Multifunctional therapeutic peptide prediction method based on deep spatio-temporal neural network

Abstract

The application provides a multifunctional therapeutic peptide prediction method based on a deep spatio-temporal neural network, which comprises the following steps: S1, multi-angle initialization embedding is performed on a multifunctional therapeutic peptide sequence to sufficiently obtain feature expression information, so as to obtain initialization features; S2, the initialization features are input into a multi-scale space feature extractor to extract spatial information of the peptide sequence, so as to obtain a peptide sequence feature matrix containing the spatial information; S3, the peptide sequence feature matrix containing the spatial information is sequentially subjected to time feature extraction through a bidirectional long short-term memory network and an AFT module; so that the time feature information of the peptide sequence is sufficiently obtained without increasing the complexity of the model; and S4, high-dimensional time feature information is mapped to a sample label space through a full connection layer, and finally, a classification result is obtained through an activation layer conversion to a probability value. The application introduces fusion coding of physical and chemical features and evolutionary information as initialization embedding of the peptide sequence, so that the inherent sequence characteristics can be better learned.

Description

Technical Field

[0001] This invention relates to the field of biomedical big data, and in particular to a multifunctional therapeutic peptide prediction method based on deep spatiotemporal neural networks. Background Technology

[0002] Deep learning has been successfully applied to the classification of multifunctional therapeutic peptides, achieving significant performance improvements. However, existing deep learning-based multifunctional therapeutic peptide classification methods have some drawbacks: (1) The amino acid sequence in biological sequences contains unique biological significance, and using only numerical embedding representation may ignore its biological information, resulting in insufficient expression of amino acid features and thus affecting prediction performance; (2) As the deep learning network deepens, the model complexity also increases, causing the model to lose some key information during feature extraction and optimization, thus affecting prediction performance; (3) There is a serious imbalance in the class samples in the multifunctional therapeutic peptide dataset, affecting the classification performance of multifunctional therapeutic peptides. Summary of the Invention

[0003] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a multifunctional therapeutic peptide prediction method based on deep spatiotemporal neural networks.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network, comprising the following steps:

[0005] S1, perform multi-angle initialization embedding on the multifunctional therapeutic peptide sequence to fully obtain feature expression information, thereby obtaining the initialization features;

[0006] S2, input the initial features into the multi-scale spatial feature extractor to extract the spatial information of the peptide sequence, thereby obtaining the peptide sequence feature matrix containing spatial information;

[0007] S3, the peptide sequence feature matrix containing spatial information is sequentially passed through a bidirectional long short-term memory network and an AFT module to extract temporal features; thus, the temporal feature information of the peptide sequence is fully obtained without increasing the model complexity;

[0008] S4, through a fully connected layer, maps high-dimensional temporal feature information to the sample label space, and finally through an activation layer to convert it into probability values ​​to obtain the classification result. The activation layer uses the sigmoid activation function. For general peptide sequence multi-classification tasks, the activation layer uses the softmax function.

[0009] Furthermore, the multi-angle initialization embedding is implemented using AAIs and BLOSUM62 descriptors, including the following steps:

[0010] S1-1, fix the length T of each peptide sequence to a set value. For peptide sequences that are shorter than the set value, the length of the set value is achieved by adding 0 to the end of the input sequence.

[0011] S1-2 are encoded using AAIs and BLOSUM62 respectively to obtain feature maps based on AAIs and feature maps based on BLOSUM62.

[0012] S1-3, the feature maps based on AAIs and the feature maps based on BLOSUM62 are concatenated to obtain the initial embedding of the peptide sequence.

[0013] AAIs represents encoding methods based on the physicochemical properties of amino acids, while BLOSUM represents a scoring method based on the comparison of the mutation probability between amino acids throughout the evolutionary process.

[0014] Furthermore, the multi-scale spatial feature extractor uses three one-dimensional convolutions with different kernel sizes k to extract the spatial features of the peptide sequence, where k is 2, 3, and 8, the number of channels is 64, and the padding is 0.

[0015] Furthermore, the multi-scale spatial feature extractor is based on residual connections, and the spatial information of the peptide sequence obtained therefrom is represented as follows:

[0016]

[0017] Where vector f MSC Spatial information of the peptide sequence;

[0018] X in It is an input vector encoded by AAIs and BLOSUM62;

[0019] O MSC (X in () represents the output of multi-scale convolution;

[0020] It is connected in series.

