Microbiome data classification method and system based on noise disturbance integrated neural network

Abstract

The invention relates to a microbiome data classification method and system based on a noise disturbance integrated neural network. The method comprises the following steps: acquiring intestinal microbiome data; inputting the intestinal microbiome data into a multi-classification model based on a noise disturbance integrated neural network to obtain a microbiome data classification result; wherein the multi-classification model based on the noise disturbance integrated neural network comprises a gradient boosting classifier and a gating circulation unit obtained by introducing noise disturbance mechanism training, and inputting the intestinal microbiome data into the gradient boosting classifier to generate a first classification result, and inputting the first classification result as priori knowledge into a gating circulation unit, and outputting a final microbiome data classification result. Compared with the prior art, the method has the advantages of high classification prediction efficiency and high accuracy.

Description

Technical Field

[0001] The present invention relates to the field of microbiome data processing, and in particular to a microbiome data classification method and system based on noise perturbation integrated neural network. Background Art

[0002] With the development of the field of microbiome, research has gradually revealed the decisive role of microorganisms in the disease process. Nevertheless, classification prediction using microbiome data still faces challenges such as data complexity, high heterogeneity, and weak model generalization ability. In addition, existing methods are often overly dependent on specific data sets and are easily interfered by adversarial sample attacks.

[0003] Given the availability of large amounts of human metagenomic data, machine learning and deep learning techniques are being used to analyze and identify relationships between different microbes and links between microbes and host diseases.

[0004] Existing microbiome data classification methods are sensitive to noise and outliers in the data, which may cause the model to produce errors or unstable results during the prediction process. When faced with new and unknown data, the generalization ability is weak, which means that the model may not adapt well on other data sets and cannot maintain good predictive performance. This limits the versatility and reliability of the algorithm in practical applications. In addition, existing feature extraction methods may not be able to fully explore the potential information in the data, or fail to capture the complex relationships between the data, thus affecting the classification prediction performance. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a microbiome data classification method and system based on noise perturbation integrated neural network with high classification prediction efficiency and high accuracy.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a method for classifying microbiome data based on a noise perturbation integrated neural network is provided, the method comprising:

[0008] Obtain gut microbiome data;

[0009] The intestinal microbiome data are input into a multi-classification model based on a noise perturbation ensemble neural network to obtain a classification result of the microbiome data; wherein the multi-classification model based on a noise perturbation ensemble neural network includes a gradient boosting classifier and a gated recurrent unit trained by introducing a noise perturbation mechanism, the intestinal microbiome data are input into the gradient boosting classifier to generate a first classification result, and the first classification result is input into the gated recurrent unit as prior knowledge to output the final microbiome data classification result.

[0010] Preferably, the method further comprises preprocessing the acquired intestinal microbiome data using a 16S processing tool, specifically:

[0011] The data were filtered to retain only sequences with a length of 101 to 150 bases;

[0012] OTUs were clustered at a 100% similarity threshold;

[0013] OTUs that appeared in less than 100 samples or had less than 10 reads per sample were removed, and OTUs with a relative abundance lower than 1% in the entire dataset were removed;

[0014] The relative abundance of each OTU was calculated by dividing the number of reads of the OTU by the total number of reads per sample;

[0015] The taxonomic information of OTUs was summarized to the genus level, that is, the relative abundances of all OTUs belonging to the same genus were added together.

[0016] Preferably, the gated recurrent unit processes sequence data through its gating mechanism and utilizes prior knowledge to enhance classification capability, and its data processing process includes:

[0017] Take the OTU abundance vector at time step t as the input x of the current time step t , where each element in the OTU abundance vector represents the relative abundance of a specific OTU in the sample, and the OTU abundance information of the previous sample is used as the hidden state h of the previous time step t-1 , calculate the update gate r t and reset gate z t , the expression is:

[0018] r t =σ(W r ·[x t ,h t-1 ]+b r ), (1)

[0019] z t =σ(W z ·[x t ,h t-1 ]+bz ), (2)

[0020] Where: W r Represents the weight matrix corresponding to the OTU abundance information that is most important for predicting the status of the current sample, W z represents the weight matrix corresponding to the new OTU information integrated into the model; b r and b z is the corresponding bias term; σ is the Sigmoid function;

[0021] According to the input x of the current time step t and the hidden state h at the previous time step t-1 , calculate the candidate hidden state

[0022]

[0023] Where: n t is a noise term, which is Gaussian noise or uniform noise, and is used to simulate the variability and noise characteristics of actual intestinal microbiome data; W c is a weight matrix used to determine how to convert the reset OTU abundance information into a candidate hidden state; b c is the bias term; tanh is the hyperbolic tangent function; ⊙ represents element-wise multiplication.

