Microbial host phenotype prediction system and method fused with taxonomy level

Abstract

The invention relates to the technical field of microbial data mining, and provides a microbial host phenotype prediction system and method fusing taxonomy levels. The method comprises a taxonomy level OTU extraction module used for obtaining metagenome data and extracting an abundance table according to taxonomy levels; the taxonomy level feature selection module is used for screening representative features of each level based on an abundance table according to a set screening standard; the taxonomy hierarchical feature optimization module is used for evaluating the importance of the representative features of each hierarchy and selecting the representative features of each hierarchy with key biomarkers; the taxonomy level image generation module is used for splicing the representative features of each level with the key biomarkers to obtain multi-scale and multi-level features; converting the multi-scale and multi-level features into a two-dimensional grayscale image; and the microbial image classification module is used for predicting to obtain a host phenotype prediction result based on the two-dimensional grayscale image.

Description

Technical Field

[0001] The present invention relates to the technical field of microbial data mining, and in particular to a microbial host phenotype prediction system and method integrating taxonomic levels. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The microbiome is becoming increasingly crucial in predicting human disease. Early studies captured microbial sequence features and applied machine learning algorithms, such as support vector machines (SVM), k-nearest neighbor (KNN), and random forests (RF), to diagnose disease. However, these methods rely on manual feature selection and are prone to subjective bias. Deep learning methods, such as convolutional neural networks (CNNs), have significantly improved diagnostic accuracy but still overlook the inherent relationships within microbial data. Existing methods focus solely on relative abundance (RA) at the genus and species levels, failing to effectively capture taxonomic hierarchies (e.g., phylum, class, order, family, genus, and species) and their higher-order relationships. Furthermore, the sparsity of metagenomic data and sample-feature imbalance (often fewer than 1,000 or even 100 samples) further complicate modeling. The success of computer vision techniques, such as object detection and image segmentation, has inspired us to apply them to the study of microbial data. However, the proper conversion of microbial data into images remains challenging: improper image construction can lead to information loss or noise. Furthermore, integrating microbial knowledge across different taxonomic levels is crucial, but current research utilizes only a single level of domain knowledge and fails to mine deep semantic relationships across multiple layers. Summary of the Invention

[0004] To address the technical problems encountered in the aforementioned background technology, the present invention provides a taxonomically integrated microbial host phenotype prediction system and method. By combining computer vision and deep learning techniques, MicroVIS successfully selects, optimizes, and integrates multi-level features of the microbiome, effectively converting them into images. This innovation not only overcomes the limitations of traditional methods in high-dimensional data processing and feature selection, but also significantly improves the accuracy of microbiome host phenotype prediction compared to traditional methods that rely solely on species abundance analysis. It also reveals the enormous potential of multi-level feature fusion and image conversion in microbiological data analysis. Future work will enhance MicroVIS to utilize biomarkers from taxonomically integrated features to predict disease treatments and improve prediction accuracy. This will enable personalized interventions to effectively modulate microbiome composition to combat disease.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The first aspect of the present invention provides a microbial host phenotype prediction system integrating taxonomic levels.

[0007] A microbial host phenotype prediction system integrating taxonomic levels, including:

[0008] The taxonomic level OTU extraction module is used to obtain metagenomic data and extract the abundance table according to the taxonomic level;

[0009] The taxonomic hierarchical feature selection module is used to select representative features of each level based on the abundance table and the set screening criteria;

[0010] The taxonomic hierarchical feature optimization module is used to evaluate the importance of representative features at each level and select representative features of each level with key biomarkers;

[0011] The taxonomic hierarchical image generation module is used to stitch together the representative features of each level with key biomarkers to obtain multi-scale and multi-level features; and convert the multi-scale and multi-level features into a two-dimensional grayscale image;

[0012] The microbial image classification module is used to predict host phenotype prediction results based on two-dimensional grayscale images.

[0013] Furthermore, the taxonomic level OTU extraction module is specifically used to:

[0014] Trim the terminal sequencing reads of the metagenomic data to obtain the initial sequence;

[0015] Demultiplexing the initial sequence;

[0016] For the demultiplexed initial sequence, the primer sequence is trimmed to remove the non-biological sequence to obtain a qualified double-ended original sequence;

[0017] Qualified double-end raw sequences were filtered, denoised, and spliced, and chimeric sequences were removed to obtain all representative sequences;

[0018] All representative sequences were annotated and abundance tables for each taxonomic level were generated.

[0019] Furthermore, the screening criteria are:

[0020] like Then discard feature j

[0021] Among them, Z j represents the number of samples in which the jth microorganism is zero in all samples Where I(x i,j =0) is the indicator function, if x i,j =0 then I(x i,j=0)=1, otherwise I(x i,j =0) = 0; if the jth feature shows zero value in more than 80% of the samples Then discard it.

[0022] Furthermore, the taxonomic hierarchical feature optimization module is specifically used to: use the Gini index based on random forest to evaluate feature importance, sort according to feature importance scores, select a certain proportion of top-ranked features, and obtain representative features of each level with key biomarkers.

