Cross-platform omics data integration and standardization method and system based on depth auto-encoder

Abstract

The invention discloses a cross-platform omics data integration and standardization method and system based on a depth auto-encoder, and belongs to the field of bioinformatics and computational biology, and the method comprises the steps: receiving cross-platform omics data and quality evaluation; correcting the batch effect by using platform information as a condition and introducing a condition variation self-encoder for resisting loss; fusing multiple omics data into a shared hidden space through a multi-head encoder and an attention mechanism; carrying out missing value filling and enhancement on the fused data based on the generative adversarial network; and outputting the standardized data and evaluating the quality. According to the method, the problems of nonlinear batch effect correction, multi-modal deep fusion and missing value intelligent processing in cross-platform omics data integration are effectively solved, and high-quality automatic data integration and standardization are realized.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, and in particular relates to a cross-platform omics data integration and standardization method and system based on deep autoencoders. Background Technology

[0002] With the rapid development of high-throughput sequencing technology, omics data is experiencing explosive growth. To integrate multi-source data from different experimental platforms, detection technologies, and batches, various cross-platform omics data integration and standardization methods have been developed. Common processing methods currently include: statistical correction methods (such as ComBat), which reduce batch-to-batch differences by linearly adjusting the data distribution; distance-based integration methods (such as MNN), which align samples between batches through nearest neighbor matching; matrix factorization-based methods (such as PMD), which extract common features across platforms using low-dimensional representations; and methods combining traditional machine learning (such as principal component analysis combined with regression models), which model technical variations linearly. These methods can, to some extent, alleviate technical biases between different platforms, improve data comparability, and provide a relatively consistent data foundation for subsequent biological analyses.

[0003] However, existing technologies still have significant shortcomings. First, traditional methods are mostly based on linear or shallow models, which have limited ability to correct for complex nonlinear batch effects and cannot completely eliminate deep technical biases caused by platform differences. Second, in terms of multi-omics data fusion, existing methods often remain at the level of shallow splicing or simple association, failing to achieve deep semantic-level fusion of multimodal data, resulting in the loss of biological association information across omics. In addition, existing solutions generally lack robust handling mechanisms for missing values ​​and usually do not have data augmentation capabilities, limiting their reliability in small sample scenarios. Finally, most existing processes lack end-to-end automated integration and standardization capabilities, and lack a systematic, multi-dimensional quality assessment system, making it difficult to comprehensively guarantee the quality and consistency of output data. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a cross-platform omics data integration and standardization method and system based on deep autoencoders, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a cross-platform omics data integration and standardization method based on a deep autoencoder, comprising: S1: Receive multi-platform omics data, and perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data; S2: Batch effect correction is performed on the cleaned multi-omics data based on conditional variational autoencoder to obtain corrected multi-omics data, in which platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. S3: The corrected multi-omics data is input into a multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. S4: Intelligent imputation of missing values ​​in the fused multi-omics data based on generative adversarial networks, and generation of enhanced samples based on the imputed complete data; S5: Based on the enhanced samples, output standardized data and evaluate the integration quality.

[0006] Preferably, in step S1, the quality assessment and cleaning includes: automatically parsing input data of different formats through a multi-platform adaptation interface; scoring the quality based on data integrity, detection rate, and signal-to-noise ratio; and using the isolated forest algorithm to identify and label samples with abnormal data quality.

[0007] Preferably, in step S2, the total loss function of the conditional variational autoencoder includes reconstruction loss, KL divergence loss, and adversarial loss.

[0008] Preferably, the adversarial loss is calculated through a discriminator network used to distinguish which platform the latent variable originates from; the goal of encoder training includes minimizing the classification ability of the discriminator network.

[0009] Preferably, in step S3, the training of the multimodal deep autoencoder adopts a hierarchical pre-training strategy, including: first, independently training the dedicated encoders and corresponding decoders for each modality; then, fixing some parameters of each dedicated encoder and jointly training the attention mechanism and the shared decoder part.

[0010] Preferably, in step S3, the multimodal depth autoencoder adopts a multi-head encoder-single decoder structure, and the fused latent variables are reconstructed through a unified decoder to reconstruct the modal data. The unified decoder contains independent branches corresponding to each modality.

