A single-cell transcriptome data integration method and system based on explicit decoupling and optimal transmission

Abstract

This invention relates to the fields of bioinformatics and computational biology, specifically to a method and system for integrating single-cell transcriptome data based on explicit decoupling and optimal transport. First, multiple batches of single-cell gene expression matrices are preprocessed. Then, a deep residual autoencoder maps the input data to a structured latent representation, explicitly segmenting it into biological feature components and batch noise components. These components are then completely separated through joint optimization using multiple loss functions. High-quality nearest neighbor pairs are selected as anchors based on a clean biological feature space. A generative adversarial network is constructed using these anchors as training samples. An optimal transport regularization term is introduced into the generator loss function, and the Wasserstein distance between distributions is minimized using the Sinkhorn algorithm, achieving accurate and geometrically smooth distribution alignment. Finally, the corrected gene expression data or low-dimensional embedding representation is output. This invention achieves efficient integration of data from different batches, platforms, and species.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and computational biology, specifically to a method and system for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission. Background Technology

[0002] Single-cell RNA sequencing technology is a core tool for studying cellular heterogeneity, developmental trajectories, and disease mechanisms. It can reveal dynamic changes in gene expression at the single-cell level, providing an unprecedented perspective for life science research. However, in practical applications, significant technical variations, known as the "batch effect," often exist between single-cell transcriptome data from different experimental batches, sequencing platforms, or donors. This batch effect is not caused by biological differences but stems from variations in experimental conditions, procedures, and equipment, severely interfering with accurate cell type identification and reliable cross-dataset comparisons. It has become a key bottleneck restricting the in-depth application of single-cell sequencing technology.

[0003] To address the batch effect problem, various methods for integrating single-cell transcriptome data have been proposed in existing technologies, which can be mainly divided into three categories: 1. Statistical correction methods based on linear models (such as ComBat): These methods estimate and eliminate batch effects by constructing linear models. However, since batch effects often exhibit complex nonlinear characteristics, these methods are difficult to handle strong nonlinear technical variations, resulting in limited correction effects.

[0004] 2. Local correction methods based on nearest neighbor (MNN) (such as Harmony and Scanorama): These methods find similar cell pairs (anchor points) between different batches in the original data space and perform local correction based on the anchor points. However, their effectiveness is highly dependent on the quality of the initial data and the reliability of the anchor points. Under strong batch effects, the cell distribution in the original data space is severely distorted, leading to inaccurate anchor point identification and thus causing the correction to fail.

[0005] 3. Generative models based on deep learning (such as scVI, iMAP): These methods typically employ an "implicit decoupling" strategy, indirectly encouraging the separation of biological signals from technical noise in the latent space through a loss function. However, these methods lack a structured separation mechanism, resulting in poor model interpretability. Furthermore, when batch effects are strong or cell types have complex structures, they are prone to problems such as undercalibration, overcalibration, or disruption of the continuity of biological structures, leading to issues like broken cell developmental trajectories and loss of rare cell subpopulations. Summary of the Invention

[0006] The purpose of this invention is to propose a method and system for integrating single-cell transcriptome data. Through an adversarial learning architecture of structured explicit decoupling and optimal transport regularization, it achieves high-precision, high-robustness, and high biological fidelity integration of single-cell transcriptome data, improves the interpretability and reliability of the integration process, and protects complex biological structures.

[0007] According to a first aspect of the present disclosure, a method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission is provided, comprising the following steps: Multiple batches of single-cell gene expression matrices were obtained, and the input data was obtained after preprocessing. A deep residual autoencoder is constructed to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. Based on the biological feature components, an approximate nearest neighbor search algorithm is used to screen the nearest neighbor pairs between different batches to form a set of highly reliable anchor points. Using the set of anchor points as training samples, a generative adversarial network is constructed. An optimal transmission regularization term is introduced into the loss function of the generator. By minimizing this loss function, the distribution alignment of the source batch and the target batch of data is achieved. Input the source batch data to be corrected into the trained generator, and output the corrected gene expression data or low-dimensional embedding representation.

[0008] In one embodiment, the comprehensive loss function is: in, To rebuild the losses, For loss of content consistency, Loss due to batch classification; The weight hyperparameters for reconstructing the loss, The weight hyperparameters for content consistency loss are... The weighted hyperparameters for batch classification loss; For input data The mathematical expectation, This represents the expression vector of the source batch of cells. For the encoder to input The mapping result shows that the vector is explicitly segmented into biological feature components. and batch noise components ; For input data and random batch labels The joint mathematical expectation, As a new biological characteristic component; To measure batch noise components conduct Activation operation, For cross-entropy loss calculation, measure Predicted batch probability distribution and actual batch label The differences.

