Exploring the gut microbiota associated with peripheral nerve invasion in colorectal cancer patients and constructing predictive models

Abstract

Objective

This study aims to explore the differences in composition, abundance, and biological functions of the gut microbiota between colorectal cancer (CRC) patients with peripheral nerve invasion (PNI) and those without peripheral nerve invasion (NPNI). Additionally, we tried to construct a machine-learning predictive model incorporating the identified microbiota characteristics to explore the impact of gut microbiota on CRC-PNI progression and to search for new non-invasive microbiological indicators for CRC-PNI. Finally, we successfully developed a predictive model to predict PNI in CRC patients through leveraging microbial biomarkers. This innovative approach is expected to offer a novel strategy for the early detection of CRC metastasis, thereby facilitating more informed decisions regarding treatment options.

Method

This study included 132 colorectal cancer (CRC) patients, who were divided into two separate groups according to whether they exhibited PNI. The gut microbiota of these participants were subjected to 16S rRNA gene sequencing, followed by a thorough analysis to identify any significant differences between the groups. We applied a cell sorting algorithm to convert the transcriptome sequencing data obtained from 8 colorectal cancer patients into a matrix representing immune cell abundance. Following this, the matrix was utilized to investigate the associations among the PNI-related distinct gut microbiota, immune cells, and immune-related genes, and PNI-related differentially expressed genes (or molecular markers, pathways), as well as their associations with KEGG pathways. Based on the differential gut microbiota, we constructed Random Forest (RF) and Multilayer Perceptron (MLP) models to predict PNI in CRC patients.

Result

Comparative analysis of α-diversity and β-diversity in the gut microbiota of CRC patients with and without PNI revealed no statistically significant differences (P > 0.05). However, Linear Discriminant Analysis effect size (LEfSe) identified 35 distinct gut microbiota, with 28 species enriched in the PNI group and 7 species significantly enriched in the NPNI group. By analyzing the gut microbiota significantly associated with PNI, we successfully constructed predictive models using RF and MLP that can predict the occurrence of PNI in CRC patients. Both models have demonstrated robust performance.

Conclusions

In the PNI and NPNI groups, 35 gut microbiota species exhibited significant variations in abundance. The differential intestinal microbiota associated with PNI in colorectal cancer may modulate the neuroinvasion process via a variety of potential biological mechanisms. The RF and MLP predictive models show considerable accuracy in predicting CRC-PNI status and are of reference value.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12866-025-04179-x.

Introduction

Colorectal cancer (CRC) ranks as the third most common cancer globally, and in developing countries, both its incidence and mortality rates are on the rise annually, with a notable trend toward earlier age of onset [1]. According to the latest cancer statistics in 2022, the incidence rate of CRC patients in China is ranked second and the mortality rate is ranked fifth [2]and there is a difference between men and women, with the incidence rate and mortality rate of men being higher than that of women [3]. CRC accounts for 10% of all cancers and cancer-related deaths diagnosed annually [4]and the global incidence of CRC patients is projected to increase to 2.5 million cases by 2035⁵. Smoking, Alcohol Consumption, and Changes in Dietary Habits as Risk Factors for the Development of CRC Patients [6]. Immune cells, cytokines and immunomodulatory factors play an important role in the development of CRC. The human gastrointestinal tract is a complex and varied environment where 100 trillion microorganisms live, and differences in the gut microbiota result in different production of immune cells, cytokines and immunomodulatory factors by the body. The colon region is a huge host of microbial communities. The gut microbiota is mainly categorized into phylum Thick-walled bacilli, phylum Anaplasma, phylum Actinobacteria, phylum Clostridium, phylum Aspergillus, and phylum Warty Microbacteria [7]. From the proximal to the distal part of the gastrointestinal tract, there is a gradual increase of about 70% in the number of microbiota [8]. The long co-evolution of the host and the gut microbiota has led to the interdependence of the two in relation to CRC, and the gut microbiota play a very important role in human health. They can have both favorable and harmful effects on the organism and the host. Intestinal commensal microbiota play an important role in resisting pathogen invasion, host immune response, nutrient metabolism and food digestion. According to the current research progress, the mechanism of harmful effects mainly includes the following aspects: (1) the effects of gut microbiota metabolites and toxins; (2) host response and inflammatory response. Disruptions in the gut microbiota equilibrium may trigger chronic inflammation and the generation of cancer-causing metabolites, potentially resulting in tumor development. Gut microbiota is also involved in the anti-tumor response in CRC [9].

Changes in gut microbiota can cause the process of colorectal cancer, so monitoring changesin gut microbiota status is important for early warning of colorectal cancer, and gut microbiota can also be used as a predictive tool for the effectiveness of tumor treatment and patient prognosis [10]. The characteristics of gut microbiota can better predict tumor growth and tumor treatment outcome, Heshiki et al. [11] explored the relationship between gut microbes and colorectal cancer using fecal specimens from healthy people, tumor patients, and it was obvious that tumor patients in the treatment-effective group had a higher intestinal microbial diversity, there was a significant difference in the composition of the gut microbiota, and there was a positive correlation between the specific microbiota and treatment outcome There was a significant difference in the composition of the gut microbiota. The risk of advanced CRC was further increased in patients with inflammatory bowel disease and Helicobacter pylori-positive chronic atrophic gastritis, compared with the healthy population [12, 13]. It is now believed that CRC is a disease in which tumor cells, non-tumor cells (i.e., stromal cells), and a large amount of gut microbiota are involved. Therefore, gut microbiota modulation is a potential strategy for colorectal cancer control.

