Colorectal cancer target multi-source sequencing data co-processing method based on federal learning

Abstract

The invention provides a colorectal cancer target multi-source sequencing data co-processing method based on federated learning, and relates to the technical field of data processing.The method comprises the steps that original sequencing data stored in a local database of each collaborator is reconstructed through a sequencing translator based on a generative adversarial network; according to a preset rule, selecting part of the reconstructed sequencing data, and mapping the reconstructed sequencing data to a public target data space; in a local training stage, original sequencing data stored in a local database and part of reconstructed sequencing data in a target data space are utilized, and a local model is trained through public parameter aggregation optimization; in the global training stage, personalized aggregation can be carried out on the model according to contribution degrees of different clients through a federal attention message passing algorithm, and the global model is trained; outputting a detection result through the trained global model; and integrating a detection result and a local interpretable analysis result, and generating a target joint evidence report.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning. Background Technology

[0002] Colorectal cancer (CRC) is one of the most common cancers worldwide, with consistently high incidence and mortality rates globally. To facilitate early diagnosis and precision treatment, scientists utilize various technologies to study the genome of colorectal cancer. Multi-source sequencing data of target sites has become a crucial basis for analyzing colorectal cancer gene variations, diagnostic biomarkers, and therapeutic targets.

[0003] Multi-source sequencing data for colorectal cancer targets involves patient data from different hospitals, medical institutions, or research organizations. The amount of data from different medical institutions is unbalanced; some institutions may have more cancer patient samples, while others may have less. Furthermore, multi-source sequencing data for colorectal cancer targets contains personal privacy information, especially genomic data, requiring protection of patient privacy. This private data cannot be directly shared between different medical institutions. This data imbalance can cause data analysis models to be biased towards certain large sample datasets, affecting the stability and accuracy of the data analysis models. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning, which can solve the technical problem that the imbalance of data quantity in the prior art can lead to the bias of data analysis models towards certain large sample data, affecting the stability and accuracy of data analysis models.

[0005] A first aspect of this invention proposes a collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning, comprising:

[0006] S1: Obtain the raw sequencing data stored in the local databases of each collaborator;

[0007] S2: The raw sequencing data stored in the local databases of each of the collaborators are reconstructed using a sequencing translator based on a generative adversarial network;

[0008] S3: Select a portion of the reconstructed sequencing data according to preset rules and map it to a common target data space;

[0009] S4: During the local training phase, the local model is trained by using the raw sequencing data stored in the local database and the partially reconstructed sequencing data in the target data space through common parameter aggregation optimization.

[0010] S5: During the global training phase, the local model is aggregated in a personalized manner based on the contribution of different clients using the federated attention message passing algorithm, and the global model is trained accordingly.

[0011] S6: Output the detection results using the completed global model;

[0012] S7: Perform federated feature selection and interpretability analysis on the detection results, and output the local interpretability analysis results;

[0013] S8: Integrate the detection results with the locally interpretable analysis results to generate a joint evidence report for the target.

[0014] A second aspect of this invention proposes a collaborative processing system for multi-source sequencing data of colorectal cancer targets based on federated learning, comprising: a processor and a memory;

[0015] The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the method for collaborative processing of multi-source sequencing data of colorectal cancer targets based on federated learning as described in the first aspect.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0017] (1) In this embodiment of the invention, the original sequencing data of each collaborator is transformed into reconstructed data by generating adversarial networks. In this way, even if the data itself is transformed and reconstructed, there is no need to exchange the original data, thereby avoiding the risk of privacy leakage.

[0018] (2) In this embodiment of the invention, by means of federated learning, the raw sequencing data stored in the local databases of each collaborator can be fully utilized to train the data analysis model, ensuring that the data analysis model is not biased towards a specific data source, avoiding model performance deviation caused by data imbalance, and improving the stability and accuracy of the data analysis model. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart illustrating a collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning, provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a collaborative processing system for multi-source sequencing data of colorectal cancer targets based on federated learning, provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] The following description, in conjunction with the accompanying drawings, details the collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning provided by the present invention through specific embodiments and application scenarios.

[0024] Reference manual attached Figure 1 The diagram illustrates a flowchart of a collaborative processing method for multi-source sequencing data of colorectal cancer targets based on federated learning, provided by an embodiment of the present invention.

