Binary classification method for predicting immune therapy response of colorectal cancer patients based on gut microbiome metagenomic data
By integrating gut microbial metagenomic data, constructing a microbial co-occurrence network, and using random walk kernel smoothing technology, the problems of unstable feature selection and insufficient prediction accuracy in existing technologies were solved, achieving higher prediction accuracy and biological interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING MEDICAL UNIVERSITY
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies for predicting immunotherapy response in colorectal cancer patients suffer from unstable feature selection, neglect of microbial interactions, insufficient predictive accuracy, and failure to effectively integrate multi-level biological information, resulting in poor model generalization ability.
By integrating gut microbial metagenomic data, a microbial co-occurrence network was constructed. Random walk kernel smoothing technology and support vector machine model were used to screen key biomarkers, and feature selection and prediction were performed by combining multi-source data.
It improves prediction accuracy and feature selection stability, can explain the microbial ecological basis of immunotherapy response, and provides more reliable biomarkers for clinical diagnosis.
Smart Images
Figure CN122090953A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data analysis technology and relates to a binary classification method for predicting the immunotherapy response of colorectal cancer patients based on gut microbiome metagenomic data. Background Technology
[0002] Colorectal cancer (CRC) is one of the leading causes of cancer-related deaths worldwide. With the advancement of precision medicine, personalized treatment based on patient molecular characteristics has become a key approach to improving treatment outcomes. However, a major challenge in clinical practice is the significant heterogeneity in patient responses to treatments such as immune checkpoint inhibitors (ICIs). Traditionally, the assessment of treatment response has relied primarily on imaging studies and clinical observation, which are inherently lagging and cannot accurately predict which patients will benefit before or early in treatment.
[0003] In recent years, the relationship between the gut microbiome and tumorigenesis, development, and treatment response has received increasing attention. Studies have shown that the composition and functional status of the gut microbiota are closely related to the sensitivity and resistance of colorectal cancer patients to immunotherapy (such as anti-PD-1 / PD-L1 antibodies) combined with chemotherapy. In particular, patients with microsatellite stable (MSS) colorectal cancer have a low response rate to immunotherapy, necessitating reliable biomarkers for efficacy prediction and population screening. Metagenomic sequencing technology can comprehensively characterize the gut microbiota at both the species and functional levels, providing a new source of molecular biomarkers for predicting immunotherapy response and identifying primary drug-resistant patients.
[0004] Currently, classification and prediction methods based on microbiome data mainly suffer from the following drawbacks:
[0005] 1. Ignoring intermicrobial interactions: Existing methods typically treat each microbial species as an independent feature, selecting features through simple statistical tests (such as t-tests, Wilcoxon rank-sum tests) or machine learning methods (such as random forests, support vector machines). These methods ignore the complex symbiotic, competitive, and metabolic interactions within the microbial ecosystem, resulting in selected features lacking biological systematicity and exhibiting poor predictive stability.
[0006] 2. Insufficient stability of feature selection: Traditional feature selection methods based on statistical significance are sensitive to sample changes. The features selected on different training sets vary greatly, lacking reproducibility and making it difficult to form reliable clinical diagnostic biomarkers.
[0007] 3. High data dimensionality and high noise: Metagenomic data is characterized by "high dimensionality and small sample size," and contains a large amount of noise features unrelated to immunotherapy response. Existing technologies lack effective prior knowledge guidance mechanisms, are prone to overfitting, and result in poor model generalization ability.
[0008] 4. Insufficient integration of multi-source data: Existing methods usually only focus on a single level of species abundance or functional genes, failing to effectively integrate multi-level biological information (such as species interaction networks and functional gene profiles).
[0009] The Network Smoothed T-Statistics SVM method, proposed by Cun and Fröhlich in 2013, was the first to introduce the Random Walk Kernel into the classification problem of gene expression data. It integrates network prior knowledge by smoothing the t-statistics on the protein-protein interaction network (PPI). This method has shown high prediction accuracy and feature stability in cancer prognosis prediction. However, this method is designed for human gene expression data and protein-protein interaction networks, and its technical solution cannot be directly applied to microbiome data for the following reasons: (1) Microbial interaction networks are constructed based on co-occurrence relationships, which is different from the biological basis of PPI networks; (2) Metagenomic data contains information at both the species and functional levels, requiring special integration strategies; (3) There are differences in the clinical problem settings of immunotherapy response prediction and cancer prognosis prediction. Summary of the Invention
[0010] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a binary classification method for predicting immunotherapy responses in colorectal cancer patients based on gut microbiome metagenomic data. This method aims to solve technical problems in existing technologies, such as unstable feature selection, neglect of microbial interactions, and insufficient prediction accuracy.
