Lipocalin classification method based on CNN-LSTM multi-feature fusion

Abstract

The invention discloses a CNN-LSTM multi-feature fusion-based lipocalin classification method, which comprises the following steps: S1, constructing a data set, and dividing the data set into training data, verification data and test data; s2, extracting protein sequence features by using three feature descriptors of K-mer, CKSAAP and CC-PSSM, and fusing the protein sequence features through a feature integration technology; s3, performing dimensionality reduction optimization on the fused features by using principal component analysis (PCA), and reserving features with a variance proportion threshold value of 0.98 to obtain low-dimensional important features; s4, the features subjected to dimension reduction are input into a CNN-LSTM fusion model, CNN comprises 128 filters, the size of a convolution kernel is 1, an activation function is ReLU, LSTM comprises 64 units, and a category probability is output through combination of a full connection layer and a Softmax function; and S5, performing classification prediction on the lipid protein according to the output probability value. According to the method, the problems of insufficient characteristic utilization and high calculation cost of a traditional method are solved, and an efficient tool is provided for lipid protein function research and disease diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics and deep learning technology, and in particular to a lipocalin classification method based on CNN-LSTM multi-feature fusion. Background Art

[0002] Lipoproteins are multifunctional proteins that transport small hydrophobic molecules and participate in other important functions, such as stress responses and allergic inflammation. They are typically composed of 165-200 residues. In recent years, many studies have used position-specific scoring matrices (PSSMs) and neural networks to identify proteins. However, conventional neural networks cannot fully utilize the potential feature information in PSSMs, and single features cannot achieve excellent accuracy.

[0003] Lipid proteins play an important role in physiological processes such as transporting hydrophobic molecules and regulating the immune system. However, their sequence homology is low, making traditional experimental methods time-consuming and costly. Existing machine learning methods (such as support vector machines (SVMs) and recurrent neural networks (RFs)) rely on single features or fail to fully integrate information from multiple sources, resulting in limited accuracy. For example, the SVM model proposed by Ramana et al. achieved an accuracy of 90.72%, while the RF model proposed by Zulfiqar et al. achieved an accuracy of 89.90%. Furthermore, traditional methods are computationally complex and struggle to process large amounts of data. Summary of the Invention

[0004] (1) Technical problems solved

[0005] To address the shortcomings of existing technologies, the present method uses three feature descriptors, K-mer, CKSAAP, and CC-PSSM, to extract local conservation, spatial distribution, and evolutionary information from protein sequences. This method, combined with principal component analysis (PCA) dimensionality reduction optimization, constructs a CNN-LSTM fusion model for classification and prediction. This method addresses the underutilization of features and high computational costs of traditional methods, providing a highly effective tool for lipid and protein functional research and disease diagnosis.

[0006] (2) Technical solution

[0007] To achieve the above object, the present invention provides the following technical solution: a lipocalin classification method based on CNN-LSTM multi-feature fusion, the method comprising the following steps:

[0008] S1. Build a dataset and divide it into training data, validation data, and test data;

[0009] S2, using K-mer, CKSAAP and CC-PSSM three feature descriptors to extract protein sequence features and fusion through feature integration technology;

[0010] S3. Use principal component analysis (PCA) to optimize the dimensionality reduction of the fused features, retain the features with a variance ratio threshold of 0.98, and obtain low-dimensional important features;

[0011] S4. Input the reduced-dimensional features into the CNN-LSTM fusion model, where the CNN contains 128 filters, the convolution kernel size is 1, the activation function is ReLU, and the LSTM contains 64 units. The class probability is output through a fully connected layer combined with the Softmax function.

[0012] S5. Classify and predict lipid proteins based on the output probability values.

[0013] As a preferred solution, the data set construction in step S1 includes:

[0014] S11, obtain lipid protein and non-lipid protein sequences from the UniProt database;

[0015] S12, using the CD-Hit tool with a threshold of 0.4 to remove homologous sequences;

[0016] S13. Exclude sequences containing non-standard amino acid characters (X, B, Z, J, O), and finally obtain training set samples and independent test set samples.

[0017] As a preferred solution, in step S2, the sequences in the data set are aligned with homologous sequences in the NCBI non-redundant (NR) database using PSI-BLAST software, and a PSSM is obtained. The PSSM is represented as follows:

[0018]

[0019] Where m is the length of the protein sequence L, a m,n Represents the amino acid residue at position m in the protein sequence.

[0020] As a preferred solution, in step S2, the CC-PSSM algorithm converts PSSM matrices of different sizes into vectors of the same length, and uses CC to calculate the property difference between two lagged separated residues. The calculation formula is as follows:

[0021]

[0022] Where i is the residue (i∈[1,20]), i1,i2 represent two different amino acids, z represents the distance between the two different residues, i.e., hysteresis, S i,j represents the PSSM score of i at point j, and the size of CC-PSSM is 380×z.

