Immune checkpoint blocking response prediction method and system based on diffusion model

Abstract

The invention discloses an immune checkpoint blocking response prediction method and system based on a diffusion model, and relates to the technical field of computational biology and biological information, and the method comprises the steps: inputting 7-dimensional splicing data into an ODE solver; and the ODE solver inputs the current time step generated by solving and the current intermediate data into the diffusion model, and feeds back the obtained current prediction score into the ODE solver, and the ODE solver iteratively inputs the next time step generated by solving and the corresponding intermediate data into the diffusion model until the ordinary differential equation in the ODE solver is completely solved, so that the current prediction score is obtained. The ODE solver carries out calculation to obtain a data sample and an integral item result which conform to Gaussian distribution, and calculates a joint probability; two joint probabilities are obtained on the basis that the curative effect of the patient on the immune checkpoint blocking therapy is in a response state and a non-response state, and a response probability is calculated on the basis of the joint probabilities; according to the response prediction method and system, more accurate prediction of ICB treatment response is realized.

Description

Technical Field

[0001] The present invention relates to the fields of computational biology and bioinformatics, and particularly to a method and system for predicting immune checkpoint blockade response based on a diffusion model. Background Art

[0002] Cancer is a type of malignant disease caused by abnormal cell proliferation with the ability of invasion and metastasis, seriously threatening human health. Due to its high heterogeneity and dynamic evolution characteristics, traditional cancer treatment methods generally have certain limitations in clinical applications. For example, surgical treatment is mainly applicable to patients with early-stage tumors that have not metastasized, and it is difficult to completely remove potential micrometastatic lesions; radiotherapy and chemotherapy can control the disease progression to a certain extent, but often have significant side effects, affecting the quality of life of patients; although targeted therapy has a certain specificity, it is only applicable to patient groups with clear molecular targets, and long-term use is prone to drug resistance.

[0003] In recent years, immunotherapy, especially immune checkpoint blockade (ICB) therapy, has brought new breakthroughs to cancer treatment. This method activates the recognition and killing of tumor cells by the host immune system by inhibiting the immune checkpoint signaling pathway between tumor cells and immune cells (such as immune checkpoint PD-1 / PD-L1, immune regulatory protein CTLA-4, etc.). However, the response rate of ICB therapy in clinical practice is only about 20% - 30%, and there are significant individual differences in the curative effect. Therefore, how to accurately predict the response of patients after receiving ICB treatment has become a key issue in improving the treatment effect, avoiding ineffective treatment from delaying the disease, and promoting the development of personalized precision medicine.

[0004] Traditional ICB efficacy prediction methods, such as those based on indicators like the PD-L1 expression level in tumor tissues or tumor mutation burden (TMB), have certain limitations. These methods can often only predict the treatment response of a part of the patients and are easily affected by factors such as tissue sampling and detection methods. Therefore, it is necessary to explore more relevant features to assist in the prediction work. Currently, the clinical, pathological, and genomic features considered to be related to ICB efficacy are: the expression of programmed cell death 1 and its ligand 1 in tumor cells (programmed cell death 1, PD-1; programmed cell death 1 ligand 1, PD-L1), microsatellite instability (MSI), human leukocyte antigen class I evolutionary divergence (HED), fraction of copy number alteration (FCNA), blood neutrophil–lymphocyte ratio (NLR), blood albumin level, body mass index (BMI), sex, and age, etc.

[0005] In the past, some studies have tried to introduce statistical learning methods (such as logistic regression, support vector machines, etc.) and conduct comprehensive modeling by combining multiple clinical features to improve the prediction accuracy. However, such methods are limited by the expression ability of the model structure and often have difficulty fully mining the complex correlation relationships between variables when processing data, and it is also difficult to capture the latent mapping rules between patient characteristics and treatment responses. Summary of the Invention

[0006] Based on the technical problems existing in the background art, the present invention proposes an immune checkpoint blockade response prediction method and system based on a diffusion model, which can effectively learn the potential representation and internal distribution in the multi-dimensional feature data of patients, so as to achieve a more accurate prediction of the ICB treatment response.