Tumor invasion and metastasis are important causes of poor prognosis in patients with colorectal cancer [14]and during tumor growth, tumor cells can spread along blood vessels, lymphatic vessels, or nerves [15]. Previous studies [16, 17] have shown that peripheral nerve invasion is closely associated with the malignant pathological phenotype and poor prognosis of colorectal cancer, and in addition, some studies [18] have shown that peripheral nerve invasion can be used as a prognostic predictor of adjuvant chemotherapy for colorectal cancer. Peripheral nerve invasion (PNI) was first identified in head and neck tumors, where tumor cells often metastasize along nerve fibers into the skull, thus affecting patient prognosis. PNI is defined as the infiltrative spread of tumor cells in or around nerve tissue, even in the absence of lymphatic infiltration (LII) or lymph node metastasis. Controls were CRC patients without peripheral nerve invasion (NPNI). NPNI is defined as no infiltrative spread of tumor cells in or around nerve tissue. Recent research indicates that in prostate and gastrointestinal malignancies, nerve fibers that infiltrate can catalyze the expansion, multiplication, and spread of cancerous cells [19–21]. This may be an important reason why PNI leads to poor prognosis in colorectal cancer patients. Previous studies have usually relied on histologic typing and pathologic staging for the prognosis of patients with colorectal cancer, which has a certain margin of error [22]. The more certain histopathological factors associated with colorectal tumor development and invasion include poor differentiation, T-staging, lymphatic and venous invasion, etc.; some studies have shown that PNI can be used as a new biological feature, and foreign studies have shown that the incidence of PNI in patients with gastrointestinal tract tumors ranges from 13 to 39%, with an average detection rate of 26% [23–25]. Exploring its role in colorectal cancer can help predict patient screening invasive and metastatic behavior of colorectal cancer and determine a more accurate step in surgical treatment [14]. This study sought to explore the variations in composition, prevalence, and biological roles of the gut microbiota among PNI-CRC patients and NPNI-CRC patients. Additionally, it aimed to develop machine learning predictive models based on distinctive microbiota features to elucidate the impact of gut microbiota on CRC-PNI progression and to identify novel non-invasive microbial biomarkers for CRC-PNI. Predicting PNI in CRC patients by microbial biomarkers provides a new strategy for early screening of CRC metastasis and determining treatment options.

Research methods

Participant data and specimen procurement

The research plan received approval from the Medical Ethics Committee of the Guangxi Medical University Cancer Hospital. Informed consent was obtained from participants before sample collection. Prior to commencing any study-related procedures, each participant gave their informed consent through the signing of the relevant documents. Between January 1, 2021, and December 31, 2021, research staff gathered pre-treatment fecal specimens from 236 colorectal cancer patients in accordance with the established inclusion and exclusion criteria. After quality assessment, a total of 198 fecal samples met the required standards and were selected for 16S rRNA sequencing. Out of the study participants, 132 colorectal cancer patients had available PNI data. Additionally, we performed transcriptome sequencing on tumor tissue samples from 17 colorectal cancer patients, among which 9 samples had 16S rRNA sequencing results of the corresponding patients’ fecal samples. We selected 8 samples with PNI information from these 9 samples for further study.

Study eligibility was assessed based on the following criteria: (1) Individuals who received surgical intervention with a confirmed pathological stage according to the AJCC CRC staging criteria, or those identified with colorectal adenocarcinoma through colonoscopic biopsy; (2) Absence of any comorbidities or history of other malignant tumors; (3) Participants were not included if they had other intestinal conditions without acute complications, such as complete bowel obstruction, bowel perforation, or pelvic abscess; (4) No previous anti-cancer therapies, such as surgery, chemotherapy, radiation therapy, immunotherapy, or traditional Chinese medicine, were administered prior to sample collection; (5) Participants who had taken antibiotics or intestinal microecological agents within the month preceding the study were excluded; (6) No consciousness deficits or additional cognitive impairments.

Stool sample collection: Stool samples were obtained from patients on the initial day of their hospital stay. Participants were advised to use sterile stool collection containers, with an emphasis on collecting the middle portion of the stool to minimize the risk of urine contamination. Subsequently, the specimens were placed into a sterile ice box. Each sample was divided into 2 mL Eppendorf (EP) tubes, containing roughly 200 mg per tube, and kept at −80 °C in a deep freezer for long-term preservation.

Collection of CRC tissue samples: Fresh CRC tissue samples were collected post-surgical resection. Tumor tissues, approximately the size of a soybean, along with adjacent paracancerous tissues, were sectioned. The specimens were then promptly dissociated and placed into liquid nitrogen for cryopreservation, with the entire process from dissociation to freezing being completed within a 30-minute timeframe.

16S rRNA sequencing

In accordance with the manufacturer’s instructions, Nucleic acids were extracted from 200 mg of fecal material using the MOBIO PowerSoil^® DNA Isolation Kit. The feces were first dissolved in Tris-EDTA buffer as per the kit’s protocol. Post-extraction, an evaluation of the DNA integrity within the samples was undertaken. Only samples meeting the quality standards proceeded to the next stages of the experimental process. The 16 S rRNA gene was targeted for amplification, focusing on the V3 and V4 regions. Primers 341 F (5′-CCTACGGGGNGGCWGCAG-3′) and 805R (5′-GACTACHVGGGGTATCTAATCC-3′) were employed to specifically amplify the target DNA sequences. The targeted DNA fragments were then subjected to polymerase chain reaction (PCR) to generate sufficient quantities for further analysis. Following For the PCR amplification process, an initial appraisal of the amplicons derived from each specimen was performed utilizing 2% agarose gel electrophoresis. The target bands for the amplified products were expected to be within the size range of approximately 300 to 350 base pairs (bp). Subsequently, we quantified the PCR products using the Quant-iT PicoGreen dsDNA Assay Kit. Based on the quantification results, all samples were pooled to achieve equimolar concentrations, aligning with the requirements for subsequent sequencing. The pooled libraries were then quantified with the KAPA Library Quantification Kit KK4824 to accurately measure DNA concentration. Following this, sequencing of the libraries was performed on an Illumina PE250 platform, employing a 2 × 250 bp sequencing chemistry. which allows for paired-end sequencing with a read length of 250 bases per end.

Transcriptome sequencing

Utilizing the Total RNA Extraction Kit, total RNA was extracted from 17 CRC tissue samples. The integrity of the RNA was assessed through electrophoresis, a standard method for visualizing RNA quality. Additionally, the purity of RNA samples was assessed using a micro UV spectrophotometer, offering a quantitative assessment of RNA purity through absorbance measurements at specific wavelengths. Following the extraction, rRNA was depleted from the total RNA samples to enrich for mRNA. Subsequently, cDNA libraries were synthesized by adhering to the protocols provided in the VAHTS™ Stranded mRNA-seq Library Prep Kit for Illumina^®. This kit is designed to generate high-quality, strand-specific sequencing libraries suitable for RNA-sequencing applications. Transcriptome library sequencing was performed on the Illumina NovaSeq 6000 platform, known for its high-throughput sequencing data generation capabilities. The raw sequencing data underwent quality assessment using the FastQC tool to ensure the integrity and quality of the reads. Subsequently, the valid data from the samples were aligned to the human reference genome (version hg38) using HISAT2, a robust alignment algorithm designed for accurate mapping of sequencing reads to a reference genome. Gene expression levels were determined using StringTie, a software tool that employs known gene models to estimate the expression of each gene. Gene expression abundance was quantified using Transcripts Per Million (TPM), a normalization method that accounts for sequencing depth and gene length variations, enabling comparison of expression levels across different samples.

