Method for predicting response of non-small cell lung cancer treatment by using multi-omics technology

Abstract

The present application relates to the field of medical data processing and biological information analysis, and discloses a prediction method for detecting the treatment responsiveness of non-small cell lung cancer by using multi-omics technology, comprising: an implicit layer mapping module, which converts the protein metabolism feature matrix and the gene expression feature matrix with dimensional differences into an implicit layer feature matrix; a feature aggregation module, which calls public biological pathway parameters to process the implicit layer feature matrix and outputs deep features; a cross-fusion module, which fuses the deep protein metabolism features and the deep transcriptome features to generate a bidirectional cross-feature matrix; and a prediction output module, which non-linearly maps the bidirectional cross-feature matrix to obtain a response probability matrix and generates a treatment response prediction report. By constructing a multi-layer graph convolution network and a cross-omics attention architecture, the present application eliminates the dimensional difference of the original feature space of multi-omics, converts the biological network topological information into deep features, and completes the cross-modal interaction, thereby overcoming the one-sided limitations caused by single-dimensional classification.

Description

Technical Field

[0001] This invention relates to the field of medical data processing and bioinformatics analysis, specifically a method for predicting the treatment response of non-small cell lung cancer using multi-omics technology. Background Technology

[0002] Treatment planning for non-small cell lung cancer (NSCLC) relies on the early prediction of treatment responsiveness in NSCLC samples. Researchers typically combine transcriptome data from high-throughput sequencing and proteome data from mass spectrometry for tumor data analysis. However, current technologies for processing transcriptome and proteome data suffer from dimensional differences in the original feature spaces of these two omics datasets, leading to inconsistencies in their initial dimensions.

[0003] Existing detection models lack the ability to incorporate gene interaction patterns recorded in public biological pathway databases when extracting multi-omics features. They fail to convert biological pathway relationships into deep features and lack steps for performing inter-layer transfer operations and addition and fusion operations in a graph network environment, resulting in a serious loss of biological network topology information.

[0004] Existing prediction systems lack the support of cross-omics attention network architectures, making it difficult to conduct multi-dimensional cross-modal interactions between deep protein metabolism features and deep transcriptome features. This results in clinical predictions for non-small cell lung cancer (NSCLC) relying solely on a single biological dimension for model classification. Furthermore, single-dimensional predictions cannot comprehensively combine cell population proportion parameters to determine the calculated incidence rate of corresponding NSCLC samples. These shortcomings lead to a severe lack of classification criteria in the output of treatment response prediction reports by existing methods, resulting in significant limitations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a predictive method for detecting the treatment response of non-small cell lung cancer using multi-omics technology. This method solves the problem that existing technologies rely on a single biological dimension feature for prediction when detecting the treatment response of non-small cell lung cancer, lack cross-modal interaction analysis between multi-omics data, and result in single classification criteria and predictive indicators.

[0006] To achieve the above objectives, this invention provides a method for predicting the treatment response of non-small cell lung cancer using multi-omics technology, comprising the following steps:

[0007] The data preprocessing module formats the protein metabolomics data and transcriptomics data into protein metabolism feature matrices and gene expression feature matrices, and the hidden layer mapping module calls a non-linear dimensionality reduction function to convert the protein metabolism feature matrices and gene expression feature matrices with initial dimensionality differences into hidden layer feature matrices with uniform dimensionality.

[0008] The feature aggregation module calls parameters from the public biological pathway database to process the hidden layer feature matrix and converts the hidden layer feature matrix into deep protein metabolism features and deep transcriptome features;

[0009] The cross-fusion module uses deep protein metabolism features and deep transcriptome features as query matrices, key matrices and value matrices respectively, and performs dot product and normalization operations to generate a bidirectional cross-feature matrix.

[0010] The prediction output module uses preset fully connected network parameters to perform nonlinear mapping calculation on the bidirectional cross feature matrix to obtain the response probability matrix. The prediction output module combines the cell population ratio parameters extracted from the transcriptome data to generate a treatment response prediction report containing the classification results and immune response scores of all sample numbers.

[0011] Single-cell RNA sequencing was performed on non-small cell lung cancer (NSCLC) samples using a high-throughput sequencer to output raw transcriptome data. Proteomics and metabolomics analysis were performed on the NSCLC samples using a mass spectrometer to output raw proteome and metabolome data. The data preprocessing module has a first set of cleaning instructions for the raw proteome and metabolome data and a second set of cleaning instructions for the raw transcriptome data. The raw proteome and metabolome data are processed using the first set of cleaning instructions, and the raw transcriptome data is processed using the second set of cleaning instructions. After processing, the data preprocessing module extracts high-confidence protein and metabolite quantification information, integrates and summarizes it to form proteome and metabolome data, and outputs standardized gene expression levels, which are then integrated and summarized to form transcriptome data.

[0012] The hidden layer mapping module's storage space is pre-written with a learnable weight matrix and a bias vector for protein metabolism, and a learnable weight matrix and a bias vector for transcriptomics. The protein metabolism feature matrix and the learnable weight matrix are multiplied together, and the protein metabolism bias vector is then added. Similarly, the gene expression feature matrix and the learnable weight matrix are multiplied together, and the transcriptomics bias vector is added. The matrix values ​​output from the multiplication and addition operations are input into a nonlinear activation function layer to generate a hidden layer feature matrix containing both the protein metabolism and transcriptomics hidden layer feature matrices.

[0013] The feature aggregation module reads gene interaction files recorded in the public biological pathway database and constructs a priori graph network model. It traverses all graph network nodes, verifies the data association status within interaction entries, and establishes edges. Based on all graph network nodes and all edges, a node adjacency matrix is ​​generated. An identity matrix is ​​added to the node adjacency matrix to define a self-loop adjacency matrix. The feature aggregation module performs a numerical summation operation on each row within the self-loop adjacency matrix, arranging the results on the main diagonal of a new square matrix of the same dimension, defining it as a degree matrix.

[0014] The feature aggregation module internally contains a multi-layer graph convolutional network environment, which includes multiple graph convolutional nodes arranged hierarchically. The feature aggregation module pushes the hidden layer feature matrix into the layer 0 node of the multi-layer graph convolutional network environment, and performs inter-layer transfer operations on the input hidden layer feature matrix by combining the self-loop adjacency matrix and the degree matrix. The feature values ​​carried within adjacent graph network nodes are added and fused along the topological connectivity path. The feature aggregation module then renames the hidden layer feature matrices detached from the multi-layer graph convolutional network environment as deep protein metabolism features and deep transcriptome features, respectively.

[0015] The cross-fusion module internally employs a cross-omics attention network architecture comprising query feature processing nodes, key feature processing nodes, and value feature processing nodes. It generates a query matrix by performing matrix multiplication on deep protein metabolism features and query weight parameters, and generates a key matrix by performing matrix multiplication on deep transcriptome features and key weight parameters. The key matrix is ​​then transposed to generate a transposed key matrix. A dot product is then performed between the transposed key matrix and the query matrix. The cross-fusion module introduces a scaling factor into the dot product result to perform numerical scaling, yielding the transcriptome-protein metabolism attention scoring matrix.

