Single cell cross-omics multi-modal prediction method

Abstract

The invention discloses a single cell cross-omics multi-modal prediction method. The method is specifically implemented according to the following steps: step 1, data collection and preprocessing; 2, training a deep learning model; step 3, extracting biological characteristics from RNA and ATAC data through the trained model, and realizing data mapping between the two modals; 4, performing automatic prediction and result evaluation; 5, carrying out model optimization and iterative improvement; and step 6, clinical application and verification. According to the method, a deep learning framework is utilized, and data from different omics levels are effectively integrated, so that deep understanding of a biological system and accurate prediction of disease states are realized. The method not only improves the analysis capability of complex biological phenomena, but also provides a new thought and tool for personalized medical treatment and precise treatment, can be widely applied to the fields of tumor classification, disease progress monitoring, drug response prediction and the like, and provides a scientific basis for clinical decision making.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioinformatics and multi-omics research methods, and specifically relates to a single-cell cross-omics multimodal prediction method. Background Art

[0002] Single-cell resolution sequencing technologies, such as single-cell RNA sequencing (scRNA-seq), high-throughput sequencing of chromatin accessibility (scATAC-seq), and related technologies for methylation and protein abundance, have greatly promoted biomedical research. scRNA-seq can depict the transcriptional characteristics of individual cells in detail and identify different cell subpopulations, while scATAC-seq reveals the heterogeneity of chromatin. These single-cell data are of great significance in deeply understanding cell-to-cell differences and biological processes, especially when studying the complexity of the immune system and the dynamics of tumor development. Studying the diversity of the tumor microenvironment through single-cell sequencing technology is crucial for developing personalized precision treatment plans. However, these cell sequencing methods are somewhat invasive, and cells will be damaged during the processing, causing changes in their internal structure and function, thereby affecting the accurate capture of molecular features, which also makes consistency between different modalities difficult.

[0003] Some existing methods for integrating single-cell data are mainly aimed at eliminating batch effects. For example, maximum mean difference manifold alignment is an unsupervised manifold alignment algorithm specifically for single-cell multi-omics datasets. However, its effect will be significantly reduced when there are significant differences or nonlinear correlations between different modalities. Cross-modal prediction methods for single-cell multi-omics data are still relatively scarce. Although the scMM method can achieve cross-omics generation, its performance between different datasets is not satisfactory. The BABEL method supports cross-modal prediction of multi-omics profiles, but has limited effect in the conversion from transcriptome to chromatin accessibility. The sciCAN model uses a shared encoder to map data of different modalities to a latent space and distinguishes the two modalities through a discriminator. This approach can simplify the model, but may lead to information loss when the modality difference is large, thus affecting the accuracy of the final prediction. Therefore, current single-cell multi-omics integration and prediction technologies still face many challenges in dealing with modality differences, mining potential common factors, and improving prediction accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a single-cell cross-omics multimodal prediction method, which is trained by matching modality data to achieve cross-modal prediction of the model on single-cell sequencing data.

[0005] The technical solution adopted by the present invention is a single-cell cross-omics multimodal prediction method, which is specifically implemented in the following steps:

[0006] Step 1, data collection and preprocessing;

[0007] Step 2: Deep learning model training;

[0008] Step 3: Extract biological features from RNA and ATAC data through the trained model and realize data mapping between the two modalities;

[0009] Step 4: Automated prediction and result evaluation;

[0010] Step 5: Model optimization and iterative improvement;

[0011] Step 6: Clinical application and verification.

[0012] The present invention is also characterized in that:

[0013] In step 1, ATAC and RNA paired sequencing data are collected, covering cell samples of different types and states, and the data are preprocessed, including denoising, normalization, image enhancement, and size standardization, to improve the training effect of the deep learning model.

[0014] Step 1 is implemented as follows:

[0015] Step 1.1, collect ATAC and RNA paired sequencing data, which cover various types of tissue cells, including tumor tissue, immune tissue, neural tissue, muscle tissue and their corresponding single-cell expression profiles;

[0016] Step 1.2: Preprocess the collected ATAC and RNA sequencing data. First, high-throughput single-cell sequencing technology is used to obtain the expression profile data of single cells. Next, denoising is performed, and the principal component analysis algorithm is used to reduce background noise, thereby enhancing signal quality. The raw data is then normalized using the z-score standardization method to make the data comparable at the same level. In addition, missing values ​​are filled in for the expression of specific genes to ensure data integrity.

[0017] In step 1.3, image enhancement processing will be applied to the standardized single-cell data. The enhancement processing includes increasing the contrast, brightness and sharpening of the image to highlight important biomarkers. Finally, size standardization is performed so that all input data have the same dimension to accommodate subsequent deep learning model training.

