Tumor drug sensitivity prediction system and method based on lightweight neural network
By adopting lightweight neural networks and a variety of advanced technologies in the tumor drug sensitivity prediction system, the existing system's problems such as low computing efficiency, poor adaptability and insufficient interpretability are solved, efficient and accurate prediction and strong generalization capabilities are achieved, and clinical application value is enhanced.
Patent Information
- Application Number
- CN202411685685.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-23
- Publication Date
- 2025-03-18
AI Technical Summary
The existing tumor drug sensitivity prediction system has problems such as inefficient computing efficiency, poor adaptability, insufficient interpretability, limited clinical application and weak generalization ability.
A prediction system based on lightweight neural network is adopted to build an efficient, flexible, interpretable and powerful generalization prediction system through technologies such as model compression, transfer learning, knowledge distillation and integrated learning.
It realizes efficient and accurate tumor drug sensitivity prediction, reduces computational complexity, improves the adaptability and interpretability of the system, and enhances its value in clinical applications.
Smart Images
Figure CN119650096A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tumor drug sensitivity prediction systems, and in particular to a tumor drug sensitivity prediction system based on a lightweight neural network and a method thereof. Background Art
[0002] With the rapid development of precision medicine, the prediction of tumor drug sensitivity has become a key link in personalized treatment. In recent years, the application of machine learning and deep learning technologies in this field has made significant progress. However, the existing prediction methods still have some problems that need to be solved.
[0003] First, traditional drug sensitivity prediction models usually rely on large-scale neural networks. Although these models have high prediction accuracy, they are computationally complex and have strict requirements on hardware resources, making them difficult to be widely used in actual clinical environments. In particular, in some areas with limited medical resources, such high-computational-cost models are almost impossible to deploy.
[0004] Secondly, existing prediction systems are often optimized for specific types of tumors or drugs, lacking sufficient versatility and flexibility. When faced with new tumor types or newly developed drugs, these systems need to re-collect a large amount of data and train from scratch, which is not only time-consuming and labor-intensive, but also greatly delays the clinical application of new drugs.
[0005] Furthermore, many current prediction models use "black box" deep learning algorithms, which can give prediction results but lack interpretability. This is unacceptable in high-risk areas such as medical decision-making, because both doctors and patients need to understand the reasons behind the prediction results.
[0006] In addition, most existing systems focus only on the prediction itself, while ignoring the application of the prediction results in clinical practice. Simply giving the prediction results of drug sensitivity without providing specific dosing recommendations is of limited practical help to clinicians.
[0007] Finally, the prediction models in the existing technologies generally have the problem of insufficient generalization ability. They perform well on the training data, but when faced with the complex situations in the real world, the prediction accuracy is often greatly reduced. This "overfitting" phenomenon seriously limits the application value of these models in actual clinical settings. Summary of the invention
[0008] In view of the above problems, the present invention proposes a tumor drug sensitivity prediction system and method based on a lightweight neural network. The system aims to solve the problems of low computational efficiency, poor adaptability, insufficient interpretability, limited clinical applicability and weak generalization ability in the prior art.
[0009] The present invention proposes a tumor drug sensitivity prediction system based on a lightweight neural network, the system comprising:
[0010] A model compression module, used to obtain a compressed model;
[0011] A transfer learning module, connected to the model compression module, for quickly training a new model according to a transfer learning framework;
[0012] A data set construction module is used to construct a data set for a new target object;
[0013] A knowledge distillation module, connected to the transfer learning module and the data set construction module, for further compressing the model volume using a knowledge distillation method;
[0014] An integrated learning module, connected to the knowledge distillation module and the data set construction module, for quickly training a new model with drug sensitivity scoring and integrating multiple models;
[0015] A prediction result output module, connected to the integrated learning module, for outputting the final tumor drug sensitivity prediction result;
[0016] A preprocessing module, connected to the data set construction module, for normalizing the training set and the test set of the model, and dividing the training set and the test set according to the labels; and
[0017] The model updating module is connected to the integrated learning module and the prediction result output module, and is used to retrain a new model based on the set results and the new test set, and calculate the loss of the model.