[0011] To achieve the above objectives, the basic solution of this invention is: a binary classification method for predicting immunotherapy response in colorectal cancer patients based on gut microbiome metagenomic data, comprising the following steps:
[0012] Raw metagenomic sequencing data and immunotherapy response records were obtained from fecal samples of colorectal cancer patients; the sequencing data were subjected to quality control, host removal, and annotation to obtain a microbial abundance matrix and a functional gene abundance spectrum; the microbial abundance matrix, functional gene abundance spectrum, and patient immunotherapy response records were integrated to construct an initial feature dataset;
[0013] Based on the microbial species abundance matrix and functional gene abundance spectrum in the initial feature dataset, Spearman rank correlation coefficients between each feature are calculated, and the adjacency matrix of the microbial co-occurrence network is constructed. Based on the immunotherapy response labels, each feature in the initial feature dataset is analyzed. Test to obtain the original absolute value of statistic , forming a statistical vector Based on the adjacency matrix Calculate the degree matrix And normalized graph Laplace matrix , build Step random walk kernel matrix Through formula Perform network smoothing on the statistical vector to obtain the network smoothed statistical vector. The significance of the network smoothing statistics is evaluated based on the permutation test. After correction by multiple false discovery rate (FDR) tests, key features are selected to form the final feature dataset.
[0014] Based on the final feature dataset, a support vector machine (SVM) binary classification model is trained to obtain a prediction model for predicting the immunotherapy response of colorectal cancer patients; the prediction model is used to process the test samples, and the output SVM decision function value is used as an immunotherapy response risk score.
[0015] The working principle and beneficial effects of this basic approach are as follows: By integrating metagenomic (species, function), metabolomics, and clinical information, this technical approach constructs comprehensive and multi-level prognostic features, overcoming the limitations of information from a single data source. It employs a multi-stage feature selection strategy, combined with prior knowledge of microbial interaction networks, and utilizes random walk kernel smoothing technology to achieve stable and interpretable screening of key biomarkers from high-dimensional metagenomic data. This results in higher prediction accuracy, full utilization of overall microbial community information, and improved biological interpretability of the binary classification model.
[0016] Furthermore, the method for quality control and host removal processing of the raw metagenomic sequencing data to obtain filtered sequencing data is as follows:
[0017] The raw sequencing data were quality controlled using the kneaddata tool. Quality trimming was performed using Trimmomatic with the sliding window parameters set to SLIDINGWINDOW:4:20. Bases with a quality value lower than 3 at the beginning and end were removed (parameters LEADING:3, TRAILING:3), and fragments shorter than 50 bp were filtered (parameter MINLEN:50).
[0018] The quality-controlled data were aligned to the human reference genome version GRCh38 using Bowtie2 alignment software with the --very-sensitive and --dovetail alignment modes. Host-derived sequences were removed to obtain filtered sequencing data.
[0019] By using preset tools and parameters, we ensure that the data preprocessing steps are standardized, reproducible, and efficient, providing a high-quality, low-noise clean sequencing data foundation for downstream analysis.
[0020] Furthermore, the steps for annotating and quantifying the species abundance of the filtered sequencing data to generate a species-level microbial abundance matrix and a functional-level relative gene abundance spectrum are as follows:
[0021] The filtered sequencing data was input into the MetaPhlAn tool and combined with the accompanying database mpa_v30_CHOCOPhlAn_201901. Species annotation was performed using default parameters to quantify the relative abundance of each microbial species or strain in each sample, generating a microbial abundance matrix. In the microbial abundance matrix, rows represent samples, columns represent microbial species or strains, and cell values are the corresponding relative abundance. The HUMAnN3 tool was used to link with the UniRef90 database to obtain the functional gene relative abundance spectrum of each microbial species or strain.
[0022] By using appropriate tools and their supporting databases, sequence-to-species and functional information can be accurately and efficiently analyzed to obtain microbial composition information and functional potential data, laying the foundation for subsequent multi-dimensional analysis of "structure-function" associations.
[0023] Furthermore, the specific steps for constructing the initial feature dataset by integrating the microbial abundance matrix, functional gene relative abundance spectrum, and immunotherapy response records are as follows:
[0024] Based on the immunotherapy response record, a binary label OS12 (patient's immunotherapy response status after 12 months of medication) was defined.
