Collagen peptide absorptivity prediction method

Abstract

The invention relates to the field of artificial intelligence, in particular to a collagen peptide absorptivity prediction method. Comprising the following steps: acquiring a collagen peptide sample data matrix, and constructing a convolutional neural network and a recurrent neural network to obtain a molecular structure feature vector and a time sequence feature vector; constructing a prediction neural network, inputting the molecular structure feature vector and the time sequence feature vector into an interactive feature fusion layer, and obtaining a fused feature vector by using an interactive feature fusion algorithm; inputting the fused feature vector into a self-adaptive decoupling layer, and calculating by using a self-adaptive feature decoupling algorithm to obtain a decoupled feature vector; and inputting the decoupled feature vector into an output layer to obtain a predicted value of the collagen peptide absorptivity. The technical problems that the molecular structure feature extraction capacity is insufficient, the complex interaction relation between local molecular structures cannot be flexibly captured, feature extraction is not deep enough, and the accuracy of collagen peptide absorptivity prediction is affected are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method for predicting the absorption rate of collagen peptides. Background Art

[0002] As people pay more and more attention to health and beauty, collagen, as an important structural protein, plays an important role in maintaining the integrity of tissues such as skin, joints and bones. In recent years, collagen peptides have gradually become an important ingredient in nutritional supplements due to their good biological activity and easy absorption properties. The absorption rate of collagen peptides directly affects the bioavailability in the body, thereby affecting the efficacy. Therefore, predicting the absorption rate of collagen peptides has important practical significance.

[0003] In the development and application of collagen peptides, the ability to predict the absorption rate of collagen peptides is crucial. Although there are methods to study and predict the absorption rate of collagen peptides, there are still certain limitations. Traditional bioavailability studies rely on animal experiments or clinical trials, which are not only time-consuming but also costly. Therefore, a fast and accurate method for predicting the absorption rate of collagen peptides has emerged. With the advancement of technology, especially the development of computational biology and artificial intelligence, it has become possible to build a scientific and accurate prediction model. The research on the prediction method of collagen peptide absorption rate not only helps to gain a deeper understanding of the biological characteristics of collagen peptides, but also provides theoretical support and practical basis for the application of collagen peptides.

[0004] However, the existing collagen peptide absorption rate prediction methods have the following technical problems: insufficient ability to extract molecular structure features, unable to flexibly capture the complex interactions between local molecular structures, resulting in insufficient depth of feature extraction, which in turn affects the accuracy of collagen peptide absorption rate prediction; poor performance when processing long time series or periodic fluctuation dynamic features, especially insufficient ability to identify and capture periodic changes, resulting in limited understanding of dynamic features; traditional feature fusion methods use simple linear superposition or feature splicing, which makes it difficult to effectively fuse features from different sources, especially unable to fully capture the complex interactions between molecular structure features and time series features; it is difficult to remove redundant information and accurately extract key features, which reduces the generalization ability and prediction performance; the existing output layer makes it difficult to capture the overall amplitude information of the feature vector and the nonlinear relationship of complex features, which reduces the understanding and expression of the overall scale and complexity of the features. Summary of the invention

[0005] The present invention provides a method for predicting the absorption rate of collagen peptides, so as to solve the problems of insufficient ability to extract molecular structure features, inability to flexibly capture the complex interactive relationship between local molecular structures, resulting in insufficient depth of feature extraction, and further affecting the accuracy of collagen peptide absorption rate prediction; poor performance when processing long time series or periodic fluctuation dynamic features, especially insufficient ability to identify and capture periodic changes, resulting in limited understanding of dynamic features; traditional feature fusion methods use simple linear superposition or feature splicing, which is difficult to effectively fuse features from different sources, especially unable to fully capture the complex interactive relationship between molecular structure features and time series features; difficult to remove redundant information and accurately extract key features, which reduces generalization ability and prediction performance; the existing output layer is difficult to capture the overall amplitude information of feature vectors and the nonlinear relationship of complex features, which reduces the understanding and expression ability of the overall scale and complexity of features.

