Multi-omics data missing interpolation method and system based on multi-view auto-encoder

Abstract

The invention provides a multi-omics data missing interpolation method and system based on a multi-view auto-encoder, and relates to the technical field of biological multi-omics data. The method comprises the following steps: firstly, acquiring multi-omics data related to a disease, and preprocessing the multi-omics data; then special encoders are constructed for different omics data; introducing a self-attention mechanism and a cross-attention mechanism to enhance the feature expression ability and realize cross-omics feature fusion based on the potential feature expressions of each group of science extracted by each encoder to obtain the fusion feature expressions of each group of science; then, adopting a hierarchical fusion strategy to further integrate feature representations of each group of chemical fusion into unified potential feature representations to realize deep fusion and compression of information; finally, specialized decoders are designed for encoders of each group for original data reconstruction; according to the method, the problem of data missing generally existing in multi-omics integration is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological multi-omics data, and particularly relates to a multi-omics data missing interpolation method and system based on a multi-view autoencoder. BACKGROUND

[0002] With the rapid development of high-throughput sequencing and various omics detection technologies, the systematic analysis from single molecular level to multi-dimensional biological data integration of genome, transcriptome, proteome and metabolome has been promoted. It provides an unprecedented opportunity to reveal the pathogenesis of complex diseases such as cancer, neurodegenerative diseases and metabolic syndrome, discover new biomarkers and achieve precise diagnosis and treatment. However, there is a phenomenon of data missing in multi-omics integration analysis. Taking the TCGA (The Cancer Genome Atlas) database as an example, among the 33 cancer types covered, only 58.3% of patients have genome, epigenetic and transcriptome data at the same time on average. The completeness of proteomic data is further reduced to 21.4%. This data missing phenomenon is mainly caused by the superimposed influence of multiple factors. At the hardware level, it is constrained by equipment resource limitation and sequencing cost. At the sample level, it involves the difficulty of tissue acquisition and preservation quality. In addition, there are differences in the development time axis of different omics technologies. Proteomic detection technology is about 10 years later than genomic technology, which leads to the lack of emerging omics data for early cases. The missing of data brings many challenges to multi-omics data integration.

[0003] In biomedical data analysis, researchers mainly adopt three methods to deal with missing data. First, sample screening. The most rigorous approach is complete case analysis, which only retains samples without any missing data. For example, Hoadley's team in the TCGA pan-cancer typing study rigorously screened 312 samples with complete genomic, transcriptomic, and epigenetic data from nearly 10,000 samples. Although this method is rigorous, it may cause sample selection bias. The relatively flexible pair deletion method selects samples according to specific analysis needs, for example, when studying gene co-expression, as long as both target genes have data. This method improves data utilization efficiency, but may affect the comparability of results because different analyses use different samples. The second method is to use simple data imputation strategies. The most basic is to replace missing values with mean or median, which Weinstein's team did in the TCGA project to handle partially missing gene expression data. However, this method weakens the correlation between variables. Hot deck imputation is to find similar samples to fill in missing values, which can better maintain the distribution characteristics of the data, but is sensitive to the calculation standard of similarity. The third is the statistical model-based method. Multiple imputation generates multiple possible data sets through Markov chain Monte Carlo simulation, which well considers the uncertainty of imputation, but has large computational load and strict requirements for data distribution. For example, the expectation-maximization algorithm (such as SVD-impute) estimates missing values by repeatedly optimizing, but it is easy to converge to a local optimal solution and sensitive to data distribution assumptions.

[0004] In recent years, deep learning methods have provided a new paradigm for multi-omics missing data reconstruction. Variational autoencoder captures data distribution characteristics through probabilistic generative model. And denoising autoencoder, such as MIDA model proposed by Gong et al., enhances model stability by artificially introducing missing noise, while contrastive learning autoencoder improves small sample scene performance through self-supervised signal. Generative adversarial network generates high-quality imputation values through adversarial training, but has the challenge of training instability. Diffusion model uses progressive denoising strategy to handle high missing rate data, although the result is stable but the inference speed is slow. Graph neural network method improves the biological rationality of imputation by integrating prior knowledge such as gene interaction network, but it is highly sensitive to the quality of the graph structure.

[0005] Despite the significant progress made by existing methods, there are still some core challenges: existing methods often rely too much on statistical models, ignoring the biological relationships and regulatory networks between molecules. Simple imputation strategies may destroy the internal structure of the data, while complex models may introduce artificial bias. And under high missing proportion of multi-omics data, the performance of traditional imputation methods decreases significantly, and even more noise may be introduced. SUMMARY

[0006] The technical problem solved by the present application is to provide a multi-omics data missing interpolation method and system based on a multi-view autoencoder to solve the problems of the prior art.

[0007] To solve the above technical problems, the technical solution adopted by the present application is: In one aspect, the present application provides a multi-omics data missing interpolation method based on a multi-view autoencoder, comprising: Obtaining multi-omics data related to a disease; Pretreating the obtained multi-omics data, including sample screening, normalization processing, and constructing a unified sample index to align the cross-omics data samples; Constructing a special encoder for different omics data, and each omics encoder together forms a multi-view autoencoder to mine the feature distribution and structure pattern of each omics, and obtain the latent feature representation of each omics; Based on the latent feature representation of each omics extracted by each encoder, introducing a self-attention mechanism and a cross-attention mechanism to enhance the feature expression ability and realize cross-omics feature fusion, and obtaining the fusion feature representation of each omics; Using a hierarchical fusion strategy, the feature representation of each omics fusion is further integrated into a unified latent feature representation to realize deep fusion and compression of information; Designing a specialized decoder for each omics encoder to reconstruct the original data; the structure of each decoder corresponds to the structure of the corresponding encoder, and the hierarchical design is consistent to ensure effective transmission of information and accurate decoding of features; each decoder receives the same latent feature representation, extracts information related to its target data type from the shared representation through parameter optimization in the supervised learning process, and realizes data reconstruction.

