Gut microbiota signatures and machine learning-based candidate feature prioritization in advanced colorectal cancer

Abstract

This single-center, cross-sectional case-control study aimed to characterize the gut microbiome profiles of patients with advanced colorectal cancer (CRC), identify candidate microbial features, and explore how well these features distinguished advanced CRC cases from healthy controls within the study dataset. Fecal samples were collected from 72 treatment-naïve patients with advanced CRC and 61 healthy controls. Microbial community structure was profiled using high-throughput sequencing of the V3-V4 region of the 16 S rRNA gene. The analytical pipeline included α/β-diversity analysis, multilevel taxonomic analysis, LEfSe, and machine-learning algorithms, including random forest (RF), gradient boosting machine (GBM), and LASSO, to identify key Amplicon Sequence Variants (ASVs) and evaluate their ability to discriminate between the two groups. Results showed significantly reduced α-diversity and distinct β-diversity in the CRC group. Key SCFA-producing genera (Faecalibacterium, Agathobacter, Roseburia) were consistently depleted. The RF feature-prioritization model achieved an out-of-bag (OOB) accuracy of 0.850 and an OOB AUC of 0.899. Nested five-fold cross-validation based on the prioritized ASVs yielded AUCs of 0.891, 0.900, and 0.891 for RF, GBM, and LASSO, respectively. These findings show internally reproducible case–control discriminatory patterns within the present cohort and support further evaluation of the prioritized microbial features in independent, clinically representative populations.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00203-026-05220-x.

Introduction

CRC is the third most commonly diagnosed cancer worldwide and represents a substantial public health burden (Bray and Laversanne 2024). CRC is characterized by substantial genetic and biological heterogeneity (Li and Ma 2021), and metastatic disease remains associated with poor prognosis, highlighting the need to characterize disease-associated biological features and complementary assessment approaches (Siegel R L, Wagle 2023). Previous 16 S rRNA surveys have demonstrated structural differences between the gut microbiota of patients with CRC and healthy individuals, including lower abundances of Roseburia and other butyrate-associated taxa (Wang et al. 2012). Similar depletion of F. prausnitzii has been reported in geographically distinct CRC cohorts, indicating that these patterns are reproducible rather than unique to the present cohort (Obuya et al. 2022). Microbial perturbations can modulate host immunity, metabolism, and cellular signaling pathways, thus driving tumor progression (Haque et al. 2025, Herlo L F, Salcudean et al. 2024, Wong C C, Yu 2023). Gut microbiota imbalance has been associated with CRC development and progression, prompting interest in microbial features as potential complements to established CRC detection approaches (Bars-Cortina D et al. 2024, Pateriya et al. 2025). Fecal occult blood testing has limitations in sensitivity, and microbiome profiling has therefore been investigated as a potential complementary approach (Young et al. 2021). However, reported performance varies across populations, sequencing platforms, and analytical workflows, and its clinical utility remains to be established through prospective validation.

Although numerous studies have investigated the role of gut microbiota in CRC, most have focused on early-stage or non-metastatic cases. Recent mCRC-focused studies have mainly examined microbiome-based differentiation of metastatic from non-metastatic CRC (Cui et al. 2023), tumor-associated microbiome features and their clinical implications in mCRC (AN H J, Partha M A, Lee et al. 2024), or gut microbiota associated with lymph-node metastasis (Wu et al. 2025). However, they did not specifically characterize fecal microbial dysbiosis in a treatment-naïve stage IV cohort or determine whether ASV-level candidates remain consistently prioritized across complementary feature-selection and classification algorithms. Thus, the microbial features distinguishing untreated advanced CRC from healthy controls remain insufficiently defined. Moreover, microbiome data are complex and high-dimensional, whereas sample sizes are often modest. Traditional statistical models may therefore be vulnerable to overfitting and feature redundancy, complicating the identification of reproducible microbial features (Rynazal et al. 2023, Porreca et al. 2024). Although machine learning can support dimensionality reduction and exploratory classification, model performance may vary across disease stages, cohorts, and sequencing platforms (Porreca et al. 2024, Ramon et al. 2024). These limitations emphasize the need to distinguish within-cohort case-control classification from clinically validated diagnostic or screening performance.