[0006] A method for predicting the absorption rate of collagen peptides of the present invention specifically includes the following technical solutions:

[0007] A method for predicting the absorption rate of collagen peptides comprises the following steps:

[0008] S1: Obtain collagen peptide sample data matrix, build a convolutional neural network based on the collagen peptide sample data matrix, perform double adaptive convolution operation, and output molecular structure feature vector; build a recurrent neural network, and use the recurrent neural network state update mechanism to obtain the time series feature vector;

[0009] s2: Construct a prediction neural network, take the molecular structure feature vector and the time series feature vector as the input of the prediction neural network, input them into the interactive feature fusion layer, and use the interactive feature fusion algorithm to obtain the fused feature vector;

[0010] S3: Input the fused feature vector into the adaptive decoupling layer, and use the adaptive feature decoupling algorithm to calculate the decoupled feature vector;

[0011] S4: Input the decoupled feature vector into the output layer to obtain the predicted value of collagen peptide absorption rate.

[0012] Preferably, the S1 specifically includes:

[0013] In the process of implementing the dual adaptive convolution operation, by introducing the adaptive convolution kernel selection mechanism and dynamic weight adjustment method, the convolution layer adjusts the size and shape of the convolution kernel according to the input two-dimensional collagen peptide sample data matrix, and the data elements in the input two-dimensional collagen peptide sample data matrix are sent to convolution kernels of different sizes and shapes for convolution operation.

[0014] Preferably, the S1 specifically includes:

[0015] Through the double adaptive convolution operation, the molecular structure feature vector is obtained. The specific implementation formula is as follows:

[0016]

[0017] in, represents the molecular structure feature vector of the l+1th convolutional layer; Leaky-ReLU represents the activation function; represents the sum of all adaptive convolution kernels, K is the number of adaptive convolution kernels; α k Represents the adaptive weight of the kth adaptive convolution kernel; represents the weight matrix of the kth adaptive convolution kernel in the lth convolution layer; * represents the convolution operator; X i represents the two-dimensional collagen peptide sample data matrix; Represents the bias term of the kth adaptive convolution kernel in the lth convolution layer.

[0018] Preferably, the S1 specifically includes:

[0019] The calculation formula of time series feature vector is:

[0020]

[0021] Among them, h t represents the hidden state at time step t, i.e., the time series feature vector; h t-1 represents the hidden state at time step t-1; x t，j represents the sample data element of the jth input at time step t; tanh(·) represents the hyperbolic tangent activation function; Represents the hidden state weight matrix of the sample data element of the jth input; Indicates the weighted sum of all input sample data elements; U h represents the weight matrix between hidden states; b h represents the bias vector of the hyperbolic tangent activation function; sin(·) represents the sine activation function; The weight matrix of the sine term of the sample data element of the jth input; U s represents the weight matrix from hidden state to sine term; b s Bias vector representing the sinusoidal part.

[0022] Preferably, the S2 specifically includes:

[0023] In the implementation process of the interactive feature fusion algorithm, the molecular structure feature vector and the time series feature vector are mapped to the same feature space through linear transformation; the element-by-element interaction mechanism and the high-order interaction mechanism are introduced for feature extraction; and the hyperbolic tangent activation function is used to map the linearly combined feature vector to a nonlinear space to obtain the fused feature vector.

[0024] Preferably, the S2 specifically includes:

[0025] The calculation formula of the fused feature vector is:

[0026]

[0027] Among them, z f represents the fused feature vector; W1, W2, W3, W4 represent the weight matrix of the fusion layer; represents the output of the Lth convolutional layer, i.e., the molecular structure feature vector; h T represents the hidden state of the Tth time step, that is, the time series feature vector; b f Represents the bias vector of the feature fusion part; ⊙ represents the element-level product; Represents the outer product.

[0028] Preferably, the S3 specifically includes:

[0029] In the implementation process of the adaptive feature decoupling algorithm, the fused feature vector is linearly transformed and nonlinearly processed using the hyperbolic tangent activation function. The results of the two linear transformations are element-by-element multiplied, and the features after element-by-element multiplication are normalized to obtain the decoupled feature vector.

