A polypeptide function prediction method and device

Abstract

The application discloses a kind of polypeptide function prediction method and device.The method includes data set construction and data preprocessing;Establish multi-label prediction model, by feature embedding module, multi-scale convolutional neural network CNN module, BiGRU module for processing context-related sequence and classification module are composed;The network model constructed is trained using the obtained data set, and the optimization algorithm is used to improve the model feature extraction and parameter prediction performance;Then the performance of the multi-label prediction model is evaluated by calculating the accuracy, coverage, accuracy, absolute true value and absolute false value five indexes.The application effectively utilizes the correlation between labels, improves the accuracy of polypeptide function prediction, the method is strong in operability, strong in practicality, can be applied to the function prediction of eight kinds of bioactive peptides simultaneously, is a reasonable and effective prediction method, the precision of eight polypeptide functions predicted by the application is as high as 80.4%.

Description

Technical Field

[0001] This invention belongs to the technical field of interdisciplinary research between computer application technology and biotechnology, and specifically relates to a method and device for predicting peptide function. Background Technology

[0002] Bioactive peptides are a class of peptides with hormone- or pharmacological activity that regulate physiological functions by binding to specific receptors on target cells. They are widely distributed in various species in nature, from lower bacteria to higher vertebrates. Bioactive peptides are generally produced by cleaving and modifying intracellularly synthesized protein precursors and possess specific biological functions. Due to the widespread presence and functional importance of bioactive peptides, many researchers have dedicated themselves to the study of these small molecule peptides, resulting in the discovery of a large number of new bioactive peptides. Compared to small molecule compounds that are low-cost, easy to transport, but have lower specificity, peptides, despite their relatively higher cost, offer higher tolerability, better specificity, and stronger interactions with proteins. This has led to the market having over 70 peptides approved as therapeutic agents in the US, Europe, and Japan, over 200 in clinical trials, and over 600 in preclinical trials. As of September 2021, there were 1264 peptide drugs globally, including peptides. An increasing number of bioactive peptides with multiple functions are being identified.

[0003] To obtain more effective peptide therapies, the identification of peptide functions is crucial. However, the speed at which experimental methods identify peptide functions cannot keep up with the rapid emergence of large-scale bioactive peptides in the genomic era. With the ever-increasing number of peptides, numerous computational methods have emerged to guide researchers in pre-screening peptide functions, making computational methods such as machine learning increasingly important in peptide function prediction. Despite significant progress, the following challenges remain. First, many machine learning-based prediction methods are affected by small sample sizes, easily leading to overfitting and poor generalization. Second, previous research has largely focused on predicting single-function bioactive peptides, with few algorithms predicting multiple functions simultaneously, and those that have emerged have low accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for predicting peptide function, addressing the aforementioned shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The first objective of this invention is to provide a method for predicting peptide function, comprising the following specific steps:

[0007] Step S1: Dataset Construction and Data Preprocessing

[0008] Eight bioactive peptides—AMPs, ACPs, AIPs, AHPs, ADPs, AOPs, IMPs, and NUPs—were extracted from the database to construct a benchmark dataset. Then, the benchmark dataset was preprocessed by analyzing redundancy and homology bias. 80% of the sequences in the preprocessed data were used to construct the training set, and 20% of the sequences were used to construct the test set.

[0009] Step S2: Establish a multi-label prediction model, which consists of a feature embedding module, a multi-scale convolutional neural network (CNN) module, a BiGRU module for processing context-related sequences, and a classification module.

[0010] The feature embedding module transforms the polypeptide biological sequence vector into a dense vector of fixed size; the amino acids {A, C, D, E, F, G, H, I, M, N, P, Q, R, S, T, V, W, Y} are assigned to natural numbers {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20}, and a zero-filling method is used to prepare a dense vector of fixed size of 500×500 dimensions; the embedding size of the feature embedding module is any one of 50, 100, and 150;

[0011] The multi-scale convolutional neural network (CNN) module consists of an encoding module and a decoding module, and is composed of multiple parallel convolutional blocks stacked together. The input of each node in the convolutional layer is a small block of the previous neural network layer. Multiple convolutional pooling layers are used to encode the sequence and extract the global features of the sequence. The pooling size of the CNN module is 3 or 5.