[0025] The microbial abundance matrix and the relative abundance spectrum of functional genes were transformed into a data matrix. ,in The sample name The feature category is represented by an OS12 vector containing binary labels indicating the patient's response to immunotherapy. The samples were then assigned values based on their prognosis (good or bad), and samples with empty OS12 label values were removed to obtain the initial feature dataset. .
[0026] Structured integration of data from different sources (species, function, metabolite-microbe associations, clinical indicators) with a defined clinical endpoint (OS12) facilitates subsequent use.
[0027] Furthermore, the adjacency matrix The construction includes:
[0028] Calculate the Spearman rank correlation coefficients among different microbial species. ;
[0029] when And corrected by the Benjamini-Hochberg method When, set As edge weight, otherwise 0;
[0030] Set diagonal elements 0, remove the self-loop connection and build an undirected network.
[0031] By specifying the use of the Spearman correlation coefficient, setting a correlation threshold (|ρ| > 0.3), and statistical significance correction (…), < 0.05), the technical effect is the construction of a microbial co-occurrence network based on robust statistical inference. This network is not arbitrarily connected, but reflects statistically significant symbiotic or competitive relationships among microorganisms. Using the absolute value of correlation as edge weights quantifies the interaction strength, providing a high-quality prior graph structure input for subsequent network smoothing algorithms. Furthermore, the degree matrix... For a diagonal matrix, its diagonal elements Indicates species The degree; based on the degree matrix and adjacency matrix Calculate the graph Laplacian matrix Normalized graph Laplace matrix The calculation formula is: ,in It is an identity matrix.
[0032] The descriptive network (adjacency matrix) is transformed into an operator suitable for spectral analysis and pairwise regularization. It is a core tool in graph signal processing. Its eigenvalues and eigenvectors reflect the overall topological properties of the network and are the necessary mathematical foundation for the next step of constructing random walk kernels and achieving network smoothing.
[0033] further, Step random walk kernel matrix Calculated using the following formula:
[0034] ,
[0035] in is a constant and , The number of random walk steps and , usually take .
[0036] Construct a p-step random walk kernel matrix K to quantify the ability of any two nodes in the network to influence each other through multi-step paths.
[0037] Furthermore, the network smoothing statistics vector The calculation formula is: Then to Sort the features and select the top-ranked features for model training.
[0038] The original t-test statistic t, based on a single microorganism, was modified based on network structure, which greatly improved the biological rationality and stability of feature screening.
[0039] Furthermore, the significance of the ranking features is evaluated using a permutation test, including the following steps:
[0040] The immunotherapy response label of the randomized samples was permuted 1000 times, and the network smoothing statistic was recalculated after each permutation. ;
[0041] Calculate each feature based on the distribution of the permuted statistics. value;
[0042] Only the network with the highest smoothness score Species substitution tests are performed to reduce computational load;
[0043] The Benjamini-Hochberg method was used for multiple test corrections to screen [the samples]. The statistically significant high ranking of species is used as the key feature.
[0044] By constructing a zero distribution through random label permutation, a rigorous statistical significance test is provided for the network smoothing statistics, ensuring that the selected features are not random noise.
[0045] Furthermore, the support vector machine employs a Gaussian kernel, and hyperparameter optimization includes: in the candidate parameter set In this process, the Span rule is used to calculate the upper bound of the leave-one-out cross-validation error, and the penalty parameter that minimizes this upper bound is selected. As the optimal hyperparameter, the performance of the prediction model is evaluated by 10 repeated 10-fold cross-validation.
[0046] The trained model is used to process the test samples, and the output SVM decision function value is used as an immunotherapy response risk score. Attached Figure Description
[0047] Figure 1 This is a flowchart illustrating the binary classification method for predicting immunotherapy response in colorectal cancer patients based on gut microbiome metagenomic data, as per the present invention. Detailed Implementation
[0048] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0049] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0050] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.
[0051] This invention discloses a binary classification method for predicting immunotherapy response in colorectal cancer patients based on gut microbiome metagenomic data. Targeting the characteristics of gut microbiome metagenomic data, it effectively integrates microbial interaction networks and multi-source omics data, resulting in higher prediction accuracy, better robustness, and the ability to be automated. It addresses technical problems in existing technologies such as unstable feature selection, neglect of microbial interaction relationships, and insufficient prediction accuracy. By introducing a network smoothing t-statistic, the stability and biological interpretability of biomarker screening are improved.