[0009] In one embodiment, the loss function of the generator is: in, It is a hyperparameter that controls the strength of regularization. This represents the expression vector of the source batch of cells. For generator functions; For the discriminator function, The distribution of target reference batch cells in the biological potential space. This represents the distribution of source batch cells in the biological latent space after generator correction. For the source batch data distribution P B Randomly selected samples For the sample The average value is taken from the results processed by the generator and discriminator. This is the optimal transmission loss.

[0010] In one embodiment, the optimal transmission loss The generated cells are obtained through the Sinkhorn algorithm with entropy regularization: for a batch of source cells sampled from nearest neighbor pairs, the generated cells are obtained through the generator. and its corresponding target reference cells The biological latent representations of these encoders trained in the first stage are extracted. and Then calculate the cost matrix. The Sinkhorn algorithm iteratively solves an entropy-regularized optimal transmission problem, obtains a soft allocation matrix (transmission plan), and calculates the minimum transmission cost as the minimum value. .

[0011] In one embodiment, the depth residual autoencoder will input Mapping to latent vectors ,Will Explicitly split into front biological components of vitae and after Batch noise components of dimensional ,Right now decoder The combined features are received and reconstructed, and the reconstructed expression spectrum is output. .

[0012] In one embodiment, the high-confidence anchor set is obtained by: normalizing the biological feature components; using the Annoy algorithm to construct an approximate nearest neighbor index, finding the k nearest neighbors of cells in other batches for each batch; and identifying cell A as a mutual nearest neighbor pair if and only if cell A is in cell B's k nearest neighbor list and cell B is also in cell A's k nearest neighbor list, and all mutual nearest neighbor pairs constitute the high-confidence anchor set M.

[0013] In one embodiment, the generative adversarial network includes a generator and a discriminator. The generator employs a fully connected layer structure with residual connections to map the expression profile of the source batch of cells to the distribution of the reference batch, generating a corrected expression profile. The discriminator employs a discrimination criterion based on Wasserstein distance and incorporates gradient penalty GP to achieve stable training, i.e., it adopts the WGAN-GP framework to distinguish between the real reference batch of cells and the generator's corrected expression profile.

[0014] According to a second aspect of the present disclosure, a single-cell transcriptome data integration system based on explicit decoupling and optimal transmission is provided, comprising: The data preprocessing module obtains single-cell gene expression matrices from multiple batches and preprocesses them to obtain input data. An explicit decoupling module constructs a deep residual autoencoder to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. The anchor point identification module, based on the biological feature components, uses an approximate nearest neighbor search algorithm to filter mutually nearest neighbor pairs between different batches to form a high-confidence anchor point set; The adversarial correction module uses the anchor set as training samples to construct a generative adversarial network. It introduces an optimal transmission regularization term into the generator's loss function and achieves data distribution alignment between the source batch and the target batch by minimizing the loss function. The data output module takes the source batch data to be corrected and inputs it into the trained generator, then outputs the corrected gene expression data or low-dimensional embedding representation.

[0015] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the single-cell transcriptome data integration method based on explicit decoupling and optimal transmission.

[0016] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission.

[0017] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention employs a structured explicit decoupling design, forcibly achieving physical isolation between biological signals and technical noise at the model architecture level, completely overcoming the "black box" limitation of implicit decoupling in current mainstream methods. The entire data correction process is logically coherent, from extracting pure biological information through feature decoupling, to identifying high-quality anchor points based on this information, and then to achieving distribution alignment through optimal transmission. The information flow in each stage is transparent and traceable throughout, facilitating researchers' understanding and verification, and solving the core pain point of insufficient interpretability in traditional deep learning integration methods.

[0018] 2. Explicit decoupling provides a pure biological representation stripped of technical noise for subsequent processes, ensuring the accuracy of nearest neighbor anchor identification from the source. Simultaneously, optimal transport regularization imposes geometric continuity constraints on the adversarial learning process, creating a synergistic effect. This enables the invention to consistently deliver excellent integration results even in highly challenging application scenarios such as strong batch effects, complex cellular heterogeneity, and cross-species data. It effectively avoids problems commonly found in traditional methods, such as insufficient correction, overcorrection, or disruption of biological structural continuity, significantly improving robustness.

