A biomarker mining model training method and device, and related equipment

Abstract

The application discloses a biomarker mining model training method and device and related equipment, comprising: providing a number of biological sample corresponding transcriptome original data and sample phenotype category label, obtaining a gene expression matrix from the transcriptome original data, processing the matrix to obtain a high-dimensional gene expression feature vector, inputting the high-dimensional gene expression feature vector corresponding to each biological sample into a biomarker mining model, the biomarker mining model obtains the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and predicts the sample phenotype category based on the confidence value. The application extracts features from transcriptome data, uses the feature vector and the corresponding sample phenotype category label as training data for training, sets a confidence calculation in the trained model, realizes quantitative evaluation of the importance of gene features, determines the marginal influence of single gene feature change on the prediction result, and thus determines the biomarker.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method, apparatus and related equipment for a biomarker discovery model. Background Technology

[0002] Biomarkers, as molecular characteristics that objectively reflect the physiological state, pathological processes, and drug responses of organisms, are the core cornerstone of modern precision medicine and occupy a crucial position in disease diagnosis and treatment as well as biopharmaceutical research and development. They enable early screening and precise staging of diseases, helping clinicians assess disease progression and treatment effectiveness, while also providing a scientific basis for personalized medication and significantly improving treatment efficacy. In the field of drug development, biomarkers serve as both potential drug targets and efficacy evaluation indicators in clinical trials, effectively shortening the research and development cycle and reducing costs. They play an irreplaceable role in improving disease diagnosis and treatment and promoting the development of the biopharmaceutical industry.

[0003] Traditional biomarker discovery methods rely heavily on manual analysis, resulting in low efficiency, high subjectivity, and difficulty in comprehensively extracting effective molecular features from high-throughput omics data. Furthermore, they lack the ability to capture complex inter-gene relationships. With the rapid development of artificial intelligence, neural network models, leveraging their powerful nonlinear feature learning and high-dimensional data processing capabilities, have been widely applied in various fields such as biomedicine, medical imaging, and clinical diagnosis. Neural network models can automate and refine the analysis of massive amounts of biological data, accurately identifying molecular features closely related to diseases through deep feature learning, significantly improving the efficiency and accuracy of biomarker discovery. Therefore, how to train biomarker discovery models has been a continuous research direction. Summary of the Invention

[0004] In view of this, this application provides a training method, apparatus and related equipment for a biomarker discovery model, so as to facilitate the training of the biomarker discovery model.

[0005] To achieve the above objectives, the following solution is proposed: A training method for a biomarker discovery model includes: Provides raw transcriptome data and phenotypic category labels for several biological samples; The raw transcriptome data corresponding to the biological samples are processed to obtain a gene expression matrix, where each row of the gene expression matrix corresponds to a biological sample and each column corresponds to a gene feature. The gene expression matrix is ​​processed to obtain a high-dimensional gene expression feature vector corresponding to each biological sample; The high-dimensional gene expression feature vectors corresponding to each biological sample are input into the biomarker discovery model to be trained to obtain the predicted sample phenotype category. The biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vectors, and to determine the sample phenotype category based on the confidence value. The parameters of the biomarker discovery model are updated until convergence, with the training objective being that the predicted sample phenotype category approaches the sample phenotype category label corresponding to the high-dimensional gene expression feature vector.

[0006] Optionally, the biomarker discovery model includes: an input layer, a weight allocation layer, and a sample phenotype category prediction layer that are cascaded in sequence; The training process of the biomarker discovery model includes: High-dimensional gene expression feature vectors are obtained through the input layer; Through the weight allocation layer, based on the high-dimensional gene expression feature vector, locally continuous gene expression patterns are extracted, and weights are allocated to each gene expression pattern and each gene locus to obtain a weighted feature tensor. The sample phenotype category prediction layer obtains the confidence value of each sample phenotype category based on the weighted feature tensor, and determines the sample phenotype category based on the confidence value. The parameters of the biomarker discovery model are updated by using the sample phenotype category labels that approximate the high-dimensional gene expression feature vector as the training target.

[0007] Optionally, the weight allocation layer includes: a feature extraction layer, a channel attention weighting layer, and a spatial attention weighting layer; The step involves extracting locally continuous gene expression patterns based on the high-dimensional gene expression feature vector through the weight allocation layer, and assigning weights to each gene expression pattern and each gene locus to obtain a weighted feature tensor, including: Through the feature extraction layer, local continuous gene expression patterns are extracted based on the high-dimensional gene expression feature vector; Through the channel attention weighting layer, global average pooling and global max pooling are performed on each gene expression mode in the gene sequence dimension to obtain global statistical information of each gene expression mode, and weights are assigned to gene expression modes based on the global statistical information. The spatial attention weighting layer performs average and maximum aggregation on the gene expression patterns with assigned weights in the channel dimension to obtain the aggregation result. Based on the aggregation result, weights are assigned to each gene site to obtain the weighted feature tensor.

[0008] Optionally, the feature extraction layer includes: a one-dimensional convolutional layer, a max pooling layer, a non-linear activation layer, a normalization layer, and a random dropout layer; The step of extracting locally continuous gene expression patterns based on the high-dimensional gene expression feature vector through the feature extraction layer includes: Through the one-dimensional convolutional layer, based on the high-dimensional gene expression feature vector, a local continuous preliminary gene expression pattern is extracted; By introducing nonlinear correlations between gene features through a nonlinear activation layer, the preliminary gene expression is transformed nonlinearly. The key gene feature patterns in the gene expression pattern after the nonlinear transformation are preserved through the max pooling layer. The key gene expression patterns are normalized using a normalization layer. By randomly discarding a portion of the normalized gene expression patterns through a dropout layer, the final gene expression pattern is obtained.

[0009] Optionally, the step of processing the raw transcriptome data corresponding to the biological sample to obtain a gene expression matrix includes: The raw transcriptome data were subjected to quality assessment analysis, and based on the assessment analysis results, the raw transcriptome data were subjected to quality control processing. The raw transcriptome data after quality control processing was compared with the pre-set reference transcriptome to obtain transcript expression level data. The expression levels at the transcript level are summarized and transformed to obtain expression levels at the gene level. Based on the gene-level expression data, a gene expression matrix is ​​constructed by selecting gene sets of preset types.

[0010] Optionally, the sample phenotype category label is a numerical label transformed based on a preset mapping relationship.

[0011] Optionally, after updating the parameters of the biomarker discovery model until convergence, the method further includes: Obtain a validation dataset, which includes high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels; The converged biomarker discovery model was validated using the validation dataset to obtain model performance evaluation data. Based on the model performance evaluation data, the evaluation results of the biomarker discovery model are determined; If the current biomarker discovery model outperforms the historical best biomarker discovery model, the current biomarker discovery model is saved as the new historical best biomarker discovery model.

[0012] Optionally, the model performance evaluation data includes: one type of model performance evaluation data or several types of model performance evaluation data; The process of determining the evaluation results of the biomarker discovery model based on the model performance evaluation data includes: The evaluation results of a biomarker discovery model are determined using model performance evaluation data. or, The average value of several model performance evaluation data is used to determine the evaluation results of the biomarker discovery model; or, The evaluation results of the biomarker discovery model are determined by using performance evaluation data from several models and combining them with preset weights.