[0052] like Figure 1 As shown, the binary classification method for predicting immunotherapy response in colorectal cancer patients based on gut microbiome metagenomic data includes the following steps:
[0053] Raw metagenomic sequencing data and immunotherapy response records were obtained from fecal samples of colorectal cancer patients; the sequencing data were subjected to quality control, host removal, and annotation to obtain a microbial abundance matrix and a functional gene abundance spectrum; the microbial abundance matrix, functional gene abundance spectrum, and patient immunotherapy response records were integrated to construct an initial feature dataset;
[0054] Based on the microbial species abundance matrix and functional gene abundance spectrum in the initial feature dataset, Spearman rank correlation coefficients between each feature are calculated, and the adjacency matrix of the microbial co-occurrence network is constructed. Based on the immunotherapy response labels, each feature in the initial feature dataset is analyzed. Test to obtain the original absolute value of statistic , forming a statistical vector Based on the adjacency matrix Calculate the degree matrix And normalized graph Laplace matrix , build Step random walk kernel matrix Through formula Perform network smoothing on the statistical vector to obtain the network smoothed statistical vector. The significance of the network smoothing statistics is evaluated based on the permutation test. After correction by multiple false discovery rate (FDR) tests, key features are selected to form the final feature dataset.
[0055] Based on the final feature dataset, a support vector machine binary classification model is trained to obtain a prediction model for predicting the immunotherapy response of colorectal cancer patients; the prediction model is used to process the test samples, and the output SVM decision function value is used as an immunotherapy response risk score.
[0056] The features screened in this invention are not isolated species, but rather form functionally related network modules. These modules often correspond to specific metabolic pathways (such as drug metabolism and DNA repair-related pathways) and have a clear biological connection with the immunotherapy response mechanism. Compared to the "black box" models of traditional machine learning, the feature set provided by this invention can explain the microbial ecological basis of the immunotherapy response state. This invention integrates microbial species abundance, functional gene abundance profiles, and microbial interaction networks, and improves the stability and biological interpretability of biomarker screening by introducing network smoothing t-statistics.
[0057] In a preferred embodiment of the present invention, the method for performing quality control and host removal processing on raw metagenomic sequencing data to obtain filtered sequencing data is as follows:
[0058] The raw sequencing data were quality controlled using the kneaddata tool. Quality trimming was performed using Trimmomatic with the sliding window parameters set to SLIDINGWINDOW:4:20. Bases with a quality value lower than 3 at the beginning and end were removed (parameters LEADING:3, TRAILING:3), and fragments shorter than 50 bp were filtered (parameter MINLEN:50).
[0059] The quality-controlled data were aligned to the human reference genome version GRCh38 using Bowtie2 alignment software with the --very-sensitive and --dovetail alignment modes. Host-derived sequences were removed to obtain filtered sequencing data.
[0060] In a preferred embodiment of the present invention, the steps of performing species annotation and abundance quantification on the filtered sequencing data to generate a microbial abundance matrix at the species level and a functional gene relative abundance spectrum at the functional level are as follows:
[0061] The filtered sequencing data was input into the MetaPhlAn tool and combined with the accompanying database mpa_v30_CHOCOPhlAn_201901. Species annotation was performed using default parameters to quantify the relative abundance of each microbial species or strain in each sample, generating a microbial abundance matrix. In the microbial abundance matrix, rows represent samples, columns represent microbial species or strains, and cell values are the corresponding relative abundance. The HUMAnN3 tool was used to link with the UniRef90 database to obtain the functional gene relative abundance spectrum of each microbial species or strain.
[0062] In a preferred embodiment of the present invention, the specific steps for constructing an initial feature dataset by integrating a microbial abundance matrix, a functional gene relative abundance spectrum, and immunotherapy response records are as follows:
[0063] Based on the immunotherapy response record, a binary label OS12 (patient's immunotherapy response status after 12 months of medication) was defined.
[0064] The microbial abundance matrix and the relative abundance spectrum of functional genes were transformed into a data matrix. ,in The sample name The feature category is represented by an OS12 vector containing binary labels indicating the patient's response to immunotherapy. The samples were then assigned values based on their prognosis (good or bad), and samples with empty OS12 label values were removed to obtain the initial feature dataset. .
[0065] This invention not only considers microbial species abundance but also integrates functional gene abundance profiles, comprehensively characterizing the microbiome features related to immunotherapy responses from two dimensions: "what microorganisms are present" and "what functions do they have." This multi-source data integration strategy can capture more comprehensive biological information compared to a single data source.