[0018] Preferably, the knowledge distillation module is further used for:
[0019] Combine the parameters of the existing network model and the parameters of the lightweight model into a new loss function;
[0020] The new loss function and the true loss function of the new model are added with a preset proportional coefficient to jointly optimize the objective functions of the two models; and
[0021] The new model is trained based on the jointly optimized objective function.
[0022] Preferably, the integrated learning module comprises:
[0023] The prediction submodule is used to make a preliminary prediction of the drug sensitivity of tumor cells;
[0024] an integration submodule, connected to the prediction submodule, for integrating multiple prediction models, including an integrated learning model constructed using the trained new model and the new test set; and
[0025] The voting submodule is connected to the integration submodule and is used to calculate the final score by using the voting method of the new integration model, wherein when the voting result is 1, it corresponds to sensitivity.
[0026] Preferably, the system further comprises:
[0027] A medication regimen generation module, connected to the prediction result output module, for generating a personalized medication regimen according to the prediction results of the prediction system;
[0028] A data input module, for receiving tumor cell or tumor drug sensitivity data as input; and
[0029] The result output module is connected to the medication regimen generation module and is used to output the final personalized medication regimen.
[0030] The method for rapidly predicting tumor drug sensitivity based on the system comprises the following steps:
[0031] Using the model compression module to obtain compressed model data;
[0032] Preprocessing the training set and the test set of the model using the preprocessing module;
[0033] Initialize a new model using the transfer learning module and train it on the training set of the new model;
[0034] Constructing a loss function of a knowledge distillation model using the knowledge distillation module;
[0035] Using the integrated learning module to predict the drug sensitivity of tumor cells and integrating multiple prediction models;
[0036] Outputting the final tumor drug sensitivity prediction result using the prediction result output module;
[0037] When the model update module detects that the loss no longer decreases within a preset time, the training is stopped and the new model is tested on a new task based on a new test set.
[0038] Preferably, the step of constructing the loss function of the knowledge distillation model using the knowledge distillation module specifically includes:
[0039] The parameters of the two models are combined into a loss function L, which is composed of the objective function L1 of the new model and the objective function L2 of the new model;
[0040] Set the scaling factor λ and construct the final loss function: L = λL1 + (1-λ)L2;
[0041] On the new data set, the drug sensitivity of tumor cells is trained so that the parameters of the new model are optimized to minimize the loss L.
[0042] Preferably, the step of predicting the drug sensitivity of tumor cells using an integrated learning module specifically includes:
[0043] Extract training and test sets of new datasets;
[0044] Initialize a new prediction model, which contains multiple weak learners, the number of weak learners is M, 1≤m≤M;
[0045] Use the training set to train each weak learner m to obtain its prediction result;
[0046] Select a data set and a weak learner m, and use the data set to retrain the weak learner m;
[0047] Set the scaling factor γ of the weak learner and construct the final weak learner total loss function based on the prediction results of the weak learner;
[0048] The ensemble model is optimized based on the final weak learner total loss function.
[0049] As an advantage, the method further comprises the following steps:
[0050] Receive tumor cell or tumor drug sensitivity data as input and construct a dataset for a new task;
[0051] Using the tumor drug sensitivity prediction system, predicting the sensitivity of tumor cells or tumor drugs in a new task;
[0052] Optimize the prediction results and get the final results;
[0053] Generate a personalized dosing regimen based on the final results.
[0054] The beneficial technical effects brought by the technical solution of the present invention are:
[0055] The core of this invention is that it innovatively integrates a variety of advanced technologies to form an efficient, flexible, interpretable and powerful prediction system. From a macro perspective, the system greatly reduces the computational complexity through model compression and knowledge distillation technology, making high-precision predictions possible, and can be smoothly deployed and run even in environments with limited computing resources. This not only improves the availability of the system, but also greatly expands its scope of application.
[0056] From a microscopic perspective, the various modules of the present invention form a collaboratively optimized whole. For example, the cooperation between the transfer learning module and the dataset construction module enables the system to quickly adapt to new tumor types or drugs, greatly improving the flexibility and efficiency of the system. The combination of the knowledge distillation module and the integrated learning module cleverly solves the contradiction between model compression and performance retention, achieving the effect of "having your cake and eating it too".