[0013] A training device for a biomarker discovery model, comprising: The training data acquisition module is used to provide raw transcriptome data and phenotypic category labels for several biological samples. The gene expression matrix acquisition module is used to process the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix. Each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. The feature vector acquisition module is used to process the gene expression matrix to obtain high-dimensional gene expression feature vectors corresponding to each biological sample; The model training module is used to input the high-dimensional gene expression feature vectors corresponding to each biological sample into the biomarker discovery model to be trained, and obtain the predicted sample phenotype category. The biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and to determine the sample phenotype category based on the confidence value. The module updates the parameters of the biomarker discovery model until convergence, with the training objective being that the predicted sample phenotype category approaches the sample phenotype category label corresponding to the high-dimensional gene expression feature vector.

[0014] A training device for a biomarker discovery model includes: a memory and a processor; The memory is used to store programs; The processor is configured to execute the program to implement the various steps of the training method for the biomarker discovery model as described in any of the preceding claims.

[0015] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a training method for a biomarker discovery model as described in any of the preceding claims.

[0016] As can be seen from the above technical solutions, the training method, apparatus, and related equipment for a biomarker discovery model provided in this application embodiment include: using raw transcriptome data and sample phenotype category labels corresponding to several biological samples; processing the raw transcriptome data corresponding to the biological samples to obtain a gene expression matrix, wherein each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature; processing the gene expression matrix to obtain a high-dimensional gene expression feature vector corresponding to each biological sample; inputting the high-dimensional gene expression feature vector corresponding to each biological sample into a biomarker discovery model to be trained to obtain a predicted sample phenotype category; the biomarker discovery model is configured to have the ability to obtain a confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and to determine the sample phenotype category based on the confidence value; and updating the parameters of the biomarker discovery model until convergence, with the predicted sample phenotype category approaching the sample phenotype category label corresponding to the high-dimensional gene expression feature vector as the training objective.

[0017] This application utilizes feature extraction from transcriptome data and trains a biomarker discovery model using high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels. The trained model then uses internal confidence calculations to quantitatively assess the importance of gene features, facilitating the determination of the marginal impact of changes in individual gene features on prediction results, thereby identifying biomarkers. This approach improves the efficiency and accuracy of biomarker discovery while enhancing the interpretability and generalizability of screening results. It provides intelligent solutions for disease diagnosis, prognostic assessment, and personalized medicine, and the model exhibits good functional generalization and versatility, making it applicable to various biological data types and disease scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a training method for a biomarker discovery model provided in this application embodiment; Figure 2 A flowchart for generating a dataset is provided as an embodiment of this application; Figure 3 A schematic diagram of an optional architecture for a biomarker discovery model provided in this application embodiment; Figure 4 A schematic diagram of an optional architecture for the weight allocation layer of a biomarker discovery model provided in this application embodiment; Figure 5 A visual ranking chart of comprehensive contributions is provided as an embodiment of this application; Figure 6 A schematic diagram of a training device for a biomarker discovery model provided in an embodiment of this application; Figure 7 This is a hardware structure block diagram of a training device for a biomarker discovery model provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Figure 1 A flowchart illustrating a training method for a biomarker discovery model provided in this application embodiment is included. The method may include: Step S100: Provide raw transcriptome data and phenotypic category labels for several biological samples.

[0022] Specifically, there is a known correlation between the raw transcriptome data corresponding to the biological sample and the sample phenotypic information corresponding to the biological sample. This known correlation means that the raw transcriptome data and the sample phenotypic information originate from the same biological sample, and the sample phenotypic information has been obtained through clinical diagnosis or pathological testing and can be used to characterize the disease state or stage of the biological sample.

[0023] To construct a high-quality transcriptome dataset for biomarker discovery, raw transcriptome sequencing data and associated phenotypic information for biological samples can be obtained from the Sequence Read Archive (SRA). This database, maintained by internationally recognized bioinformatics institutions, provides table-transformed high-throughput sequencing data and associated biological sample annotation information. The obtained raw transcriptome data originates from biological samples with clear clinical or pathological diagnoses. The phenotypic information characterizes the disease state or stage of the sample; for example, sample phenotypes may include three categories: early fibrosis (F0–2), severe fibrosis (F3), and cirrhosis (F4). The phenotypic information can be converted into numerical labels using predefined mapping relationships, such as mapping F0–2 to category 0, F3 to category 1, and F4 to category 2.

[0024] Step S101: Process the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix.

[0025] Specifically, each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. By processing the raw transcriptome data, errors, inconsistencies, redundancies, or incomplete information can be identified and addressed. Combined with transcriptome data processing standards, the raw transcriptome data undergoes quality control, quantitative analysis, and structured organization, ultimately generating a high-quality, usable gene expression matrix. The gene expression matrix is ​​aligned with the sample phenotypic information according to the sample dimensions, laying a comprehensive and reliable dataset foundation for machine learning tasks.

[0026] Step S102: Process the gene expression matrix to obtain the high-dimensional gene expression feature vector corresponding to each biological sample.

[0027] Specifically, each row in the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. The multidimensional gene features corresponding to a single sample are combined in a preset order to form a high-dimensional numerical vector that can characterize the overall transcriptome features of the biological sample. The numerical vector is then standardized using a standardization function to obtain a high-dimensional feature vector suitable for model input. The standardization function can be expressed as follows:

[0028] in, This represents the original expression level of the j-th gene in the i-th biological sample; This represents the average expression level of the j-th gene in the biological sample being tested; This represents the standard deviation of the expression level of the j-th gene in the biological sample being tested; This represents the standardized expression feature value corresponding to the j-th gene in the i-th biological sample after standardization.

[0029] By using standardized functions, the dimensional differences between different samples can be eliminated, and the samples can be converted into high-dimensional gene expression feature vectors suitable for model input. This ensures that the data structure, parameter dimensions, and information integrity are highly matched with the model's acceptance standards, providing a data foundation for the smooth progress of subsequent biomarker discovery processes and the accuracy of results.

[0030] The parameters of the standardization function are calculated solely based on the test data, including the mean and standard deviation of each gene feature in the test sample. By applying the standardization parameters to each test data set using a consistent feature transformation, the gene expression features in the dataset are mapped to a uniform numerical distribution space, thereby improving the stability of model training and ensuring the accuracy of model generalization performance evaluation.

[0031] Step S103: Input the high-dimensional gene expression feature vector corresponding to each biological sample into the biomarker discovery model to be trained to obtain the predicted sample phenotype category.

[0032] Specifically, the biomarker discovery model is configured to obtain confidence values ​​for each sample phenotypic category based on high-dimensional gene expression feature vectors, and to determine the sample phenotypic category based on these confidence values. The biomarker discovery model uses internal confidence calculations to quantitatively assess the importance of gene features, facilitating the determination of the marginal impact of changes in individual gene features on prediction results, thereby identifying biomarkers. The biomarker discovery model is trained, validated, and optimized. Model training can utilize the AdamW optimizer, executing 500 training epochs. In each epoch, the model first performs forward propagation computation. During model training, the training data can be divided into training, validation, and test datasets. For example, 60% can be used as training data, 20% as validation data, and the remaining 20% ​​as test data. Training the biomarker discovery model based on these divided datasets ultimately yields a biomarker discovery model. This model can learn more general features, rather than merely fitting specific data samples, thereby improving its generalization ability and enabling it to discover potential biomarkers in various scenarios. The process of obtaining training, validation, and test sets using gene expression matrices and corresponding sample phenotypic categories can be found in [reference needed]. Figure 2 As shown, Figure 2 This is a flowchart illustrating a dataset generation process provided in an embodiment of this application.