[0023] As a preferred solution, in step S2, the CKSAAP feature takes into account the situation where the amino acid pairs are separated by any number of "K" amino acids (K can be 0, 1, 2, 3, 4, or 5), and the calculation formula is as follows:

[0024]

[0025] It is used to calculate the relative frequency of a specific amino acid pair (xy) when considering "K" amino acids apart; where: Q x,y : relative frequency of a specific amino acid pair; V x,y : the number of specific amino acid pairs; N: the total number of amino acids; K: the number of intervals between amino acids.

[0026] As a preferred solution, in step S2, the principle of K-mer encoding is to calculate the frequency of k nucleotides in the protein sequence. The protein sequence R can be converted into a vector with 4k dimensions by using the K-mer frequency. The formula is as follows:

[0027]

[0028] where f i k-mer is the normalized frequency of occurrence of the i-th Kmers nucleotide in the sample sequence, T represents the transposition of the matrix; f i k-mer It can be expressed as; where N(t) is the number of Kmers of type t in the RNA sequence R.

[0029] As a preferred solution, in step S3, the data set D = {x1, x2, x3, ... x i ,…,x n}, (1≤i≤n), where n is the total number of samples, and each sample is a d-dimensional vector, the formula is as follows:

[0030]

[0031] Where x1=(v 1,1 ,v 1,2 ,…v 1,d ), then the data after dimensionality reduction can be expressed as Y=XW, where W is a d×l matrix, and the variance of the data after dimensionality reduction is:

[0032]

[0033] where λ i Is the eigenvalue of the covariance matrix Σ of X. In order to maximize the variance Var(Y) of the reduced-dimensional data Y, it is necessary to perform eigenvalue decomposition on the covariance matrix Σ and convert the obtained eigenvalue λ iArrange them from large to small, and then select the eigenvectors ω1, ω2, ...ω corresponding to the first l eigenvalues selected by the present invention. l , form the transformation matrix W, and then obtain the reduced dimension data through Y=XW;

[0034] Through the Lagrangian method, it is transformed into the corresponding Lagrangian function as follows:

[0035]

[0036] where λ i is the Lagrange multiplier, by i By finding the partial derivative and making it equal to zero, we can get the corresponding eigenvalue and eigenvector. In order to maximize the variance, we need to get the largest eigenvalue l and the corresponding eigenvector ω. i ; Dimension l is calculated by parameter variance ratio threshold Var ratio (l) to find the most suitable dimension l, the formula is as follows:

[0037]

[0038] As a preferred solution, in step S4, a maximum pooling layer is set after the CNN layer, the pooling window size is 2, and zero padding is used to keep the input and output dimensions consistent; a Dropout layer with a dropout rate of 0.2 is set after the LSTM layer to prevent overfitting.

[0039] As a preferred solution, in step S4, the model training adopts the Adam optimizer, the loss function is the cross entropy loss function, the training rounds are 50, and the batch size is 8.

[0040] As a preferred solution, the model performance is evaluated by the following indicators:

[0041] Accuracy (ACC), specificity (SP), sensitivity (SN), Matthew correlation coefficient (MCC), and AUC value; an accuracy of 93.5% was achieved on the independent test set, and the AUC value was 0.935.

[0042] (3) Beneficial effects

[0043] Compared with the prior art, the present invention provides a lipocalin classification method based on CNN-LSTM multi-feature fusion, which has the following beneficial effects:

[0044] The method of the present invention extracts local conservation, spatial distribution, and evolutionary information of protein sequences using three feature descriptors: K-mer, CKSAAP, and CC-PSSM. Combined with principal component analysis (PCA) dimensionality reduction optimization, a CNN-LSTM fusion model is constructed for classification prediction. The CNN layer extracts local features, the LSTM layer captures temporal dependencies, and ultimately outputs the classification probability through a Softmax function. Experiments show that the accuracy of the present invention on an independent test set reached 93.5%, with an AUC value of 0.935, significantly outperforming existing models. This method addresses the issues of underutilized features and high computational costs of traditional methods, providing a highly efficient tool for lipid protein function research and disease diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the research method of the present invention;

[0046] Figure 2 Generate a flow chart for CC-PSSM of the present invention;

[0047] Figure 3 Generate a flow chart for the CKSAAP of the present invention;

[0048] Figure 4 This is a schematic diagram showing the performance comparison of different feature extraction method combinations under the CNN-LSTM classifier of the present invention;

[0049] Figure 5 Schematic diagram of the distribution of the data set of the present invention;

[0050] Figure 6 This is a schematic diagram showing the importance comparison of fusion features in the present invention;

[0051] Figure 7 Schematic diagram of the PRC and ROC_AUC curves of CNN_LSTM on the training set, test set and validation set of the present invention;

[0052] Figure 8 Schematic diagram of PRC and ROC_AUC curves of different strategies of the present invention on the test set;

[0053] Figure 9 Schematic diagram of the confusion matrix of the test set and validation set of the present invention. DETAILED DESCRIPTION

[0054] In order to better understand the purpose, structure and function of the present invention, the lipocalin classification method based on CNN-LSTM multi-feature fusion of the present invention will be further described below with reference to the accompanying drawings and specific examples.

