Radiosensitivity and toxic and side effect detection system based on multiple omics

Abstract

The invention discloses a radiotherapy sensitivity and toxic and side effect detection system based on multiple omics, and relates to the technical field of radiotherapy, and the system comprises the steps: obtaining genomics data, proteomics data and metabonomics data of a patient at different stages, and generating a patient multi-omics time series data set; based on the patient multi-omics time sequence data set, extracting patient multi-omics time sequence features, and constructing a patient radiotherapy sensitivity dynamic prediction model; combining clinical manifestation and treatment history in clinical data of the patient, generating an adversarial network by utilizing a cGAN condition, generating toxicity simulation data under different radiotherapy doses, and establishing a patient toxic and side effect risk prediction model; and according to the patient radiotherapy sensitivity dynamic prediction model and the toxic and side effect risk prediction model, obtaining an optimal radiotherapy dose interval of the patient, and generating a patient personalized radiotherapy digital twinning scheme. The method has the beneficial effects that the treatment safety and effect of patients are improved, and the method has higher clinical application value and personalized treatment potential.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, specifically to a multi-omics-based system for detecting radiotherapy sensitivity and toxic side effects. Background Technology

[0002] Existing multi-omics-based radiosensitivity and toxicity detection systems face challenges such as data heterogeneity and lack of standardization. Insufficient utilization of single-omics data leads to limited predictive capabilities. Furthermore, inadequate processing of high-dimensional data and insufficient algorithmic support affect the system's accuracy and real-time performance. Despite initial applications, there are still technical barriers in clinical translation and personalized treatment, and a lack of universally applicable solutions. Summary of the Invention

[0003] To address the aforementioned technical issues, a multi-omics-based radiosensitivity and toxicity detection system is provided. This solution resolves the problems of data heterogeneity and lack of standardization faced by existing multi-omics-based radiosensitivity and toxicity detection systems, as well as the insufficient utilization of single-omics data, resulting in limited predictive capabilities. Furthermore, the lack of processing and algorithmic support for high-dimensional data affects the system's accuracy and real-time performance. Although preliminary applications have been made, technical barriers remain in clinical translation and personalized treatment, and a lack of universally applicable solutions persists.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A multi-omics-based method for detecting radiosensitivity and toxicity includes:

[0006] Based on the collection of patient clinical data, genomic data, proteomics data and metabolomics data of patients at different stages are obtained to generate patient multi-omics time series datasets;

[0007] Based on the patient multi-omics time series dataset, extract the patient multi-omics time series features, assess the patient radiotherapy sensitivity in real time, and construct a dynamic prediction model for the patient radiotherapy sensitivity.

[0008] By combining the clinical manifestations and treatment history in the patient's clinical data, cGAN conditional generative adversarial network is used to generate toxicity simulation data under different radiotherapy doses, and a patient toxicity risk prediction model is established.

[0009] Based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects, the correlation between radiotherapy sensitivity and toxic side effects is balanced to obtain the optimal radiotherapy dose range for patients and generate personalized digital twin plans for radiotherapy.

[0010] Preferably, based on the collection of patient clinical data, the sample data collection time is recorded, and genomic data, proteomics data and metabolomics data of patients at different stages are obtained. The data are normalized and outliers are linearly imputed.

[0011] Based on the patient's genomic, proteomic, and metabolomics data at different stages, the time axis was aligned, and Euclidean distance was used to calculate the data differences at each time point, generating a multi-omics data distance matrix, as shown in the following formula:

[0012]

[0013] in, For time points With time point The Euclidean distance between them For genomic data, For proteomics data, For genomic data in the first Feature vectors at each time point For proteomics data in the first Feature vectors at each time point For genomic data in the first The time point, the first Normalized values ​​of each feature For proteomics data in the first The time point, the first Normalized values ​​of each feature The number of common features;

[0014] Using dynamic programming, the minimum cumulative distance path between different time points of multi-omics data is calculated to obtain the optimal alignment path of multi-omics data.

