DNA site identification method and device based on multi-model dynamic ensemble

Abstract

This application relates to the field of DNA site identification technology, and particularly to a DNA site identification method and apparatus based on multi-model dynamic ensemble. The method includes: acquiring a DNA sequence; using the DNA sequence as input, running a first basis model to obtain a class probability vector P. j The first base model is a machine learning model based on the deep forest algorithm, where j is the cascade layer number of the first base model; the class probability vector P j To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction result Q. p The second base model is a machine learning model based on support vector machines. Compared with traditional support vector machines for binary classification, the model provided in this application embodiment can provide more accurate confidence information based on the binary classification results. Compared with traditional machine learning schemes, the scheme provided in this application embodiment adds a certain amount of performance adjustment space.

Description

Technical Field

[0001] This application belongs to the field of DNA site identification technology, and in particular relates to a DNA site identification method and apparatus based on multi-model dynamic integration. Background Technology

[0002] With the widespread application of machine learning in bioinformatics, mainstream technologies for tasks such as gene sequence analysis and protein function prediction have also been updated. However, existing methods still have significant shortcomings in methylation data processing, model synergy, and noise robustness. For example, traditional DNA sequence encoding often uses one-hot encoding or k-mer frequency statistics, which fails to effectively preserve the local structural information of CpG islands in methylated sequences; while ensemble schemes based on neural networks can capture nonlinear relationships, they suffer from significant overfitting issues with small-sample methylation data and have poor model interpretability.

[0003] Therefore, there is an urgent need to provide a DNA site identification method with better performance. Summary of the Invention

[0004] This application provides a DNA site identification method and apparatus based on multi-model dynamic integration, which can solve the problem of poor identification accuracy (accuracy, precision, recall) of existing methods.

[0005] In a first aspect, embodiments of this application provide a DNA site identification method based on multi-model dynamic integration, comprising:

[0006] Obtain the DNA sequence;

[0007] Using the DNA sequence as input, the first basis model is run to obtain the class probability vector P. j The first base model is a machine learning model based on the deep forest algorithm, and j is the cascade layer number of the first base model.

[0008] Using the class probability vector P j To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction result Q. p The second base model is a machine learning model based on support vector machines.

[0009] The first base model and the second base model constitute at least a part of the ensemble model, which is a machine learning model trained based on samples and labels; the dynamic update refers to the update of the input feature space of the second base model as the cascaded output of the j-th layer of the first base model is updated.

[0010] In one possible implementation of the first aspect, the ensemble model further includes a meta-model; the meta-model is a machine learning model obtained through cross-validation; the cross-validation refers to:

[0011] Using at least two mutually exclusive subsets of the enhanced meta-training set as input, run candidate models, quantize the output of the candidate models based on the weights corresponding to the subsets, and determine the candidate model with the optimal quantization value as the meta-model;

[0012] The enhanced meta-training set is a training set obtained by splicing the features of the DNA sequence to the output of the first base model and / or the second base model, and injecting adaptive Gaussian noise; the candidate models include machine learning models based on the deep forest algorithm and machine learning models based on support vector machines.

[0013] In one possible implementation of the first aspect, the adaptive Gaussian noise includes directional Gaussian noise for the methylation boundary.

[0014] In one possible implementation of the first aspect, the prediction result Q is obtained. p Following the steps, it also includes:

[0015] splicing the prediction result Q p Based on the characteristics of the DNA sequence, the input feature Q is obtained. * ;

[0016] Based on the input feature Q * Running the meta-model yields the first prediction F1;

[0017] The first prediction F1 is corrected based on the K-nearest neighbor algorithm to obtain the second prediction F2;

[0018] By combining the first prediction F1 and the second prediction F2, a third prediction F3 is obtained.

[0019] In one possible implementation of the first aspect, the DNA sequence is characterized by sequence weights; the sequence weights are determined based on information gain ratio, and the sequence weight values ​​of CpG island regions in the DNA sequence are higher than those of non-CpG island regions.

[0020] In one possible implementation of the first aspect, the third prediction F3 satisfies:

[0021] F3=γG1(F2)+(1-γ)·G2(F1,F2)

[0022] Where G1 and G2 are both preset functions; γ is a preset function related to Acc(Q) * The fusion weights γ are given by γ, and the fusion weights γ satisfy:

[0023] If the meta-model is a machine learning model based on the deep forest algorithm, then the fusion weight γ is also related to the total number L of the cascaded layers of the meta-model, and Acc(Q) * The weighted accuracy is correlated with the meta-model.

[0024] If the meta-model is a machine learning model based on support vector machines, then Acc(Q) * The weighted interval is associated with the meta-model; wherein the interval is positively correlated with the average distance from the support vector to the hyperplane, and the interval is negatively correlated with the maximum distance from the support vector to the hyperplane.

[0025] The beneficial effects of the embodiments in this application compared with the prior art are:

[0026] 1. Compared to traditional support vector machines used for binary classification, the model provided in this application embodiment can provide more accurate confidence information based on the binary classification results, specifically:

[0027] Through the coupling and synergy of the first and second base models, the cascaded layers of the deep forest-like model (first base model) output class probability vectors (P). j This can be introduced as additional probabilistic information into the support vector machine-like model (second basis model); based on this, the symbolic distance from the sample output by the support vector machine-like model to the hyperplane is used to represent the confidence level, and combined with the introduced probabilistic information, a more reliable confidence level information is output.

[0028] 2. Compared to traditional machine learning solutions, the solution provided in this application provides additional performance adjustment flexibility, specifically:

[0029] For machine learning models, although the present application has optimized their interpretability through output parameters such as confidence, in practice—especially during training—there are still many performance optimization bottlenecks that cannot be explored for various reasons. In this case, it is essential to adjust the model parameters (e.g., adjust weight initialization, introduce gradient clipping, add BN layers, etc.). However, the bottlenecks that still exist in traversing known methods may also exist. In this regard, the embodiments of the present application provide new optimization possibilities that are different from the prior art based on the newly introduced dynamic update feature.

[0030] This application utilizes the cascading expansion characteristic of deep forest-like models to provide multiple outputs with probabilistic information (i.e., class probability vectors P). j Theoretically, for every possible value of j, P j All information can be introduced into a support vector machine-like model (thus achieving dynamic updates), while for a deep forest-like model structure, different specific model architectures result in different output P values ​​at different levels.j The parameters such as information density, effectiveness, and confidence vary, and questions such as which of these differences are introduced into support vector machine-like models, when the cascading expansion of deep forest-like models stops, and whether / how the results of two base models are fused will provide new possibilities for optimizing model performance bottlenecks.