[0007] The immune checkpoint blockade response prediction method based on a diffusion model proposed by the present invention includes: Obtain the multi-dimensional index features of a patient and splice them to obtain 7-dimensional spliced data, and then input it into an ODE solver; The ODE solver inputs the currently solved time step and the current intermediate data into the trained diffusion model, and feeds the obtained current prediction score back to the ODE solver. The ODE solver iteratively inputs the next time step solved and the intermediate data corresponding to the next time step into the diffusion model until the ordinary differential equation in the ODE solver is solved. The ODE solver calculates a data sample conforming to a Gaussian distribution and the result of the integral term, where the intermediate data is the noise perturbation feature corresponding to the current time step; The joint probability is calculated using the sum of the log-likelihood of the data sample and the result of the integral term; Based on the two states of the patient's response and non-response to the immune checkpoint blockade therapy, two joint probabilities are obtained. The ratio of the joint probability in the response state to the sum of the two joint probabilities is used as the response probability of the patient to the immune checkpoint blockade therapy.

[0008] Further, the obtaining of the multi-dimensional index features of the patient and splicing them to obtain 7-dimensional spliced data is specifically as follows: The 6-dimensional index features of the patient are obtained and normalized as the original 6-dimensional data; The preset efficacy of the patient for the immune checkpoint blockade therapy is used as the 7th dimension data and normalized and then spliced at the end of the original 6-dimensional data to obtain 7-dimensional spliced data.

[0009] Further, during the process of the ODE solver solving the ordinary differential equation, one path of the ODE solver gradually accumulates and calculates the result of the integral term based on the prediction scores fed back by the diffusion model corresponding to each time step; the other path of the ODE solver is used to add noise to the 7-dimensional spliced data at each time step to obtain intermediate data. After all time steps are completed, a data sample conforming to a Gaussian distribution is obtained, and the log-likelihood of the data sample is calculated.

[0010] Further, the diffusion model uses a feed-forward network as the noise prediction network, and the feed-forward network includes two hidden layers connected in series and one output layer; The training process of the diffusion model is as follows: The 6-dimensional index features of the patient and the preset efficacy of the patient for the immune checkpoint blockade therapy are obtained to obtain 7-dimensional spliced data; randomly sample time steps , and after adding noise to the 7-dimensional spliced data, it is used as a noisy sample to construct a training dataset, where is a uniform distribution, is the maximum end point of time; The training dataset is input into the first hidden layer of the feed-forward network; An overlay layer is set between two adjacent hidden layers, an overlay layer is set between the second hidden layer and the output layer, and in the two overlay layers, the corresponding time step is embedded into the noisy sample through Gaussian Fourier projection; Construct the loss function of the diffusion model to adjust the trainable parameters in the diffusion model.

[0011] Further, the Gaussian Fourier projection operation at time step is as follows: ; where ; where are the parameters of the linear layer network, represents the phase information of the concatenated sine and cosine functions, is a randomly generated Gaussian matrix, is the concatenation operation.

[0012] Further, when constructing the training dataset with the noisy samples after adding noise to the 7-dimensional concatenated data, the operation corresponding to the time step for the noisy samples is as follows: ; where ; where represents the noise intensity corresponding to the time step , σ is a hyperparameter representing the maximum noise intensity, represents the standard Gaussian noise sampled randomly, represents the original sample without noise, represents the noisy sample after adding noise.

[0013] Further, the loss function of the diffusion model is as follows: ; where represents the predicted score of the diffusion model, represents the noisy sample after adding noise.

[0014] Further, the log-likelihood of the data sample is calculated as follows: ; where is the likelihood of the data sample of the Gaussian distribution, is a hyperparameter, is the data sample conforming to the Gaussian distribution, is the two-norm of.