[0057] It is particularly worth mentioning that the present invention enhances the interpretability of the model by introducing a rule-based expert system. This not only meets the high requirements for transparency in the medical field, but also provides a new idea for integrating machine learning models with traditional medical knowledge. At the same time, the addition of a dosing regimen generation module directly converts the prediction results into clinically applicable recommendations, greatly enhancing the practical value of the system.
[0058] In addition, the integrated learning strategy adopted by the present invention effectively improves the generalization ability of the model. By integrating the prediction results of multiple sub-models, the system can better cope with complex situations in actual clinical environments and significantly reduce the risk of overfitting.
[0059] In general, the present invention not only solves many problems in the prior art, but also achieves a qualitative leap in efficiency, flexibility, interpretability and clinical applicability. Its successful application will greatly promote the development of precision medicine, bring more personalized and accurate treatment plans to cancer patients, and ultimately improve the treatment effect and patient prognosis. The significance of this system is not limited to the technical level, but also reflects the huge potential of the deep application of artificial intelligence technology in the medical field, and points out the direction for the transformation of future medical models. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is the overall logic block diagram of the system of the present invention.
[0061] Figure 2 This is the logic block diagram of the knowledge distillation module of the present invention.
[0062] Figure 3 This is the logic block diagram of the integrated learning module of the present invention.
[0063] Figure 4 This is the logic block diagram of the model update module of the present invention.
[0064] Figure 5 This is a logic block diagram of the preprocessing module of the present invention. DETAILED DESCRIPTION
[0065] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.
[0066] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0067] See also Figure 1-5 The purpose of the present invention is to provide a tumor drug sensitivity prediction system and method based on a lightweight neural network to solve the problem of low efficiency and low accuracy of tumor drug sensitivity prediction in the prior art. To achieve the above purpose, the present invention provides the following technical solutions:
[0068] First, the present invention proposes a tumor drug sensitivity prediction system based on a lightweight neural network, which includes multiple functional modules that work together to achieve efficient and accurate tumor drug sensitivity prediction. Specifically, the system includes: a model compression module 1, a transfer learning module 2, a data set construction module 3, a knowledge distillation module 4, an integrated learning module 5, a prediction result output module 6, a preprocessing module 7, and a model update module 8.
[0069] The model compression module 1 is used to obtain a compressed model. In one embodiment of the present invention, the module uses pruning technology to compress the original neural network model, reducing the model size to about 30% of the original size while maintaining the model performance without significant degradation. The advantage of this is that the computational complexity and storage requirements of the model are greatly reduced, allowing the model to run quickly in a resource-constrained environment.
[0070] Next, the transfer learning module 2 is connected to the model compression module 1 to quickly train a new model according to the transfer learning framework. Specifically, the module uses a model pre-trained on a large-scale data set as a starting point, and then fine-tunes it on specific tumor drug sensitivity data. This method can significantly reduce training time and improve the performance of the model on new tasks. Preferably, the present invention adopts a fine-tuning strategy, and only fine-tunes the last few layers of the model, which not only retains the underlying feature extraction capabilities, but also can adapt to the characteristics of new tasks.
[0071] Secondly, the data set construction module 3 is used to construct a data set for a new target object. In the task of predicting tumor drug sensitivity, data quality directly affects the accuracy of the prediction results. Therefore, this module adopts a multi-source data fusion method, combining gene expression data, drug molecular structure information and clinical response data to construct a comprehensive data set. Preferably, in one embodiment of the present invention, a 10-fold cross-validation method is used to evaluate the model performance to ensure the reliability of the results.
[0072] Thirdly, the knowledge distillation module 4 is connected with the transfer learning module 2 and the data set construction module 3 to further compress the model size using the knowledge distillation method. The core idea of this module is to use a complex "teacher" model to guide the learning of a simple "student" model. Specifically, a large neural network model is first trained as a teacher model, and then the output of this teacher model is used to train a small student model. This method can not only reduce the model size, but also maintain or even improve the model performance to a certain extent.
[0073] The integrated learning module 5 is connected to the knowledge distillation module 4 and the data set construction module 3, and is used to quickly train a new model with drug sensitivity scoring and integrate multiple models. This module adopts two main integrated learning methods, Bagging and Boosting. The Bagging method trains multiple basic models in parallel, and then averages or votes on their prediction results to get the final result. The Boosting method trains multiple models sequentially, and each new model attempts to correct the errors of the previous model. The combination of these two methods greatly improves the generalization ability and prediction accuracy of the model.