Analysis of immune cell infiltration in tumors

Tumor immune cell infiltration was analyzed using the CIBERSORT R script version 1.03. This algorithm leverages microarray data to construct a feature matrix that enables the translation of the Transcripts Per Million (TPM) matrix into a comprehensive profile of 22 distinct immune cell types. This includes various cell types and functional states, offering a detailed representation of the immune landscape within the tumor microenvironment [26].

Functional enrichment analysis of transcriptome sequencing associated with PNI

The single-sample gene set enrichment analysis (ssGSEA) algorithm assigns a score to each sample, reflecting the extent of enrichment for a particular gene set as determined by a rank-ordered methodology [27]. We obtained the necessary gene set files in GMT format, including ‘c2.cp.kegg.v2022.1.Hs.symbols.gmt’ and ‘c5.go.v2022.1.Hs.symbols.gmt’. We then calculated the gene set scoring matrix for each sample using the ssGSEA algorithm, which is part of the GSVA package version 1.46.0. Subsequently, we utilized the NPNI group as a control to assess differential Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways against the experimental groups. Data retrieval and preprocessing were performed using the TCGAbiolinks package (version 2.25.3). Differential abundance analysis was subsequently conducted with the limma algorithm (version 3.54.0). Gene Ontology (GO) analysis covered three categories: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC). The threshold for identifying differentially expressed genes was established at a p-value below 0.05 and an absolute log2 fold change (|log2FC|) exceeding 0. A log2FC greater than 0 signified up-regulation in the PNI group, while a log2FC less than 0 indicated up-regulation in the NPNI group. The results of this screening were visualized using volcano plots, which were generated with the ggplot2 package version 3.4.0.

Constructing machine learning models and identifying gut microbial biomarkers

We employed Random Forest (RF) model and Multilayer Perceptron (MLP) model to identify gut microbial biomarkers for predicting PNI in CRC patients.

RF and MLP models are esteemed machine learning techniques, renowned for their robust predictive capabilities [28–30]. RF modeling is a synergistic machine learning approach that harnesses the power of an ensemble of decision trees, working in unison to make robust predictions [31]. The RF method amalgamates numerous decision trees, transcending the limitations of a singular model. As the ensemble expands, the generalization error diminishes, thereby mitigating the risk of overfitting [31]. Distinct from other models, the RF model adeptly manages complex interactions among numerous independent variables without being susceptible to multicollinearity issues [31, 32]. Consequently, RF models often exhibit greater accuracy compared to other machine learning approaches [31, 32]. The MLP model, namely the multilayer perceptron model, is a type of feedforward artificial neural network (ANN), which functions to transform a set of input vectors into corresponding output vectors. This model can be viewed as a directed graph structure made up of multiple layers, with each layer’s nodes being fully interconnected with the nodes of the next layer. MLP model, as an extension of the perceptron, addresses its limitation of being unable to recognize linearly inseparable data. The MLP model features several key aspects: (1) highly parallel processing capabilities; (2) robust nonlinear global functionality; (3) strong fault tolerance; (4) associative memory capabilities; (5) exceptional self-adaptation and self-learning capabilities. Therefore, we decided to use the above two models for prediction. We utilized the Python programming language and the scikit-learn library(https://scikit-learn.org/stable/) to install the necessary packages for constructing and evaluating machine learning models. In constructing the models, we followed the standard parameter configurations as prescribed by the library. The gut microbiota data from 132 CRC patients with PNI information served as the basis for our analysis. These data were divided into a training subset and a testing subset. We utilized the microbial species present in the gut microbiota as features for our models. The RF and MLP models were constructed and utilized for prediction based on gut microbiota species that exhibited significant differences and were distinctly associated with PNI. The effectiveness of these models was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC) as metrics. The AUC was used to assess the accuracy of the machine learning models on both the training and validation datasets.

Statistical approaches for analyzing 16S rRNA sequencing data

The subsequent analyses were performed using R software version 3.5.1, where all p-values were computed as two-tailed and deemed statistically significant if p < 0.05. We initiated our workflow by acquiring FASTQ format raw sequencing data for each sample. Quality filtering of the raw data was accomplished using Quantitative Insights Into Microbial Ecology version 2 (QIIME2). Subsequently, species annotation was conducted utilizing the Greengenes database version 13.8. For the extraction of gut microbiota ASVs/OTUs, we employed the phyloseq package version 1.26.1. Initially, we assessed the species diversity within each gut microbiota sample using α-diversity metrics. The Chao1 and ACE indices were used to assess the richness and evenness of the microbiota, respectively, whereas the Shannon and Simpson indices were utilized to evaluate the diversity of the microbial community. β-diversity was measured to capture the variability in species composition across gut microbiota samples within the same ecosystem. For β-diversity analyses, we conducted ADONIS and ANOSIM analyses using the vegan package version 2.5.6. LEfSe analysis supports high-dimensional categorical comparisons [33] and can be used to pinpoint species most likely responsible for the observed differences between groups. Subsequently, we used LEfSe software version 1.0.0 to perform LEfSe analysis. This analysis was used to assess the results, particularly to measure the effect size of each species showing significant differences between groups (i.e., The LDA score, which is logarithmically transformed with base 10 by default, indicates the discriminative power of each species. The greater the absolute value of the LDA score, the more pronounced the distinction between the groups); Using|LDA|>2 with P < 0.05 as the thresholds, we identified species with significantly different abundances between groups. The outcomes were visualized in bar graphs. Ultimately, we conducted KEGG pathway prediction for intergroup enrichment analysis on the sequenced samples utilizing PICRUSt2 software version 2.3.0. The nonparametric Mann-Whitney U rank sum test, aided by the vegan package version 2.5.6, was used to compare α and β diversity indices and KEGG pathways between groups. The results were then visualized as bar graphs using the ggplot2 package version 3.4.0.