[0030] Preferably, the S3 specifically includes:

[0031] The calculation formula of the decoupled eigenvector is:

[0032]

[0033] Among them, z d represents the decoupled eigenvector; W d1 represents the first set of weight matrices of the decoupled layer; b d1 represents the bias vector of the first linear transformation of the decoupling layer; ⊙ represents the element-wise product; W d2 represents the second set of weight matrices of the decoupled layer; b d2 represents the bias vector of the second linear transformation of the decoupling layer; ||·||2 represents the L2 norm; ∈ represents a very small constant.

[0034] Preferably, the S4 specifically includes:

[0035] The calculation formula for the predicted value of collagen peptide absorption rate is:

[0036]

[0037] in, represents the predicted value of collagen peptide absorption rate; Sigmoid represents the activation function; W out represents the output layer weight matrix; b out represents the bias vector of the output layer; η represents the adjustment parameter used to adjust the influence of the nonlinear term; ln(·) represents the natural logarithm function; ||z d ||2 represents the L2 norm of the decoupled eigenvector.

[0038] The beneficial effects of the technical solution of the present invention are:

[0039] 1. By constructing a convolutional neural network and performing double adaptive convolution operations, the deep features of the collagen peptide molecular structure are extracted layer by layer, which can capture complex local molecular structure information and improve the richness and expression ability of feature representation; by introducing the adaptive convolution kernel selection mechanism and dynamic weight adjustment method, the convolution layer can flexibly adjust the size and shape of the convolution kernel according to the input sample data, thereby improving the accuracy and adaptability of feature extraction. The Leaky-ReLU activation function is used to introduce nonlinear transformation, which enhances the fitting ability of complex feature relationships and avoids the problem of neuron "death".

[0040] 2. By utilizing the recurrent neural network state update mechanism and combining the nonlinear activation of the hyperbolic tangent function and the sine function, we can enhance our understanding of the dynamic characteristics of periodicity and volatility in time series, effectively capture the dependencies between time steps, and improve our ability to capture the complex dynamic relationships of time series data.

[0041] 3. The molecular structure feature vector and the time series feature vector are linearly transformed, element-by-element interacted and outer product operated to generate a fused feature vector. The fine-grained and complex implicit relationships between the features are revealed through element-by-element multiplication and outer product operations, thereby enhancing the ability to understand feature combinations.

[0042] 4. Through the adaptive feature decoupling algorithm, the redundant information and key features of the fused feature vector are removed to generate the decoupled feature vector, which effectively removes the redundant information in the feature, retains the key useful information, and improves the effectiveness of feature representation.

[0043] 5. By cleverly combining linear and nonlinear transformations and capturing the linear and nonlinear relationship of features, accurate prediction of collagen peptide absorption rate is achieved, and the predicted value is mapped to between 0 and 1 using the Sigmoid activation function to ensure that the output result is within a reasonable range of biomedical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of a method for predicting collagen peptide absorption rate according to the present invention. DETAILED DESCRIPTION

[0045] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0047] The specific scheme of the method for predicting the absorption rate of collagen peptides provided by the present invention is described in detail below with reference to the accompanying drawings.

[0048] See attached Figure 1 , which shows a flow chart of a method for predicting collagen peptide absorption rate provided by an embodiment of the present invention, the method comprising the following steps:

[0049] S1. Obtain collagen peptide sample data matrix, construct convolutional neural network based on collagen peptide sample data matrix, perform double adaptive convolution operation, and output molecular structure feature vector; construct recurrent neural network, and use recurrent neural network state update mechanism to obtain time series feature vector;

[0050] Get the collagen peptide sample data matrix, which is specifically expressed as:

[0051]

[0052] Among them, X i represents the two-dimensional collagen peptide sample data matrix; x t，j represents the jth input sample data element at time step t, such as molecular weight, amino acid sequence; T represents the length of the time series; m represents the dimension of the sample data input at each time step;

[0053] Based on the collagen peptide sample data matrix, a convolutional neural network with L layers is constructed. Each layer of the convolutional neural network applies a double adaptive convolution operation, and the output of each layer is used as the input of the next layer to extract the molecular structure features layer by layer. After L layers of double adaptive convolution operations, the molecular structure feature vector is output. Molecular structure feature vector It contains the deep features of the molecular structure of collagen peptides, as well as complex local molecular structure information;