Combining 16 S rRNA gene sequencing with machine learning offers a useful exploratory framework for microbial feature prioritization and group classification. 16 S rRNA sequencing enables profiling of microbial communities in fecal or tumor samples, and is more cost-effective than whole-genome sequencing for many clinical research settings (Bars-Cortina D et al. 2024, Konishi et al. 2022). Meanwhile, machine learning algorithms such as RF and GBM can reduce data dimensionality, rank features by importance, and identify microbial patterns that distinguish predefined groups. In a cross-sectional case-control study, however, such analyses primarily quantify separation between sampled cases and controls and do not by themselves establish clinical diagnostic utility (Young et al. 2021, Wu et al. 2025, Porreca et al. 2024). Candidate features identified through this approach require independent validation in intended-use populations, preferably with integration of multi-omics and relevant clinical variables.

Accordingly, this study aimed to characterize gut microbiota alterations in treatment-naïve patients with advanced CRC and prioritize candidate microbial features associated with case-control discrimination. Fecal samples from patients and healthy controls were analyzed by 16 S rRNA sequencing, integrating α/β-diversity, multilevel taxonomic profiling, LEfSe, RF-based feature ranking, and exploratory machine-learning classification. The resulting candidate microbial features are intended to support subsequent mechanistic investigations and independent external-validation studies.

Materials and methods

Study subjects

This was a single-center, cross-sectional case-control study with prospective participant recruitment. The study was approved by the Medical Research Ethics Review Committee of the General Hospital of Ningxia Medical University (Approval No.: KYLL-2025-2274). From March 2022 to December 2024, 72 treatment-naïve patients with established advanced CRC (CP) and 61 healthy controls (HC) were enrolled at the Ningxia Medical University General Hospital, a tertiary public hospital in Yinchuan, Ningxia, China. All participants provided written informed consent prior to sample collection.

Inclusion criteria for the mCRC group were: (1) age > 18 years; (2) histopathologically confirmed colorectal adenocarcinoma, diagnosed by biopsy or surgical specimen based on WHO classification criteria; (3) confirmed distant metastasis (stage IV per AJCC 8th edition). Healthy controls were individuals undergoing routine health examinations, aged > 18 years, with no history of malignancy. Exclusion criteria for both groups included: (1) prior systemic chemotherapy, radiotherapy, targeted therapy, or immunotherapy; (2) antibiotic, probiotic, or prebiotic use within three months; (3) active infection within one week; (4) inflammatory bowel disease or other chronic gastrointestinal disorders; (5) major gastrointestinal surgery. Baseline data, including demographics, medical history, and tumor characteristics, were recorded. All fecal samples were self-collected, transported within one hour, and stored at −80 °C.

Fecal DNA extraction

Genomic DNA was extracted from fecal samples following the manufacturer’s instructions of the QIAamp DNA Stool Mini Kit (Qiagen, Hilden, Germany). DNA integrity was assessed using 1% agarose gel electrophoresis, and DNA concentration and purity were measured using a NanoDrop ND-1000 spectrophotometer (Thermo Fisher Scientific, USA).

16 S rRNA sequencing

The V3-V4 hypervariable region was amplified using primers 338 F (5’-ACTCCTACGGGAGGCAGCAG-3’) and 805R (5’-GGACTACHVGGGTWT CTAAT-3’). Purified amplicons were pooled in equimolar amounts and sent to Shanghai Mubo Biomedical Technology Co., Ltd. for sequencing on the Illumina MiSeq PE300 platform (Illumina, USA) using standard protocols and the MiSeq Reagent Kit v3 (2 × 300, 600 cycles). Raw sequencing data were primarily processed using the DADA2 pipeline for denoising and ASV generation, which was adopted as the main analytical basis for all downstream analyses.

Alpha and beta diversity analysis

α-Diversity analysis was conducted to assess the richness and within-sample diversity. ACE, Chao1, and Observed Species indices were calculated as richness measures, whereas Shannon and Simpson D indices were calculated as diversity measures; for the Simpson D index used here, a higher value indicates lower diversity. Between-group differences were assessed using two-sided Wilcoxon rank-sum tests. Consistent with the non-parametric analysis, α-diversity indices were summarized using medians and interquartile ranges (IQRs). To address multiplicity across the five α-diversity indices, P values were adjusted using the Benjamini–Hochberg false-discovery-rate (FDR) procedure. Calculations were primarily performed using mothur (version 1.42.1), with statistical testing and graphing conducted in R. All α-diversity calculations were based on ASV abundance tables generated by DADA2.

β-Diversity analysis was performed to assess between-sample differences in overall microbial community composition. Bray-Curtis and Jaccard distance matrices were calculated from rarefied ASV abundance tables using the vegan and phyloseq packages in R (version 4.3.0). Principal coordinate analysis (PCoA) was used to visualize community dissimilarities in two and three dimensions. Overall between-group compositional differences were assessed by permutational multivariate analysis of variance (PERMANOVA; 999 permutations) using the adonis2 function. Homogeneity of multivariate dispersion was evaluated using betadisper followed by a permutation test. All β-diversity analyses were based on rarefied ASV tables to ensure equal sequencing depth across samples.