[0074] The prediction result output module 6 is connected to the integrated learning module 5 and is used to output the final tumor drug sensitivity prediction result. This module not only outputs the predicted sensitivity score, but also provides a credibility assessment of the prediction. Preferably, in one embodiment of the present invention, a probability calibration technique is used to convert the original output of the model into a more reliable probability estimate, which is particularly important for clinical decision support systems.
[0075] The preprocessing module 7 is connected to the data set construction module 3, and is used to normalize the training set and test set of the model, and divide the training set and test set according to the label. Data preprocessing is a key step in machine learning and has a significant impact on model performance. This module first standardizes the numerical features and scales all features to the same scale. Then, the categorical features are uniquely encoded. Finally, the stratified sampling method is used to divide the training set and the test set to ensure that the proportion of each type of sample in the two sets is consistent with the original data set.
[0076] The model update module 8 is connected to the ensemble learning module 5 and the prediction result output module 6, and is used to retrain the new model based on the set results and the new test set, and calculate the loss of the model. This module realizes the online learning ability of the model, and can continuously learn from new data to improve the prediction performance of the model. Specifically, when new test data arrives, the model will calculate the loss based on the prediction results and the actual labels, and then use the gradient descent method to update the model parameters. Preferably, in one embodiment of the present invention, an Adam optimizer is used, which can adaptively adjust the learning rate and accelerate the convergence of the model.
[0077] Next, the present invention further refines the function of the knowledge distillation module 4. This module is not just a simple knowledge transfer from the teacher model to the student model, but adopts an innovative loss function design. Specifically, the knowledge distillation module 4 is also used to:
[0078] First, the parameters of the existing network model and the parameters of the lightweight model are combined into a new loss function. This new loss function takes into account two aspects: one is the prediction accuracy of the student model for the original label, and the other is the similarity between the output of the student model and the output of the teacher model.
[0079] Then, the new loss function and the true loss function of the new model are added with a preset proportional coefficient to jointly optimize the objective functions of the two models. The preset proportional coefficient here is a hyperparameter that needs to be tuned, usually between 0.5 and 0.7. Preferably, in one embodiment of the present invention, a grid search method is used to determine the optimal proportional coefficient, and 0.6 is finally selected as the optimal value.
[0080] Finally, the new model is trained based on the jointly optimized objective function. This training method not only allows the student model to learn the knowledge of the original task, but also learns the "dark knowledge" of the teacher model, that is, the subtle differences in the teacher model's judgments on each category.
[0081] The present invention also describes in detail the internal structure of the integrated learning module 5. The module includes three submodules: a prediction submodule, an integration submodule and a voting submodule.
[0082] The prediction submodule is used to make a preliminary prediction of the drug sensitivity of tumor cells. This submodule contains multiple basic models, each of which independently predicts the input data. Preferably, in one embodiment of the present invention, multiple machine learning algorithms such as random forest, gradient boosting tree and support vector machine are used as basic models.
[0083] The integration submodule is connected to the prediction submodule and is used to integrate multiple prediction models, including an integrated learning model constructed using a trained new model and a new test set. This submodule adopts the Stacking integration method, that is, the prediction results of the basic model are used as features to train a new model to obtain the final prediction. Preferably, in one embodiment of the present invention, logistic regression is used as a meta-learner for Stacking.
[0084] The voting submodule is connected to the integration submodule, and is used to calculate the final score using the voting method of the new integrated model. Specifically, when the voting result is 1, it corresponds to sensitive. The voting here adopts a weighted voting method, and the weight of each model is dynamically adjusted according to its performance on the validation set. Preferably, in one embodiment of the present invention, 0.5 is set as the threshold for sensitivity judgment, that is, when the weighted voting result is greater than 0.5, it is judged to be sensitive.
[0085] In addition, the present invention also adds a medication regimen generation module 9 to the system, which is connected to the prediction result output module 6 and is used to generate a personalized medication regimen according to the prediction results of the prediction system. The addition of this module enables the system to not only predict drug sensitivity, but also directly provide medication recommendations to clinicians, greatly improving the practicality of the system.