[0008] Further, the method extracts three different types of omics data: copy number variation CNV data at the genomic level, mRNA data, and RPPA data.

[0009] Further, the method constructs a CNV encoder for copy number variation CNV data at the genomic level; the CNV encoder adopts a multi-layer perceptron structure and introduces an activation function with a learnable slope parameter; Constructing an mRNA encoder for RNA sequencing data at the transcriptome level; the mRNA encoder adopts a multi-layer feedforward neural network structure and enters a cross-layer residual connection to alleviate the gradient vanishing problem of deep network structure, and combines LeakyReLU and ELU activation functions to enhance nonlinear expression ability; The RPPA encoder is constructed for reverse-phase protein array (RPPA) data at the proteome level; the RPPA encoder adopts a multi-layer perceptron structure and combines instance normalization and a Gaussian error linear unit activation function to improve the stability of noise data and the extraction capability of key features.

[0010] Further, the method is to solve the problem of missing values commonly existing in multi-omics data, and all encoders integrate missing mask information at the input layer, and use the input features and missing mask information to generate corresponding latent feature representations.

[0011] Further, based on the latent feature representations of each omics extracted by each encoder, a self-attention mechanism and a cross-attention mechanism are introduced to enhance the feature expression capability and realize cross-omics feature fusion, and the specific method for obtaining the fusion feature representation of each omics is as follows: The self-attention mechanism is applied to the latent feature representation of each omics data to enhance the internal feature representation of the omics and obtain the enhanced feature representation of each omics; The cross-attention mechanism is designed to realize directional information exchange between different omics data; the cross-attention mechanism enables one omics data to actively query related information in other omics data; On the basis of the cross-attention mechanism, an adaptive gating mechanism is introduced to avoid excessive interference from the omics features; Finally, the integrated enhanced feature representation of each omics and the cross-attention information from all other omics are obtained.

[0012] Further, a hierarchical fusion strategy is adopted to further integrate the feature representation of each omics into a unified latent feature representation to realize deep fusion and compression of information, and the specific method is as follows: The integrated features of all omics are spliced to form an initial fusion feature representation ; The fusion network is adopted to gradually map the initial fusion feature representation to a shared latent space through multiple layers of nonlinear transformation to obtain the latent feature representation, as shown in the following formula: ; ; Among them, represents the initial input data of the entire fusion network; represents the latent feature representation output by the qth layer of the fusion network, , represents the number of layers of the fusion network; is the learnable weight matrix and bias term of the qth layer of the fusion network; represents the batch normalization operation; ReLU is an activation function.

[0013] Further, the output feature representation of the decoder corresponding to each omics is expressed as: ; ; ; wherein, represents the feature input of the decoder, which is equal to the unified latent feature representation of omics t obtained through the fusion network ; represents the output feature of the pth layer of the decoder of omics t, p P represents the number of layers of the decoder; is the activation function of the pth layer of the decoder of omics t; and respectively represent the weight matrix and the bias term of the pth layer and the final layer of the decoder of omics t; represents the reconstructed omics data after decoding.

[0014] Further, the method considers the heterogeneity and differential statistical characteristics of multi-omics data, designs an adaptive loss function framework, and optimizes the imputation performance for different omics data types, specifically: For the characteristics of each omics data, determine the loss function corresponding to each omics data; introduce a learnable loss weight parameter for each type of omics data , and ensure the loss weight normalization of each omics through the Softmax function; The total loss function comprehensively considers the weighted loss of each omics, and simultaneously introduces an entropy regularization term to prevent the weight from being excessively concentrated on a single omics type.

[0015] On the other hand, the present application provides a multi-omics data missing imputation system based on a multi-view autoencoder, comprising a data acquisition module, a preprocessing module, an encoder construction module, a feature fusion module, a feature integration module and a decoder module. The data acquisition module is used to acquire multi-omics data related to diseases; The preprocessing module pre-processes the acquired multi-omics data, including sample screening, normalization processing and constructing a unified sample index to align the cross-omics data samples; The encoder construction module constructs a dedicated encoder for different omics data, and each omics encoder together forms a multi-view autoencoder to mine the feature distribution and structural pattern of each omics, and obtains a latent feature representation of each omics; The feature fusion module introduces a self-attention mechanism and a cross-attention mechanism based on the latent feature representation of each omics extracted by each encoder to enhance the feature expression ability and realize cross-omics feature fusion, and obtains a fusion feature representation of each omics. The feature integration module adopts a hierarchical fusion strategy to further integrate the features of each omics into a unified potential feature representation to realize deep fusion and compression of information. The decoder module designs a specialized decoder for the encoder of each omics to reconstruct the original data; the structure of the decoder is echoed with the corresponding encoder and remains consistent in hierarchical design to ensure effective transmission of information and accurate decoding of features.

[0016] In a third aspect, the present application provides an electronic device, comprising: one or more processors, and a memory for storing instructions, when the instructions are executed by the one or more processors, causing the one or more processors to execute the multi-omics data missing interpolation method based on the multi-view autoencoder.