ASV generation and distribution analysis

High-quality sequences were denoised, merged, and used to generate an ASV table via the DADA2 algorithm. This ASV-based pipeline served as the unique and definitive analytical framework for all subsequent bioinformatics analyses. Representative ASV sequences were taxonomically annotated using the SILVA database (Release 138). To compare similarities and uniqueness in microbial composition between groups, a Venn diagram was generated. Shared ASVs were defined as those detected in at least one sample in each group. No prevalence-based core microbiome definition was applied. This analysis was completed within the QIIME 2 environment, and the Venn diagram was generated using relevant R packages.

Taxonomic composition analysis

To reveal differences in community structure at various taxonomic levels, this study systematically analyzed the relative abundance of taxa from phylum to genus. First, based on ASV annotation results, the relative abundance of each taxon per sample was calculated at each taxonomic rank and visualized using bar charts. Interactive hierarchical pie charts were also generated using the Krona tool to display species composition proportions across multiple taxonomic levels. Between-group abundance differences were assessed using the two-sided Wilcoxon rank-sum test. P values were adjusted using the Benjamini-Hochberg FDR procedure within each prespecified taxonomic level, and the number of taxa included in each correction family was recorded. All statistical analyses and graphing were performed using R.

LEfSe analysis

The non-parametric Kruskal-Wallis rank-sum test was first used to screen for taxa with significant abundance differences between groups. Subsequently, Linear Discriminant Analysis (LDA) was performed on significantly different taxa to calculate an LDA score for each, quantifying its contribution to distinguishing the groups. Taxa with an LDA score greater than 2 and a p-value less than 0.05 were considered differential taxonomic features. Results were presented as a cladogram and a directionally signed LDA score plot, with positive and negative values indicating enrichment in the corresponding groups.

RF-based feature prioritization and OOB internal evaluation

RF analysis was used to prioritize ASVs associated with separation between the CP and HC groups. An RF classifier containing 1,000 trees was fitted to the ASV abundance table using the randomForest package in R. Feature importance was quantified by mean decrease in accuracy (MDA), defined as the increase in OOB prediction error after permutation of an individual ASV. Higher MDA values indicated a greater contribution to case-control classification. ASVs with an MDA value greater than 0.001 were retained as candidate microbial features. Because interpretation of MDA depends on the performance of the fitted model, OOB class probabilities and predictions were extracted to calculate the confusion matrix, error rate, accuracy, sensitivity, specificity, and AUC with its 95% confidence interval. CP was defined as the positive class. These metrics were treated as internal OOB performance estimates rather than external validation. The abundance distributions of the retained ASVs were visualized in a heatmap.

Exploratory internal evaluation of selected features

In the original workflow, RF-based feature prioritization was conducted before the dataset was divided into a 70% training set and a 30% test set, which could introduce optimistic bias because feature selection was not restricted to the training data. Therefore, a fully nested five-fold cross-validation procedure was implemented for leakage-controlled internal evaluation. Rarefied ASV counts were transformed using log (1 + count), and the random seed was set to 20,260,729. In each outer fold, one subset was held out exclusively for evaluation, while the remaining samples were used for feature selection, model tuning, and fitting. An RF selector with 1,000 trees was fitted only to the outer training data, and ASVs with an MDA value greater than 0.001 were retained for subsequent model training and evaluation within that fold. The default mtry setting of the randomForest package was used.

Hyperparameters were optimized by inner five-fold cross-validation. For GBM, 54 combinations were evaluated: n.trees = 100, 300, or 500; interaction.depth = 1, 2, or 3; shrinkage = 0.01, 0.05, or 0.10; and n.minobsinnode = 5 or 10. The optimal configuration was selected by maximizing cross-validated ROC AUC. LASSO logistic regression was fitted using a binomial family with alpha = 1; the penalty parameter was selected using five-fold cross-validation with AUC as the optimization measure, and lambda.1se was used for final prediction. The optimized RF, GBM, and LASSO models were evaluated on the untouched outer test folds, and pooled out-of-fold probabilities were used to calculate final AUCs.