[0015] Based on the optimal alignment path of multi-omics data, proteomics and metabolomics time points are mapped to the time axis of genomics to generate patient multi-omics time series datasets.

[0016] Preferably, based on the patient multi-omics time-series dataset, causal convolution operation is used to extract the local time-series features of the patient multi-omics dataset;

[0017] Using the patient's multi-omics time-series features as input, a linear transformation maps these features to query, key, and value vectors, respectively. The similarity between the query and key is calculated, and attention weights are generated using a softmax function to establish dependencies between time points. The formula is as follows:

[0018]

[0019] in, For attention weights, For a point in time, For query vectors, For the key vector, To score the similarity, Scaling factor It is an exponential function. The total number of time steps. For summation index.

[0020] Preferably, the value vector is weighted and summed based on attention weights to obtain the final representation at each time point, and the patient's multi-omics time-series features are extracted;

[0021] Based on the multi-omics temporal characteristics of patients, and taking the patients' radiotherapy sensitivity as the target, the likelihood function is defined as a normal distribution. Using Bayes' theorem, the posterior distribution is calculated by combining the prior distribution and the likelihood function.

[0022] As a further development, the prior distribution includes: previous clinical experience; the likelihood function includes: patient time series data;

[0023] Based on the calculated posterior distribution, the radiotherapy sensitivity of patients at different time points is dynamically assessed, and the prediction results are adjusted in real time to construct a dynamic prediction model for patient radiotherapy sensitivity.

[0024] Preferably, a generator is designed based on the patient's clinical manifestations and treatment history in the clinical data to generate toxic reactions as output;

[0025] The discriminant is designed to determine whether the input sample data is authentic by inputting real patient data and combining causal loss to ensure that there is a monotonically increasing correlation between toxic reactions and radiotherapy dose.

[0026] Adversarial training is performed using a generator and a discriminator, and the loss functions of the generator and discriminator are alternately optimized to improve the ability to generate toxicity simulation data. Through iterative training, toxicity simulation data under different radiotherapy doses are generated.

[0027] Based on the toxicity simulation data generated under different radiotherapy doses, and combined with the dynamic prediction model of patient radiotherapy sensitivity, a feature vector of toxic side effects risk is extracted.

[0028] Preferably, a logistic regression model is constructed based on the toxic side effect risk feature vector, with the patient's toxic side effect risk as the target variable, and the cross-entropy loss function is used to calculate the error between the model's prediction results and the true label;

[0029] Cross-validation was used to evaluate the model's performance on different training and test sets. Probability predictions were made, and a threshold was set according to the requirements to determine the risk of patient side effects. If the predicted probability was greater than the threshold, the patient was determined to have a risk of side effects; if the predicted probability was less than the threshold, the patient was determined to have no risk of side effects. A patient side effect risk prediction model was thus established.

[0030] Preferably, based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects, a correlation model between patient radiotherapy sensitivity and toxic side effects risk is constructed. Radiotherapy sensitivity and toxic side effects risk are defined as two optimization objectives. Using the NSGA-II non-dominated sorting genetic algorithm, a set of dose distribution schemes is randomly generated. The constraints are defined as organ limit and dose smoothness. According to the objective function and constraints, the solution is divided into multiple Pareto front levels, which are then non-dominated and sorted. The crowding degree is calculated according to the target spatial distance. Through selection, crossover, and mutation operations, the dose distribution scheme is iteratively optimized until the termination condition is met, and the optimal radiotherapy dose range for the patient is obtained.

[0031] Preferably, based on the patient's optimal radiotherapy dose range, a federated learning algorithm is used. Hospital nodes update local model parameters locally and upload encrypted gradients. The central server aggregates the uploaded encrypted gradients, updates the global model, and performs continuous learning and model updates to construct a personalized digital twin model of radiotherapy for the patient.

