Disease subtype prediction method, system, device and medium based on small sample

Abstract

The present invention provides a disease subtype prediction method, system, device and medium based on small samples. The method includes obtaining gene expression data of a first data set; predicting the gene expression data through a prediction model to obtain the disease subtype of the sample to be predicted; the prediction model includes a sample selection network, a feature selection layer and a meta-learner; the method trains a prediction model for disease subtype prediction through a meta-learner, learns from related clinical tasks through meta-learning technology to extract valuable information, so as to help the model be well generalized to the prediction task of disease subtypes; in the training process of the prediction model, through the process of feature selection and sample reweighting, by adaptively removing noise data, the dimensionality curse problem is solved, and it can be widely used in the field of machine learning technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular to a method, system, device and medium for predicting disease subtypes based on small samples. Background Art

[0002] Disease subtype prediction is the identification of subgroups of similar patients, which can guide treatment decisions for specific individuals. For example, in the past 15 years, five subtypes of breast cancer have been identified and studied in depth. At the molecular biology level, using gene expression data to predict disease subtypes is of great significance for improving the accuracy of disease diagnosis and identifying potential disease-related genes. However, a challenging problem is that gene expression data is well-known for small sample data, that is, for each disease subtype, we only have relatively few samples. In recent years, more and more machine learning research has begun to tend to small sample learning.

[0003] However, unlike images, gene expression data is difficult to analyze due to its high dimensionality and high noise characteristics. The dimensionality problem often makes prediction more challenging because a large number of redundant features are involved in decision making. In addition, gene expression data inevitably has high noise, which can easily cause the model to overfit, resulting in poor generalization performance of the model. Summary of the invention

[0004] In view of this, in order to at least partially solve one of the above-mentioned technical problems, an embodiment of the present invention aims to provide a disease subtype prediction method based on a small sample, by adding two steps of feature selection and sample selection, so that it can filter out disease-related genes and remove noise data, thereby achieving accurate prediction of disease subtypes; at the same time, an embodiment of the present invention also provides a system, device and medium that can implement a disease subtype prediction method based on a small sample.

[0005] In a first aspect, an embodiment of the present invention provides a method for predicting disease subtypes based on a small sample, which comprises the following steps: obtaining gene expression data of a first data set;

[0006] The gene expression data is predicted by a prediction model to obtain the disease subtype of the first data set; the prediction model includes a sample selection network, a feature selection layer, and a meta-learner;

[0007] The prediction model is trained through the following steps:

[0008] Constructing a first sample data set; the sample data set includes gene expression data of the second data set;

[0009] Obtaining a feature weighted vector of the training data set through a feature selection layer; constructing a second sample data set based on the feature weighted vector;

[0010] Inputting the second sample data set into the sample selection network to obtain sample weights;

[0011] The meta-learner is trained according to the second sample data set and the sample weights to obtain a trained prediction model.

[0012] In some embodiments of the present invention, the step of constructing a sample data set specifically includes:

[0013] Acquire a number of gene expression data and obtain a first sample data set through meta-analysis;

[0014] Alternatively, a number of gene expression data are obtained, and integrated to obtain a first sample data set through batch correction and machine learning.

[0015] In some embodiments of the present invention, the step of constructing a sample data set further includes:

[0016] Extracting a plurality of support samples and a plurality of query samples from the second data set according to categories of the gene expression data;

[0017] The support set is constructed based on the support samples; the query set is constructed based on the query samples.

[0018] In some embodiments of the present invention, the step of inputting the second sample data set into the sample selection network to obtain the sample weight is specifically as follows:

[0019] The confidence of the sample data in the second sample data set is determined, and the sample weight is assigned according to the confidence.

[0020] In some embodiments of the present invention, the step of determining the confidence of the sample data in the second sample data set further includes:

[0021] Determine a loss function for sample data in a second sample data set, embed the loss function into the sample data,

[0022] The confidence of the sample data is obtained by fitting the weighted function through the neural network.

[0023] In some embodiments of the present invention, the step of training a meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model includes:

[0024] Remove noise samples in the support set according to sample weights; determine the sample categories in the support set after denoising;

[0025] The Euclidean distance between the samples and sample categories in the denoised support set is obtained, and the Euclidean distance is normalized to obtain the output of the meta-learner.