[0012] The BiGRU module for processing context-related sequences receives forward input and learns inverse input to extract sequence context information through two gated recurrent units (GRUs) – forward and backward; the GRU unit value of the BiGRU module is any one of 50, 100, and 150.

[0013] The classification module uses a fully connected layer as the classification module, which converts the output of the convolutional layer into a vector pattern, and each output neuron is connected to the input neuron; the fully connected dimension of the classification module is 128 or 256.

[0014] Step S3: Train the constructed multi-label prediction model using the acquired dataset, and

[0015] An optimization algorithm is used to improve the model's feature extraction and parameter prediction performance; the training learning rate is 0.01 or 0.001.

[0016] Step S4: The performance of the multi-label prediction model is evaluated by calculating five indicators: accuracy, coverage, precision, absolute true value, and absolute false value.

[0017] Furthermore, in step S1, the database includes DRAMP, AntiCP, PreAIP, MAHTPRED, BioDADpep, and Uniprot; the preprocessing process involves using the protein sequence or nucleic acid sequence clustering tool CD-HIT to delete sequences with a similarity greater than 90%, avoiding redundancy and homology bias within each functional peptide dataset.

[0018] Furthermore, in step S2, before implementing the embedding mechanism, the peptide is converted into a filling vector and then input into the multi-tag prediction model. The peptide has a length of 5 to 500 amino acids, and filling vectors of equal length can be processed as input files.

[0019] Furthermore, in step S2, in the BiGRU module that processes context-related sequences, the formula for calculating the activation of the j-th hidden unit in the GRU layer is as follows:

[0020]

[0021] Where σ is the logistic sigmoid function, Wz and Uz represent different learned weight matrices, Xt is the input, and ht-1 is the previous hidden state. The output of the multi-scale convolutional network layer is used as the input.

[0022] At time t, the reset gate r of the j-th hidden unit is similarly calculated by the following formula:

[0023]

[0024] Then, the activation of the j-th hidden unit at time t is calculated by the following formula:

[0025]

[0026] The candidate activations are calculated accordingly using the following formula:

[0027]

[0028] Where tanh is the hyperbolic tangent function, and rt is the element-wise multiplication and a set of reset gates, respectively.

[0029] Furthermore, the classification module uses a fully connected layer as the classification module, which converts the output of the convolutional layer into a vector pattern, and each output neuron is connected to the input neuron.

[0030] Furthermore, in step S4, the five indicators—precision, coverage, accuracy, absolute true value, and absolute false value—are defined as follows:

[0031]

[0032]

[0033]

[0034]

[0035] Where N is the total number of multifunctional bioactive peptides, ∪ represents the union in set theory, ∩ represents the intersection in set theory, || represents the operation of counting the number of elements, and L... i Let represent the subset of the i-th sample that has true labels. This represents a subset of samples with predicted labels for the i-th sample.

[0036] Furthermore, in step S4, the sensitivity (SEN) and specificity (SPE) were used to evaluate the predictive performance of different multi-labeling methods for different peptide functions. The calculation formulas are as follows:

[0037]

[0038]

[0039] TP, TN, FP, and FN correspond to the number of true positives, true negatives, false positives, and false negatives, respectively.

[0040] The second objective of this invention is to provide a peptide function prediction device, comprising a dataset construction and data preprocessing module, a feature embedding module, a multi-scale convolutional neural network (CNN) module, a BiGRU module for processing context-dependent sequences, and a classification module. The dataset construction and data preprocessing module constructs a benchmark dataset by extracting eight bioactive peptides (AMPs, ACPs, AIPs, AHPs, ADPs, AOPs, IMPs, and NUPs) from a database. Then, redundancy and homology bias are addressed using CD-HIT, a tool for clustering protein or nucleic acid sequences, deleting sequences with similarity greater than 90%. 80% of the preprocessed sequences are used to construct the training set, and 20% are used to construct the test set. The feature embedding module uses peptide length as the basis for processing. A padding vector of 5-500 amino acids is used, and padding vectors of equal length can be processed as input files to transform the polypeptide biological sequence vector into a dense vector of fixed size. The multi-scale convolutional neural network (CNN) module consists of an encoding module and a decoding module, which is composed of multiple parallel convolutional blocks stacked together. The input of each node in the convolutional layer is a small block of the previous layer of the neural network. Multiple convolutional pooling layers are used to encode the sequence and extract the global features of the sequence. The BiGRU module for processing context-related sequences uses two gated recurrent units (GRUs) for forward and backward directions to receive the forward input and learn the backward input to extract the sequence context information. The classification module uses a fully connected layer as the classification module, which transforms the output of the convolutional layer into a vector pattern, and each output neuron is connected to the input neuron.