[0032] Furthermore, a multi-omics-based radiosensitivity and toxicity detection system is provided to implement the aforementioned multi-omics-based radiosensitivity and toxicity detection method, comprising:

[0033] The module includes a time-series dataset module, a dynamic prediction model module, a toxic side effect risk prediction module, and a personalized radiotherapy module.

[0034] The time-series dataset module is used to collect genomic data, proteomics data and metabolomics data of patients at different stages based on the collection of patient clinical data, and generate a multi-omics time-series dataset of patients.

[0035] The dynamic prediction model module is electrically connected to the time series dataset module. The dynamic prediction model module is used to extract multi-omics time series features of patients based on the multi-omics time series dataset of patients, evaluate the radiotherapy sensitivity of patients in real time, and construct a dynamic prediction model of patients' radiotherapy sensitivity.

[0036] The toxic side effect risk prediction module is electrically connected to the time series dataset module. The toxic side effect risk prediction module is used to combine the clinical manifestations and treatment history in the patient's clinical data, use the cGAN conditional generative adversarial network to generate toxicity simulation data under different radiotherapy doses, and establish a patient toxic side effect risk prediction model.

[0037] The personalized radiotherapy module is electrically connected to the toxic side effect risk prediction module. The personalized radiotherapy module is used to balance the correlation between radiotherapy sensitivity and toxic side effects based on the patient's radiotherapy sensitivity dynamic prediction model and toxic side effect risk prediction model, obtain the patient's optimal radiotherapy dose range, and generate a personalized radiotherapy digital twin plan for the patient.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention proposes a multi-omics-based approach to detect radiosensitivity and toxicity. By integrating genomics, proteomics, and metabolomics data, this method enables precise assessment and dynamic prediction of patients' radiosensitivity and toxicity. Furthermore, by utilizing generative adversarial networks to generate toxicity simulation data at different doses, combined with patient clinical manifestations and treatment history, it can more comprehensively predict patients' radiotherapy responses and toxicity risks. Finally, by optimizing the radiotherapy dose range, personalized radiotherapy plans can be implemented, maximizing treatment efficacy, reducing side effects, and improving patient safety and outcomes, demonstrating higher clinical application value and potential for personalized treatment. Attached Figure Description

[0040] Figure 1 This is a flowchart of a multi-omics-based method for detecting radiosensitivity and toxic side effects.

[0041] Figure 2 This is a framework diagram of a radiotherapy sensitivity and toxicity detection system based on multi-omics. Detailed Implementation

[0042] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0043] Reference Figure 1 As shown, a multi-omics-based method for detecting radiosensitivity and toxicity includes:

[0044] S1. Based on the collection of patient clinical data, obtain genomic data, proteomics data and metabolomics data of patients at different stages, and generate patient multi-omics time series datasets;

[0045] Step S1 includes the following:

[0046] Based on the collection of patient clinical data, the sample data collection time was recorded, and genomic data, proteomics data and metabolomics data of patients at different stages were obtained. The data were normalized and outliers were linearly imputed.

[0047] Based on the patient's genomic, proteomic, and metabolomics data at different stages, the time axis was aligned, and Euclidean distance was used to calculate the data differences at each time point, generating a multi-omics data distance matrix, as shown in the following formula:

[0048]

[0049] in, For time points With time point The Euclidean distance between them For genomic data, For proteomics data, For genomic data in the first Feature vectors at each time point For proteomics data in the first Feature vectors at each time point For genomic data in the first The time point, the first Normalized values ​​of each feature For proteomics data in the first The time point, the first Normalized values ​​of each feature The number of common features;

[0050] Using dynamic programming, the minimum cumulative distance path between different time points of multi-omics data is calculated to obtain the optimal alignment path of multi-omics data.

[0051] Based on the optimal alignment path of multi-omics data, proteomics and metabolomics time points are mapped to the time axis of genomics to generate patient multi-omics time series datasets.