[0026] In some embodiments of the present invention, the step of training a meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model further includes:

[0027] According to the query set, the output of the meta-learner is verified, and the trained prediction model is output based on the verification results.

[0028] In a second aspect, the technical solution of the present invention also provides a system for predicting disease subtypes based on small samples, including a data acquisition unit, a model building unit, and a prediction output unit, wherein:

[0029] A data acquisition unit, used to acquire gene expression data of a first data set;

[0030] The model building unit is used to train the prediction model, which includes: a sample selection network, a feature selection layer, and a meta-learner; the training steps of the prediction model include:

[0031] Constructing a first sample data set; the sample data set includes gene expression data of a second data set; obtaining a feature weight vector of a training data set through a feature selection layer; constructing a second sample data set according to the feature weight vector; inputting the second sample data set into a sample selection network to obtain a sample weight; training a meta-learner according to the second sample data set and the sample weight to obtain a trained prediction model;

[0032] The prediction output unit is used to predict the gene expression data through the prediction model to obtain the disease subtype of the first data set sample.

[0033] In a third aspect, the technical solution of the present invention also provides a device for predicting disease subtypes based on a small sample, which comprises:

[0034] at least one processor;

[0035] at least one memory for storing at least one program;

[0036] When at least one program is executed by at least one processor, the at least one processor implements the disease subtype prediction method based on small samples in the first aspect.

[0037] In a fourth aspect, the technical solution of the present invention further provides a storage medium, in which a program executable by a processor is stored. When the program executable by the processor is executed by the processor, it is used to implement the method in the first aspect.

[0038] The advantages and beneficial effects of the present invention will be partially given in the following description, and the other parts can be understood through the specific embodiments of the present invention:

[0039] The small sample-based disease subtype prediction method provided in the embodiment of the present invention obtains a prediction model for disease subtype prediction through a meta-learner, and learns from related clinical tasks through meta-learning technology to extract valuable information to help the model be well generalized to the prediction task of disease subtypes; in the training process of the prediction model, through the process of feature selection and sample reweighting, the noise data is adaptively removed to solve the dimensionality curse problem, and it has certain advantages in predicting disease subtypes and identifying potential disease correlations. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 This is a flowchart of the steps of a method for predicting disease subtypes based on a small sample according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of the steps of a training method for a disease subtype prediction model based on a small sample in an embodiment of the present invention;

[0043] Figure 3 A curve diagram of the training loss of Select-ProtoNet under unbiased data and the training loss of ProtoNet under unbiased data in an embodiment of the present invention;

[0044] Figure 4 A curve diagram showing the accuracy of Select-ProtoNet under unbiased data and the training accuracy of ProtoNet under unbiased data according to an embodiment of the present invention;

[0045] Figure 5 This is the sample weight distribution of the training data of Select-ProtoNet under a noise rate of 30% according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0047] The technical solution provided by the embodiment of the present invention solves the problem of predicting disease subtypes with small samples. A commonly used method in genomic data is to expand the sample size by aggregating data from multiple studies under comparable conditions or treatments. Due to the complex characteristics of gene expression data, methods usually encounter bottlenecks, and the fusion of data from different platforms or experiments will inevitably be affected by batch effects, heterogeneity and other deviations. In addition, existing methods usually only consider subtype prediction tasks for specific diseases, without taking into account several clinical variables that doctors and clinicians often pay attention to. In order to simulate the process of doctors and clinicians studying disease subtype prediction, the embodiment of the present invention introduces meta-learning technology to develop a new data-efficient model, which can extract shared experience or knowledge from a series of related tasks and quickly transfer it to new tasks. Therefore, the basic idea of ​​the prediction model of the embodiment is to learn from related clinical tasks through meta-learning technology to extract valuable information to help the model be well generalized to the prediction task of disease subtypes.

[0048] In the first aspect, Figure 1 As shown, the embodiment of the present invention provides a disease subtype prediction method based on a small sample, which mainly includes steps S01-S03:

[0049] S01. Obtain gene expression data of a first data set. In the embodiment, the first data set is a data set consisting of gene expression data to be predicted.