[0031] 3. Compared to traditional DNA site identification schemes, the scheme provided in this application has a more targeted performance optimization approach, specifically:

[0032] A three-dimensional feature matrix is ​​generated by binary encoding, and higher weights are assigned to CpG island regions. Two base models, deep forest (cascaded structure adaptive termination) and SVM (hybrid kernel function dynamic optimization), are integrated. The model prediction results are concatenated with the original features and directional Gaussian noise targeting the methylation boundary is injected. Then, a meta-model is constructed through triple cross-validation weighted evaluation. Finally, a three-stage prediction fusion strategy (including residual correction, nearest neighbor median adjustment and weighted mode fusion) is adopted for the test set, which significantly improves the classification accuracy and generalization ability of DNA methylation site identification.

[0033] Secondly, embodiments of this application provide a DNA site identification device based on multi-model dynamic integration, comprising:

[0034] Sequence acquisition module, used to acquire DNA sequences;

[0035] The first execution module is used to run the first base model with the DNA sequence as input to obtain the class probability vector P. j The first base model is a machine learning model based on the deep forest algorithm, and j is the cascade layer number of the first base model.

[0036] The second execution module is used to execute the class probability vector P. j To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction result Q. p The second base model is a machine learning model based on support vector machines.

[0037] The first base model and the second base model constitute at least a part of the ensemble model, which is a machine learning model trained based on samples and labels; the dynamic update refers to the update of the input feature space of the second base model as the cascaded output of the j-th layer of the first base model is updated.

[0038] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the DNA site identification method based on multi-model dynamic integration as described in any of the first aspects above.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the DNA site identification method based on multi-model dynamic integration as described in any of the first aspects above.

[0040] Fifthly, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute the DNA site identification method based on multi-model dynamic integration as described in any of the first aspects.

[0041] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the DNA site identification method based on multi-model dynamic integration provided in the embodiments of this application.

[0044] Figure 2 This is a schematic diagram of the DNA site recognition device based on multi-model dynamic integration provided in the embodiments of this application.

[0045] Figure 3 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application.

[0046] Figure 4 This is a schematic diagram providing an overview of the DNA methylation site identification method based on multi-model dynamic integration and noise enhancement provided in the embodiments of this application.

[0047] Figure 5 This is a schematic diagram of the binary encoding of DNA data.

[0048] Figure 6 This is a schematic diagram of the DNA methylation site process.

[0049] Figure 7 This is a schematic diagram of the multi-model container construction and training process.

[0050] Figure label:

[0051] Terminal equipment 30;

[0052] Processor 301;

[0053] Memory 302;

[0054] Computer program 303. Detailed Implementation

[0055] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0056] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0057] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0058] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0059] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0060] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0061] like Figure 1 As shown in the embodiments of this application, a DNA site identification method based on multi-model dynamic integration is provided, including:

[0062] Step 102, Obtain the DNA sequence;

[0063] Step 104: Using the DNA sequence as input, run the first basis model to obtain the class probability vector P. j The first base model is a machine learning model based on the deep forest algorithm, and j is the cascade layer number of the first base model.

[0064] Step 106, using the class probability vector P j To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction result Q. p The second base model is a machine learning model based on support vector machines.

[0065] The first base model and the second base model constitute at least a part of the ensemble model, which is a machine learning model trained based on samples and labels; the dynamic update refers to the update of the input feature space of the second base model as the cascaded output of the j-th layer of the first base model is updated.

[0066] In various embodiments of this application, the model based on support vector machine, the support vector machine-like model, SVM, SupportVector Machine, or related concepts are limited to a model that performs binary classification based on a hyperplane in the feature space. In some preferred embodiments, it is further limited to a model in which the distance from the feature point in the feature space to the hyperplane can reflect the confidence level.

[0067] Compared to traditional support vector machines used for binary classification, the model provided in this application embodiment can provide more accurate confidence information based on the binary classification results, specifically:

[0068] Through the coupling and synergy of the first and second base models, the cascaded layers of the deep forest-like model (first base model) output class probability vectors (P). j This can be introduced as additional probabilistic information into the support vector machine-like model (second basis model); based on this, the symbolic distance from the sample output by the support vector machine-like model to the hyperplane is used to represent the confidence level, and combined with the introduced probabilistic information, a more reliable confidence level information is output.

[0069] Similarly, in the embodiments of this application, the machine learning model based on the Deep Forest algorithm, the deep forest-like model, or related concepts are limited to machine learning models based on decision tree ensembles. These can include generalized deep forest algorithm models, completely randomized deep forest algorithm models, and other variant models based on decision tree ensembles. The core limitation of the embodiments of this application lies in their cascading characteristics, enabling them to output intermediate results P with certain information (e.g., confidence levels) in a hierarchical manner. j .

[0070] Compared to traditional machine learning solutions, the solution provided in this application provides additional performance tuning flexibility, specifically:

[0071] For machine learning models, although the present application has optimized their interpretability through output parameters such as confidence, in practice—especially during training—there are still many performance optimization bottlenecks that cannot be explored for various reasons. In this case, it is essential to adjust the model parameters (e.g., adjust weight initialization, introduce gradient clipping, add BN layers, etc.). However, the bottlenecks that still exist in traversing known methods may also exist. In this regard, the embodiments of the present application provide new optimization possibilities that are different from the prior art based on the newly introduced dynamic update feature.

[0072] This application utilizes the cascading expansion characteristic of deep forest-like models to provide multiple outputs with probabilistic information (i.e., class probability vectors P). j Theoretically, for every possible value of j, P j All information can be introduced into a support vector machine-like model (thus achieving dynamic updates), while for a deep forest-like model structure, different specific model architectures result in different output P values ​​at different levels. j The parameters such as information density, effectiveness, and confidence vary, and questions such as which of these differences are introduced into support vector machine-like models, when the cascading expansion of deep forest-like models stops, and whether / how the results of two base models are fused will provide new possibilities for optimizing model performance bottlenecks.

[0073] The following will provide a more detailed description of this embodiment.

[0074] In a preferred embodiment, the DNA sequence data in step 102 uses encoded genomic methylation data. In different embodiments, the DNA sequence used as model input data can use different encoding formats, for example, binary encoding: A→00, T→01, C→10, G→11.

[0075] The first and second base models involved in steps 104 and 106 together constitute at least a part of the ensemble model, thus limiting this embodiment to implementation relying on a multi-model prediction container. The foregoing limitations have clarified the actual referents of the deep forest-like model and the support vector machine-like model. Based on this, a preferred implementation of the specific architecture of the base models is provided below. This embodiment does not constitute an additional limitation on the scope of protection of this application.