[0017] In a fourth aspect, the present application provides a computer-readable storage medium storing executable instructions, which when executed cause a processor to execute the multi-omics data missing interpolation method based on the multi-view autoencoder.

[0018] In a fifth aspect, the present application provides a computer program product comprising a computer program or instructions, which when executed by a processor implement the multi-omics data missing interpolation method based on the multi-view autoencoder.

[0019] The technical scheme has the beneficial effects that the multi-omics data missing interpolation method and system based on the multi-view autoencoder provided by the present application solve the problem of data missing in multi-omics integration. The method adopts a multi-view autoencoder structure, constructs a customized specific omics encoder network architecture according to the unique statistical characteristics of different omics data types (genomic, transcriptomic, and proteomic), and accurately captures the inherent characteristics and complex patterns. In addition, the dual attention mechanism of the self-attention mechanism and the cross-attention mechanism is introduced to realize deep feature learning, which realizes the capture of complex biological associations between different omics data in the shared latent space. Finally, an adaptive loss function strategy is introduced to dynamically adjust the loss weight according to the characteristics of different omics data types, effectively handling the challenge of the heterogeneity of different omics data feature distribution.

[0020] The method and system of the present application maintain the biological association between omics data, and exhibit significant advantages on various missing rates, missing modes, and cancer data sets, with an average filling error reduced by 25-35% compared with existing methods, and performance close to that of using original complete data in downstream tasks such as molecular subtype classification. It provides a higher quality and more reliable multi-omics data basis for downstream analysis (such as disease mechanism analysis and biomarker mining). BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 Structure block diagram of the multi-view autoencoder provided for the embodiment 1 of the present application; Figure 2 Comparison chart of data set interpolation performance test results provided for the embodiment 1 of the present application. DETAILED DESCRIPTION

[0022] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0023] Embodiment 1:

[0024] The method of the present application is implemented based on a multi-view autoencoder architecture, which is composed of four interrelated core functional modules, each of which is carefully designed for specific challenges of multi-omics data analysis. The architecture follows the overall process of encoding-fusion-decoding, effectively capturing the biological interactions between different omics levels, and using these relationships to infer missing values.

[0025] Data type-specific encoder module: provides customized representation learning mechanism for different omics data types such as CNV, mRNA and RPPA. This specific design is based on the fact that different types of omics data have significantly different statistical characteristics and biological significance. CNV data is represented as a discrete genomic structural variation pattern; mRNA data is characterized by high-dimensional sparse distribution and wide expression dynamic range; RPPA data is often affected by batch effects and follows a non-standard distribution. By adjusting the internal structure of each encoder, including specific activation function selection, normalization strategy and network topology, each encoder can effectively capture the intrinsic characteristics of its corresponding data type, significantly improving the quality of initial feature extraction and laying a solid foundation for subsequent representation integration.

[0026] Self-attention feature enhancement module: introduces an intra-feature self-attention mechanism, which enables each omics data to dynamically evaluate the importance of its internal features. This mechanism can identify and enhance feature combinations with high biological relevance, while suppressing the influence of non-informational or redundant features.

[0027] Cross-attention integration module: designs a fully connected cross-attention network to realize structured information exchange between different omics data types. The cross-attention module uses learnable integration coefficients to dynamically adjust the strength of information exchange between different omics data types, enabling the model to identify relevant molecular events across omics levels and infer them through inter-molecular regulation relationships in the case of partial data loss.

[0028] Potential representation and decoding reconstruction module: the multi-omics representation enhanced by self-attention and cross-attention is integrated into a unified potential representation through feature concatenation and nonlinear transformation, which contains both shared information and specific information. In the decoding stage, a method combining shared potential space and specific decoder is adopted. Each decoder receives the same integrated potential representation, and through parameter optimization in the supervised learning process, it forms the ability to extract information related to its target data type from the shared representation. This mechanism is based on the representation separation ability of deep neural networks, allowing different decoders to naturally form sensitivity to different regions or feature combinations of the potential space during training without explicitly specifying the separation mechanism.