[0050] S02, predicting the gene expression data through the prediction model to obtain the disease subtype of the sample to be predicted; in this embodiment, the constructed model includes a sample selection network, a feature selection layer and a meta-learner. First of all, meta-learners have a prominent history in meta-learning in machine learning, and many meta-learning methods have been applied to solve small sample learning. Meta-learning methods can be roughly divided into two main types: first, a meta-learning method based on gradient optimization, in which the method learns a meta-learner in an outer loop to initialize the base learner of the inner loop, and then trains on a new and less task; second, a metric-based meta-learning method, in which the method needs to understand the metric space of sample features, in which effective classification can be performed with very few samples. Most of the current meta-learning methods rely only on image classification tasks. There is a lack of systematic research on using meta-learning to solve the problem of predicting small sample disease subtypes. In this embodiment, an extended meta-learner is selected, which is a prototype network. In addition, the model of the embodiment is attached with two modules that are robust to high-dimensional and high-noise gene expression data, which are a sample selection network and a feature selection layer.

[0051] In an embodiment, Figure 2 As shown, the training process of the prediction model includes steps S021-S024:

[0052] S021, construct a first sample data set; the sample data set includes the gene expression data of the second data set; wherein the constructed first sample data set is the training set of the model, and the second data set is the clinical data of the gene expression data, and its historical data all carry labels of corresponding disease subtypes, and the training set of the model is obtained by arranging the clinical data carrying labels. In step S021, in order to solve the problem of small sample learning, a data fusion strategy can usually be used to increase the sample size, that is, to improve the generalization ability of the model from the perspective of data and avoid overfitting, which can be completed by step S0211 or step S0212:

[0053] S0211. Acquire a number of gene expression data and obtain a first sample data set through meta-analysis; that is, independently analyze each data set through meta-analysis, and finally combine the statistical results to identify genes related to the disease.

[0054] S0212. Obtain a number of gene expression data, and integrate them to obtain a first sample data set through batch correction and machine learning. That is, using the integrative analysis via data merging method, the experimental data from different platforms are regarded as a single data set from the same experiment, the batch effect is eliminated through the batch correction method, and then the integrated data set is analyzed using the machine learning method.

[0055] In addition, in the step of constructing the first sample data set in step S021, it also includes steps S0213 and S0214:

[0056] S0213. Extracting a plurality of support samples and a plurality of query samples from the second data set according to the categories of the gene expression data;

[0057] S0214. Construct a support set based on the support samples; construct a query set based on the query samples.

[0058] Specifically, since FSL involves building a model using training data of known categories, the model can classify unknown new categories using a small number of samples. In the FSL setting, the embodiment obtains a large sample set Ds from a source class Cs of a set of labeled gene expression data, obtains a small sample set Dt from a target class Ct, and the test set T is derived from Ct, and The goal of small sample learning is to train a classification model with Cs so that it can generalize well to task T. Furthermore, the embodiment adopts an episodic training strategy commonly used in small sample learning models based on meta-learning. Specifically, a series of n-way k-shot tasks randomly sampled from Ds are defined, and each n-way k-shot task is defined as an episode D = (S; Q). n-way k-shot means that there are n classes and k samples in each class. S is called the support set containing n classes and k samples of each class, and Q is called the query set with the same n classes. In an embodiment, a new small sample data set is constructed for the prediction of disease subtypes in the field of biology. Episode D can be constructed by the following process: First, a small number of source classes C = {C i |i=1,2,3…,n c} is obtained from Cs. Then, k support samples and q query samples are randomly selected from each class in C to generate support set S and query set Q respectively. Therefore, we have support set S = {(x i ,y i )|y i ∈C, i=1,2,3…,m s} and query set Q = {(x i ,y i )|y i ∈C, i=1,2,3…,m q}, where m s =n×k,m q represents the number of query samples in each episodic, and

[0059] The meta-learner to be trained is then expressed as Its parameters are It can be understood as the probability that the output sample of the meta-learner belongs to class y given a data sample x. On the query set Q of different episodes, the classifier is trained between its predicted label and the true label so that it can be generalized to other data sets. The objective function is as follows:

[0060]

[0061] In formula (1), different forms Different types of meta-learning algorithms can be defined.