[0076] In a preferred embodiment, the base model consists of only two elements: a deep forest and a support vector machine (SVM), and:

[0077] Deep forests employ a cascaded structure, with each cascaded layer containing:

[0078] Completely Random Forest: When splitting a tree node, a feature f is randomly selected. j Where j is the number of tree levels, and the splitting threshold τ ranges from (min f) j ,max f j ).

[0079] Random Forest: When splitting a node, the feature f with the largest decrease in Gini coefficient is selected from c candidate features. The calculation formula is:

[0080]

[0081] SVM uses a hybrid kernel function formula: K(x, y) = θ × K linear +(1-θ)×K rbf , where K linear and K rbf These represent the linear kernel and the radial basis function kernel, respectively. The value of θ ranges from [0,1] and is optimized through grid search.

[0082] Furthermore, in this embodiment, deep forest and SVM are tightly coupled and collaborated, and the class probability vector P output by the j-th layer of deep forest is concatenated. j As an additional feature dimension of SVM, the input feature space of SVM is dynamically updated as X'=[X|P j ], where P j It is updated as the number of cascaded layers j increases.

[0083] The above embodiments are basic embodiments of this application. The beneficial effects mentioned can be achieved by relying solely on the above basic embodiments. Based on this, this application provides more preferred embodiments as follows.

[0084] The following is an example of a preferred scheme for an integrated model (or multi-model container), in which:

[0085] The ensemble model further includes a meta-model; the meta-model is a machine learning model obtained through cross-validation; the cross-validation refers to:

[0086] Using at least two mutually exclusive subsets of the enhanced meta-training set as input, run candidate models, quantize the output of the candidate models based on the weights corresponding to the subsets, and determine the candidate model with the optimal quantization value as the meta-model;

[0087] The enhanced meta-training set is a training set obtained by splicing the features of the DNA sequence to the output of the first base model and / or the second base model, and injecting adaptive Gaussian noise; the candidate models include machine learning models based on the deep forest algorithm and machine learning models based on support vector machines.

[0088] In this embodiment, the base model and the meta-model together constitute the ensemble model. In an optional implementation, the overall architecture of the ensemble model can be based on a Stacking meta-learner architecture, but the model collaboration aspect differs. Specifically, in this embodiment, the overall architecture is still a Stacking-like structure based on the base model outputting intermediate results and the meta-model outputting the final results, but the approach to model collaboration is not entirely the same as the stacking, voting, and averaging methods of traditional Stacking meta-learners.

[0089] The differences include at least the following:

[0090] There is interaction and collaboration among the base models, and the determination of the meta-model pattern depends on the performance of the base models. That is, the specific architecture of the meta-model is consistent with the best-performing base models. Here, "best-performing" means that at least one inference metric meets a preset condition. The condition can be that the F1-score is greater than a preset value (or the F1-score is the best among all base models), the interpretability is optimal during model training / fine-tuning, etc. Therefore, unlike the model inference prediction process described in the main solutions of the various embodiments of this application, the step of determining the meta-model pattern is completed during the training process. Thus, the subsequent implementation of some embodiments is based on the model training steps and is carried out before the execution of the solutions of the various embodiments of this application.

[0091] In the traditional stacking meta-learner architecture, the output of the base model serves as the input of the meta-model, which then outputs the final result. Data transmission here is linear or linearly superimposed, and the cooperation between models is limited to the exchange of input and output data. However, in the ensemble model architecture provided in this embodiment, at least two base models have some data interaction before outputting the result (e.g., P...). j Furthermore, depending on the specific architecture of the metamodel, data interaction (non-input-output interaction) between the base models and the metamodel is not prohibited. For example, when the metamodel is a support vector machine, the P of the deep forest base model... j It can be dynamically updated in the support vector machine base model and support vector machine meta-model.

[0092] The following is an example of a specific optimization scheme for DNA site recognition tasks. In this example:

[0093] The adaptive Gaussian noise includes directional Gaussian noise for methylation boundaries.

[0094] The DNA sequence is characterized by sequence weights; the sequence weights are determined based on information gain ratio, and the sequence weights of CpG island regions in the DNA sequence are higher than those of non-CpG island regions.

[0095] It is worth noting that the optimizations provided in this embodiment are based on DNA methylation data. Due to the biological characteristics of DNA methylation data (such as CpG island features, class imbalance, and boundary ambiguity), these optimizations can significantly improve model performance.

[0096] The biological characteristics of DNA methylation data will at least bring the following problems to the task of model site identification:

[0097] Traditional noise injection (such as uniform Gaussian noise) blindly covers features, easily drowning out weak signals at methylation boundaries (such as CpG island edges), leading to category ambiguity and increased misclassification rate.

[0098] Traditional DNA coding (such as k-mer frequency) ignores the local structure of CpG islands, resulting in insufficient importance of methylation-sensitive features (such as "C" / "G" sites).

[0099] More specifically, noise injection techniques, as an important means of improving model robustness, often employ uniformly distributed or fixed-intensity Gaussian noise in existing schemes, lacking quantitative perception of the confidence level of methylation sites. Studies have shown that blindly injecting noise may mask effective biosignature signals; for example, weakly correlated features of methylation sites (such as CpG island edges) are easily submerged by noise.

[0100] In terms of model ensemble strategies, the commonly used static weight allocation or simple voting mechanisms are difficult to adapt to the dynamic changes in the distribution of methylation data (such as CpG island boundary mutations). Common multi-model weighted fusion methods fix the base model weights only through validation set accuracy, without considering local correlations in the methylation feature space, leading to a sharp drop in generalization performance under sequence mutations or noise interference scenarios. In addition, traditional models such as Support Vector Machines (SVM) in multi-model frameworks often limit their boundary adaptability due to the single kernel function selection, especially when dealing with high-dimensional sparse methylation data, where the individual use of linear kernels and RBF kernels may lead to marginal effect decay.

[0101] To overcome these problems, the optimization approach of directional Gaussian noise targeting methylation boundaries and the optimization approach of weight enhancement for CpG island regions (sequence weights dynamically calculated based on information gain rate) are introduced. Based on this, this embodiment provides a decision-making method that integrates dynamic weight adjustment, adaptive noise enhancement, and heterogeneous model collaboration to overcome the limitations of existing technologies in methylation sequence analysis. Combined with other embodiments, this embodiment significantly improves the robustness and accuracy of multi-model integration in methylation data by integrating DNA binary encoding, hybrid kernel function optimization, and multi-stage validation strategies, providing technical support for applications such as gene diagnostics and drug target prediction.