[0054] The dual adaptive convolution operation is designed to enhance the ability to extract complex molecular structure features. By introducing an adaptive convolution kernel selection mechanism and a dynamic weight adjustment method, the convolution layer can flexibly adjust the size and shape of the convolution kernel according to the input two-dimensional collagen peptide sample data matrix, thereby improving the accuracy of feature extraction;

[0055] Specifically, the sample data elements in the input two-dimensional collagen peptide sample data matrix are respectively sent to convolution kernels of different sizes and shapes for convolution operation, and the features extracted by each convolution kernel are multiplied by the corresponding weight to adjust the weight contribution, and all the features processed by the convolution kernel are weighted summed, and the Leaky-ReLU activation function is further applied to introduce nonlinear transformation. The Leaky-ReLU activation function can not only retain part of the negative value information and avoid the problem of neuron "death", but also enhance the nonlinear expression ability and fit complex feature relationships;

[0056] The double adaptive convolution operation formula is as follows:

[0057]

[0058] in, represents the molecular structure feature vector of the l+1th convolutional layer; Leaky-ReLU represents the activation function, which is used to increase the nonlinear expression ability; represents the sum of all adaptive convolution kernels, K is the number of adaptive convolution kernels; α k Represents the adaptive weight of the kth adaptive convolution kernel, which is used to dynamically adjust the contribution of different convolution kernels; represents the weight matrix of the kth adaptive convolution kernel in the lth convolution layer; * represents the convolution operator, which is used for convolution operation; X i represents the two-dimensional collagen peptide sample data matrix; Represents the bias term of the kth adaptive convolution kernel in the lth convolution layer;

[0059] The flexibility and accuracy of processing complex molecular structure features are significantly enhanced through the adaptive convolution kernel selection mechanism and dynamic weight adjustment method.

[0060] Based on the collagen peptide sample data matrix, a recurrent neural network is constructed, and the hidden state of each time step is obtained using the recurrent neural network state update mechanism. After processing the entire time series T, the hidden state h of the Tth time step is obtained. T , i.e., the time series feature vector, which integrates the feature interactions and dynamic change information of all time steps;

[0061] The recurrent neural network state update mechanism is designed to improve the ability to capture complex dynamic relationships in time series data. By combining the nonlinear activation of hyperbolic tangent function and sine function, the understanding of periodic and fluctuating dynamic characteristics is enhanced, thereby improving the prediction accuracy of collagen peptide absorption rate.

[0062] Specifically, at each time step, each input sample data element is multiplied by the corresponding weight matrix to adjust the influence of the input sample data element in the hidden state update. The hidden state of the previous time step is also linearly transformed through the weight matrix between the hidden states to capture the dependency between time steps. The linear combination of all input sample data elements adjusted by the weight matrix and the hidden state of the previous time step, plus the bias vector of the hyperbolic tangent activation function, is nonlinearly transformed through the hyperbolic tangent activation function tanh(·), which introduces nonlinear expression capabilities and can handle complex patterns and relationships.

[0063] In order to further enhance the ability to capture periodic or fluctuating dynamic features in time series, the sinusoidal activation function sin(·) is introduced as a different nonlinear activation method. The input sample data elements and the hidden state of the previous time step are linearly combined through their respective weight matrices, and the bias vector of the sinusoidal term is added. Then, the nonlinear transformation is performed through the sinusoidal function, which can identify periodic changes and improve the sensitivity to dynamic features.

[0064] The hidden state at each time step is calculated as:

[0065]

[0066] Among them, h t represents the hidden state at time step t, i.e., the time series feature vector; ht-1 represents the hidden state at time step t-1; x t，j represents the sample data element of the jth input at time step t; tanh(·) represents the hyperbolic tangent activation function, which is used to introduce nonlinear transformation; Represents the hidden state weight matrix of the sample data element of the jth input; Indicates the weighted sum of all input sample data elements; U h represents the weight matrix between hidden states; bh represents the bias vector of the hyperbolic tangent activation function, which is used to adjust the output of the linear combination; sin(·) represents the sine activation function, which is used to introduce periodic nonlinear transformation and capture the periodic or fluctuating dynamic characteristics in the time series; The weight matrix of the sine term of the sample data element of the jth input; U s represents the weight matrix from hidden state to sine term; b s The bias vector representing the sinusoidal term is used to adjust the output of the linear combination;

