstrated in the independent cohort II. The combination of the five bacteria also showed an increase in AUC (0.928, Figure 3D ). In patients with colorectal cancer, the relative abundance of the five bacteria was significantly higher than that in the healthy control group (P < 0.001, Figure 3C ). It showed better diagnostic performance in the combination of the five bacteria than Gm or Pm alone. Therefore, the combination of Gm, Pm, Pc, Cs, and Ps improves the diagnostic ability of Gm and Pm to identify colorectal cancer and healthy control groups.

FIT stool specimens were tested from 206 patients with colorectal cancer and 112 normal controls in cohort I. We found that 56.31% of the stool specimens from colorectal cancer patients were FIT positive. In this subcohort of colorectal cancer patients, the detection rate of FIT was lower than the combination of five bacteria (67.96%, P < 0.05). Pairwise comparison of the ROC curve showed that the combinations of five bacteria were significantly superior to Gm or Pm alone for the diagnosis of colorectal cancer (both P < 0.05). FIT was slightly related with TNM staging (P = 0.076), while the relative abundances of the five bacteria or Gm and Pm were independent of TNM staging. Assessment of colorectal cancer detection in the TNM staging subgroup showed that the quantification of bacterial markers was much more sensitive than FIT in stage I. Similarly, in stage II and III, we observed more prominent detection rates by bacteria than FIT ( Figure 4 ). It indicated that the sensitivity of the quantitative detection bacterial markers is significantly higher than FIT, especially for non-metastatic colorectal cancer.

After the combination of bacterial markers and FIT, the sensitivity of FIT increased from 56.31% to 80.58% and the five-bacteria group from 67.96% to 80.58%, PPV and NPV were both improved, and the specificity was almost unchanged ( Table 2 ). According to the I–IV stages of TNM staging, the sensitivity of the binding of bacterial markers to FIT was significantly higher than FIT alone. It suggested that the joint detection of bacterial markers and FIT had elevated sensitivity and specificity for the non-invasive diagnosis of colorectal cancer.

To further demonstrate the clinical application prospects of the five target bacteria, cohort I was used as the training group to construct the colorectal cancer risk model, and cohort II was analyzed as the testing group to test the feasibility of the cohort I-dependent testing model (Dong et al., 2019; Sun et al., 2019). The clinical decision curve and clinical impact curve of the colorectal cancer risk prediction model combined with FIT showed a significant clinical benefit of the prediction model ( Figures 5A, B ). To facilitate the clinical application of the colorectal cancer risk prediction model for the combination of five bacteria, we further constructed a nomogram based on this model ( Figure 6A ). The relative abundance of the five bacteria combinations and FIT was assigned, and the total score was calculated to determine the likelihood of complications. When the total points are up to 50, the predicted model predicts a probability of 0.4 ( Figure 6A ). Taking the prediction of the possibility of colorectal cancer as the transverse coordinate and the actual positive rate of the case as the longitudinal coordinate, the prediction calibration curve (calibration curve) of the model showed the overall prediction with the actual situation ( Figures 6B, C ). The prediction accuracy of the model was not significantly reduced with the change of the population, which proved the reliability of the colorectal cancer risk model based on the five target bacteria.