[0062] S022. Obtain the feature weight vector of the training data set through the feature selection layer; construct the second sample data set based on the feature weight vector; since the prototype network learns the prototype of each class in the support set S, classify each sample in the query set Q according to the distance between the query set Q and different prototypes. as follows:

[0063]

[0064] In formula (2), d represents the Euclidean distance in the feature space, C n Represents the prototype in each episode. Formula (2) shows that the prototype network outputs about and class prototype C n The distance between generates the class distribution of the query sample x, where is the embedding function that maps sample x to a feature vector Each prototype C n All of them are embedded by supporting samples S belonging to class n. n The vector is averaged to calculate, and the form is as follows:

[0065]

[0066] In formula (3), S n ∈S, represents the support sample set belonging to category n. In gene expression data, each sample x∈R p All have features with high feature dimension p. Usually, feature selection methods try to find a selection vector β = (β1, β2, ..., β p ), that is, the corresponding element is multiplied by the data x to filter out useless features and obtain the data x new The new representation of , which is more conducive to the subsequent steps, namely:

[0067] x new =β⊙x,β j ∈[0,1] (4)

[0068] Regularization technology is an effective way to solve the problem of dimensionality curse. Usually, a specific form of regularization is manually set under certain assumptions about the training data. However, when the basic knowledge of gene expression data is poor, corresponding improvements are needed. In order to overcome the problems caused by this situation, the embodiment selects vector modeling to obtain a softmax network layer, which can learn from the data to obtain an adaptive feature weight vector:

[0069]

[0070] In formula (5), θ∈Rp is the parameter vector of the softmax layer, and exp is the element-wise exponential operator. Formula (5) can be easily embedded into formula (2) to obtain:

[0071]

[0072] In formula (6), C n It can be further expressed as:

[0073]

[0074] Formula (6) allows θ to be expressed as It is also learned efficiently via automatic differentiation techniques and can be easily extended to other problems.

[0075] S023, input the second sample data set into the sample selection network to obtain sample weights; wherein the second sample data set is the training set or support set and query set obtained after processing in step S022. Due to the high noise in the gene expression data, it is easy to cause the generalization performance of the model to decline. In order to overcome this problem, the embodiment does not adopt the simple averaging strategy in the prototype network, but attempts to assign weights to the support samples to characterize the confidence of the samples, in order to suppress the influence of extreme noise samples. Instead, the prototype C obtained in the end n It is calculated by weighting all embedded support samples of class n. In some embodiments, the step of inputting the second sample data set into the sample selection network to obtain the sample weight is specifically: determining the confidence of the sample data in the second sample data set, and assigning the sample weight according to the confidence.

[0076] Specifically, prototype C n The weighted representation of all embedded support samples of class n is:

[0077]

[0078] In formula (8), v i Reflects the support sample x i Confidence of being a clean sample. In the embodiment, a larger weight v means that it is more likely to be a clean sample with high confidence.

[0079] In addition, the step of determining the confidence of the sample data in the second sample data set, in some embodiments, is specifically expanded as follows: determining the loss function of the sample data in the second sample data set, embedding the loss function into the sample data, fitting the weighted function through a neural network, and obtaining the confidence of the sample data.

[0080] Specifically, to determine v, inspired by the current adaptive sample weighting strategy Meta-Weight-Net (MW-Net), the embodiment learns a weighting function to distinguish clean and noisy samples. More specifically, MW-Net models the sample weight v as an MLP network V(ι; Θ) with only one hidden layer, which is a universal approximator of any continuous function, so it can fit a variety of weighting functions, whose input is the loss function of the sample and the output is the weight of the sample. However, since the prototype network does not calculate the loss of the support sample, the embodiment uses an embedded representation As the input of MW-Net. Therefore, formula (8) can be further expanded as follows:

[0081]

[0082] In formula (9), Θ represents the parameters of MW-NetV2 (different from the original MW-Net). The architecture of MW-NetV2 is the same as that of MW-Net, except that the input is the embedded representation of each sample instead of the loss in the network.