[0033] The dataset partitioning method can be configured to reduce the impact of sample class distribution differences on model training, validation, and performance evaluation by controlling or constraining the consistency or similarity of sample class distributions in different data subsets. For example, during dataset partitioning, the sample order can be randomly shuffled first to eliminate potential biases that may exist in the original data organization order. The random shuffling process is based on a preset random seed to ensure the repeatability of the experiment. After the sample order is randomly shuffled, the training data is stratified according to the numerical labels corresponding to the samples, so that the distribution ratio of different phenotypic class samples in the training dataset, validation dataset, and test dataset remains consistent or approximately consistent, thereby reducing the impact of sample class distribution differences on model training and performance evaluation results.

[0034] Step S104: Update the parameters of the biomarker discovery model until convergence.

[0035] Specifically, the training objective is to ensure that the predicted sample phenotype category approximates the sample phenotype category label corresponding to the high-dimensional gene expression feature vector. The parameters of the biomarker discovery model are then updated until convergence. For any training batch, the model generates a predicted output based on the input feature vector, calculates the loss value based on the predicted output and the corresponding label, and updates the model parameters using the gradient through backpropagation. This process is performed sequentially on all training batches within a training cycle to obtain the model parameter update results for that training cycle.

[0036] Whether a biomarker discovery model has converged can be determined by calculating the loss function value for the predicted sample phenotype category and sample phenotype category label. The loss function value is obtained by judging whether the loss function value reaches the preset value.

[0037] The converged biomarker discovery model is configured to obtain confidence values ​​for each sample phenotype category based on high-dimensional gene expression feature vectors, and to determine the sample phenotype category based on the confidence values. This allows the trained biomarker discovery model to obtain confidence values ​​for each sample phenotype category, quantify the marginal impact of changes in individual gene features on the prediction results, and thus identify biomarkers.

[0038] After obtaining the confidence values ​​of each sample phenotypic category using the trained biomarker discovery model and quantifying the marginal impact of changes in individual gene features on the prediction results, model interpretation methods can be used to determine the marginal contribution of changes in individual gene features to the prediction results. Using the marginal contribution of changes in individual gene features to the prediction results, the contribution of each gene feature to the sample phenotypic category, such as disease state or disease stage, can be determined. The contribution can then be quantitatively evaluated and ranked to ultimately identify the biomarker.

[0039] The model interpretation methods may include: sensitivity analysis of model output to changes in input features; quantitative analysis of the importance of input features based on model gradient information; evaluation of changes in model prediction results based on feature perturbation or sampling methods; calculation of feature contribution based on model internal parameters or structural information; and estimation of the relationship between input features and output results based on model-independent methods. The model interpretation process may employ at least one model interpretation method, based on all samples, a portion of samples, or a single sample, and may be performed for a single classification category or for multiple classification categories separately. The overall feature importance is then constructed based on the interpretation results for each classification category.

[0040] Deep feature learning enables the precise identification of molecular features closely related to diseases, significantly improving the efficiency and accuracy of biomarker discovery. Furthermore, by leveraging the interpretability analysis of the model, the contribution of each molecular feature can be quantified, making the discovered biomarkers more reliable and clinically valuable. This provides strong technical support for the implementation of precision medicine and propels biomarker discovery into a new stage of intelligence and efficiency.

[0041] By implementing this technical solution, a deep learning model is introduced to automatically extract features from transcriptome data. Combined with a weighted analysis mechanism, the importance of gene features is quantitatively assessed, facilitating the determination of the marginal impact of changes in individual gene features on prediction results. This leads to the identification of biomarkers, improving the efficiency and accuracy of biomarker discovery while enhancing the interpretability and universality of screening results. This not only provides an intelligent solution for disease diagnosis, prognosis assessment, and personalized medicine, but the model also exhibits good functional generalization and universality, making it applicable to various biological data types and disease scenarios.

[0042] The above embodiment provides a training method for a biomarker discovery model, comprising: providing raw transcriptome data and phenotypic category labels corresponding to several biological samples; processing the raw transcriptome data corresponding to the biological samples to obtain a gene expression matrix, wherein each row of the gene expression matrix corresponds to a biological sample and each column corresponds to a gene feature; processing the gene expression matrix to obtain a high-dimensional gene expression feature vector corresponding to each biological sample; inputting the high-dimensional gene expression feature vector corresponding to each biological sample into a biomarker discovery model to be trained to obtain a predicted phenotypic category; the biomarker discovery model is configured to have the ability to obtain a confidence value of each phenotypic category based on the high-dimensional gene expression feature vector and to determine the phenotypic category based on the confidence value; and updating the parameters of the biomarker discovery model until convergence, with the training objective being that the predicted phenotypic category approaches the phenotypic category label corresponding to the high-dimensional gene expression feature vector. This application utilizes feature extraction from transcriptome data and high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels as training data to train a biomarker discovery model. The trained model then employs an internal weighted analysis mechanism and confidence score calculation to quantitatively assess the importance of gene features, facilitating the determination of the marginal impact of changes in individual gene features on prediction results, thereby identifying biomarkers. This approach improves the efficiency and accuracy of biomarker discovery while enhancing the interpretability and generalizability of screening results. It not only provides intelligent solutions for disease diagnosis, prognostic assessment, and personalized medicine, but also demonstrates good functional generalization and versatility, making it applicable to various biological data types and disease scenarios.

[0043] In some embodiments of this application, the biomarker discovery model may include: an input layer, a weight allocation layer, and a sample phenotype category prediction layer that are cascaded in sequence; The training process of the biomarker discovery model may include: S11. Obtain high-dimensional gene expression feature vectors through the input layer.

[0044] Specifically, the high-dimensional gene expression feature vector obtained from the input layer can be a high-dimensional gene expression feature vector that includes the batch size and feature length. Where B represents the batch size and L represents the length of the gene sequence expression feature.

[0045] S12. Through the weight allocation layer, based on the high-dimensional gene expression feature vector, locally continuous gene expression patterns are extracted, and weights are allocated to each gene expression pattern and each gene locus to obtain a weighted feature tensor.

[0046] Specifically, the obtained high-dimensional gene expression feature vectors can be analyzed... Feature mapping is performed by expanding the dimension of the input high-dimensional gene expression feature vector through the Unsqueeze operation. This adds a new dimension to the tensor without changing the original data values. The final output is a tensor containing batch size, 1, and feature length, satisfying the input requirements of a one-dimensional convolutional neural network for extracting locally continuous gene expression patterns. Here, the position of 1 represents the number of feature channels, i.e., the number of gene expression patterns. The feature mapping formula can be:

[0047] S13. Through the sample phenotype category prediction layer, based on the weighted feature tensor, the confidence value of each sample phenotype category is obtained, and the sample phenotype category is determined based on the confidence value.

[0048] Specifically, the feature tensor obtained in the above steps is passed through a feature extraction layer for feature extraction again. The feature tensor is then flattened and mapped to a one-dimensional feature vector h, which is then input into the fully connected prediction layer. The output format is as follows:

[0049] Where W and b represent the weight parameters and bias parameters of the fully connected layer, respectively. This represents the model's prediction score for each sample belonging to different phenotypic categories, where N is the number of phenotypic categories. The phenotypic category prediction layer can contain a feature extraction layer, a flattening layer, a fully connected layer, and a prediction layer. (See reference...) Figure 3 As shown, Figure 3 This is a schematic diagram of an optional architecture for a biomarker discovery model provided in an embodiment of this application, wherein several weight allocation modules can be cascaded in sequence.

[0050] After the sample phenotype category prediction layer obtains a weighted high-order feature tensor containing batch size, number of feature channels, and feature length, it can be further extracted by the feature extraction layer. This high-order feature tensor is then flattened into a one-dimensional feature vector and input into the fully connected prediction layer. Through linear mapping, the feature dimension is mapped to a preset category space. The final output tensor can be a prediction result tensor containing batch size and number of categories, used to represent the model's prediction score for different sample phenotype categories, serving as a confidence value. This enables automatic modeling of high-dimensional features of the transcriptome and disease phenotype discrimination learning.