[0086] Preferably, in one embodiment of the present invention, the medication regimen generation module 9 adopts a method combining a rule-based expert system and a machine learning model. The expert system contains a large number of clinical medication rules, while the machine learning model can learn the optimal medication strategy from historical data. The combination of the two ensures the safety of the medication regimen and improves its personalization.
[0087] The medication regimen generation module 9 is matched with a data input module 10 and a result output module 11. The data input module 10 is used to receive tumor cell or tumor drug sensitivity data as input. This module supports input of multiple data formats, including but not limited to gene expression spectrum data, drug molecular structure data, etc. The result output module 11 is connected to the medication regimen generation module 9 and is used to output the final personalized medication regimen. The output regimen includes not only the recommended drug type and dosage, but also the expected therapeutic effect and possible side effects.
[0088] Next, the present invention proposes a method for rapidly predicting tumor drug sensitivity based on the above system. The method comprises the following steps:
[0089] First, the compressed model data is obtained by using the model compression module 1. This step is usually completed in the offline stage of the system, with the purpose of providing an efficient basic model for subsequent online prediction.
[0090] Then, the training set and test set of the model are preprocessed using the preprocessing module 7. This step is the data preparation stage to ensure the quality and consistency of the input data. Preferably, in one embodiment of the present invention, the continuous features are z-score standardized and the categorical features are one-hot encoded.
[0091] Secondly, a new model is initialized using the transfer learning module 2 and trained on the training set of the new model. This step greatly reduces the time and computing resources required for training the model from scratch through the transfer learning technology. Preferably, in one embodiment of the present invention, a fine-tuning strategy is adopted to fine-tune only the last few layers of the model and freeze the parameters of other layers.
[0092] Again, the loss function of the knowledge distillation model is constructed using the knowledge distillation module 4. This step is the core of knowledge distillation. Through a carefully designed loss function, the "knowledge" of the large model is effectively transferred to the small model.
[0093] Next, the drug sensitivity of tumor cells is predicted using the integrated learning module 5, and multiple prediction models are integrated. This step improves the accuracy and stability of the final prediction by integrating the prediction results of multiple models.
[0094] Then, the final tumor drug sensitivity prediction result is outputted using the prediction result output module 6. This step not only outputs the prediction result, but also provides a credibility assessment of the prediction.
[0095] Finally, when the model update module 8 detects that the loss no longer decreases within a preset time, the training is stopped, and the new model is tested on a new task based on a new test set. This step realizes automatic updating and performance evaluation of the model. Preferably, in one embodiment of the present invention, an EarlyStopping mechanism is set, and when the loss on the validation set does not decrease within 10 consecutive epochs, the training is stopped.
[0096] The present invention also describes in detail the steps of constructing the loss function of the knowledge distillation module 4. Specifically, it includes:
[0097] First, the parameters of the two models are combined into a loss function L, which consists of the objective function L1 of the new model and the objective function L2 of the new model. Here L1 is usually the cross entropy loss, which measures the gap between the model prediction and the true label; L2 is the KL divergence, which measures the gap between the output of the student model and the output of the teacher model.
[0098] Then, the scaling factor λ is set to construct the final loss function: L = λL1 + (1-λ)L2. This λ is a hyperparameter that needs to be tuned to balance the weights of the two losses. Preferably, in one embodiment of the present invention, the optimal value of λ is determined to be 0.6 through cross-validation.
[0099] Finally, on the new data set, the drug sensitivity of tumor cells is trained so that the parameters of the new model are optimized to minimize the loss L. This step updates the model parameters by back propagation algorithm and gradient descent method. Preferably, in one embodiment of the present invention, an Adam optimizer is used, the learning rate is set to 0.001, and the batch size is 64.
[0100] The present invention also describes in detail the steps of predicting the drug sensitivity of tumor cells by the integrated learning module 5. Specifically, it includes:
[0101] First, extract the training set and test set of the new data set. This step usually adopts the method of random stratified sampling to ensure that the ratio of each type of sample in the training set and the test set is consistent with the original data set. Preferably, in one embodiment of the present invention, the training set and the test set are divided in a ratio of 8:2.