[0011] Preferably, in step S4, when performing missing value imputation based on generative adversarial network, a mask matrix is ​​used to ensure that the observed feature values ​​of the sample are not modified during the imputation process.

[0012] Preferably, in step S4, the method of generating enhanced samples includes: performing linear interpolation on the latent variable representation of the samples in the shared latent space.

[0013] Preferably, in step S5, outputting standardized data and evaluating the integration quality includes: Z-score standardization is performed on each feature, and quantile normalization is used to align the feature distributions across different batches. The study calculated batch effect residual assessment indices, verified the preservation of known biological relationships in the integrated data, and calculated sample cluster profile coefficients.

[0014] Secondly, the present invention also provides a cross-platform omics data integration and standardization system based on a deep autoencoder, for implementing the method described in the first aspect, the system comprising: A cross-platform omics data interface and quality assessment module is used to receive multi-platform omics data, and to perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data. The batch effect correction module is used to perform batch effect correction on the cleaned multi-omics data based on the conditional variational autoencoder to obtain the corrected multi-omics data. The platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. A multimodal deep autoencoder fusion module is used to input the corrected multi-omics data into the multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. The missing value intelligent imputation and data augmentation module is used to intelligently imput missing values ​​in the fused multi-omics data based on generative adversarial networks, and generate augmented samples based on the imputed complete data; The standardized data output and verification module is used to output standardized data and evaluate the integration quality based on the enhanced samples.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention utilizes a batch effect correction module based on a conditional variational autoencoder, incorporating adversarial loss during model training. By treating platform information as a conditional variable and employing an adversarial training mechanism, it forces data from different sources to converge in the latent space. This technical solution can model and eliminate complex, nonlinear inter-platform technical biases. Compared to traditional linear correction methods, it more thoroughly removes technical noise, significantly improving the comparability and consistency of cross-platform omics data and providing a cleaner biological signal foundation for subsequent analysis.

[0016] This invention employs a multi-head encoder combined with an attention mechanism, utilizing a "multimodal deep autoencoder fusion module, including a dedicated encoder for each type of omics data. Each dedicated encoder maps the corresponding modality data to a shared latent space, and the mapped latent variables are weighted and fused through an attention mechanism." This technical solution can map multi-omics data, such as genomics and transcriptomics, to a unified shared latent space and dynamically learn the contribution weight of each modality to the overall representation. It achieves deep semantic fusion that goes beyond simple splicing or shallow association, effectively preserving the inherent biological connections between cross-omics data.

[0017] This invention integrates Generative Adversarial Networks (GANs) into the batch correction and data fusion process through a "missing value intelligent imputation and data augmentation module," which imputes missing values ​​in fused data based on a generative adversarial network and generates augmented samples based on the imputed complete data. This technical solution can intelligently impute missing values ​​that better conform to the overall data distribution using high-quality feature representations that have been corrected and fused, while simultaneously generating synthetic samples with reasonable biological characteristics. This effectively repairs data defects, expands the dataset size, and enhances the generalization ability of downstream analytical models.

[0018] This invention achieves fully automated processing from raw data access, quality assessment, batch correction, multimodal fusion, data repair and enhancement to standardized output and verification. This technical solution integrates multiple key steps into a coherent system framework, significantly reducing manual intervention and operation switching, ensuring the standardization and repeatability of the data processing workflow, and significantly improving the efficiency of large-scale cross-platform omics data integration.

[0019] This invention, through "assessing integration quality" and a further defined assessment module, can automatically perform multi-dimensional quality checks, including batch effect residual assessment, biological signal retention verification, and clustering effect evaluation. This technical solution provides comprehensive assessment indicators from technical correction effects to the retention of biological significance, forming a complete quality closed-loop verification system. This ensures the reliability of the final output standardized data and meets the stringent requirements of subsequent advanced bioinformatics analyses. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a cross-platform omics data integration and standardization method based on a deep autoencoder, according to an embodiment of the present invention. Figure 2 This is a structural diagram of the batch effect correction model for the conditional variational autoencoder according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the multimodal deep autoencoder fusion architecture according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the working principle of intelligent missing value imputation and data augmentation in an embodiment of the present invention. Figure 5 This is a visualization diagram illustrating the standardized data output and quality assessment of an embodiment of the present invention; Figure 6 This is a system framework diagram of cross-platform omics data integration and standardization based on deep autoencoders, according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] Example 1 like Figure 1 As shown, this embodiment provides a cross-platform omics data integration and standardization method based on a deep autoencoder, including: S1: Receive multi-platform omics data, and perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data; Furthermore, in step S1, the quality assessment and cleaning includes: automatically parsing input data of different formats through a multi-platform adaptation interface; scoring the quality based on data integrity, detection rate, and signal-to-noise ratio; and using the isolated forest algorithm to identify and label samples with abnormal data quality.