Statistical analysis methods

Bioinformatics and statistical analyses were performed using R (version 4.3.0) and QIIME 2 (version 2023.5). Continuous baseline variables are presented as mean ± standard deviation (SD), and categorical variables as number (percentage). Age and BMI were compared using independent-samples t tests, whereas categorical baseline variables were compared using Pearson’s chi-square test or Fisher’s exact test, as appropriate. α-Diversity indices are presented as median (IQR) and were compared using two-sided Wilcoxon rank-sum tests. P values across the five indices were adjusted using the Benjamini-Hochberg FDR procedure. Shared-ASV differential-abundance P-values were adjusted using the Benjamini-Hochberg FDR procedure. β-diversity was assessed by PERMANOVA with 999 permutations and reported using pseudo-F, R², and permutation P values; multivariate dispersion was assessed using betadisper followed by a permutation test. Covariate-adjusted Bray–Curtis PERMANOVA included CP/HC group, age, sex, and BMI. RF-based MDA ranking was used for candidate-feature prioritization. Wilson score intervals were used to calculate the 95% confidence intervals for OOB accuracy, sensitivity, and specificity. All tests were two-sided, and P < 0.05 was considered statistically significant unless an FDR-adjusted threshold was specified. ASV-level confounding was evaluated independently of RF selection using separate multivariable logistic models across the full ASV set, with case-control status as the outcome and each ASV, age, sex, and BMI as predictors. ASV-term P values were BH-corrected across all tested ASVs.

Results

Gut microbial α-diversity and β-diversity in advanced CRC and healthy controls

A total of 72 treatment-naïve patients with advanced CRC and 61 healthy controls were enrolled. There were no significant between-group differences in age, sex, BMI, smoking history, alcohol drinking history, hypertension, or diabetes mellitus (all P > 0.05) Table 1. Tumor location, histological differentiation, and treatment status were recorded descriptively for the CRC group only and were not assessed in healthy controls. Therefore, no corresponding between-group comparisons were performed. Among the CRC patients, 40 (55.6%) had colon cancer and 32 (44.4%) had rectal cancer; 45 (62.5%) had well- or moderately differentiated tumors, and 27 (37.5%) had poorly differentiated tumors.

Table 1.

Baseline characteristics of patients with advanced CRC and healthy controls

Characteristic	Advanced CRC group (CP, n = 72)	Healthy control group (HC, n = 61)	P-value
Age (years), Mean ± SD	59.6 ± 7.2	58.9 ± 6.8	0.588
Sex, n (%)			0.742
Male	41 (56.9)	33 (54.1)
Female	31 (43.1)	28 (45.9)
BMI (kg/m²), Mean ± SD	23.8 ± 2.1	24.1 ± 2.3	0.446
Smoking history, n (%)	22 (30.6)	17 (27.9)	0.731
Alcohol drinking history, n (%)	18 (25.0)	15 (24.6)	0.954
Hypertension, n (%)	25 (34.7)	20 (32.8)	0.820
Diabetes mellitus, n (%)	11 (15.3)	9 (14.8)	0.929
Tumor location, n (%)		NA	NA
Colon	40 (55.6)	NA
Rectum	32 (44.4)	NA
Differentiation, n (%)		NA	NA
Well/moderately differentiated	45 (62.5)	NA
Poorly differentiated	27 (37.5)	NA
Treatment status, n (%)	Newly diagnosed, treatment-naïve, 72 (100%)	NA	NA

Tumor-related variables were recorded only for advanced CRC; therefore, no between-group comparisons were performed. NA, not applicable. “Newly diagnosed” denotes histologically confirmed stage IV CRC at enrollment; “treatment-naïve” denotes no prior anticancer therapy before fecal sampling

Rarefaction curves for all 133 samples (CP, n = 72; HC, n = 61) approached plateaus at high sequencing depths (Fig. S1). All 133 samples included in the rarefied analyses were rarefied to 11,176 reads. Mean Good’s coverage was 0.9996 overall, 0.9996 in CP, and 0.9995 in HC, supporting adequate sequencing depth for the present analyses. Median observed ASV counts were 104.5 in CP and 128 in HC (Wilcoxon rank-sum test, P = 0.0014). The ACE, Chao1, and Observed Species indices were lower in CP than in HC (all P < 0.01; Fig. 1A-C). The Shannon index was also lower in CP (3.06 ± 0.86 vs. 3.35 ± 0.90, P = 4.42e-05), whereas the Simpson D index was higher (0.12 ± 0.047 vs. 0.073 ± 0.09, P < 0.001; Fig. 1D, E), consistently indicating lower within-sample diversity in the CP group. β-Diversity analysis demonstrated statistically significant between-group compositional differences. PCoA based on Bray-Curtis distances showed partial separation with substantial overlap (PERMANOVA P = 0.016; Fig. 1F), and PCoA based on Jaccard distances showed a similar pattern (PERMANOVA P = 0.001; Fig. 1G). These findings indicate group-associated differences in microbial community composition within the present cohort.