[0102] Secondly, a new prediction model is initialized, which includes multiple weak learners, and the number of weak learners is M, 1≤m≤M. The weak learners here can be different types of machine learning algorithms, or different variants of the same type of algorithms. Preferably, in one embodiment of the present invention, M=10 is set, including 3 random forests, 3 gradient boosting trees, 2 support vector machines and 2 K nearest neighbor classifiers.
[0103] Then, each weak learner m is trained using the training set to obtain its prediction results. This step is performed in parallel, and each weak learner is trained independently to improve efficiency.
[0104] Next, a data set and a weak learner m are selected, and the weak learner m is retrained using the data set. This step is mainly to further improve the performance of the weak learner. Preferably, in one embodiment of the present invention, a k-fold cross validation method is used, and k is set to 5.
[0105] Again, the scaling factor γ of the weak learner is set, and the final weak learner total loss function is constructed according to the prediction results of the weak learner. Here, γ is a weight vector used to adjust the importance of different weak learners in the final prediction. Preferably, in one embodiment of the present invention, all γ are initially set to 1 / M, and then dynamically adjusted according to the performance of each weak learner on the validation set.
[0106] Finally, the ensemble model is optimized based on the final weak learner total loss function. This step usually uses gradient descent to update γ to minimize the overall loss.
[0107] The present invention also proposes a method for optimizing a personalized drug administration regimen for tumor cells, the method comprising the following steps:
[0108] First, tumor cell or tumor drug sensitivity data is received as input and a data set for a new task is constructed. This step is completed by a data input module 10, which can process input data in various formats, including but not limited to gene expression profile data, drug molecular structure data, etc.
[0109] Secondly, the tumor drug sensitivity prediction system is used to predict the sensitivity of tumor cells or tumor drugs in the new task. This step makes full use of the various modules of the prediction system described above, including model compression, transfer learning, knowledge distillation and ensemble learning.
[0110] Then, the prediction results are optimized to obtain the final results. This step is mainly completed by integrating the learning module 5 and the prediction result output module 6. Preferably, in one embodiment of the present invention, the Softmax function is used to calibrate the probability of the original prediction score, so that the output result is more interpretable.
[0111] Finally, a personalized medication regimen is generated based on the final result. This step is completed by the medication regimen generation module 9, which combines a rule-based expert system and a machine learning model. Preferably, in one embodiment of the present invention, the medication regimen includes not only the recommended drug type and dosage, but also the expected therapeutic effect and possible side effects, as well as matters that require special attention during the treatment process.
[0112] In addition, the present invention also proposes a tumor cell or tumor drug sensitivity prediction system for new tasks or new data sets. The system includes:
[0113] The big model is obtained by training the new test set and the new model using the above method. This big model is usually a complex neural network with powerful feature extraction and prediction capabilities.
[0114] The small model is also obtained by training the new test set and the new model using the above method. This small model is obtained by extracting key knowledge from the large model through knowledge distillation technology, and has a smaller model size and faster reasoning speed.
[0115] A model selection module is connected to the large model and the small model, and is used to select the large model, the small model, or one or more of their prediction results for prediction according to preset conditions. The function of this module is to make a trade-off between accuracy and efficiency. Preferably, in one embodiment of the present invention, when system resources are sufficient and the prediction accuracy requirement is extremely high, the large model is selected for use; when a fast response is required or when operating in a resource-constrained environment, the small model is selected for use.
[0116] The prediction result output module is connected to the model selection module and is used to output the final prediction result of tumor cells or tumor drug sensitivity. This module not only outputs the prediction result, but also provides a credibility assessment of the prediction. Preferably, in one embodiment of the present invention, the Monte Carlo Dropout technique is used to estimate the uncertainty of the prediction.
[0117] In general, the tumor drug sensitivity prediction system and method based on lightweight neural network proposed in the present invention realizes efficient and accurate prediction of tumor drug sensitivity by innovatively combining advanced technologies such as model compression, transfer learning, knowledge distillation and ensemble learning. The system can not only quickly adapt to new prediction tasks, but also flexibly select models of different scales according to actual needs, providing reliable decision support for clinicians. In addition, the system also has the ability to generate personalized dosing regimens, further improving its application value in actual medical scenarios. The successful application of this system is expected to greatly improve the effect of personalized tumor treatment and bring better treatment results to patients.