[0024] According to the update gate r t , reset gate z t and candidate hidden states Calculate the new hidden state h of the prediction of the current sample OTU abundance t :

[0025]

[0026] Preferably, the gradient boosting classifier captures the nonlinear relationship of the intestinal microbiome data by constructing a decision tree to generate a first classification result.

[0027] Preferably, in the multi-classification model based on noise perturbation ensemble neural network, a self-attention mechanism is introduced to calculate the relationship between each position in the sequence and assign a weight to each position to focus on the most important information for the current task.

[0028] Preferably, the training process of the gradient boosting classifier includes:

[0029] For the training set (x i ,y i ), input feature x i Represents the relative abundance of each OTU in the sample, corresponding to the target variable y iIndicates the relative abundance of an OTU in a specific sample;

[0030] Construct a prediction model F(x) to represent the initial estimate of the average abundance of the microbial community in a sample, and then make the prediction value F(x i ) as close as possible to the true label y i ;

[0031] Initialize the model:

[0032]

[0033] Iteratively train multiple weak learners, and fit a new model each iteration to correct the residual of the previous model to represent the difference between the actual observed OTU abundance and the OTU abundance predicted by the model. The residual of the current model is:

[0034] r it =y i -F t-1 (x i ), (6)

[0035] In the residual r it Fit a new weak learner on it to predict the structure of the microbial community in the sample:

[0036]

[0037] Among them, L is the loss function, which is used to measure the difference between the predicted value and the true value;

[0038] Combine the new model with the previous model through the learning rate and update the model prediction results:

[0039] F t (x) = F t-1 (x)+γh t (x), (8)

[0040] The training is repeated until a preset number of iterations is reached or the residual meets the set conditions.

[0041] Preferably, the weak learner is a decision tree.

[0042] According to a second aspect of the present invention, a microbiome data classification system based on a noise perturbation integrated neural network is provided, characterized in that it comprises:

[0043] Input module to obtain gut microbiome data;

[0044] A classification module, wherein the multi-classification model based on the noise perturbation integrated neural network includes a gradient boosting classifier and a gated recurrent unit trained by introducing a noise perturbation mechanism, the intestinal microbiome data is input into the gradient boosting classifier to generate a first classification result, and the first classification result is input into the gated recurrent unit as prior knowledge to output the final microbiome data classification result.

[0045] According to a third aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the program, any one of the methods described above is implemented.

[0046] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, any one of the methods described above is implemented.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention designs a multi-classification model based on a noise perturbation integrated neural network. The GB module is used to capture the nonlinear relationship of intestinal microbiome data by constructing a decision tree. The classification results output by the GB module are input into the GRU module as prior knowledge. The GRU processes sequence data through its gating mechanism and uses the prior knowledge to enhance the classification ability. Noise perturbation is introduced during the GRU training process to improve the adaptability and robustness of the multi-classification model to uncertain microbiome input data. Such a structure enables the GB and GRU to work together through a joint training strategy, and the classification results of the GB provide additional supervision signals for the GRU, while the noise perturbation promotes the model to learn more robust feature representations, thereby improving the generalization ability and classification accuracy of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the method of the present invention. ;

[0050] Figure 2 It is the average AUC value of different methods on different data sets in the embodiment. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0052] Example

[0053] Accurate classification of microbiome data is of great significance for accurate prediction of subsequent disease phenotypes. Currently, microbiome data classification still faces challenges such as data complexity, high heterogeneity, and weak model generalization ability. In addition, existing methods are often over-reliant on specific data sets and are easily interfered by adversarial sample attacks.

[0054] To address these problems, this embodiment proposes a microbiome data classification method based on a noise perturbation ensemble neural network (NPENN). The method can achieve accurate classification of microbiome data by constructing a multi-classification model based on the noise perturbation ensemble neural network. The multi-classification model based on the noise perturbation ensemble neural network specifically includes: a gradient boosting classifier (GB) and a gated recurrent unit (GRU) working together in combination with a noise perturbation mechanism, wherein the GB captures the nonlinear relationship of the intestinal microbiome data by constructing a decision tree, and the classification result output by the GB is input into the GRU as prior knowledge. The GRU processes the sequence data through its gating mechanism and uses the prior knowledge to enhance the classification ability. At the same time, noise perturbation is introduced in the GRU training process to improve the adaptability and robustness of the model to the uncertainty of the input data. Such a structure enables the GB and GRU to work together through a joint training strategy. The classification result of the GB provides an additional supervision signal for the GRU, and the noise perturbation promotes the model to learn a more robust feature representation, thereby improving the generalization ability and classification accuracy of the overall model.