[0052] When using it, refer to the content of step S1 above:

[0053] Existing technologies mainly focus on the temporal alignment of single-omics data, which suffers from insufficient handling of heterogeneity in multi-omics data, scale differences, and neglect of noise. Furthermore, they have high computational complexity and are difficult to process large-scale data. This step solves the above problems by calculating the differences between multi-omics data through Euclidean distance and optimizing the alignment path using dynamic programming. This method accurately aligns different omics data, overcomes the influence of noise, and improves computational efficiency, providing comprehensive support for precision medicine and disease prevention.

[0054] S2. Based on the patient multi-omics time series dataset, extract the patient multi-omics time series features, evaluate the patient's radiotherapy sensitivity in real time, and construct a dynamic prediction model for the patient's radiotherapy sensitivity.

[0055] Step S2 includes the following:

[0056] Based on the patient multi-omics time series dataset, causal convolution operation is used to extract the local time series features of the patient multi-omics dataset;

[0057] Using the patient's multi-omics time-series features as input, a linear transformation maps these features to query, key, and value vectors, respectively. The similarity between the query and key is calculated, and attention weights are generated using a softmax function to establish dependencies between time points. The formula is as follows:

[0058]

[0059] in, For attention weights, For a point in time, For query vectors, For the key vector, To score the similarity, Scaling factor It is an exponential function. The total number of time steps. For summation index;

[0060] The Value vector is weighted and summed based on attention weights to obtain the final representation of each time point, and the multi-omics time-series features of patients are extracted.

[0061] Based on the multi-omics temporal characteristics of patients, and taking the patients' radiotherapy sensitivity as the target, the likelihood function is defined as a normal distribution. Using Bayes' theorem, the posterior distribution is calculated by combining the prior distribution and the likelihood function.

[0062] As a further development, the prior distribution includes: previous clinical experience; the likelihood function includes: patient time series data;

[0063] Based on the calculated posterior distribution, the radiotherapy sensitivity of patients at different time points is dynamically assessed, and the prediction results are adjusted in real time to construct a dynamic prediction model for patient radiotherapy sensitivity.

[0064] When using it, refer to the content of step S2 above:

[0065] Existing methods for predicting radiotherapy sensitivity mostly rely on single-omics data or static models, which cannot update and accurately assess patients' radiotherapy responses in real time. They also lack consideration for multi-omics data fusion and temporal dependence, affecting prediction accuracy and clinical decision-making. This step extracts multi-omics temporal features through causal convolution and self-attention mechanisms, and combines them with Bayesian updates to assess patients' radiotherapy sensitivity in real time, improving prediction accuracy and personalized treatment capabilities, and enhancing the model's adaptability and reliability in clinical practice.

[0066] S3. Combining the clinical manifestations and treatment history in the patient's clinical data, using the cGAN conditional generative adversarial network, generate toxicity simulation data under different radiotherapy doses, and establish a patient toxicity risk prediction model.

[0067] Step S3 includes the following:

[0068] Based on the clinical manifestations and treatment history in the patient's clinical data, a generator is designed to generate toxic reactions as output;

[0069] The discriminant is designed to determine whether the input sample data is authentic by inputting real patient data and combining causal loss to ensure that there is a monotonically increasing correlation between toxic reactions and radiotherapy dose.

[0070] Adversarial training is performed using a generator and a discriminator, and the loss functions of the generator and discriminator are alternately optimized to improve the ability to generate toxicity simulation data. Through iterative training, toxicity simulation data under different radiotherapy doses are generated.

[0071] Based on the generation of toxicity simulation data under different radiotherapy doses, combined with a dynamic prediction model of patient radiotherapy sensitivity, a feature vector of toxic side effects risk is extracted.

[0072] Based on the toxic side effect risk feature vector, a logistic regression model is constructed, with the patient's toxic side effect risk as the target variable. The cross-entropy loss function is used to calculate the error between the model's prediction results and the true label.