[0029] In this embodiment, the multi-omics data missing imputation method based on multi-view autoencoder includes the following steps: Step 1: Obtain multi-omics data related to diseases; In this embodiment, the multi-omics data of breast infiltrating carcinoma and pan-kidney carcinoma is collected from the TCGA database of the UCSC Xena browser (https: / / xenabrowser.net / ). For each cancer type, three different types of omics data are extracted: copy number variation (CNV) data at the genomic level, RNA sequencing (mRNA) data at the transcriptomic level, and reverse phase protein array (RPPA) data at the proteomic level; Step 2: Preprocess the multi-omics data obtained in step 1, including sample screening, normalization processing, and constructing sample index to align cross-omics data samples; Step 2.1: Sample screening: use Venn diagram analysis to determine the samples common to the three types of omics data to ensure consistency of multi-omics data; for each cancer type, only keep the samples present in all three types of omics data; and remove samples with more than 10% missing values to ensure data integrity; Step 2.2: Normalization processing: for copy number variation CNV data at the genomic level, use min-max scaling to standardize the value range to the [0, 1] interval; for RNA sequencing data at the transcriptomic level, implement two-stage conversion, first apply logarithmic transformation to alleviate severe right-skewed distribution, then perform Z-score standardization to eliminate scale differences between genes. For reverse phase protein array RPPA data at the proteomic level, apply median-based Robust Z-score standardization processing; each type of omics data after specific normalization achieves a standard distribution with mean 0 and standard deviation 1; Step 2.3: For the three types of omics data normalized in step 2.2, construct a unified sample index to align cross-omics data samples, solving the problem of inconsistent sample IDs between different omics data sets; Firstly, the sample ID set of each omics dataset is extracted, and the samples commonly owned by multiple omics are obtained by using the intersection operation to construct a unified sample index list; then, according to the unified sample index list, the data of each omics is subset filtered and sorted, so as to ensure the sample alignment of cross-omics data; Step 3: Due to the significant heterogeneity of different omics data in statistical characteristics and biological mechanisms, a special encoder is constructed for CNV, mRNA and RPPA, and the omics encoders together form a multi-view autoencoder, as shown in Figure 1 , to mine the feature distribution and structure pattern of each omics, and obtain the potential feature representation of each omics; the design aims to fully mine the feature distribution and structure pattern of each omics through a modular representation learning structure, so as to improve the expression ability of the omics-specific latent space; Step 3.1: A CNV encoder is constructed for copy number variation CNV data at the genomic level; for the discrete characteristics and spatial continuity of CNV data, the CNV encoder adopts a multi-layer perceptron structure and introduces an activation function (Parametric Rectified Linear Unit, PReLU) with a learnable slope parameter; the feature representation obtained by the multi-layer perceptron structure in the CNV encoder at the first l layer can be defined as: ; wherein, represents the feature representation obtained by the multi-layer perceptron structure in the CNV encoder at the first l layer; represents batch normalization, and are the weight matrix and bias vector of the multi-layer perceptron structure in the CNV encoder at the first layer, respectively; the PReLU activation function allows the slope parameter in the negative value interval to be adaptively learned, enhancing the sensitivity to discrete state transitions; The structure of the CNV encoder adopts a hierarchical shrinkage design, which gradually maps the original high-dimensional CNV feature vector (input dimension) of each CNV data sample to a latent space of 256 and 128 dimensions, forming a pyramid structure; this design can effectively extract multi-scale features of CNV data, while gradually compressing redundant information and retaining key structure variation patterns.

[0030] Step 3.2: Constructing mRNA encoder for RNA sequencing data at transcriptome level; considering the high dimensionality, collinearity and complex co-expression patterns of mRNA data, the mRNA encoder adopts a multi-layer feedforward neural network structure with cross-layer residual connection to alleviate the gradient vanishing problem of deep network structure, and combines LeakyReLU (Leaky Rectified Linear Unit) and ELU (Exponential Linear Unit) activation functions to enhance the non-linear expression ability; the feature representation obtained by the m-th layer feedforward neural network in the mRNA encoder is denoted as The feature representation obtained by the m-th layer feedforward neural network in the mRNA encoder is denoted as ; wherein, The feature representation obtained by the m-th layer feedforward neural network in the mRNA encoder is denoted as The feature representation obtained by the m-th layer feedforward neural network in the mRNA encoder is denoted as is the LeakyReLU (Leaky Rectified Linear Unit) and ELU (Exponential Linear Unit) activation function used alternately, The feature representation obtained by the m-th layer feedforward neural network in the mRNA encoder is denoted as n The weight matrix and bias vector of the m-th layer feedforward neural network in the mRNA encoder are denoted as is the residual connection, which is enabled when the input and output dimensions are the same; it alleviates the gradient vanishing problem by directly passing the information of the previous layer. This encoder adopts a hierarchical structure of [512, 256], with the first layer performing a large degree of dimension compression and the second layer further refining the features. The residual connection between adjacent layers not only effectively alleviates the gradient vanishing problem commonly encountered in deep network training, but also preserves the original co-expression patterns between genes.

[0031] Step 3.3: Constructing RPPA encoder for RPPA data at proteome level; considering the moderate dimensionality and high noise level of RPPA data, the RPPA encoder adopts a multi-layer perceptron structure combined with instance normalization (Instance Normalization) and Gaussian error linear unit (GELU) activation function to improve the stability of noise data and the extraction ability of key features; the feature representation obtained by the m-th layer of the multi-layer perceptron structure in the RPPA encoder is denoted as The feature representation obtained by the m-th layer of the multi-layer perceptron structure in the RPPA encoder is denoted as ; wherein, The feature representation obtained by the m-th layer of the multi-layer perceptron structure in the RPPA encoder is denoted as is the instance normalization, which effectively eliminates the technical variation between individuals by normalizing each sample independently; The GELU activation function provides smoother nonlinear transformation than the ReLU (Rectified Linear Unit) and has a natural inhibitory effect on noise signals for the weight matrix and bias vector of the m-th layer of the multi-layer perceptron structure in the RPPA encoder. The RPPA encoder adopts a [128, 64] compact structure to avoid overfitting caused by excessive parameterization, which is particularly important for proteomics data with limited sample size and high noise. This design can effectively extract the key protein activity patterns in RPPA data while reducing the interference of batch effects and technical noise. To handle the missing value problem commonly existing in multi-omics data, all encoders integrate missing mask information at the input layer. The integrated input feature representation is: ; wherein, is the integrated feature representation obtained by the t-th type of omics data at the input layer of the corresponding encoder after introducing the missing mask; represents the input feature of the t-th type of omics data; is the corresponding missing mask matrix; represents the element-wise multiplication operation; this strategy ensures that each omics encoder only models the observed feature values, avoiding noise introduced by missing data.