[0055] This embodiment achieves effective integration of neural networks by introducing a noise mechanism and combining prior knowledge of a gradient boosting classifier (GB). NPENN performs well in both prediction accuracy and generalization ability, and has obvious advantages over traditional methods in dealing with data complexity and heterogeneity. It not only enhances the diversity of data, improves the robustness and reliability of the model, but also guides the neural network to learn data features and laws more efficiently through the introduction of GB prior knowledge.

[0056] 1. Microbiome data preprocessing

[0057] In this embodiment, the obtained intestinal microbiome data is preprocessed using a 16S processing tool, specifically:

[0058] The data were filtered to retain only sequences with a length of 101 to 150 bases;

[0059] OTUs were clustered at a 100% similarity threshold;

[0060] OTUs that appeared in less than 100 samples or had less than 10 reads per sample were removed, and OTUs with a relative abundance lower than 1% in the entire dataset were removed;

[0061] The relative abundance of each OTU was calculated by dividing the number of reads of the OTU by the total number of reads per sample;

[0062] The taxonomic information of OTUs was summarized to the genus level, that is, the relative abundances of all OTUs belonging to the same genus were added together.

[0063] 2. Noise perturbation integrated neural network model NPENN

[0064] The noise perturbation integrated neural network model NPENN in this embodiment adopts the GRU structure as a whole and is integrated with the GB method to obtain better prediction performance. The GRU structure can better learn the local features and patterns in the data, while the GB method can better integrate the relationship between different features and patterns, so that the model can obtain sufficient information and features at different levels, thereby making more accurate predictions. The results of GB are provided to the GRU model as prior information to enhance the learning ability of the model, and the performance and effect of the model are improved by introducing domain knowledge and prior assumptions.

[0065] During the training process, Gaussian noise and uniform noise are introduced to simulate the data changes and noise conditions in reality. Adding Gaussian noise to the data means that the disturbance added to each data point is randomly drawn from a fixed normal distribution, which can simulate the random errors in the data measurement process. Introducing uniform noise can help the model adapt to data changes within a certain range and improve its ability to handle unseen data. In this way, the model is forced to learn not only specific features in the training data, but also more generalized patterns, thereby improving its ability to predict new data. In addition, the NPENN model also introduces a self-attention mechanism in the GRU to increase the model's attention to the input data and better capture the complex relationships between the data. It calculates the relationship between each position in the sequence and assigns a weight to each position, allowing the model to focus on the most important information for the current task.

[0066] In summary, the NPENN model combines multiple technical means such as GRU, GB, noise perturbation and self-attention mechanism to build a powerful framework suitable for handling complex disease phenotype prediction tasks. Figure 1 Flowchart showing the NPENN model.

[0067] 2.1 Noise Perturbation Mechanism

[0068] Covers 14 datasets covering 9 different types of diseases in the medical field. This means that each sample in the dataset contains a large amount of microbial composition information, and these microorganisms may have different effects on different types of diseases. In addition, each disease may have different pathogenesis, clinical manifestations, and treatment responses.

[0069] Therefore, different disease datasets have different characteristics and data distributions, and there may be differences between datasets, such as sample size, data quality, sample source, and experimental conditions.

[0070] Considering the complexity of the above data in terms of biomedicine and mathematical statistics, this embodiment introduces two noise disturbance mechanisms: Gaussian noise and uniform noise.

[0071] Gaussian noise is a random, continuous noise whose distribution follows a Gaussian distribution. The introduction of this noise helps make the model more robust, avoid overfitting, and improve the generalization ability of the model.

[0072] Assume Gaussian noise is N(μ,σ 2 ), where μ is the mean, σ 2 is the variance. Its probability density function is:

[0073]

[0074] Among them, x represents a random variable, and its value range is the entire real number axis.

[0075] Uniform noise is a random, uniformly distributed noise that can be introduced in neural network training by adding uniformly distributed random values ​​to model parameters. Unlike Gaussian noise, the values ​​of uniform noise are distributed in a specified interval, and the probability of each value is equal. By introducing uniform noise, the diversity of the model can be increased, and its adaptability to uncertainty and noise can be improved, thereby enhancing the robustness and generalization ability of the model. Both types of noise help improve the performance of the model in real environments.