[0025] This invention provides a method for collaborative processing of multi-source sequencing data for colorectal cancer targets based on federated learning. The multi-source sequencing data includes at least two of the following: WES sequencing data, WGS sequencing data, targeted panel sequencing data, RNA-seq sequencing data, single-cell RNA-seq sequencing data, and ctDNA sequencing data. The collaborative processing method may include the following steps:

[0026] S1: Obtain the raw sequencing data stored in the local databases of each collaborator.

[0027] S2: The raw sequencing data stored in the local databases of each collaborator is reconstructed using a sequencing translator based on a generative adversarial network.

[0028] In one possible implementation, S2 specifically includes sub-steps S201 to S203:

[0029] S201: Define the common target data space.

[0030] This invention proposes two methods for selecting the target data space, each with its own advantages.

[0031] In one possible implementation, S201 specifically includes sub-steps S2011 to S2013:

[0032] S2011: Sparse reconstruction of raw sequencing data stored in the local databases of each collaborator is performed using a sparse autoencoder.

[0033] Among them, sparse autoencoders are a variant of autoencoders whose goal is to learn a low-dimensional representation of data. They can effectively capture key features of input data during the learning process, thereby achieving more efficient data compression and denoising.

[0034] S2012: Calculate the complexity of each collaborator's local database using the L2 distance between the sparsely reconstructed sequencing data and the original sequencing data.

[0035] S2013: Select the collaborator's local database with the lowest complexity and define the target data space.

[0036] In this embodiment of the invention, selecting the collaborator's local database with the lowest complexity to define the target data space can ensure data consistency and alignment, reduce data redundancy and noise interference, and thus improve the accuracy and efficiency of data analysis.

[0037] In another possible implementation, S201 specifically includes sub-steps S2011 and S2013:

[0038] S2011: Learns domain-invariant representations from the intersection and union features of raw sequencing data stored in the local databases of all collaborators using an adversarial domain-invariant encoder.

[0039] Adversarial domain-invariant encoders (ADIEs) are models that learn domain-invariant features through adversarial training (similar to generative adversarial networks). They extract representations of the input data through an encoder and then use an adversarial network to ensure that these representations remain invariant across different domains (e.g., samples from different datasets or distributions). This approach is widely used in cross-domain transfer learning, such as when transferring knowledge between datasets from different sources, to ensure that the model learns domain-invariant features rather than relying solely on dataset-specific features.

[0040] S2012: Define the target data space based on the domain-invariant representation.

[0041] Specifically, domain-invariant representations eliminate differences in sequencing platforms, batch effects, or sample distributions among collaborators' data, resulting in similar feature distributions across different data sources. Therefore, this representation space can be directly used as the target data space, meaning that all collaborators' data are mapped to this unified representation space during subsequent reconstruction, alignment, and training. This defined target data space maximizes the preservation of biological information from each data source while enabling seamless integration of cross-institutional data, providing a unified data foundation for collaborative processing under federated learning.

[0042] The applicable scenarios for defining the target data space using either the collaborator's local database with the lowest complexity or using a domain-invariant representation are as follows: When the data from different collaborators is highly homogeneous (e.g., sequencing platforms and batch differences are small), and the differences between different data sources mainly lie in data volume and noise levels, defining the target data space using the collaborator's local database with the lowest complexity is suitable. Conversely, when the data from different collaborators is highly heterogeneous (e.g., different hospitals using different sequencing platforms, different batch effects, or even different population samples), leading to significant biases in data distribution, defining the target data space using a domain-invariant representation can effectively eliminate platform differences and distribution biases.

[0043] In this embodiment of the invention, it is possible to ensure that the model shares the same representation across datasets from different sources, thereby improving cross-domain transfer learning capabilities and the model's generalization performance. The target data space is defined based on the learned domain-invariant representation, enabling data from different medical institutions to be analyzed and collaboratively processed under the same standard. This improves the consistency and accuracy of data integration, providing strong support for subsequent joint modeling and accurate prediction.

[0044] S202: For each collaborator, by generating adversarial networks, important features in the original sequencing data are effectively captured, and the original sequencing data is reconstructed to obtain reconstructed sequencing data.

[0045] Generative Adversarial Networks (GANs) are a type of deep learning model that consists of a generator and a discriminator. The generator aims to generate realistic fake data by learning the data distribution, while the discriminator is responsible for distinguishing between real data and generated fake data. Both train adversarially, continuously optimizing their strategies.