[0118] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. A tumor drug sensitivity prediction system based on a lightweight neural network, characterized in that: The system comprises: A model compression module, used to obtain a compressed model; A transfer learning module, connected to the model compression module, for quickly training a new model according to a transfer learning framework; A data set construction module is used to construct a data set for a new target object; A knowledge distillation module, connected to the transfer learning module and the data set construction module, for further compressing the model volume using a knowledge distillation method; An integrated learning module, connected to the knowledge distillation module and the data set construction module, for quickly training a new model with drug sensitivity scoring and integrating multiple models; A prediction result output module, connected to the integrated learning module, for outputting the final tumor drug sensitivity prediction result; A preprocessing module, connected to the data set construction module, for normalizing the training set and the test set of the model, and dividing the training set and the test set according to the labels; and The model updating module is connected to the integrated learning module and the prediction result output module, and is used to retrain a new model based on the set results and the new test set, and calculate the loss of the model.
2. The tumor drug sensitivity prediction system according to claim 1, characterized in that: The knowledge distillation module is also used to: Combine the parameters of the existing network model and the parameters of the lightweight model into a new loss function; The new loss function and the true loss function of the new model are added with a preset proportional coefficient to jointly optimize the objective functions of the two models; and The new model is trained based on the jointly optimized objective function.
3. The tumor drug sensitivity prediction system according to claim 1, characterized in that: The integrated learning module includes: The prediction submodule is used to make a preliminary prediction of the drug sensitivity of tumor cells; an integration submodule, connected to the prediction submodule, for integrating multiple prediction models, including an integrated learning model constructed using the trained new model and the new test set; and The voting submodule is connected to the integration submodule and is used to calculate the final score by using the voting method of the new integration model, wherein when the voting result is 1, it corresponds to sensitivity.
4. The tumor drug sensitivity prediction system according to claim 1, characterized in that: The system further comprises: A medication regimen generation module, connected to the prediction result output module, for generating a personalized medication regimen according to the prediction results of the prediction system; A data input module, for receiving tumor cell or tumor drug sensitivity data as input; and The result output module is connected to the medication regimen generation module and is used to output the final personalized medication regimen.
5. A method for rapidly predicting tumor drug sensitivity based on the system according to any one of claims 1 to 4, characterized in that: The following steps are involved: Using the model compression module to obtain compressed model data; Preprocessing the training set and the test set of the model using the preprocessing module; Initialize a new model using the transfer learning module and train it on the training set of the new model; Constructing a loss function of a knowledge distillation model using the knowledge distillation module; Using the integrated learning module to predict the drug sensitivity of tumor cells and integrating multiple prediction models; Outputting the final tumor drug sensitivity prediction result using the prediction result output module; When the model update module detects that the loss no longer decreases within a preset time, the training is stopped and the new model is tested on a new task based on a new test set.
6. The method according to claim 5, characterized in that The step of using the knowledge distillation module to construct the loss function of the knowledge distillation model specifically includes: The parameters of the two models are combined into a loss function L, which is composed of the objective function L1 of the new model and the objective function L2 of the new model; Set the scaling factor λ and construct the final loss function: L = λL1 + (1-λ)L2; On the new data set, the drug sensitivity of tumor cells is trained so that the parameters of the new model are optimized to minimize the loss L.
7. The method according to claim 5, characterized in that The step of using the integrated learning module to predict the drug sensitivity of tumor cells specifically includes: Extract training and test sets of new datasets; Initialize a new prediction model, which contains multiple weak learners, the number of weak learners is M, 1≤m≤M; Use the training set to train each weak learner m to obtain its prediction result; Select a data set and a weak learner m, and use the data set to retrain the weak learner m; Set the scaling factor γ of the weak learner and construct the final weak learner total loss function based on the prediction results of the weak learner; The ensemble model is optimized based on the final weak learner total loss function.
8. The method according to claim 5, characterized in that The following steps are also included: Receive tumor cell or tumor drug sensitivity data as input and construct a dataset for a new task; Using the tumor drug sensitivity prediction system, predicting the sensitivity of tumor cells or tumor drugs in a new task; Optimize the prediction results and get the final results; Generate a personalized dosing regimen based on the final results.