[0076] Assume that uniform noise is U(a,b), where a and b are the starting and ending values ​​of the interval respectively, and its probability density function is:

[0077]

[0078] Here, x also represents a random variable, and its value range is in the interval [a,b].

[0079] 2.2. GB is used to generate prior information

[0080] GB is a powerful and flexible machine learning algorithm that builds a strong classifier by combining multiple weak classifiers. It performs well in processing various types of data and complex nonlinear relationships and has strong generalization ability.

[0081] First, the output of the integrated GB is introduced into the neural network as prior knowledge to provide the model with additional data feature information. This strategy significantly enhances the model's understanding of the data structure and its learning efficiency.

[0082] Utilizing the prior information generated by GB helps the neural network model to generalize to novel data samples more effectively, thereby improving prediction accuracy. In practical application scenarios, especially in the case of disease data with limited sample size or sparse features, the performance and stability of the model may be challenged. Using GB output as prior knowledge can effectively alleviate the problems caused by data sparsity and enhance the data adaptability of the model. In addition, the integrated learning mechanism of the GB model performs well in reducing the risk of overfitting, providing the neural network with rich data feature information and enhanced learning potential.

[0083] A powerful ensemble model GB is constructed by iteratively training multiple weak learners.

[0084] For the training set (x i ,y i ), input feature x i Represents the relative abundance of each OTU in the sample, corresponding to the target variable y i Indicates the relative abundance of an OTU in a specific sample;

[0085] Construct a prediction model F(x) to represent the initial estimate of the average abundance of the microbial community in a sample, and then make the prediction value F(x i ) as close as possible to the true label y i ;

[0086] Initialize the model:

[0087]

[0088] Then, multiple weak learners are iteratively trained, and each iteration fits a new model to correct the residual of the previous model to represent the difference between the actual observed OTU abundance and the OTU abundance predicted by the model. The residual of the current model is:

[0089] r it =y i -F t-1 (x i ). (4)

[0090] Then, in the residual r itFit a new weak learner (a decision tree is used in this implementation) to minimize the residual error as much as possible:

[0091]

[0092] Among them, L is the loss function, which is used to measure the difference between the predicted value and the true value.

[0093] Finally, the new model is combined with the previous model through the learning rate to update the model's prediction results:

[0094] F t (x) = F t-1 (x)+γh t (x). (6)

[0095] Repeat the above steps until the preset number of iterations is reached or the residual meets certain conditions.

[0096] 2.3 GRU Model

[0097] The GRU model has significant advantages in disease phenotype prediction tasks.

[0098] First, the GRU model, through its focus on sequence modeling, can effectively capture the temporal information of microbial composition, thereby more accurately understanding the evolution of microbial communities.

[0099] Secondly, the self-attention layer is introduced into the GRU model, which enables it to focus on the key features in the OTU table, thereby improving the model's ability to identify and utilize important microbial populations.

[0100] In addition, the GRU model is also excellent in handling sparse features and can effectively utilize the large amount of zero-value information in microbial composition data.

[0101] The GRU model has strong diversity modeling capabilities and can flexibly adapt to the characteristics of different disease types and data distributions, thereby improving the applicability and generalization ability of the model and providing reliable technical support for the realization of disease phenotype prediction tasks.

[0102] For the GRU model, the OTU abundance vector at time step t is used as the input x of the current time step t , where each element in the OTU abundance vector represents the relative abundance of a specific OTU in the sample, and the OTU abundance information of the previous sample is used as the hidden state h of the previous time step t-1 , calculate the update gate r t and reset gate z t , the expression is:

[0103] r t =σ(Wr ·[x t ,h t-1 ]+b r ), (7)

[0104] z t =σ(W z ·[x t ,h t-1 ]+b z ), (8)

[0105] Where: W r Represents the weight matrix corresponding to the OTU abundance information that is most important for predicting the status of the current sample, W z represents the weight matrix corresponding to the new OTU information integrated into the model; b r and b z is the corresponding bias term; σ is the Sigmoid function;

[0106] Next, according to the input x of the current time step t and the hidden state h at the previous time step t-1 , calculate the candidate hidden state

[0107]

[0108] Where: n t is a noise term, which is Gaussian noise or uniform noise, and is used to simulate the variability and noise characteristics of actual intestinal microbiome data; W c is a weight matrix used to determine how to convert the reset OTU abundance information into a candidate hidden state; b c is the bias term; tanh is the hyperbolic tangent function; ⊙ represents element-wise multiplication.