[0023] Furthermore, the taxonomic hierarchical image generation module is specifically used to: splice the representative features of each level with key biomarkers at different classification levels according to certain rules to obtain multi-scale multi-level features; arrange the multi-scale multi-level features into a square matrix; if the length of the multi-scale multi-level feature is not enough to fill the square matrix, fill the end of the multi-scale multi-level feature with zero values to ensure that a complete square matrix is obtained; normalize the square matrix so that its value range is mapped to the grayscale interval of [0,255] and convert it into a two-dimensional grayscale image.

[0024] Furthermore, the microbial image classification module is specifically used to: obtain host phenotype prediction results based on two-dimensional grayscale images using the ResNet network; wherein the specific processing process of the ResNet network includes: based on the two-dimensional grayscale image, performing preliminary feature extraction through a 3×3 convolution layer to generate a feature map; inputting the feature map into four levels of residual block stacking in sequence, each level containing num_blocks[l] residual blocks, wherein the feature map size of the lth layer is The number of channels is C l After four levels of residual blocks are stacked, the spatial dimension of the feature map is reduced to 1x1 through the global average pooling layer, and it is mapped to the logits of the number of categories through the fully connected layer; after the logits are softmaxed, the predicted probability distribution of each category is obtained.

[0025] The second aspect of the present invention provides a method for predicting microbial host phenotypes that integrates taxonomic levels.

[0026] A taxonomically integrated microbial host phenotype prediction method, including:

[0027] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0028] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0029] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0030] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0031] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0032] A third aspect of the present invention provides a computer device, comprising:

[0033] a processor adapted to execute a computer program;

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0035] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0036] Based on the abundance table and the set screening criteria, representative features of each level are screened;

[0037] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0038] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0039] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0040] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is suitable for being loaded by a processor and executing the following steps:

[0041] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0042] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0043] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0044] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0045] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0046] A fifth aspect of the present invention provides a computer program product or computer program.

[0047] The present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the following steps:

[0048] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0049] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0050] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0051] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0052] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention provides a microbial host phenotype prediction system and method that integrates taxonomic hierarchies. The system includes a taxonomic-level OTU extraction module, which acquires metagenomic data and extracts an abundance table based on the taxonomic hierarchy. A taxonomic-level feature selection module, based on the abundance table and set screening criteria, selects representative features from each hierarchy. A taxonomic-level feature optimization module, which assesses the importance of representative features from each hierarchy and selects those with key biomarkers. A taxonomic-level image generation module, which concatenates representative features from each hierarchy with key biomarkers to generate multi-scale, multi-level features, converts these features into two-dimensional grayscale images, and finally, a microbial image classification module, which predicts host phenotypes based on the two-dimensional grayscale images. The MicroVIS framework converts multi-level microbial data into multi-scale images, leveraging microbiology domain knowledge and advanced computer vision technology. MicroVIS first integrates taxonomic domain expertise to reveal complex, high-level semantic relationships within the microbiome. A conversion strategy is then employed to convert the multi-level microbial features into two-dimensional images. Finally, it leverages advanced visual deep learning methods to learn feature embeddings to predict associations between the microbiome and disease, providing a novel perspective on domain knowledge and microbial data visualization. MicroVIS facilitates the identification of biologically meaningful biomarkers, highlighting its potential for broader disease prediction applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0056] Figure 1 1 is a structural diagram of a microbial host phenotype prediction system integrating taxonomic levels, as shown in an embodiment of the present invention;

[0057] Figure 2 1 is a network structure diagram of a microbial host phenotype prediction system integrating taxonomic levels, as shown in an embodiment of the present invention;

[0058] Figure 3 is a structural diagram of a BasicBlock residual block shown in an embodiment of the present invention;

[0059] Figure 4 is a schematic diagram of eight data sets shown in an embodiment of the present invention;

[0060] Figure 5 Schematic diagram of AUC comparison results on 16s rRNA dataset shown in the embodiment of the present invention;

[0061] Figure 6Schematic diagram of the performance distribution of MicroVIS and the benchmark method on the 16s rRNA dataset shown in the embodiment of the present invention;

[0062] Figure 7 Schematic diagram of the comparison results of the MicroVIS model and the benchmark model on eight real metagenome multimodal datasets shown in an embodiment of the present invention;

[0063] Figure 8 Schematic diagram of ablation experiment results on a 16s rRNA dataset shown in an embodiment of the present invention;

[0064] Figure 9 Schematic diagram of ablation experiment results on a metagenomic multimodal dataset shown in an embodiment of the present invention;

[0065] Figure 10 It is a structural diagram of a computer device shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0067] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0068] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0069] In order to facilitate understanding of the technical solution of the present invention, some technical terms involved in the present invention are introduced below.

[0070] Random Forest (RF) is a supervised learning algorithm based on ensemble learning, mainly used for classification and regression tasks. It improves the accuracy and robustness of the model by constructing multiple decision trees and combining their prediction results.

[0071] Stochastic Gradient Descent (SGD) is an optimization algorithm used to minimize an objective function (such as a loss function).