[0102] Specifically, in a preferred embodiment of adaptive Gaussian noise, the noise intensity coefficient α can be set as follows:

[0103]

[0104] In the formula, Acc represents the accuracy of the prediction result, and the value of β ranges from [0.5, 1.5].

[0105] Simultaneously, N(μ,α) was added to the positive class samples (methylation sites). 2 Noise (μ = 0.065), negative samples are padded with N(-μ, α). 2 ), targeting CpG island boundary bias (e.g., positive bias to enhance weak signals for positive classes, and negative bias to suppress false positives for negative classes).

[0106] Based on the above work, the problems of class imbalance and boundary ambiguity have been overcome (or, in other words, the imbalance and boundary ambiguity have been mitigated), thus possessing the following advantages:

[0107] Improve boundary clarity (enhance noise robustness): Noise direction compensates for the sparsity of methylation data (e.g., few positive class samples), reducing low-confidence misclassifications (see subsequent comparison examples; Table 3 shows that the misclassification rate drops to 15.2% when α = 0.65). Adaptive intensity prevents noise from masking effective features.

[0108] A preferred implementation of sequence weighting:

[0109] The genomic methylation dataset was used as the DNA sample. Normalization preprocessing was performed on the input dataset D, with DNA sequences encoded in binary: A→00, T→01, C→10, G→11. CpG island regions (continuous "CG" sequences) were assigned a weight of 1.2 to enhance methylation feature representation. The weights were calculated based on the information gain ratio, using the following formula:

[0110]

[0111] The implementation approach here is:

[0112] Because CpG islands have higher information gain, the model focuses more on important regions, thereby improving accuracy.

[0113] In the weighted binary encoding, the weight of CpG island regions is multiplied by 1.2 (non-island regions by 1.0), highlighting the methylation-sensitive characteristics.

[0114] The following advantages are achieved through this implementation method:

[0115] Overall, the feature representation is improved (enhancing generalization ability), and high-weight regions guide the model to focus on CpG islands (e.g., deep forests prioritize high InfoGain features during splitting), thereby improving classification accuracy (see subsequent comparison examples; Table 2 shows that the AUC-ROC of deep forests reaches 0.915). On the other hand, the weights are dynamically calculated to adapt to different sequence distributions.

[0116] Meanwhile, the noise direction and intensity are optimized for the CpG island boundary.

[0117] The following is an optional embodiment of multi-model fusion, in which:

[0118] The predicted result Q is obtained. p Following the steps, it also includes:

[0119] splicing the prediction result Q p Based on the characteristics of the DNA sequence, the input feature Q is obtained. * ;

[0120] Based on the input feature Q * Running the meta-model yields the first prediction F1;

[0121] The first prediction F1 is corrected based on the K-nearest neighbor algorithm to obtain the second prediction F2;

[0122] By combining the first prediction F1 and the second prediction F2, a third prediction F3 is obtained.

[0123] The third prediction F3 satisfies:

[0124] F3=γG1(F2)+(1-γ)·G2(F1,F2)

[0125] Where G1 and G2 are both preset functions; γ is a preset function related to Acc(Q) * The fusion weights γ are given by γ, and the fusion weights γ satisfy:

[0126] If the meta-model is a machine learning model based on the deep forest algorithm, then the fusion weight γ is also related to the total number L of the cascaded layers of the meta-model, and Acc(Q) * The weighted accuracy is correlated with the meta-model.

[0127] If the meta-model is a machine learning model based on support vector machines, then Acc(Q) * The weighted interval is associated with the meta-model; wherein the interval is positively correlated with the average distance from the support vector to the hyperplane, and the interval is negatively correlated with the maximum distance from the support vector to the hyperplane.

[0128] This embodiment essentially provides a three-stage prediction. In practice, the prediction process can be specifically as follows:

[0129] In the first stage, the original prediction F1 = BestModel is output, where BestModel is the base model with the highest Acc(Q*).

[0130] In the second stage, the formula for calculating the residual correction is F2=F1+ΔF, where The formula for calculating the residual is given. In the KNN algorithm, K is set to 5 and the spatial clustering of methylation sites is considered by filtering nearest neighbors through Euclidean distance in the feature space.

[0131] In the third stage, the calculation formula for the result fusion is F3=γF2+(1-γ)·Mode({F1,F2}), where γ=0.6+0.1·Acc(Q*). When the optimal meta-model is a deep forest, it is adjusted according to the number of cascaded layers L. The adjustment formula is: γ=0.6+0.1×(Acc(Q*)-0.1×min(j,4)). If the optimal meta-model is an SVM, the fusion coefficient γ is increased by a compensation weight of 0.1 to compensate for the ambiguity of the methylation boundary, that is, γ'=min(γ+0.1,0.9). The mode is taken for classification tasks and the weighted average is taken for regression tasks.

[0132] Compared to traditional DNA site identification schemes, the preferred schemes provided in this application have a more targeted performance optimization approach, specifically:

[0133] A three-dimensional feature matrix is ​​generated by binary encoding, and higher weights are assigned to CpG island regions. Two base models, deep forest (cascaded structure adaptive termination) and SVM (hybrid kernel function dynamic optimization), are integrated. The model prediction results are concatenated with the original features and directional Gaussian noise targeting the methylation boundary is injected. Then, a meta-model is constructed through triple cross-validation weighted evaluation. Finally, a three-stage prediction fusion strategy (including residual correction, nearest neighbor median adjustment and weighted mode fusion) is adopted for the test set, which significantly improves the classification accuracy and generalization ability of DNA methylation site identification.

[0134] In the above preferred solutions:

[0135] G1(F2)=F2

[0136] G2(F1,F2)=Mode({F1,F2})

[0137] In other words, G2 based on the mode and G1 that directly outputs F2 itself are the better solutions in practice. Some suboptimal solutions can still be implemented here, for example:

[0138] 1. Weighted average fusion

[0139] G1(F2) = weighted average(F2), G2(F1,F2) = weighted average(F1,F2), with weights based on confidence levels (e.g., Acc(Q*)).

[0140] Advantages: Smooths out predictive fluctuations, suitable for regression tasks (the documentation mentions using a weighted average in regression). Example: F3 = γ·mean(F2) + (1-γ)·mean(F1,F2).

[0141] 2. Maximum confidence fusion

[0142] G1(F2) = F2 (direct output), G2(F1,F2) = max(F1,F2) (take the high confidence result).

[0143] Advantages: Prioritizes high-confidence predictions under low-quality data, reducing misclassifications (e.g., CpG island boundary samples). Example: F3 = γ·F2 + (1-γ)·max(F1,F2).