[0109] Finally, according to the update gate r t , reset gate z t and candidate hidden states Calculate the new hidden state h of the prediction of the current sample OTU abundance t :

[0110]

[0111] By updating the gate, resetting the gate, and calculating the candidate hidden state, the hidden state h in the GRU model t can be obtained. In this process, n t is the added noise term, which can be Gaussian or uniform.

[0112] 3. Experimental setup

[0113] 3.1 Experimental Environment

[0114] This experiment was conducted on an Intel Core i7-10750H laptop with a six-core 3.0GHz processor and 32GB of memory. Specifically, the operating system was Windows 11, the integrated development environment was Visual Studio Code, the GPU was NVIDIA GeForce RTX 3060, the CUDA version was 11.4, the CUDNN version was 8.2, and the Python version was 3.7. This experiment used PyTorch as the main deep learning framework, and its version was 1.8.

[0115] 3.2 Dataset

[0116] All datasets in this example are from the Zendo database, an online platform for storing and sharing scientific research data, code, and other related materials (https: / / zenodo.org / ). Most of the data were downloaded from online repositories (such as SRA) or links provided in the original study, and studies that required additional ethics committee approval or authorized access were excluded. This example specifically analyzes 16S data from stool samples and excludes studies with fewer than 15 cases. In cases where there are multiple control groups in CRC studies, such as healthy groups and non-CRC adenoma groups, healthy patients are used as controls for all comparisons. For studies with non-healthy controls, such as non-IBD patients, these patients were used as controls, consistent with the original paper.

[0117] Public 16S processing tools were used to process the raw data. Usually, the sequences are truncated to the first base with a quality score less than 25 based on the quality score. However, some data sets failed to meet this strict quality standard (the generated OTU table either lacked most of the content in the original sample, or the read depth was significantly lower than that reported in the original paper). If all reads are trimmed to 200 bases, this length trimming will discard most of the sequences. Therefore, the threshold was lowered to 150 or 101 bases to ensure that most of the sequences are not lost. The OTUs were clustered at 100% similarity using USEARCH software. For each data set, samples with less than 100 reads and OTUs with less than 10 reads were deleted, and OTUs with a frequency of less than 1% in the study were deleted. The relative abundance of each OTU was calculated by dividing the value of each OTU by the total number of reads per sample using USEARCH software, and the OTUs were collapsed to the genus level. All statistical analyses in this example were performed based on relative abundance data at the genus level. Finally, 14 data sets covering 9 diseases were processed for disease phenotype prediction. The detailed information of each dataset is shown in Table 1.

[0118] Table 1 Sample quantity information in each data set

[0119]

[0120] 3.3 Performance Evaluation Indicators

[0121] AUC is a commonly used evaluation metric, usually used to evaluate the performance of classification models. It reflects the predictive ability of the model under different thresholds. In the field of disease phenotype prediction, AUC is widely used to measure the accuracy and reliability of the model in classifying different diseases. The AUC value ranges from 0 to 1. The closer the value is to 1, the better the model performance is, while the closer the value is to 0.5, the closer the model performance is to random prediction. For example, for a prediction model for a specific disease, if its AUC value is close to 1, it means that the model shows a high accuracy in distinguishing between diseased and healthy samples; if the AUC value is close to 0.5, it means that the classification effect of the model is no different from random guessing and the performance is poor.

[0122] Compared with other classification indicators, AUC has three main advantages: First, when faced with a dataset with imbalanced categories, such as when the number of diseased samples and healthy samples is unbalanced in disease prediction, the AUC value can measure the performance of the model well and will not be affected by the imbalanced category; Second, the calculation of the AUC value does not depend on the specific threshold selection, but comprehensively considers the model performance under all possible thresholds. Therefore, it is insensitive to changes in model performance under different thresholds and has strong robustness; Third, the AUC value is a general evaluation indicator that can effectively evaluate model performance in various classification tasks. It is not restricted by data type and field, so it has a wide range of applicability.

[0123] Therefore, this embodiment selects the AUC value as the evaluation indicator in order to intuitively and comprehensively evaluate the disease phenotype prediction performance of each prediction method, thereby providing a reliable evaluation and reference for its application in clinical practice.