[0024] Specifically, the system integrates microbial sample data from various water bodies such as lakes and rivers. This data, generated by sequencing platforms like Illumina and BGI, includes microbial genome data, metatranscriptome data, and metabolome data. The system receives data through a standard interface. It automatically verifies data integrity, calculates feature detection rates, compares data distribution differences across platforms, and uses an isolated forest algorithm to identify and filter low-quality samples caused by sampling or library construction anomalies, outputting a clean, uniformly formatted data matrix.

[0025] In this embodiment, multi-platform omics data is received through a standardized interface, quality assessment is performed, unqualified samples and features are filtered out, and a clean preliminary data matrix is ​​output.

[0026] This embodiment features a unified application programming interface (API) that supports automatic parsing of various data formats, including FASTQ, BAM, CEL, and RAW. The interface includes a built-in format detector that can identify the data source platform (such as Illumina HiSeq, BGIMGISEQ, Thermo Fisher Ion Torrent, etc.) and invoke the corresponding parser.

[0027] Quality assessments were performed across multiple dimensions for each dataset, specifically including: Data integrity check: Verify the integrity of sample identifiers, feature identifiers, and numerical matrices.

[0028] Detection rate calculation: For each sample, calculate the feature detection rate and mark samples with low detection rates.

[0029] Distribution consistency assessment: The Kolmogorov-Smirnov test was used to compare the distribution differences of the same type of data on different platforms.

[0030] Outlier detection: The Isolation Forest algorithm is used to identify outlier samples based on feature representation distribution.

[0031] Preliminary cleaning and formatting: Remove substandard samples and features, convert the data into a uniform matrix format (sample × feature), and perform basic preprocessing such as logarithmic transformation.

[0032] S2: Batch effect correction is performed on the cleaned multi-omics data based on conditional variational autoencoder to obtain corrected multi-omics data, in which platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. Furthermore, in step S2, the total loss function of the conditional variational autoencoder includes reconstruction loss, KL divergence loss, and adversarial loss.

[0033] Furthermore, the adversarial loss is calculated through a discriminator network used to distinguish which platform the latent variables originate from; the goal of encoder training includes minimizing the classification ability of the discriminator network.

[0034] Specifically, the conditional variational autoencoder is trained conditioned on platform type. The encoder learns to extract platform-independent microbial community features, while the decoder learns to reconstruct data under given platform conditions. Through optimization incorporating adversarial loss, the model forces the distribution of samples from different platforms to converge in the latent space. After training, inputting data from any platform, the encoder extracts platform-independent features, which are then reconstructed by a decoder on a specified "reference platform," yielding corrected and comparable feature data. Visualization shows that before correction, data points clustered according to sequencing platform, while after correction, they are redistributed more according to aquatic environment type or dominant microbial communities.

[0035] This embodiment uses a conditional variational autoencoder model to learn platform-independent biometric representations based on platform information. Platform-specific bias is eliminated through an encoder-decoder architecture and adversarial training.

[0036] like Figure 2 As shown, the structure diagram of the batch effect correction model of the conditional variational autoencoder is presented. In this embodiment, a deep generative model is designed with platform information as a condition. The model input is the original feature vector x and the platform label c. The encoder network learns the posterior distribution q(z|x,c) of the latent variable z, and then the decoder network reconstructs the data p(x|z,c) from the latent variable.

