Gut microbiota alterations and their association with tumorigenic pathways in colorectal cancer: insights from a pooled analysis of 109 microbiome datasets

The 16S rRNA gut metagenomics datasets (Accession No. PRJEB7774) from CRC patients were obtained from the Sequence Read Archive (SRA; https://www.ncbi.nlm.nih.gov/sra) [26]. This study included a total of 109 datasets, with faecal samples from 63 healthy individuals (26 females and 37 males) and 46 CRC cases (18 females and 28 males). Detailed information of the datasets, including control and patient characteristics are presented in Supplementary Tables S1 and S2 [26].

The downloaded 16S rRNA metagenomics data was subjected to analysis using Quantitative Insights Into Microbial Ecology (QIIME; version 2022.8) [27]. Within QIIME, the forward and reverse reads of each sample were initially combined, followed by demultiplexing and quality-filtering at a Q-score of 25. The DADA2 algorithm was employed to derive high-quality amplicon sequence variants (ASVs) [28]. Taxonomic profiling was conducted against the Greengenes v13.8 reference database [29], and was converted into relative abundance levels across the categories of phyla, classes, orders, families, genera, and species. Subsequently, utilizing the Greengenes reference database, open-reference operational taxonomic units (OTUs) were selected from non-chimeric sequences with a similarity threshold of 97%. Subsequent to filtering chimeric reads, an average of 52,950,993 clean reads per sample (range 35,322,543 − 75,218,590) was obtained. The entirety of the refined sequences was employed for clustering analysis, resulting in the discovery of operational taxonomic units (OTUs), post elimination of solitary OTUs. Within individual OTUs, the most frequently appearing read was designated as its representative read. Taxonomical attribution of the OTUs was done based upon their alignment with the Greengenes database. Assessment of the variance in the percentage of analysable read counts between HFF and HFS was performed based upon the p-values computed through the Wilcoxon signed-rank test. OTUs with < 10 counts across all samples were eliminated and the remaining OTUs were converted into relative abundance values across phyla, classes, orders, families, genera, and species classifications.

Analyses for intersample (beta; β) and intrasample (alpha; α) diversities were conducted [30, 31]. The α diversity was assessed through multiple metrics: Observed, Chao1, ACE, Shannon index [32], Simpson index [33], Fisher statistics, and phylogenetic diversity [34]. Data pertaining to bacterial genera was provided. To determine β diversity, Bray-Curtis distances were calculated using both UniFrac and Bray-Curtis distance matrices. This resulted in two-dimensional Principal Coordinate Analysis (PCoA) plots. Establishment of the β diversity among the samples was done based upon the weighted and unweighted UniFrac distance matrices [30]. Statistical significances of α diversity measures were evaluated using Mann-Whitney or Kruskal-Wallis tests, while differences in β diversity were determined through permutational Multivariate Analysis Of Variance (MANOVA).

Linear Discriminant Analysis Effect Size (LEfSe) [35] was utilized to detect bacterial species that displayed notable increases or decreases in relative abundance within each specific phenotypic class. A Benjamini-Hochberg false discovery rate (FDR) adjusted p-value threshold of 0.05 was applied, along with a logarithmic linear discriminant analysis (LDA) score cut-off of 2. Subsequently, bar plots depicting the results of LEfSe analysis were generated on the MicrobiomeAnalyst platform [36].

To determine the greatest impactful microbial features, a Random Forest (RF) classification analysis was performed using the MicrobiomeAnalyst platform. The model was constructed with 500 trees (ntree = 500), a minimum node size of 5, and a maximum tree depth of 10. Prior to RF modeling, the data were transformed using the centered log-ratio (CLR) method to account for the compositional nature of microbiome data. As CLR transformation is sensitive to zero values, a small pseudo-count was added to all features to replace zeros and ensure numerical stability during log transformation. Only features with at least 10% prevalence across samples were included. Assessment of the model performance was done by applying a 10-fold cross-validation method. Microbial taxa with a mean decrease in accuracy (MDA) score greater than 0.005 were considered important predictors and were retained as significant features for downstream interpretation.

To explore probable functional associations between gut microbes and host genes, we utilized the MIAOME [37] and GIMICA [38] databases. These resources compile both experimentally validated and computationally predicted microbe-host associations, including microbial metabolites, secreted proteins, and immune-modulatory factors. We input the list of differentially abundant microbial taxa into both databases to retrieve associated human genes based on immune signaling relevance and prior literature evidence. Only genes consistently predicted in both platforms and relevant to inflammation or cancer were retained. These predicted interactions were further validated by comparing gene expression profiles in CRC vs. control tissues using GEPIA2 and TCGA transcriptomic datasets.