[0041] A third objective of the present invention is to provide a peptide function prediction device, comprising a processor and a memory, wherein the memory stores an application program executable by the processor, for causing the processor to perform the aforementioned peptide function prediction method.

[0042] A fourth objective of the present invention is to provide a computer-readable storage medium storing computer-readable instructions for performing the above-described peptide function prediction method.

[0043] Compared with the prior art, the beneficial effects of the technical solution provided by the present invention are as follows:

[0044] (1) The method of this invention includes dataset construction and data preprocessing; establishing a multi-label prediction model, consisting of a feature embedding module, a multi-scale convolutional neural network (CNN) module, a BiGRU module for processing context-related sequences, and a classification module; training the constructed network model using the acquired dataset, and employing optimization algorithms to improve the model's feature extraction and parameter prediction performance; and then evaluating the performance of the multi-label prediction model using five indicators: calculation accuracy, coverage, accuracy, absolute true value, and absolute false value. In this process, this invention effectively utilizes the correlation between labels to improve the accuracy of peptide function prediction. This method is highly operable and practical, and can be applied simultaneously to the functional prediction of eight bioactive peptides: AMPs, ACPs, AIPs, AHPs, ADPs, AOPs, IMPs, and NUPs. It is a reasonable and effective prediction method, and the accuracy of this invention in predicting the functions of eight peptides reaches 80.4%. Attached Figure Description

[0045] Figure 1 This is a ranking graph comparing the accuracy and absolute truth of models on the training and test sets.

[0046] Figure 2 This is a comparison chart of the accuracy of the prediction results of the test machine for AIP, AMP, AHP, ACP, ADP, AMP, IMP, and NUP. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments and accompanying drawings are described in further detail below. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in this field or according to the product instructions. Reagents or instruments whose manufacturers are not specified are all commercially available conventional products.

[0048] The database information used in this invention is as follows:

[0049] DRAMP: Antimicrobial peptide database, website: http: / / dramp.cpu-bioinfor.org / ;

[0050] AntiCP: Anticancer peptide database, URL: https: / / webs.iiitd.edu.in / raghava / anticp2 / ;

[0051] PreAIP: URL: http: / / kurata14.bio.kyutech.ac.jp / PreAIP / ;

[0052] mAHTPred: URL: http: / / thegleelab.org / mAHTPred;

[0053] BioDADpep: A bioinformatics database of antidiabetic peptides, available at http: / / omicsbase.com / BioDADPep / .

[0054] Uniprot (Swiss-prot): A protein database, available at https: / / www.uniprot.org / .

[0055] Example 1

[0056] S1. Dataset Construction and Data Preprocessing

[0057] S1-1, Determination of the benchmark dataset: Literature search was conducted using 'bioactive peptide' as the keyword, and eight health-related functional peptides were collected from the literature, namely antimicrobial peptide (AMP), anticancer peptide (ACP), anti-inflammatory peptide (AIP), antihypertensive peptide (AHP), antidiabetic peptide (ADP), antioxidant peptide (AOP), immunomodulatory peptide (IMP), and neuroactive peptide (NUP). Since the amount of functional peptide data was insufficient to train the deep learning model, this invention used eight bioactive peptides (AMPs, ACPs, AIPs, AHPs, ADPs, AOPs, IMPs, and NUPs) with a quantity greater than 500 to train the deep learning model.