Fig. 1.

Gut microbial α-diversity and β-diversity in the CP and HC groups (CP, n = 72; HC, n = 61). A-C ACE, Chao1, and Observed Species indices. D, E Shannon and Simpson D indices; a higher Simpson D value indicates lower diversity. α-Diversity comparisons used two-sided Wilcoxon rank-sum tests with Benjamini–Hochberg correction across the five indices. F Two-dimensional PCoA based on Bray-Curtis distance; PC1 and PC2 explain 9.72% and 4.87% of the variance, respectively. G Three-dimensional PCoA based on Jaccard distance; PC1, PC2, and PC3 explain 8.83%, 4.23%, and 3.79% of the variance, respectively. The ordination plots show statistically significant but partially overlapping group distributions (CP, red; HC, blue)

Analysis of shared and unique ASV distribution between groups

Venn diagram analysis revealed 790 shared ASVs between the CP patient group and the healthy control group. These ASVs represent taxa detected in both groups, while each group also contained a distinct number of unique ASVs (Fig. 2A, B). To further characterize the distributional patterns of these 790 shared ASVs, we performed differential abundance analysis using the Wilcoxon rank-sum test with FDR correction. A total of 32 shared ASVs were significantly differentially abundant between the CP and HC groups (FDR-adjusted p < 0.05), of which 7 were significantly enriched in CP and 25 were significantly depleted in CP (Fig. 2C). The depleted ASVs were taxonomically assigned primarily to short-chain fatty acids (SCFAs)-producing genera, including Faecalibacterium, Enterocloster, and Agathobacter. Detailed statistical results are available in Supplementary Table S1. These findings describe cohort-specific abundance differences among ASVs detected across both groups.

Fig. 2.

Distribution and differential abundance of shared and group-specific ASVs. A UpSet-style summary of shared and group-specific ASV counts. B Venn diagram showing 790 shared ASVs, 215 CP-specific ASVs, and 233 HC-specific ASVs. C Bar chart comparing the average abundance of 32 shared ASVs that differed significantly between CP and HC by Wilcoxon rank-sum testing with Benjamini-Hochberg FDR correction. Statistical labels denote FDR-adjusted P values. *p < 0.05, **p < 0.01, ***p < 0.001

Differences in multi-level taxonomic composition and analysis of potential functional taxa

Subsequently, compositional differences in the gut microbiota between HC and CP groups were systematically compared at the domain, phylum, class, order, family, and genus levels. Circular plots (Fig. 3A, B) showed that within the bacterial domain, the phylum Bacillota (Firmicutes) occupied the major proportion, and its relative abundance changed significantly between the HC and CP groups. Further quantitative analysis (Fig. S2, S3, Fig. 3C) indicated that at the phylum level, community structure already showed clear distinctions between the two groups. At the class level, Clostridia was highly abundant in both groups, but significantly higher in the HC group than in the CP group. At the order level, multiple dominant orders exhibited distinct abundance patterns between groups. At the family level, Wilcoxon rank-sum test identified numerous significantly differential taxa between CP and HC (Fig. S3). Several dominant families showed significant or highly significant differences, indicating marked microbial dysbiosis at the family level. At the genus level (Fig. 3C), the abundances of Faecalibacterium, Agathobacter, Roseburia, and Coprococcus were consistently and significantly reduced in the CP group. These genera include taxa commonly associated with short-chain fatty acid production.

Fig. 3.

Comparison of gut microbiota composition and relative abundance between the CP and HC groups. A, B Krona interactive circular plot showing the taxonomic composition of the gut microbiota in the HC and CP groups from phylum to genus. C Relative abundances of 20 displayed genera compared using the Wilcoxon rank-sum test, with Benjamini–Hochberg correction for multiple comparisons. Points and horizontal intervals represent between-group differences and 95% confidence intervals. Statistical labels denote FDR-adjusted P values; *p < 0.05, **p < 0.01, ***p < 0.001

Identification of key differential microbial taxa associated with CRC via LEfSe analysis

To systematically identify key differential species in the gut microbiota of CP patients versus HC, we employed the non-parametric Kruskal-Wallis rank-sum test to screen for microbial taxa with significant abundance differences between groups (Fig. 4A). Subsequently, LDA was used to further quantify the contribution of each differential species to group separation, with the results presented as LDA scores (Log10) (Fig. 4B). The analysis revealed significant differential taxa across multiple taxonomic levels distinguishing the two groups. Among these, genera such as Faecalibacterium, Agathobacter, Roseburia, and Gemmiger exhibited high absolute LDA scores. These taxa should be regarded as cohort-associated differential features because their abundance may also reflect metastatic heterogeneity, nutritional or immune status, and other systemic effects of advanced disease.