[0072] The cross entropy loss function is a commonly used loss function in machine learning to measure the difference between the model's predicted probability distribution and the true distribution. It is widely used in classification tasks and is particularly suitable for situations where the predicted output is a probability distribution.

[0073] Based on this, Figure 1 、 Figure 2 As shown, the present invention provides a microbial host phenotype prediction system integrating taxonomic levels, comprising:

[0074] The taxonomic level OTU extraction module is used to obtain metagenomic data and extract the abundance table according to the taxonomic level;

[0075] The taxonomic hierarchical feature selection module is used to select representative features of each level based on the abundance table and the set screening criteria;

[0076] The taxonomic hierarchical feature optimization module is used to evaluate the importance of representative features at each level and select representative features of each level with key biomarkers;

[0077] The taxonomic hierarchical image generation module is used to stitch together the representative features of each level with key biomarkers to obtain multi-scale and multi-level features; and convert the multi-scale and multi-level features into a two-dimensional grayscale image;

[0078] The microbial image classification module is used to predict host phenotype prediction results based on two-dimensional grayscale images.

[0079] Specifically, the system described in this embodiment includes:

[0080] Taxonomic OTU extraction module: extracts multi-level operational taxonomic units (OTUs) from microbial 16S rRNA sequencing data;

[0081] Taxonomic hierarchical feature selection module: selects representative features of each level through sparsity filtering;

[0082] Taxonomic hierarchical feature optimization module: Gini index based on random forest (RF) is used to evaluate feature importance and retain key biomarkers;

[0083] Taxonomy hierarchical image generation module: concatenates multi-level features into 1D vectors and converts them into 2D grayscale images;

[0084] Microbial image classification module: uses an improved ResNet architecture for host phenotype prediction.

[0085] The present invention utilizes microbial data visualization to convert the one-dimensional relative abundance of microorganisms into a two-dimensional image, takes advantage of the powerful feature learning ability of deep learning, and alleviates overfitting by using a residual neural network.

[0086] In some embodiments, operational taxonomic units (OTUs) are extracted from different taxonomic levels. The extraction process for each raw sample obtained from microbiome sequencing includes the following steps:

[0087] Data preprocessing: Trim the end sequencing reads of each downloaded sample to maintain high data quality;

[0088] Sequence import and demultiplexing: The original sequence (FASTQ format) is imported into QIIME 2 and demultiplexed in QIIME 2. In this embodiment, the QIIME 2 software package can be used, and other software that achieves the same function can also be used.

[0089] Primer trimming: Using the Cutadapt plugin, primer sequences were trimmed to remove non-biological sequences such as primers, sequencing adapters, and PCR spacers.

[0090] Quality control: The paired-end raw data were quality controlled using DADA2 using the default parameters of QIIME 2. This step also filters out phiX reads that are common in Illumina sequence data and removes chimeric sequences;

[0091] Sequence annotation: All representative sequences were annotated using the QIIME 2 software package and the SILVA database to generate OTU tables at each taxonomic level (including species, genus, family, order, class, and phylum).

[0092] These steps are consistent with established guidelines and standards for effectively training and validating classifiers;

[0093] The 16S rRNA sequencing dataset was collected through legal and compliant means and includes 95 subjects, including 35 HF (heart failure) patients with depressive symptoms (DS), 36 HF patients without depressive symptoms (NDS), and 24 healthy controls (HC). For ease of description, this dataset will be referred to as MDDHC.

[0094] In one or more embodiments, for each downloaded raw sample, the end sequencing reads are trimmed for quality control. (1) The raw sequence (FASTQ) data is imported into QIIME 2, and the primer sequences are trimmed for quality control using Cutadapt software. Low-quality reads (Phred quality score <20) are removed using Sickle33; (2) According to the default parameters of QIIME2, qualified double-end raw data are subjected to high-quality filtering, noise reduction, splicing, and de-chimerization processing using DADA2, which not only reduces sequencing errors but also removes sequence redundancy; (3) Representative sequences of each ASV are selected using the QIIME2 software package, and all representative sequences are evaluated and annotated using the Silver database. Amplicon sequence variations and operational taxonomic units (OTUs), as well as abundance tables at each taxonomic level (species, genus, family, order, class, phylum), are obtained. All methods are performed in accordance with relevant guidelines and regulations. This is used to train and validate the classifier.