[0073] Cross-validation was used to evaluate the model's performance on different training and test sets. Probability predictions were made, and a threshold was set according to the requirements to determine the risk of patient side effects. If the predicted probability was greater than the threshold, the patient was determined to have a risk of side effects; if the predicted probability was less than the threshold, the patient was determined to have no risk of side effects. A patient side effect risk prediction model was thus established.

[0074] When using it, refer to the content of step S3 above:

[0075] Existing domestic and international technologies for predicting the risk of toxic side effects largely rely on traditional statistical and machine learning methods. These methods depend on data labeling and manual feature extraction, and struggle to handle complex nonlinear relationships and high-dimensional data. They are also susceptible to imbalanced data, feature selection bias, and overfitting, resulting in insufficient prediction accuracy and robustness. Introducing conditional generative adversarial networks (GANs) can automatically generate toxicity simulation data under different radiotherapy doses, avoiding manual intervention. Through adversarial training, complex data patterns and potential correlations can be captured, and combined with the patient's radiotherapy sensitivity model, more comprehensive toxic side effect risk characteristics can be extracted. Accurate predictions can be made based on logistic regression models, and cross-validation can improve the accuracy and robustness of predictions, thereby providing personalized risk assessment support for clinical practice.

[0076] S4. Based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects, balance the correlation between radiotherapy sensitivity and toxic side effects, obtain the optimal radiotherapy dose range for the patient, and generate a personalized digital twin plan for radiotherapy for the patient.

[0077] Step S4 includes the following:

[0078] Based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects risk, a correlation model between patient radiotherapy sensitivity and toxic side effects risk is constructed. Radiotherapy sensitivity and toxic side effects risk are defined as two optimization objectives. Using the NSGA-II non-dominated sorting genetic algorithm, a set of dose distribution schemes is randomly generated. The constraints are defined as organ limit and dose smoothness. According to the objective function and constraints, the solution is divided into multiple Pareto front levels, which are then non-dominated sorted. The crowding degree is calculated according to the target spatial distance. Through selection, crossover, and mutation operations, the dose distribution scheme is iteratively optimized until the termination condition is met, and the optimal radiotherapy dose range for the patient is obtained.

[0079] Based on the patient's optimal radiotherapy dose range, a federated learning algorithm is used to update local model parameters locally and upload encrypted gradients. The central server aggregates the uploaded encrypted gradients, updates the global model, and performs continuous learning and model updates to build a personalized digital twin model of radiotherapy for the patient.

[0080] When using it, refer to the content of step S4 above:

[0081] Existing technologies for personalized digital twin models of radiotherapy suffer from issues such as data privacy protection, heterogeneity, optimization of multi-objective balance, and communication efficiency. This step combines optimization of radiotherapy sensitivity and toxic side effects risk, utilizes federated learning to ensure data privacy and achieve cross-hospital data sharing, continuously updates the model to adapt to changes in the patient's condition, and improves optimization efficiency through a non-dominated sorting genetic algorithm, thereby achieving personalized treatment, maximizing treatment effectiveness, and minimizing side effects.

[0082] Reference Figure 2 As shown, a multi-omics-based radiosensitivity and toxicity detection system is characterized by comprising:

[0083] The module includes a time-series dataset module, a dynamic prediction model module, a toxic side effect risk prediction module, and a personalized radiotherapy module.

[0084] The time-series dataset module is used to collect genomic data, proteomics data and metabolomics data of patients at different stages based on the collection of patient clinical data, and generate a multi-omics time-series dataset of patients.

[0085] The dynamic prediction model module is electrically connected to the time series dataset module. The dynamic prediction model module is used to extract multi-omics time series features of patients based on the multi-omics time series dataset of patients, evaluate the radiotherapy sensitivity of patients in real time, and construct a dynamic prediction model of patients' radiotherapy sensitivity.

[0086] The toxic side effect risk prediction module is electrically connected to the time series dataset module. The toxic side effect risk prediction module is used to combine the clinical manifestations and treatment history in the patient's clinical data, use the cGAN conditional generative adversarial network to generate toxicity simulation data under different radiotherapy doses, and establish a patient toxic side effect risk prediction model.