[0051] Through the above structural design, the biomarker discovery model realizes an end-to-end automated modeling process from high-dimensional gene expression data of the transcriptome to sample phenotype prediction results. It can automatically learn the expression features of key genes that are highly correlated with disease phenotypes without the need for manual feature screening, providing a reliable basis for subsequent biomarker screening and biological interpretation.

[0052] S14. Using the sample phenotype category labels that are close to the high-dimensional gene expression feature vector as the training target, update the parameters of the biomarker discovery model.

[0053] Specifically, during the model training phase, the system takes the sample phenotype category and sample phenotype category label output by the model as input and calculates the classification loss function between them. This loss function is used to measure the difference between the model's prediction result and the sample phenotype category label. The preferred loss function is the cross-entropy loss function, which guides the model to gradually optimize the network parameters during backpropagation and improve its ability to distinguish between different sample phenotype categories.

[0054] Furthermore, the entire forward propagation process returns a tensor of the model's predicted class score for each sample. This tensor contains the batch size and the number of classes, with the value of each dimension representing the prediction confidence level of the corresponding class. Through the alternating execution of forward and backward propagation, the model achieves end-to-end learning from the input transcriptome feature data to the sample phenotype prediction results, forming a complete closed-loop modeling process.

[0055] In some embodiments of this application, the weight allocation layer may include: a feature extraction layer, a channel attention weighting layer, and a spatial attention weighting layer, as shown in the reference. Figure 4 As shown, Figure 4 This is a schematic diagram of an optional architecture for the weight allocation layer of a biomarker discovery model provided in an embodiment of this application.

[0056] Based on this, S12, through the weight allocation layer, based on the high-dimensional gene expression feature vector, extracts locally continuous gene expression patterns, and assigns weights to each gene expression pattern and each gene locus to obtain a weighted feature tensor, which may include: S21. Through the feature extraction layer, based on the high-dimensional gene expression feature vector, extract locally continuous gene expression patterns.

[0057] Specifically, local continuous gene expression patterns are extracted along the gene expression sequence dimension as feature channels, and the feature dimension is progressively reduced while suppressing noise interference. After the feature extraction layer, the model can introduce an attention-weighted layer to enhance the expressive power of key features. The attention-weighted layer can employ a convolutional block attention mechanism and may include channel attention layers and spatial attention layers.

[0058] By performing feature mapping on the high-dimensional gene expression feature vector obtained from the input layer, which includes batch size and gene expression sequence length, we obtain a processed high-dimensional gene expression feature vector containing batch size, number of feature channels, and gene expression sequence length. Where B represents the batch size, C represents the number of feature channels, and L represents the length of the gene expression sequence. The feature extraction layer can extract local continuous gene expression patterns from the processed high-dimensional gene expression feature vector. The specific formula can be:

[0059] in, This represents a one-dimensional convolution operation. This represents a non-linear activation function.

[0060] S22. Through the channel attention weighting layer, global average pooling and global max pooling are performed on each gene expression mode in the gene sequence dimension to obtain global statistical information of each gene expression mode, and weights are assigned to gene expression modes based on the global statistical information.

[0061] Specifically, for the input convolutional feature tensor F, global average pooling and global max pooling operations are performed on the gene expression sequence dimension to obtain global statistical information for each feature channel. The calculation process can be represented as follows:

[0062] in, This represents the global average pooling operator, used to perform average aggregation on the input feature tensor along the gene expression sequence dimension. This represents the global max pooling operator, used to perform maximum aggregation on the input feature tensor along the gene expression sequence dimension.

[0063] Global statistics are input into a nonlinear mapping network to generate channel attention weights:

[0064] in, This represents a mapping function consisting of one-dimensional convolutional layers. This represents the Sigmoid activation function.

[0065] Channel attention weights are used to weight the feature tensor along the channel dimension, and their calculation form can be:

[0066] Here, ⊙ represents element-wise multiplication.

[0067] This step adaptively enhances gene expression patterns that contribute significantly to the phenotypic category of a sample, measuring the relative importance of different feature channels, i.e., different gene expression patterns, in the phenotypic classification task. Specifically, the process of weighting gene expression patterns based on global statistical information involves inputting this information into a non-linear mapping network composed of one-dimensional convolutional layers. First, channel dimension compression is performed, then channel dimension restoration is completed. Finally, a sigmoid activation function is used to generate channel attention weights ranging from 0 to 1, generating weight coefficients for each feature channel and assigning weights to the gene expression patterns.

[0068] S23. Through the spatial attention weighting layer, the gene expression patterns with assigned weights are averaged and maximum aggregated in the channel dimension to obtain the aggregation result. Based on the aggregation result, weights are assigned to each gene site to obtain the weighted feature tensor.

[0069] Specifically, after channel weighting, the model can model the importance of features in the positional dimension using a spatial attention layer. This involves modifying the feature tensor in the channel dimension. By performing average aggregation and maximum aggregation operations separately, and then concatenating them along the channel dimension, a feature representation containing sequence position information can be formed:

[0070] in, This represents the average aggregation operator along the channel dimension, used to calculate the mean of the feature tensor along the channel dimension. This represents the maximum aggregation operator in the channel dimension, used to calculate the maximum value of the feature tensor in the channel dimension.

[0071] Subsequently, the concatenated features are mapped using a one-dimensional convolution operation, and spatial attention weights are generated using a Sigmoid activation function.

[0072] in, This represents the Sigmoid activation function. This represents a one-dimensional convolution operation.

[0073] Spatial attention weights are used to weight the feature tensor along the sequence position dimension, resulting in a spatially enhanced feature representation:

[0074] Used to characterize the contribution of different positions in the gene expression sequence to the discrimination task.

[0075] This step highlights the role of key gene loci in the discrimination of sample phenotypic categories. The resulting weighted feature tensor can be represented as a high-order feature tensor with batch size, number of feature channels, and feature length. This feature tensor preserves global transcriptome information while strengthening the focus on the spatial distribution of gene features in the discrimination line. Specifically, weighting each gene locus based on the aggregation result can include concatenating the aggregation result along the channel dimension to form a feature representation containing spatial location information. This concatenated feature is then mapped through a one-dimensional convolutional layer, and a sigmoid activation function is used to generate spatial attention weights for different positions in the gene expression sequence. Weighting is then applied to each gene locus to obtain a weighted feature tensor, which characterizes the contribution of different positions in the gene expression sequence to the discrimination task.

[0076] The channel attention-weighted layer and spatial attention-weighted layer of the attention-weighted layer adaptively weight key genes at both the gene expression pattern level and the gene expression sequence position level. This allows the model to maintain global transcriptome information while focusing on key gene features that play a decisive role in disease discrimination. This highlights gene expression patterns that are significant for disease-related biomarkers, improving the model's ability to model complex biological data. The model can more accurately identify disease-related key gene expression patterns in the high-dimensional transcriptome feature space, thereby improving overall prediction performance. The channel attention-weighted layer and spatial attention-weighted layer can be executed sequentially according to the example, in reverse order, or in parallel or cascaded combinations.

[0077] Furthermore, the aforementioned feature extraction layer, channel attention weighting layer, and spatial attention weighting layer can be cascaded multiple times in the model to extract more abstract and discriminative higher-order gene expression features layer by layer.