[0016] The cross-fusion module calls the exponential normalization function, inputting the transcriptome-protein metabolism attention scoring matrix into the function for nonlinear numerical transformation calculation, outputting a normalized attention matrix. It then performs matrix multiplication on deep transcriptome features and value weight parameters to generate a transcriptome value matrix. Finally, it performs matrix multiplication between the normalized attention matrix and the transcriptome value matrix to output transcriptome-protein metabolism fusion features. Following parallel topological paths and computational instructions within the cross-omics attention network architecture, the cross-fusion module processes deep transcriptome features and deep protein metabolism features, outputting protein metabolism-transcriptome fusion features. The merging processing node calls the matrix merging calculation function to perform same-dimensional concatenation, outputting a bidirectional cross-feature matrix.

[0017] The hardware structure of the prediction output module internally deploys a multi-layer neural network architecture, with fully connected network layers at the ends of the multi-layer neural network architecture. The prediction output module controls the bidirectional cross-feature matrix and the output weight matrix to trigger matrix multiplication operations and superimpose the output bias vector to generate a fused scoring result. The fused scoring result is input into the Sigmoid activation function to perform a monotonic nonlinear mapping, generating an immune response scoring matrix, i.e., a response probability matrix.

[0018] The prediction output module retrieves a preset judgment threshold and performs a row-by-row comparison of the response probability matrix to determine the relationship between the probability value and the preset judgment threshold. When the difference between the probability value and the preset judgment threshold exceeds a preset confidence interval, and the probability value is greater than or equal to the preset judgment threshold, it is classified into the responder category array; when the difference between the probability value and the preset judgment threshold exceeds a preset confidence interval, and the probability value is less than the preset judgment threshold, it is classified into the no-responder category array; when the difference between the probability value and the preset judgment threshold is within a preset confidence interval, a second verification calculation is triggered and it is marked as a suspected state.

[0019] For CD8 recorded within the responder classification array and in the cell population proportion parameter + The proportion of MAIT cells is greater than or equal to the first preset proportion threshold or CXCR6. + CD8 + For non-small cell lung cancer samples with a MAIT cell ratio greater than or equal to a second preset threshold, a preferred strategy label is written into the treatment response prediction report, and they are marked as predicted beneficiaries. For non-small cell lung cancer samples recorded within the non-responder classification array, an alternative strategy label is written into the treatment response prediction report. After the treatment response prediction report is generated, the prediction output module packages the generated treatment response prediction report and sends it to the display terminal on the system's periphery.

[0020] This invention provides a predictive method for detecting treatment response in non-small cell lung cancer using multi-omics technology. It offers the following advantages:

[0021] 1. This invention assigns a learnable weight matrix for protein metabolism and a learnable weight matrix for transcriptome to the protein metabolism feature matrix and gene expression feature matrix through a hidden layer mapping module. After a dot product operation, the protein metabolism bias vector and the transcriptome bias vector are superimposed respectively and sent to a nonlinear activation function layer to perform matrix mapping, generating a hidden layer feature matrix of uniform dimension. This eliminates the dimensional differences in the original feature spaces of multiple omics and solves the defect of inconsistent initial dimensions between transcriptome data and protein metabolome data.

[0022] 2. This invention constructs a priori graph network model by reading gene interaction files recorded in a public biological pathway database through a feature aggregation module. Based on the graph network nodes and edges, a self-loop adjacency matrix and a degree matrix are generated. The hidden layer feature matrix is ​​input into a multi-layer graph convolutional network environment to perform inter-layer transfer operations, which causes the feature values ​​carried inside adjacent graph network nodes to be added and fused along the topological connection path, thus converting the biological pathway association into deep protein metabolism features and deep transcriptome features.

[0023] 3. This invention utilizes a cross-omics attention network architecture loaded within the cross-fusion module to convert deep protein metabolism features and deep transcriptome features into query matrices and key matrices. Dot product and normalization operations are performed to generate a bidirectional cross-feature matrix. This matrix is ​​then used in conjunction with a fully connected network layer deployed within the prediction output module to perform nonlinear mapping to obtain a response probability matrix. The incidence rate is calculated by combining cell population ratio parameters. Based on a preset judgment threshold, a treatment response prediction report containing classification results and preferred strategy labels is output, overcoming the limitations of using only a single biological dimension for classification. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0025] Figure 2 This is a schematic diagram of the prediction system architecture and data flow path of the present invention;

[0026] Figure 3 This is a comparison chart of the receiver operating characteristic (ROC) curves of the present invention;

[0027] Figure 4 This is a two-dimensional distribution diagram of the immune response score and key cell proportions of the present invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See attached document Figure 1 and attached Figure 2 This invention provides a predictive method for detecting the treatment response of non-small cell lung cancer using multi-omics technology, which relies on a predictive system for execution;

[0030] The prediction system includes data acquisition equipment and a central server, with the data acquisition equipment communicating with the central server. The data acquisition equipment includes a high-throughput sequencer and a mass spectrometer, which analyze samples from non-small cell lung cancer patients and output transcriptome data and proteome data, respectively.

[0031] The central server includes a communication interface, a central processing unit (CPU), and memory. The CPU is electrically connected to both the communication interface and the memory. The communication interface receives proteome and transcriptome data from the data acquisition equipment and transmits it to the CPU. The memory contains a set of computer instructions and a public biological pathway database including the Kyoto Encyclopedia of Genes and Genomes.

[0032] The central processing unit reads the set of computer instructions from the memory, controls the data preprocessing module to call the protein metabolomics data and transcriptomics data for calculation, and extracts features to construct the protein metabolism feature matrix and gene expression feature matrix.

[0033] The central server integrates a graphics processing unit (GPU), which is connected to the central processing unit (CPU). The CPU transmits protein metabolism feature matrices, gene expression feature matrices, and parameters from the public biological pathway database to the GPU. The GPU is internally divided into multiple computational nodes, which read the protein metabolism feature matrices and gene expression feature matrices in parallel and perform dimensionality reduction mapping and cross-attention fusion scoring operations in parallel.

[0034] The graphics processor returns the fused joint omics feature matrix to the central processing unit; the central processing unit reads the response probability to the treatment plan returned by the prediction output module and controls the communication interface to send the response probability to the display terminal.

[0035] See attached document Figure 2 The central processing unit and the graphics processing unit are divided into different data flow paths; multiple computing flow nodes are respectively set as data preprocessing module, hidden layer mapping module, feature aggregation module, cross-fusion module and prediction output module.

[0036] The communication interface routes the received protein metabolomics and transcriptomics data to the data preprocessing module; the data preprocessing module formats the protein metabolomics and transcriptomics data into protein metabolism feature matrices and gene expression feature matrices, and then transmits the protein metabolism feature matrices and gene expression feature matrices into the hidden layer mapping module.

[0037] The hidden layer mapping module is equipped with a non-linear dimensionality reduction function. The non-linear dimensionality reduction function is called to convert the protein metabolism feature matrix and gene expression feature matrix with initial dimensionality differences into a hidden layer feature matrix with uniform dimensionality, and then push the hidden layer feature matrix to the feature aggregation module.