[0018] In step 2, based on the preprocessed data, a multimodal data integration model was built using a deep learning framework. The model was trained using matched ATAC and RNA sequencing data. The model parameters were adjusted using the back propagation algorithm to minimize the reconstruction error and classification error. The validation set was used to tune the model and select the optimal hyperparameter configuration to improve model performance.

[0019] Step 2 is implemented as follows:

[0020] Step 2.1, divide the processed data set into training set and test set in the ratio of 8:2, set the batch size to 32, select ADAM optimizer for model training, use Relu activation function to increase the nonlinear ability of the model, introduce two dropout layers in the network architecture to prevent overfitting, that is, randomly discard some neurons during training to ensure the generalization ability of the model, and select loss function to evaluate model performance;

[0021] Step 2.2: During the training process, monitor the performance indicators of the model. When the loss function converges or reaches the set number of iterations, the training process terminates.

[0022] Step 2.3, save the trained weight parameters and model structure for subsequent verification and application, to ensure that users can reproduce the results under different experimental conditions;

[0023] In step 2.4, the trained weight parameters are used to evaluate the model on the test set, and statistical indicators such as the accuracy, sensitivity, specificity, and AUC area are calculated. Finally, the model parameters are further optimized or the network structure is adjusted based on the evaluation results to improve the predictive ability of the model.

[0024] In step 4, automated predictions are performed on new single-cell multi-omics data, and the trained model is used to infer cross-modal biological processes and predict cell states. The prediction results are evaluated and compared with the experimental results for consistency, and performance indicators are calculated to verify the reliability of the model.

[0025] In step 5, based on the feedback between the predicted results and the actual observed data, the model is continuously improved and optimized, and the training set is updated and the model is retrained using new experimental data to enhance its generalization ability in various cell types and conditions.

[0026] In step 6, the method is applied to actual single-cell research projects and compared with other methods to verify its practicality and effectiveness in biomedical research.

[0027] The beneficial effects of the present invention are

[0028] 1) Accurate detection and analysis: The present invention uses single-cell sequencing technology to efficiently capture the genome, transcriptome and proteome information of each cell, reveal the heterogeneity between cells, and help researchers identify specific cell types and states that play a key role in the development of diseases. Through the integrated analysis of multi-omics data, it is possible to systematically understand how cells influence each other at different omics levels, thereby more comprehensively understanding the regulatory mechanisms in biological processes;

[0029] 2) Promotion of personalized medicine: The present invention provides precise omics information that covers the dynamic changes of cell functions and phenotypes, allowing doctors to develop personalized treatment plans based on the unique biological characteristics of patients, improve efficacy, and reduce unnecessary side effects;

[0030] 3) Early disease monitoring: The single-cell cross-omics translation technology of the present invention has high sensitivity and can identify potential biomarkers in the early stages of the disease. It is particularly important for rapidly progressing diseases such as cancer and neurodegenerative diseases, and can provide support for early intervention. By continuously monitoring the omics changes of cells, doctors can adjust treatment plans in time to respond to the progression or recurrence of the disease, which improves the clinical response speed compared to previous detection methods;

[0031] 4) Multi-dimensional data fusion: The present invention uses a data fusion algorithm to integrate omics data from different sources to form a comprehensive bioinformatics map. This multi-dimensional data processing capability not only improves the efficiency of data utilization, but also enhances the understanding of complex biological systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the process of the single-cell cross-omics multimodal prediction method of the present invention;

[0033] Figure 2 This is a flowchart of the preprocessing in the single-cell cross-omics multimodal prediction method of the present invention. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] The present invention provides a single-cell cross-omics multimodal prediction method, such as Figure 1 As shown, the specific implementation steps are as follows:

[0037] Step 1, data collection and preprocessing;

[0038] Step 2: Deep learning model training;

[0039] Step 3: Use the trained model to extract biological features from RNA and ATAC data and achieve data mapping between the two modalities; this not only helps understand the association between the modalities, but also provides rich feature information for downstream analysis;

[0040] Step 4: Automated prediction and result evaluation;

[0041] Step 5: Model optimization and iterative improvement;

[0042] Step 6: Clinical application and verification.

[0043] Example 2

[0044] The single-cell cross-omics multimodal prediction method, in which in step 1, ATAC and RNA paired sequencing data published by the sequencing organization 10XGenomics are collected, covering cell samples of different types and states, and the data are preprocessed, including denoising, normalization, image enhancement, and size standardization, to improve the training effect of the deep learning model. The quality and consistency of the input data are ensured, so that the model can learn features more effectively.