[0021] By introducing residual connections into the multi-scale spatial feature extractor, it is possible to extract the spatial information of peptide sequences by fusing multi-scale representations and original information without increasing the complexity of the model.

[0022] Furthermore, the peptide sequences containing spatial information are sequentially passed through a bidirectional long short-term memory network and an AFT module for temporal feature extraction, specifically including the following steps:

[0023] S3-1, the spatially extracted feature matrix is ​​fed into a bidirectional long short-term memory network to capture long-distance dependencies;

[0024] S3-2, The extracted feature map is input into the AFT module, and the AFT module outputs the temporal feature information of the peptide sequence; the transformation process of the feature map in the AFT module includes:

[0025] The AFT module first linearly transforms the data into Q, K, and V, which are obtained by transforming the same features output by BILSTM through a linear layer. Q represents the query vector, V represents the vector of the queried information, and K represents the relevance vector of the queried information to other information. Then, K is combined with the learned pairwise positional bias, weighted and averaged with V, and then multiplied element-wise with Q. Finally, a T×200-dimensional feature map is generated through the sigmoid activation function, which fully captures the temporal feature information of the peptide sequence without increasing the model complexity.

[0026] Furthermore, this also includes introducing a loss function during training:

[0027]

[0028] Where p represents the number of positive samples;

[0029] n represents the number of negative samples;

[0030] p i This represents the i-th positive sample;

[0031] p j This represents the j-th negative sample;

[0032] α i This represents the weight of the i-th positive sample;

[0033] α j This represents the weight of the j-th negative sample;

[0034] γ is an adjustable parameter used to adjust the weights of difficult samples. When γ is high, the minority class will receive a larger weight than the majority class, and vice versa. α i α j Both γ and γ are hyperparameters. These are used to weight and fuse the extracted features, resulting in a more accurate feature representation, a better understanding of sequence information, and improved classification accuracy and stability.

[0035] By introducing the loss function of formula (6), the problem of imbalance in the number of categories can be alleviated, making the classification results more accurate.

[0036] In summary, due to the adoption of the above technical solution, the present invention has the following advantages:

[0037] (1) The fusion encoding of physicochemical features and evolutionary information is introduced as the initial embedding of peptide sequences, which changes the single information input represented only by numerical values. The special initial feature representation can provide sufficient biological information, enabling the model to better learn its intrinsic sequence characteristics.

[0038] (2) A model of spatiotemporal neural network based on AFT module is proposed. As an important part of spatiotemporal network, AFT module uses element-wise multiplication instead of dot product operation. Without increasing the complexity of the model, it can maintain the global interaction between Q and V like multi-head attention, fully obtain the global interaction information of peptide sequence, and improve the performance of classifier.

[0039] (3) An effective loss function strategy is introduced to optimize the class imbalance problem of the model. This strategy aims to increase the model's attention to difficult-to-classify peptide sequences. The algorithm reduces the weight of easily classified samples by reshaping the loss and focuses the training on difficult-to-classify samples, thereby improving training efficiency and the model's classification performance.

[0040] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0042] Figure 1 This is a schematic diagram of the overall structure of the TP-STNN model proposed in this invention.

[0043] Figure 2 This is a schematic diagram of the multi-scale spatial feature extraction of the present invention.

[0044] Figure 3 This is a diagram of the temporal feature extraction architecture of the present invention.

[0045] Figure 4 This represents the different metric results of different initialization feature methods on the test set. Figure 4 (a) Comparison of Precision index results Figure 4 (b) Comparison of Coverage index results Figure 4 (c) Comparison of Accuracy index results. Figure 4 (d) Comparison of Absolute true index results Figure 4 (e) Comparison of Absolute false index results.

[0046] Figure 5 This represents the box plot results of different models on the test set. Figure 5(a) Comparison of Precision index results Figure 5 (b) Comparison of Coverage index results Figure 5 (c) Comparison of Accuracy index results. Figure 5 (d) Comparison of Absolute true index results.