[0038] The feature aggregation module has access to synchronously read parameters from the public biological pathway database in memory. The feature aggregation module calls the public biological pathway database parameters to process the hidden layer feature matrix and converts the hidden layer feature matrix into deep protein metabolism features and deep transcriptome features.

[0039] Deep protein metabolism features and deep transcriptome features are input into the cross-fusion module. The cross-fusion module provides a cross-attention computing environment, where deep protein metabolism features and deep transcriptome features undergo mutual weight matrix multiplication. The multidimensional deep protein metabolism features and deep transcriptome features are integrated into a parallel bidirectional cross-feature matrix.

[0040] The prediction output module receives the bidirectional cross-feature matrix output by the cross-fusion module. The prediction output module contains fully connected network parameters, which perform numerical mapping calculations on the bidirectional cross-feature matrix. The calculated non-small cell lung cancer response probability is output by the prediction output module.

[0041] See attached document Figure 1 The present invention provides a method for predicting the treatment response of non-small cell lung cancer using multi-omics technology, comprising the following steps:

[0042] S1. The data acquisition device generates multi-omics raw data covering single-cell transcriptomics, proteomics, and metabolomics features and pushes it to the data preprocessing module. The data preprocessing module performs batch correction, low-quality cell removal, and standardization on the multi-omics raw data, and combines it with CD8 extracted from the transcriptomics data. + The algorithm identifies the proportion of immune cells such as MAIT cells, generating a protein metabolism feature matrix containing pathway activity and metabolic characteristics, and a gene expression feature matrix containing cell population proportion characteristics. These two matrices are then sent to the hidden layer mapping module. The hidden layer mapping module uses a nonlinear dimensionality reduction function to compress and align the feature dimensions of the protein metabolism and gene expression feature matrices, and then sends the dimension-unified multi-omics hidden layer feature matrix to the feature aggregation module. This provides a standardized input basis for predicting the responsiveness of non-small cell lung cancer to immunotherapy combined with chemotherapy.

[0043] S2. The feature aggregation module receives the hidden layer feature matrix, retrieves the public biological pathway database parameters from the memory, distributes the hidden layer feature matrix to different nodes of the multi-layer graph convolutional network, and performs aggregation and update operations with the public biological pathway database parameters. The feature aggregation module has a preset multi-layer graph convolutional network environment, which contains multiple graph convolutional network nodes arranged in a hierarchical manner. The feature aggregation module attaches topological correlation features to the hidden layer feature matrix through the multi-layer graph convolutional network environment, which evolves into deep protein metabolism features and deep transcriptome features.

[0044] S3. The feature aggregation module imports deep protein metabolism features and deep transcriptome features into the cross-fusion module in parallel. The cross-fusion module uses deep protein metabolism features and deep transcriptome features as query matrices and key-value matrices, performs dot product and normalization operations to generate a bidirectional cross-feature matrix, presenting a serial state of protein metabolism layer and transcription layer dimensions.

[0045] S4. The cross-fusion module imports the bidirectional cross-feature matrix into the prediction output module. The prediction output module uses preset fully connected network parameters to perform nonlinear mapping calculation on the bidirectional cross-feature matrix to obtain the response probability matrix, and transmits the response probability value to the communication interface.

[0046] See attached document Figure 1 and attached Figure 2 The data acquisition device generates multi-omics raw data and initiates multi-omics data preprocessing and initial feature mapping nodes. In this embodiment, to ensure the consistency of the pathological state of the multi-omics data over time, the sample source for non-small cell lung cancer patients is limited to tumor tissue biopsy samples and paired peripheral blood mononuclear cell samples from patients with advanced driver gene-negative non-small cell lung cancer who are receiving first-line treatment with immune checkpoint inhibitors combined with chemotherapy. As a preferred approach, the time interval between the collection of tumor tissue biopsy samples and peripheral blood mononuclear cell samples is strictly limited to within 24 hours to eliminate metabolite fluctuation interference caused by sampling time differences. The collected tumor tissue biopsy samples and paired peripheral blood mononuclear cell samples are placed inside the data acquisition device.

[0047] Data acquisition equipment includes a high-throughput sequencer and a mass spectrometer. The high-throughput sequencer performs single-cell RNA sequencing on the samples to establish a baseline immune cell atlas and extracts CD8+. + MAIT cells and their CXCR6 + The system analyzes immune cell atlas information, including subpopulation percentages, and outputs raw transcriptome data. Mass spectrometry performs proteomics and metabolomics analyses on the samples, describing immunometabolic characteristics and outputting raw proteome and metabolome data. The raw transcriptome and raw proteome and metabolome data are collectively referred to as multi-omics raw data.

[0048] Raw proteome and transcriptome data are sent from the data acquisition device to the data preprocessing module. The data preprocessing module has a pre-defined set of cleaning instructions for the raw proteome data, including a baseline denoising algorithm and a missing value imputation algorithm. The raw proteome data undergoes baseline denoising to remove background noise, and then missing value imputation to fill in undetected low-abundance signals. The processed feature files are then used to extract high-confidence protein and metabolite quantification information, which is then integrated and summarized to form the proteome data.

[0049] The data preprocessing module has an independently configured set of second sequence cleaning instructions for the raw transcriptome data. It performs adapter sequencing fragment stripping and low-quality base removal on the raw transcriptome data, mapping the cleaned raw transcriptome data to a reference transcriptome sequence. A counter within the data preprocessing module counts the total number of fragment reads mapped to individual gene blocks and sends this count to the built-in sequence normalization unit. The sequence normalization unit calculates the number of transcription reads per thousand bases per million mapped reads and outputs standardized gene expression levels. These standardized gene expression levels are then integrated to form the transcriptome data. The proteome and metabolome data are in a state of pending extraction within the data preprocessing module.

[0050] The data preprocessing module reads the protein metabolomics data to be extracted and performs two-dimensional numerical permutation and recombination on the protein metabolomics data. The data preprocessing module integrates a multi-stage feature filter. The multi-stage feature filter uses statistical analysis to screen candidate feature variables that differ between historical response and non-response groups. It calls the LASSO regression algorithm to sparsify the candidate feature variables and retain variables with non-zero coefficients. It calls the Boruta algorithm based on random forest to evaluate the importance of each non-zero coefficient variable. It uses SHAP value analysis to remove redundant variables with contributions below a preset threshold and extracts a high-confidence feature set.

[0051] The data preprocessing module sets the number of samples participating in sequencing as the row dimension parameter and the number of features retained in the high-confidence feature set as the column dimension parameter. The processed values ​​are sequentially filled into the first data array defined by the row and column dimension parameters, filling row by row according to the sample entry number order, forming a two-dimensional real-valued protein metabolism feature matrix where each row represents the characteristic state of a single non-small cell lung cancer sample and each column represents the distribution of a single feature. After completion, this matrix is ​​cached in an independent memory block, and the transcriptome data in the pending extraction state is retrieved. The data preprocessing module counts the number of non-small cell lung cancer samples contained within the transcriptome data and simultaneously extracts the number of independent gene blocks contained within the transcriptome data.