[0046] In one possible implementation, privacy-preserving noise is added to the gradient of the discriminator:

[0047]

[0048]

[0049] Where N represents Gaussian noise, This indicates that the mean is 0 and the variance is . Gaussian noise, σ n c represents the variance of the fundamental noise. g Let I represent the gradient penalty constant, q represent the identity matrix, and n represent the sampling frequency. dδ represents the update frequency of the discriminator, log represents the logarithmic function to the base 10, δ represents the failure probability, and ε represents the privacy budget parameter.

[0050] In this embodiment of the invention, adding Gaussian noise to the gradient of the discriminator effectively achieves differential privacy protection during training. While ensuring the model can learn data distribution characteristics, it prevents the discriminator from leaking sensitive information about individual samples (such as genomic variant sites) from the gradient, reducing privacy risks such as member inference attacks. Furthermore, by adjusting the gradient penalty constant and privacy budget parameters, a balance can be achieved between privacy protection and model accuracy, thereby enabling secure and reliable collaborative training using multi-source sequencing data.

[0051] In this embodiment of the invention, by reconstructing the raw sequencing data stored in the local databases of each collaborator using a sequencing translator based on a generative adversarial network (GAN), data privacy protection and cross-institutional data alignment can be effectively achieved. This method ensures that the original data remains locally, while extracting and reconstructing key features from the data through the GAN, guaranteeing that the reconstructed data retains a distribution similar to the original data. This not only effectively alleviates the problem of data heterogeneity, ensuring collaborative analysis of data from different sources in a unified data space, but also enables effective data sharing and model training while guaranteeing privacy.

[0052] S3: Select a portion of the reconstructed sequencing data according to preset rules and map it to a common target data space.

[0053] Optionally, the preset rules specifically include: in the feature space, the feature similarity between the features of the selected reconstructed sequencing data and the features of the original sequencing data stored in the local database is greater than a first preset similarity. Furthermore, in the data space, the data similarity between the selected reconstructed sequencing data and the original sequencing data stored in the local database is less than a second preset similarity.

[0054] Those skilled in the art can set the magnitude of the first preset similarity and the second preset similarity according to the actual situation, and the present invention does not limit them.

[0055] In this embodiment of the invention, the similarity requirement in the feature space ensures that the selected reconstructed data maintains consistency with the original data in key features, thereby preserving the biological significance of the data. Meanwhile, the similarity constraint in the data space prevents over-reliance on redundant information in the original data, avoiding overfitting or noise interference. This approach effectively improves the quality and usability of the reconstructed data, while ensuring that collaborators can perform efficient model training and optimization in a unified data space during data sharing and collaborative training.

[0056] S4: During the local training phase, the local model is trained by using the raw sequencing data stored in the local database and some reconstructed sequencing data in the target data space through common parameter aggregation optimization.

[0057] In one possible implementation, S4 specifically includes sub-steps S401 to S404:

[0058] S401: Obtain the common parameter space of the local and global models of each collaborator.

[0059] S402: Based on the global model's detection results of the sequencing data, calculate the importance of each common parameter in the common parameter space:

[0060]

[0061] Among them, w u L represents the importance parameter of the u-th common parameter. g θ represents the loss function during global model training. u Let represent the u-th common parameter in the common parameter space, and ∂ represent the partial derivative operation.

[0062] It's important to note the significance of calculating common parameters—specifically, evaluating the contribution of each parameter to the global model's training loss using partial derivatives. This helps the model understand which parameters are crucial for global prediction performance. This facilitates dynamic parameter updates, ensuring that key parameters with significant impact on the global model are not easily altered during local training, thus maintaining model stability and efficient convergence. This step provides crucial information for subsequent regularization, enabling the model to be adjusted and optimized more effectively.

[0063] S403: Based on the importance of each common parameter, perform common parameter aggregation optimization to construct the loss function for local training:

[0064]

[0065] Among them, L all L represents the loss function. local λ1 represents the loss function used for local model training when using raw sequencing data stored in the local database and partially reconstructed sequencing data from the target data space, where λ1 represents the first regularization parameter. This represents the u-th common parameter in the global model. Let represent the u-th common parameter in the local model, and n represent the total number of common parameters in the common parameter space.

[0066] It's important to note that by introducing an importance matrix and weighting the model training according to the importance of each common parameter, the consistency between the local and global models can be effectively constrained. This step, by applying regularization to the differences in common parameters, ensures that the local model prioritizes maintaining the key parameters of the global model during optimization, reducing unnecessary bias or over-updates, thereby improving the accuracy and stability of the local model. Simultaneously, this aggregation optimization method can handle data heterogeneity issues in multi-source data scenarios, improving the reliability of cross-institutional collaboration.