[0083] By adding a feature selection layer (FS-Net) and a sample selection network (MW-NetV2) to the prototype network, the method of the embodiment can selectively select disease-related genes that are helpful for classification and suppress the negative impact of noise data. The learning process is similar to the prototype network, and the parameters of FS-Net and MW-NetV2 are updated simultaneously with the parameters of the prototype network. The final loss function is:

[0084]

[0085]

[0086] In formula (11),

[0087] S024, training a meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model; step S024 further includes:

[0088] Step S0241, removing noise samples in the support set according to the sample weights; determining the sample categories in the support set after denoising;

[0089] Step S0242: obtain the Euclidean distance between the samples and the sample categories in the support set after denoising, normalize the Euclidean distance, and obtain the output of the meta-learner.

[0090] Specifically, after step S022 and step S023, the prototype network is trained. In the process of feature extraction, the original features are converted into a set of new features through subspace learning. However, it is difficult to obtain a reasonable biological explanation from the learned feature subspace. Feature selection removes irrelevant and redundant features and selects a set of important features that are closely related to the target. Then, a sample weight strategy is adopted to apply weights to the samples according to their training reliability, and finally a trained prediction model is obtained.

[0091] Necessarily, step S024 also includes step S0243: verifying the output of the meta-learner according to the query set, and outputting the trained prediction model according to the verification result.

[0092] In a second aspect, the technical solution of the present invention also provides a system for predicting disease subtypes based on small samples, including a data acquisition unit, a model building unit, and a prediction output unit, wherein:

[0093] A data acquisition unit, used to acquire gene expression data of a first data set;

[0094] The model building unit is used to train the prediction model, which includes: a sample selection network, a feature selection layer, and a meta-learner; the training steps of the prediction model include:

[0095] Constructing a first sample data set; the sample data set includes gene expression data of a second data set; obtaining a feature weight vector of a training data set through a feature selection layer; constructing a second sample data set according to the feature weight vector; inputting the second sample data set into a sample selection network to obtain sample weights; training a meta-learner according to support samples and sample weights to obtain a trained prediction model;

[0096] The prediction output unit is used to predict the gene expression data through the prediction model to obtain the disease subtype of the sample to be predicted.

[0097] In a third aspect, an embodiment of the present invention also provides a device for predicting disease subtypes based on small samples, which includes at least one processor; at least one memory for storing at least one program; when the at least one program is executed by at least one processor, the at least one processor implements the disease subtype prediction method based on small samples as in the first aspect.

[0098] An embodiment of the present invention further provides a storage medium storing a program, and the program is executed by a processor as in the method of the first aspect.

[0099] In addition to the above methods, systems, devices and other embodiments, the following are the results obtained by the specific implementation of the embodiments of the present invention; in the process, predictions were made through the embodiments and simulated data sets and real gene expression data sets, and the results of predictions made by other methods or means were compared, as shown in Table 1:

[0100] Table 1

[0101]

[0102]

[0103] In Table 1, the present embodiment (Select-ProtoNet) is compared with the ordinary metric-based meta-learning method ProtoNet and ProtoNet with one of the two additional modules. SelectF-ProtoNet and SelectS-ProtoNet have a feature selection layer and a sample selection network, respectively. In the process, four scenarios in which the sample class labels are damaged on the support set are set. The class label of each support sample is independently converted to a random class with probability P, where P = 0%; 10%; 30%; 50%. Table 1 shows the accuracy (%) ± variance of Select-ProtoNet and its baseline method on the simulated dataset after 30 random runs under different experimental settings. It can be seen that both additional modules contribute to the performance improvement, and the proposed method achieves the best results. As the noise rate increases, the advantage of Select-ProtoNet over ProtoNet becomes more obvious. As shown Figure 3 and Figure 4 As shown in the figure, the training loss curves and accuracy between ProtoNet and Select-ProtoNet are compared. It can be seen that ProtoNet takes nearly twice as long to achieve the best classification accuracy compared to Select-ProtoNet in this example. Figure 5 As shown, Figure 5 The weight distribution of clean and noisy training support samples is plotted in . It can be seen that almost all the larger weights belong to clean samples, and the weight values ​​of noisy samples are smaller than those of clean samples, which shows that MW-NetV2 can distinguish between clean and noisy samples.