[0032] Each omics data type encoder generates the corresponding latent feature representation using the input feature and missing mask information, with the formula: ; wherein, represents the encoder designed for the t-th type of omics data, which outputs the latent feature representation of the omics through the input feature and the missing mask; Step 4: Based on the latent feature representation of each omics extracted by each encoder, self-attention mechanism and cross-attention mechanism are introduced to enhance feature expression ability and realize cross-omics feature fusion, obtaining the fusion feature representation of each omics; For each omics latent feature representation, the self-attention mechanism is introduced to enable each omics data to dynamically assess the importance of its internal features, identify and strengthen feature combinations related to biological functions, and suppress the influence of non-informational or redundant features; The cross-attention integration module is constructed by designing a fully connected cross-attention network to realize structured information exchange between different omics data types. The cross-attention integration module uses learnable integration coefficients to dynamically adjust the strength of information exchange between different omics data types, enabling the multi-view autoencoder to identify relevant molecular events across omics levels and make inferences through inter-molecular regulation relationships in the case of partial data missing; Step 4.1: Apply self-attention mechanism to the latent feature representation of each omics data, enhance the feature representation within the omics, and obtain the enhanced feature representation of each omics; The multi-head self-attention design projects the latent feature representation of each omics data into the query (Q), key (K), and value (V) space, and calculates the correlation strength between features through dot product similarity. The self-attention mechanism calculation process is as follows: ; where, is the attention function, Q, K, and V are the query, key, and value, respectively, is the scaling factor, which is equal to the feature dimension divided by the number of heads; this scaling design effectively prevents the gradient vanishing problem caused by the large dot product value. The core advantage of multi-head self-attention is that it allows the multi-view autoencoder to simultaneously focus on different aspects of the feature space and the projection subspace, greatly enriching the information capacity of the feature representation: ; where X is the input feature matrix; for each , , and are the learnable projection matrices corresponding to the fth attention head; h is the number of attention heads; To prevent the information introduced by the multi-head attention mechanism from distorting the original features, for each encoder of each omics type, a residual connection architecture is introduced to add the attention output features to the encoded features of the omics, and LayerNorm is used to ensure training stability; the final enhanced feature representation is calculated as: ; where, represents the feature representation obtained by the encoder for the tth omics data; represents the multi-head attention feature calculated based on the encoder output feature .

[0033] Step 4.2: In biological systems, there are complex regulatory relationships between different molecular levels of omics data, such as CNV affecting mRNA expression, and mRNA expression regulating protein levels. To model these cross-omics associations, a cross-attention mechanism is designed to enable directional information exchange between different omics data; the cross-attention mechanism enables one omics data to actively query related information in other omics data; through the cross-attention mechanism, omics j can actively query information in omics k; the cross-attention mechanism feature is calculated as: ; where a, b are omics categories including CNV, mRNA and RPPA and a b; is a cross-attention function, and denotes the enhanced feature representation of omics a and omics b; 、 and are learnable weight matrices for computing query (Q), key (K) and value (V). The cross-attention mechanism takes the enhanced feature of omics a as query, while the enhanced feature of omics b provides both key and value, allowing omics a to focus on the most relevant part of omics b; To avoid excessive interference from source omics features, especially when the correlation between two omics is weak, an adaptive gating mechanism is introduced on the basis of the cross-attention mechanism , the calculation process is as follows: ; where, is a sigmoid activation function that limits the gating value to the interval [0, 1], is a learnable gating weight matrix; this gating design allows the encoder to determine how much information from other omics should be absorbed according to the characteristics of the data itself, achieving true adaptive information flow control. The gating mechanism performs well when dealing with high missing rate data, effectively blocking the spread of unreliable information.

[0034] Finally, the source omics a integrates the enhanced feature representation and cross-attention information from all other omics b to obtain the fused feature representation of each omics: ; where, denotes element-wise multiplication, achieving fine-grained information flow control; Step 5: Adopt a hierarchical fusion strategy to further integrate the fused features of each omics into a unified latent feature representation, achieving deep fusion and compression of information after obtaining the fused feature representation of each omics in step 4; First, the integrated features of all omics are spliced to form the initial fusion feature representation : ; This direct splicing preserves the complete identity of each omics in the feature space, facilitating the identification of omics-specific information by subsequent fusion layers. Subsequently, a fusion network is used to gradually map the initial fusion feature representation to a shared latent space through multiple layers of nonlinear transformation, obtaining the latent feature representation: ; ; wherein, represents the initial input data of the entire fusion network; represents the latent feature representation output by the qth layer of the fusion network, represents the number of layers of the fusion network; is the learnable weight matrix and bias term of the qth layer of the fusion network; represents the batch normalization operation; ReLU is the activation function; In this embodiment, a 3-layer fusion network is adopted, and the layer widths are 512, 256 and 128 respectively, forming a gradient structure of gradually concentrated information. Compared with the basic fusion methods such as simple splicing or averaging, this multi-layer progressive fusion architecture can learn complex cross-omics feature interaction patterns and high-order nonlinear relationships, improving the information density and biological interpretability of the latent representation.

[0035] Step 6: Design a specialized decoder for each omics encoder to reconstruct the original data; In order to realize the original data reconstruction of each omics data, a special decoder is designed for each omics encoder. The structure of each decoder corresponds to the corresponding encoder, and the hierarchical design is consistent to ensure the effective transmission of information and accurate decoding of features; each decoder receives the same latent feature representation, extracts the information related to its target data type from the shared representation through parameter optimization in the supervised learning process, and realizes data reconstruction.