[0047] Figure 6 This is a schematic diagram illustrating how the present invention visualizes samples in the test dataset using t-SNE. Figure 6 (a) is a visualization of t-SNE during the training period of epoch=1. Figure 6 (b) is a visualization of t-SNE during the training period of epoch=12. Figure 6 (c) is a visualization of t-SNE during the training period of epoch=24. Figure 6 (d) is a visualization of t-SNE during the training period of epoch=36. Figure 6 (e) is a visualization of t-SNE during the training period of epoch=48. Figure 6 (f) is a visualization of t-SNE during the training period of epoch=60. Detailed Implementation

[0048] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0049] To effectively extract sequence features of therapeutic peptides, we propose a spatiotemporal neural network model, TP-STNN. Figure 1This paper describes the model overview. Part I covers data collection and organization, including data preprocessing and dividing the preprocessed data into training and testing sets. The preprocessing specifically includes: 1. Removing peptides containing non-standard amino acid sequences. 2. Removing peptides with sequence lengths less than 5 bp or greater than 50 bp, as long peptides generally have high toxicity and low stability, while very short peptide sequences have poor activity. 3. Peptides with too few classes cannot effectively train the deep neural network; therefore, peptides with fewer than 40 classes are removed. After these preprocessing steps, a benchmark dataset with 9841 multifunctional therapeutic peptides is obtained. Here, bp is a unit of biological sequence length. Part II details the descriptor extraction and feature matrix filling. Part III presents the architecture of the prediction model. Part IV describes the model tuning process, including optimizing the model's class imbalance using a loss function and testing the model with the test set. Based on a spatiotemporal neural network model, a multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network is proposed. This method mainly consists of multi-angle initialization embedding in Part II, multi-scale spatial feature extraction and temporal feature extraction in Part III, and the functional prediction part in Part IV. The entire process is represented in Formula 1. We first let X... in Let M and represent the input and output features of our model. The feature representation of the model is formulated as follows:

[0050] M = g FL ((f TFE (f MSC (X in )) (1)

[0051] Where X in It is a fusion of embedded features that possess physicochemical characteristics and evolutionary information.

[0052] f MSC (·) represents the transformation function of a multi-scale convolutional neural network (MSC) incorporating residual connections.

[0053] f TFE (·) represents the transformation function for temporal feature extraction based on AFT.

[0054] g FL (·) represents the transition function of a fully connected network with activation units, which integrates global and local features of the sequence.

[0055] 1. Multi-angle initialization embedding (MIE)

[0056] Therapeutic peptide sequences are known to be one-dimensional sequences of varying lengths, making them difficult to handle in machine learning. Discriminative initial feature representations play a crucial role in building robust predictors. Simply considering the sequence encoding of a peptide may not provide sufficient information to learn its intrinsic sequence characteristics. Traditional encoding methods convert the sequence into a fixed-size numerical representation, called a feature or descriptor, before training, such as pseudo-amino acid composition, pseudo-reduced amino acid composition, and compositional transition distribution encoding. Over time, various encoding methods have been developed. Based on the nature of the encoding, these methods can be mainly divided into sequence-based and structure-based methods. For peptide function prediction, our goal is to develop interpretable models. Therefore, sequence encoding methods that preserve residue position information are desirable. To this end, we convert the amino acids in the sequence into corresponding physicochemical and evolutionary information representations without losing residue position information:

[0057] AAIs represent a 31-dimensional vector (composed of 0 / 1 elements) that replaces each amino acid in a peptide sequence with its physicochemical properties. For each amino acid, a unique 31-dimensional vector is formed by concatenating a 10-bit vector and a 21-bit vector. The elements of the 10-bit vector describe the membership of the specific amino acid in 10 overlapping groups based on its physicochemical properties. The elements of the 21-bit vector are determined by the membership of the specific amino acid in 3 groups (7*3=21 groups) based on 7 physicochemical properties, including polarity, normalized van der Waals volume, hydrophobicity, secondary structure, solvent accessibility, charge, and polarizability. Therefore, each peptide is encoded as a 31×T matrix.

[0058] BLOSUM represents the probability of mutation between amino acids throughout evolution. It is a symmetric 20×20 matrix, where each of the 20 rows is a vector containing 20 elements describing the evolutionary relationship of a specific amino acid to other amino acids. It has been widely used for protein or peptide prediction. The mathematical representation of each element K(i,j) in the matrix is ​​as follows:

[0059]

[0060] Where, p ij f represents the probability that amino acids 'i' and 'j' are homologous sequences. i f is the probability that amino acid 'i' appears in any protein sequence. j λ represents the probability that amino acid 'j' appears in any protein sequence, and λ is a scaling factor used to round the elements in the matrix to convenient integer values. Among different BLOSUM variants, we chose the most widely used BLOSUM62 matrix to compare sequences with 62% similarity.