[0052] The data preprocessing module determines the number of samples and the number of independent gene blocks as the row and column dimensions of the second data array, respectively. Then, it writes the standard gene expression values ​​row by row according to the sample number, filling in a two-dimensional real-valued gene expression feature matrix in which rows represent single non-small cell lung cancer samples and columns represent single gene blocks. The data preprocessing module internally constructs the gene expression feature matrix loaded with continuous floating-point values ​​and transfers it to another independent memory block, so that the protein metabolism feature matrix and the gene expression feature matrix are in a state of waiting for the hidden layer mapping module to initiate a read request.

[0053] Within the hidden layer mapping module, a read instruction is generated for the data preprocessing module. The protein metabolism feature matrix and gene expression feature matrix, which are in a waiting state, are read into the computation buffer of the hidden layer mapping module. The protein metabolism feature matrix is ​​filled with discrete binary values, while the gene expression feature matrix is ​​filled with continuous floating-point values. Furthermore, there is a misalignment in the column dimension parameters of the protein metabolism feature matrix and the gene expression feature matrix.

[0054] To eliminate the differences in data scale and feature dimension between the protein metabolism feature matrix and the gene expression feature matrix, the hidden layer mapping module calls a nonlinear dimensionality reduction function to compress and unify the matrix dimensions. The hidden layer mapping module's storage space is pre-written with a learnable weight matrix and a bias vector for protein metabolism, and simultaneously with a learnable weight matrix and a bias vector for transcriptomics. The protein metabolism feature matrix and the learnable weight matrix are multiplied together, and the protein metabolism bias vector is then added. Similarly, the gene expression feature matrix and the learnable weight matrix are multiplied together, and the transcriptomics bias vector is added. The matrix values ​​output after the multiplication and addition operations are then fed into the nonlinear activation function layer. The calculation of the hidden layer mapping features follows the following mathematical expression:

[0055] ;

[0056] ;

[0057] in: This represents the completed protein metabolism hidden layer feature matrix; This represents the calculated hidden layer feature matrix of the transcriptome; This represents a non-linear activation function, and in this scheme, the non-linear activation function is limited to the ReLU function; Represents the protein metabolism feature matrix; This represents a learnable weight matrix representing protein metabolism. Represents the protein metabolism bias vector; Represents the gene expression characteristic matrix; This represents the learnable weight matrix of the transcriptome; This represents the transcriptome bias vector.

[0058] The learnable weight matrix for protein metabolism defines the transformation relationship from the initial feature dimension of protein metabolism to the target hidden layer dimension, while the learnable weight matrix for transcriptome defines the transformation relationship from the initial feature dimension of transcriptome to the same target hidden layer dimension. The protein metabolism bias vector corresponds to the numerical offset of the transformed protein metabolism feature matrix, and the transcriptome bias vector corresponds to the numerical offset of the transformed gene expression feature matrix.

[0059] The hidden layer mapping module uses a nonlinear dimensionality reduction function to unify the protein metabolism feature matrix and gene expression feature matrix, which have differences in initial dimension and numerical range, to the same feature dimension. The protein metabolism hidden layer feature matrix and the transcriptome hidden layer feature matrix have completely equal numerical values ​​in the column dimension and are jointly defined as the hidden layer feature matrix. After generation, the hidden layer feature matrix is ​​pushed into the data receiving port of the feature aggregation module.

[0060] The feature aggregation module receives the protein metabolism hidden layer feature matrix and the transcriptome hidden layer feature matrix, sends a data read request to the central server's storage; reads gene interaction files recorded in the public biological pathway database and constructs a prior graph network model; the prior graph network model is internally divided into multiple discrete graph network nodes; extracts the feature dimension labels contained in the protein metabolism hidden layer feature matrix and the transcriptome hidden layer feature matrix, maps the feature dimension labels one-to-one to the discrete graph network nodes, and counts the total number of graph network nodes contained in the prior graph network model, which is set as the node count.

[0061] The feature aggregation module traverses all graph network nodes, retrieves interaction entries recorded in the gene interaction file, and verifies the data association status of any two graph network nodes within the interaction entries. If the first and second graph network nodes have an interaction record within the interaction entry, the feature aggregation module establishes an edge between them. If the first and second graph network nodes lack an interaction record within the interaction entry, the feature aggregation module keeps them disconnected.

[0062] The feature aggregation module generates a node adjacency matrix based on all graph network nodes and all edges. The node adjacency matrix is ​​a square matrix of equal dimension recording the set of real numbers, with both row and column parameters equal to the number of nodes. The feature aggregation module performs numerical filling operations on the row and column intersection coordinates within the node adjacency matrix. When there is an edge between two graph network nodes corresponding to the row and column intersection coordinates, the feature aggregation module fills in the value 1 at the intersection coordinate. When the two graph network nodes corresponding to the row and column intersection coordinates are disconnected, the feature aggregation module fills in the value 0 at the intersection coordinate. The filled node adjacency matrix is ​​stored in the computational memory area of ​​the feature aggregation module, and the node adjacency matrix carries topological connectivity information at the gene and transcript dimensions.

[0063] The feature aggregation module reads the node adjacency matrix stored in the computation memory area, adds identity matrix values ​​to the node adjacency matrix, and sets all values ​​on the main diagonal of the node adjacency matrix to 1. The node adjacency matrix after adding identity matrix values ​​is defined as a self-loop adjacency matrix. The self-loop adjacency matrix is ​​traversed and counted in both rows and columns, and the values ​​in each row of the self-loop adjacency matrix are summed separately. The summed values ​​are arranged sequentially on the main diagonal of a new equal-dimensional square matrix, and the non-main diagonal positions in the new equal-dimensional square matrix are filled with the value 0. The new equal-dimensional square matrix filled with values ​​is defined as the degree matrix, which reflects the total number of edges that a single graph network node has in the graph structure.

[0064] The feature aggregation module internally contains a multi-layer graph convolutional network environment, which includes multiple graph convolutional nodes arranged hierarchically. The data input is set to the layer 0 node, and the protein metabolism hidden feature matrix within the hidden layer feature matrix is ​​first pushed into the layer 0 node of the multi-layer graph convolutional network. The multi-layer graph convolutional network performs inter-layer transfer operations on the input protein metabolism hidden feature matrix using the self-loop adjacency matrix and the degree matrix. The inter-layer update transformation of the feature matrix within the multi-layer graph convolutional network follows the following mathematical expression:

[0065] ;

[0066] in: Indicates the first Layer feature matrix; Represents a nonlinear activation function; Degree matrix; Represents a self-loop adjacency matrix; Indicates the first Layer feature matrix; Indicates the first The learnable weight matrix of the layer.

[0067] The first one defined in the formula The learnable weight matrix of a layer internally stores the neuron parameters set for the current layer of a multi-layer graph convolutional network. The learnable weight matrix of a layer performs a column-dimensional numerical mapping on the input features. The self-loop adjacency matrix provides the topological connectivity paths between nodes in the graph network. The degree matrix performs symmetric normalization on the self-loop adjacency matrix, and the feature values ​​carried within adjacent graph network nodes are added and fused along the topological connectivity paths. The learnable weight matrix of the layer synchronously adjusts the scale of the fused feature values, and the topological features of the graph network nodes are extracted and recorded in the first layer. Within the layer feature matrix.