Fig. 4.

Identification of intergroup differential species via LEfSe analysis. A Cladogram showing differential taxa and their phylogenetic relationships from phylum to genus. B LDA score plot of differential taxa across multiple taxonomic levels. Bar direction and color indicate the group in which each taxon was enriched, while bar length represents the absolute LDA effect size. Differential features were defined by |LDA score| > 2 and P < 0.05

RF-Based ASV prioritization and internally evaluated case-control classification

RF analysis was performed to prioritize ASVs contributing to separation between the CP and HC groups. In the exploratory full-cohort RF analysis, 41 ASVs met the selection criterion of MDA > 0.001 and were retained as candidate microbial features (Supplementary table S2). These ASVs included taxa assigned to genera such as Faecalibacterium, Agathobacter, Roseburia, and Gemmiger. Their group-associated abundance patterns were descriptively visualized in a heatmap (Fig. 5A). The fitted RF model achieved an OOB accuracy of 0.850 and an OOB error rate of 0.1504 (Table 2). With CP defined as the positive class, the sensitivity was 0.917 and the specificity was 0.771. The OOB AUC was 0.899 (95% CI, 0.843–0.956). The OOB confusion matrix showed that 66 CP and 47 HC samples were correctly classified, whereas 6 CP and 14 HC samples were misclassified. These results indicate that the fitted RF model contained sufficient group-related discriminatory information to support interpretation of the MDA ranking, although they do not establish clinical diagnostic performance.

Fig. 5.

Abundance patterns of RF-prioritized ASVs and internally evaluated classification performance across machine-learning models. A Heatmap showing Z-score-standardized relative abundances of RF-prioritized ASVs across advanced colorectal cancer (CP) and healthy control (HC) samples; the top annotation denotes group membership. B ROC curves from nested five-fold cross-validation for RF, GBM, and LASSO. AUCs were 0.891 (95% CI 0.830–0.951), 0.900 (0.842–0.958), and 0.891 (0.834–0.948), respectively. Corresponding cutoffs were 0.554, 0.461, and 0.583; sensitivities were 0.852, 0.820, and 0.852; specificities were 0.833, 0.903, and 0.806. These results indicate consistent internal case–control discrimination across models

Table 2.

Out-of-bag internal performance and confusion matrix of the random forest model

Category	Metric	Result
Sample characteristics	Total sample size	133
	Advanced CRC cases	72
	Healthy controls	61
Model parameter	Number of decision trees (ntree)	1000
OOB internal performance	OOB error rate	0.150
	OOB accuracy	0.850 (95% CI, 0.779–0.900)
	Sensitivity for CRC	0.917 (95% CI, 0.830–0.961)
	Specificity for healthy controls	0.771 (95% CI, 0.651–0.858)
	OOB AUC (95% CI)	0.899 (95% CI, 0.843–0.956)
Confusion matrix	True positives (TP)	66
	True negatives (TN)	47
	False positives (FP)	14
	False negatives (FN)	6

In the nested five-fold cross-validation analysis, the MDA > 0.001 selection criterion was reapplied independently within each outer training fold. Exploratory nested five-fold cross-validation, with feature selection repeated within each outer training fold, yielded AUCs of 0.891 for RF, 0.900 for GBM, and 0.891 for LASSO (Fig. 5B). Compared with the original single-split estimates, the AUCs differed by −0.077 for RF, − 0.060 for GBM, and +0.052 for LASSO. The lower nested estimates for RF and GBM indicate measurable optimistic bias in the original workflow, whereas the opposite change in LASSO suggests sensitivity to the original data split. The nested estimates provide a more conservative assessment of internal group-discriminatory information but still require independent external validation.

Potential confounding by age, sex, and BMI was further evaluated using multivariable logistic regression. In a selection-independent analysis of the full ASV set, separate multivariable logistic regression models adjusted for age, sex, and BMI identified 94 ASVs associated with case–control status after Benjamini–Hochberg correction across all tested ASVs (FDR < 0.05); 36 of these overlapped with the 41 RF-prioritized ASVs (Supplementary table S3). Covariate-adjusted PERMANOVA based on Bray–Curtis distances showed no significant effects of age (R² = 0.0061, P = 0.795), sex (R² = 0.0079, P = 0.297), or BMI (R² = 0.0065, P = 0.708), whereas CP/HC group membership remained significantly associated with community structure after adjustment. These findings indicate that most RF-prioritized ASVs were supported by a selection-independent covariate-adjusted analysis and that the overall community difference was not materially explained by the measured covariates within this cohort.