[0095] The abundance table of each level of the microbial metasystem and its class labels are used as input to the model. For each sample in the dataset, the abundance table captures the relative abundance curves at the species, genus, family, order, class, and phylum levels as follows:

[0096] OTU * ∈{OTU Species ,OTU Genus ,OTU Family ,OTU Order ,OUT Class ,OTU Phylum}

[0097] For simplicity, S,…,P are used to replace Species,…,Phylum respectively. * ∈{OTU S ,OTU G ,OTU F ,OTU O ,OTU C ,OTU P}. OTU * is considered as the characteristic of the sample. Let n be the total number of samples in the dataset. Then, the sample set in the dataset is recorded as {s1, s2, s3, ..., s n}. Sample s in the dataset i Indicated as s i =(S i ,G i ,F i ,O i ,Ci ,P i ,y i ), where S i -P i For s i The abundance feature vector from Species to Pylum, y i For s i Class labels of samples s i Species-level eigenvectors The length is l S , Indicates the jth microbial feature at the species level in the i-th sample s i The value in . And so on (that is, each sample contains six feature vectors of different classification levels, with lengths of l S ,l G ,l F ,l O ,l C ,l P , a class label y i ). Considering the distribution pattern of microbial abundance characteristics, which is usually dense in the early stage and sparse in the later stage, the present invention adopts the following screening criteria:

[0098] like Then discard feature j

[0099] Define Z j represents the number of samples in which the jth microorganism is zero in all samples Where I(x i,j =0) is the indicator function, if x i,j =0 then I(x i,j =0)=1, otherwise I(x i,j =0) = 0. If the jth feature shows a zero value in more than 80% of the samples This feature has low information content and can be discarded. This screening strategy can effectively reduce the data dimension and focus on those features that are more representative and different, thereby improving the accuracy and efficiency of subsequent analysis.

[0100] Due to the high imbalance between sample and feature dimensions, and in order to clearly represent microbial markers of biological insights, the present invention also introduces a feature screening step to reduce sample feature dimensions, which helps reduce model complexity, improve model performance, and supplement the model's poor interpretability.

[0101] The present invention uses a feature screening strategy based on random forest to evaluate feature importance and reduce feature dimension. Each classification tree in the random forest is a binary tree, and the root node contains all the training self-help samples. According to certain principles, at each node, the variable that minimizes the "impurity" of the node after branching is selected from a set of randomly selected variables as the branch variable, and splits into left nodes and right nodes, which respectively contain a subset of the training data. The nodes after splitting continue to split according to the same rules until the branching stop rule is met and stops growing. The Gini index is used here as an evaluation indicator to measure "impurity". Variable screening uses variable importance score statistics to compare the contribution of features and retain features with greater contributions. The present invention uses GI to represent the Gini index, and uses Represents the characteristic vector of the sample at the species level, with l S variables, we need to calculate l S Score statistics. Represents the variable x j The score statistic is used, that is, the average change in node split impurity of the j-th feature in all decision trees of RF. The calculation formula of the Gini index of node q in the m-th tree is:

[0102]

[0103] Among them, C means there are C categories, p qc Indicates the proportion of category c in node q. Intuitively speaking, it is the probability that two samples randomly drawn from node q have inconsistent category labels. Feature x j The importance of node q in the mth tree, that is, the change in the Gini index before and after the node q branches is:

[0104]

[0105] in, and Respectively represent the Gini index of the two new nodes after branching.

[0106] If feature x j The nodes that appear in the decision tree m are set Q, then x j The importance of the i-th tree is:

[0107]

[0108] There are M trees in RF

[0109]

[0110] Finally, all the obtained importance scores are normalized.

[0111]

[0112] Sort the obtained feature importance scores and retain the top 80% of the features to obtain the species-level feature vector S", which has a length of l" after feature screening. S The feature vectors at other levels are also updated through the above process.

[0113] For traditional deep learning network applications, deeper convolutional neural networks can capture richer and more complex features and generalize well to new tasks. Microbial data often has highly complex features that may require deep networks to fully capture. However, deep networks often encounter gradient vanishing or gradient exploding problems during training, resulting in degraded model performance. ResNet effectively alleviates these problems by introducing residual blocks and shortcut connections, allowing the construction and training of deeper networks to better learn and represent complex patterns in microbial data.

[0114] In the task of microbial disease prediction, features at a single classification level may not be able to fully capture the complexity of microbial data. In order to better utilize information at different levels, this paper adopts a multi-level and multi-scale feature fusion method. Specifically, first, take sample s i For example, the features S at different classification levels i ,G i ,F i ,O i ,C i ,P i According to certain rules, they are spliced together to form new multi-scale and multi-level features X with a length of L. These features contain information at different classification levels and can capture the multi-scale structure of the data. Then, the spliced feature vectors X are arranged into a square matrix A k×k The side length k is:

[0115]

[0116] in, Represents the rounding function, k is not less than The total capacity of the square matrix is k 2 , if the length of X is not enough to fill the square matrix, that is, k 2 >L, it is necessary to fill the end of the vector with zeros to ensure that a complete square matrix can be formed. The specific filling rules are as follows: (i) Starting from the first element of the eigenvector, fill the square matrix in a zigzag order. (ii) When filling to the edge of the matrix, change the filling direction and continue filling until the square matrix is filled. Then, the characteristic matrix A after the square matrix is arranged k×kNormalize the image so that its value range is mapped to the grayscale interval of [0,255] and convert it into a grayscale image. The specific normalization steps are: find the maximum value max(A) and the minimum value min(A) in the feature matrix and perform normalization:

[0117]

[0118] The normalized matrix A norm Convert to grayscale image I gray , each matrix element corresponds to the grayscale value of a pixel.