[0066] In a preferred embodiment of the present invention, the adjacency matrix The construction includes:
[0067] Calculate the Spearman rank correlation coefficients among different microbial species. ;
[0068] when And corrected by the Benjamini-Hochberg method When, set As edge weight, otherwise 0;
[0069] Set diagonal elements 0, remove the self-loop connection and build an undirected network.
[0070] In a preferred embodiment of the present invention, the degree matrix For a diagonal matrix, its diagonal elements Indicates species The degree; based on the degree matrix and adjacency matrix Calculate the graph Laplacian matrix Normalized graph Laplace matrix The calculation formula is: ,in It is an identity matrix.
[0071] In a preferred embodiment of the present invention, Step random walk kernel matrix Calculated using the following formula:
[0072] ,
[0073] in is a constant and , The number of random walk steps and , usually take .
[0074] In a preferred embodiment of the present invention, the network smoothing statistics vector The calculation formula is: Then to Sort the features and select the top-ranked features for model training.
[0075] In a preferred embodiment of the present invention, evaluating the significance of ordination features using a permutation test includes the following steps:
[0076] The immunotherapy response label of the randomized samples was permuted 1000 times, and the network smoothing statistic was recalculated after each permutation. ;
[0077] Calculate each feature based on the distribution of the permuted statistics. value;
[0078] Only the network with the highest smoothness score Species substitution tests are performed to reduce computational load;
[0079] The Benjamini-Hochberg method was used for multiple test corrections to screen [the samples]. The statistically significant high ranking of species is used as the key feature.
[0080] Preferably, by introducing a microbial co-occurrence network and a random walk kernel smoothing mechanism, this invention combines the statistical significance of a single microorganism with its topological position in the network. Compared to methods based solely on t-tests, the coefficient of variation of the smoothed statistics across different training subsets is reduced, and the consistency of feature selection is significantly improved. This is because the network smoothing effect can identify microorganisms that are statistically marginally significant but occupy key positions in the network (such as hub species connecting multiple functional modules), and these species are more likely to be biomarkers related to immunotherapy responses.
[0081] Preferably, the support vector machine uses a Gaussian kernel, and hyperparameter optimization includes: in the candidate parameter set In this process, the Span rule is used to calculate the upper bound of the leave-one-out cross-validation error, and the penalty parameter that minimizes this upper bound is selected. As the optimal hyperparameter, the performance of the prediction model is evaluated by 10 repeated 10-fold cross-validation.
[0082] Based on key feature sets Construct the final feature dataset The binary classification model was trained using the Support Vector Machine (SVM) algorithm.
[0083] Kernel function: Gaussian radial basis kernel (RBF Kernel), parameters Set as ( (characteristic number);
[0084] Hyperparameter optimization: Candidate penalty parameters The upper bound of the leave-one-out cross-validation error is calculated using the Span rule, and the method that minimizes this upper bound is selected. As the optimal parameter, this method is more efficient than grid search and is suitable for biomedical data with limited sample size;
[0085] Data standardization: Z-score standardization is performed on each feature (based on the mean and standard deviation of the training set), and the parameters are applied to the test set.
[0086] The probability of a patient's response to immunotherapy is quantified by the decision function value of a support vector machine, which helps clinicians develop individualized treatment plans.
[0087] This invention employs a permutation test to evaluate the significance of species with the highest predetermined proportion (e.g., 10%-20%) in the network smoothness score, significantly reducing computational complexity. Furthermore, the final SVM model relies on only a few key features, resulting in low detection costs.
[0088] For example, this embodiment uses a subgroup of colorectal cancer (CRC) from the PRJNA1023797 dataset to validate the method of the present invention. The original data comes from the NCBI Sequence Read Archive (SRA) database, BioProject number PRJNA1023797 (Iebba et al., Custom scoring based on ecological topology of gutmicrobiota associated with cancer immunotherapy outcome, Cell, 2024).
[0089] Download raw sequencing data from the NCBI SRA database (accession numbers can be found on the PRJNA1023797 project page). Quality control was performed using kneaddata; trimmomatic was used to remove adapter sequences and low-quality bases with parameters set to LEADING:3 and TRAILING:3; Bowtie2 was used to align the quality-controlled sequences to the human reference genome GRCh38 (--very-sensitive, --dovetail mode), removing aligned host DNA sequences and retaining unaligned sequences for subsequent microbiome analysis.