The Gene Expression Profiling Interactive Analysis (GEPIA data tool; http://gepia.cancer-pku.cn/) facilitates the exploration of functional genomic datasets to identify connections between genomic and phenotypic factors. CRC tumor samples (n = 275) and normal colon tissues (n = 349) were selected using default parameters. To minimize batch effects between datasets, the “Match TCGA normal and GTEx data” option was enabled, ensuring reliable comparison of gene expression levels between CRC and control groups. In this study, the tool was utilized to investigate potential associations between the expression of crucial genes and the expression profiles of CRC patients within the TCGA dataset. Validation of these key genes was carried out through box plots, alongside an examination of the pathological stage and transcript per million data. Statistical significance was determined at the cut-off of p < 0.05. The gene expression analyses were compared under default normalization settings, with statistical cut-offs set at |log2 fold change| ≥ 1 and p-value < 0.01.

16S rRNA amplicon sequencing data collected from healthy controls and CRC cases was employed to discern diversity and abundances of the resident microbial communities. Significant alterations in the gut microbial diversity between CRC cases and undiseased controls were observed. The β diversity assessed through PCoA plots indicated a clear distinction in bacterial population patterns between control and CRC faecal microbiota, supported by substantial Bray-Curtis distances (R² = 0.02, p = 0.003), Jensen-Shannon divergence (R² = 0.03, p = 0.002) and Jaccard distances (R² = 0.02, p = 0.003) (Fig. 1A). Regarding α diversity, notable decreases were exhibited in CRC faecal microbiota compared to control samples (Fig. 1B), implying reductions in the variety of inhabiting bacterial species during disease progression. This could be attributed to the adverse impact of a dysbiotic microbiota environment on certain healthy bacteria. Additionally, the α diversity assessed through box plot indicated a clear distinction in bacterial population structures between control and CRC faecal microbiota, supported by a substantial Observed richness (p = 0.0002), Chao1 (p = 0.0001), Fisher (p = 0.0002) and Shannon (p = 0.025).

The overall composition of the microbiota was altered in the patients with CRC cancer compared with controls, and the abundances of several taxa were elevated in samples obtained from the faecal microbiota. At the phylum level, Firmicutes, Actinobacteria, and Proteobacteria were found to be greatly enriched in CRC cases (Fig. 2A). Additionally, At the class level, the most enriched bacterial classes in CRC patients included Deinococci, Candidatus, Saccharibacteria, Betaproteobacteria, Deltaproteobacteria, Erysipelotrichia, Bacilli, and Negativicutes (Fig. 2B). At the order level, taxa such as Clostridiales, Bacteroidales, Bifidobacteriales, Enterobacteriales, and Verrucomicrobiales showed increased abundance in CRC samples (Fig. 2C) (Figure S1-S6). At the family level, prominent groups enriched in CRC patients included Ruminococcaceae, Lachnospiraceae, Bifidobacteriaceae, Enterobacteriaceae, Bacteroidaceae, and Eubacteriaceae. At the genus level, key genera such as Ruminococcus, Bifidobacterium, Subdoligranulum, Escherichia, and Bacteroides were more abundant in the CRC group (Supplementary Figures S1–S6). At the species level, several taxa were enriched in CRC patients, including Subdoligranulum (unclassified), Escherichia coli, Ruminococcus bromii, Bifidobacterium adolescentis, Faecalibacterium prausnitzii, Ruminococcus sp. 5_1_39BFAA, Akkermansia muciniphila, Eubacterium rectale, Prevotella copri, Methanobrevibacter smithii, Dialister invisus, Alistipes putredinis, and Bacteroides uniformis (Fig. 2D).

For enhanced understanding of the potential impact of modified microbiota in CRC patients, our subsequent analyses focused on discerning the variations in the gut microbial composition between controls and CRC cases. Compositional methods were employed to evaluate the distance patterns in order to recognize the correlations of taxa in relation to control and CRC faecal groups. The correlations were evaluated by comparison of relative microbial abundances using LEfSe, which explores variations in the abundance of microbiome features (such as clades, OTUs, etc.) between controls and cases. LEfSe analysis validated the distinct separation of microbiome features enriched in normal controls vs. CRC cases (Fig. 3A). Taxa reported as enriched in CRC patients were identified using LEfSe analysis, with significance determined based on an FDR-adjusted p-value < 0.05 and an LDA score > 2. The RF classification model demonstrated effective separation of CRC and control samples, with the error rate notably lower for the control group (green line) and CRC group (blue line), compared to the overall classification error (red line) across 500 trees (Fig. 3B). Feature importance analysis (Fig. 3C) highlighted Prevotella copri, Gemella morbillorum, Parvimonas unclassified, Ruminococcaceae bacterium D16, and Dialister invisus as the top microbial species contributing to CRC classification, based on their Mean Decrease in Accuracy (MDA). Several of these taxa also overlapped with those identified in LEfSe analysis, underscoring their potential as robust microbial biomarkers for CRC diagnosis and risk stratification.