[0119] Residual blocks are the fundamental building blocks of residual networks, used to address degradation issues in deep neural networks. Deep networks often encounter vanishing or exploding gradients during training, causing model performance to decline as network depth increases. ResNet alleviates this problem by introducing residual learning.

[0120] The present invention uses BasicBlock residual blocks and adopts residual learning for every few stacked layers. Figure 3 shown.

[0121] The building blocks are defined as:

[0122] y=F(x,W i )+x

[0123] Here x and y are the input and output vectors of the layer under consideration. The function F(x,W i )+x represents the residual mapping to be learned. The residual block in the figure contains two consecutive convolutional layers, denoted as F=W2σ(W1x), each of which is followed by a batch normalization layer and a ReLU activation function, where σ represents the ReLU activation function. In order to simplify the notation, the bias is omitted. The operation F+x is performed by a shortcut connection and element addition. The present invention adopts the second nonlinearity after addition (i.e., σ(y)). The dimensions of x and F are the same in y=F(x,W i )+x must be equal. If they are not equal (for example, when changing the input / output channels), a linear projection W is performed through the shortcut connection s To match the dimensions:

[0124] y=F(x,W i )+W s x

[0125] Where y is the output, x is the input, F is the residual function, and W iThe advantage of this is that it can increase the depth and performance of the network without affecting the convergence speed and stability of the network. The identity mapping is sufficient to solve the degradation problem and is economical, so W s Only used when matching dimensions.

[0126] The overall structure of ResNet used in the present invention includes four levels of stacking, each level consists of multiple residual blocks. The size of the input image is k×k, and the number of channels is 1. The number of residual blocks at each level is represented by the parameters num_blocks[0] to num_blocks[3]. In the implementation, the network first receives the input image and performs preliminary feature extraction through a 3×3 convolutional layer to generate a feature map. Then, the input is sequentially passed through four levels of residual block stacking, each level contains num_blocks[l] residual blocks, where the feature map size of the lth layer is The number of channels is C l These layers progressively reduce the size of the feature map and increase its depth to better capture the abstract features of the image. After four layers of residual blocks are stacked, the spatial dimension of the feature map is reduced to 1x1 through a global average pooling layer and mapped to the number of class logits through a fully connected layer. These logits are then softmaxed to obtain the predicted probability distribution for each class.

[0127] In this embodiment, the cross entropy loss function is used to calculate the difference between the predicted output and the actual label. The formula is as follows:

[0128]

[0129] Among them, N is the batch size, C is the number of categories, here is 2, target i,c is the true label of the cth class of the i-th sample, softmax(x i ) c is the model’s predicted probability for the cth class of the i-th sample. This loss function trains the model by minimizing the cross entropy between the predicted probability distribution and the true label, enabling the model to predict categories more accurately.

[0130] In order to optimize the model parameters, the present invention adopts stochastic gradient descent, by calculating the gradient of each parameter and updating the parameter value in the opposite direction of the gradient. During the training process, the present invention sets the learning rate to 10 -3(i.e. 0.001). The learning rate controls the step size of each parameter update. Too large a learning rate may cause oscillations, while too small a learning rate may cause slow convergence. In the experiments of the embodiments, a learning rate of 0.001 is generally able to effectively converge the model while maintaining training stability. In addition, the present invention also applies momentum to accelerate the optimization process. Momentum can help SGD continue to advance in the relevant direction, thereby speeding up the convergence rate and reducing oscillations.

[0131] To evaluate the performance of the model on established benchmarks, we used 11 publicly available datasets covering a variety of disease samples for host phenotype prediction tasks.

[0132] NDS and HC dataset: 36 heart failure patients without depressive symptoms (NDS) and 24 healthy controls (HC).

[0133] DS-HC dataset: 35 patients with heart failure (HF) and depressive symptoms (DS) and 24 healthy controls (HC).

[0134] NDS and DS dataset: 35 patients with heart failure (HF) and depressive symptoms (DS) and 36 patients with heart failure but no depressive symptoms (NDS).

[0135] The IBD dataset focuses on inflammatory bowel disease samples.

[0136] The EW-T2D dataset provides type 2 diabetes samples from region A.

[0137] The C-T2D dataset contains type 2 diabetes samples from region B.

[0138] The Obesity dataset is specifically designed for samples related to obesity.

[0139] The Liver Cirrhosis dataset contains samples related to liver cirrhosis.

[0140] Colorectal cancer and early stage colorectal cancer - EMBL datasets provide samples of colorectal cancer.

[0141] The hypertension dataset contains samples of chronic hypertension.

[0142] Among them, NDS to HC, DS to HC, and NDS to DS are a set of 16S rRNA quantities. In order to highlight the importance of microbial taxonomy, the present invention performs multi-level OTU extraction on these three data sets, such as Figure 4This process yielded OTU signatures at different taxonomic levels within the microbial classification system. Eight other public datasets were metagenomic and multimodal, derived from baseline references. These datasets contained markers for the relative abundance of gut microbial species and at the strain level. Figure 4 A detailed overview of all datasets was provided, including sample sizes of affected and control subjects and data sources.