[0055] Example 1

[0056] refer to Figure 1-9The present invention provides a lipocalin classification method based on CNN-LSTM multi-feature fusion, which comprises the following steps:

[0057] S1. Build a dataset and divide it into training data, validation data, and test data;

[0058] S2, using K-mer, CKSAAP and CC-PSSM three feature descriptors to extract protein sequence features and fusion through feature integration technology;

[0059] S3. Use principal component analysis (PCA) to optimize the dimensionality reduction of the fused features, retain the features with a variance ratio threshold of 0.98, and obtain low-dimensional important features;

[0060] S4. Input the reduced-dimensional features into the CNN-LSTM fusion model, where the CNN contains 128 filters, the convolution kernel size is 1, the activation function is ReLU, and the LSTM contains 64 units. The class probability is output through a fully connected layer combined with the Softmax function.

[0061] S5. Classify and predict lipid proteins based on the output probability values.

[0062] Specifically, the dataset construction in step S1 includes:

[0063] S11, obtain lipid protein and non-lipid protein sequences from the UniProt database;

[0064] S12, using the CD-Hit tool with a threshold of 0.4 to remove homologous sequences;

[0065] S13. Exclude sequences containing non-standard amino acid characters (X, B, Z, J, O), and finally obtain training set samples and independent test set samples.

[0066] A high-quality data set is the basis for establishing a reliable and accurate model. The data set is a collection of 307 positive samples and 307 negative samples from the open source database UniProt. Then, CD-Hit was used to set the threshold to 0.4 to exclude highly homologous protein sequences to avoid affecting the results. Finally, the present invention obtained 211 lipid proteins and 211 non-lipid proteins. Then the present invention excluded sequences containing non-amino acid characters such as "X", "B", "Z", "J" and "O". Finally, the present invention obtained processed data samples including 209 lipid protein sequences and 210 non-lipid protein sequences. In addition, the present invention also established an independent data set including 42 lipid protein sequences and 51 non-lipid protein sequences as a test set to verify the effectiveness of the model.

[0067] In this study, the original dataset was preprocessed using a series of feature extraction methods. Experimental and comparative analysis confirmed that features derived from K-mer and CKSAAP, as well as CC-PSSM features extracted from PSSM, were the most practical. These features were selected because they provide a more comprehensive representation of the data and extract meaningful information from the analysis, thereby improving the accuracy and applicability of the classification model.

[0068] Example 2

[0069] Further, in step S2 of the present invention, PPSM is a standard bioinformatics method for predicting protein structure and identifying sequence similarity. It compares the similarity of the query sequence with a set of aligned sequences and calculates a PSSM score. PSSM reflects the evolutionary information of the alignment of proteins with a set of homologous sequences. The present invention uses PSI-BLAST software to align the sequences in the data set with the homologous sequences in the non-redundant (NR) database of NCBI, and obtains PSSM. PSI-BLAST software is used to align the sequences in the data set with the homologous sequences in the non-redundant (NR) database of NCBI, and obtains PSSM. PSSM is expressed as follows:

[0070]

[0071] Where m is the length of the protein sequence L, a m,n represents the probability that the amino acid residue at position m in the protein sequence will be converted into another type of amino acid residue n over time. m,n The larger the value, the greater the possibility that the amino acid at position m will produce an n-type amino acid residue during the conversion process. m,n This indicates that the transformation may be beneficial to the function or stability of the protein, or that the amino acid at that position is more conserved.

[0072] Example 3

[0073] In step S2 of the present invention, the CC-PSSM algorithm converts PSSM matrices of different sizes into vectors of the same length, and uses CC to calculate the property difference between two lagged separated residues. The calculation formula is as follows:

[0074]

[0075] Where i is the residue (i∈[1,20]), i1,i2 represent two different amino acids, z represents the distance between the two different residues, i.e., hysteresis, S i,j Indicates the PSSM score of i at point j. The size of CC-PSSM is 380×z. The detailed process is as shown in the attached manual. Figure 2shown.

[0076] Specifically, the present invention CC-PSSM is a cross-covariance method based on the Position Specific Scoring Matrix (PSSM) generated by protein sequence alignment, which is used for protein structure prediction and analysis. It uses the NR database of PSI-BLAST

[37] and NCBI

[38] to align protein sequences with local information to determine the PSSM matrix information of the protein. Using the PSSM matrix, protein sequences can be predicted based on evolution. The CC-PSSM algorithm converts PSSM matrices of different sizes into vectors of the same length and uses CC to calculate the property differences between two lagged residues.