[0078] In some embodiments of this application, the feature extraction layer may include: a one-dimensional convolutional layer, a max-pooling layer, a non-linear activation layer, a normalization layer, and a random dropout layer. The one-dimensional convolutional layer is used to extract local pattern features along the gene expression sequence dimension; the max-pooling layer is used to progressively compress the sequence length and enhance the translation invariance of features; the normalization layer is used to stabilize the feature distribution; and the random dropout layer is used to suppress overfitting and improve the model's generalization ability. As the number of convolutional layers increases, the model gradually completes the abstraction from low-level local expression features to high-level discriminative features.

[0079] Based on this, S21, extracting locally continuous gene expression patterns through the feature extraction layer based on the high-dimensional gene expression feature vector may include: extracting locally continuous preliminary gene expression patterns through the one-dimensional convolutional layer based on the high-dimensional gene expression feature vector; introducing nonlinear correlations between gene features through a nonlinear activation layer to perform nonlinear transformation on the preliminary gene expression patterns; retaining key gene feature patterns in the nonlinearly transformed gene expression patterns through a max pooling layer; normalizing the key gene expression patterns through a normalization layer; and randomly discarding some normalized gene expression patterns through a random discarding layer to obtain the final gene expression pattern.

[0080] In some embodiments of this application, step S101, processing the raw transcriptome data corresponding to the biological sample to obtain a gene expression matrix, may include: S41. Perform quality assessment analysis on the raw transcriptome data, and based on the assessment analysis results, perform quality control processing on the raw transcriptome data.

[0081] Specifically, quality control processing can remove low-quality sequences, sequences with a high proportion of indeterminate bases, and adapter contamination sequences.

[0082] S42. The raw transcriptome data after quality control processing is compared with the preset reference transcriptome to obtain transcript expression data.

[0083] S43. Summarize and transform the expression data at the transcript level to obtain the expression data at the gene level.

[0084] S44. Based on gene expression data, select gene sets of preset types to construct a gene expression matrix.

[0085] Specifically, the gene expression matrix is ​​arranged with biological samples as rows and gene features as columns, with each matrix element representing the expression level of the corresponding gene feature in the corresponding biological sample.

[0086] In some embodiments of this application, after updating the parameters of the biomarker discovery model until convergence, the model can be validated. During the validation phase, based on the model's prediction results for samples in the validation dataset, at least one model performance evaluation index can be calculated. A comprehensive evaluation result of the model is constructed based on the model performance evaluation index, and the comprehensive evaluation result corresponding to the current model is compared with the comprehensive evaluation result of historical models. When the comprehensive evaluation result of the current model is better than the comprehensive evaluation result of the historical best model, the parameter state of the current model is saved as the optimal model. Based on this, after step S104, updating the parameters of the biomarker discovery model until convergence, the following may also be included: S51. Obtain the validation dataset, which includes high-dimensional gene expression feature vectors and corresponding sample phenotype category labels.

[0087] S52. Using the aforementioned validation dataset, validate the converged biomarker discovery model to obtain model performance evaluation data.

[0088] Specifically, the model performance evaluation data can be one type of model performance evaluation data or several types of model performance evaluation data. These data can include: classification accuracy, recall, precision, F1 score, and area under the curve (AUC), etc. Classification accuracy measures the overall correctness of the model's prediction results; the F1 score comprehensively reflects the model's performance in balancing precision and recall; recall assesses the model's overall ability to identify samples from different categories; and precision measures the reliability of the model's prediction results. When the model is used for multi-class tasks, the AUC can be calculated by breaking down the multi-class prediction problem into multiple binary prediction sub-problems.

[0089] S53. Based on the model performance evaluation data, determine the evaluation results of the biomarker discovery model.

[0090] Specifically, model performance evaluation data can be one type of model performance evaluation data or several types of model performance evaluation data. The evaluation result of the biomarker discovery model can be determined using one type of model performance evaluation data; or the evaluation result can be determined by calculating the average value of several types of model performance evaluation data; or the evaluation result can be determined by using several types of model performance evaluation data and combining them with preset weights. The preset weights can be determined based on the number of samples of each phenotypic category in the validation dataset, assigning weights to the model performance evaluation data corresponding to different sample phenotypic categories to reduce the impact of imbalanced sample category distribution on the model performance evaluation results.

[0091] By using the above-mentioned multi-indicator joint evaluation method, the predictive performance and stability of the model in multi-class biological data scenarios can be reflected more comprehensively and objectively.

[0092] To obtain a more stable and comprehensive evaluation of the model's overall performance, the five performance metrics mentioned above are averaged with equal weights to construct a comprehensive evaluation metric for the model. For the k-th training epoch, the corresponding comprehensive evaluation metric is defined as:

[0093] in, , , , and These represent the area under the curve, classification accuracy, F1 score, recall, and precision calculated by the model based on all training samples during the k-th training period, respectively.

[0094] During the model validation phase, the validation dataset is input into the model parameter state corresponding to the current training cycle, and multiple performance evaluation metrics and their corresponding comprehensive evaluation metrics are calculated in the same manner as in the training phase. By monitoring the comprehensive evaluation metrics obtained from the validation dataset in real time, the comprehensive evaluation metrics corresponding to the current training cycle are compared with the best comprehensive evaluation metrics recorded in historical training cycles during model training.

[0095] S54. If the evaluation results of the current biomarker discovery model are better than the historical best biomarker discovery model, save the current biomarker discovery model as the new historical best biomarker discovery model.

[0096] Specifically, when the comprehensive evaluation index corresponding to the current training cycle is better than the historical best comprehensive evaluation index, the parameter state of the current model is immediately saved and marked as the optimal model parameters. Through this method, dynamic optimal model selection based on comprehensive evaluation of multiple performance indicators is achieved during model training, thereby avoiding model selection bias caused by fluctuations in a single performance indicator and improving the model's generalization performance and stability.

[0097] In this embodiment, the target biomarker discovery model is compared with various comparative models to verify its feasibility and superiority in biological phenotype classification tasks. The comparative models include: a one-dimensional convolutional neural network model without an attention mechanism, a classification model built based on traditional machine learning methods, and a model variant using the same network structure but without attention weighting.

[0098] In the comparative experiments, the biomarker discovery model and all comparison models were trained and tested on the same transcriptome dataset constructed in this embodiment to ensure the fairness and comparability of the experimental results. Each model was evaluated under the same data partitioning strategy and training configuration. Evaluation metrics included classification accuracy, F1 score, and other indicators that reflect the model's discriminative performance and generalization ability.

[0099] Compared to baseline models without weighted attention mechanisms and other comparative methods, the biomarker discovery model of this application exhibits higher classification accuracy and F1 score on the test set, while having a lower loss function value. This further verifies the significant technical effect of attention weighting in improving the model's prediction accuracy and stability, indicating that the model can more effectively extract discriminant features highly correlated with disease phenotypes from high-dimensional transcriptome gene expression data, thereby improving the overall performance of disease state or disease stage prediction.

[0100] The biomarker discovery model trained in the above embodiments can be used to discover biomarkers. The following describes the process of determining biomarkers using a biomarker discovery model trained in the above embodiments. The biomarker discovery method may include: S61. Provide raw transcriptome data corresponding to several biological samples.

[0101] Specifically, to construct high-quality transcriptome datasets for biomarker discovery, raw transcriptome sequencing data and associated phenotypic information of biological samples can be obtained from the Sequence Read Archive (SRA). This database, maintained by internationally recognized bioinformatics institutions, provides high-throughput sequencing data transformed from tables and associated biological sample annotation information. The obtained raw transcriptome data all originate from biological samples with clear clinical or pathological diagnoses. The phenotypic information of these samples is used to characterize the disease state or stage of the disease; for example, sample phenotypes can include three categories: early fibrosis (F0–2), severe fibrosis (F3), and cirrhosis (F4).