[0015] Further, in calculating the joint probability using the sum of the log-likelihood of the data sample and the result of the integral term, when the patient's response to immune checkpoint blockade therapy is a response, the logarithm of the joint probability The specific calculation formula is: ; where is the joint probability in the response population , is the original sample without noise, is the noise perturbation feature at time during the diffusion process, is that the patient's response to immune checkpoint blockade therapy is a response, is the prediction score, is the divergence of the prediction score, represents the model parameters of the diffusion model after training, and are the start time and end time of diffusion respectively, is the cumulative divergence effect of the integral from to , is a hyperparameter representing the maximum noise intensity, is the integration variable.

[0016] An immune checkpoint blockade response prediction system based on a diffusion model includes a data acquisition module, a prediction module, a joint probability module, and a response probability module; The data acquisition module is used to obtain the multi-dimensional index features of the patient and splice them to obtain 7-dimensional spliced data, which is then input into the ODE solver; In the prediction module, the ODE solver inputs the current time step and the current intermediate data generated by the solution into the trained diffusion model, and feeds back the obtained current prediction score to the ODE solver. The ODE solver inputs the next time step and the intermediate data corresponding to the next time step generated by the solution as the current time step into the diffusion model iteratively until the ordinary differential equation in the ODE solver is solved. The ODE solver calculates the data sample conforming to the Gaussian distribution and the result of the integral term. The intermediate data is the noise perturbation feature corresponding to the current time step; In the joint probability module, the joint probability is calculated using the sum of the log-likelihood of the data sample and the result of the integral term; In the response probability module, based on the two states of the patient's response and non-response to immune checkpoint blockade therapy, two joint probabilities are obtained. The ratio of the joint probability in the response state to the sum of the two joint probabilities is used as the response probability of the patient to immune checkpoint blockade therapy.

[0017] The advantages of the immune checkpoint blockade response prediction method and system based on the diffusion model provided by the present invention are as follows: It can effectively learn the latent representation and internal distribution in the multi-dimensional feature data of patients, so as to achieve a more accurate prediction of the ICB treatment response; The work of using the diffusion model to generate 7-dimensional medical data and calculate the sample conditional probability for ICB efficacy prediction fills a part of the blank of the application of the diffusion model in the low-dimensional data space; Theoretically, an accurate calculation method for the response probability of predicting the efficacy is given. Compared with other methods that only give the confidence level in the prediction problem, this embodiment has more advantages in the interpretability of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of the present invention; Figure 2 is a schematic diagram of the ROC curve and AUC performance corresponding to the prediction results of this embodiment on the test set; Figure 3 is a schematic diagram of the precision-recall curve and AUPRC performance corresponding to the prediction results of this embodiment on the test set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions of the present invention will be described in detail through specific embodiments. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0020] As Figures 1 to 3 shown, the immune checkpoint blockade response prediction method based on the diffusion model proposed by the present invention includes: Step 1: Obtain the multi-dimensional index features of the patient and splice them to obtain 7-dimensional spliced data, and then input it into the ODE solver; Step 2: The ODE solver inputs the currently solved time step and the current intermediate data into the pre-trained diffusion model, and feeds the obtained current prediction score back to the ODE solver. The ODE solver takes the next time step and the intermediate data corresponding to the next time step generated by the solution as the current time step and iteratively inputs them into the diffusion model until the ordinary differential equation in the ODE solver is solved. The ODE solver calculates the data sample conforming to the Gaussian distribution and the integral term result. The intermediate data is the noise perturbation feature corresponding to the current time step; Step 3: Calculate the joint probability by using the sum of the log-likelihood of the data sample and the integral term result; Step 4: Based on the two states of the patient's response and non - response to immune checkpoint blockade therapy, two joint probabilities are obtained. The ratio of the joint probability in the response state to the sum of the two joint probabilities is used as the response probability of the patient to immune checkpoint blockade therapy.

[0021] Given the first 6 dimensions (patient metrics) of a vector data and hoping to know the 7th dimension (the patient's response to ICB therapy); in this embodiment, by using a combination of an ODE solver (i.e., a differential equation solver) and a diffusion model, and combining the joint probability generation process, it is possible to effectively learn the latent representation and internal distribution in the patient's multi - dimensional feature data, thereby achieving a more accurate prediction of the response to ICB treatment.