Statistical methods

All subsequent analyses were conducted using R software version 4.2.2, where all p-values were computed as two-tailed and deemed statistically significant if p < 0.05. For continuous clinical data, t-tests were employed, while Pearson’s chi-square test was applied to categorical data (Table 1). These calculations were performed using SPSS software version 23.0. Pearson’s correlation was utilized to determine the relationships between gut microbiota and the abundance of immune cells and immune-related genes. In contrast, Spearman’s correlation was applied to assess the associations between dominant microbial subgroups and between gut microbiota and KEGG pathways. These analyses were conducted using the Hmisc package version 4.7-1. Spearman’s correlation was used to assess the relationships between gut microbiota and Biological Process (BP) as well as Molecular Function (MF) items. These correlations were computed using the ggcorrplot package version 0.1.4. Subsequently, the correlation matrix was visualized as heatmaps, which were generated with the Igraph package version 1.3.5 and Cytoscape software version 3.7.2.

Table 1.

CRC patient demographics and clinical features categorized by PNI status

		CRC patients with PNI(n = 73)	CRC patients without PNI (n = 59)	P value	Test
Age (years, mean (SD))		58.05 ± 11.59	57.80 ± 10.38	0.183	T-Test
Age (%)	≥ 60	32(43.8)	34(57.6)	0.115	Pearson Chi-square
	< 60	41(56.2)	25(42.4)
Gender (%)	male	45(61.6)	36(61.0)	0.941	Pearson Chi-square
	female	28(38.4)	23(39.0)
TNM stage (%)	early (0 ~ 2)	24(32.9)	34(58.6)	0.003**	Pearson Chi-square
	advanced (3 ~ 4)	49(67.1)	24(41.4)
LVI invasion (%)	YES	41(56.2)	5(8.5)	< 0.001***	Pearson Chi-square
	NO	32(43.8)	54(91.5)

Result

Colorectal cancer patient demographics and clinical features categorized by PNI status

After the data collection, this study included 132 CRC patients with information regarding PNI. Among them, 73 patients exhibited PNI, whereas the other 59 patients did not have PNI. No statistically significant differences were found between CRC patients with and without PNI regarding age and gender, suggesting that the baseline data were well-balanced and comparable; there were statistically significant differences in terms of TNM staging (P = 0.003) and PNI (P < 0.001). (See Table 1 for specific data)

Microbiota diversity comparison between CRC patients in the PNI and NPNI groups

We assessed the gut microbiota diversity in CRC patients, with differences in six alpha diversity indices of gut microbiota samples between the PNI and NPNI groups illustrated in Fig. 1A, however, these differences did not achieve statistical significance (P > 0.05). Figure 1B illustrate the beta diversity profiles of CRC patients in both the PNI and NPNI groups. The gut microbial compositions of CRC patients in the PNI and NPNI groups showed no significant differences in Bray (P = 0.7487) and Jaccard test indices (P = 0.7230) (Supplementary Tables 1 and 2). The aforementioned findings indicate that there were no statistically significant differences in α-diversity and β-diversity of the gut microbiota between CRC patients in the PNI and NPNI groups. This suggests that the microbial community structure and species diversity did not significantly differ between the two sample sets.

Fig. 1.

A Comparison of the alpha diversity index of gut microbiota in the PNI and NPNI groups of CRC patients.B Comparison of beta diversity index of gut microbiota in PNI group and NPNI group of CRC patients. The horizontal axis denotes the group, the vertical axis shows the community diversity index value for samples within that group, and the color codes the group

Identification of gut microbiota associated with PNI

We performed LEfSe analysis of the abundance of the fecal microbiota of CRC patients in the PNI group (Y group) versus the NPNI group (N group), and found that the abundance of a total of 35 species was significantly different. Among these, 28 species were found to have higher abundance in the PNI group, while 7 species had higher abundance in the NPNI group (Supplementary Table 3, Figure 2A). Figure 2B presents an LDA bar chart depicting the LDA scores from the LEfSe analysis for each significantly different bacterial group, where higher scores denote a more substantial impact of the respective taxon.

Fig. 2.

Differential gut microbiota analysis of CRC patients in the PNI and NPNI groups. A LEFSe analysis: evolutionary relationship diagram. Node size reflects species abundance, with larger nodes indicating greater abundance. Node color denotes group classification: yellow nodes signify species with no significant abundance differences between groups; red nodes highlight species with significantly higher abundance in the PNI group; green nodes highlight species with significantly higher abundance in the NPNI group. Nodes in each layer represent taxonomic ranks from phylum to species (inside-out), while annotations for species markers in each layer provide taxonomic information from species to phylum (outside-in). B LDA bar diagram originating from 16S rRNA gene sequencing. The hue of the bar diagram signifies groups exhibiting markedly elevated abundance, and the length of each bar is proportional to the extent of the LDA score. C Correlation network diagram of PNI-associated gut microbiota. Each node in the graph corresponds to a distinct species, with the color of the node indicating its group affiliation. The size of a node is proportional to the number of connections it has within the network; larger nodes signify a greater number of connections. The lines between nodes indicate significant correlations, with blue lines representing negative correlations (Spearman correlation coefficient values less than 0) and red lines indicating positive correlations (Spearman correlation coefficient values greater than 0). The thickness of the line is directly related to the strength of the Spearman correlation coefficient, with thicker lines denoting stronger correlations

Furthermore, we depicted graphs of the relevance between the dominant microbiota of the PNI group and those of the NPNI group (Fig. 2C), The bacteria f__Ruminococcaceae.g__Ruminococcus, o__Clostridiales.f__Clostridiales_Incertae_Sedis__XIII, f__Peptostreptococcacea.g__Romboutsia, f__Lachnospiraceae.g__Clostridium_XIVa and g__Holdemania.s__uncultured_bacterium are the top 5 bacteria with the most connections to other nodes, showing the closest correlation with other dominant bacteria. A significant negative correlation was observed between the dominant microbiota in the NPNI group and that in the majority of the PNI groups. The above findings indicate a possible regulatory association among these dominant groups.