[0143] All traditional machine learning (ML) models and their ensembles were deployed using the default parameters from the scikit-learn library. The performance of each model was evaluated using a 5-fold cross-validation approach. The method was validated using a 5-fold cross-validation scheme with an 8:2 split between training and testing data. The hyperparameters for model training were as follows: stochastic gradient descent (SGD) was used as the optimizer, a learning rate of 0.001, a regularization parameter θ of 0.8, a batch size of 2, 50 training epochs, and a binary classification setting.

[0144] The present invention aims to predict the disease status of an individual based on their microbial profile. The present invention treats this as a binary classification task, where "true" indicates a person with a disease (patient) and "false" indicates a healthy individual (control group). The present invention uses the area under the receiver operating characteristic curve (ROC curve) (AUC) as a key metric for evaluating model performance. Higher AUC values indicate better model performance, and its value ranges from 0.5 for random guessing to 1 for perfect classification. The AUC metric is well suited for handling imbalanced datasets, as seen in the hypertension (class imbalance ratio of 3.9:1) and inflammatory bowel disease (class imbalance ratio of 3.4:1) datasets in the study, see for details. Figure 4 The AUC metric measures the model's ability to distinguish different categories by evaluating its performance at various classification thresholds, thus providing a comprehensive measure of discriminative ability.

[0145] MicroVIS was compared with seven benchmark models covering traditional and state-of-the-art methods for predicting microbe-host phenotypes. These models included standard machine learning techniques such as support vector machines (SVM), logistic regression (LR), random forests (RF), and gradient boosted decision trees (GBDT), as well as advanced deep learning models such as DeepMicro, PopPhy-CNN, and MVIB.

[0146] On 9 of the 11 datasets, the model designed by this invention outperformed all baseline methods, with an average performance of 10.5% higher than the second-best method. Because 16SrRNA datasets rely on taxonomic-level OTUs as input features, while metagenomic multimodal datasets use species relative abundance and strain-level markers, this paper analyzes the results of these two dataset types separately.

[0147] (1) Comparison based on 16S rRNA datasets: We conducted a series of experiments comparing our model with traditional machine learning methods and with the latest microbiome-based disease prediction techniques. The MicroVIS model performed strongly in the 16S rRNA-based microbiome-host phenotype prediction task. Figure 5 The performance comparison of MicroVIS with seven baseline methods on three 16S rRNA datasets is summarized. Figure 6 The boxplots in Figure 3 show the distribution of performance across the ten runs. The results consistently show that MicroVIS outperforms other models on all datasets, with significant performance improvements.

[0148] like Figure 5 As shown, NDS vs. HC: MicroVIS achieved an AUC 14.7% higher than the second-best method, DeepMicro. DS vs. HC: MicroVIS significantly outperformed the second-best model, PopPhy-CNN, with a 17.7% performance improvement. NDS vs. DS: MicroVIS achieved an AUC 15% higher than the second-best method, RF. With the exception of RF, all baseline methods performed poorly, with an average AUC of only 0.559, close to the 0.5 threshold for random guessing. This suggests a general inadequacy in capturing the relationship between the microbiome and the host. This discrepancy highlights the difficulty in distinguishing heart failure patients who present with similar phenotypes, particularly those with and without depressive symptoms.

[0149] Overall, MicroVIS achieved a 15% improvement over other methods, highlighting its strong ability to distinguish patient groups with subtle clinical symptoms. Its outstanding performance on the 16s rRNA dataset demonstrates its adaptability and effectiveness in capturing complex disease features.

[0150] (2) Comparison of metagenomic datasets: After the above experiments confirmed the effectiveness of MicroVIS combined with integrated taxonomic information, the present invention further used eight different metagenomic multimodal datasets to confirm the superiority of the model on other modalities. Figure 7 The performance of MicroVIS is compared with baseline methods on a series of host phenotype prediction tasks. The results show that MicroVIS outperforms various baseline models on 6 out of 8 datasets, reaching the state-of-the-art level.

[0151] On the colorectal cancer dataset, MicroVIS demonstrated significant performance gains, with an AUC value 17.3% higher than the next-best method, PopPhy-CNN. On the cirrhosis dataset, MicroVIS outperformed its closest competitor, MVIB, by 4.5%. Although improvements on the EW-T2D and obesity datasets were less significant, MicroVIS still outperformed the next-best method by 0.4% and 1.8%, respectively.

[0152] On the Early-Colorectal-EMBL dataset, baseline methods achieved an average AUC of 0.539, while on the Hypertension dataset, they achieved a slightly higher score of 0.576. In contrast, MicroVIS significantly outperformed these baseline methods, achieving AUC values of 73.2% and 73.5% on the corresponding datasets, respectively, which are 15% and 7.7% higher than the second-best method. These significant improvements highlight MicroVIS's ability to address these complex prediction challenges. The lower AUC values of existing models indicate the difficulty in capturing microbial changes in these diseases, which may be due to the subtle nature of the biological signals or the limitations of current measurement technologies. Therefore, the Early-Colorectal-EMBL and Hypertension datasets pose particularly challenging tasks for host phenotype prediction.