[0022] In this embodiment, Step 1: Obtain the multi - dimensional metric features of the patient and splice them to obtain 7 - dimensional spliced data, and then input it into the ODE solver. Specifically: Obtain the 6 - dimensional metric features of the patient and perform normalization processing as the original 6 - dimensional data; Preset the efficacy of the patient to immune checkpoint blockade therapy as the 7th dimension data and splice it at the end of the original 6 - dimensional data after normalization to obtain 7 - dimensional spliced data.

[0023] Among them, the original 6 - dimensional data includes tumor mutation burden (TMB), patient systemic treatment history (PSTH), blood albumin, neutrophil - lymphocyte ratio (NLR), age, and cancer type.

[0024] In this embodiment, the efficacy of the patient to immune checkpoint blockade therapy is divided into two states: "response" and "non - response". By using the ODE solver and the diffusion model, the joint probabilities in these two states are calculated, and then the response probability of the patient to immune checkpoint blockade therapy is calculated.

[0025] It should be noted that after the ordinary differential equation in the ODE solver is solved in this embodiment, two ODE solvers are set. The solution ranges of both ODE solvers are [,0.59]. One ODE solver gradually accumulates and calculates the integral term result based on the prediction scores fed back by the diffusion model corresponding to each time step. This integral term result is used for subsequent calculation of the joint probability, that is: ; Among them, is the prediction score, is the divergence of the prediction score, σ is a hyperparameter representing the maximum noise intensity, is the original sample without noise.

[0026] Another ODE solver adds noise to the 7-dimensional concatenated data at each time step to obtain intermediate data. After all time steps are completed, a data sample conforming to the Gaussian distribution is obtained, and the log-likelihood of the data sample is calculated, i.e.: ; wherein, is the noise perturbation characteristic at time during the diffusion process, represents the parameters of the trained diffusion model.

[0027] Solve the above two ODE solvers in the time range of [0, 0.59] to obtain a data sample conforming to the Gaussian distribution and the result of the integral term . Calculate the log-likelihood of the data sample ; wherein, is the likelihood of the data sample of the Gaussian distribution, is the data sample conforming to the Gaussian distribution, is the two-norm of.

[0028] Thus, the logarithm of the joint probability is obtained according to the following formula: ; wherein, is the joint probability in the response population , is the input data sample, is the noise perturbation characteristic at time during the diffusion process, is the response of the patient to the efficacy of immune checkpoint blockade therapy, is the prediction score, is the divergence of the prediction score, represents the model parameters of the trained diffusion model, and are the start time and end time of diffusion respectively.

[0029] The first part corresponds to the log-likelihood of the data sample of the Gaussian distribution at the end of the diffusion process. The second part is the integral term , is the cumulative divergence effect of the integral from to . In this embodiment, is preferably 0, is preferably 0.59.

[0030] In the above formula for calculating the joint probability, is used as the base likelihood to evaluate the matching degree between the patient's baseline characteristics and the response population, and the integral term is used as a dynamic correction to quantify the likelihood adjustment of the evolution of biometric characteristics during the treatment process. Generally, when it is a positive adjustment, and the characteristic evolution conforms to the response pattern. When it is a negative adjustment, and the characteristic deviates from the response pattern; finally, the joint probability The higher it is, the greater the likelihood that the patient will respond to the ICB (immune checkpoint blockade) therapy.

[0031] Based on the logarithm of the joint probability , the joint probability is obtained: .

[0032] Since the calculation formula of the joint probability in the non-response population is the same as the calculation process of , it will not be elaborated in this embodiment. Based on and , the response probability can be calculated: ; This response probability is the prediction of the patient's treatment effect. Specifically, it means that, on the premise of knowing the patient's relevant indicators, the response probability that the patient can "respond" to the ICB therapy.

[0033] In this embodiment, the diffusion model is transformed from an implicit generation model to an explicit probability density calculation model, and the calculation process has a solid theoretical basis and the calculation results are guaranteed. At the same time, an ODE solver is used in combination, and the convergence is faster and the accuracy is controllable.