[0144] 3. Model-based dynamic fusion

[0145] G1 and G2 are implemented using lightweight models (such as logistic regression), with inputs (F1, F2) and outputs a fused value, and are optimized using cross-validation during training.

[0146] Advantages: Adaptive data distribution enhances the flexibility of data fusion.

[0147] The training phase of the ensemble model will be described in detail below with an example.

[0148] The following embodiments still provide a decision-making method based on multi-model dynamic integration and noise enhancement, which relates to the fields of machine learning and bioinformatics technology, specifically for DNA methylation site identification and optimized for its CpG island features.

[0149] This method encodes DNA sequences in binary form and optimizes the encoding weights for CpG island features of methylation sites (increasing the weight of CpG island regions by 20%) to construct a multi-model prediction container that includes deep forest and support vector machine (SVM). A dynamic weight correction strategy is employed to optimize the cascaded training process of the deep forest, and a hybrid kernel function (linear kernel and RBF kernel) is designed to improve the adaptability of SVM to fuzzy methylation boundary patterns. During the base model training phase, deep forest and SVM achieve tight coupling and synergy by constructing a joint optimization objective, with the objective function incorporating the importance weights of methylation features (based on CpG island density). An adaptive Gaussian noise injection mechanism enhances the robustness of the meta-training set; the noise intensity coefficient is negatively correlated with the prediction confidence of the base model for methylation sites, and adjustments are made for class imbalance in methylation data. Triple cross-validation selects the optimal meta-model, and three-stage prediction (original prediction, residual correction, and dynamic fusion) is performed, where residual correction optimizes the spatial correlation of methylation sites (such as CpG island clusters). This method improves accuracy in gene sequence methylation classification tasks and solves the problems of noise sensitivity, feature redundancy and insufficient dynamic adjustment in traditional multi-model integration of methylation data. It is suitable for biomedical methylation data analysis and high-dimensional heterogeneous data processing scenarios.

[0150] The methods specifically include:

[0151] 1. A genomic methylation dataset was selected as the DNA sample. Normalization preprocessing was performed on the input dataset D. The DNA sequences were encoded using binary encoding: A→00, T→01, C→10, G→11. CpG island regions (continuous "CG" sequences) were assigned a weight of 1.2 to enhance the representation of methylation features. The weight calculation was based on the information gain ratio, and the formula was as follows: Generate a 3D feature matrix

[0152] 2. Construct a multi-model prediction container, initialize the storage space Predictions = {}, and include two base models: Deep Forest and Support Vector Machine (SVM);

[0153] 3. For the base model, deep forests employ a cascaded structure, with each cascaded layer containing: Completely Random Forest: Each tree node randomly selects a feature f when splitting. j Where j is the number of tree levels, and the splitting threshold τ ranges from (min f) j ,max fj The value range is based on the methylation feature distribution adjustment, prioritizing high information gain points until the purity of leaf node y satisfies Gini(y) < 0.1. Random forest: when splitting a node, selects the feature f with the largest decrease in Gini coefficient from c candidate features, calculated using the following formula. The feature candidate set preferentially includes CpG island-related features (such as "C" and "G" encoded bits). The SVM uses a hybrid kernel function formula K(x,y)=θ×K linear +(1-θ)×K rbf The K mentioned linear and K rbf These represent the linear kernel and the radial basis function kernel, respectively. The value of θ ranges from [0,1] and is optimized through grid search. The value of the regularization parameter M ranges from [0.1,10] and is dynamically adjusted based on the class imbalance of the methylation data (the C value increases in the dense region of positive class samples).

[0154] 4. During the base model training phase, Deep Forest and SVM are tightly coupled and work together, concatenating the class probability vector P output by the j-th layer of Deep Forest. j As an additional feature dimension of SVM, the input feature space of SVM is dynamically updated as X'=[X|P j ], where P j Updated as the number of cascaded layers j increases, while setting a joint optimization objective:

[0155]

[0156] Where F meth I represents the methylation-sensitive feature set. f Let λ be the methylation importance weight of feature f, and λ = 0.3 + 0.1·j. λ increases with the importance of methylation features (e.g., the weight of CpG island features in layer j increases). This prompts the node splitting of the deep forest to consider both its own purity (Gini coefficient) and the requirement to maximize the interval of the SVM. When the SVM's validation set interval improvement rate in the current cascade layer is <1%, the cascade expansion of the deep forest is forcibly terminated to avoid redundant layers interfering with the methylation boundary division.

[0157] 5. Concatenate the base model prediction result Q with the D* feature to form Q*, and inject adaptive Gaussian noise. The formula for calculating the noise intensity coefficient is as follows: In the formula, Acc represents the accuracy of the prediction result, and the value of β ranges from [0.5, 1.5]. For positive samples, N(μ, α) is added. 2 In the formula, μ = 0.1α (adjusted based on the sparsity of positive methylation samples), and N(-μ,α) is added to negative samples. 2This improves the clarity of methylation site category boundaries (noise direction and intensity are optimized for CpG island boundaries), and the mean μ of directional Gaussian noise is optimized based on methylation boundary characteristics: μ = k × (1 - Conf boundary )×a, where k=0.1 is the methylation boundary adjustment factor, Conf boundary This represents the confidence level of the model on CpG island boundary samples, where the noise follows the order N(0, α). 2 )distributed;

[0158] 6. Triple cross-validation is used to evaluate the performance of the meta-model. The training set of the enhanced meta-model is divided into mutually exclusive subsets {S1, S2, S3}. When the deep forest is used as the meta-model, the weighted accuracy is calculated as Acc(Q*) = ∑w i ×Acc(S i ), in the formula To correspond to the weights, when SVM is used as the meta-model, the margin maximization criterion is used instead of the accuracy evaluation formula. In the formula, Margin is the average distance from the support vector to the hyperplane, and methylation boundary samples are optimized first in the calculation;

[0159] 7. Perform three-stage prediction on the test set T. In the first stage, output the original prediction F1 = BestModel, where BestModel is the base model with the highest Acc(Q*). In the second stage, the residual correction is calculated using the formula F2 = F1 + ΔF, where... The formula for calculating the residual is as follows: In the KNN algorithm, K is set to 5, and the spatial clustering of methylation sites is considered by filtering nearest neighbors through Euclidean distance in the feature space. In the third stage, the formula for calculating the result fusion is F3=γF2+(1-γ)·Mode({F1,F2}), where γ=0.6+0.1·Acc(Q*). When the optimal meta-model is a deep forest, it is adjusted according to the number of cascaded layers L. The adjustment formula is: γ=0.6+0.1×(Acc(Q*)-0.1×min(j,4)). If the optimal meta-model is an SVM, the fusion coefficient γ is increased by a compensation weight of 0.1 to compensate for the ambiguity of the methylation boundary, i.e. γ'=min(γ+0.1,0.9). The mode is taken for classification tasks, and the weighted average is taken for regression tasks.