Discussion

This study characterized gut microbiota dysbiosis in treatment-naïve patients with advanced CRC using 16 S rRNA sequencing, taxonomic analysis, and exploratory machine learning. Advanced CRC cases showed reduced community richness and evenness, separation of overall community structure from healthy controls, and depletion of several short-chain fatty acid-producing genera, including Faecalibacterium, Agathobacter, Roseburia, and Coprococcus. The RF feature-prioritization model showed an OOB AUC of 0.899 and an OOB error rate of 0.150, while nested five-fold cross-validation yielded AUCs of 0.891–0.900 across RF, GBM, and LASSO. These findings provide exploratory internal support for the selected features.

CRC imposes a substantial and rising public health burden in China. The age-standardized incidence increased from 2.75/100,000 in 1972 to 19.39/100,000 in 2019 (Zhang et al. 2024), and over 611,000 new cases were estimated in 2023 (BRAY and LAVERSANNE 2024). A recent regional study also documented a steady increase in CRC incidence in Gansu Province from 2013 to 2023 (He et al. 2026). Within this context, recruitment of 72 treatment-naïve patients with advanced CRC over three years at a single tertiary center provided a feasible cohort for exploratory analysis. This study found that gut microbial richness and evenness were significantly reduced in patients with advanced CRC compared with healthy controls, consistent with previous reports of CRC-associated dysbiosis (Alrahawy et al. 2022, Yu and Zhao 2026). However, the cross-sectional design provides no temporal information and therefore cannot determine whether microbial dysbiosis contributes to CRC development and progression or occurs secondarily to stage IV disease. The observed reduction in microbial diversity may therefore reflect stage IV disease and its accompanying systemic changes, including alterations in metabolism, immune status, and nutritional condition, rather than a causal microbial effect (Yu and Zhao 2026). Conversely, loss of SCFA-producing bacteria may impair intestinal homeostasis and promote inflammation (Vescovo et al. 2025, Touchefeu et al. 2021), whereas bacterial genotoxins such as colibactin can contribute to DNA damage and colorectal carcinogenesis (Eraghieh Farahani H, Pourhajibagher 2025). Thus, the observed microbial alterations may reflect a bidirectional relationship with advanced CRC and should be interpreted as disease-associated patterns rather than evidence of causality.

At the genus level, the lower abundances of Faecalibacterium and Roseburia observed here are consistent with previous CRC studies rather than newly identified taxonomic signals (Wang et al. 2012, Obuya et al. 2022). Their biological relevance is supported by experimental evidence showing that F. prausnitzii can suppress HCT116 cell proliferation and reduce colorectal tumorigenesis in vivo (Dikeocha I J et al. 2022). Faecalibacterium and Roseburia include major SCFA-producing species, particularly butyrate producers, with reported anti-inflammatory and epithelial-protective effects (Vescovo et al. 2025, Touchefeu et al. 2021, Hou et al. 2022, Kang et al. 2025). Consistent with this functional relevance, R. intestinalis-derived butyrate has been shown to enhance cytotoxic CD8 + T-cell activity and anti-PD-1 efficacy in CRC models (Kang et al. 2023). Stage-associated shifts in the CRC microbiome have been reported, although the taxa and direction of change vary across studies (Dikeocha I J et al. 2022, Gao et al. 2022, GONZáLEZ et al. 2024). However, CRC-associated microbial patterns may also be influenced by intestinal transit time, intestinal inflammation, and body mass index (Tito R Y, Verbandt et al. 2024). Accordingly, these taxa should be interpreted as candidate microbial features associated with the present case-control comparison rather than features uniquely attributable to CRC. Moreover, differential analysis of shared ASVs showed that several taxa, including the established SCFA producer Faecalibacterium and members assigned to Enterocloster, were less abundant in the CP group. A reduction in established SCFA-producing taxa could contribute to lower SCFA availability, potentially impairing intestinal barrier integrity and promoting chronic inflammation.