Compared to controls, a total of 35 microbial taxa were significantly altered in CRC cases, including 28 that were downregulated and 7 that were upregulated (Fig. 4A). Differential abundance was determined using the LEfSe algorithm, applying an FDR-adjusted p-value < 0.05 and a logarithmic LDA score > 2 to identify statistically and biologically meaningful changes. With regards to the differential microbiome features ranked according to effect sizes, the taxa Prevotella copri, Methanobrevibacter smithii, Bacteroides eggerthii, Dialister invisus, Alistipes onderdonkii, Bacteroides caccae, and Parabacteroides merdae were enriched in CRC cases while Pseudomonas unclassified, Clostridium hathewayi, Clostridium symbiosum, Bifidobacterium animalis, Streptococcus thermophilus, Ruminococcus_sp.5_1_39BFA were repressed (Fig. 4B). This is in consistent with LEfSe-driven analysis which also aligns with the conclusion that CRC faecal microbiota contain fewer bacterial features. Furthermore, we performed the association of microbial biomarker with clinical phenotypes or diseases.

Finally, the association between differential microbiome changes in CRC cases (vs. controls) and their functional annotation were predicted. The interactions between key resident microbial changes and host genes were predicted in accordance with previous literatures findings using MIAOME and GIMICA databases [37]. Regarding microbial biomarker association with clinical phenotypes or diseases, conditions such as cirrhosis disease, inflammatory bowel disease, type 2 diabetes, fatty liver, obesity, NAFLD, colorectal neoplasma and CRC were found to relate predominantly with Bacteroides caccae, Bacteroides eggerthii, Alistipes onderdonkii, Bifidobacterium animalis, Clostridium hathewayi, Clostridium symbiosum, Dialister invisus, Methanobrevibacter smithii, Parabacteroides merdae, Prevotella copri, Rothia dentocariosa, Ruminococcus sp 5 1 39BFAA and Streptococcus thermophilus (Fig. 5A). Further, the interaction of key microbes (viz., Streptococcus thermophilus, Bacteroides caccae, Bacteroides eggerthii, Bifidobacterium animalis, Methanobrevibacter smithii, Parabacteroides merdae, Prevotella copri, Rothia dentocariosa, Ruminococcus sp 5 1 39BFAA) and host genes identified key inflammatory and immune-metabolic pathways related host genes (Fig. 5B).

To explore the downstream impact of microbiome alterations on host gene expression, we validated the expression of six inflammation-associated genes CD44, CXCL8, DUSP16, FOXP3, IFNGR2, and IL18 predicted from MIAOME and GIMICA interactions (Fig. 5A–B). Gene expression data from TCGA for the CRC cases (n = 275) and undiseased controls (n = 349) were analyzed using GEPIA2. All six genes showed significantly greater expression in CRC cases versus controls (Fig. 5C; p < 0.01). Notably, CXCL8 and IL18, key cytokines in the inflammatory cascade, exhibited the most pronounced upregulation. These findings support the idea that dysbiosis of the resident microbiome may influence host transcriptional responses through immune-regulatory mechanisms. The observed expression changes are consistent with predicted microbial-host gene associations shown in Fig. 5A-B, wherein taxa such as Prevotella copri, Methanobrevibacter smithii, and Bacteroides eggerthii were linked to immune and inflammatory gene regulation.

Further, GO functional enrichment analyses of predicted host genes was undertaken. The significant pathways which were identified related to inflammation, NF-κB signalling, cellular response to lipopolysaccharide, Toll-like receptor signalling, positive regulation of JNK cascade, inflammatory bowel disease signalling, alcoholic liver disease, lipid and atherosclerosis, and apoptosis (Fig. 6A). Further, the host genes were also involved in important biological processes that were associated with the tumorigenesis and metabolic dysfunction, and included prominent biological process related to inflammation, immune response, cell proliferation, cytokine production, apoptosis and cell differentiation (Fig. 6B).