[0068] The protein metabolism hidden layer feature matrix penetrates layer by layer in the multi-layer graph convolutional network (MLN) to reach the output node of the MLM. The feature aggregation module renames the protein metabolism hidden layer feature matrix that has exited the MLM to a deep protein metabolism feature. The feature aggregation module synchronously schedules the transcriptome hidden layer feature matrix within the hidden layer feature matrix, and the transcriptome hidden layer feature matrix is ​​pushed into the layer 0 node of the same MLM. The transcriptome hidden layer feature matrix reaches the output node after undergoing identical inter-layer graph convolutional operations, and the transcriptome hidden layer feature matrix exits the MLM and is renamed a deep transcriptome feature. The feature aggregation module sends a data-ready communication signal to the cross-fusion module, and the deep protein metabolism feature and deep transcriptome feature are pushed into the input port of the cross-fusion module by the feature aggregation module.

[0069] The cross-fusion module receives transmitted deep protein metabolism features and deep transcriptome features; internally, it has a cross-omics attention network architecture that includes query feature processing nodes, key feature processing nodes, and value feature processing nodes; based on the attention allocation requirements from transcriptome to protein metabolism, it establishes a mapping connection configuration between deep protein metabolism features and deep transcriptome features within the cross-omics attention network architecture; it imports deep protein metabolism features into the query feature processing node, and simultaneously imports deep transcriptome features into the key feature processing node and the value feature processing node.

[0070] The query feature processing node has query weight parameters set internally. Deep protein metabolism features are multiplied with these query weight parameters to generate a query matrix of independent dimensions. The key feature processing node has key weight parameters set internally. Deep transcriptome features are multiplied with these key weight parameters to generate a key matrix of independent dimensions.

[0071] Before performing cross-omics attention interaction, a scaling constraint rule is introduced based on the general rule that dot product operations tend to amplify the variance of high-dimensional vectors. The cross-fusion module extracts the key matrix and performs a matrix transpose operation on it. The transposed key matrix is ​​then subjected to a dot product operation with the query matrix. A scaling factor is introduced into the dot product result, and numerical scaling is performed on the result based on this scaling factor. The transcriptome-to-protein metabolism attention scoring matrix is ​​obtained during the scaling operation. The transcriptome-to-protein metabolism attention scoring matrix follows the following mathematical expression:

[0072] ;

[0073] in: This represents the transcriptome's attention score matrix for protein metabolism; Indicates the characteristics of deep protein metabolism; Indicates the query weight parameter; Indicates deep transcriptome characteristics; Indicates the key weight parameter; Indicates the matrix transpose identifier; This represents the scaling factor.

[0074] The query weight parameter is set as the set of neural network parameters that maps deep protein metabolism features from the original feature dimension to the target vector dimension. The key weight parameter is set as the set of neural network parameters that maps deep transcriptome features from the original feature dimension to the same target vector dimension. Deep protein metabolism features and deep transcriptome features are then subjected to matrix multiplication using the query weight parameter and the key weight parameter, bringing them into a unified data comparison space.

[0075] During calculation, the result is divided by the square root of the target vector dimension as a scaling factor. The cross-fusion module uses scaling to limit the numerical range of the dot product output, preventing the dot product result from exceeding the floating-point processing limit or causing gradient vanishing in subsequent derivative calculations. The transcriptome-protein metabolism attention scoring matrix contains correlation coefficients of the projection from deep transcriptome features to deep protein metabolism features. This matrix is ​​temporarily stored in the cache register of the cross-fusion module awaiting subsequent aggregation operations.

[0076] The cross-fusion module reads the transcriptome-protein metabolism attention scoring matrix temporarily stored in the cache register. Within the cross-omics attention network architecture, it calls the exponential normalization function, inputting the transcriptome-protein metabolism attention scoring matrix into the exponential normalization function. The exponential normalization function performs a row-wise nonlinear numerical transformation calculation on the transcriptome-protein metabolism attention scoring matrix. The discrete correlation coefficients recorded within the transcriptome-protein metabolism attention scoring matrix are mapped to continuous floating-point numbers distributed within a closed interval of 0 to 1. After the nonlinear numerical transformation calculation, the sum of the continuous floating-point numbers in each row of the transcriptome-protein metabolism attention scoring matrix is ​​strictly equal to the value 1. The exponential normalization function outputs a normalized attention matrix, which contains standardized probability distribution weight parameters. The processing objective of the exponential normalization function is defined as eliminating absolute numerical differences caused by different feature dimensions and generating a normalized relative weight distribution.

[0077] The cross-fusion module synchronously reads deep transcriptome features located within the value feature processing node. Value feature processing nodes have pre-defined value weight parameters. The deep transcriptome features and value weight parameters undergo matrix multiplication to generate a transcriptome value matrix. The cross-fusion module controls the matrix multiplication operation between the normalized attention matrix and the transcriptome value matrix. In this matrix multiplication, the probability distribution weight parameters loaded within the normalized attention matrix are multiplied by the feature values ​​carried within the transcriptome value matrix. The multiplication operation between the normalized attention matrix and the transcriptome value matrix outputs transcriptome-protein metabolism fusion features.

[0078] The cross-fusion module processes deep transcriptome features and deep protein metabolism features according to completely parallel network topology paths and computational instructions. Based on the attention allocation requirements of protein metabolism to the transcriptome, it generates a protein metabolism-transcriptome attention scoring matrix and performs a nonlinear numerical transformation calculation on this matrix using an exponential normalization function. Deep protein metabolism features undergo matrix-dot multiplication with value weight parameters within their corresponding value feature processing nodes to generate a protein metabolism value matrix. The cross-fusion module multiplies the transformed protein metabolism-transcriptome attention scoring matrix with the protein metabolism value matrix, outputting the fused protein metabolism-transcriptome features.

[0079] The calculation of transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features follows the following mathematical expression:

[0080] ;

[0081] ;

[0082] in: This indicates the transcriptome's fusion characteristics with protein metabolism; This indicates the role of protein metabolism in transcriptome fusion features; Represents the exponential normalization function; This represents the transcriptome's attention score matrix for protein metabolism; This represents the attention scoring matrix of protein metabolism on the transcriptome; Indicates deep transcriptome characteristics; Indicates the characteristics of deep protein metabolism; The value weight parameter represents the weight.

[0083] The value weight parameter is set as a numerical matrix defining the projection relationship between deep transcriptome features and deep protein metabolism features onto a unified target value space. By multiplying the deep transcriptome features and deep protein metabolism features by the value weight parameter matrix, the internal values ​​of the original omics features are rescaled and rearranged. The rescaled and rearranged transcriptome value matrix and protein metabolism value matrix each receive probability distribution weight parameters passed from the normalized attention matrix. Larger probability distribution weight parameters amplify and preserve highly correlated cross-omics feature values ​​within the fusion feature, while smaller probability distribution weight parameters reduce and suppress less correlated cross-omics feature values ​​within the fusion feature. The cross-fusion module internally includes a merge processing node; the cross-fusion module generates transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features and pushes them to the merge processing node.