[0160] In the above training steps, some better schemes and parameters are as follows:

[0161] In step 3, the output of the cascaded layers of the deep forest and the distance between the margin hyperplanes of the SVM are used as meta-features. The concatenation dimension is: [class probability vector | margin distance | original DNA binary code], where j is 4 and c is 103, containing 100 original features and 3 methylation-specific features. The weights θ of the SVM's hybrid kernel function are optimized using gradient descent, and the update formula is: In the formula, η = 0.01 is the learning rate, and Acc is... valid When the value is >0.8, a linear kernel weight is added for the linearly separable methylation feature, and the value of M is 8.2.

[0162] In the above method, preferably, in step 4, the value of α is 0.65, the result of Acc(Q) is 0.89, the value of μ is 0.065, and the value of β is 0.8.

[0163] In the above method, preferably, in step 6, the value of γ is 0.685.

[0164] Taking the training method into account, this application also has at least the following beneficial effects:

[0165] By employing dynamic multi-model ensemble and adaptive noise enhancement, this method significantly improves the accuracy and generalization ability of DNA methylation site identification. Compared to traditional single-model methods (such as SVM or random forest) or static ensemble methods, its deep forest's cascaded adaptive termination mechanism and SVM's dynamic optimization of mixed kernel functions effectively capture multi-scale patterns in methylation sequences (such as CpG island boundaries). By splicing model prediction results with original features and injecting directional Gaussian noise, the model's robustness to methylation data noise is enhanced, and class boundaries are optimized. Furthermore, the three-stage prediction fusion strategy (residual correction, nearest neighbor median adjustment, and weighted mode fusion) further improves the stability of methylation site results. Compared to techniques such as ChIP-seq and DAP-seq that rely on antibodies or complex experimental procedures, this method only requires methylation sequence information to achieve high-precision prediction. Moreover, through a triple cross-validation weighted evaluation mechanism, its overall performance on methylation benchmark datasets outperforms existing mainstream algorithms, while reducing dependence on computational resources and sample size.

[0166] Figure 4 This method demonstrates its application to genomic methylation data (methylation processes such as...). Figure 7 As shown, the DNA sequence is first encoded into a fixed-length binary matrix and weighted according to CpG islands. Then, a hybrid model of deep forest and SVM is constructed for training. Efficiency is improved by optimizing the dynamic parameters of methylation features and the hierarchical termination condition. Directional Gaussian noise is injected into the model output to enhance the boundary of methylation categories. After selecting the optimal meta-model by cross-validation, the original prediction, residual correction and result fusion are performed on the test set in stages. At the same time, a secondary validation mechanism is triggered for samples with low confidence in methylation (such as the edge of CpG islands), and finally the classification performance is gradually optimized.

[0167] The input data for this method mainly includes methylation-related FASTA and TSC files, located in the directories: train / sequences.fasta (training sequences) and train / labels.tsv (labels). Data preprocessing is performed on these input files. The output files include CSV and BED files, located in the directory: final_predictions.csv with fields: sequence_ID, predicted_label, confidence_score; and low_confidence_samples.bed with fields: chromosome number, start position, end position, and sample ID.

[0168] This identification method can be divided into the following steps:

[0169] Input data preparation and preprocessing

[0170] This embodiment selects a genome methylation dataset (methylation process such as...). Figure 6 The DNA sequence sample (as shown) includes DNA sequence data (in FASTA format, each sequence begins with a ">" to identify a unique ID, followed by a line containing the base sequence composed of A / T / C / G. The sequence length is fixed at 50bp, with any insufficient length padded with "N" and marked as invalid, such as the 50bp sequence identified by ">chr1:1000-1050") and a tag file (in TSV format (.tsv), containing two columns: "sequence_ID" and "label," with the label being a binary value (1 indicates a site, 0 indicates a non-site)). The tag file must be strictly aligned with the ID in the FASTA file, and the ratio of positive to negative samples should be controlled between 1:1 and 1:3. The preprocessing stage involves performing the following operations: FASTA sequences are processed according to... Figure 5 The rules shown are encoded as a three-dimensional binary matrix, and consecutive "CG" sequences are assigned a weight of 1.2 (simulating CpG island features). The calculation formula is as follows: (InfoGain of CpG island location) CpG =0.85), for example, the sequence "ATCG" is converted to [[0,0],[0,1],[1,0],[1,1]], and finally a matrix of shape 3200×50×2 is generated; invalid samples containing "N" are filtered at the same time to ensure data quality.

[0171] Table 1: Statistics on input dataset preprocessing

[0172]

[0173] 2. Multi-model container construction and training

[0174] The model container includes two base models: Deep Forest and SVM (e.g., Deep Forest and SVM). Figure 7 As shown):

[0175] Deep Forest: The cascaded structure has an upper limit of 10 layers. Each layer contains 4 completely random forests and 4 random forests. The feature subspace sampling rate is √d (d = 100). The feature subspace sampling preferentially includes methylation sensitive sites (such as "C" / "G" in positional encoding).

[0176] SVM: The initial parameter θ = 0.5 of the hybrid kernel function is dynamically optimized through gradient descent (learning rate η = 0.01), as shown in the formula. Optimization was performed on the linear boundary of methylation, and θ was eventually stabilized at 0.68, with the regularization parameter M optimized to 8.2.