[0037] Encoder network structure: Input layer (feature dimension) → fully connected layer (512 nodes, ReLU) → fully connected layer (256 nodes, ReLU) → μ layer (latent variable dimension) and σ layer (latent variable dimension); Decoder network structure: Input layer (latent variable dimension + platform label dimension) → fully connected layer (256 nodes, ReLU); Fully connected layer (512 nodes, ReLU) → Output layer (feature dimension); Multi-task loss function: The model optimization objective consists of three parts: Reconstruction loss: Measures the difference between the reconstructed data and the original data, using mean squared error. (1) KL divergence loss: The constrained latent variable distribution approximates a standard normal distribution. (2) Adversarial loss: Ensures consistent latent space distribution of samples from different platforms. A discriminator network D(z) is introduced to attempt to distinguish which platform the latent variable z originates from. The discriminator's classification ability is minimized during encoder training. (3) The total loss function is: , where β and γ are hyperparameters.

[0038] After training, for input data from any platform, the latent variable representation z is obtained through the encoder, and then the data is reconstructed using the decoder of the "standard platform" (such as specifying a representative platform) to obtain the corrected feature representation.

[0039] S3: The corrected multi-omics data is input into a multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. Furthermore, in step S3, the training of the multimodal deep autoencoder adopts a hierarchical pre-training strategy, including: first, independently training the dedicated encoders and corresponding decoders for each modality; then, fixing some parameters of each dedicated encoder and jointly training the attention mechanism and the shared decoder part.

[0040] Furthermore, in step S3, the multimodal depth autoencoder adopts a multi-head encoder-single decoder structure. The fused latent variables are reconstructed through a unified decoder to reconstruct the modal data. The unified decoder contains independent branches corresponding to each modality.

[0041] Specifically, for the corrected genomic, transcriptomic, and metabolomic data, the multimodal fusion module is activated. Each modality's data is mapped to a shared latent space via a dedicated encoder, and an attention mechanism dynamically calculates the weight of each modality in the overall representation. The fused latent variables are reconstructed through independent decoder branches for each modality, achieving deep semantic integration of cross-omics information.

[0042] like Figure 3 The diagram shows a multimodal deep autoencoder fusion architecture. In this embodiment, the corrected multi-omics data is mapped to a shared latent space through a multi-head encoder, weighted and fused using an attention mechanism, and then reconstructed through a shared decoder to achieve deep feature fusion.

[0043] The corrected multi-omics data (such as genome, transcriptome, methylome) are integrated into a unified low-dimensional space.

[0044] Specific implementation method: Multi-head encoder-shared latent space architecture: Dedicated encoder networks are designed for each type of omics data, and all encoder outputs are directed to the same latent space.

[0045] For M types of omics data, the encoder output for the m-th omics is: (4) Latent space fusion representation is achieved through weighted summation using an attention mechanism: (5) Attention weight It is learned through a small network.

[0046] Shared decoder design: Reconstructing modal data from fused latent variables: (6) Each modality has an independent decoder branch, but they share the latent variable input.

[0047] Layered training strategy: Modality-specific pre-training: Train the autoencoder for each modality separately to learn the data features of each modality.

[0048] Joint fine-tuning: Fix the parameters of the first half of the encoder, and jointly train the attention weights and the shared decoder.

[0049] End-to-end optimization: Fine-tuning the entire network with a smaller learning rate to optimize the multimodal reconstruction loss. (7) in The importance weights for each modality.

[0050] S4: Intelligent imputation of missing values ​​in the fused multi-omics data based on generative adversarial networks, and generation of enhanced samples based on the imputed complete data; Furthermore, in step S4, when performing missing value imputation based on generative adversarial network, a mask matrix is ​​used to ensure that the observed feature values ​​of the sample are not modified during the imputation process.

[0051] Furthermore, in step S4, the method of generating enhanced samples includes: performing linear interpolation on the latent variable representation of the samples in the shared latent space.

[0052] Specifically, for cases where some samples lack certain omics data or feature values, the missing value imputation module, based on a generative adversarial network, predicts and generates reasonable missing values ​​using existing observational features. Simultaneously, it generates synthetic samples with reasonable microbial community characteristics through latent space interpolation, or adds noise to the feature space to enhance dataset diversity and model robustness. For example... Figure 4 The diagram shows the working principle of the intelligent missing value imputation and data augmentation module.