[0104] In addition to the comparison with the metric-based meta-learning method ProtoNet, the process also compares with existing machine learning methods, including prototype network (ProtoNet), majority class prediction (Majority), logistic regression (LogisticRegression) and neural network (Neural Network). The real dataset used in the comparison process is TCGA Meta-Dataset, which is a public benchmark dataset in the field of gene expression analysis. It contains 174 clinical tasks derived from the Cancer Genome Atlas (TCGA) and can be used in a multi-task learning framework. However, it cannot be directly used as a benchmark dataset for small sample learning because some tasks have very unbalanced classes. Therefore, a new dataset mini-TCGA Meta-Dataset is used in the comparison process. This dataset consists of 68 TCGA benchmark clinical tasks, each of which has two categories and each category has at least 60 samples, as shown in Table 2:

[0105] Table 2

[0106]

[0107]

[0108] Table 2 shows the comparison results of the model test accuracy (%) using different noise rate settings on the miniTCGA Meta-Dataset. The average accuracy (± variance) of 30 repetitions is recorded. The best results are highlighted in bold as the classification accuracy under different noise rate settings in the miniTCGA Meta-Dataset. It can be seen that the prediction model of the embodiment improves on its baseline method, has a significant improvement for all noise rate settings, and outperforms the three traditional supervised methods.

[0109] As shown in Table 3, the lung cancer subtype prediction accuracy of all compared methods on the TCGA Meta-Dataset with task id Expression_Subtype,LUNG is reflected. In particular, ProtoNet and Select-ProtoNet consider the interrelationship of clinical tasks, use all clinical task samples of miniTCGA Meta-Dataset to build shared experience or knowledge, and transfer it to help predict lung cancer subtypes, while traditional methods only consider lung cancer subtype tasks.

[0110] Table 3

[0111]

[0112] As shown in Table 3, we can see that the prediction model of the embodiment achieved the best results, with an accuracy improvement of more than 20% compared to the baseline method and an accuracy improvement of more than 29% compared to the best result of the traditional method. This means that our model significantly improves the prediction performance of disease subtypes by incorporating various relevant clinical tasks.

[0113] From the above specific implementation process, it can be concluded that the technical solution provided by the present invention has the following advantages or strengths compared with the prior art:

[0114] 1. The embodiment provided by the present invention, by additionally adding two modules, a feature selection layer and a sample selection network, to the prototype network model, the model can adaptively select important features and clean samples, and can learn and generalize well from small sample data.

[0115] 2. In the embodiment provided by the present invention, the two additional modules and the prototype network are updated in a unified framework, which can be more easily implemented on the basis of the original prototype network.

[0116] 3. The superiority of the method of the present invention in predicting disease subtypes and identifying potential disease-related genes was demonstrated through verification with simulated and real gene expression data.

[0117] In some selectable embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided by way of example, for the purpose of providing a more comprehensive understanding of technology. The disclosed method is not limited to the operation and logic flow presented herein. Selectable embodiments are expected, wherein the order of various operations is changed and the sub-operation of a part for which is described as a larger operation is performed independently.

[0118] In addition, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise specified, one or more of the functions and / or features can be integrated into a single physical device and / or software module, or one or more functions and / or features can be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the present invention. More specifically, in view of the properties, functions and internal relationships of the various functional modules in the device disclosed herein, the actual implementation of the module will be understood within the conventional skills of the engineer. Therefore, those skilled in the art can implement the present invention set forth in the claims without excessive experimentation using ordinary techniques. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0119] Among them, if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0120] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0121] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0122] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0123] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0124] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