[0034] In one of the embodiments, the training process of the diffusion model is as follows: S100. Obtain the 6-dimensional index characteristics of the patient and the preset treatment effect of the patient on the immune checkpoint blockade therapy to obtain 7-dimensional concatenated data; randomly sample the time step , and after adding noise to the 7-dimensional concatenated data, use it as a noisy sample to construct a training dataset, where is a uniform distribution, is the maximum end point of time; First, normalize the 6-dimensional index characteristics, use the treatment effect of the patient on the immune checkpoint blockade therapy as the 7th dimension data, normalize the 7th dimension data, and concatenate the two normalized characteristics to obtain 7-dimensional concatenated data.

[0035] The normalization process of the 6D index features is the same as that of the 7th - dimensional data, and its formula is: ; Where, is the original data, that is, the corresponding 6D index features or the 7th - dimensional data, is the normalized feature, and are the mean and variance of the training set used to train the diffusion model, respectively.

[0036] S200. Input the training data set into the first hidden layer of the feed - forward network; Set a stacking layer between two adjacent hidden layers, set a stacking layer between the second hidden layer and the output layer, and embed the corresponding time step into the noisy sample through Gaussian Fourier projection in the two stacking layers; Time step The Gaussian Fourier projection operation of is: ; Where, ; Where, is the linear - layer network parameter, represents the phase information of the concatenated sine - cosine function, is a randomly generated Gaussian matrix, is the concatenation operation.

[0037] After adding noise to the 7 - D concatenated data to construct the training data set as a noisy sample, the operation of calculating the time step corresponding to the noisy sample is as follows: ; Where, ; Where, represents the noise intensity corresponding to the time step , σ is a hyper - parameter representing the maximum noise intensity, represents the randomly sampled standard Gaussian noise, , is the normal distribution, is the identity matrix, represents the original sample without noise, represents the noisy sample after adding noise.

[0038] During the training process of the diffusion model, essentially, the model is made to learn how to perform noise prediction on noise samples at different time steps and with different noise intensities for denoising. The time step is embedded in the stacking layer, enabling the diffusion model to more accurately predict the corresponding score for the input noisy sample.

[0039] S300. Construct the loss function of the diffusion model to adjust the trainable parameters in the diffusion model.

[0040] Loss function of the diffusion model As follows: ; Wherein, represents the predicted score of the diffusion model, represents the noisy sample after adding noise.

[0041] As an embodiment, in this embodiment, the training process of the diffusion model is as follows: (a1) Obtain patient data; In this embodiment, the publicly available Chowell_train dataset is used for training the diffusion model. This dataset includes 11 indicators of 964 patients; in the test phase, the diffusion model is used to predict the curative effect on the Chowell_test dataset and evaluate the prediction results. This test set also includes 11 indicators of 515 patients.

[0042] (a2) Train the diffusion model; In this embodiment, the diffusion model is trained in the environment of Python3.10.14 and torch2.5.0, and the Adam optimizer is used for optimization. The relevant hyperparameter settings are: the epoch is set to 1200, the learning rate is set to 0.0001, and the maximum noise intensity is set to 25. Epoch refers to the process of one forward propagation and one backward propagation of the entire training dataset input into the diffusion model completely.

[0043] (a3) Predict the curative effect of the patient; According to Figure 1Process: Extract six indicators (Tumor Mutation Burden (TMB), Patient Systemic Treatment History (PSTH), blood albumin, Neutrophil-Lymphocyte Ratio (NLR), age, cancer type) of the patients in the test set as the original six-dimensional data; after performing normalized data processing, concatenate them with the seventh-dimensional data to obtain seven-dimensional concatenated data; input the seven-dimensional concatenated data into an ODE solver, and the solver uses the trained diffusion model to solve the ODE; after calculating the joint probability using the solution result, calculate the response probability as the prediction of whether the patient "responds" to the ICB efficacy.

[0044] (a4)Prediction result; To evaluate the prediction result of the diffusion model, this embodiment uses two indicators for evaluation: Area Under the ROC Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC).