[0177] During the base model co-training phase, deep forests and support vector machines achieve tight coupling and synergy through bidirectional feature penetration and joint optimization. The specific implementation process is as follows: After the j-th layer of the deep forest is cascaded and trained, the class probability vector P output by that layer is extracted. j ∈R n ×k (n is the number of samples, k is the number of classes), which is then used as dynamic features and concatenated to the right side of the original input feature matrix X of the SVM to form an extended feature space X′=[X|P j ]∈R n ×(d+k), where the original feature dimension d = 100, and is updated layer by layer as the cascade hierarchy increases; a joint loss function is constructed simultaneously, and a joint optimization objective is set. Where λ = 0.3 + 0.1·j is dynamically adjusted with the number of layers, and λ increases the weight of methylation features in layer j—when j = 1, DF purity dominates (λ = 0.4), and when j = 5, the SVM margin weight increases to 70% (λ = 0.8)—forcing the deep forest to select the split point from 10 candidate features that maximizes the decrease in the Gini coefficient and synchronously increases the SVM margin distance when splitting nodes; after each first-level layer training is completed, the margin distance improvement rate of the SVM on the validation set is calculated. If ΔMargin < 1% for two consecutive layers, cascade termination is triggered (at layer 6, Margin6 = 1.25 and Margin5 = 1.24, then ΔMargin = 0.8%, forcibly stopping expansion), thereby suppressing redundant layers from interfering with the boundary partitioning of the SVM, and the actual training terminates at layer 6. This collaborative mechanism enables the two types of models to form a bidirectional optimization loop in the hierarchical progression: the deep forest guides the SVM to focus on high-discrimination feature regions through probability vector injection, while the SVM constrains the feature splitting direction of the deep forest through interval feedback, ultimately producing an inseparable synergistic gain in the DNA sequence classification task.

[0178] After training, the output features of the two models are concatenated: the 2D class probability vector of the deep forest, the 1D margin distance of the SVM, and the original 100D binary encoding, to form a 103-dimensional feature matrix.

[0179] Table 2: Comparison of Base Model Training Performance

[0180]

[0181] 3. Noise Injection and Meta-model Training

[0182] Inject directional Gaussian noise into the prediction results of the basis model:

[0183] Noise intensity When the accuracy of the base model Acc(Q) = 0.86, α = 0.65;

[0184] Positive samples (methylation sites) are perturbed with N(0.065, 0.652), and negative samples are perturbed with N(-0.065, 0.652) to enhance the distinguishability of methylation category boundaries (the noise direction is optimized for CpG island boundaries).

[0185] The metamodel was evaluated using triple cross-validation: the augmented dataset was divided into mutually exclusive subsets S1 / S2 / S3 in a 4:4:2 ratio, and the weighted accuracy Acc(β) was calculated as ∑w i ×Acc(S i ), weight We prioritize selecting the SVM with the largest weighted interval or the deep forest with the highest accuracy as the meta-model (optimized for methylation boundaries).

[0186] Table 3: Impact of noise enhancement on meta-model performance

[0187]

[0188] 4. Three-stage prediction for the test set

[0189] Input the test set (500 FASTA sequences) and execute the following procedure:

[0190] Original predictions: Initial F1 results were generated using the optimal base model (deep forest, Acc = 0.86);

[0191] Residual correction: Calculate the 5 nearest neighbors of F1 (Euclidean distance ≤ 1.2) in the feature space, take the median of the predicted residuals ΔF, and the correction result F2 = F1 + ΔF;

[0192] Result fusion: The fusion coefficient was calculated as γ = 0.6 + 0.1Acc(β) = 0.685. If the meta-model is SVM, γ is increased by 0.1 to compensate for 0.785. The final result is F3 = 0.785F2 + 0.215*Mode{F1,F2} (mode is taken for classification tasks).

[0193] Table 4: Validation of the Three-Stage Prediction Process

[0194] Prediction phase accuracy F1 score Low confidence sample proportion <![CDATA[Original prediction (F1)]]> 86.4％ 0.892 18.3％ <![CDATA[Residual correction (F2)]]> 88.1％ 0.906 13.6％ <![CDATA[Result fusion (F3)]]> 89.3％ 0.917 11.2％

[0195] 5. Processing of low-confidence samples

[0196] When SVM detects a low-confidence sample with a margin < 0.5 (such as the edge of a CpG island), it triggers a deep forest secondary layer cascade validation: two additional validation layers are added in the 6th layer cascade. If the difference between the secondary result and the original prediction is > 0.3, the secondary result is used to overwrite the original output.

[0197] Implementation effect verification

[0198] In the HGU133 methylation dataset (3000 test sequences), the method of this invention achieved an accuracy of 89.3%, a significant improvement over single-model SVM (85.1%) and deep forest (86.4%); the misclassification rate of low-confidence samples was reduced to 15.2% (from 28.0%), and the F1-score reached 0.917. Output files include final_predictions.csv (predicted labels and confidence levels) and low_confidence_samples.bed (genomic coordinates of low-confidence samples).

[0199] The effectiveness of the method of this invention was verified on the HGU133 methylation dataset, through accuracy... The calculated classification accuracy is 89.3%, a significant improvement over traditional support vector machines (85.1%) and deep forests (86.4%). For low-confidence samples (samples with prediction confidence <0.7), the misclassification rate is... The decrease from 28.0% in the comparison algorithm to 15.2% indicates that the tight coupling collaboration and noise enhancement strategies effectively improved the model's robustness. Key metrics The AUC-ROC score reached 0.917, outperforming the single-model methods of 0.826 (SVM) and 0.892 (Deep Forest), demonstrating that the multi-model dynamic ensemble strategy improves the class imbalance problem through weight optimization (the complementarity of boundary partitioning in SVM and feature interaction in Deep Forest). Further analysis shows that the model's ability to distinguish between positive and negative samples (AUC-ROC area under the curve of 0.932) and its overall performance index, the Matthews correlation coefficient, are significantly higher. Both are higher than those of Random Forest (AUC-ROC 0.897, MCC 0.801) and XGBoost (AUC-ROC 0.908, MCC 0.832). The output file final_predictions.csv provides predicted labels and confidence data, supporting downstream interpretability analysis (such as SHAP value attribution), while the genomic coordinates marked in low_confidence_samples.bed can be directly used for CRISPR-Cas9 site-specific editing experiments, providing a high-precision tool for the study of DNA-protein interaction mechanisms.

[0200] In this example, the evaluation metrics for different algorithms on the HGU133 methylation dataset are as follows:

[0201]

[0202]

[0203] In this example, the parameters for comparing the performance of different algorithms are set as follows:

[0204]

[0205] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0206] Corresponding to the DNA site identification method based on multi-model dynamic integration described in the above embodiments, Figure 2 The diagram shows a structural block diagram of a DNA site recognition device based on multi-model dynamic integration provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiments of this application are shown.

[0207] Reference Figure 2 The device includes:

[0208] Sequence acquisition module 201, used to acquire DNA sequences;

[0209] The first running module 202 is used to run the first base model with the DNA sequence as input to obtain the class probability vector P. j The first base model is a machine learning model based on the deep forest algorithm, and j is the cascade layer number of the first base model.

[0210] The second operation module 203 is used to execute the class probability vector P. j To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction result Q. pThe second base model is a machine learning model based on support vector machines.