[0019] 3. The smooth alignment mechanism, built upon optimal transport theory, can specifically protect key biological features in single-cell data. Important biological processes, such as detailed cell subtype structures, scarce cell populations, and dynamic, continuous cell development trajectories, are preserved to the maximum extent during batch effect elimination. This provides a more reliable data foundation for downstream bioinformatics research, including cell type annotation, developmental trajectory inference, and differentially expressed gene analysis, ensuring the scientific rigor and accuracy of subsequent analytical results.

[0020] 4. The core framework of this invention does not rely on specific prior biological knowledge and possesses strong scenario adaptability. It can be flexibly extended to meet various single-cell transcriptome data integration needs across experimental platforms, tissues, disease states, and species. Simultaneously, the framework's algorithm design takes into account large-scale data processing capabilities, efficiently handling single-cell atlas data at the million-cell scale, meeting the application demands of current single-cell sequencing technology towards high-throughput and large-scale processing, and possessing broad practical value. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0022] Figure 1 Flowchart of a method for integrating single-cell transcriptome data; Figure 2 This is a schematic diagram of a single-cell transcriptome data integration system. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0027] Example 1: This embodiment provides a method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission, including the following steps: S1. Obtain single-cell gene expression matrices from multiple batches, and preprocess them to obtain input data; Specifically, preprocessing can be performed using a standard single-cell analysis workflow (such as Scanpy) as follows: First, perform basic quality filtering of cells and genes; second, normalize the library size for gene expression counts of each cell to a standard of 10,000 total counts per cell; then, perform a log(1+x) transformation to stabilize the data variance; finally, based on the joint data of all batches, select 2000 highly variable genes to form the input feature matrix of the model.

[0028] S2. Construct a deep residual autoencoder to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. Specifically, the encoder maps the input high-dimensional gene expression vector into a structured latent vector. This latent vector is explicitly and rigidly divided into two parts: the first L dimensions are biological feature components, which specifically encode only inherent biological identity information such as cell type, state, and developmental stage; the last K dimensions are batch noise components (K is the number of experimental batches), which specifically encode only technical variation information such as experimental source, sequencing platform, and donor. The two parts are functionally independent and physically isolated in the feature space.

[0029] The decoder receives the structured latent vector obtained from the above combination, reconstructs the gene expression profile, and outputs the reconstructed expression profile. By jointly optimizing the following three types of loss functions, a complete source separation is forced between the biological feature components and the batch noise components: Reconstruction loss: Calculated using mean squared error (MSE) to ensure the model accurately captures the original gene expression patterns; Content consistency loss: To enforce the batch invariance of biological feature components, the component is combined with a randomly generated batch noise vector (corresponding to random batch labels), and reconstructed into a "pseudo-cell" expression vector by the decoder. This vector is then input into the encoder to extract new biological feature components. By constraining the new feature components to be consistent with the original biological feature components, content consistency constraints are achieved. Batch classification loss: By strengthening the encoding ability of batch noise components to batch sources through supervised learning, the batch noise components are input into the classifier to predict their batch labels, minimizing the cross-entropy loss between the predicted labels and the true batch labels.

[0030] Through the above multi-objective optimization training, the encoder can learn a pure biological feature representation that is batch invariant.

[0031] S3. Based on the biological feature components, an approximate nearest neighbor search algorithm is used to screen the nearest neighbor pairs between different batches to form a set of highly reliable anchor points; Specifically, this invention calculates the nearest neighbor pairs between different batches of cells in a pure biological feature space stripped of technical noise (rather than the original data space containing technical variations). These cell pairs selected in the denoised feature space have higher biological matching accuracy and become interpretable and highly reliable anchor points connecting the feature learning and distribution correction stages.

[0032] The implementation process is as follows: Based on the biological feature components of all cells obtained in the first stage, normalization is first performed to eliminate the influence of scale differences; then, an efficient approximate nearest neighbor search algorithm (such as Annoy) is used to match the k nearest neighbors of each batch of cells in the pure biological feature space; a pair of cells is identified as a high-quality mutual nearest neighbor pair if and only if cell A is in the k nearest neighbor list of cell B, and cell B is also in the k nearest neighbor list of cell A; all cell pairs that satisfy this condition constitute the anchor set M, which serves as a reliable supervision signal for the subsequent distribution correction process.