[0067] S2, construct a prediction neural network, take the molecular structure feature vector and the time series feature vector as the input of the prediction neural network, input them into the interactive feature fusion layer, and use the interactive feature fusion algorithm to obtain the fused feature vector;

[0068] Construct a prediction neural network, including an interactive feature fusion layer, an adaptive decoupling layer, and an output layer; use the molecular structure feature vector and the time series feature vector as inputs of the prediction neural network and input them into the interactive feature fusion layer;

[0069] In the interactive feature fusion layer, the interactive feature fusion algorithm is used to obtain the fused feature vector;

[0070] The interactive feature fusion algorithm aims to effectively integrate the molecular structure feature vector and the time series feature vector of the collagen peptide to improve the accuracy of the prediction of the absorption rate of the collagen peptide;

[0071] The molecular structure feature vector and the time series feature vector are mapped to the same feature space through linear transformation, ensuring that the features from different neural networks are compared and fused at the same scale and dimension;

[0072] After completing the linear transformation, an element-by-element interaction mechanism is introduced. By performing element-by-element multiplication of the molecular structure feature vector and the time series feature vector, the element-by-element correlation between the two is captured, which can reveal the fine-grained interaction relationship between features and enhance the ability to understand feature combinations. The element-by-element interaction results are linearly transformed through the weight matrix of the interactive feature fusion layer to further extract meaningful interactive features.

[0073] Furthermore, a high-order interaction mechanism is introduced, which can capture complex and implicit feature relationships by calculating the outer product of the molecular structure feature vector and the time series feature vector. Linear transformation is also performed through the weight matrix to extract effective information;

[0074] The hyperbolic tangent activation function tanh(·) was applied to map the linearly combined features into a nonlinear space, which enhanced the ability to capture complex patterns. The fused feature vector has extremely high expressiveness and relevance, which significantly improved the prediction accuracy and robustness of collagen peptide absorption rate.

[0075] The calculation formula of the fused feature vector is:

[0076]

[0077] Among them, z f represents the fused feature vector; tanh(·) represents the hyperbolic tangent activation function, which is used to introduce nonlinear transformation; W1, W2, W3, and W4 represent the weight matrix of the fusion layer, which is used for linear transformation; represents the output of the Lth convolutional layer, i.e., the molecular structure feature vector, which reflects the local spatial characteristics of the peptide segment; h T represents the hidden state of the Tth time step, that is, the time series feature vector, which captures the dynamic change information of the peptide segment with the time step; b f Represents the bias vector of the feature fusion part, which is used to adjust the linear output after feature fusion; ⊙ represents the element-level product, which is used for element-by-element interaction to capture the relationship between features in the local dimension; Represents the outer product, which is used to capture the complex relationship between features;

[0078] S3, input the fused feature vector into the adaptive decoupling layer, and use the adaptive feature decoupling algorithm to calculate the decoupled feature vector;

[0079] The fused feature vector is input into the adaptive decoupling layer, and the decoupled feature vector is calculated using the adaptive feature decoupling algorithm;

[0080] The adaptive feature decoupling algorithm aims to improve the prediction performance of collagen peptide absorption rate and the effectiveness of feature representation by removing redundant information and accurately extracting key features;

[0081] Perform linear transformation on the fused feature vector. Use two independent weight matrices to perform different linear transformations on the fused feature vector respectively, and add corresponding bias terms after each linear transformation. The double linear transformation can capture different aspects and potential relationships in the fused feature vector and enhance the expressiveness of the feature.

[0082] Applying the hyperbolic tangent activation function tanh(·) for nonlinear processing can introduce complex feature interactions and improve the ability to capture nonlinear patterns, making the feature representation rich and discriminative.

[0083] The results of the two linear transformations are multiplied element by element to further strengthen the interaction between different features, which can capture the subtle correlation between features and enhance the complexity and expressiveness of feature representation;

[0084] In order to ensure the stability of feature representation and numerical calculation, the features after element-by-element multiplication are normalized. Specifically, the second norm of the linear transformation result is calculated and a very small constant is added to the denominator to prevent division by zero errors.

[0085] The final output is the decoupled feature vector, which not only retains the key useful information but also removes unnecessary redundancy.