[0124] 3.4 Experimental Results

[0125] 3.4.1. Method Comparison

[0126] Table 2 shows the performance evaluation results of five different prediction methods (SVM, RF, MLP, MVIB, NPENN) under different data sets, where the performance evaluation results are expressed as AUC indicators. It can be seen from Table 2 that in most cases, the AUC values ​​of each method on different data sets vary within a relatively consistent range. For example, for the ASD-Kang data set, the AUC values ​​of all algorithms are higher than those on other data sets. In each data set, the performance of different algorithms varies. For example, in the CRC-Xiang data set, the AUC values ​​of all algorithms are relatively high, while in the OB-Ross data set, the AUC values ​​of all algorithms are relatively low. In the OB-Ross data set, the AUC value of the NPENN method is 0.63, while the AUC of RF is 0.67, and the prediction result is the best. In the OB-Jumpertz data set, the AUC value of the NPENN method reaches 0.98, which is 25% higher than the SVM and 12% higher than the MVIB algorithm. On the PAR-Scheperjans dataset, the MVIB method achieved an AUC value of 0.66, while the AUC value of NPENN was only 0.62, and the prediction effect was not as good as RF and MLP. In the CDI-Vincent dataset, all methods performed well, with AUC values ​​ranging from 0.8 to 0.9. Overall, NPENN and RF performed well on most datasets, followed by MVIB, while the performance of SVM and MLP was relatively low. Among the 14 datasets, NPENN achieved an AUC value of more than 0.9 on 4 datasets, while MVIB achieved an AUC value of more than 0.9 on only 2 datasets.

[0127] Table 2 AUC values ​​of each method on different datasets

[0128]

[0129]

[0130] Figure 2The average AUC value of each method on different data sets is shown. As can be seen from the figure, the average AUC value of the NPENN method is the highest, which is 0.8059, showing the best classification performance. The MVIB method is second, with an average AUC value of 0.7471. The average AUC value of the RF method is better than that of the MLP. Overall, the prediction effect of NPENN on multiple disease phenotype data sets can reach more than 80%, and the prediction effect of MVIB, RF and MLP on multiple disease phenotype data sets can reach more than 70%. First, the AUC values ​​of all algorithms on the ASD-Kang data set are generally high, which may mean that there is a more obvious correlation between the features in the ASD-Kang data set and the target variable, so that various algorithms can effectively capture this correlation and achieve a higher prediction accuracy. In contrast, the generally low AUC value on the OB-Ross data set indicates that the data set may contain more noise or the relationship between the features and the target is not obvious enough, which brings challenges to the learning of the model.

[0131] MVIB performs worse than NPENN and RF on some datasets, probably because the MVIB model relies more on the multi-view characteristics of the data. If the different views of the dataset are not complementary enough or the quality of some views is poor, the performance of the MVIB model may be limited. From the overall trend, the NPENN method has the highest average AUC value, which shows that the method has strong adaptability and stability when considering the characteristics and challenges of different datasets.

[0132] 3.4.2 Ablation Experiment

[0133] Considering that the NPENN model introduces the prior information of the noise mechanism and the gradient boosting method, it is necessary to set up an ablation experiment to verify the effectiveness of the model.

[0134] Ablation experiments can help evaluate the impact of noise mechanisms on model performance, determine the role of prior information of gradient boosting methods in the model, evaluate the contribution of each module, and verify the effectiveness of neural network integration. By conducting ablation experiments, we can have a deeper understanding of the role of each module of the model, optimize the model design, and improve the performance and generalization ability of the model. In addition, ablation experiments can also help discover potential problems in the model and provide guidance for model improvement.

[0135] This embodiment sets five ablation models, namely GB, NPENN (Noise-; GB+), NPENN (Noise+; GB-), NPENN (Noise-; GB-) and NPENN.

[0136] Specifically, the GB model is a baseline model that uses only the gradient boosting method for prediction without introducing a noise mechanism; the NPENN (Noise-; GB+) model does not introduce a noise mechanism but combines the prior information of the gradient boosting method; the NPENN (Noise+; GB-) model introduces a noise mechanism but does not combine the prior information of the gradient boosting method; the NPENN (Noise-; GB-) model neither introduces a noise mechanism nor combines the prior information of the gradient boosting method; NPENN is a complete model that includes a noise mechanism and prior information combined with the gradient boosting method. Table 3 shows the AUC values ​​of each model on different datasets.