[0036] The output features of the decoder corresponding to each omics can be represented as: ; ; ; wherein, represents the feature input of the decoder, which is equal to the unified latent feature representation obtained by the omics t through the fusion network (that is, the latent feature representation output by the Qth layer of the fusion network); represents the output feature of the pth layer of the decoder of omics t, p P represents the number of layers of the decoder; is the activation function of the pth layer of the decoder of omics t; and respectively represent the weight matrix and bias term of the pth layer and the final layer of the decoder of omics t; represents the reconstructed omics data after decoding; Specifically, the CNV decoder uses the PReLU activation function to enhance the reconstruction accuracy of discrete states, the mRNA decoder uses the alternating LeakyReLU and ELU activation functions, and the RPPA decoder uses the GELU activation function and instance normalization to reduce noise and improve the robustness of protein expression level reconstruction.

[0037] In the multi-view autoencoder training process, the reconstruction accuracy and loss function optimization of the decoder for missing positions are very important. The mask-weighted loss function ensures that the multi-view autoencoder optimization target is for missing value prediction. The final decoded reconstruction data is represented as: ; wherein, respectively represent the reconstructed data of different omics by the decoder; the decoder combines the reconstruction errors of different omics to achieve the filling of missing positions and the recovery of missing values.

[0038] Step 7: Considering the heterogeneity and differential statistical characteristics of multi-omics data, an adaptive loss function framework is designed to optimize the imputation performance for different omics data types. CNV data loss function: CNV data is usually concentrated in a few discrete states, and on the other hand, it shows smooth transition characteristics in the continuous region of the genome. For this complex characteristic, a hybrid loss function is designed by combining the weighted combination of mean squared error (MSE) and mean absolute error (MAE) : ; wherein, MSE and MAE are only calculated at the observed positions to ensure that the loss calculation is not affected by the missing values: ; ; wherein, is the mask matrix corresponding to the CNV data, used to mark the missing data positions; and are the true value and the predicted value of the decoder reconstruction of the CNV data, respectively; w and r represent the row index and the column index, respectively; the MSE part focuses on the overall amplitude accuracy and imposes stronger punishment on large deviations, which helps to learn the overall distribution characteristics of the CNV data, while the MAE part is linearly sensitive to error size, enhancing the accurate positioning ability of the discrete transition boundary.

[0039] mRNA data loss function: mRNA data is characterized by high dimension, wide distribution, and extreme expression values. The expression range of different genes may differ by several orders of magnitude. In view of these characteristics, the mRNA data loss function uses a Huber-like loss function, as shown in the following formula: ; wherein, are the true value and the decoder reconstructed prediction value of mRNA data, respectively; is the mask matrix corresponding to the mRNA data; is a threshold parameter, when the error is less than , the MSE is used to provide accurate gradient information, and when it is greater than , the linearly increasing MAE is used, which effectively prevents the excessive influence of outliers on the training of the encoder. This design is particularly suitable for the high expression of cancer-specific genes that may exist in gene expression data, avoiding the risk of these extreme values dominating the entire training process.

[0040] RPPA data loss function: Compared with the previous two types of omics data, RPPA data has moderate dimension and has been standardized. Considering these characteristics of RPPA data, the RPPA data loss function uses the standard MSE loss to ensure reconstruction accuracy: ; wherein, are the true value and the decoder reconstructed prediction value of RPPA data, respectively; is the mask matrix corresponding to the RPPA data; The importance of different omics data is related to factors such as information quality, reliability, and prediction difficulty, and may vary significantly in different sample groups or disease backgrounds. To solve this problem, an adaptive weight mechanism is introduced, allowing the multi-view autoencoder to dynamically learn the optimal contribution proportion of different omics data in the overall loss. A learnable loss weight parameter is introduced for each type of omics data, defined as follows: ; wherein, are the loss weights of each omics data, initialized as equal weights; To ensure the comparability and numerical stability of the weights, the loss weights of each omics data are normalized by the Softmax function: ; wherein, denotes the weight of omics t, and t denotes different omics; denotes the normalized weight of each omics data. This design ensures that all loss weights are 1 and the weight values are non-negative, forming a probability distribution over the loss of each omic. In actual training, the weight parameters are optimized through backpropagation together with other network parameters, so as to automatically adjust the contribution of each omic in the total loss according to the characteristics of the data.

[0041] The total loss function takes into account the weighted loss of each omic, and an entropy regularization term is introduced to prevent the weights from being excessively concentrated on a single omic type. The total loss function can be expressed as: ; wherein, is the loss function corresponding to omic t; is the regularization strength; is the normalized loss weight of all omics; is the normalized loss weight of all omics is the entropy function of the normalized loss weight of all omics, and the calculation formula is: ; The entropy regularization term makes the multi-view autoencoder balance the use of information of each omic and prevents a certain omic data from dominating the multi-view autoencoder learning in the training process. When the reconstruction difficulty of a certain omic data is significantly higher than that of other omics, the multi-view autoencoder may tend to give up this omic and focus on the omic that is easier to reconstruct, and the entropy regularization term effectively alleviates this tendency.

[0042] In order to evaluate the effectiveness of the imputation strategies in different data, this embodiment compares the statistical-based mean, median filling method and K-Nearest Neighbor (KNN) machine learning algorithm with K value of 5. To ensure the objectivity of the evaluation, all methods are tested under a unified data division and evaluation framework. The root mean square error (RMSE) is used as the core indicator for performance evaluation, and the lower the value, the smaller the deviation between the imputed value and the true value, and the better the imputation effect.