To analyze complex, high-dimensional microbiome data, we employed RF and LEfSe for feature prioritization and RF, GBM, and LASSO for exploratory classification. Microbiome studies are affected by population, regional, and sampling heterogeneity (Li et al. 2024, Shi et al. 2024, EL Leithy A A, Youssef A S E et al. 2025), whereas machine-learning approaches can integrate high-dimensional microbial patterns for exploratory classification (Young et al. 2021, Rynazal et al. 2023, Porreca et al. 2024, Konishi et al. 2022). In the context of advanced CRC, for which stage-associated microbiota alterations have been reported (Wu et al. 2025), this approach revealed taxonomically relevant features with biological significance and provided an internally reproducible analytical framework for future external evaluation of these findings. Importantly, the fitted RF feature-prioritization model achieved an OOB accuracy of 0.850, an OOB error rate of 0.150, and an OOB AUC of 0.899. These results indicate that the model captured group-related information, providing a model-level basis for interpreting the MDA feature ranking. The downstream RF, GBM, and LASSO classifiers were used to assess whether the prioritized ASVs retained internal discriminatory information across different modeling frameworks. Nested five-fold cross-validation yielded AUCs of 0.891, 0.900, and 0.891, respectively, providing exploratory internal support for the selected features.

The microbial features identified in this study, such as reduced Faecalibacterium, provide additional clues to the gut microbiota characteristics associated with advanced CRC and may support the future development of microbiome-based classification approaches. Stool-based microbiome analysis has also been proposed as a potentially complementary, non-invasive method alongside conventional examinations such as colonoscopy (Herlo L F, Salcudean et al. 2024, Wu et al. 2025). Nevertheless, the present findings should be regarded as exploratory observations within the studied cohort, and their broader clinical value requires further validation. Preclinical evidence suggests that fecal microbiota transplantation may suppress CRC progression by reshaping microbial dysbiosis and antitumor immune responses (Wu et al. 2025), whereas evidence for durable clinical benefit and improvement in patient outcomes remains limited (Yu et al. 2023). Recent proteomic evidence further suggests that R. intestinalis may modulate acetylation and lactylation patterns in CRC cells, providing an additional mechanistic context for its biological effects (Li et al. 2025). Nevertheless, the clinical relevance of these findings remains to be determined. Therefore, these microbial features may be more appropriately considered as candidates for subsequent classification and translational studies rather than as tools ready for immediate clinical application. Furthermore, given evidence that the gut microbiome may influence responses or resistance to CRC therapy (Xia et al. 2022), these findings may provide a basis for future studies of patient stratification and microbiota-targeted adjunctive interventions.

Several important limitations should be emphasized. This single-center cross-sectional study included only 133 participants and lacked any independent external or prospective validation cohort. Metastatic sites and overall disease burden were not stratified, so the observed taxa may reflect heterogeneous metastatic biology, nutritional/immune status, bowel function, or systemic illness rather than patterns uniquely attributable to CRC. Comparing advanced CRC cases with healthy controls restricted the disease spectrum and did not include clinically relevant disease controls. In multivariable models adjusting for age, sex, and BMI, most RF‑prioritized ASVs remained significant; however, detailed dietary intake, commonly used medications, bowel habits, inflammatory parameters, nutritional indices, and other potentially relevant covariates were unavailable, precluding comprehensive multivariable microbiome analysis incorporating these additional confounders. Residual confounding by unmeasured clinical and lifestyle factors therefore remains possible. Nested cross‑validation reduced information leakage, but all performance estimates derived from the same dataset may still be affected by overfitting and optimistic bias. Future studies should include larger multicenter cohorts, clinically relevant disease controls, stratification of metastatic sites and disease burden, comprehensive confounder assessment, and independent prospective validation, with metagenomic and metabolomic analyses where feasible.

Conclusion

This single-center study identified gut microbial alterations and candidate ASVs associated with treatment-naïve advanced CRC. The OOB performance of the fitted RF model supports the internal coherence of the MDA-based feature ranking, while nested cross-validation indicates that the prioritized ASVs retain exploratory group-discriminatory information across RF, GBM, and LASSO. These findings support further evaluation of the selected microbial features but do not establish microbial features uniquely attributable to CRC or clinical diagnostic models. Independent external validation in larger and clinically representative populations is required.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Na Luo: Conceptualization, Methodology, Formal analysis, Writing-original draft, Writing-review & editing. HuiYan: Investigation, Data curation, Validation, Writing-original draft. Nan Wang: Resources, Supervision, Project administration. Weining Fan: Methodology, Software, Visualization. Ping Chen: Supervision, Writing-review & editing.

Funding

Ningxia Natural Science Foundation (2024AAC03713) and Study on the Role and Mechanism of Karyopherin Alpha 2 (KPNA2)-Mediated Tumor-Associated Macrophage Polarization in Gastric Cancer Progression (2024-10-2026.10).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Conflict of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.