[0045] Example 3

[0046] Single-cell cross-omics multimodal prediction methods, such as Figure 2 As shown, step 1 is specifically implemented according to the following steps:

[0047] Step 1.1, collect ATAC and RNA paired sequencing data published by 10X Genomics. These data cover various types of tissue cells, including tumor tissue, immune tissue, neural tissue, muscle tissue and their corresponding single-cell expression profiles. Ensure that the samples are representative so as to effectively capture key biological information;

[0048] Step 1.2, pre-process the collected ATAC and RNA sequencing data. First, high-throughput single-cell sequencing technology is used to obtain the expression profile data of single cells. Next, denoising is performed, and the principal component analysis (PCA) algorithm is used to reduce background noise and enhance signal quality. Then, the raw data is normalized and the z-score normalization method is used to make the data comparable at the same level. In addition, missing values ​​are filled in for the expression of specific genes to ensure data integrity;

[0049] In step 1.3, image enhancement processing will be applied to the standardized single-cell data. The enhancement processing includes increasing the contrast, brightness and sharpening of the image to highlight important biomarkers. This process ensures that the data input into the deep learning model has high quality and consistency. Finally, size standardization is performed so that all input data have the same dimension to accommodate subsequent deep learning model training.

[0050] Example 4

[0051] Single-cell cross-omics multimodal prediction method, in which in step 2, a multimodal data integration model is constructed based on the preprocessed data using a deep learning framework (variational autoencoder). The model is trained using matched ATAC and RNA sequencing data, and the model parameters are adjusted using the back propagation algorithm to minimize the reconstruction error and classification error. The validation set is used to tune the model and select the best hyperparameter configuration, such as learning rate, batch size, and network structure, to improve model performance.

[0052] Example 5

[0053] The single-cell cross-omics multimodal prediction method, wherein step 2 is specifically implemented according to the following steps:

[0054] Step 2.1, divide the processed data set into training set and test set in a ratio of 8:2. Set the batch size to 32, select ADAM optimizer for model training, and use Relu activation function to increase the nonlinear ability of the model. In the network architecture, introduce two dropout layers to prevent overfitting, that is, randomly discard some neurons during training to ensure the generalization ability of the model. Select an appropriate loss function (reconstruction loss) to evaluate the model performance;

[0055] Step 2.2: During the training process, monitor the performance indicators of the model. When the loss function converges or reaches the set number of iterations, the training process terminates. At this time, the key features of the single-cell multimodal data are extracted. These features can reflect the transcriptome and epigenomic information of the cell, providing a basis for subsequent analysis.

[0056] RNA sequencing prediction loss function L for deep learning networks (RNA) for:

[0057]

[0058] The loss is a weighted sum of the reconstruction loss and the Kullback-Leibler (KL) divergence term, where y represents the original RNA input, t y represents the embedding representation of y, d y Indicates sequencing depth, q y (t y |y,d y ) represents the encoder, p y (y|t y ,d y ) represents a decoder;

[0059] Similarly, the ATAC sequencing prediction loss function L of the deep learning network (ATAC) for:

[0060]

[0061] Where x represents the original ATAC input, t x represents the embedding representation of x, d x Indicates sequencing depth, q x (t x |x,d x ) represents the encoder, p x (x|t x ,d x ) represents a decoder;

[0062] Step 2.3, save the trained weight parameters and model structure for subsequent verification and application, to ensure that users can reproduce the results under different experimental conditions;

[0063] Step 2.4, use the trained weight parameters to evaluate on the test set, and calculate the model's accuracy, sensitivity, specificity, AUC area and other statistical indicators. These indicators can comprehensively evaluate the model performance and provide a basis for clinical application. Finally, based on the evaluation results, further optimize the model parameters or adjust the network structure to improve the model's predictive ability.

[0064] Example 6

[0065] Single-cell cross-omics multimodal prediction method, in which in step 4, new single-cell multi-omics data are automatically predicted, and the trained model is used to perform cross-modal biological process inference and cell state prediction. The prediction results are evaluated, compared with the experimental results, and performance indicators (accuracy, sensitivity and specificity) are calculated to verify the reliability of the model.

[0066] In step 5, the model is continuously improved and optimized based on the feedback between the predicted results and the actual observed data. New experimental data is used to update the training set and retrain the model to improve its generalization ability in various cell types and conditions.

[0067] In step 6, the developed method is applied to actual single-cell research projects, and its practicality and effectiveness in biomedical research are verified by comparison with other methods. In addition, visualization tools are provided to help researchers understand the correspondence between model output and biological meaning, and promote the transformation of scientific research results.