[0153] These results confirm the effectiveness of MicroVIS and highlight its robust performance in host phenotype prediction tasks. It is worth noting that PopPhy-CNN fails to produce results for the Early-Colorectal-EMBL and Hypertension datasets due to an infinite loop problem when pruning the phylogenetic tree in the original implementation.

[0154] To evaluate the individual contribution of each module to the model performance, the present invention conducted a series of ablation experiments to evaluate the independent contribution of each module to the model performance. The specific modifications are as follows: (1) MicroVIS-IC variant: In the microbial image classification module, the original ResNet-based computer vision method is replaced by a CNN network; (2) MicroVIS-FO variant: The classification-level feature optimization module is removed; (3) MicroVIS-TH variant: All operations in the classification-level feature selection module, classification-level feature optimization module, and classification-level image generation module are simultaneously removed, finally forming this variant model.

[0155] 1) Results of 16s rRNA dataset: Figure 5The experimental results in Figure 2 clearly demonstrate the performance of MicroVIS on the three datasets: NDS vs. HC, DS vs. HC, and NDS vs. DS. All results are the average of five-fold cross validation. For clarity, Figure 6 Visualized. Removing classification-level operations from all modules in MicroVIS-TH resulted in significant performance drops of 26.7%, 24.5%, and 37% on the NDS vs. HC, DS vs. HC, and NDS vs. DS datasets, respectively. Switching MicroVIS-IC's computer vision module from a ResNet-based to a CNN-based architecture resulted in performance drops of 19.3%, 25.5%, and 18.8% on the NDS vs. HC, DS vs. HC, and NDS vs. DS datasets, respectively.

[0156] MicroVIS-FO does not adopt the classification-level feature optimization step but uses all selected features, which leads to a performance drop of 5.8%, 3.2%, and 5% on the NDS vs. HC, DS vs. HC, and NDS vs. DS datasets, respectively.

[0157] MicroVIS achieves the best performance on all datasets without removing any modules, which highlights the importance of the combined effects of these modules.

[0158] Among the tested models, MicroVIS achieved exceptional performance, demonstrating its outstanding performance. This achievement underscores the importance of each module and the rationality and effectiveness of the model architecture. The synergistic effect of its components enables MicroVIS to provide accurate host phenotype predictions, particularly in complex datasets, resulting in impressive classification results. In summary, each module contributes uniquely to the overall model performance, and removing any module results in degraded results.

[0159] 2) Results on metagenomic datasets: We conducted ablation experiments to confirm the impact of each module in MicroVIS on multimodal metagenomic datasets. Figure 8 and Figure 9. Compared with MicroVIS, the MicroVIS-IC variant showed an average performance drop of 20.3% on the eight datasets. In the same datasets, the MicroVIS-FO variant resulted in a relatively small average drop of 6.4%. MicroVIS, which combines these two modules, achieved the best performance in various disease classification tasks. The ablation study results highlight the complementary roles of the FO module (taxonomic hierarchical feature optimization module) and the IC module (microbial image data classification module replaced by ResNet with CNN) in different datasets. The FO module improves performance by reducing feature redundancy in complex high-dimensional data. The IC module excels at identifying subtle feature relationships, especially in multimodal datasets such as metagenomes. The combination of the two significantly improves the classification accuracy of the model.

[0160] Combination of the above Figure 1 The microbial host phenotype prediction method integrating taxonomic levels provided by the embodiment of the present invention is introduced in detail. Next, the microbial host phenotype prediction system integrating taxonomic levels provided by the embodiment of the present invention will be introduced with reference to the accompanying drawings.

[0161] In some embodiments, the present invention provides a method for predicting microbial host phenotypes that integrates taxonomic levels, including: obtaining metagenomic data, extracting an abundance table according to the taxonomic level; based on the abundance table, screening representative features of each level according to set screening criteria; evaluating the importance of the representative features of each level, and selecting representative features of each level with key biomarkers; splicing the representative features of each level with key biomarkers to obtain multi-scale and multi-level features; converting the multi-scale and multi-level features into a two-dimensional grayscale image; and predicting the host phenotype prediction result based on the two-dimensional grayscale image.

[0162] See also Figure 10 The structural diagram of the computer device shown in FIG. 1 includes a processor, a communication interface, and a computer-readable storage medium. The processor, the communication interface, and the computer-readable storage medium can be connected via a bus or other means. The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the memory of the computer device, and the computer-readable storage medium is used to store a computer program. The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer-readable storage medium. The processor (or CPU (Central Processing Unit)) is the computing core and control core of the computer device, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the following steps:

[0163] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0164] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0165] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0166] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0167] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0168] This embodiment provides a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device within a computer device that is used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media within the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device.

[0169] Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by the processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage; optionally, it may be at least one computer-readable storage medium located away from the aforementioned processor.