[0102] S62. Process the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix.

[0103] Specifically, each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. By processing the raw transcriptome data, errors, inconsistencies, redundancies, or incomplete information can be identified and addressed. In conjunction with transcriptome data processing standards, the raw transcriptome data undergoes quality control, quantitative analysis, and structured organization, ultimately generating a high-quality, usable gene expression matrix. The gene expression matrix is ​​aligned with the sample phenotypic information according to the sample dimensions.

[0104] S63. Process the gene expression matrix to obtain the high-dimensional gene expression feature vector corresponding to each biological sample.

[0105] Specifically, each row in the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. The multidimensional gene features corresponding to a single sample are combined in a preset order to form a high-dimensional numerical vector that can characterize the overall transcriptome features of the biological sample. The numerical vector is then standardized using a standardization function to obtain a high-dimensional feature vector suitable for model input. The standardization function can be expressed as follows:

[0106] in, This represents the original expression level of the j-th gene in the i-th biological sample; This represents the average expression level of the j-th gene in the biological sample being tested; This represents the standard deviation of the expression level of the j-th gene in the biological sample being tested; This represents the standardized expression feature value corresponding to the j-th gene in the i-th biological sample after standardization.

[0107] By using standardized functions, the dimensional differences between different samples can be eliminated, and the samples can be converted into high-dimensional gene expression feature vectors suitable for model input. This ensures that the data structure, parameter dimensions, and information integrity are highly matched with the model's acceptance standards, providing a data foundation for the smooth progress of subsequent biomarker discovery processes and the accuracy of results.

[0108] The parameters of the standardization function are calculated solely based on the test data, including the mean and standard deviation of each gene feature in the test sample. By applying the standardization parameters to each test data set using a consistent feature transformation, the gene expression features in the dataset are mapped to a uniform numerical distribution space, thereby improving the stability of model training and ensuring the accuracy of model generalization performance evaluation.

[0109] S64. For each sample phenotype category, based on the biomarker discovery model, using the high-dimensional gene expression feature vector corresponding to each biological sample, while keeping other gene features unchanged, determine the average marginal contribution value of each gene feature to the sample phenotype category predicted by the biomarker discovery model.

[0110] Specifically, the biomarker discovery model is trained using high-dimensional gene expression feature vectors corresponding to biological samples as training samples and phenotypic category labels corresponding to these high-dimensional gene expression feature vectors as training labels. Determining the marginal contribution of each gene feature to the phenotypic category predicted by the biomarker discovery model can be achieved through various methods, such as sensitivity analysis of the model output to changes in input features; quantitative analysis of the importance of input features based on model gradient information; evaluation of changes in model prediction results based on feature perturbation or sampling methods; calculation of feature contribution based on model internal parameters or structural information; and estimation of the relationship between input features and output results using model-independent methods.

[0111] Among them, sample phenotypic information is used to characterize the disease state or stage of a biological sample, and can be obtained through clinical diagnosis or pathological testing. Sample phenotypic information can be converted into numerical labels using a preset mapping relationship, such as mapping F0–2 to category 0, F3 to category 1, and F4 to category 2.

[0112] S65. Based on the average marginal contribution value corresponding to each gene characteristic, determine the biomarkers.

[0113] Specifically, all genes are sorted in descending order according to their average marginal contribution value. Based on the average marginal contribution value corresponding to each gene feature, each gene feature can be sorted. Based on the gene feature sorting, genes with higher contribution values ​​are selected as candidate biomarkers. The sorting method can be descending order, ascending order, or other sorting methods.

[0114] In some embodiments of this application, S62, processing the raw transcriptome data corresponding to the biological sample to obtain a gene expression matrix may include: S71. Perform quality assessment analysis on the raw transcriptome data, and based on the assessment analysis results, perform quality control processing on the raw transcriptome data.

[0115] Specifically, quality control processing can remove low-quality sequences, sequences with a high proportion of indeterminate bases, and adapter contamination sequences.

[0116] S72. The raw transcriptome data after quality control processing is compared with the preset reference transcriptome to obtain transcript expression data.

[0117] S73. The expression data at the transcript level are summarized and transformed to obtain the expression data at the gene level.

[0118] S74. Based on gene-level expression data, select gene sets of preset types to construct a gene expression matrix.

[0119] Specifically, the gene expression matrix is ​​arranged with biological samples as rows and gene features as columns, with each matrix element representing the expression level of the corresponding gene feature in the corresponding biological sample.

[0120] In some embodiments of this application, S64, for each sample phenotype category, based on the biomarker discovery model, using the high-dimensional gene expression feature vector corresponding to each biological sample, and while keeping other gene features unchanged, determining the average marginal contribution value of each gene feature to the sample phenotype category predicted by the biomarker discovery model may include: S81. Based on the biomarker discovery model, using the high-dimensional gene expression feature vectors corresponding to each biological sample, while keeping other gene features unchanged, determine the marginal contribution value of each gene feature to the phenotypic category predicted by the biomarker discovery model for each sample.

[0121] Specifically, the biomarker discovery model can be interpreted and analyzed to determine the marginal contribution of a change in a certain gene feature to the model's prediction of the sample's phenotypic category, while keeping other gene features unchanged. The interpretation method can be the gradient-based SHAP interpretation method.

[0122] S82. For each sample phenotype category, determine the average marginal contribution of each gene feature to the sample phenotype category predicted by the biomarker discovery model.

[0123] Specifically, after determining the marginal contribution value, the average marginal contribution value of each gene feature across all tested samples can be calculated for each sample phenotypic category. This average marginal contribution value serves as the feature importance of that gene feature within the corresponding sample phenotypic category. The formula for calculating the average marginal contribution value is as follows:

[0124] in, Let N represent the marginal contribution value of the i-th test sample in the j-th gene feature dimension corresponding to the c-th category, and N represent the number of test samples.

[0125] In some embodiments of this application, step S104 involves calculating the comprehensive contribution degree for each gene feature based on the average marginal contribution value of the gene feature to each sample phenotypic category. The comprehensive contribution degree can be used to obtain biological features with stable contributions in the overall prediction of multi-classification tasks. For each gene feature, the average marginal contribution value of the gene feature to each sample phenotypic category can be averaged to calculate the comprehensive contribution degree. The specific calculation formula is as follows:

[0126] Where c represents the number of categories corresponding to the sample phenotypic category.

[0127] In some embodiments of this application, S65, determining biomarkers based on the average marginal contribution value corresponding to each gene feature may include: S91. For each gene feature, the comprehensive contribution is calculated based on the average marginal contribution value of the gene feature to each sample phenotypic category.

[0128] Specifically, for each gene feature, the overall contribution can be calculated by using its average marginal contribution value across all sample phenotypic categories, which can eliminate single-class bias to some extent.

[0129] S92. Based on the comprehensive contribution of each gene characteristic, biomarkers are determined.

[0130] Specifically, all genes are sorted in descending order according to their overall contribution. Based on the overall contribution of each gene feature, each gene feature can be sorted. Based on the sorting of gene features, genes with higher contributions are selected as candidate biomarkers. The sorting method can be descending order, ascending order, or other sorting methods. Figure 5 This application provides a visualization chart for ranking comprehensive contributions. Figure 5The overall contribution ranking results are presented graphically, intuitively showing the importance distribution of each gene feature. This helps researchers understand the role of different gene features in the disease development process from the model decision-making level, and provides a reliable basis for subsequent biomarker screening, biological mechanism analysis and clinical research.