[0137] Table 3 AUC values ​​of each ablation model on different datasets

[0138]

[0139] As can be seen from Table 3, the AUC values ​​of NPENN(Noise+; GB-) on 8 data sets are higher than those of NPENN(Noise-; GB+), while the AUC values ​​of NPENN(Noise-; GB+) on 4 data sets are higher than those of NPENN(Noise+; GB-). Therefore, it can be preliminarily inferred that the contribution of the noise perturbation mechanism to the model performance is higher than the prior information of GB. Specifically, the AUC values ​​of the NPENN(Noise+; GB-) model on most data sets are higher than those of the NPENN(Noise-; GB+) model, which indicates that the introduction of the noise mechanism has a positive effect on improving the model performance. Although the NPENN(Noise-; GB+) model performs better than the NPENN(Noise+; GB-) model on a few data sets, considering the overall trend, the noise mechanism plays a more important role in improving the model performance.

[0140] The results show that when processing microbiome datasets, the noise mechanism plays a key role in the generalization ability and robustness of the model, which can help the model better adapt to the complexity and heterogeneity of the data, thereby improving the predictive performance of the model on unknown microbiome data.

[0141] 3.5 Summary and Discussion

[0142] This embodiment proposes a new disease phenotype prediction model NPENN. The NPENN model combines the ideas of neural networks and ensemble learning to improve the prediction performance on microbiome datasets. First, the NPENN model uses neural networks as the basic architecture, taking advantage of its advantages in processing complex data and learning nonlinear relationships. Specifically, GRU is used as the basic component of the neural network to learn the complex relationship between the representation of input data and features. Most importantly, the NPENN model introduces a noise perturbation mechanism to deal with noise and uncertainty in microbiome data. During the model training process, the complexity of the model is constrained by adding noise penalty terms to prevent overfitting and improve generalization ability. This generalization ability is reflected in the fact that this embodiment uses a variety of microbial disease phenotype data from different intestinal sites of different samples, and achieves good prediction results. In this embodiment, AUC is used as a performance evaluation indicator to evaluate the accuracy and reliability of different prediction methods in disease phenotype prediction. The selection of AUC values ​​is based on its wide application in classification model performance evaluation and a comprehensive reflection of the model prediction ability under different thresholds. The experimental results show that the AUC values ​​of each method on different data sets change relatively consistently, but in each data set, the performance of different methods varies. For example, in the CRC-Xiang dataset, the AUC values ​​of all algorithms are relatively high, while they are lower in the OB-Ross dataset. Overall, NPENN and RF perform better on most datasets, followed by MVIB, while SVM and MLP have relatively low performance. In addition, by calculating the average AUC value, we found that the NPENN method performs best on all datasets, followed by MVIB, which indicates that the NPENN method has high performance in predicting multiple disease phenotypes.

[0143] The results of ablation experiments showed that the introduction of noise mechanism had a positive effect on improving model performance, although on some data sets, the model without the introduction of noise mechanism performed slightly better. This suggests that the introduction of noise mechanism helps the model better adapt to the complexity and heterogeneity of the data. The results of feature importance analysis showed that for most disease data sets, the first few predictive features contributed most significantly to the results, while the subsequent features contributed less. In addition, the box plot showed the degree of variation of features in each disease, indicating that there were differences in microbial composition between different diseases. Through PCA and PCoA analysis, the samples were projected into a low-dimensional space to observe their differences, but the results showed that the discrimination of samples from different diseases in the space was low. Finally, the number of significantly changed microorganisms in CRC, CDI and IBD was shown by the microbial volcano map, which may be related to the pathogenesis of the disease. We verified microbial markers associated with specific diseases through the Peryton database. The results showed that Bacteroides appeared in a variety of diseases, while Faecalibacterium was associated with IBD and PD. These microorganisms may become disease markers, which will help to deeply understand the disease mechanism and formulate personalized medical strategies.

[0144] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0145] Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunication networks.

[0146] The processing unit performs the various methods and processes described above. For example, in some embodiments, the method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the method in any other appropriate manner (e.g., by means of firmware).

[0147] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip systems (SOCs), load programmable logic devices (CPLDs), and the like.

[0148] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0149] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0150] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for classifying microbiome data based on noise perturbation ensemble neural network, characterized in that: The method includes: Obtain gut microbiome data; The intestinal microbiome data are input into a multi-classification model based on a noise perturbation ensemble neural network to obtain a classification result of the microbiome data; wherein the multi-classification model based on a noise perturbation ensemble neural network includes a gradient boosting classifier and a gated recurrent unit trained by introducing a noise perturbation mechanism, the intestinal microbiome data are input into the gradient boosting classifier to generate a first classification result, and the first classification result is input into the gated recurrent unit as prior knowledge to output the final microbiome data classification result.