[0170] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes the one or more instructions stored in the computer-readable storage medium to implement the following steps:

[0171] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0172] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0173] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0174] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0175] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0176] This embodiment provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the following steps:

[0177] Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy;

[0178] Based on the abundance table and the set screening criteria, the representative features of each level are screened;

[0179] Assess the importance of representative features at each level and select representative features of each level with key biomarkers;

[0180] The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image;

[0181] Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

[0182] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0183] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0184] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0185] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0186] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0187] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A microbial host phenotype prediction system integrating taxonomic levels, characterized by: include: The taxonomic level OTU extraction module is used to obtain metagenomic data and extract the abundance table according to the taxonomic level; The taxonomic hierarchical feature selection module is used to select representative features of each level based on the abundance table and the set screening criteria; The taxonomic hierarchical feature optimization module is used to evaluate the importance of representative features at each level and select representative features of each level with key biomarkers; The taxonomic hierarchical image generation module is used to stitch together the representative features of each level with key biomarkers to obtain multi-scale and multi-level features; and convert the multi-scale and multi-level features into a two-dimensional grayscale image; The microbial image classification module is used to predict host phenotype prediction results based on two-dimensional grayscale images.

2. The microbial host phenotype prediction system integrating taxonomic levels according to claim 1, characterized in that: The taxonomic level OTU extraction module is specifically used to: Trim the terminal sequencing reads of the metagenomic data to obtain the initial sequence; Demultiplexing the initial sequence; For the demultiplexed initial sequence, the primer sequence is trimmed to remove the non-biological sequence to obtain a qualified double-ended original sequence; Qualified double-end raw sequences were filtered, denoised, and spliced, and chimeric sequences were removed to obtain all representative sequences; All representative sequences were annotated and abundance tables for each taxonomic level were generated.

3. The microbial host phenotype prediction system integrating taxonomic levels according to claim 1, characterized in that: The screening criteria are: like Then discard feature j Among them, Z j represents the number of samples in which the jth microorganism is zero in all samples Where I(x i,j =0) is the indicator function, if x i,j =0 then I(x i,j =0)=1, otherwise I(x i,j =0) = 0; if the jth feature shows zero value in more than 80% of the samples Then discard it.

4. The microbial host phenotype prediction system integrating taxonomic levels according to claim 1, characterized in that: The taxonomic hierarchical feature optimization module is specifically used to: use the Gini index based on random forest to evaluate feature importance, sort according to feature importance scores, select a certain proportion of top-ranked features, and obtain representative features of each level with key biomarkers.

5. The microbial host phenotype prediction system integrating taxonomic levels according to claim 1, characterized in that: The taxonomic hierarchical image generation module is specifically used to: splice representative features of each level with key biomarkers at different classification levels according to certain rules to obtain multi-scale multi-level features; arrange the multi-scale multi-level features into a square matrix; if the length of the multi-scale multi-level feature is not enough to fill the square matrix, fill the end of the multi-scale multi-level feature with zero values to ensure that a complete square matrix is obtained; normalize the square matrix so that its value range is mapped to the grayscale interval of [0,255], and convert it into a two-dimensional grayscale image.

6. The microbial host phenotype prediction system integrating taxonomic levels according to claim 1, characterized in that: The microbial image classification module is specifically used to obtain host phenotype prediction results based on a two-dimensional grayscale image using a ResNet network; wherein the specific processing process of the ResNet network includes: based on the two-dimensional grayscale image, performing preliminary feature extraction through a 3×3 convolution layer to generate a feature map; inputting the feature map into four levels of residual block stacking in sequence, each level containing num_blocks[l] residual blocks, wherein the feature map size of the lth layer is The number of channels is C l After four levels of residual blocks are stacked, the spatial dimension of the feature map is reduced to 1x1 through the global average pooling layer, and it is mapped to the logits of the number of categories through the fully connected layer; after the logits are softmaxed, the predicted probability distribution of each category is obtained.

7. A method for predicting microbial host phenotypes integrating taxonomic levels, characterized in that: include: Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy; Based on the abundance table and the set screening criteria, the representative features of each level are screened; Assess the importance of representative features at each level and select representative features of each level with key biomarkers; The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image; Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

8. A computer device, characterized in that: a processor adapted to execute a computer program; A computer-readable storage medium having a computer program stored therein, wherein when the computer program is executed by the processor, the following steps are implemented: Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy; Based on the abundance table and the set screening criteria, the representative features of each level are screened; Assess the importance of representative features at each level and select representative features of each level with key biomarkers; The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image; Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

9. A computer-readable storage medium, characterized in that The computer readable storage medium stores a computer program, which is suitable for being loaded by a processor and executed to implement the following steps: Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy; Based on the abundance table and the set screening criteria, the representative features of each level are screened; Assess the importance of representative features at each level and select representative features of each level with key biomarkers; The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image; Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the computer program implements the following steps: Obtain metagenomic data and extract abundance tables based on taxonomic hierarchy; Based on the abundance table and the set screening criteria, the representative features of each level are screened; Assess the importance of representative features at each level and select representative features of each level with key biomarkers; The representative features of each level with key biomarkers are spliced together to obtain multi-scale and multi-level features; the multi-scale and multi-level features are converted into a two-dimensional grayscale image; Based on the two-dimensional grayscale image, the host phenotype prediction results are obtained.

Images