[0068] The single-cell cross-omics multimodal prediction method of the present invention optimizes the single-cell cross-omics process in experimental design, applies automation technology and high-throughput sequencing, and greatly reduces the manpower and material costs required for the experiment. In addition, the fast data analysis pipeline also shortens the cycle from sample collection to result acquisition, making scientific research and clinical applications more efficient. This cost-reducing and efficiency-increasing feature is particularly important for large-scale screening and high-throughput testing, and can still conduct in-depth scientific research under limited resources. The present invention is not only suitable for basic scientific research, but also has good clinical transformation potential. By simplifying the process from basic research to clinical application, it enables cutting-edge technology to serve clinical practice more quickly and improve the health of patients. In addition, the present invention also provides a new perspective for clinical research, can generate a large amount of biological data of cell subpopulations, and help discover new therapeutic targets and biomarkers.

Claims

1. Single-cell cross-omics multimodal prediction method, characterized by: Please follow the steps below to implement: Step 1, data collection and preprocessing; Step 2: Deep learning model training; Step 3: Extract biological features from RNA and ATAC data through the trained model and realize data mapping between the two modalities; Step 4: Automated prediction and result evaluation; Step 5: Model optimization and iterative improvement; Step 6: Clinical application and verification.

2. The single-cell cross-omics multimodal prediction method according to claim 1, characterized in that: In step 1, ATAC and RNA paired sequencing data are collected, covering cell samples of different types and states, and the data are preprocessed, including denoising, normalization, image enhancement, and size standardization, to improve the training effect of the deep learning model.

3. The single-cell cross-omics multimodal prediction method according to claim 2, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1, collect ATAC and RNA paired sequencing data, which cover various types of tissue cells, including tumor tissue, immune tissue, neural tissue, muscle tissue and their corresponding single-cell expression profiles; Step 1.2: Preprocess the collected ATAC and RNA sequencing data. First, high-throughput single-cell sequencing technology is used to obtain the expression profile data of single cells. Next, denoising is performed, and the principal component analysis algorithm is used to reduce background noise, thereby enhancing signal quality. The raw data is then normalized using the z-score standardization method to make the data comparable at the same level. In addition, missing values ​​are filled in for the expression of specific genes to ensure data integrity. In step 1.3, image enhancement processing will be applied to the standardized single-cell data. The enhancement processing includes increasing the contrast, brightness and sharpening of the image to highlight important biomarkers. Finally, size standardization is performed so that all input data have the same dimension to accommodate subsequent deep learning model training.

4. The single-cell cross-omics multimodal prediction method according to claim 1, characterized in that: In step 2, based on the preprocessed data, a multimodal data integration model is constructed using a deep learning framework, the model is trained using matched ATAC and RNA sequencing data, the model parameters are adjusted using a back propagation algorithm to minimize reconstruction error and classification error, and the model is tuned using a validation set and the optimal hyperparameter configuration is selected to improve model performance.

5. The single-cell cross-omics multimodal prediction method according to claim 4, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, divide the processed data set into training set and test set in the ratio of 8:2, set the batch size to 32, select ADAM optimizer for model training, use Relu activation function to increase the nonlinear ability of the model, introduce two dropout layers in the network architecture to prevent overfitting, that is, randomly discard some neurons during training to ensure the generalization ability of the model, and select loss function to evaluate model performance; Step 2.2: During the training process, monitor the performance indicators of the model. When the loss function converges or reaches the set number of iterations, the training process terminates. Step 2.3, save the trained weight parameters and model structure for subsequent verification and application, to ensure that users can reproduce the results under different experimental conditions; In step 2.4, the trained weight parameters are used to evaluate the model on the test set, and the accuracy, sensitivity, specificity, and AUC statistical indicators of the model are calculated. Finally, the model parameters are further optimized or the network structure is adjusted based on the evaluation results to improve the predictive ability of the model.

6. The single-cell cross-omics multimodal prediction method according to claim 1, characterized in that: In step 4, the new single-cell multi-omics data are automatically predicted, and the trained model is used to perform cross-modal biological process inference and cell state prediction. The prediction results are evaluated and compared with the experimental results for consistency, and performance indicators are calculated to verify the reliability of the model.

7. The single-cell cross-omics multimodal prediction method according to claim 1, characterized in that: In step 5, the model is continuously improved and optimized based on the feedback between the predicted results and the actually observed data, and the training set is updated and the model is retrained using new experimental data to enhance its generalization ability in various cell types and conditions.

8. The single-cell cross-omics multimodal prediction method according to claim 1, characterized in that: In step 6, the method is applied to an actual single-cell research project, and its practicality and effectiveness in biomedical research are verified by comparison with other methods.

Images