2. A method for classifying microbiome data based on noise perturbation ensemble neural network according to claim 1, characterized in that: The method further includes preprocessing the acquired intestinal microbiome data using a 16S processing tool, specifically: The data were filtered to retain only sequences with a length of 101 to 150 bases; OTUs were clustered at a 100% similarity threshold; OTUs that appeared in less than 100 samples or had less than 10 reads per sample were removed, and OTUs with a relative abundance lower than 1% in the entire dataset were removed; The relative abundance of each OTU was calculated by dividing the number of reads of the OTU by the total number of reads per sample; The taxonomic information of OTUs was summarized to the genus level, that is, the relative abundances of all OTUs belonging to the same genus were added together.

3. A method for classifying microbiome data based on noise perturbation ensemble neural network according to claim 2, characterized in that: The gated recurrent unit processes sequence data through its gating mechanism and utilizes prior knowledge to enhance classification capabilities. Its data processing process includes: Take the OTU abundance vector at time step t as the input x of the current time step t , where each element in the OTU abundance vector represents the relative abundance of a specific OTU in the sample, and the OTU abundance information of the previous sample is used as the hidden state h of the previous time step t-1 , calculate the update gate r t and reset gate z t , the expression is: r t =σ(W r ·[x t ,h t-1 ]+b r ), (1) z t =σ(W z ·[x t ,h t-1 ]+b z ), (2) Where: W r Represents the weight matrix corresponding to the OTU abundance information that is most important for predicting the status of the current sample, W z represents the weight matrix corresponding to the new OTU information integrated into the model; b r and b z is the corresponding bias term; σ is the Sigmoid function; According to the input x of the current time step t and the hidden state h at the previous time step t-1 , calculate the candidate hidden state Where: n t is a noise term, which is Gaussian noise or uniform noise, and is used to simulate the variability and noise characteristics of actual intestinal microbiome data; W c is a weight matrix used to determine how to convert the reset OTU abundance information into a candidate hidden state; b c is the bias term; tanh is the hyperbolic tangent function; ⊙ represents element-wise multiplication. According to the update gate r t , reset gate z t and candidate hidden states Calculate the new hidden state h of the prediction of the current sample OTU abundance t :

4. The method for classifying microbiome data based on noise perturbation ensemble neural network according to claim 1, characterized in that: The gradient boosting classifier captures the nonlinear relationship of the intestinal microbiome data by constructing a decision tree to generate a first classification result.

5. The method for classifying microbiome data based on noise perturbation ensemble neural network according to claim 1, characterized in that: In the multi-classification model based on noise perturbation ensemble neural network, a self-attention mechanism is introduced to calculate the relationship between each position in the sequence and assign a weight to each position to focus on the most important information for the current task.

6. The method for classifying microbiome data based on noise perturbation ensemble neural network according to claim 1, characterized in that: The training process of the gradient boosting classifier includes: For the training set (x i ,y i ), input feature x i Represents the relative abundance of each OTU in the sample, corresponding to the target variable y i Indicates the relative abundance of an OTU in a specific sample; Construct a prediction model F(x) to represent the initial estimate of the average abundance of the microbial community in a sample, and then make the prediction value F(x i ) as close as possible to the true label y i ; Initialize the model: Iteratively train multiple weak learners, and fit a new model each iteration to correct the residual of the previous model to represent the difference between the actual observed OTU abundance and the OTU abundance predicted by the model. The residual of the current model is: r it =y i -F t-1 (x i ), (6) In the residual r it Fit a new weak learner on it to predict the structure of the microbial community in the sample: Among them, L is the loss function, which is used to measure the difference between the predicted value and the true value; Combine the new model with the previous model through the learning rate and update the model prediction results: F t (x)=F t-1 (x)+γh t (x), (8) The training is repeated until a preset number of iterations is reached or the residual meets the set conditions.

7. The method for classifying microbiome data based on noise perturbation integrated neural network according to claim 6, characterized in that: The weak learner is a decision tree.

8. A microbiome data classification system based on noise perturbation ensemble neural network, characterized in that: include: Input module to obtain gut microbiome data; A classification module, wherein the multi-classification model based on the noise perturbation integrated neural network includes a gradient boosting classifier and a gated recurrent unit trained by introducing a noise perturbation mechanism, the intestinal microbiome data is input into the gradient boosting classifier to generate a first classification result, and the first classification result is input into the gated recurrent unit as prior knowledge to output the final microbiome data classification result.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images