[0084] The merging processing node receives transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features, and loads them into a pre-defined contiguous memory address space. Both sets of fusion features have identical row and column dimension values; the row dimension value is equal to the number of non-small cell lung cancer samples participating in sequencing, and the column dimension value is equal to the target vector dimension set by the cross-omics attention network architecture.

[0085] The merge processing node calls the matrix merging calculation function to perform same-dimensional splicing of transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features. It then performs horizontal arrangement and combination of transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features in the column dimension, amplifies the data dimensions, outputs the joint omics feature matrix, and keeps the row sequence numbering of transcriptome-protein metabolism fusion features and protein metabolism-transcriptome fusion features unchanged.

[0086] The merging processing node uses the sample number as the alignment index and horizontally splices the protein metabolism pair transcriptome feature numerical sequence corresponding to the same sample number to the end of the transcriptome pair protein metabolism feature numerical sequence. It traverses all sample numbers to complete the splicing task of bidirectional fusion features and generates a complete joint omics feature matrix.

[0087] The merging processing node performs a horizontal concatenation operation along the column dimension, ensuring that the row dimension values ​​of the joint omics feature matrix are equal to the row dimension values ​​of the transcriptome-protein metabolism fusion features, thus preserving all non-small cell lung cancer samples on their respective row axes. The column dimension values ​​of the joint omics feature matrix generate numerical superposition and amplification; the column dimension values ​​of the joint omics feature matrix are strictly equal to the sum of the column dimension values ​​of the transcriptome-protein metabolism fusion features and the protein metabolism-transcriptome fusion features. In terms of specific data volume, the column dimension values ​​of the joint omics feature matrix are equal to twice the target vector dimension.

[0088] The joint omics feature matrix internally contains integrated cross-omics feature parameters calculated using bidirectional interactive attention weighting. A hardware data transmission path connects the merging processing node to the system's underlying prediction output module. Once the merging processing node determines that the joint omics feature matrix construction is complete, it pushes the entire joint omics feature matrix into the input port of the prediction output module via the hardware data transmission path. The joint omics feature matrix then awaits feature classification signal extraction from the fully connected layer.

[0089] The prediction output module extracts the joint omics feature matrix stored within the input port. The hardware structure of the prediction output module internally deploys a multi-layer neural network architecture, with fully connected network layers at the ends. Each fully connected network layer has both input and output neurons. The number of input neurons in the fully connected network layer is strictly equal to the column dimension of the joint omics feature matrix. The number of output neurons in the fully connected network layer is limited to 1, and each output neuron is responsible for outputting a single-dimensional classification scalar. The fully connected network layer is pre-loaded with the output weight matrix and output bias vector that constitute the fully connected network parameters. The output weight matrix maps the joint omics feature matrix to the prediction space.

[0090] The prediction output module controls the input of the joint omics feature matrix into the fully connected network layer. The joint omics feature matrix and the output weight matrix trigger matrix multiplication and are superimposed with the output bias vector to generate a fused score. This fused score is then input into the Sigmoid activation function for monotonic nonlinear mapping. The nonlinear mapping operation forcibly compresses the values ​​into a closed interval between 0 and 1, generating an immune response score (IR-score) matrix, i.e., a response probability matrix. The prediction output module internally calculates the response probability matrix using the following mathematical expression:

[0091] ;

[0092] in: Represents the response probability matrix; This represents the Sigmoid activation function; Represents the joint omics feature matrix; This represents the output weight matrix; This represents the output bias vector.

[0093] The response probability matrix is ​​a two-dimensional matrix with one column dimension, and the row dimension corresponds to the number of all non-small cell lung cancer (NSCLC) samples participating in the sequencing. Each individual probability value within the response probability matrix represents the calculated incidence of a positive response to a treatment in a single NSCLC sample. A preset judgment threshold is pre-written into the built-in storage unit of the prediction output module. This preset judgment threshold is a fixed floating-point value, determined using a Youden exponent maximization optimization method based on a large amount of historical clinical response data. The value of the preset judgment threshold is limited to a closed interval between 0.48 and 0.53. The prediction output module retrieves the preset judgment threshold and performs a row-by-row comparison operation on the response probability matrix.

[0094] The prediction output module extracts the first row of probability values ​​corresponding to the first sample number and determines the relationship between the first row of probability values ​​and the preset judgment threshold. To prevent misjudgments due to minor fluctuations caused by a single absolute boundary, the system introduces a tolerance judgment interval. When the first row of probability values ​​is greater than or equal to the preset judgment threshold, the non-small cell lung cancer sample corresponding to the first sample number is classified into the responder classification array; when the first row of probability values ​​is less than the preset judgment threshold and the deviation exceeds the confidence interval, the non-small cell lung cancer sample corresponding to the first sample number is classified into the non-responder classification array; when the difference between the first row of probability values ​​and the preset judgment threshold is within the ±0.05 confidence interval, a second review calculation is triggered and the sample is marked as suspected, and the probability values ​​corresponding to all non-small cell lung cancer samples are processed one by one according to the same comparison instruction.

[0095] After all classification and alignment operations are completed, the prediction output module, combined with the cell population proportion parameters extracted from the transcriptome data, generates a treatment response prediction report containing the classification results corresponding to all sample numbers and the immune response score (IR-score). The prediction output module retrieves transcriptome data and statistically analyzes the CD8+ levels in each non-small cell lung cancer sample. + MAIT cells account for a significant portion of total CD8+ cells. + The proportion of T cells and CD8 + CXCR6 in MAIT cells + The proportion of subgroups. For those recorded within the responder classification array and CD8 + MAIT cell proportion greater than or equal to 15.1% or CXCR6 + CD8 + For non-small cell lung cancer samples with a MAIT cell ratio of ≥35.9%, the prediction output module will write a preferred strategy label for providing first-line chemotherapy combined with immune checkpoint inhibitor therapy to the corresponding non-small cell lung cancer sample in the treatment response prediction report, and mark it as a predicted beneficiary population.

[0096] For non-small cell lung cancer samples recorded within the non-responder classification array, the prediction output module includes a pause in providing the current immune checkpoint inhibitor combined with chemotherapy regimens, and recommends alternative strategies such as targeted therapy or other chemotherapy drug combinations, within the treatment response prediction report. This is achieved by fusing the probability scores from the multimodal cross-omics network output with key cell subpopulations (CD8+) in the immune atlas. + The MAIT ratio improves the robustness and reliability of predicting the efficacy of immunotherapy.

[0097] Once the assessment report is generated, the prediction output module packages the generated treatment response prediction report and sends it to the external display terminal of the system, where it is displayed in the clinical decision support system. At the same time, the treatment response prediction report is backed up to the large-capacity storage hard drive connected to the prediction output module.