[0053] In this embodiment, a generative adversarial network (GAN) is used to intelligently impute missing data. Interpolation is performed in the latent space to generate synthetic samples, expanding the dataset and improving the model's robustness.

[0054] This approach addresses the common problem of missing values ​​in real-world data and generates synthetic samples to augment the dataset. Specific implementation methods include: (1) Missing value imputation based on Generative Adversarial Networks (GANs): Generator G: Takes partial observed features xobs and random noise nn as input, and outputs the complete feature vector xcomplete.

[0055] Discriminator D: Determines whether the input features are real, complete data or generated data.

[0056] Imputation mechanism: For samples with missing values, the observed features are input into the generator, and the generated feature values ​​are used to fill in the missing positions. A mask matrix is ​​used to ensure that the observed values ​​are not modified.

[0057] Generator loss function: (8) Where M is a mask matrix (observed positions are 1, missing positions are 0), and ⊙ represents element-wise multiplication.

[0058] (2) Data augmentation strategies: Latent space interpolation: Linear interpolation is performed on similar samples in the latent space to generate synthetic samples with reasonable biological characteristics.

[0059] Feature-level noise injection: Adding appropriate noise to the complete sample to enhance the robustness of the model.

[0060] Adversarial example generation: FGSM (Fast Gradient Sign Method) is used to generate adversarial examples to improve the model's ability to handle abnormal situations.

[0061] S5: Based on the enhanced samples, output standardized data and evaluate the integration quality.

[0062] Furthermore, in step S5, outputting standardized data and evaluating the integration quality includes: Z-score standardization is performed on each feature, and quantile normalization is used to align the feature distributions across different batches. The study calculated batch effect residual assessment indices, verified the preservation of known biological relationships in the integrated data, and calculated sample cluster profile coefficients.

[0063] Specifically, all processed data undergoes Z-score standardization and quantile normalization to ensure the output matrix meets downstream analysis requirements. The system automatically generates an evaluation report: PCA plots demonstrate batch effect elimination; silhouette coefficients assess the clustering effect of samples by environmental type; and the preservation of known microbial metabolic pathway associations in the integrated data is calculated and validated. The final standardized data package can be directly used for studies such as microbial community structure analysis, functional gene mining, and environmental driver association analysis. Figure 5 As shown, this is a visualization diagram of standardized data output and quality assessment.

[0064] The integrated data is standardized to conform to a standard normal distribution. The integration effect is verified through multiple dimensions, including batch effect residual assessment, biological signal preservation assessment, and cluster quality assessment, and finally, a standardized data package is output.

[0065] Ensure that the output data meets standardization requirements and comprehensively evaluate the integration quality. Specific implementation methods include: (1) Data standardization processing: Feature scaling: Z-score standardization is applied to each feature so that the mean is 0 and the standard deviation is 1.

[0066] Distribution alignment: Quantile normalization is used to ensure that the feature distribution is consistent across different batches.

[0067] Outlier pruning: Scaling values ​​that exceed specified quantiles (such as 1% and 99%) to the boundary.

[0068] (2) Quality assessment index system: Residual batch effect assessment: Calculate the average batch distance (ABD) of data before and after correction. (9) in For the sample set of the kth batch, This is the mean of the batch of samples.

[0069] Biological signal retention assessment: Verify the biological consistency of the corrected data using known biological relationships (such as pathway co-expression and gene regulatory relationships). Calculate the proportion of biological relationships retained in the data.

[0070] Clustering quality assessment: The silhouette coefficient is used to assess the degree of clustering of samples of the same biological category in the feature space.

[0071] (3) Standardized data output: The output includes: 1) a fully standardized feature matrix; 2) a quality assessment report; 3) visualization charts (such as PCA charts and t-SNE charts); 4) metadata documents.

[0072] Example 2 like Figure 6 As shown, this embodiment provides a cross-platform omics data integration and standardization system based on a deep autoencoder, used to implement the method described in Embodiment 1, including: A cross-platform omics data interface and quality assessment module is used to receive multi-platform omics data, and to perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data. The batch effect correction module is used to perform batch effect correction on the cleaned multi-omics data based on the conditional variational autoencoder to obtain the corrected multi-omics data. The platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. A multimodal deep autoencoder fusion module is used to input the corrected multi-omics data into the multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. The missing value intelligent imputation and data augmentation module is used to intelligently imput missing values ​​in the fused multi-omics data based on generative adversarial networks, and generate augmented samples based on the imputed complete data; The standardized data output and verification module is used to output standardized data and evaluate the integration quality based on the enhanced samples.