[0090] Species-level annotation: Using the MetaPhlAn tool and the database mpa_vOct22_CHOCOPhlAnSGB_202403, species annotation was performed on the filtered sequences to obtain the relative abundance matrix of microbial species in each sample. This tool identifies species based on specific marker genes, down to the species or strain level.
[0091] Functional level annotation: Pathway abundance profiles were obtained using HUMANN in conjunction with the UniRef90 and MetaCyc databases. .
[0092] Tag Definition: Treatment Response Tag {0: No response, 1: Response};
[0093] Dataset Construction: Integrating Species Abundance Matrix Functional abundance matrix and label vector Remove samples with missing labels. Perform a log10 transformation on the abundance data to reduce data skewness and construct the initial feature dataset. .
[0094] Based on the species abundance matrix in the initial feature dataset Functional abundance matrix Calculate the Spearman rank correlation coefficients among all pairs of microbial species. Compared to Pearson correlation, Spearman correlation is more robust to non-normally distributed microbial abundance data.
[0095] Network construction standards:
[0096] when And corrected by the Benjamini-Hochberg method At that time, it was assumed that there was a significant co-occurrence relationship between the two species, and adjacency matrix elements were set. ;otherwise 0; Set diagonal 0 (remove self-loops) to construct an undirected weighted network.
[0097] The adjacency matrix obtained in this step This reflects the ecological interactions among microorganisms, rather than simple physical interactions.
[0098] Based on adjacency matrix Calculate the degree matrix (diagonal elements) , indicating species (connectivity in the network), and then calculate the graph Laplacian matrix: And the normalized graph Laplacian matrix: .
[0099] Constructing the p-step random walk kernel: ,in =1 (attenuation factor), =2 (number of random walk steps, considering second-order neighbor relationships).
[0100] The kernel matrix Capture any two nodes (microorganisms) in the network through a length not exceeding The topological similarity of the paths is considered, taking into account the contributions of multiple alternative paths rather than simply the shortest path distance.
[0101] Each feature (species or functional gene) in the initial feature dataset is based on an immunotherapy response label. Perform an independent two-sample t-test to obtain the absolute value of the raw t-statistic. , forming a vector .
[0102] Calculate the network smoothed t-statistic using network smoothing: The biological significance of this formula lies in the fact that even for a certain species The differential expression itself is not significant, but if its neighboring species Significant differences and with If the connections are tight, then the value of that species is... It can be boosted by contributions from its neighbors. This "smoothing" effect can identify microbes in key positions in the network (such as hub species that connect multiple differentially expressed modules), which may be key functional nodes regulating the response to immunotherapy.
[0103] Since the statistics after network smoothing no longer follow a standard t-distribution, this invention uses a permutation test to assess significance:
[0104] Pre-screening: To reduce computational cost, only the top 10% of features based on the network smoothing statistic t are selected;
[0105] Permutation test: Random permutation response label A total of 1000 iterations were performed, each time recalculating the network smoothing statistics to construct the null distribution.
[0106] Value calculation: Based on the distribution of the permuted statistics, calculate the empirical p-value for each feature: p = 1000, the number of times the permuted statistic is greater than or equal to the actual statistic;
[0107] Multiple test correction: The Benjamini-Hochberg method was used to control the false discovery rate (FDR) at the 5% level and to screen for significant features.
[0108] Finally, the key feature set is obtained. It includes microbial biomarkers enhanced with network topology information.
[0109] Based on key feature sets Construct the final feature dataset The binary classification model was trained using the Support Vector Machine (SVM) algorithm.
[0110] Kernel function: Gaussian radial basis kernel (RBF Kernel), parameters Set as ( (characteristic number);
[0111] Hyperparameter optimization: Candidate penalty parameters The upper bound of the leave-one-out cross-validation error is calculated using the Span rule, and the method that minimizes this upper bound is selected. As the optimal parameter, this method is more efficient than grid search and is suitable for biomedical data with limited sample size;
[0112] Data standardization: Z-score standardization is performed on each feature (based on the mean and standard deviation of the training set), and the parameters are applied to the test set.
[0113] The model performance was evaluated using 10 repeated 10-fold cross-validation: each time the data was randomly divided into 10 folds; 9 folds were used as the training set (including network construction, feature selection and SVM training) and 1 fold was used as the test set in turn; this was repeated 10 times to obtain a total of 100 performance evaluation results.