[0098] Specific application examples:

[0099] Based on clinical data and multi-omics feature extraction processes from patients with advanced driver gene-negative non-small cell lung cancer, the constants and initial parameters set in the method of this invention are as follows:

[0100] The scaling factor used to apply scaling constraints to the output features of the cross-omics attention network architecture is set to 64 (corresponding to the square of the target vector dimension); the preset decision threshold determined by the Youden exponent maximization optimization method is set to 0.50; the confidence interval to prevent fluctuations in a single absolute boundary is set to ±0.05; and the CD8 alpha of the extracted biological validation indicators is used. + The minimum percentage of MAIT cells is strictly set at 15.1%.

[0101] During the multi-omics feature unification and fusion computation performed by the central processing unit (CPU), the hidden layer mapping module performs dimensionality reduction and scale alignment on the initial input matrix. At a certain feature extraction time, the sum of the discrete protein metabolism feature node numerical sequences extracted by the computational flow node for a single patient sample is 12.5, and the sum of the corresponding continuous gene expression feature node numerical sequences is 8.2. The pre-written protein metabolism learnable weight matrix in the hidden layer mapping module's storage space has a mapping multiplier of 0.8 at the current node, and the extracted protein metabolism bias vector value is 0.2. The CPU substitutes the values ​​into the hidden layer mapping feature calculation formula to obtain the unified mapping features for the current node:

[0102] ;

[0103] The central processing unit uses the calculated value of 10.2 as the aligned single-node hidden layer value and appends it to the protein metabolism hidden layer feature matrix, thereby eliminating the differences in initial data scale and dimensions caused by different sequencing platforms and mass spectrometry equipment.

[0104] When the method actively performs cross-attention calculations to capture the interaction between deep transcriptome features and deep protein metabolism features, the cross-fusion module dynamically quantifies the correlation between the two feature spaces. Based on the attention allocation requirements from the transcriptome to protein metabolism, the arithmetic unit extracts the dot product of the current sample's key feature processing node and the query feature processing node, which yields a result of 24.8. The central processing unit then performs numerical scaling on the core terms of the transcriptome-protein metabolism attention scoring matrix formula.

[0105] ;

[0106] The calculated scaling correlation degree of 3.1 is then input into the exponential normalization function, which transforms it into a continuous floating-point number in the closed interval [0,1] as a probability distribution weight parameter, ensuring that the gradient vanishing phenomenon will not occur due to dimensionality expansion during subsequent fusion calculations.

[0107] As the transcriptome-protein metabolism fusion features and the protein metabolism-transcriptome fusion features are horizontally concatenated along the column dimension at the merging processing node, the joint omics feature matrix is ​​fully pushed into the prediction output module. Inside the fully connected network layer, for a specific non-small cell lung cancer sample number, the sum of the product of its joint omics feature matrix and output weight matrix is ​​0.35, and the extracted output bias vector is 0.12. The prediction output module substitutes the values ​​into the formula to perform nonlinear mapping calculations to generate the corresponding single probability values ​​in the response probability matrix:

[0108] ;

[0109] Since the calculated probability value of 0.615 is greater than the preset judgment threshold of 0.50 and exceeds the upper limit of the ±0.05 confidence interval, the prediction output module directly classifies this non-small cell lung cancer sample into the responder classification array. Simultaneously, the transcriptome data of this sample is retrieved, and its CD8+ is statistically analyzed. + MAIT cells account for a significant portion of total CD8+ cells. + The proportion of T cells was 18.2%. Since 18.2% is greater than the biological threshold of 15.1%, the method's final judgment condition was triggered by two factors. The prediction output module added a preferred strategy label for providing the corresponding patient with first-line chemotherapy combined with immune checkpoint inhibitor therapy to the treatment response prediction report for this non-small cell lung cancer sample, and the prediction process ended.

[0110] Actual operation tests were conducted and data were compared. Specific verification results are detailed in the appendix. Figure 3 and attached Figure 4 The experiment compared the traditional clinical single-omics (transcriptome-only) prediction method with the method of this invention in a clinical validation cohort of 120 patients with advanced driver gene-negative non-small cell lung cancer in the same batch. The experimental data are shown in Table 1.

[0111] Table 1. Comparison of operational data between traditional single-omics prediction methods and the method of this invention

[0112]

[0113] See attached document Figure 3 And Table 1, Appendix Figure 3 This figure compares receiver operating characteristic (ROC) curves for predicting the efficacy of immunotherapy in non-small cell lung cancer in a validation cohort. The horizontal axis represents the false positive rate, dimensionless, ranging from 0 to 1. The vertical axis represents the true positive rate, dimensionless, ranging from 0 to 1. The figure includes ROC curves from both the conventional method and the method of this invention.

[0114] During the climb of the false positive rate from 0 to 0.4, traditional methods, relying on only a single feature source and failing to incorporate parameters from a public biological pathway database for graph network correction via a feature aggregation module, exhibit poor ability to capture weak response signals, resulting in a slow curve climb and ultimately an AUC area under the curve of 0.72. The method of this invention calculates a bidirectional cross-feature matrix through a cross-fusion module and performs lateral splicing operations using a merging processing node, deeply mining the nonlinear interaction features at the transcriptional and metabolic levels.

[0115] The ROC curve of the method of this invention shows an extremely steep upward trend, achieving a high true positive rate in an extremely low false positive range. The area under the AUC curve eventually reaches 0.91, and the overall prediction accuracy increases from 74.2% to 92.6%, completely eliminating the feature bias and misjudgment phenomenon that are very easy to be generated by single-omics models.

[0116] See attached document Figure 4 And Table 1, Appendix Figure 4 This diagram displays the two-dimensional distribution of immune response scores and key cell proportions in non-small cell lung cancer samples during the final prediction output. The horizontal axis represents the immune response score, which is dimensionless and ranges from 0.2 to 0.9. The vertical axis represents CD8+. + MAIT cell percentage, expressed as a percentage (%), ranging from 0 to 30. The graph includes scatter plots of real clinical response samples, scatter plots of real clinical non-responder samples, preset judgment thresholds, and key cell percentage boundaries.

[0117] Traditional methods tend to produce a large number of confusing classifications when the immune response score falls within the critical grayscale range of 0.4 to 0.6. The method of this invention introduces a two-dimensional hard constraint into the prediction output module, constructing a highly discriminative distribution envelope in two-dimensional space. (See attached...) Figure 4 As can be seen, all real clinical non-responder sample scatter points are strictly compressed to the left of the preset judgment threshold boundary line (0.50 on the horizontal axis) and below the key cell proportion boundary line (15.1% on the vertical axis); while real clinical responder samples are highly enriched in the upper right region.

[0118] The method of this invention uses a dual-judgment boundary to cut off the outbound path of false positive samples, reducing the false positive rate from 18.5% to 4.3%. It demonstrates strong robustness and reliability in screening and predicting the beneficiaries, providing a highly confident decision-making basis for adjusting clinical treatment plans.