[0045] The ROC curve is a curve plotted with the False Positive Rate (FPR) as the horizontal axis and the True Positive Rate (TPR) as the vertical axis. The value of AUC can evaluate the overall ability of the diffusion model to distinguish whether a patient responds to treatment, with a value range of [0, 1], where 0.5 represents random classification and 1.0 represents perfect classification.

[0046] The Precision-Recall curve is a curve plotted with the Recall as the horizontal axis and the Precision as the vertical axis. The value range of AUPRC is [0, 1], and the baseline is the proportion of positive samples. AUPRC reflects the model's ability to identify positive samples in highly imbalanced data.

[0047] Random guessing, Decision Tree, and Support Vector Machine (SVM) are used as comparative examples. Random guessing means that the classifier has no preference for all samples, and positive and negative samples are predicted as positive or negative with equal probability. Therefore, in the ROC curve, the True Positive Rate (TPR) and the False Positive Rate (FPR) are always equal. In the Precision-Recall curve, the precision is a constant value, which is consistent with the proportion of positive samples in the dataset.

[0048] The datasets used by the decision tree and the support vector machine are exactly the same as those used by the model in this embodiment. The specific implementation depends on the DecisionTreeClassifier function and the support vector machine classifier function (SVC function) provided by the sklearn toolkit; among them, DecisionTreeClassifier is a function in the scikit toolkit for constructing a decision tree classification model, and it adopts the CART algorithm (Classification and Regression Tree algorithm).

[0049] According to Figure 2 , it can be seen that the performance metric AUC of the method in this embodiment (corresponding to Ours in the figure) is significantly better than random guessing and also better than traditional classification methods such as decision trees and support vector machines, verifying the effectiveness of the method in this embodiment. Further, the ROC curve of the method in this embodiment rises significantly in the range of low false positive rates, indicating that the method in this embodiment is more sensitive to the recognition of positive samples.

[0050] According to Figure 3 , it can be seen that the performance metric AUPRC of the method in this embodiment (corresponding to Ours in the figure) is significantly better than random guessing and also better than decision trees and support vector machines, verifying the ability of the method in this embodiment to recognize positive samples on an unbalanced dataset. And the precision-recall curve has a very high precision in the range of low recall rates, indicating that the method in this embodiment is highly reliable when recognizing samples confirmed as positive.

[0051] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes an equivalent substitution or change, and should be covered by the protection scope of the present invention.

Claims

1. A method for predicting immune checkpoint blockade response based on a diffusion model, characterized in that, Including: Obtain the multi-dimensional index features of the patient and splice them to obtain 7-dimensional spliced data, and then input it into the ODE solver; The ODE solver inputs the currently solved time step and the current intermediate data into the trained diffusion model, and feeds back the obtained current prediction score to the ODE solver. The ODE solver iteratively inputs the next time step and the intermediate data corresponding to the next time step generated by the solution into the diffusion model until the ordinary differential equation in the ODE solver is solved. The ODE solver calculates the data sample and the integral term result that conform to the Gaussian distribution, and the intermediate data is the noise perturbation feature corresponding to the current time step; Calculate the joint probability by using the sum of the log-likelihood of the data sample and the integral term result; Based on the two states of the patient's response and non-response to the immune checkpoint blockade therapy, two joint probabilities are obtained. The ratio of the joint probability in the response state to the sum of the two joint probabilities is used as the response probability of the patient to the immune checkpoint blockade therapy.

2. The response prediction method according to claim 1, wherein The obtaining of the multi-dimensional index features of the patient and splicing them to obtain 7-dimensional spliced data is specifically as follows: Obtain the 6-dimensional index features of the patient and perform normalization processing as the original 6-dimensional data; Preset the efficacy of the patient's immune checkpoint blockade therapy as the 7th-dimensional data and splice it at the end of the original 6-dimensional data after normalization to obtain 7-dimensional spliced data.

3. The response prediction method according to claim 1, wherein During the process of the ODE solver solving the ordinary differential equation, one path of the ODE solver gradually accumulates and calculates the integral term result based on the prediction scores fed back by the diffusion model corresponding to each time step; the other path of the ODE solver is used to add noise to the 7-dimensional spliced data at each time step to obtain intermediate data. After all time steps are completed, a data sample that conforms to the Gaussian distribution is obtained, and the log-likelihood of the data sample is calculated.