Differential functional pathways in the gut microbiota of CRC patients in the PNI and NPNI groups

In the fecal samples of CRC patients from the PNI and NPNI groups, significant differences in the abundance of three KEGG pathways were observed among those identified through microbial gene co-identification. These pathways were ko00514: Other types of O-glycan biosynthesis (P = 0.007), ko00601: Glycosphingolipid biosynthesis - lacto and neolacto series (P = 0.030), and ko04080: Neuroactive ligand-receptor interaction (see Supplementary Tables 4 and Supplementary Fig. 1 for details). Among them, the abundance of the first two pathways was significantly higher in the PNI group than in the NPNI group. Whether these two pathways are associated with CRC-PNI requires further investigation.

Correlation between PNI-related gut microbiota differences and tumor-infiltrating immune cells

Tumor-infiltrating immune cells are a vital part of the tumor microenvironment and help regulate the local immune response within tumors, making them a potential target for immunotherapy. The strip chart illustrates the composition of 22 immune cells in 8 CRC patients who have both PNI and transcriptome sequencing data (e.g., Fig. 3A). Overall, patients in the NPNI group had more plasma cells than those in the PNI group. Conversely, patients in the PNI group showed higher levels of memory B cells and naive B cells compared to the NPNI group. Figures 3B-D show the correlation of 22 immune cells in the predominant microbiota of the PNI and NPNI groups. In the PNI group, f__Oxalobacteraceae.g__Oxalobacter and g__Bacteroides.s__Bacteroides_salyersiae showed a significant positive correlation with Macrophages M1. Additionally, f__Peptostreptococcaceae.g__Romboutsia, g__Lachnospiracea_incertae_sedis.s__Eubacterium_eligens_ATCC_27750, o__Clostridiales.f__Clostridiaceae_1, and f__Clostridiaceae_1.g__Clostridium_sensu_stricto were significantly positively correlated with T cells CD4 naive. Furthermore, g__Lachnospiracea_incertae_sedis.s__Eubacterium_eligens_ATCC_27750, o__Clostridiales.f__Clostridiaceae_1, and f__Clostridiaceae_1.g__Clostridium_sensu_stricto, as well as g__Bacteroides.s__Bacteroides_salyersiae, showed significant positive correlations with Monocytes. In addition, other gut microbiota also exhibit a positive correlation with tumor-infiltrating immune cells(e.g., Fig. 3C). In the NPNI group, g__Anaerobacterium.s__uncultured_bacterium exhibited a significant positive correlation with Monocytes and Macrophages M1.

Fig. 3.

Correlation of PNI-related differential gut microbiota with tumor immune infiltrating cells. A Bar chart of immune cell abundance in CRC patients by PNI status. Each bar represents an individual sample, with the vertical axis displaying the predicted relative abundance of immune cells. The relative abundances of all immune cells within a sample sum to 1. The graph uses distinct colors to represent each type of immune cell. B Correlation network diagram between PNI-associated gut microbiota differences and immune cells. In the graph, each node denotes a distinct intestinal bacterium or immune cell. The color of the node signifies its group classification. Interconnecting lines indicate significant correlations between pairs of nodes, with blue lines representing negative correlations (Pearson correlation coefficient values less than 0) and red lines signifying positive correlations (Pearson correlation coefficient values greater than 0). C Heat map of the correlation between the dominant microbiota and the abundance of immune cells in the PNI group. D Heat map of the correlation between the dominant microbiota and the abundance of immune cells in the NPNI group. The heatmap plots immune cells along the horizontal axis and bacteria along the vertical axis. Red colors within the graph denote positive correlations, while blue colors indicate negative correlations. The intensity of the colors reflects the magnitude of the Pearson correlation coefficients, with shades ranging from light to dark corresponding to coefficients increasing from small to large. The "*" in the graph indicates the size of P-value: no * for P-value ≥ 0.05, * for 0.01 ≤ P<0.05, ** for 0.001 ≤ P<0.01, *** for P<0.001

Correlation of different gut microbiota with immune-related genes

In the PNI group, dominant bacteria such as g__Clostridium_IV.s__uncultured_Clostridiaceae_bacterium, g__Christensenella.s__Christensenella_minuta, g__Bacteroides.s__Bacteroides_salyersiae, and f__Oxalobacteraceae.g__Oxalobacter showed significant positive correlations with multiple checkpoints (e.g., Fig. 4A), immune activation genes (e.g., Supplementary Fig. 2), immunosuppressive genes (e.g., Supplementary Fig. 3), chemokines (e.g., Fig. 4B), and chemokine receptors (e.g., Supplementary Fig. 4). In the NPNI group, the dominant bacterium g__Anaerobacterium.s__uncultured_bacterium exhibited significant correlations with multiple checkpoints (e.g., Fig. 4C), immune activation genes (e.g., Supplementary Fig. 5), immunosuppressive genes (e.g., Supplementary Fig. 6), chemokines (e.g., Fig. 4D), and chemokine receptors (e.g., Supplementary Fig. 7). The aforementioned findings indicate that PNI-associated variations in gut microbiota could affect the expression of immune-related genes.

Fig. 4.

Correlation of PNI-related differential gut microbiota with immune-related genes. A Heatmap of correlation between predominant microbiota and checkpoints in the PNI group. B Heatmap of correlation between predominant microbiota and chemokines in the PNI group. C Heatmap of correlation between predominant microbiota and checkpoints in the NPNI group. D Heatmap of correlation between predominant microbiota and chemokines in the NPNI group. In the heatmap, genes are represented on the horizontal axis and bacteria on the vertical axis. Correlations are indicated by color, with red denoting positive correlations and blue indicating negative correlations. The intensity of the color reflects the magnitude of the Pearson correlation coefficients, transitioning from light to dark as the coefficients increase in value from small to large. The "*" in the graph indicates the P-value size: no * for P-value ≥0.05, * for 0.01 ≤ P<0.05, ** for 0.001 ≤P<0.01,*** for P<0.001

Detecting distinct pathways in PNI and NPNI CRC patients and their link with differential microbiota

We transformed the RNA sequencing data from tumor samples of 8 patients, who also had gut microbiota 16S rRNA sequencing data, into corresponding scoring matrices using GO and KEGG analyses. Subsequently, a differential analysis of the GO and KEGG pathway scoring matrices for the PNI and NPNI groups revealed that the PNI group had 308 up-regulated GOs and 21 KEGGs, while the NPNI group had 212 up-regulated GOs and 1 KEGG (e.g., Fig. 5A-B). Respective detailed enrichment lists for GO terms and KEGG pathways can be found in Supplementary Tables 5 and 6. Furthermore, correlation analyses were conducted between the colony counts of 39 differential colonies from 8 patients and the corresponding GO and KEGG pathway score matrices. Significant correlations were identified between certain differential colonies and specific BPs, MFs, and KEGGs ( Fig. 5C, Supplementary Table 7), suggesting that the differential colonies might influence neuroinvasion in CRC through some potential biological functions.