[0043] To test the generalization ability of each method, experiments are carried out on three datasets with different characteristics: BRCA (breast cancer), KIPAN (pan-kidney cancer) and GBM (glioblastoma). The experimental results are as follows: Figure 2As shown, the results show that the multi-view autoencoder architecture of the present application achieves the lowest RMSE value on all datasets, showing the best performance. Specifically, compared with the suboptimal method, its error on the BRCA, KIPAN and GBM datasets is reduced by about 5.13%, 14.25% and 6.07% respectively. It is worth noting that on the GBM dataset with the most complex data structure, the simple mean filling method is better than the KNN algorithm, which reflects the limitations of traditional machine learning methods in dealing with highly heterogeneous data. In contrast, the multi-view autoencoder architecture of the present application performs stably in all tests, proving its strong adaptability and potential as a standardized preprocessing tool.

[0044] Embodiment 2: In this embodiment, the multi-omics data missing imputation system based on multi-view autoencoder includes a data acquisition module, a preprocessing module, an encoder construction module, a feature fusion module, a feature integration module and a decoder module. The data acquisition module is used to acquire multi-omics data related to diseases; The preprocessing module preprocesses the acquired multi-omics data, including sample screening, normalization processing and constructing a unified sample index to align the samples across omics data; The encoder construction module constructs a dedicated encoder for different omics data, and the omics encoders together form a multi-view autoencoder, which mines the feature distribution and structure pattern within each omics to obtain the latent feature representation of each omics; The feature fusion module introduces a self-attention mechanism and a cross-attention mechanism based on the latent feature representation of each omics extracted by each encoder to enhance the feature expression ability and realize cross-omics feature fusion, and obtains the fused feature representation of each omics; The feature integration module adopts a hierarchical fusion strategy to further integrate the feature representation of each omics fusion into a unified latent feature representation to realize deep fusion and compression of information; The decoder module designs a specialized decoder for the encoder of each omics to reconstruct the original data; the structure of the decoder corresponds to the corresponding encoder, and remains consistent in hierarchical design to ensure effective transmission of information and accurate decoding of features.

[0045] Embodiment 3: The embodiment provides an electronic device, which comprises one or more processors and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the multi-omics data missing imputation method based on the multi-view autoencoder.

[0046] The electronic device can be a mobile phone, a computer, a tablet computer, or the like, and includes a memory and a processor. The memory stores a computer program which, when executed by the processor, implements the multi-omics data missing imputation method based on a multi-view autoencoder as described in the embodiments. It can be understood that the electronic device can further include an input / output (I / O) interface and a communication component.

[0047] The processor is configured to execute all or part of the steps of the multi-omics data missing imputation method based on a multi-view autoencoder as described in the embodiments. The memory is configured to store various types of data, which can include, for example, instructions of any application program or method in the electronic device, and application-related data.

[0048] The processor can be an Application Specific Integrated Cricuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic elements, and is configured to execute the multi-omics data missing imputation method based on a multi-view autoencoder as described in the embodiments.

[0049] Embodiment 4 The embodiment provides a computer-readable storage medium storing executable instructions which, when executed, can be stored in one computer-readable storage medium if implemented in the form of a software functional unit and sold or used as an independent product.

[0050] The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the multi-omics data missing imputation method based on a multi-view autoencoder as described in the embodiments.

[0051] The aforementioned storage medium includes a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD (Secure Digital Memory Card) or a DX (an abbreviation of Memory Data Register, MDR) memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an APP (an abbreviation of Application) application store, and the like, which can store a program check code, and on which a computer program is stored, the computer program being executed by a processor to implement each step of the multi-omics data missing imputation method based on a multi-view autoencoder.

[0052] Embodiment 5 The embodiment provides a computer program product, including a computer program or instructions, which are executed by a processor to implement the multi-omics data missing imputation method based on a multi-view autoencoder.

[0053] Based on such understanding, the technical solution of the present application or the part of the technical solution that essentially contributes to the prior art or the part of the technical solution can be embodied in the form of a computer program product.

[0054] Each embodiment in the present application is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

[0055] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the present application.

Claims

1. A multi-omics data missing imputation method based on multi-view autoencoder, characterized in that, The method comprises the following steps: Obtaining disease-related multi-omics data; Preprocessing the obtained multi-omics data, including sample screening, normalization processing, and constructing a unified sample index to align the samples across different omics data; Constructing a special encoder for different omics data, and the omics encoders together form a multi-view autoencoder to mine the feature distribution and structural pattern of each omics, and obtain the latent feature representation of each omics; Based on the latent feature representation of each omics extracted by each encoder, a self-attention mechanism and a cross-attention mechanism are introduced to enhance the feature expression ability and realize cross-omics feature fusion, and the fusion feature representation of each omics is obtained; A hierarchical fusion strategy is adopted to further integrate the feature representation of each omics into a unified latent feature representation to realize deep fusion and compression of information; Designing a specialized decoder for each omics encoder to reconstruct the original data; the structure of each decoder corresponds to the structure of the corresponding encoder, and the hierarchical design is consistent to ensure effective information transmission and accurate decoding of features; each decoder receives the same latent feature representation, extracts the information related to the target data type from the shared representation through parameter optimization in the supervised learning process, and realizes data reconstruction. 2.The multi-omics data missing imputation method based on multi-view autoencoder according to claim 1, wherein, The method extracts three different types of omics data: copy number variation CNV data at the genomic level, mRNA data, and RPPA data. 3.The multi-omics data missing imputation method based on multi-view autoencoder according to claim 2, characterized in that, The method constructs a CNV encoder for copy number variation CNV data at the genomic level; the CNV encoder adopts a multi-layer perceptron structure and introduces an activation function with a learnable slope parameter; An mRNA encoder is constructed for RNA sequencing data at the transcriptome level; the mRNA encoder adopts a multi-layer feedforward neural network structure and enters a cross-layer residual connection to alleviate the gradient vanishing problem of deep network structure, and combines LeakyReLU and ELU activation functions to enhance the non-linear expression ability; An RPPA encoder is constructed for reverse phase protein array RPPA data at the proteome level; the RPPA encoder adopts a multi-layer perceptron structure and combines instance normalization and Gaussian error linear unit activation functions to improve the stability of noise data and the extraction ability of key features.