[0086] The calculation formula of the decoupled eigenvector is:

[0087]

[0088] Among them, z d represents the decoupled eigenvector; W d1 represents the first set of weight matrices of the decoupled layer; b d1 represents the bias vector of the first linear transformation of the decoupling layer, which provides additional degrees of freedom for capturing offset features; ⊙ represents the element-wise product, which is used to enhance the interaction between features; tanh(·) represents the hyperbolic tangent activation function, which is used to introduce nonlinearity and help capture complex patterns; W d2 represents the second set of weight matrices of the decoupled layer; b d2 Represents the bias vector of the second linear transformation of the decoupling layer, which provides additional degrees of freedom for capturing offset features; ||·||2 represents the L2 norm, which is used for normalization to ensure stability and prevent gradient explosion; ∈ represents a very small constant to prevent division by zero errors;

[0089] S4, inputting the decoupled feature vector into the output layer to obtain the predicted value of collagen peptide absorption rate;

[0090] In the output layer, the decoupled feature vectors of the input are linearly combined through an output layer weight matrix and an output layer bias vector to adjust the importance and influence of the features;

[0091] The nonlinear term is introduced. The L2 norm of the decoupled input feature vector is taken as logarithm and then multiplied by the adjustment parameter. This is to capture the overall scale and complexity of the features, thereby improving the understanding and expression of the nonlinear relationship between features. Specifically, the L2 norm provides the overall amplitude information of the feature vector, while the logarithmic transformation converts the amplitude information into a scale suitable for model processing, avoiding the adverse effects of maximum or minimum values ​​on model training.

[0092] The design of the entire output layer is not just a simple superposition of linear or nonlinear transformations, but through clever combination and parameter adjustment, it achieves effective fusion and expression of features. The linear transformation part is responsible for capturing the linear relationship of features, while the logarithmic transformation enhances the understanding of feature scale and complexity. The Sigmoid activation function maps the results of the comprehensive transformation into a reasonable prediction range, achieving accurate prediction of the absorption rate of collagen peptides.

[0093] The calculation formula for the predicted value of collagen peptide absorption rate is:

[0094]

[0095] in, represents the predicted value of collagen peptide absorption rate; Sigmoid(·) represents the activation function, which is used to map the result of linear combination to between 0 and 1 to ensure that the output predicted value is within a reasonable range; W out represents the output layer weight matrix, which is used for linear transformation to extract feature combinations useful for predicting collagen peptide absorption rate; b out represents the bias vector of the output layer, which is used to adjust the result after linear transformation; η represents the adjustment parameter used to adjust the influence of nonlinear terms; ln(·) represents the natural logarithmic function, which performs logarithmic transformation to capture the amplitude change of features and further enhance the nonlinear expression ability; ||z d ||2 represents the L2 norm of the decoupled eigenvector, which is used to measure the overall amplitude of the eigenvector;

[0096] The above steps achieve accurate prediction of collagen peptide absorption rate, ensuring the robustness and accuracy of the prediction results in biomedical applications.

[0097] In summary, a method for predicting the absorption rate of collagen peptides was completed.

[0098] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0099] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for predicting collagen peptide absorption rate, characterized in that: The following steps are involved: S1: Obtain collagen peptide sample data matrix, build a convolutional neural network based on the collagen peptide sample data matrix, perform double adaptive convolution operation, and output molecular structure feature vector; build a recurrent neural network, and use the recurrent neural network state update mechanism to obtain the time series feature vector; S2: Construct a prediction neural network, take the molecular structure feature vector and the time series feature vector as the input of the prediction neural network, input them into the interactive feature fusion layer, and use the interactive feature fusion algorithm to obtain the fused feature vector; S3: Input the fused feature vector into the adaptive decoupling layer, and use the adaptive feature decoupling algorithm to calculate the decoupled feature vector; S4: Input the decoupled feature vector into the output layer to obtain the predicted value of collagen peptide absorption rate.