[0077] Example 4

[0078] In step S2 of the present invention, the CKSAAP feature takes into account the situation where the amino acid pairs are separated by any number of "K" amino acids (K can be 0, 1, 2, 3, 4, or 5), and the calculation formula is as follows:

[0079]

[0080] It is used to calculate the relative frequency of a specific amino acid pair (xy) when considering "K" amino acids apart; where: Q x,y : relative frequency of a specific amino acid pair; V x,y : the number of specific amino acid pairs; N: the total number of amino acids; K: the number of intervals between amino acids.

[0081] Example 5

[0082] In step S2 of the present invention, the principle of K-mer encoding is to calculate the frequency of k nucleotides in the protein sequence. The protein sequence R can be converted into a vector with 4k dimensions by using the K-mer frequency. The formula is as follows:

[0083]

[0084] where f i k-mer is the normalized frequency of occurrence of the i-th Kmers nucleotide in the sample sequence, T represents the transposition of the matrix; f i k-mer It can be expressed as: Where N(t) is the number of kmers of type t in the RNA sequence R. A kmer is a sequence segment of length k, obtained by cutting a portion of the original sequence before encoding. When k is an odd number, that is, k = n (where n is an odd number), it is called an n-mer. For example, the original sequence before kmer coding is CTAGAGCTGG. After kmer coding (k = 1, 3, 5, 7), the resulting sequences are shown in Table 1 below.

[0085] sequence GKWLSLLKHILK K＝1 GKWLSLLKHILK K＝3 GKWKWLWLSLSSLLLLLLKLKHKHIHILILK K＝5 GKWLSKWLSLWLSLLLLSLLKSLLKHLLKHILKHILKHILK K＝7 GKWLSLLKWLSLLKWLSLLKHLSLLKHISLLKHILLLKHILK

[0086] Table 1

[0087] Example 6

[0088] In the principal component analysis in step S3 of the present invention, the importance of the feature is expressed by the variance. Assume that the present invention has a data set using the data set D = {x1, x2, x3, ... x i ,…,x n}, (1≤i≤n), where n is the total number of samples, and each sample is a d-dimensional vector, the formula is as follows:

[0089]

[0090] Where x1=(v 1,1 ,v 1,2 ,…v 1,d ), then the data after dimensionality reduction can be expressed as Y=XW, where W is a d×l matrix, and the variance of the data after dimensionality reduction is:

[0091]

[0092] where λ i Is the eigenvalue of the covariance matrix Σ of X. In order to maximize the variance Var(Y) of the reduced-dimensional data Y, it is necessary to perform eigenvalue decomposition on the covariance matrix Σ and convert the obtained eigenvalue λ i Arrange them from large to small, and then select the eigenvectors ω1, ω2, ...ω corresponding to the first l eigenvalues selected by the present invention. l , form the transformation matrix W, and then obtain the reduced dimension data through Y=XW;

[0093] Through the Lagrangian method, it is transformed into the corresponding Lagrangian function as follows:

[0094]

[0095] where λ i is the Lagrange multiplier, by iBy finding the partial derivative and making it equal to zero, we can get the corresponding eigenvalue and eigenvector. In order to maximize the variance, we need to get the largest eigenvalue l and the corresponding eigenvector ω. i ; Dimension l is calculated by parameter variance ratio threshold Var ratio (l) to find the most suitable dimension l, the formula is as follows:

[0096]

[0097] Furthermore, in step S4 of the present invention, a maximum pooling layer is set after the CNN layer, with a pooling window size of 2 and zero padding to maintain consistent input and output dimensions. A Dropout layer with a dropout rate of 0.2 is set after the LSTM layer to prevent overfitting. In step S4, the model is trained using the Adam optimizer, the cross-entropy loss function, 50 training rounds, and a batch size of 8. Model performance is evaluated using the following metrics: accuracy (ACC), specificity (SP), sensitivity (SN), Matthew correlation coefficient (MCC), and AUC value; on an independent test set, an accuracy of 93.5% is achieved, with an AUC value of 0.935.

[0098] Example 7

[0099] In the present invention, based on the fusion classifier of CNN-LSTM, classification is a form of supervised learning and plays an important role in the decision-making process. In this work, predicting lipid protein is a typical binary classification problem. Traditional classification methods can not predict lipid protein well, and the classification method based on deep neural network can make full use of the potential features of data, so as to achieve satisfactory classification effect. Therefore, in order to better explore the prediction model and analyze the features, the present invention selects CNN-LSTM classifier to establish a model for identifying lipid protein. In the method of the present invention, the integrated features are used as the input of CNN, and the features are further extracted by convolution layer and maximum pooling layer at first, and then the dependency information between the data is captured by long short-term memory network, so as to realize the classification of lipid protein.