[0211] The first base model and the second base model constitute at least a part of the ensemble model, which is a machine learning model trained based on samples and labels; the dynamic update refers to the update of the input feature space of the second base model as the cascaded output of the j-th layer of the first base model is updated.

[0212] According to the above embodiments, in one possible implementation, the ensemble model further includes a meta-model; the meta-model is a machine learning model obtained through cross-validation; the cross-validation refers to:

[0213] Using at least two mutually exclusive subsets of the enhanced meta-training set as input, run candidate models, quantize the output of the candidate models based on the weights corresponding to the subsets, and determine the candidate model with the optimal quantization value as the meta-model;

[0214] The enhanced meta-training set is a training set obtained by splicing the features of the DNA sequence to the output of the first base model and / or the second base model, and injecting adaptive Gaussian noise; the candidate models include machine learning models based on the deep forest algorithm and machine learning models based on support vector machines.

[0215] According to the above embodiments, in one possible implementation, the adaptive Gaussian noise includes directional Gaussian noise for methylation boundaries.

[0216] According to the above embodiments, in one possible implementation, the apparatus further includes:

[0217] The stitching module is used to stitch together the prediction result Q. p Based on the characteristics of the DNA sequence, the input feature Q is obtained. * ;

[0218] The first prediction module is used to predict based on the input feature Q. * Running the meta-model yields the first prediction F1;

[0219] The second prediction module is used to correct the first prediction F1 based on the K-nearest neighbor algorithm to obtain the second prediction F2;

[0220] The third prediction module is used to fuse the first prediction F1 and the second prediction F2 to obtain the third prediction F3.

[0221] According to the above embodiments, in one possible implementation, the DNA sequence features include sequence weights; the sequence weights are determined based on information gain ratio, and the sequence weight values ​​of CpG island regions in the DNA sequence are higher than those of non-CpG island regions;

[0222] According to the above embodiments, in one possible implementation, the third prediction f3 satisfies:

[0223] F3=γG1(F2)+(1-γ)·G2(F1,F2)

[0224] Where G1 and G2 are both preset functions; γ is a preset function related to Acc(Q) * The fusion weights γ are given by γ, and the fusion weights γ satisfy:

[0225] If the meta-model is a machine learning model based on the deep forest algorithm, then the fusion weight γ is also related to the total number L of the cascaded layers of the meta-model, and Acc(Q) * The weighted accuracy is correlated with the meta-model.

[0226] If the meta-model is a machine learning model based on support vector machines, then Acc(Q) * The weighted interval is associated with the meta-model; wherein the interval is positively correlated with the average distance from the support vector to the hyperplane, and the interval is negatively correlated with the maximum distance from the support vector to the hyperplane.

[0227] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0228] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0229] This application also provides a terminal device, such as... Figure 3As shown, the terminal device 30 includes: at least one processor 301, a memory 302, and a computer program 303 stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the steps in any of the above-described method embodiments.

[0230] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0231] This application provides a computer program product that, when run on a mobile terminal, enables the mobile terminal to implement the steps described in the above-described method embodiments.

[0232] If the integrated unit is implemented as 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, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0233] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0234] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0235] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0236] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0237] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A DNA site identification method based on multi-model dynamic ensemble, characterized in that, include: Obtain the DNA sequence; Using the DNA sequence as input, the first basis model is run to obtain the class probability vector. The first base model is a machine learning model based on the deep forest algorithm. The cascade layer number of the first base model; With the class probability vector To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction results. The second base model is a machine learning model based on support vector machines. Wherein, the first base model and the second base model constitute at least a part of the ensemble model, and the ensemble model is a machine learning model trained based on samples and labels; the dynamic update refers to the change in the input feature space of the second base model as the first base model updates. Cascaded output updates of layers.

2. The DNA site identification method based on multi-model dynamic ensemble as described in claim 1, characterized in that, The ensemble model further includes a meta-model; the meta-model is a machine learning model obtained through cross-validation; the cross-validation refers to: Using at least two mutually exclusive subsets of the enhanced meta-training set as input, run candidate models, quantize the output of the candidate models based on the weights corresponding to the subsets, and determine the candidate model with the optimal quantization value as the meta-model; The enhanced meta-training set is a training set obtained by splicing the features of the DNA sequence to the output of the first base model and / or the second base model, and injecting adaptive Gaussian noise; the candidate models include machine learning models based on the deep forest algorithm and machine learning models based on support vector machines.

3. The DNA site identification method based on multi-model dynamic integration as described in claim 2, characterized in that, The adaptive Gaussian noise includes directional Gaussian noise for methylation boundaries.

4. The DNA site identification method based on multi-model dynamic integration as described in claim 2, characterized in that, The predicted result is obtained. Following the steps, it also includes: splicing the prediction results The input features are obtained by combining the characteristics of the DNA sequence. ; Based on the input features Running the meta-model yields the first prediction. ; The first prediction is corrected based on the K-nearest neighbor algorithm. The second prediction was obtained. ; Fusion of the first prediction and the second prediction The third prediction was obtained. .

5. The DNA site identification method based on multi-model dynamic integration as described in claim 4, characterized in that, The DNA sequence is characterized by sequence weights; The sequence weights are determined based on information gain ratio, and the sequence weights of CpG island regions in the DNA sequence are higher than those of non-CpG island regions.

6. The DNA site identification method based on multi-model dynamic integration as described in claim 4, characterized in that, The third prediction satisfy: in, , All are preset functions; For preset, related to The fusion weights, and the fusion weights satisfy: If the meta-model is a machine learning model based on the deep forest algorithm, then the fusion weights It is also related to the total number of cascaded layers in the metamodel. ,and Weighted accuracy associated with the meta-model; If the meta-model is a machine learning model based on support vector machines, then The weighted interval is associated with the meta-model; wherein the interval is positively correlated with the average distance from the support vectors to the hyperplane, and the interval is negatively correlated with the maximum distance from the support vectors to the hyperplane.

7. A DNA site recognition device based on multi-model dynamic integration, characterized in that, include: Sequence acquisition module, used to acquire DNA sequences; The first execution module is used to run the first base model with the DNA sequence as input to obtain the class probability vector. The first base model is a machine learning model based on the deep forest algorithm. The cascade layer number of the first base model; The second execution module is used to execute the class probability vector. To add additional feature dimensions, a second base model is run and updated dynamically to obtain the prediction results. The second base model is a machine learning model based on support vector machines. Wherein, the first base model and the second base model constitute at least a part of the ensemble model, and the ensemble model is a machine learning model trained based on samples and labels; the dynamic update refers to the change in the input feature space of the second base model as the first base model updates. Cascaded output updates of layers.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

10. A computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images