[0131] In some embodiments of this application, to improve the stability of the overall contribution, multiple biomarker discovery models can be obtained by repeatedly modeling the training process. This repeated modeling can be achieved by introducing at least one variable training factor. Therefore, before step S92, which determines the biomarker based on the overall contribution corresponding to each gene feature, the following may also be included: Using several different biomarker discovery models, the following steps are repeatedly performed for each sample phenotype category: based on the biomarker discovery model, using the high-dimensional gene expression feature vector corresponding to each biological sample, while keeping other gene features unchanged, to determine the average marginal contribution value of each gene feature to the sample phenotype category predicted by the biomarker discovery model; for each gene feature, based on the average marginal contribution value of the gene feature to each sample phenotype category, to calculate the comprehensive contribution value, resulting in several comprehensive contribution values.

[0132] Specifically, different biomarker discovery models adjust at least one variable training factor during the training process. These variable training factors may include: training data composition method, training data sampling method, training data partitioning method, model initialization parameters, model hyperparameter configuration, and training round configuration.

[0133] In training different biomarker discovery models, variations in the composition of training data can be achieved by adjusting the selection range, proportion, combination, or inclusion relationships of samples in the training dataset to construct different training datasets and thus obtain different biomarker discovery models. Changes in data sampling or partitioning methods can also lead to different biomarker discovery models. For example, the training dataset can be reconstructed by randomly shuffling the sample order or by stratifying the samples based on class labels after random shuffling. Furthermore, the dataset can be partitioned multiple times using K-fold cross-validation. Finally, variations in model initialization parameters can be achieved by sampling different model initialization parameters. Different biomarker discovery models can be obtained by initializing them with numerical values, different random seeds, or different weight initialization strategies. Based on changes in model hyperparameter configuration, the model training process can be configured by adjusting network structure parameters, learning rate parameters, regularization parameters, batch size parameters, or optimizer-related parameters, resulting in different biomarker discovery models. Based on changes in training epoch configuration, the model training process can be controlled by adjusting the number of training epochs, iteration termination conditions, or model parameter update strategies, resulting in different biomarker discovery models. Alternatively, different biomarker discovery models can be trained by combining any one or more of the above methods.

[0134] Several comprehensive contribution values ​​are statistically aggregated to obtain the final comprehensive contribution value. Then, based on the comprehensive contribution value corresponding to each gene feature, biomarkers are determined, which can include: determining biomarkers based on the final comprehensive contribution value corresponding to each gene feature.

[0135] The training apparatus for the biomarker discovery model provided in the embodiments of this application will be described below. The training apparatus for the biomarker discovery model described below can be referred to in correspondence with the training method for the biomarker discovery model described above.

[0136] Figure 6 This application provides a schematic diagram of a training device for a biomarker discovery model, which may include: Training data acquisition module 10 is used to provide raw transcriptome data and sample phenotypic category labels for several biological samples; The gene expression matrix acquisition module 20 is used to process the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix. Each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. The feature vector acquisition module 30 is used to process the gene expression matrix to obtain high-dimensional gene expression feature vectors corresponding to each biological sample; The model training module 40 is used to input the high-dimensional gene expression feature vectors corresponding to each biological sample into the biomarker discovery model to be trained, and obtain the predicted sample phenotype category. The biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and to determine the sample phenotype category based on the confidence value. The model updates the parameters of the biomarker discovery model until convergence, with the training objective being that the predicted sample phenotype category approaches the sample phenotype category label corresponding to the high-dimensional gene expression feature vector.

[0137] The above embodiment provides a training device for a biomarker discovery model, comprising: a training data acquisition module 10, used to acquire raw transcriptome data and sample phenotype category labels corresponding to several biological samples; a gene expression matrix acquisition module 20, used to process the raw transcriptome data corresponding to the biological samples to obtain a gene expression matrix, wherein each row of the gene expression matrix corresponds to a biological sample and each column corresponds to a gene feature; a feature vector acquisition module 30, used to process the gene expression matrix to obtain a high-dimensional gene expression feature vector corresponding to each biological sample; and a model training module 40, used to input the high-dimensional gene expression feature vector corresponding to each biological sample into the biomarker discovery model to be trained to obtain the predicted sample phenotype category, wherein the biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and to determine the sample phenotype category based on the confidence value; and to update the parameters of the biomarker discovery model until convergence, with the predicted sample phenotype category approaching the sample phenotype category label corresponding to the high-dimensional gene expression feature vector as the training objective. This application utilizes feature extraction from transcriptome data and high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels as training data to train a biomarker discovery model. The trained model then employs an internal weighted analysis mechanism and confidence score calculation to quantitatively assess the importance of gene features, facilitating the determination of the marginal impact of changes in individual gene features on prediction results, thereby identifying biomarkers. This approach improves the efficiency and accuracy of biomarker discovery while enhancing the interpretability and generalizability of screening results. It not only provides intelligent solutions for disease diagnosis, prognostic assessment, and personalized medicine, but also demonstrates good functional generalization and versatility, making it applicable to various biological data types and disease scenarios.

[0138] Optionally, the biomarker discovery model may include: an input layer, a weight allocation layer, and a sample phenotype category prediction layer that are cascaded in sequence; The model training module 40 executes the training process of the biomarker discovery model, which may include: High-dimensional gene expression feature vectors are obtained through the input layer; Through the weight allocation layer, based on the high-dimensional gene expression feature vector, locally continuous gene expression patterns are extracted, and weights are allocated to each gene expression pattern and each gene locus to obtain a weighted feature tensor. The sample phenotype category prediction layer obtains the confidence value of each sample phenotype category based on the weighted feature tensor, and determines the sample phenotype category based on the confidence value. The parameters of the biomarker discovery model are updated by using the sample phenotype category labels that approximate the high-dimensional gene expression feature vector as the training target.

[0139] Optionally, the weight allocation layer may include: a feature extraction layer, a channel attention weighting layer, and a spatial attention weighting layer; The model training module executes the process of extracting locally continuous gene expression patterns based on the high-dimensional gene expression feature vector through the weight allocation layer, and assigning weights to each gene expression pattern and each gene locus to obtain a weighted feature tensor. This process may include: Through the feature extraction layer, local continuous gene expression patterns are extracted based on the high-dimensional gene expression feature vector; Through the channel attention weighting layer, global average pooling and global max pooling are performed on each gene expression mode in the gene sequence dimension to obtain global statistical information of each gene expression mode, and weights are assigned to gene expression modes based on the global statistical information. The spatial attention weighting layer performs average and maximum aggregation on the gene expression patterns with assigned weights in the channel dimension to obtain the aggregation result. Based on the aggregation result, weights are assigned to each gene site to obtain the weighted feature tensor.

[0140] Optionally, the feature extraction layer may include: a one-dimensional convolutional layer, a max pooling layer, a non-linear activation layer, a normalization layer, and a random dropout layer; The model training module performs the process of extracting local continuous gene expression patterns based on the high-dimensional gene expression feature vector through the feature extraction layer, which may include: Through the one-dimensional convolutional layer, based on the high-dimensional gene expression feature vector, a local continuous preliminary gene expression pattern is extracted; By introducing nonlinear correlations between gene features through a nonlinear activation layer, the preliminary gene expression is transformed nonlinearly. The key gene feature patterns in the gene expression pattern after the nonlinear transformation are preserved through the max pooling layer. The key gene expression patterns are normalized using a normalization layer. By randomly discarding a portion of the normalized gene expression patterns through a dropout layer, the final gene expression pattern is obtained.