[0100] Specifically, about Convolutional Neural Networks (CNN)

[0101] In the method of the present invention, the integrated features after dimensionality reduction are first input into the one-dimensional convolution layer x c In , the implicit correlation in the input data is extracted through the convolution kernel matrix, where the expression of the convolution layer is as follows:

[0102]

[0103] The number of filters in the convolutional layer is 128, and the size is 1. Each filter can capture different features of the input data. Ultimately, the present invention will obtain 128 feature maps. The activation function of the present invention is ReLU, which is defined as follows:

[0104] f(x)=max(0,x)

[0105] The convolution is followed by the maximum pooling layer. Since the amount of lipid protein data is not large, the present invention performs zero padding during pooling to ensure that the spatial dimensions of the input and output features remain unchanged. Here, the pooling window size of the present invention is 2, which not only retains useful information but also reduces the possibility of overfitting.

[0106] Long Short-Term Memory (LSTM)

[0107] In order to capture the dependency information in the output data of the convolutional neural network, the present invention introduces LSTM. LSTM is designed by changing the network module (cell) of RNN, and four nonlinear neural network layers are designed inside it. The state s of the cell is added, and the hidden state h and the cell state s are input to the next cell. In this cell, the present invention needs to use the forget gate f (t) To decide how to discard information, the formula is as follows:

[0108] f (t) =σ(b 1 +U 1 x (t) +W 1 h (t-1) )

[0109] where b 1 、U 1 and W 1 Represent the bias, input weight and the recurrent weight of the forget gate respectively. Then through the external input gate g (t) To determine the updated information. The formula is as follows:

[0110] g (t) =σ(b 2 +U 2 x (t) +W 2 h (t-1) )

[0111] The corresponding output gate q (t) =σ(b 3 +U 3 x (t) +W 3 h (t-1) ) determines the final output information, and the final cell state s (t) and output h(t) See the above formula.

[0112] s (t) =f (t) s (t-1) +g (t) tanh(b+Ux (t) +Wh (t-1) )

[0113] h (t) =tanh(s (t) )q (t)

[0114] The model of the present invention uses a cross entropy loss function to minimize the training error, and the present invention uses the cross entropy loss function as follows:

[0115]

[0116] where y i is the label value of each sample, Is the model's predicted value for the sample, and its error Loss is used to measure the quality of the model parameters. In order to reduce the possibility of overfitting, the present invention adds a dropout rate with a value of 0.2.

[0117] At the same time, to demonstrate the superiority of the model established by the present invention, seven classification methods were compared: long short-term memory network (LSTM), support vector machine (SVM), recurrent neural network (RNN), convolutional neural network (CNN), deep learning network (DNN), K-nearest neighbor algorithm (KNN), and random forest (RF). Detailed results are shown in Table 4. The CNN-LSTM deep learning classification model combines a convolutional neural network and a long short-term memory network. The convolutional neural network extracts local features of the input data and obtains feature maps. The long short-term memory network then further learns and captures the temporal characteristics of the input data. This allows not only to extract information from the local features, but also to integrate this information through the LSTM layer and make time series-based predictions.

[0118] Evaluation indicators

[0119] In the present invention, accuracy (ACC), sensitivity (SN), specificity (SP), Matthew's correlation coefficient (MCC), F1 are used as indicators to measure model performance [39-42], as well as the area under the receiver operating characteristic curve (AUC). Using the given number of true positive samples (TP), true negative samples (TN), false positive sample size (FP), and false negative sample size (FN), it is expressed by the following formula:

[0120]

[0121] A true positive is a lipid protein, while a false positive indicates a non-lipid protein that is classified as a lipid protein. On the other hand, a true negative represents a non-lipid protein, while a false negative indicates a lipid protein that is classified as a non-lipid protein. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) are used to measure the effectiveness of the prediction model. A perfect classifier has an AUC of 1, while random behavior has an AUC of 0.5.

[0122] Results and discussion of the experiments of this invention

[0123] Performance of individual feature extraction methods

[0124] In this study, we preprocessed the original dataset using a series of feature extraction methods. We conducted a preliminary selection of these features using an experimental CNN-LSTM classifier. The comparative results are shown in Table 2 below. Through experimental comparison, we determined that features 1, 2, and 3, obtained using the K-mer, CC-PSS, and CKSAAP feature extraction methods, were the most practical for our research.