Claims

1. A predictive method for detecting treatment response in non-small cell lung cancer using multi-omics technology, characterized in that, Includes the following steps: The data preprocessing module formats the protein metabolomics data and transcriptomics data into protein metabolism feature matrices and gene expression feature matrices, and the hidden layer mapping module calls a non-linear dimensionality reduction function to convert the protein metabolism feature matrices and gene expression feature matrices with initial dimensionality differences into hidden layer feature matrices with uniform dimensionality. The feature aggregation module calls parameters from the public biological pathway database to process the hidden layer feature matrix and converts the hidden layer feature matrix into deep protein metabolism features and deep transcriptome features; The cross-fusion module uses the deep protein metabolism features and the deep transcriptome features as query matrices, key matrices and value matrices respectively, and performs dot product and normalization operations to generate a bidirectional cross-feature matrix. The prediction output module uses preset fully connected network parameters to perform nonlinear mapping calculation on the bidirectional cross feature matrix to obtain the response probability matrix. The prediction output module combines the cell population ratio parameters extracted from the transcriptome data to generate a treatment response prediction report containing the classification results and immune response scores of all sample numbers.

2. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 1, characterized in that, The data preprocessing module formats the protein metabolomics data and transcriptomics data into protein metabolism feature matrices and gene expression feature matrices, including the following steps: Single-cell RNA sequencing was performed on non-small cell lung cancer samples using a high-throughput sequencer to output raw transcriptome data. Proteomics and metabolomics detection were performed on the non-small cell lung cancer samples using a mass spectrometer to output raw proteome and metabolome data. The raw proteome data is processed using a set of first cleaning instructions, and the raw transcriptome data is processed using a set of second sequence cleaning instructions. After processing, high-confidence protein and metabolite quantitative information is extracted, integrated, and summarized to form the protein metabolomics data. Standard gene expression values ​​are output, integrated, and summarized to form the transcriptome data.

3. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 1, characterized in that, The hidden layer mapping module calls a non-linear dimensionality reduction function to convert the protein metabolism feature matrix and the gene expression feature matrix, which have initial dimensionality differences, into a hidden layer feature matrix of uniform dimension, including the following steps: The learnable weight matrix of protein metabolism and the bias vector of protein metabolism, as well as the learnable weight matrix of transcriptome and the bias vector of transcriptome, are pre-written into the storage space of the hidden layer mapping module. The protein metabolism feature matrix and the protein metabolism learnable weight matrix are multiplied by a dot product and then the protein metabolism bias vector is added. The gene expression feature matrix and the transcriptome learnable weight matrix are multiplied by a dot product and then the transcriptome bias vector is added. The matrix values ​​output by the dot product operation and the superposition operation are input into the nonlinear activation function layer to generate the hidden layer feature matrix containing the protein metabolism hidden layer feature matrix and the transcriptome hidden layer feature matrix.

4. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 1, characterized in that, The feature aggregation module calls parameters from the public biological pathway database to process the hidden layer feature matrix, including the following steps: Read the gene interaction files recorded in the public biological pathway database and construct a priori graph network model. Traverse all graph network nodes to check the data association status within the interaction entries and establish connections. A node adjacency matrix is ​​generated based on all the nodes and edges of the graph network. An identity matrix is ​​added inside the node adjacency matrix to define a self-loop adjacency matrix. The numerical summation operation is performed on each row inside the self-loop adjacency matrix, and the result is arranged on the main diagonal of a new equal-dimensional square matrix, which is defined as the degree matrix.

5. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 4, characterized in that, The process of converting the hidden layer feature matrix into deep protein metabolism features and deep transcriptome features includes the following steps: The feature aggregation module contains a multi-layer graph convolutional network environment with multiple graph convolutional nodes arranged hierarchically. The hidden layer feature matrix is ​​pushed into the 0th layer node of the multi-layer graph convolutional network environment, and inter-layer transfer operation is performed on the input hidden layer feature matrix in combination with the self-loop adjacency matrix and the degree matrix. The feature values ​​carried within adjacent graph network nodes are added and merged along the topological connectivity path, and the hidden layer feature matrices that are removed from the multi-layer graph convolutional network environment are renamed as the deep protein metabolism features and the deep transcriptome features, respectively.

6. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 1, characterized in that, The process of generating a bidirectional cross feature matrix by performing dot product and normalization operations includes the following steps: The cross-fusion module contains a cross-omics attention network architecture that includes query feature processing nodes, key feature processing nodes, and value feature processing nodes. The deep protein metabolism features and query weight parameters are multiplied by a matrix to generate a query matrix, and the deep transcriptome features and bond weight parameters are multiplied by the matrix to generate a bond matrix. The key matrix is ​​transposed to generate a transposed key matrix. The transposed key matrix is ​​then multiplied by the query matrix. A scaling factor is introduced into the result of the dot product operation to perform numerical scaling and obtain the transcriptome attention score matrix for protein metabolism.

7. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 6, characterized in that, It also includes the following steps: The exponential normalization function is invoked, and the transcriptome attention score matrix for protein metabolism is input into the exponential normalization function to perform nonlinear numerical transformation calculations, and the normalized attention matrix is ​​output. The deep transcriptome features and value weight parameters are multiplied by a matrix to generate a transcriptome value matrix. The normalized attention matrix and the transcriptome value matrix are then multiplied to output the transcriptome protein metabolism fusion features. Deep transcriptome features and deep protein metabolism features are processed according to the parallel topological paths and computational instructions in the cross-omics attention network architecture. The protein metabolism-transcriptome fusion features are output, and the matrix merging calculation function is called by the merging processing node to perform same-dimensional splicing and output the bidirectional cross-feature matrix.

8. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 1, characterized in that, The response probability matrix is ​​obtained by nonlinearly mapping the bidirectional cross feature matrix using preset fully connected network parameters, including the following steps: A fully connected network layer is arranged at the end of the multi-layer neural network architecture of the prediction output module; The bidirectional cross feature matrix and the output weight matrix are multiplied together and the output bias vector is superimposed to generate a fused scoring result. The fusion scoring result is input into the Sigmoid activation function to perform a monotonic nonlinear mapping, generating an immune response scoring matrix, which serves as the response probability matrix.

9. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 8, characterized in that, It also includes the following steps: A preset judgment threshold is retrieved, and a row-by-row comparison is performed on the response probability matrix to determine the relationship between the judgment probability value and the preset judgment threshold. When the difference between the probability value and the preset judgment threshold exceeds the preset confidence interval, and the probability value is greater than or equal to the preset judgment threshold, it is classified into the responder classification array. When the difference between the probability value and the preset judgment threshold exceeds the preset confidence interval, and the probability value is less than the preset judgment threshold, it is classified into the non-responder classification array. When the difference between the probability value and the preset judgment threshold is within the preset confidence interval, a second review calculation is triggered and the status is marked as suspected.

10. The method for predicting the treatment response of non-small cell lung cancer using multi-omics technology according to claim 9, characterized in that, It also includes the following steps: For the samples recorded inside the said responder classification array and the CD8 + MAIT cell proportion greater than or equal to a first preset proportion threshold or CXCR6 + CD8 + The non-small cell lung cancer samples with MAIT cell proportion greater than or equal to a second preset proportion threshold are written with a preferred strategy label in the treatment response prediction report and marked as a predicted benefit population; For the non-small cell lung cancer samples recorded within the non-responder classification array, alternative strategy labels are written into the treatment response prediction report; Once the treatment response prediction report is generated, the prediction output module packages the generated treatment response prediction report and sends it to the display terminal on the periphery of the system.

Images