[0061] Therefore, the multi-angle initialization embedding process is as follows: First, the length T of each peptide sequence is fixed at 50 to ensure that the size of the output matrix is ​​constant. In practical applications, the length T is set empirically, generally based on the longest input peptide sequence. However, if the peptide sequence is too long, a relatively reasonable T value is set, and the excessively long peptide sequence is divided into multiple segments for processing. For peptide sequences < 50, zeros are padded to the end of the input sequence to achieve a length of 50. Then, encoding is performed using AAIs and BLOSUM62 respectively to obtain feature maps based on AAIs and BLOSUM62. Finally, the feature maps generated by AAIs and BLOSUM62 are concatenated and the initial embedding of the peptide sequence is represented in 51×T dimension.

[0062] 2. Multi-scale spatial feature extraction (MSC) based on residual connections

[0063] To effectively mitigate the gradient vanishing problem as the number of network layers increases, enabling the network to learn deeper features, this work employs a multi-scale convolutional neural algorithm to extract features from peptide sequences of varying lengths for complex peptide initialization feature patterns. Simultaneously, residual connections are introduced to fuse multi-scale representations and original information to extract spatial information from the peptide sequence without increasing model complexity. Gradients not only propagate backward layer by layer but also flow back to the starting layer via identity shortcuts (a form of residual network). The representation vector f obtained from the residual mapping... MSC Formulated as:

[0064]

[0065] Where vector f MSC Spatial information of the peptide sequence; X in It is an input vector encoded by AAIs and BLOSUM62, O MSC (X in () represents the output of multi-scale convolution. It is connected in series.

[0066] like Figure 2 As shown, for N feature matrices of size 51×T, we use three one-dimensional convolutions with different kernel sizes k to extract complex patterns of biopeptides, where k is 2, 3, and 8, the number of channels is 64, and the padding is 0. The model learns and obtains patterns from X through convolution. in To O MSC (X in The feature map size is 192×T, and then the input X of the residual block is concatenated. inFinally, spatial feature extraction generates a 243×T feature map containing multi-scale representations and original information. In practical applications, other multi-scale representations are also possible; the value of k is determined by observing experimental results and is usually not very large. For this invention, using three different kernel sizes k, with k values ​​of 2, 3, and 8 respectively, yields the best results.

[0067] 3. Temporal Feature Extraction (TFE)

[0068] Multi-head self-attention (MHSA) performs well in capturing the intrinsic correlations of features, but the dot product operation is computationally and storage-intensive. AFT is a variant of dot product attention that eliminates the need for it. AFT does not use matrix multiplication while maintaining the global interaction between Q and V, similar to MHSA. Therefore, we propose... Figure 3 The temporal feature extraction method shown, based on AFT and a bidirectional long short-term memory (BILSTM) network, reduces dependence on external information, adaptively focuses on important elements of the data, and learns global interaction information. First, the spatially extracted feature matrix is ​​fed into the BILSTM network to capture long-distance dependencies. This process generates a T×200 feature map. Then, the extracted feature map is input into the AFT module. The AFT module first linearly transforms it into Q, K, and V, which are obtained by linearly transforming the same features output by BILSTM and have the same feature size as the BILSTM output, forming a T×200 vector matrix. Q represents the query vector, V represents the vector of the queried information, and K represents the correlation vector between the queried information and other information. Then, K is combined with the learned pairwise positional bias, weighted and averaged with V, and then multiplied element-wise with Q. Finally, a T×200 dimensional feature map is generated through a sigmoid activation function, fully capturing the temporal feature information of the peptide sequence without increasing model complexity. The core of temporal feature extraction is AFT, which automatically assigns weights to the internal relationships of the sequence. The mathematical representation of this mechanism is as follows:

[0069]

[0070]

[0071] Where Q, K, and V are matrices composed of the query vector, key vector, and value vector, respectively. t This indicates that the query vector at position t is output, K t′ V represents the key vector at the t′-th input position. t′ W represents the value vector at the t′-th position of the input; Q W K W VLet X and t' represent the parameter matrices of the corresponding linear mapping, respectively. Let X be the input feature, and t, t' ∈ [1, T] be the positions of the output and input vector sequences, respectively. AFT t Let be the attention matrix at position t, ⊙ be the element-wise multiplication, and σ be the attention matrix at position t. q It is a non-linear activation function of sigmoid, ω t,t′ To output the t-th position, we need to consider the pairwise positional deviation of the input t′ position, where T represents the length of the peptide sequence.