[0125] Model ACC F1 AUC Sp Sn MCC K-mer CNN_LSTM 0.845 0.857 0.941 0.800 0.886 0.690 CC-PSSM CNN_LSTM 0.964 0.966 0.988 0.950 0.977 0.929 CKSAAP CNN_LSTM 0.905 0.895 0.964 0.977 0.829 0.817 AAC CNN_LSTM 0.750 0.727 0.849 0.854 0.651 0.514 CTDT CNN_LSTM 0.714 0.76 0.850 0.512 0.927 0.479 GAAC CNN_LSTM 0.679 0.703 0.724 0.556 0.821 0.386

[0126] Table 2

[0127] Through experimental analysis, it was found that the model of the present invention achieved ACCs of 0.929, 0.964, and 0.857 when selecting the features K-mer, CC-PSSM, and CKSAAP, respectively. Combined with other indicators, a comprehensive comparative analysis showed that the model performance indicators brought by these three feature data were superior to those of other features.

[0128] Performance of hybrid feature representation methods

[0129] In order to explore the influence of these three features on the model of the present invention and further improve the accuracy of the model of the present invention in lipid protein classification, different feature combinations of the features obtained by the three feature extraction methods of K-mer, CC-PSSM and CKSAAP were performed on the CNN-LSTM model and compared respectively. The results are shown in Figure 2. Figure 3 The present invention found that the model performance of the fusion feature is better than the model performance brought by a single feature. Figure 4 The results show that the model performance of the three feature combinations of K-mer, CC-PSSM and CKSAAP is the best.

[0130] Optimizer selection

[0131] Based on the established model, the present invention selected the Adam optimizer to optimize the training process and compared the performance of other optimizers. The results are shown in Table 3 below. The present invention found that among all the optimizers, Adam had the highest accuracy of 96.4%. The accuracy of Adamax, Experimentaladagrad and RMSprop was lower than that of the other optimizers, although there were other indicators that were slightly higher than Adam and Nadam. However, the impact on the recognition results of the present invention was very small. Nadam showed a relatively balanced performance among all the optimizers. Its performance was relatively stable, balanced in various indicators, and had no obvious shortcomings. However, its overall performance was lower than that of Adam, and through experiments, the present invention found that it had strong performance in classification tasks and could effectively handle most situations.

[0132] Model ACC F1 AUC Sp Sn MCC adamax 0.929 0.927 0.995 0.952 0.905 0.858 nadam 0.952 0.952 0.993 0.952 0.952 0.905 adam 0.964 0.964 0.993 0.976 0.952 0.929 rmsprop 0.917 0.911 0.98 0.976 0.857 0.839 Experimentaladagrad 0.857 0.842 0.969 0.952 0.762 0.728

[0133] Table 3

[0134] Overall, Adam is the most suitable optimizer for lipid protein binary classification tasks because it performs well and well in all key indicators.

[0135] Dimensionality reduction

[0136] Through the three feature extraction methods of K-mer, CKSAAP and CC-PSSM, the dimensions of the feature data obtained are 400, 760 and 2400 respectively. After the feature integration technology, a 3560-dimensional feature vector can be obtained. The data obtained by these feature extraction technologies need to delete some feature information that is useless for the recognition of the present invention, and remove the data that have the same impact on model recognition, so as to reduce the amount of calculation of the present invention and improve the efficiency of the model. Through the analysis of experimental results, PCA can retain the information that is useful for the recognition of lipid proteins of the present invention to the greatest extent after the dimensionality reduction of the data of the present invention. The model can obtain the best accuracy when the 3560 dimension is reduced to 180 dimensions. Therefore, the present invention selects the comprehensive variable as 180. The present invention uses principal component analysis (PCA) to visualize the balanced data set, such as Figure 5 After analyzing the fused features, each of them has a specific meaning in its dimension, such as Figure 6 The present invention calculates the feature importance of the fusion feature and Figure 9 They are highlighted in .

[0137] Parameter Adjustment

[0138] Parameters that affect model performance must be adjusted to achieve optimal performance. In this study, the present invention focused on adjusting six parameters in the convolutional neural network. After experimental selection, the present invention finally determined the optimal parameters. The best effect is achieved when the number of filters in the convolution layer is 128 and the size of the convolution kernel is 1. The activation function of the present invention selects ReLU, and the maximum pooling operation is performed on a window with a pooling window size of 2, and the input and output information sizes are kept consistent through the padding operation. In order to reduce model overfitting, the dropout rate of the Dropout layer of the present invention is 0.2. The parameter indicators in the long short-term memory network are as follows: the number of units in the LSTM layer is 64, indicating the output dimension of the layer, the number of units in the output layer is 2, and softmax is selected as the activation function. Since the model of the present invention is used to process classification tasks, the loss function of the present invention selects the classification cross entropy loss function. Finally, through comparative experimental analysis, the present invention finally chooses to use the Adam optimization algorithm to minimize the loss function, and the classification effect at this time reaches the best. The parameter of principal component analysis is 0.98.