[0073] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-platform omics data integration and standardization method based on deep autoencoders, characterized in that, Includes the following steps: S1: Receive multi-platform omics data, and perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data; S2: Batch effect correction is performed on the cleaned multi-omics data based on conditional variational autoencoder to obtain corrected multi-omics data, in which platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. S3: The corrected multi-omics data is input into a multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. S4: Intelligent imputation of missing values ​​in the fused multi-omics data based on generative adversarial networks, and generation of enhanced samples based on the imputed complete data; S5: Based on the enhanced samples, output standardized data and evaluate the integration quality.

2. The method according to claim 1, characterized in that, In step S1, the quality assessment and cleaning includes: automatically parsing input data of different formats through a multi-platform adaptation interface; scoring the quality based on data integrity, detection rate, and signal-to-noise ratio; and using the isolated forest algorithm to identify and label samples with abnormal data quality.

3. The method according to claim 1, characterized in that, In step S2, the total loss function of the conditional variational autoencoder includes reconstruction loss, KL divergence loss, and adversarial loss.

4. The method according to claim 3, characterized in that, The adversarial loss is calculated through a discriminator network used to distinguish which platform the latent variables originate from; the goal of encoder training includes minimizing the classification ability of the discriminator network.

5. The method according to claim 1, characterized in that, In step S3, the training of the multimodal deep autoencoder adopts a hierarchical pre-training strategy, including: first, independently training the dedicated encoders and corresponding decoders for each modality; then, fixing some parameters of each dedicated encoder and jointly training the attention mechanism and the shared decoder part.

6. The method according to claim 1, characterized in that, In step S3, the multimodal depth autoencoder adopts a multi-head encoder-single decoder structure. The fused latent variables are reconstructed through a unified decoder to reconstruct the modal data. The unified decoder contains independent branches corresponding to each modality.

7. The method according to claim 1, characterized in that, In step S4, when performing missing value imputation based on generative adversarial network, a mask matrix is ​​used to ensure that the observed feature values ​​of the sample are not modified during the imputation process.

8. The method according to claim 1, characterized in that, In step S4, the method of generating enhanced samples includes: performing linear interpolation on the latent variable representation of the samples in the shared latent space.

9. The method according to claim 1, characterized in that, In step S5, outputting standardized data and evaluating the integration quality includes: Z-score standardization is performed on each feature, and quantile normalization is used to align the feature distributions across different batches. The study calculated batch effect residual assessment indices, verified the preservation of known biological relationships in the integrated data, and calculated sample cluster profile coefficients.

10. A cross-platform omics data integration and standardization system based on a deep autoencoder, characterized in that, The system for implementing the method according to any one of claims 1-9 comprises: A cross-platform omics data interface and quality assessment module is used to receive multi-platform omics data, and to perform quality assessment and cleaning on the multi-platform omics data to obtain cleaned multi-omics data. The batch effect correction module is used to perform batch effect correction on the cleaned multi-omics data based on the conditional variational autoencoder to obtain the corrected multi-omics data. The platform information is used as a conditional variable input to the model, and adversarial loss is introduced during model training to make the distribution of samples from different platforms consistent in the latent space. A multimodal deep autoencoder fusion module is used to input the corrected multi-omics data into the multimodal deep autoencoder for fusion. The multimodal deep autoencoder includes a dedicated encoder set for each type of omics data. Each dedicated encoder maps the corresponding modal data to a shared latent space and performs weighted fusion of the mapped latent variables through an attention mechanism to obtain the fused multi-omics data. The missing value intelligent imputation and data augmentation module is used to intelligently imput missing values ​​in the fused multi-omics data based on generative adversarial networks, and generate augmented samples based on the imputed complete data; The standardized data output and verification module is used to output standardized data and evaluate the integration quality based on the enhanced samples.

Images