Fig. 5.

Detecting PNI-associated differential pathways and their correlation with PNI-associated differential gut microbiot. A PNI-associated differential expression: GO volcano map. B PNI-associated differential expression: KEGG volcano map. The horizontal axis represents the log2 fold change, with genes farther from the center exhibiting greater differential expression. The vertical axis denotes the significance of expression differences, shown as -log10 of the p-value, with points closer to the top signifying more significant differences. Each point on the plot corresponds to a differentially expressed gene, color-coded as follows: red for up-regulated genes, blue for down-regulated genes, and gray for genes with no significant differential expression. C Correlation graph of PNI-associated differential BP and MF with PNI-associated differential gut microbiota. In the graph, the horizontal axis shows bacterial species, and the vertical axis represents Gene Ontology (GO) labels. The color scheme denotes the correlation direction: red for positive and blue for negative correlations. The color intensity corresponds to the Spearman's correlation coefficient, where darker shades indicate higher values. The "x" symbol in the graph denotes the P-value: an "x" indicates P-value ≥ 0.05, while its absence suggests P<0.05

Constructing machine learning models to predict PNI condition in CRC patients

Using 35 differential microbial taxa identified by LEfSe, we constructed Random Forest (RF) and Multilayer Perceptron (MLP) models to predict PNI in colorectal cancer (CRC) patients. In the RF model, the training set confusion matrix (Fig. 6A) revealed true positive (TP = 50) and true negative (TN = 22) counts significantly exceeding false positive (FP = 16) and false negative (FN = 4). This yielded a recall (sensitivity) of 92.6% (high ability to detect PNI-positive cases), precision of 75.8%, and F1-score of 83.5%, with an ROC AUC of 0.885 (Fig. 6C), indicating excellent training performance. However, the validation set (Fig. 6B) showed limited generalization: TP = 3, TN = 21 marginally outperformed FP = 0, FN = 16, resulting in a recall of 15.8% (high false-negative rate), 100% precision (no false positives), F1-score of 27.3%, and ROC AUC of 0.764.

Fig. 6.

Assessing the effectiveness of RF and MLP predictive models. A RF predictive model training set confusion matrix. B RF predictive model validation set confusion matrix. D MLP predictive model training set confusion matrix. E MLP predictive model validation set confusion matrix. The x-axis represents the model's predictions, while the y-axis represents the actual observations. The value'1' indicates a correct prediction, and'0' signifies an incorrect prediction. The numbers within the boxes represent the count of samples for each prediction category. C ROC curves for RF predictive model in training and validation sets.F ROC curves for MLP predictive model in training and validation sets. The horizontal axis represents the false positive rate as predicted by the model, while the vertical axis denotes the true positive rate. The area under the curve (AUC) quantifies the model's discriminatory ability; a higher AUC value indicates superior diagnostic performance of the model

For the MLP model, the training set (Fig. 6D) displayed TP = 48, TN = 26 surpassing FN = 6, FP = 12, with recall = 88.9%, precision = 80%, F1-score = 84.1%, and ROC AUC = 0.821 (Fig. 6F). The validation set (Fig. 6E) showed TP = 11, TN = 17 vs. FN = 8, FP = 4, yielding recall = 57.9%, precision = 73.3%, F1-score = 64.9%, and ROC AUC = 0.734. While MLP’s validation recall requires improvement, its F1-score was 37.6% higher than RF’s validation set, demonstrating more stable generalization.

Both models achieved AUC > 0.8 (training) and AUC > 0.7 (validation), confirming predictive utility. The RF model’s high training recall (92.6%) highlights its potential for clinical PNI prediction via gut microbiota analysis. Future efforts to expand sample size will optimize validation performance, enhancing clinical applicability.

Discussion

Our research concentrated on examining the disparities in the structure and abundance of the gut microbiota among CRC patients in the PNI and NPNI groups, with the aim of identifying microbial markers of predictive or discriminative importance through analysis of their differential gut microbiota. We ascertained the proportional abundance of each bacterial group in the stool samples from both patient groups and investigated their associated biofunctional pathways via 16S rRNA sequencing, respectively.

This study initially conducted a comparative analysis of the α-diversity and β-diversity of the gut microbiota between the PNI and NPNI groups of CRC patients. The findings indicated that there were no substantial variations in either α-diversity or β-diversity across the two groups. This implies that there is no marked disparity in the number of microbial community species and the composition of microorganisms between the two sample groups. What we would like to further explore is whether the presence of specific microbiota within the gut microbiota of CRC patients correlates with the PNI by CRC. Consequently, we proceeded to determine the abundance of each gut microbial community within the two groups, sought out gut microbiota that exhibited significant differences in abundance between these groups, and examined the potential correlation between these microbiota and PNI in colorectal cancer.