[0139] Comparison of different classifiers

[0140] The present invention selects the optimal feature set used in the above-mentioned model. In this section, the present invention compares the CNN-LSTM-based classification model of the present invention with other different classifiers. The fused optimal features are then input into other classifier methods, including CNN, LSTM, RNN, RF, SVM, and KNN classifiers. The parameters of CNN are set to 128 filters, the size of the convolution kernel is 1, the size of the pooling window is 2, and the dropout rate of the Dropout layer is 0.2. The parameters of LSTM are set to 128 units and the dropout rate is 0.2. The parameters of RNN are set to 128 units and the dropout rate is 0.2. The parameter n of RF is 50, the regularization parameter c of SVM is 5, the tolerance parameter is 0.01, the parameter k of KNN is 2, the maximum number of leaf node samples is 20, and the distance parameter p is 10. Regarding the traditional machine learning models RF, SVM, and KNN, the present invention divides the data set into 80% training set and 20% validation set. For the deep learning models CNN, RNN, and LSTM, the model training rounds are 50, with 8 training samples each time, and the test set is used as the validation set to evaluate the model performance at the end of each round. The model performance results are shown in Table 4 below, and the PR and ROC curves are plotted. Figure 7As shown in the figure, the present invention notes that the model combining the CNN-LSTM classifier has higher accuracy than other classifier models, with an ACC 2.4%-9.5% higher and an AUC 0.6%-5.1% higher than other classifier models. Clearly, the present invention's model, Lipo-CNN_LSTM, outperforms existing traditional classifiers and single deep models in most evaluation metrics, indicating that the CNN-LSTM classifier is more suitable for lipid protein prediction.

[0141] Model ACC F1 AUC Sp Sn MCC Val_CNN_LSTM 0.976 0.976 0.997 0.976 0.976 0.952 CNN 0.952 0.951 0.991 0.976 0.929 0.906 LSTM 0.952 0.951 0.99 0.976 0.929 0.906 RNN 0.94 0.94 0.982 0.952 0.929 0.881 RF 0.952 0.951 0.989 0.976 0.929 0.906 Support Vector Machine 0.905 0.907 0.969 0.881 0.929 0.81 KNN 0.881 0.868 0.946 0.976 0.786 0.776

[0142] Table 4

[0143] Comparison with Existing Models on Independent Datasets: This section compares the proposed model with other existing methods on independent datasets to examine their efficiency and performance. The results on independent datasets demonstrate that the proposed model achieves 3.6%-7.77% higher accuracy than existing models. Table 5 shows a comparison of the proposed model with existing models.

[0144] Methon Acc MCC Sn Sp AUC reference Lipocalin-Pred 85.73 0.776 88.41 90.11 0.922

[18] Lipo-RF 89.90 0.868 91.73 92.66 0.956

[22] Lipo-CNN_LSTM 93.50 0.872 88.10 98.00 0.935 Our Study

[0145] Table 5

[0146] in conclusion

[0147] Lipoproteins are responsible for transporting small, hydrophobic molecules such as steroids, retinoids, and lipids. They share regions of sequence homology and a common tertiary structure [49, 50]. They have been applied in several fields, including stress response, homeostasis, candidate markers of renal function, and allergic infections. Therefore, it is necessary to develop models to identify lipoproteins. In this work, we developed a deep CNN-LSTM-based model to distinguish lipoproteins from non-lipoproteins. The dataset was converted into feature vectors using three feature extraction methods: Kmer, CKSAAP, and CC-PSSM. Each feature descriptor was then evaluated using a CNN-LSTM-based classifier. To further improve prediction accuracy, we implemented different fusion strategies for these three features and used principal component analysis to select the optimal feature set. Finally, the effectiveness of our model for lipoprotein classification was verified using a test set. Table 4 shows the results of the multi-feature fusion method on each classifier. Our results show that in the proposed model, protein sequences are characterized by the three descriptors Kmer, CKSAAP, and CC-PSSM, which improves the accuracy of the model. Next, feature fusion and feature dimensionality reduction were performed to identify the optimal feature set that yielded the highest accuracy. Based on this optimal feature set, the CNN-LSTM-based classifier achieved the best prediction results. The accuracy achieved on an independent dataset reached 93.5%, surpassing the performance of all previous models. However, this study has certain limitations. The database of the present invention contains relatively small amounts of lipid protein data. When faced with large data sets, the accuracy of the present invention's model may decline, requiring the present invention to further expand its dataset and identify more representative lipid protein information.

[0148] The work of this invention will provide assistance for subsequent research. The subsequent research will focus on creating a user-friendly web server for lipid protein prediction and will adopt additional feature selection methods and algorithms to provide valuable information for the identification of lipid proteins and the diagnosis of related diseases, while also providing assistance for the research of related drugs.