[0141] Optionally, the gene expression matrix acquisition module 20 performs the process of processing the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix, which may include: The raw transcriptome data were subjected to quality assessment analysis, and based on the assessment analysis results, the raw transcriptome data were subjected to quality control processing. The raw transcriptome data after quality control processing was compared with the pre-set reference transcriptome to obtain transcript expression level data. The expression levels at the transcript level are summarized and transformed to obtain expression levels at the gene level. Based on the gene-level expression data, a gene expression matrix is ​​constructed by selecting gene sets of preset types.

[0142] Optional, may also include: The validation set acquisition module is used to acquire the validation dataset, which includes high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels. The evaluation data acquisition module is used to validate the converged biomarker discovery model using the validation dataset to obtain model performance evaluation data. The evaluation result acquisition module is used to determine the evaluation result of the biomarker discovery model based on the model performance evaluation data. The comparison module is used to save the current biomarker discovery model as the new historical best biomarker discovery model if the evaluation result of the current biomarker discovery model is better than that of the historical best biomarker discovery model.

[0143] This application also provides a training device for a biomarker discovery model. Figure 7 The diagram shows the hardware structure of the training device for the biomarker discovery model. (Refer to...) Figure 7 The hardware structure of the training device for the biomarker discovery model may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4. In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4; Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device; The memory stores a program, which the processor can call. The program is used to implement the various processing steps in the training method of the aforementioned biomarker discovery model.

[0144] This application embodiment also provides a storage medium that can store a program suitable for execution by a processor, the program being used to implement various processing flows in the aforementioned training method for biomarker discovery models.

[0145] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0146] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined with each other, and the same or similar parts can be referred to each other.

[0147] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for a biomarker discovery model, characterized in that, include: Provides raw transcriptome data and phenotypic category labels for several biological samples; The raw transcriptome data corresponding to the biological samples are processed to obtain a gene expression matrix, where each row of the gene expression matrix corresponds to a biological sample and each column corresponds to a gene feature. The gene expression matrix is ​​processed to obtain a high-dimensional gene expression feature vector corresponding to each biological sample; The high-dimensional gene expression feature vectors corresponding to each biological sample are input into the biomarker discovery model to be trained to obtain the predicted sample phenotype category. The biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vectors, and to determine the sample phenotype category based on the confidence value. The parameters of the biomarker discovery model are updated until convergence, with the predicted sample phenotype category approaching the sample phenotype category label corresponding to the high-dimensional gene expression feature vector as the training objective.

2. The method according to claim 1, characterized in that, The biomarker discovery model includes: a sequentially cascaded input layer, a weight allocation layer, and a sample phenotype category prediction layer; The training process of the biomarker discovery model includes: The high-dimensional gene expression feature vector is obtained through the input layer; Through the weight allocation layer, based on the high-dimensional gene expression feature vector, locally continuous gene expression patterns are extracted, and weights are allocated to each gene expression pattern and each gene locus to obtain a weighted feature tensor. The sample phenotype category prediction layer obtains the confidence value of each sample phenotype category based on the weighted feature tensor, and determines the sample phenotype category based on the confidence value. The parameters of the biomarker discovery model are updated until convergence, with the determined sample phenotype category approaching the sample phenotype category corresponding to the high-dimensional gene expression feature vector as the training target.

3. The method according to claim 2, characterized in that, The weight allocation layer includes: a feature extraction layer, a channel attention weighting layer, and a spatial attention weighting layer; The step involves extracting locally continuous gene expression patterns based on the high-dimensional gene expression feature vector through the weight allocation layer, and assigning weights to each gene expression pattern and each gene locus to obtain a weighted feature tensor, including: Through the feature extraction layer, local continuous gene expression patterns are extracted based on the high-dimensional gene expression feature vector; Through the channel attention weighting layer, global average pooling and global max pooling are performed on each gene expression mode in the gene sequence dimension to obtain global statistical information of each gene expression mode, and weights are assigned to gene expression modes based on the global statistical information. The spatial attention weighting layer performs average and maximum aggregation on the gene expression patterns with assigned weights in the channel dimension to obtain the aggregation result. Based on the aggregation result, weights are assigned to each gene site to obtain the weighted feature tensor.

4. The method according to claim 1, characterized in that, The feature extraction layer includes: a one-dimensional convolutional layer, a max pooling layer, a non-linear activation layer, a normalization layer, and a random dropout layer; The step of extracting locally continuous gene expression patterns based on the high-dimensional gene expression feature vector through the feature extraction layer includes: Through the one-dimensional convolutional layer, based on the high-dimensional gene expression feature vector, a local continuous preliminary gene expression pattern is extracted; By introducing nonlinear correlations between gene features through a nonlinear activation layer, the initial gene expression pattern is transformed nonlinearly. The key gene feature patterns in the gene expression pattern after the nonlinear transformation are preserved through the max pooling layer. The key gene expression patterns are normalized using a normalization layer. By randomly discarding a portion of the normalized gene expression patterns through a dropout layer, the final gene expression pattern is obtained.

5. The method according to claim 1, characterized in that, The process of processing the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix includes: The raw transcriptome data were subjected to quality assessment analysis, and based on the assessment analysis results, the raw transcriptome data were subjected to quality control processing. The raw transcriptome data after quality control processing was compared with the pre-set reference transcriptome to obtain transcript expression data. The expression levels at the transcript level are summarized and transformed to obtain expression levels at the gene level. Based on the gene-level expression data, a gene set of a preset type is selected to construct a gene expression matrix.

6. The method according to claim 1, characterized in that, The sample phenotype category label is a numerical label converted based on a preset mapping relationship.

7. The method according to claim 1, characterized in that, After updating the parameters of the biomarker discovery model until convergence, the following steps are also included: Obtain a validation dataset, which includes high-dimensional gene expression feature vectors and corresponding sample phenotypic category labels; The converged biomarker discovery model was validated using the validation dataset to obtain model performance evaluation data. Based on the model performance evaluation data, the evaluation results of the biomarker discovery model are determined; If the current biomarker discovery model outperforms the historical best biomarker discovery model, the current biomarker discovery model is saved as the new historical best biomarker discovery model.

8. The method according to claim 7, characterized in that, The model performance evaluation data includes: one type of model performance evaluation data or several types of model performance evaluation data; The process of determining the evaluation results of the biomarker discovery model based on the model performance evaluation data includes: The evaluation results of a biomarker discovery model are determined using model performance evaluation data. or, The average value of several model performance evaluation data is used to determine the evaluation results of the biomarker discovery model; or, The evaluation results of the biomarker discovery model are determined by using performance evaluation data from several models and combining them with preset weights.

9. A training device for a biomarker discovery model, characterized in that, include: The training data acquisition module is used to provide raw transcriptome data and phenotypic category labels for several biological samples. The gene expression matrix acquisition module is used to process the raw transcriptome data corresponding to the biological sample to obtain the gene expression matrix. Each row of the gene expression matrix corresponds to a biological sample, and each column corresponds to a gene feature. The feature vector acquisition module is used to process the gene expression matrix to obtain high-dimensional gene expression feature vectors corresponding to each biological sample; The model training module is used to input the high-dimensional gene expression feature vectors corresponding to each biological sample into the biomarker discovery model to be trained, and obtain the predicted sample phenotype category. The biomarker discovery model is configured to have the ability to obtain the confidence value of each sample phenotype category based on the high-dimensional gene expression feature vector, and to determine the sample phenotype category based on the confidence value. The module updates the parameters of the biomarker discovery model until convergence, with the training objective being that the predicted sample phenotype category approaches the sample phenotype category label corresponding to the high-dimensional gene expression feature vector.

10. A training device for a biomarker discovery model, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement the various steps of the training method for the biomarker discovery model as described in any one of claims 1-8.

11. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the various steps of the training method for the biomarker discovery model as described in any one of claims 1-8.

Images