[0072] 4. Loss Function

[0073] In this work, the multi-label dataset is imbalanced, with some therapeutic peptides being very frequent (e.g., the number of therapeutic peptides in the largest class (ABP) is 2154), while some therapeutic peptides are quite rare (e.g., the number of the smallest number (AEP) is 58). In cases of highly imbalanced data, misclassification of minority samples can cause the model to accurately predict the majority class but fail to correctly predict the minority class, which is detrimental to the minority class and hinders the performance of classification tasks. While the binary cross-entropy loss (BCE) addresses the imbalance between positive and negative samples, it does not distinguish between easily distinguishable and difficult samples. When there are a large number of easily distinguishable negative samples, the entire training process will revolve around these samples, overwhelming the positive samples and causing significant loss. To address this issue, the focal loss function, originally introduced for object detection, is employed. Due to its sparsity, it is also suitable for classification tasks on imbalanced datasets. The key idea of ​​the focal loss function is to reshape the loss, reducing the weight of easily distinguishable samples and focusing the training on difficult samples. The definition of Focal Loss is as follows:

[0074]

[0075] Where p represents the number of positive samples and n represents the number of negative samples, p i Let p represent the i-th positive sample. j Let α represent the j-th negative sample. i Let α represent the weight of the i-th positive sample. j α represents the weight of the j-th negative sample, and γ is an adjustable parameter used to adjust the weight of difficult samples. When γ is high, the minority class will receive a larger weight than the majority class, and vice versa. i α j Both γ and γ are hyperparameters. These are used to weight and fuse the extracted features, resulting in a more accurate feature representation, a better understanding of sequence information, and improved classification accuracy and stability.

[0076] To further illustrate the effects of the improvements in this invention, the following comparative experiments were conducted:

[0077] 1. Comparison with different initialization embedding representation methods

[0078] We compare our method with existing popular sequence representation methods, including numerical representation (NR) of peptide sequences using peptide sequence position information, k-mer, one-hot vectors, AAIs, BLOSUM62, and the initialization embedding representation method AAIs+BLOSUM62 proposed in this invention. k-mer is a popular sequence feature encoding method, where k=4. One-hot vectors are bit vectors that include all zeros except for a single dimension. Experimental comparisons and analysis show the following results: Figure 4 As shown in the figure. On the test set, our proposed initialization embedding representation method achieved better results overall, with an accuracy of 65.3%, coverage of 66.7%, precision of 61.5%, and an absolute error rate of 3.9%. This indicates that although NR does not require known prior information, it ignores the biological information within the amino acid sequence. Our representation has biological meaning, while the specific meaning of each dimension of the vector obtained by the embedding method is unknown. Furthermore, the initialization embedding representation method proposed in this paper has certain advantages compared to other embedding methods, suggesting that using certain prior information is beneficial for therapeutic peptide prediction.

[0079] 2. Selection of different feature extractors

[0080] To achieve an efficient multi-labeled therapeutic peptide prediction model, we compared the performance of models based on different classical machine learning methods (such as BR, CLR, RAKEL, and RBRL) and deep learning models, such as... Figure 5As shown, the box plot results of different models on the test set are presented. M1 to M8 represent different deep learning models: M1 (CNN), M2 (AFT), M3 (MHSA), M4 (CNN+AFT), M5 (CNN+BILSTM), M6 (BILSTM+MHSA), M7 (CNN+BILSTM+MHSA), and M8 (Our Model). To ensure the fairness of the model comparison, we adopted two strategies during the training of other deep learning models: (1) uniformly encoding the peptide sequences according to the optimal AAIs+BLOSUM62 encoding method and using it as the input vector for all models; (2) using the average of ten repeated experiments as the prediction results of each deep learning method. Table 1 shows that the M8 model performed best in comparison with other models. The accuracy and absolute true rate on the test set were improved by 1.1% and 1.7%, respectively. This is mainly because deep learning models automatically extract the implicit feature information of peptide sequences, using CNNs in spatial feature extraction for local and global feature extraction, and further filtering features using BILSTM and AFT in temporal feature extraction, thus improving model performance. Therefore, we adopt a CNN+BiLSTM+AFT model for multifunctional therapeutic peptide prediction.