[0067] S404: Train the local model based on the loss function used during local training.

[0068]

[0069] in, This represents the model parameters at the (t+1)th iteration. Let represent the model parameters at the t-th iteration, and η1 represent the first learning rate.

[0070] It should be noted that updating model parameters through gradient descent can optimize the performance of the local model while maintaining consistency between local and global parameters, thus ensuring a balance between the local and global models.

[0071] In this embodiment of the invention, the local model is trained through common parameter aggregation optimization. On the one hand, the original data ensures that the model can capture the unique features of the institution. On the other hand, the reconstructed data and common parameter constraints keep the model aligned with the global model, avoiding bias caused by excessive data heterogeneity. Simultaneously, by introducing importance-weighted regularization terms to the common parameters, the stability of key parameters is preferentially preserved, mitigating "catastrophic forgetting" and improving the model's generalization ability and convergence performance. Ultimately, this optimization method not only improves the accuracy and robustness of the local model but also lays a more solid foundation for subsequent global model aggregation.

[0072] S5: During the global training phase, the federated attention message passing algorithm can be used to aggregate local models in a personalized way based on the contribution of different clients, and then train the global model.

[0073] Federated Attention Message Passing Algorithm (FedAMP) is an improved federated learning method designed for personalized model aggregation. FedAMP introduces an attention mechanism to perform weighted aggregation based on the similarity between different client models and the global model. Specifically, it evaluates the distance between client models and the target model, using this similarity to dynamically adjust the contribution of each client, thereby achieving personalized model updates. FedAMP effectively handles heterogeneity across institutions and data sources, improving the accuracy and adaptability of the global model. This invention improves upon the Federated Attention Message Passing Algorithm.

[0074] In one possible implementation, S5 specifically includes sub-steps S501 to S503:

[0075] S501: To reduce the loss function of each collaborator's local model and improve the effectiveness of collaboration among collaborators, an optimization objective function for the global model is constructed:

[0076]

[0077]

[0078] Where min represents minimization, G represents the objective function, and A represents the global set of model parameters. a i Let a represent the local model parameters of the i-th collaborator. j Let L represent the local model parameters of the j-th collaborator, m represent the total number of collaborators, and L represent the total number of collaborators. i Let λi represent the loss function of the local model of the i-th collaborator, λ2 represent the second regularization parameter, and B represent the attention-inducing function that evaluates the effectiveness of collaboration among collaborators. Let e ​​represent the Euclidean distance between the local model parameters of the i-th collaborator and the j-th collaborator, e represent the natural constant, and σ represent the similarity metric sensitivity parameter.

[0079] It should be noted that by constructing a global model to optimize the objective function, and combining the local model loss and collaboration effectiveness of each collaborator, the global model can optimize the cross-client collaboration effect while minimizing local loss.

[0080] S502: Employs alternating optimization, based on the attention-induced function in the objective function, and updates the proximal terms through gradient descent.

[0081]

[0082] Among them, U t U represents the proximal term at the t-th iteration. t-1Let ηt represent the proximal term at the (t-1)th iteration, and η2 represent the second learning rate. Represents the gradient operator, A t-1 This represents the set of global model parameters at the (t-1)th iteration.

[0083] It should be noted that updating proximal terms through gradient descent can more accurately guide the optimization direction of the global model, avoid excessive updates at once, and ensure the stability and gradualness of the training process.

[0084] S503: Using the approximate point algorithm, based on the proximal terms, update the global model parameters:

[0085]

[0086] Among them, A t Let represent the set of global model parameters at the t-th iteration, arg min represents the function to take its minimum value, and λ3 represents the third regularization parameter.

[0087] It should be noted that updating the global model parameters through the approximation point algorithm ensures that the global model takes into account the contributions of all clients in each iteration and balances local data characteristics with global consistency.

[0088] In this embodiment of the invention, a federated attention message passing algorithm is introduced during the global training phase. This algorithm uses an attention mechanism to perform personalized aggregation based on the similarity and contribution of different clients to the global model, effectively addressing the problem of data heterogeneity among institutions. It avoids the drawbacks of traditional average aggregation methods, where "good data is diluted and bad data is amplified," enabling closer collaboration between similar data sources and allowing models from different clients to achieve differentiated optimization. Simultaneously, while maintaining overall performance, the global model also improves its adaptability to local data from various institutions, ultimately enhancing the model's generalization ability, stability, and personalized prediction accuracy.

[0089] S6: Output the detection results using the completed global model.