As a result of the study, it was discovered that there are 35 species exhibiting notable disparities in abundance within the gut microbiota of the PNI group compared to the NPNI group. Among them, 28 species were found to have increased abundance in the PNI group, while 7 species showed higher abundance in the NPNI group. (Supplementary Tables 3 and Fig. 2A-B). f__Ruminococcaceae. g__Ruminococcus, o__Clostridiales.f__Clostridiales. Incertae_Sedis__XIII, f__Peptostreptococcaceae. g__Romboutsia, f__Lachnospiraceae. g__Clostridium_XlVa, g__Holdemania. s__uncultured_ bacterium was the one with markedly greater abundance in the PNI group (Fig. 2A, B). It has been shown in other studies that the development of CRC may be accompanied by an increase in Ruminococcus and Lachnospiraceae, which are enriched in the normal tissues of the colon of CRC patients, and that these Ruminococcus play a role in degrading lysoglycerophospholipids, which promote tumor growth, in the tissues, and that the tissue Lachnospiraceae promote the immune surveillance function of CD8 + T cells and control CRC progression [34, 35]. We speculate that increased abundance of these bacteria may have an inhibitory effect on colorectal cancer progression. Secondly, a study showed that the abundance of Peptostreptococcaceae and Clostridiaceae was higher in CRC rats than in CTRL rats by analyzing the gut microbiota in animal studies [36]. There are other studies showing that the abundance of these microbiota is associated with colorectal cancer. However, our analysis of immune cells in the tumor microenvironment of the PNI group revealed a positive correlation between certain dominant microbiota (f__Peptostreptococcaceae.g__Romboutsia, f__Clostridiaceae) and T cells CD4 naive (Fig. 3C). It has been shown that T cells CD4 naive infiltrating breast tumors differentiate in situ into Treg, a heterogeneous subpopulation of immunosuppressive T cells that usually favors tumor progression. We can speculate that these increased abundances of gut microbiota increase the infiltration of treg in tumor tissues by recruiting T cells CD4 naive, which enhances the invasiveness of tumor tissues and leads to the development of PNI, but the exact mechanism remains to be further investigated [37, 38]. Based on the above, existing studies have used bacteria such as Lachnospiraceae, Clostridiales, Ruminococcus, and Clostridiaceae as microbial biomarkers between CRC patients and healthy controls [39]. It serves as a reference for our research.

The results of microbial gene co-identification of KEGG pathways detected in fecal samples from CRC patients in the PNI group versus the NPNI group (Supplementary Table 4) suggest that compared to the NPNI group, the ko00514:Other types of O-glycan biosynthesis, ko00601:Glycosphingolipid biosynthesis - lacto and neolacto series and ko04080:Neuroactive ligand-receptor interaction pathway were significantly enriched in the PNI group. o-Glycans are oligosaccharides abnormally linked to serine and threonine residues via n-acetylgalactosamine. o-Glycosylation promotes carcinogenesis signaling. Glycosylation promotes oncogenic signaling and tumor progression, and has potential as a biomarker for tumor development and staging [40–42]. Glycosphingolipid (GSL) is a class of sphingolipids, and clusters of GSL antigens with microstructural domains of transducing molecules form a structural and functional unit involved in adhesion and signal transduction of tumor cells, which in turn increases the metastatic potential and invasiveness of cancer cells [43–45]. From this we know that the pathways ko00514:Other types of O-glycan biosynthesis, ko00601:Glycosphingolipid biosynthesis - lacto and neolacto series increase o-glycan and GSL content respectively to increase tumor invasiveness, which leads to PNI, but the exact process remains unknown and requires further study. Neuroactive ligand-receptor interaction encompasses a collection of neuroreceptor genes, including dopamine receptors and proto-oncogenes. These genes play a role in processing environmental information and facilitating the interaction of signaling molecules. Research has demonstrated that inhibiting the neuroactive ligand-receptor interaction pathway can boost the immunotherapeutic response in colon cancer [46]. This pathway exhibited notable enrichment in the PNI group, leading to the conjecture that it may be substantially amplified within the tumor tissues of CRC patients, thereby intensifying PNI.

We analyzed the tumor microenvironment of 22 immune cells in 8 CRC patients with both PNI and RNA sequencing information. The findings indicated (Fig. 3A) that, in contrast to those in the NPNI group, patients in the PNI group exhibited a greater quantity of B cells memory and B cells naive, but a lower number of plasma cells. Plasma cells, also called effector B cells, can be differentiated from B cells naive or from B cells memory through secondary reactions [47]. Tumour-infiltrating B lymphocytes (including tumour-infiltrating B cells and plasma cells) have a key synergistic role in tumor control, and have been progressively investigated for immunotherapy as they promote anti-tumor immunity through their unique antigenic presentation to T cells [48, 49]. B-cell linker protein (BLNK) serves as a pivotal linker protein that bridges B-cell receptor-associated kinases to various signaling pathways, which play a role in modulating B-cell function and development. Research has indicated that BLNK protein expression correlates with poor prognosis in colorectal cancer, and that excessive expression of BLNK protein serves as an independent risk factor for CRC relapse and appears to be oncogenic in colorectal cancer [50]. There are data in this study suggesting that PNI is more common when BLNK protein is highly expressed, which is consistent with the trend of change we observed [51]. We therefore hypothesize that BLNK protein affects the proliferation and differentiation of B cells naive in the tumor microenvironment. When BLNK protein is overexpressed, B cells memory and B cells naive are more inclined to proliferate rather than differentiate into plasma cells, and in this case immune function is altered and PNI is more likely to occur in CRC patients. When BLNK protein is less expressed, B cells memory and B cells naive are more inclined to differentiate into plasma cells, and the probability of developing PNI is reduced.

Colorectal cancer development is associated with many immune checkpoints, immune activation genes, immune-suppressing genes, chemokines and chemokine receptors. Changes in gut microbiota may indirectly lead to their alteration and thus promote colorectal cancer. Therefore, we examined the association between the differential gut microbiota and immunity-associated genes. Within the predominant bacteria of both the PNI and NPNI groups, certain dominant bacteria (see 3.6) exhibited notable positive correlations with a variety of immune checkpoints, immune activation genes, immune-suppressing genes, chemokines, and chemokine receptors. However, according to the current study, it only suggests that the gut microbiota influence these processes, but it is not possible to determine exactly what role is played by the dominant bacteria involved, and there have been no relevant reports in this regard. The association of gut microorganisms with immune checkpoints, immune activation genes, immune-suppressing genes, chemokines, and chemokine receptors in the development of colorectal cancer, and whether or not these processes influence PNI, deserves to be further explored.