[0081] Table 1 Results of different models on the test set

[0082]

[0083] 3. Performance comparison of different loss functions

[0084] In this section, we discuss the performance of different data balancing methods. We compare the Focal Loss loss function with several classic algorithms: cross-entropy, MLSMOTE, CW1, and CW2. In recent years, cross-entropy loss has been widely used for training deep learning-based biopeptide predictors. Focal Loss has achieved great success in multiple fields such as image classification and entity recognition. The MLSMOTE, CW1, and CW2 data are from the PrMFTP paper. Experimental results are shown in Table 2, demonstrating that Focal Loss has a significant advantage in improving prediction accuracy. Under the same model architecture and initial embedding, the model trained using Focal Loss achieves a Coverage value of 0.667, a 4% improvement over cross-entropy loss. In the experiments, we also attempted to apply an exponential decay strategy to control the speed of parameter updates. The results show that the Focal Loss loss function performs best. This loss reduces the contribution of well-classified samples and increases the contribution of poorly classified or misclassified samples, allowing the model to better handle class imbalance problems.

[0085] Table 2. Performance comparison of different loss functions on the test set.

[0086]

[0087] 4. Comparison of different methods for predicting multifunctional therapeutic peptides

[0088] To further demonstrate the model's universality, we compared it with existing methods. We randomly selected 80% of the dataset as a subset and repeated this process 10 times, resulting in 10 subsets. The model was tested on each subset based on five evaluation metrics, and the average performance of the subsets was used as the final result for each model (the test results for MLBP predictions were directly obtained from the literature PrMFTP). The results are shown in Table 3. On the test set, TP-STNN outperformed PrMFTP and MLBP in precision, coverage, accuracy, and absolute correctness, with values ​​of 0.653, 0.667, 0.615, and 0.535, respectively. Overall, TP-STNN exhibits better test performance than the PrMFTP model, providing important features for prediction. Furthermore, this method demonstrates better performance in predicting multifunctional peptides. This is mainly because the inclusion of richer biological information makes it easier for the model to learn its intrinsic characteristics. By reshaping the loss of simple and difficult samples, the focus is shifted to difficult samples, thereby improving the model's classification performance.

[0089] Table 3 compares the performance with existing methods and tools. * indicates the average result of 10 replicate experiments.

[0090]

[0091] 5. Ablation test

[0092] The effects of different network modules on peptide sequence feature extraction were compared through ablation experiments, and the results are shown in Table 4. To verify the superiority of each module for peptide sequence feature extraction, we compared three neural network frameworks: (i) a model without a CNN module, (ii) a model without a BILSTM module, (iii) a model without an AFT module, and (iiii) a model without a FL module. Among them, CNN and BILSTM networks are commonly used for peptide sequence feature extraction. As can be seen from the table, the removal of any module leads to a performance decrease. Comparing the results with / o CNN, / o BiLSTM, / o AFT, and TP-STNN, the features extracted by the CNN layer are beneficial to improving the model's prediction performance. Comparing the results with / o FL and the model, we found that the focal loss function is beneficial to improving the model's performance. Removing AFT leads to a decrease in model performance, indicating that AFT can optimize the extracted features and improve model performance.

[0093] Table 4 Ablation Experiment

[0094]

[0095] 6. Visualization of feature representation during training

[0096] Deep learning-based methods can automatically extract features. To verify the effectiveness of the model's feature extraction, we further visualize the optimized model. We use t-distributed random neighborhood embedding (t-SNE) dimensionality reduction to reduce the model's high-dimensional latent features to 2D for easier visualization. We output the results of the last layer of the network (i.e., the second fully connected layer) to obtain the high-dimensional features. High-dimensional features cannot be directly visualized. t-SNE maps high-dimensional data to a low-dimensional space by minimizing the KL divergence between data points. Figure 6 (a)~ Figure 6 (f) shows the discriminative effect of the automatically extracted features on the test set, with epoch[x] in the title of each subfigure indicating the corresponding training period. Initially (epoch = 0), samples from each class are mixed together because the neural network parameters are randomly initialized. As the number of training iterations increases (epoch = 60), the samples become increasingly distinct, indicating that our model can learn sufficient information through backpropagation optimization.