4. The multi-omic data missing imputation method based on multi-view autoencoder according to claim 3, characterized in that, To deal with the missing value problem commonly existing in multi-omics data, all encoders integrate missing mask information at the input layer, and use the input features and missing mask information to generate the corresponding latent feature representation.

5. The multi-omic data missing imputation method based on multi-view autoencoder according to claim 4, characterized in that, The specific method based on the latent feature representation of each omics extracted by each encoder, introducing a self-attention mechanism and a cross-attention mechanism to enhance the feature expression ability and realize cross-omics feature fusion, and obtaining the fusion feature representation of each omics is as follows: Apply the self-attention mechanism to the latent feature representation of each omics data to enhance the internal feature representation of the omics and obtain the enhanced feature representation of each omics; Design a cross-attention mechanism to realize directional information exchange between different omics data; the cross-attention mechanism enables one omics data to actively query related information in other omics data; On the basis of the cross-attention mechanism, an adaptive gating mechanism is introduced to avoid excessive interference from the source omics features; Finally, the enhanced feature representation of each omics and cross-attention information from all other omics are integrated to obtain the fused feature representation of each omics.

6. The multi-omic data missing imputation method based on multi-view autoencoder according to claim 5, characterized in that, The hierarchical fusion strategy is adopted to further integrate the feature representation of each omics into a unified latent feature representation to realize the deep fusion and compression of information. concatenating the integrated features of all omics to form an initial fused feature representation ; The fusion network is adopted to gradually map the initial fused feature representation to a shared latent space through multiple layers of nonlinear transformation to obtain the latent feature representation, as shown in the following formula: ； ； wherein, represents the initial input data of the entire fusion network; represents the latent feature representation of the output of the qth layer in the fusion network, , represents the number of layers of the fusion network; is the learnable weight matrix and bias term of the qth layer of the fusion network; represents the batch normalization operation; ReLU is the activation function.

7. The multi-omic data missing imputation method based on multi-view autoencoder according to claim 6, characterized in that, The output feature representation of the decoder corresponding to each omics is: ； ； ； wherein, represents the feature input of the decoder, which is equal to the unified latent feature representation of omics t obtained by the fusion network ; represents the output feature of the pth layer of the decoder of omics t, p P represents the number of layers of the decoder; is the activation function of the pth layer of the decoder of omics t; and respectively represent the weight matrix and the bias term of the pth layer and the final layer of the decoder of omics t; represents the reconstructed omics data after decoding.

8. The multi-omic data missing imputation method based on multi-view autoencoder according to claim 7, characterized in that, The method considers the heterogeneity and differential statistical characteristics of multi-omics data, designs an adaptive loss function framework, and optimizes the imputation performance for different omics data types, specifically: The loss function corresponding to each omics data is determined according to the characteristics of the omics data; a learnable loss weight parameter is introduced for each kind of omics data and the loss weights of the omics are normalized through the Softmax function. The total loss function comprehensively considers the weighted loss of each omics, and introduces an entropy regularization term to prevent the weights from being excessively concentrated on a single omics type.

9. A multi-omics data missing imputation system based on multi-view autoencoder, implemented based on the method of claim 1, characterized in that: The method comprises a data acquisition module, a preprocessing module, an encoder construction module, a feature fusion module, a feature integration module, and a decoder module. The data acquisition module is configured to acquire multi-omics data related to a disease. The preprocessing module is configured to preprocess the acquired multi-omics data, including sample screening, normalization processing, and constructing a unified sample index to align the cross-omics data samples. The encoder construction module is configured to construct a dedicated encoder for different omics data, and the encoders of each omics together form a multi-view autoencoder to mine the feature distribution and structure pattern within each omics and obtain the latent feature representation of each omics. The feature fusion module is configured to extract the latent feature representation of each omics based on the encoders, introduce a self-attention mechanism and a cross-attention mechanism to enhance the feature expression ability and realize cross-omics feature fusion, and obtain the fused feature representation of each omics. The feature integration module is configured to adopt a hierarchical fusion strategy to further integrate the fused feature representation of each omics into a unified latent feature representation to realize the deep fusion and compression of information. The decoder module is configured to design a specialized decoder for the encoder of each omics to reconstruct the original data; the structure of the decoder is in correspondence with the corresponding encoder and maintains consistency in hierarchical design to ensure effective information transmission and accurate feature decoding.

10. A computer program product for performing the multi-omic data missing imputation method based on multi-view autoencoder according to any one of claims 1-8, characterized in that: The computer program or instructions are executed by a processor to implement the multi-omics data missing imputation method based on the multi-view autoencoder.

Images