[0149] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A lipocalin classification method based on CNN-LSTM multi-feature fusion, characterized by: The method comprises the following steps: S1. Build a dataset and divide it into training data, validation data, and test data; S2, using K-mer, CKSAAP and CC-PSSM three feature descriptors to extract protein sequence features and fusion through feature integration technology; S3. Use principal component analysis (PCA) to optimize the dimensionality reduction of the fused features, retain the features with a variance ratio threshold of 0.98, and obtain low-dimensional important features; S4. Input the reduced-dimensional features into the CNN-LSTM fusion model, where the CNN contains 128 filters, the convolution kernel size is 1, the activation function is ReLU, and the LSTM contains 64 units. The class probability is output through a fully connected layer combined with the Softmax function. S5. Classify and predict lipid proteins based on the output probability values.

2. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 1, characterized in that: The data set construction in step S1 includes: S11, obtain lipid protein and non-lipid protein sequences from the UniProt database; S12, using the CD-Hit tool with a threshold of 0.4 to remove homologous sequences; S13. Exclude sequences containing non-standard amino acid characters (X, B, Z, J, O), and finally obtain training set samples and independent test set samples.

3. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 2, characterized in that: In step S2, the sequences in the dataset were aligned with the homologous sequences in the NCBI non-redundant (NR) database using PSI-BLAST software, and PSSM was obtained. The PSSM is represented as follows: Where m is the length of the protein sequence L, a m,n Represents the amino acid residue at position m in the protein sequence.

4. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 3, characterized in that In step S2, the CC-PSSM algorithm converts PSSM matrices of different sizes into vectors of the same length, and uses CC to calculate the property difference between two lagged separated residues. The calculation formula is as follows: Where i is the residue (i∈[1,20]), i1,i2 represent two different amino acids, z represents the distance between the two different residues, i.e., hysteresis, S i,j represents the PSSM score of i at point j, and the size of CC-PSSM is 380×z.

5. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 1, characterized in that In step S2, the CKSAAP feature considers the situation where the amino acid pairs are separated by any number of "K" amino acids (K can be 0, 1, 2, 3, 4, or 5), and the calculation formula is as follows: It is used to calculate the relative frequency of a specific amino acid pair (xy) when considering "K" amino acids apart; where: Q x,y : relative frequency of a specific amino acid pair; V x,y : the number of specific amino acid pairs; N: the total number of amino acids; K: the number of intervals between amino acids.

6. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 1, characterized in that: In step S2, the principle of K-mer encoding is to calculate the frequency of k nucleotides in the protein sequence. The protein sequence R can be converted into a vector with 4k dimensions by using the K-mer frequency. The formula is as follows: where f i k-mer is the normalized frequency of occurrence of the i-th Kmers nucleotide in the sample sequence, T represents the matrix change; f i k-mer It can be expressed as; where N(t) is the number of Kmers of type t in the RNA sequence R.

7. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 1, characterized in that In step S3, the dataset D={x1,x2,x3,…x i ,…,x n }, (1≤i≤n), where n is the total number of samples, and each sample is a d-dimensional vector, the formula is as follows: Where x1=(v 1,1 ,v 1,2 ,…v 1,d ), then the data after dimensionality reduction can be expressed as Y=XW, where W is a d×l matrix, and the variance of the data after dimensionality reduction is: where λ i Is the eigenvalue of the covariance matrix Σ of X. In order to maximize the variance Var(Y) of the reduced-dimensional data Y, it is necessary to perform eigenvalue decomposition on the covariance matrix Σ and convert the obtained eigenvalue λ i Arrange them from large to small, and then select the eigenvectors ω1, ω2, ...ω corresponding to the first l eigenvalues selected by the present invention. l , form the transformation matrix W, and then obtain the reduced dimension data through Y=XW; Through the Lagrangian method, it is transformed into the corresponding Lagrangian function as follows: where λ i is the Lagrange multiplier, by i By finding the partial derivative and making it equal to zero, we can get the corresponding eigenvalue and eigenvector. In order to maximize the variance, we need to get the largest eigenvalue l and the corresponding eigenvector ω. i ; Dimension l is calculated by parameter variance ratio threshold Var ratio (l) to find the most suitable dimension l, the formula is as follows:

8. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 1, characterized in that In step S4, a maximum pooling layer is set after the CNN layer, the pooling window size is 2, and zero padding is used to keep the input and output dimensions consistent; a Dropout layer with a dropout rate of 0.2 is set after the LSTM layer to prevent overfitting.

9. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 8, characterized in that In step S4, the model training adopts Adam optimizer, the loss function is cross entropy loss function, the training rounds are 50, and the batch size is 8.

10. The lipocalin classification method based on CNN-LSTM multi-feature fusion according to claim 9, characterized in that ,The model performance is evaluated by the following metrics: Accuracy (ACC), specificity (SP), sensitivity (SN), Matthew correlation coefficient (MCC), and AUC value; an accuracy of 93.5% was achieved on the independent test set, and the AUC value was 0.935.

Images