[0090] Optionally, the training model of the present invention can be a model with prediction function, such as an early diagnosis model for colorectal cancer, an early diagnosis model for colorectal cancer, a targeted therapy response prediction model, a colorectal cancer recurrence and metastasis prediction model, etc.

[0091] S7: Perform federated feature selection and interpretability analysis on the detection results, and output the local interpretability analysis results.

[0092] In one possible implementation, S7 specifically includes sub-steps S701 to S704:

[0093] S701: Introduces L1 regularization constraints. Each site calculates the L1 coefficient of its local features. After aggregation by the federated server, features with L1 coefficients greater than the preset coefficient are selected, and features strongly correlated with the target are filtered.

[0094] It's worth noting that L1 regularization possesses inherent feature selection capabilities, automatically sparsifying the parameter matrix and compressing the weights of irrelevant or redundant features to zero. By calculating L1 coefficients at each site and aggregating them on the federated server, important features consistent across institutions can be screened out without revealing the original data. This helps remove noisy features, highlighting molecular biomarkers strongly correlated with colorectal cancer targets, and improving the model's stability, generalization ability, and clinical reliability.

[0095] It should be noted that those skilled in the art can set the value of the preset coefficient according to actual needs, and this invention does not limit this.

[0096] S702: Calculate the SHAP value of each feature to quantify the contribution of each feature to the detection result.

[0097] SHAP (Shapley Values) is a model interpretability method derived from Shapley values ​​in cooperative game theory. Its purpose is to interpret a model's predictions by quantifying the contribution of each input feature to the model's output. SHAP calculates the "marginal contribution" of each feature in a single prediction and ensures a fair distribution of these contributions. SHAP can be applied to any machine learning model and provides interpretability to the model's decision-making process, making it particularly important in fields such as medicine and finance.

[0098] It's important to note that SHAP values ​​quantify the marginal contribution of each feature to the model's predictions in a fair and interpretable way. This not only clarifies why the model makes a certain prediction but also helps researchers or doctors identify key genes or mutation sites driving those predictions.

[0099] S703: Use causal subgraphs to analyze causal relationships between features.

[0100] A causal subgraph is a graphical model used to represent causal relationships between variables. In this type of graph, each node represents a variable, and edges represent causal relationships between variables. Causal subgraphs are commonly used to study and model causal inferences. They help us understand how variables influence each other and to make causal inferences, such as how other variables change when one variable changes. Causal subgraphs are widely used in causal inference, policy analysis, and the evaluation of the effectiveness of medical interventions.

[0101] It's important to note that causal subgraphs reveal the causal structure between features, not just their correlations. For example, some gene mutations may be upstream drivers, while changes in gene expression are their consequences. Causal analysis helps distinguish between directly acting features and indirectly related features, preventing the model from being misled by spurious correlations. This not only enhances the scientific explanatory power of the analysis but also helps identify key interventionable factors in clinical decision-making.

[0102] S704: Output locally interpretable analysis results based on feature importance ranking, SHAP value plot, and causal subgraph.

[0103] It's important to note that by combining feature importance ranking, SHAP value graph, and causal subgraph as output, the model's predictions can be interpreted from multiple dimensions: including global importance ranking, explanations of individual predictions, and the revelation of causal mechanisms between features. This output is both intuitive and scientific, making it convenient for clinicians to use in decision-making and enhancing the model's usability and reliability in real-world medical scenarios.

[0104] S8: Integrate detection results with locally interpretable analysis results to generate a joint evidence report for the target.

[0105] Specifically, by combining the detection results of the global model with locally interpretable analysis results (such as SHAP values ​​and causal relationship analysis), the report presents the feature contributions, causal relationships, and biological significance of the prediction results from multiple perspectives. By integrating feature importance rankings, SHAP contribution maps, and causal subgraphs, the report not only provides accurate prediction results but also clearly explains the role of each feature in the diagnosis of colorectal cancer targets, helping physicians understand how the model draws conclusions and improving the transparency and credibility of decision-making. This joint report provides physicians with more comprehensive and reliable diagnostic evidence and promotes the implementation of personalized treatment.

[0106] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0107] (1) In this embodiment of the invention, the original sequencing data of each collaborator is transformed into reconstructed data by generating adversarial networks. In this way, even if the data itself is transformed and reconstructed, there is no need to exchange the original data, thereby avoiding the risk of privacy leakage.

[0108] (2) In this embodiment of the invention, by means of federated learning, the raw sequencing data stored in the local databases of each collaborator can be fully utilized to train the data analysis model, ensuring that the data analysis model is not biased towards a specific data source, avoiding model performance deviation caused by data imbalance, and improving the stability and accuracy of the data analysis model.