4. The response prediction method according to claim 1, wherein The diffusion model uses a feedforward network as the noise prediction network, and the feedforward network includes two hidden layers connected in series and one output layer; The training process of the diffusion model is as follows: Obtain the 6-dimensional index features of the patient and the preset efficacy of the patient for immune checkpoint blockade therapy to obtain 7-dimensional concatenated data; randomly sample time steps , and after adding noise to the 7-dimensional concatenated data, use it as a noisy sample to construct a training dataset, where is a uniform distribution, is the maximum end point of time; Input the training data set into the first hidden layer of the feedforward network; Set a stacking layer between two adjacent hidden layers, set a stacking layer between the second hidden layer and the output layer, and embed the corresponding time step into the noisy sample through Gaussian Fourier projection in the two stacking layers; Construct the loss function of the diffusion model to adjust the trainable parameters in the diffusion model.

5. The response prediction method according to claim 4, wherein Time step Gaussian Fourier projection operation is as follows: Wherein, Among them, are the linear layer network parameters, represents the phase information of the concatenated sine and cosine functions, is a randomly generated Gaussian matrix, is the concatenation operation.

6. The response prediction method according to claim 4, wherein In constructing the training dataset with the noisy samples obtained by adding noise to the 7-dimensional concatenated data, the time step is calculated The operations corresponding to the noisy samples are as follows: Wherein, Among them, represents the time step corresponding noise intensity, σ is a hyperparameter representing the maximum noise intensity, represents the standard Gaussian noise of random sampling, represents the original sample without noise addition, represents the noisy sample after noise addition.

7. The response prediction method according to claim 6, wherein Loss function of the diffusion model As follows: Among them, represents the prediction score of the diffusion model, represents the noisy sample after adding noise.

8. The response prediction method according to claim 1, wherein The log-likelihood of the data sample The calculation formula is as follows: Among them, is the likelihood of data samples with a Gaussian distribution, is a hyperparameter, is a data sample conforming to a Gaussian distribution, is the two-norm of.

9. The response prediction method according to claim 8, wherein In the calculation of the joint probability obtained by using the sum of the log-likelihood of the data sample and the integral term result, when the patient's response to the immune checkpoint blockade therapy is a response, the logarithm of the joint probability The specific calculation formula is as follows: wherein, is the joint probability in the response population , is the original sample without noise, is the noise perturbation feature at time during the diffusion process, is the response of the patient's efficacy to immune checkpoint blockade therapy, is the prediction score, is the divergence of the prediction score, represents the model parameters of the diffusion model after training, and are the diffusion start time and diffusion end time respectively, is the integral from to the cumulative divergence effect, is a hyperparameter representing the maximum noise intensity, is the integration variable.

10. An immune checkpoint blockade response prediction system based on a diffusion model, characterized in that, Including a data acquisition module, a prediction module, a joint probability module, and a response probability module; The data acquisition module is used to obtain the multi-dimensional index features of the patient and splice them to obtain 7-dimensional spliced data, and then input it into the ODE solver; In the prediction module, the ODE solver inputs the currently solved time step and the current intermediate data into the trained diffusion model, and feeds the obtained current prediction score back to the ODE solver. The ODE solver takes the next time step solved and the intermediate data corresponding to the next time step as the current time step and iteratively inputs them into the diffusion model until the ordinary differential equation in the ODE solver is solved. The ODE solver calculates a data sample conforming to the Gaussian distribution and the integral term result, and the intermediate data is the noise perturbation feature corresponding to the current time step; In the joint probability module, the joint probability is calculated using the sum of the log-likelihood of the data sample and the integral term result; In the response probability module, based on the two states of the patient's response and non-response to the immune checkpoint blockade therapy, two joint probabilities are obtained, and the ratio of the joint probability in the response state to the sum of the two joint probabilities is used as the patient's response probability to the immune checkpoint blockade therapy.

Images