[0114] Performance evaluation metrics
[0115] Primary metric: AUC-ROC (Area Under the Receiver Operating Characteristic Curve); Secondary metrics: Accuracy, Sensitivity / Recall, Specificity, F1 score;
[0116] Stability Metric: Feature Selection Stability Index (SI) is defined as the weighted average of the frequencies in which each feature is selected over 100 validations. ,in The feature set that has been selected at least once. Features The number of times it was selected.
[0117] Validation results on the PRJNA1023797 dataset show:
[0118] Prediction performance: Using 10 repeated 10-fold cross-validation, the average AUC of the method in this invention is [value missing]. ( Although the absolute AUC value is at a moderate level, this performance is within an acceptable range for clinical decision support, given the complexity of predicting immunotherapy responses and the high heterogeneity of the gut microbiome.
[0119] Feature selection stability: The stability index (SI) reaches 90.5%, significantly outperforming traditional machine learning methods (such as SVM-RFE, whose SI is typically lower). In 100 validation runs: 20 features in the core feature set (including key metabolic pathways such as folic acid conversion and arginine synthesis) were consistently selected (selection frequency 100%); feature selection consistency: average selection per run... One characteristic (coefficient of variation) This demonstrates the high stability of the number of features.
[0120] Biological function annotations were performed on the selected high-frequency features (selection frequency ≥ 90%), and these features formed a function-species association network:
[0121] One-carbon metabolism and immune regulation
[0122] Key feature: 1CMET2-PWY: folate transformations III (E. coli) |
[0123] g__Butyricimonas.s__Butyricimonas_synergistica (select frequency 100%).
[0124] Biological significance: This study detected significant activity signals of the genus *Butyricimonas* in this pathway.
[0125] Folic acid (vitamin B9) is involved in one-carbon metabolism, providing methyl donors for DNA synthesis and methylation. This finding is consistent with the mechanism by which microbial one-carbon metabolism affects the function of host immune cells reported by Smith et al. (Cell Metabolism, 2014), suggesting that this microbial community may potentially influence immunotherapy responses by regulating local folic acid availability.
[0126] Arginine metabolism and T cell function
[0127] Key feature: ARGSYN-PWY: L-arginine biosynthesis I|
[0128] g__Blautia.s__Blautia_obeum, ARGININE-SYN4-PWY (select frequency 100%).
[0129] Biological significance: This study found that *Blautia obeum* highly co-occurs with the arginine synthesis pathway. Arginine is a key substrate for T cell activation and mTOR signaling. This finding is consistent with the mechanism by which arginine regulates T cell survival and function reported by Geiger et al. (Nature, 2016), suggesting that *Blautia* may potentially participate in anti-tumor immune regulation by affecting local arginine metabolism in the gut.
[0130] Amino acid metabolism and gut microbiota homeostasis
[0131] Key features: GLUTORN-PWY: L-ornithine biosynthesis I|g__Blautia.s__Blautia_obeum, Anaerostipes hadrus (selection frequency 100%).
[0132] Biological significance: Both *Blautia* and *Anaerostipes* are core butyrate producers in the gut. Although *GLUTORN-PWY* belongs to the ornithine synthesis pathway (amino acid metabolism), these two genera are known to maintain intestinal barrier function through other metabolic pathways, such as butyrate synthesis. Literature reports that butyrate can regulate the Treg / Th17 balance by inhibiting histone deacetylase (HDAC) (Smith et al., 2013, *Nature Communications*), suggesting that these stably enriched butyrate-producing bacteria may synergistically influence the immunotherapy microenvironment through multiple metabolic pathways.
[0133] Network Feature Summary: These high-frequency features are interconnected through pivotal species such as Blautia, Butyricimonas, and Anaerostipes, forming functionally synergistic microbial ecological units. The network smoothing step successfully identified these "functional hubs" located at key positions in the network, validating the effectiveness of this method in capturing biologically relevant features.
[0134] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.