[0033] S4. Using the set of anchor points as training samples, construct a generative adversarial network, introduce an optimal transmission regularization term into the loss function of the generator, and achieve distribution alignment between the source batch and the target batch of data by minimizing the loss function; Specifically, using the anchor set M as training samples, a Generative Adversarial Network (GAN) is constructed, comprising a generator and a discriminator. The generator employs a fully connected layer structure with residual connections, and its core function is to learn to map the expression profiles of the source batch of cells to the distribution space of the reference batch (usually the batch with the largest number of cells), outputting the corrected cell expression profile. The discriminator uses a discrimination criterion based on Wasserstein distance, and introduces gradient penalty (GP) to ensure training stability (i.e., using the WGAN-GP framework). Its core function is to distinguish whether the input cell expression profile is the real data of the reference batch or the corrected data output by the generator.

[0034] A regularization term based on optimal transport theory is introduced into the generator's loss function. This regularization term, calculated using the Sinkhorn algorithm with entropy regularization, accurately quantifies and minimizes the Wasserstein distance between the source batch and the target reference batch of cells in the biological latent representation space, enabling a geometrically smooth "soft assignment" during the distribution alignment process. Guided by high-quality biological anchors, this design achieves distribution alignment between the source and target batches while effectively protecting the continuity of the cell state space and local topology (such as continuous developmental trajectories). This avoids the overcorrection and biological structure destruction problems that are prone to occur in traditional hard matching or adversarial learning methods. Simultaneously, it guides the generator to align the source distribution to the target distribution with minimal geometric deformation, ensuring the effectiveness of batch difference elimination and the integrity of biological features. During training, samples are sampled from the set M of nearest neighbors, and the discriminator and generator are trained iteratively and alternately until the model converges.

[0035] S5. Input the source batch data to be corrected into the trained generator, and output the corrected gene expression data or low-dimensional embedding representation.

[0036] Specifically, after the adversarial correction training is completed, all source batches of data to be corrected are input into the trained generator, which outputs the batch-corrected gene expression matrix or a processed low-dimensional embedding representation. This output enables thorough mixing of different experimental batches, sequencing platforms, and even cells of the same species, while preserving as much as possible the fine cell type structure, rare cell subpopulations, and continuous cell developmental trajectories, among other key biological information.

[0037] The corrected gene expression matrix can be directly used for various downstream bioinformatics analyses such as unsupervised clustering, cell type annotation, differential expression analysis, and pseudo-temporal analysis, and can ensure the comparability and reliability of analysis results between different batches.

[0038] Example 2: This embodiment provides a single-cell transcriptome data integration system based on explicit decoupling and optimal transmission, including: The data preprocessing module obtains single-cell gene expression matrices from multiple batches and preprocesses them to obtain input data. An explicit decoupling module constructs a deep residual autoencoder to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. The anchor point identification module, based on the biological feature components, uses an approximate nearest neighbor search algorithm to filter mutually nearest neighbor pairs between different batches to form a high-confidence anchor point set; The adversarial correction module uses the anchor set as training samples to construct a generative adversarial network. It introduces an optimal transmission regularization term into the generator's loss function and achieves data distribution alignment between the source batch and the target batch by minimizing the loss function. The data output module takes the source batch data to be corrected and inputs it into the trained generator, then outputs the corrected gene expression data or low-dimensional embedding representation.

[0039] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0040] Example 3: An electronic device is provided for running the aforementioned "a method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain single-cell gene expression matrices from multiple batches, and preprocess them to obtain input data; S2. Construct a deep residual autoencoder to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. S3. Based on the biological feature components, an approximate nearest neighbor search algorithm is used to screen the nearest neighbor pairs between different batches to form a set of highly reliable anchor points; S4. Using the set of anchor points as training samples, construct a generative adversarial network, introduce an optimal transmission regularization term into the loss function of the generator, and achieve distribution alignment between the source batch and the target batch of data by minimizing the loss function; S5. Input the source batch data to be corrected into the trained generator, and output the corrected gene expression data or low-dimensional embedding representation.

[0041] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0042] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0043] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0044] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0045] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for single-cell transcriptome data integration based on explicit decoupling and optimal transport, the method comprising: Includes the following steps: Multiple batches of single-cell gene expression matrices were obtained, and the input data was obtained after preprocessing. A deep residual autoencoder is constructed to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. Based on the biological feature components, an approximate nearest neighbor search algorithm is used to screen the nearest neighbor pairs between different batches to form a set of highly reliable anchor points. Using the set of anchor points as training samples, a generative adversarial network is constructed. An optimal transmission regularization term is introduced into the loss function of the generator. By minimizing this loss function, the distribution alignment of the source batch and the target batch of data is achieved. Input the source batch data to be corrected into the trained generator, and output the corrected gene expression data or low-dimensional embedding representation.