[0109] Reference manual attached Figure 2The diagram shows a schematic of the structure of a collaborative processing system for multi-source sequencing data of colorectal cancer targets based on federated learning, provided in an embodiment of the present invention.

[0110] This invention provides a collaborative processing system 20 for multi-source sequencing data of colorectal cancer targets based on federated learning, including a processor 201 and a memory 202.

[0111] The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-described method for collaborative processing of multi-source sequencing data of colorectal cancer targets based on federated learning, and can achieve the same technical effect. To avoid repetition, the present invention will not repeat the above.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for collaborative processing of multi-source sequencing data of colorectal cancer targets based on federated learning, characterized in that, The method comprises the following steps: S1: obtaining local database stored raw sequencing data of each collaborator; S2: reconstructing the local database stored raw sequencing data of each collaborator by using a sequencing translator based on a generative adversarial network; S3: selecting part of the reconstructed sequencing data according to a preset rule and mapping the part of the reconstructed sequencing data to a common target data space; S4: in a local training stage, training a local model by using the local database stored raw sequencing data and part of the reconstructed sequencing data in the target data space through common parameter aggregation optimization; S5: in a global training stage, training a global model by using a federated attention message passing algorithm and according to the contribution degree of different clients to perform personalized aggregation on the local model; S6: outputting a detection result by using the trained global model; S7: performing federated feature selection and explainability analysis on the detection result to output a local explainability analysis result; S8: integrating the detection result and the local explainability analysis result to generate a target joint evidence report.

2. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The multi-source sequencing data comprises at least two of WES sequencing data, WGS sequencing data, targeted panel sequencing data, RNA-seq sequencing data, single-cell RNA-seq sequencing data and ctDNA sequencing data.

3. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The S2 specifically comprises: S201: defining a common target data space; S202: for each collaborator, reconstructing the raw sequencing data by using the generative adversarial network to effectively capture important features in the raw sequencing data, and obtaining reconstructed sequencing data.

4. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 3, characterized in that, The S201 specifically comprises: S2011: performing sparse reconstruction on the local database stored raw sequencing data of each collaborator by using a sparse autoencoder; S2012: calculating the complexity of the local database of each collaborator by using the L2 distance between the sparse reconstructed sequencing data and the raw sequencing data; S2013: selecting the local database of the collaborator with the lowest complexity to define the target data space.

5. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 3, characterized in that, The S201 specifically comprises: S2011: learning a domain-invariant representation in the intersection-union features of the local database stored raw sequencing data of all collaborators by using an adversarial domain-invariant encoder; S2012: defining the target data space according to the domain-invariant representation.

6. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The generative adversarial network comprises a generator and a discriminator, wherein the gradient of the discriminator is added with noise for privacy protection.

7. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The preset rule specifically comprises: In a feature space, the feature similarity between the features of the selected reconstructed sequencing data and the features of the local database stored raw sequencing data is greater than a first preset similarity; and in a data space, the data similarity between the selected reconstructed sequencing data and the local database stored raw sequencing data is less than a second preset similarity.

8. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The S4 specifically comprises: S401: obtaining a common parameter space of the local model and the global model of each collaborator; S402: calculating importance of each public parameter in the public parameter space according to a detection result of the global model on the sequencing data; S403: performing public parameter aggregation optimization based on the importance of each public parameter, and constructing a loss function in local training; S404: training the local model based on the loss function in local training.

9. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The S5 specifically comprises: S501: constructing an optimization objective function of the global model, aiming to reduce the loss function of the local model of each collaborator and improve the collaboration effectiveness among the collaborators; S502: updating a proximal term by gradient descent based on an attention inducing function in the optimization objective function by using alternating optimization; S503: updating the global model parameters based on the proximal term by using an approximation point algorithm.

10. The federated learning based colorectal cancer target multi-source sequencing data collaborative processing method according to claim 1, characterized in that, The S7 specifically comprises: S701: introducing L1 regularization constraint, each site calculates L1 coefficient of local feature, and the federal server selects features with L1 coefficient greater than a preset coefficient after aggregation, so as to screen features strongly related to the target point; S702: calculating SHAP value of each feature to quantify contribution of each feature to the detection result; S703: analyzing causal relationship among features by using a causal subgraph; S704: outputting the local explainable analysis result according to feature importance ranking, SHAP value graph and the causal subgraph.

Images