[0097] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multifunctional therapeutic peptide prediction method based on deep spatiotemporal neural networks, characterized in that, Includes the following steps: S1, perform multi-angle initialization embedding on the multifunctional therapeutic peptide sequence to fully obtain feature expression information, thereby obtaining the initialization features; Multi-angle initialization embedding is implemented using AAIs and BLOSUM62 descriptors; S2, the initial features are input into the multi-scale spatial feature extractor. The multi-scale spatial feature extractor uses a multi-scale convolutional neural algorithm to satisfy the feature extraction of peptide sequences of different lengths. At the same time, residual connections are introduced to integrate multi-scale representation and original information to extract the spatial information of peptide sequences without increasing the complexity of the model, thereby obtaining a peptide sequence feature matrix containing spatial information. S3, the peptide sequence feature matrix containing spatial information is sequentially passed through a bidirectional long short-term memory network and an AFT module to extract temporal features; thus, the temporal feature information of the peptide sequence is fully obtained without increasing the model complexity; S4 maps high-dimensional temporal feature information to the sample label space through a fully connected layer, and finally transforms it into probability values ​​through an activation layer to obtain the classification result.

2. The multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network according to claim 1, characterized in that, Multi-angle initialization embedding is implemented using AAIs and BLOSUM62 descriptors, and includes the following steps: S1-1, fix the length T of each peptide sequence to a set value. For peptide sequences that are shorter than the set value, the length of the set value is achieved by adding 0 to the end of the input sequence. S1-2 are encoded using AAIs and BLOSUM62 respectively to obtain feature maps based on AAIs and feature maps based on BLOSUM62. S1-3, the feature map based on AAIs and the feature map based on BLOSUM62 are concatenated to obtain the initial embedding of the peptide sequence; AAIs represents encoding methods based on the physicochemical properties of amino acids, while BLOSUM62 represents a scoring method based on the comparison of the mutation probability between amino acids throughout the evolutionary process.

3. The multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network according to claim 1, characterized in that, The multi-scale spatial feature extractor uses three one-dimensional convolutions with different kernel sizes k to extract the spatial features of the peptide sequence, where k is 2, 3, and 8, the number of channels is 64, and the padding is 0.

4. The multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network according to claim 1, characterized in that, The multi-scale spatial feature extractor is based on residual connections, and the spatial information of the peptide sequence obtained therefrom is represented as follows: Where vector f MSC Spatial information of the peptide sequence; X in It is an input vector encoded by AAIs and BLOSUM62; O MSC (X in () represents the output of multi-scale convolution; It is connected in series.

5. The multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network according to claim 1, characterized in that, The peptide sequence feature matrix containing spatial information is sequentially passed through a bidirectional long short-term memory network and the AFT module for temporal feature extraction, specifically including the following steps: S3-1, the spatially extracted feature matrix is ​​fed into a bidirectional long short-term memory network to capture long-distance dependencies; S3-2, The extracted feature map is input into the AFT module, and the AFT module outputs the temporal feature information of the peptide sequence; the transformation process of the feature map in the AFT module includes: The AFT module first linearly transforms the data into Q, K, and V, which are obtained by transforming the same features output by BILSTM through a linear layer. Q represents the query vector, V represents the vector of the queried information, and K represents the relevance vector of the queried information to other information. Then, K is combined with the learned pairwise positional bias, weighted and averaged with V, and then multiplied element-wise with Q. Finally, a T×200-dimensional feature map is generated through the sigmoid activation function, which fully captures the temporal feature information of the peptide sequence without increasing the model complexity.

6. The multifunctional therapeutic peptide prediction method based on a deep spatiotemporal neural network according to claim 1, characterized in that, This also includes introducing a loss function during training: Where p represents the number of positive samples; n represents the number of negative samples; p j This represents the i-th positive sample; p j This represents the j-th negative sample; α i This represents the weight of the i-th positive sample; α j This represents the weight of the j-th negative sample; γ is an adjustable parameter used to adjust the weights of difficult samples.

Images