2. The method according to claim 1, wherein, The comprehensive loss function is: in, To rebuild the losses, For loss of content consistency, Loss due to batch classification; The weight hyperparameters for reconstructing the loss, The weight hyperparameters for content consistency loss are... The weighted hyperparameters for batch classification loss; For input data The mathematical expectation, This represents the expression vector of the source batch of cells. For the encoder to input The mapping result shows that the vector is explicitly segmented into biological feature components. and batch noise components ; For input data and random batch labels The joint mathematical expectation, As a new biological characteristic component; To measure batch noise components conduct Activation operation, For cross-entropy loss calculation, measure Predicted batch probability distribution and actual batch label The differences.

3. The method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission according to claim 1, characterized in that, The loss function of the generator is: in, It is a hyperparameter that controls the strength of regularization. This represents the expression vector of the source batch of cells. For generator functions; For the discriminator function, The distribution of target reference batch cells in the biological potential space. This represents the distribution of source batch cells in the biological latent space after generator correction. For the source batch data distribution P B Randomly selected samples For the sample The average value is taken from the results processed by the generator and discriminator. This is the optimal transmission loss.

4. The method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission according to claim 3, characterized in that, The optimal transmission loss The generated cells are obtained through the Sinkhorn algorithm with entropy regularization: for a batch of source cells sampled from nearest neighbor pairs, the generated cells are obtained through the generator. and its corresponding target reference cells The biological latent representations of these encoders trained in the first stage are extracted. and ; Then calculate the cost matrix. The Sinkhorn algorithm iteratively solves an entropy-regularized optimal transmission problem, obtains a soft-assignment matrix, and calculates the minimum transmission cost as the minimum value. .

5. The method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission according to claim 1, characterized in that, The depth residual autoencoder will input Mapping to latent vectors ,Will Explicitly split into front biological components of vitae and after Batch noise components of dimensional ,Right now ; decoder The combined features are received and reconstructed, and the reconstructed expression spectrum is output. .

6. The method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission according to claim 1, characterized in that, The high-confidence anchor set is obtained by: normalizing the biological feature components; using the Annoy algorithm to construct an approximate nearest neighbor index, finding the k nearest neighbors of cells in other batches for each batch; and identifying cell A as a mutual nearest neighbor pair if and only if cell A is in cell B's k nearest neighbor list and cell B is also in cell A's k nearest neighbor list. All mutual nearest neighbor pairs constitute the high-confidence anchor set M.

7. The method for integrating single-cell transcriptome data based on explicit decoupling and optimal transmission according to claim 1, characterized in that, The generative adversarial network includes a generator and a discriminator. The generator uses a fully connected layer structure with residual connections to map the expression profile of the source batch of cells to the distribution of the reference batch, generating a corrected expression profile. The discriminator uses a discrimination criterion based on Wasserstein distance and adds gradient penalty GP to achieve stable training, that is, it adopts the WGAN-GP framework to distinguish between the real reference batch of cells and the generator's corrected expression profile.

8. A single-cell transcriptome data integration system based on explicit decoupling and optimal transmission, characterized in that, include: The data preprocessing module obtains single-cell gene expression matrices from multiple batches and preprocesses them to obtain input data. An explicit decoupling module constructs a deep residual autoencoder to map the input data into a structured latent representation. The latent representation is explicitly segmented into biological feature components and batch noise components. The two types of components are completely separated by minimizing the comprehensive loss function. The anchor point identification module, based on the biological feature components, uses an approximate nearest neighbor search algorithm to filter mutually nearest neighbor pairs between different batches to form a high-confidence anchor point set; The adversarial correction module uses the anchor set as training samples to construct a generative adversarial network. It introduces an optimal transmission regularization term into the generator's loss function and achieves data distribution alignment between the source batch and the target batch by minimizing the loss function. The data output module takes the source batch data to be corrected and inputs it into the trained generator, then outputs the corrected gene expression data or low-dimensional embedding representation.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the single-cell transcriptome data integration method based on explicit decoupling and optimal transmission as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a single-cell transcriptome data integration method based on explicit decoupling and optimal transmission as described in any one of claims 1-7.

Images