Finally, this study further data-mined the 16S rRNA sequencing data of the gut microbiota from CRC patients in the PNI and NPNI groups to construct a fecal microbial biomarker for differentiating between the PNI and NPNI groups. We integrated 35 characteristic bacteria associated with neuroinvasion in CRC, constructed both the RF and MLP models, and compared their discriminative efficacy. Our findings revealed that the RF predictive model achieved an AUC value of 0.885 for the training-set ROC curve and 0.764 for the validation-set ROC curve. In contrast, the MLP predictive model had an AUC value of 0.821 for the training-set ROC curve and 0.734 for the validation-set ROC curve, which indicated that both models had better diagnostic efficacy. The above results suggest that fecal biomarkers based on dominant microbiota could serve as a means to distinguish CRC patients in the PNI group from those in the NPNI group. Further construction of the correlation between gut microbiota and CRC-PNI through analysis using machine learning models and construction of relevant discriminative models can assist clinical diagnosis and treatment. Nonetheless, the influence process of many microbiota on CRC as well as PNI is still unknown, and further studies are needed to elucidate it.

Undeniably, the current research has certain limitations. Firstly, this study was cross-sectional, lacked systematic continuity, ignored individual differences, and was easily confounded by regional group effects. Secondly, the deficiency of healthy control subjects in this study is not favorable for ascertaining the connection between the gut microbiota disparities among CRC patients at varying stages of progression and the healthy population. Next, our study only focused on the fecal microbiota, i.e., the microbiota of the intestinal mucosa’s mucus layer. However, we recognize that the development of CRC might be intimately linked to the microbiota of the entire digestive system, including the digestive tract and digestive glands, and if we can analyze the microbial communities of these sites together, we would attain a more profound comprehension of how the human body’s commensal microbiota impact the progression of CRC. Finally, this study only superficially measured the differences in intestinal bacterial abundances between the PNI and NPNI groups. The results of the LEfSe analysis were not corrected for multiple hypothesis testing, and the relevant conclusions should be regarded as potential clues. Their statistical significance requires further validation through independent cohorts or relevant basic experiments to confirm the association between intestinal microbiota and CRC-PNI. In addition, although the RF and MLP models can reduce the risk of overfitting, the final RF and MLP predictive models constructed in this study still have some bias due to the lack of cross-validation with external data from other centers. Therefore, their predictive results can only be used for reference. Moreover, due to the lack of basic experimental investigation, this study cannot reveal the mechanisms underlying the association between these characteristic microbiota and PNI. We will endeavor to surmount these deficiencies in order to attain better outcomes in the subsequent research.

Conclusion

In the PNI and NPNI groups, there were 35 types of gut microbiota that showed significant differences in species abundance. Among them, four types of gut microbiota with significantly increased abundance (Lachnospiraceae, Clostridiales, Ruminococcus, and Clostridiaceae) have been used as biomarkers for distinguishing colorectal cancer patients from healthy controls. The increased abundance of Streptococcaceae and Clostridiaceae increases the infiltration of treg in tumor tissues by recruiting T cells CD4 naive, which may enhances the invasive properties of tumor tissues and leads to the development of PNI. However, the specific mechanisms of action still need to be further studied. The RF and MLP predictive models, constructed using gut microbiota as fecal microbial biomarkers, demonstrate considerable diagnostic capabilities and can serve as tools for predicting the CRC-PNI status. Their predictive results are of reference significance.

Authors' contributions

Chuanbin Chen, Qingmin Chen, Shenghai Liu, Guoxi Li, Jiawei Zhao, Jingting Huang, Tianyi Ye, Xinting Yang, Zigui Huang, Zhen Wang, Fuhai He, Mingjian Qin, Chenyan Long, Binzhe Tang, Yongqi Huang, Weizhong Tang, Jungang Liu, Xiaoliang Huang: conceptualized and planned the experiments; Chuanbin Chen, Qingmin Chen, Shenghai Liu, Guoxi Li, Jiawei Zhao, Jingting Huang, Tianyi Ye, Xinting Yang, Zigui Huang, Zhen Wang, Fuhai He, Mingjian Qin, Chenyan Long, Binzhe Tang, Yongqi Huang, Weizhong Tang, Jungang Liu, Xiaoliang Huang: examined and interpreted the data; Chuanbin Chen, Qingmin Chen, Shenghai Liu, Guoxi Li, Jiawei Zhao, Jingting Huang, Tianyi Ye, Xinting Yang, Zigui Huang, Zhen Wang, Fuhai He, Mingjian Qin, Chenyan Long, Binzhe Tang, Yongqi Huang, Weizhong Tang, Jungang Liu, Xiaoliang Huang: assisted with reagents, materials, and analysis tools; Chuanbin Chen, Qingmin Chen, Shenghai Liu, Guoxi Li, Jiawei Zhao, Jingting Huang, Tianyi Ye, Xinting Yang, Zigui Huang, Zhen Wang, Fuhai He, Mingjian Qin, Chenyan Long, Binzhe Tang, Yongqi Huang, Weizhong Tang, Jungang Liu, Xiaoliang Huang: contributed to manuscript writing. All authors reviewed the manuscript.

Funding

Middle-aged and Young Teachers’ Basic Ability Promotion Project of Guangxi (Basic Ability Promotion Project of Guangxi) (2021KY0087); Youth Research Project of Guangxi Medical University Affiliated Cancer Hospital (yuanqingji2023-10hao); Youth Science Foundation of Guangxi Medical University (GXMUYSF202402); Natural Science Foundation of Guangxi Province (Guangxi Natural Science Foundation) (2021GXNSFAA196008); China Postdoctoral Science Foundation (2023MD734155); Youth Science Foundation of Guangxi Medical University (GXMUYSF202357). Guangxi Key Laboratory of Basic and Translational Research for Colorectal Cancer.

Data availability

The datasets generated during the current study are available in the [GSA for Human] repository, [1.16S rRNA: https://ngdc.cncb.ac.cn/gsa-human/browse/HRA010085. 2.Transcriptome sequencing date: https://ngdc.cncb.ac.cn/gsa-human/browse/HRA011890]. Currently, this dataset is not fully publicly available, additional inquiries should be directed to the corresponding authors.

Declarations

Ethics approval and consent to participate

This study was approved by the Medical Ethics Committee of Guangxi Medical University Cancer Hospital, which was in accordance with the Declaration of Helsinki. The ethics approval number is KY2025580. All participants provided written informed consent. All methods were performed in accordance with the relevant guidelines and regulations in ethics approval and consent to participate.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.