[0125] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A disease subtype prediction method based on a small sample, characterized in that: The following steps are involved: Obtaining gene expression data of a first data set; Predicting the gene expression data by using a prediction model to obtain the disease subtype of the first data set; the prediction model includes a sample selection network, a feature selection layer and a meta-learner; The prediction model is trained by the following steps: Constructing a first sample data set; the first sample data set includes gene expression data of a second data set; wherein the second data set is clinical data of the gene expression data, and the clinical data carries a label of a disease subtype; Obtaining a feature weighted vector of the training data set through the feature selection layer; constructing a second sample data set according to the feature weighted vector; Inputting the second sample data set into the sample selection network to obtain sample weights; Training the meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model; The step of constructing the first sample data set specifically includes: Acquire a number of the gene expression data, and obtain the first sample data set through meta-analysis; or obtaining a plurality of the gene expression data, and integrating them to obtain the first sample data set through batch correction and machine learning; The step of constructing the first sample data set includes: Extracting a plurality of support samples and a plurality of query samples from the second data set according to the categories of the gene expression data; Construct a support set according to the support samples; construct a query set according to the query samples; The step of inputting the second sample data set into the sample selection network to obtain sample weights is specifically as follows: Determining the confidence of the sample data in the second sample data set, and assigning the sample weight according to the confidence; The step of determining the confidence of the sample data in the second sample data set comprises: Determine a loss function of sample data in the second sample data set, embed the loss function into the sample data, and obtain the confidence of the sample data by fitting a weighting function through a neural network; The step of training the meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model, wherein the step of training the meta-learner includes: Remove the noise samples in the support set according to the sample weights; determine the sample categories in the support set after denoising; Obtaining the Euclidean distance between the sample in the support set after denoising and the sample category, normalizing the Euclidean distance, and obtaining the output of the meta-learner; The step of training the meta-learner according to the second sample data set and the sample weight to obtain a trained prediction model includes: According to the query set, the output of the meta-learner is verified, and the trained prediction model is output according to the verification result.

2. A system for predicting disease subtypes based on small samples, characterized in that: It includes a data acquisition unit, a model building unit, and a prediction output unit, wherein: The data acquisition unit is used to acquire gene expression data of the first data set; The model building unit is used to train a prediction model, wherein the prediction model includes: a sample selection network, a feature selection layer, and a meta-learner; the training steps of the prediction model include: Constructing a first sample data set; the first sample data set includes gene expression data of a second data set; wherein the second data set is clinical data of gene expression data, and the clinical data carries a label of a disease subtype; obtaining a feature weight vector of a training data set through the feature selection layer; constructing a second sample data set according to the feature weight vector; inputting the second sample data set into the sample selection network to obtain a sample weight; training the meta-learner according to the second sample data set and the sample weight to obtain a trained prediction model; The prediction output unit is used to predict the gene expression data through a prediction model to obtain the disease subtype of the first data set sample; The training steps of the prediction model specifically include: The step of constructing the first sample data set specifically includes: Acquire a number of the gene expression data, and obtain the first sample data set through meta-analysis; or obtaining a plurality of the gene expression data, and integrating them to obtain the first sample data set through batch correction and machine learning; The step of constructing the first sample data set includes: Extracting a plurality of support samples and a plurality of query samples from the second data set according to the categories of the gene expression data; Construct a support set according to the support samples; construct a query set according to the query samples; The step of inputting the second sample data set into the sample selection network to obtain sample weights is specifically as follows: Determining the confidence of the sample data in the second sample data set, and assigning the sample weight according to the confidence; The step of determining the confidence of the sample data in the second sample data set comprises: determining a loss function of sample data in the second sample data set, embedding the loss function into the sample data, Fitting a weighted function through a neural network to obtain the confidence level of the sample data; The step of training the meta-learner according to the second sample data set and the sample weights to obtain a trained prediction model, wherein the step of training the meta-learner includes: Remove the noise samples in the support set according to the sample weights; determine the sample categories in the support set after denoising; Obtaining the Euclidean distance between the sample in the support set after denoising and the sample category, normalizing the Euclidean distance, and obtaining the output of the meta-learner; The step of training the meta-learner according to the second sample data set and the sample weight to obtain a trained prediction model includes: According to the query set, the output of the meta-learner is verified, and the trained prediction model is output according to the verification result.

3. A device for predicting disease subtypes based on small samples, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the disease subtype prediction method based on a small sample as described in claim 1.

4. A storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to implement the disease subtype prediction method based on small samples as claimed in claim 1 when executed by the processor.

Images