[0087] The personalized radiotherapy module is electrically connected to the toxic side effect risk prediction module. The personalized radiotherapy module is used to balance the correlation between radiotherapy sensitivity and toxic side effects based on the patient's radiotherapy sensitivity dynamic prediction model and toxic side effect risk prediction model, obtain the patient's optimal radiotherapy dose range, and generate a personalized radiotherapy digital twin plan for the patient.

[0088] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for detecting radiosensitivity and toxic side effects based on multi-omics, characterized in that, include: S1. Based on the collection of patient clinical data, obtain genomic data, proteomics data and metabolomics data of patients at different stages, and generate patient multi-omics time series datasets; S2. Based on the patient multi-omics time series dataset, extract the patient multi-omics time series features, evaluate the patient's radiotherapy sensitivity in real time, and construct a dynamic prediction model for the patient's radiotherapy sensitivity. S3. Combining the clinical manifestations and treatment history in the patient's clinical data, using the cGAN conditional generative adversarial network, generate toxicity simulation data under different radiotherapy doses, and establish a patient toxicity risk prediction model. S4. Based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects, balance the correlation between radiotherapy sensitivity and toxic side effects, obtain the optimal radiotherapy dose range for the patient, and generate a personalized digital twin plan for radiotherapy for the patient.

2. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 1, characterized in that, S1 includes: Based on the collection of patient clinical data, the sample data collection time was recorded, and genomic data, proteomics data and metabolomics data of patients at different stages were obtained. The data were normalized and outliers were linearly imputed. Based on the patient's genomic, proteomic, and metabolomics data at different stages, the time axis was aligned, and Euclidean distance was used to calculate the data differences at each time point, generating a multi-omics data distance matrix, as shown in the following formula: in, For time points With time point The Euclidean distance between them For genomic data, For proteomics data, For genomic data in the first Feature vectors at each time point For proteomics data in the first Feature vectors at each time point For genomic data in the first The time point, the first Normalized values ​​of each feature For proteomics data in the first The time point, the first Normalized values ​​of each feature The number of common features; Using dynamic programming, the minimum cumulative distance path between different time points of multi-omics data is calculated to obtain the optimal alignment path of multi-omics data. Based on the optimal alignment path of multi-omics data, proteomics and metabolomics time points are mapped to the time axis of genomics to generate patient multi-omics time series datasets.

3. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 2, characterized in that, S2 includes: Based on the patient multi-omics time series dataset, causal convolution operation is used to extract the local time series features of the patient multi-omics dataset; Using the patient's multi-omics time-series features as input, a linear transformation maps these features to query, key, and value vectors, respectively. The similarity between the query and key is calculated, and attention weights are generated using a softmax function to establish dependencies between time points. The formula is as follows: in, For attention weights, For a point in time, For query vectors, The key vector. To score the similarity, Scaling factor It is an exponential function. The total number of time steps. For summation index.

4. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 3, characterized in that, S2 further includes: The Value vector is weighted and summed based on attention weights to obtain the final representation of each time point, and the multi-omics time-series features of patients are extracted. Based on the multi-omics temporal characteristics of patients, and taking the patients' radiotherapy sensitivity as the target, the likelihood function is defined as a normal distribution. Using Bayes' theorem, the posterior distribution is calculated by combining the prior distribution and the likelihood function. As a further development, the prior distribution includes: previous clinical experience; the likelihood function includes: patient time series data; Based on the calculated posterior distribution, the radiotherapy sensitivity of patients at different time points is dynamically assessed, and the prediction results are adjusted in real time to construct a dynamic prediction model for patient radiotherapy sensitivity.

5. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 2, characterized in that, S3 includes: Based on the patient's clinical manifestations and treatment history in the clinical data, a generator is designed to generate toxic reactions as output; The discriminant is designed to determine whether the input sample data is authentic by inputting real patient data and combining causal loss to ensure that there is a monotonically increasing correlation between toxic reactions and radiotherapy dose. Adversarial training is performed using a generator and a discriminator, and the loss functions of the generator and discriminator are alternately optimized to improve the ability to generate toxicity simulation data. Through iterative training, toxicity simulation data under different radiotherapy doses are generated. Based on the toxicity simulation data generated under different radiotherapy doses, and combined with the dynamic prediction model of patient radiotherapy sensitivity, a feature vector of toxic side effects risk is extracted.

6. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 5, characterized in that, S3 further includes: Based on the toxic side effect risk feature vector, a logistic regression model is constructed, with the patient's toxic side effect risk as the target variable. The cross-entropy loss function is used to calculate the error between the model's prediction results and the true label. Cross-validation was used to evaluate the model's performance on different training and test sets. Probability predictions were made, and a threshold was set according to the requirements to determine the risk of patient side effects. If the predicted probability was greater than the threshold, the patient was determined to have a risk of side effects; if the predicted probability was less than the threshold, the patient was determined to have no risk of side effects. A patient side effect risk prediction model was thus established.

7. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 6, characterized in that, S4 includes: Based on the dynamic prediction model of patient radiotherapy sensitivity and the prediction model of toxic side effects risk, a correlation model between patient radiotherapy sensitivity and toxic side effects risk is constructed. Radiotherapy sensitivity and toxic side effects risk are defined as two optimization objectives. Using the NSGA-II non-dominated sorting genetic algorithm, a set of dose distribution schemes is randomly generated. The constraints are defined as organ limit and dose smoothness. According to the objective function and constraints, the solution is divided into multiple Pareto front levels, which are then non-dominated sorted. The crowding degree is calculated according to the target spatial distance. Through selection, crossover, and mutation operations, the dose distribution scheme is continuously iterated and optimized until the termination condition is met, thus obtaining the optimal radiotherapy dose range for the patient.

8. The method for detecting radiosensitivity and toxic side effects based on multi-omics according to claim 7, characterized in that, S4 further includes: Based on the patient's optimal radiotherapy dose range, a federated learning algorithm is used to update local model parameters locally and upload encrypted gradients. The central server aggregates the uploaded encrypted gradients, updates the global model, and performs continuous learning and model updates to build a personalized digital twin model of radiotherapy for the patient.

9. A multi-omics-based system for detecting radiosensitivity and toxic side effects, characterized in that, To implement any one of claims 1-8, the method for detecting radiosensitivity and toxicity based on multi-omics comprises: The module includes a time-series dataset module, a dynamic prediction model module, a toxic side effect risk prediction module, and a personalized radiotherapy module. The time-series dataset module is used to collect genomic data, proteomics data and metabolomics data of patients at different stages based on the collection of patient clinical data, and generate a multi-omics time-series dataset of patients. The dynamic prediction model module is electrically connected to the time series dataset module. The dynamic prediction model module is used to extract multi-omics time series features of patients based on the multi-omics time series dataset of patients, evaluate the radiotherapy sensitivity of patients in real time, and construct a dynamic prediction model of patients' radiotherapy sensitivity. The toxic side effect risk prediction module is electrically connected to the time series dataset module. The toxic side effect risk prediction module is used to combine the clinical manifestations and treatment history in the patient's clinical data, use the cGAN conditional generative adversarial network to generate toxicity simulation data under different radiotherapy doses, and establish a patient toxic side effect risk prediction model. The personalized radiotherapy module is electrically connected to the toxic side effect risk prediction module. The personalized radiotherapy module is used to balance the correlation between radiotherapy sensitivity and toxic side effects based on the patient's radiotherapy sensitivity dynamic prediction model and toxic side effect risk prediction model, obtain the patient's optimal radiotherapy dose range, and generate a personalized radiotherapy digital twin plan for the patient.

Images