Claims
1. A binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data, characterized in that, Comprising the following steps: Obtaining raw metagenomic sequencing data of fecal samples of colorectal cancer patients and immunotherapy response records; performing quality control, host removal and annotation on the sequencing data to obtain a microbial abundance matrix and a functional gene abundance profile; integrating the microbial abundance matrix, the functional gene abundance profile and the patient immunotherapy response records to construct an initial feature dataset; Based on the microbial species abundance matrix and functional gene abundance spectrum in the initial feature dataset, Spearman rank correlation coefficients between each feature are calculated, and the adjacency matrix of the microbial co-occurrence network is constructed. Based on the immunotherapy response labels, each feature in the initial feature dataset is analyzed. Test to obtain the original absolute value of statistic , forming a statistical vector Based on the adjacency matrix Calculate the degree matrix And normalized graph Laplace matrix , build Step random walk kernel matrix Through formula Perform network smoothing on the statistical vector to obtain the network smoothed statistical vector. The significance of the network smoothing statistics is evaluated based on the permutation test. After correction by multiple false discovery rate (FDR) tests, key features are selected to form the final feature dataset. Based on the final feature dataset, training a support vector machine binary classification model to obtain a prediction model for predicting the immunotherapy response of colorectal cancer patients; processing the test sample using the prediction model and outputting the SVM decision function value as the immunotherapy response risk score.
2. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome data according to claim 1, wherein, The method for quality control and host removal processing of raw metagenomic sequencing data is as follows: Use the kneaddata tool to perform quality control on the raw sequencing data, perform quality trimming through Trimmomatic, set the sliding window parameter as SLIDINGWINDOW:4:20, remove bases with quality values less than 3 at the beginning and end (parameters LEADING:3, TRAILING:3), and filter out fragments with lengths less than 50 bp (parameter MINLEN:50); The quality-controlled data is aligned to the human reference genome version GRCh38 using Bowtie2 alignment software in --very-sensitive and --dovetail alignment mode to remove host-derived sequences and obtain filtered sequencing data.
3. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome data according to claim 2, wherein, The steps for species annotation and abundance quantification of the filtered sequencing data to generate a microbial abundance matrix at the species level and a functional gene relative abundance profile at the functional level are as follows: Input the filtered sequencing data into the MetaPhlAn tool, combine the supporting database mpa_v30_CHOCOPhlAn_201901, and use the default parameters for species annotation to quantify the relative abundance of each microbial species or strain in each sample, generate a microbial abundance matrix, and the microbial abundance matrix represents samples in rows, microbial species or strains in columns, and the cell value is the corresponding relative abundance; through the HUMAnN3 tool associated with the UniRef90 database, the functional gene relative abundance profile of each microbial species or strain is obtained.
4. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome data according to claim 1, wherein, The specific steps for integrating the microbial abundance matrix, the functional gene relative abundance profile, and the immunotherapy response records to construct an initial feature dataset are as follows: Based on the immunotherapy response records, define a binary classification label OS12 (immunotherapy response of patients 12 months after drug administration); Converting the microbial abundance matrix, the relative abundance profile of functional genes into a data matrix wherein is the sample name, is the feature category, and the binary label OS12 vector containing the patient's immunotherapy response and is assigned according to the sample's prognosis good and bad, removing the samples with empty OS12 label value, obtaining the initial feature dataset .
5. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, An adjacency matrix The construction of the adjacency matrix comprises: Spearman rank correlation coefficients between each pair of microbial species were calculated ; When and after Benjamini-Hochberg method correction , set as edge weight, otherwise 0; Setting diagonal elements 0, remove self-loop connections, construct undirected network.
6. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, degree matrix is a diagonal matrix whose diagonal elements represent the degrees of the species ; the graph Laplacian matrix is computed based on the degree matrix and the adjacency matrix , and the normalized graph Laplacian matrix is computed as follows: where is the identity matrix.
7. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, Step random walk kernel matrix is calculated by the following equation: ， wherein is a constant and , is a random walk step number and , typically taken .
8. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, Network smooth statistics vector The calculation formula is Then sort The top-ranked features are selected for model training.
9. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, Use the permutation test to evaluate the significance of the ranked features, including the following steps: immunotherapy response signature 1000 times, re-computing the network smooth statistics after each permutation ; Based on the distribution of the permuted statistics, the value of each feature is calculated value; Only the network with the highest smoothness score Species substitution tests are performed to reduce computational load; Benjamini-Hochberg method for multiple testing correction to screen statistically significant high-ranking species as the key features.
10. The binary classification method for predicting the immunotherapy response of a colorectal cancer patient based on gut microbiome metagenomic data according to claim 1, wherein, Support Vector Machines (SVM) employ a Gaussian kernel, hyperparameter optimization includes: computing the upper bound of leave-one-out cross-validation error using Span rules over a candidate parameter set , selecting the penalty parameter that minimizes this upper bound as the optimal hyperparameter; the performance of the predictive model is evaluated by 10 repeated 10-fold cross-validation.