2. A method for predicting collagen peptide absorption rate according to claim 1, characterized in that: The S1 specifically includes: In the process of implementing the dual adaptive convolution operation, by introducing the adaptive convolution kernel selection mechanism and dynamic weight adjustment method, the convolution layer adjusts the size and shape of the convolution kernel according to the input two-dimensional collagen peptide sample data matrix, and the data elements in the input two-dimensional collagen peptide sample data matrix are sent to convolution kernels of different sizes and shapes for convolution operation.

3. A method for predicting collagen peptide absorption rate according to claim 2, characterized in that: The S1 specifically includes: Through the double adaptive convolution operation, the molecular structure feature vector is obtained. The specific implementation formula is as follows: in, represents the molecular structure feature vector of the l+1th convolutional layer; Leaky-ReLU represents the activation function; represents the sum of all adaptive convolution kernels, K is the number of adaptive convolution kernels; α k Represents the adaptive weight of the kth adaptive convolution kernel; represents the weight matrix of the kth adaptive convolution kernel in the lth convolution layer; * represents the convolution operator; X i represents the two-dimensional collagen peptide sample data matrix; Represents the bias term of the kth adaptive convolution kernel in the lth convolution layer.

4. The method for predicting collagen peptide absorption rate according to claim 1, characterized in that: The S1 specifically includes: The calculation formula of time series feature vector is: Among them, h t represents the hidden state at time step t, i.e., the time series feature vector; h t-1 represents the hidden state at time step t-1; x t，j represents the sample data element of the jth input at time step t; tanh(·) represents the hyperbolic tangent activation function; Represents the hidden state weight matrix of the sample data element of the th input; Indicates the weighted sum of all input sample data elements; U h represents the weight matrix between hidden states; b h represents the bias vector of the hyperbolic tangent activation function; sin(·) represents the sine activation function; The weight matrix of the sine term of the sample data element of the jth input; U s represents the weight matrix from hidden state to sine term; b s Bias vector representing the sinusoidal part.

5. The method for predicting collagen peptide absorption rate according to claim 1, characterized in that: The S2 specifically includes: In the implementation process of the interactive feature fusion algorithm, the molecular structure feature vector and the time series feature vector are mapped to the same feature space through linear transformation; the element-by-element interaction mechanism and the high-order interaction mechanism are introduced for feature extraction; and the hyperbolic tangent activation function is used to map the linearly combined feature vector to a nonlinear space to obtain the fused feature vector.

6. A method for predicting collagen peptide absorption rate according to claim 5, characterized in that: The S2 specifically includes: The calculation formula of the fused feature vector is: Among them, z f represents the fused feature vector; W1, W2, W3, W4 represent the weight matrix of the fusion layer; represents the output of the Lth convolutional layer, i.e., the molecular structure feature vector; h T represents the hidden state of the tth time step, that is, the time series feature vector; b f Represents the bias vector of the feature fusion part; ⊙ represents the element-level product; Represents the outer product.

7. The method for predicting collagen peptide absorption rate according to claim 1, characterized in that: The S3 specifically includes: In the implementation process of the adaptive feature decoupling algorithm, the fused feature vector is linearly transformed and nonlinearly processed using the hyperbolic tangent activation function. The results of the two linear transformations are element-by-element multiplied, and the features after element-by-element multiplication are normalized to obtain the decoupled feature vector.

8. A method for predicting collagen peptide absorption rate according to claim 7, characterized in that: The S3 specifically includes: The calculation formula of the decoupled eigenvector is: Among them, z d represents the decoupled eigenvector; W d1 represents the first set of weight matrices of the decoupled layer; b d1 represents the bias vector of the first linear transformation of the decoupling layer; ⊙ represents the element-wise product; W d2 represents the second set of weight matrices of the decoupled layer; b d2 represents the bias vector of the second linear transformation of the decoupling layer; ||·||2 represents the L2 norm; ∈ represents a very small constant.

9. The method for predicting collagen peptide absorption rate according to claim 1, characterized in that: The S4 specifically includes: The calculation formula for the predicted value of collagen peptide absorption rate is: in, represents the predicted value of collagen peptide absorption rate; Sigmoid represents the activation function; W out represents the output layer weight matrix; b out represents the bias vector of the output layer; η represents the adjustment parameter used to adjust the influence of the nonlinear term; ln(·) represents the natural logarithm function; ||z d ||2 represents the L2 norm of the decoupled eigenvector.

Images