[0058] S1-2 and AMP are basic polypeptides with in vivo antibacterial activity. ACP is a series of peptides that can inhibit cancer cell proliferation or migration, or inhibit cancer angiogenesis. AIP is an endogenous peptide with immunotherapeutic capabilities that inhibit antigen-specific T(H)1-driven responses and regulatory T cell generation. AHP is a peptide with the potential to improve hypertension by scavenging free radicals and inhibiting the activity of angiotensin-converting enzyme and reninase. ADP is a class of peptides that not only act on β cells or T cells to regulate insulin production, but also help assess the symptoms of diabetes. AOP is a class of peptides that inhibit or delay lipid oxidation and scavenge free radical damage. IMP is a class of peptides with immune functions present in the body, mainly applicable to diseases such as cancer, lung cancer, leukemia, and osteosarcoma. NUP is a class of peptides that can interact with α, δ, and γ receptors in the body as hormones and neurotransmitters, and can play a role in analgesia, regulation of respiration, and body temperature. Preparation of the benchmark dataset: AMP, ACP, AIP, AHP, ADP, AOP, IMP, and NUP were extracted from DRAMP, AntiCP, PreAIP, MAHTPRED, BioDADpep, and Uniprot (Swiss-prot), respectively. To avoid redundancy and homology bias within each functional peptide dataset, this application used a protein or nucleic acid sequence clustering tool (CD-HIT) to remove sequences with similarity >90%, ultimately obtaining a total of 1891 AMPs, 864 ACPs, 1651 AIPs, 1421 AHPs, 992 ADPs, 1201 AOPs, 1192 IMPs, and 1097 NUPs. Finally, this application selected 80% of the sequences for constructing the training set and 20% for constructing the test set. Table 1 lists the details of the dataset.

[0059] Table 1. Details of the dataset

[0060]

[0061] S2. Construction based on feature prediction algorithm

[0062] S2-1, Model Overview: In this work, this application uses a variety of deep neural network algorithms to construct a multi-label predictor. Figure 1 This paper describes the construction of a proposed multi-label predictor for multifunctional bioactive peptides. It consists of four main parts. The first part is a feature embedding module, which transforms sequence vectors into dense vectors of a fixed size. The second part is a multi-scale convolutional neural network (CNN) module. The third part is a BiGRU module for processing context-sensitive sequences. The final part is a classification module. In each output unit, a sigmoid function is used as the activation function, outputting a score between 0 and 1. These modules will be described in the next section. More details of the experiments will be given in the implementation details section.

[0063] S2-2, Feature Embedding Module: Before implementing the embedding mechanism, the peptides need to be converted into padding vectors and then input into the prediction model. First, the amino acids {A, C, D, E, F, G, H, I, M, N, P, Q, R, S, T, V, W, Y} are assigned natural numbers {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20}. The lengths of these peptides range from 5 (the minimum length of all sequences) to 500 (the maximum length of all sequences), but only peptide vectors of equal length can be processed as input files. Therefore, a zero-padding method is used to prepare peptides of a fixed length of 500 amino acids. Finally, the sequences are encoded into 500-dimensional vectors. After obtaining the sequence vectors, the positive integers of the vectors can be converted into dense, continuous feature vectors of a fixed size using the embedding layer. This mechanism is commonly used in natural language processing. In this study, the weights of the embedding layer can be updated during training. Finally, we learn to use embedding matrices to represent peptide sequences.

[0064] S2-3, Multi-scale Convolutional Neural Network (CNN) Module: It mainly consists of an encoding module and a decoding module, and is composed of multiple parallel convolutional blocks. The input of each node in the convolutional layer is a small piece of the previous layer of the neural network. Multiple convolutional pooling layers are used to encode the sequence and extract the global features of the sequence, which helps to obtain more abstract features through deeper analysis. After obtaining the convolutional feature matrix, deconvolutional layers are used to extract feature maps at different scales. By reducing the number of features, the max pooling mechanism is used to prevent excessive overlap.

[0065] S2-4. BiGRU Module for Handling Context-Related Sequences: Bidirectional GRUs, particularly in recurrent neural networks (RNNs), have achieved significant success in classification tasks. Using Bidirectional GRUs to analyze sequence streams, two gated recurrent units (GRUs) in both the forward and backward directions facilitate receiving forward inputs and learning backward inputs to extract sequence context information. In this study, the BiGRU uses its internal state to process sequence vectors, fully utilizing sequence context information from both directions. Furthermore, the gated recurrent unit (GRU) is the main part of the BiGRU, used to dynamically remember or forget sequence information. To gain more detail about the GRU, the activation of the j-th hidden unit in the GRU layer is described below. First, the update gate z of the j-th hidden unit at time t is calculated using the following formula:

[0066]

[0067] Where σ is the logistic sigmoid function, Wz and U z X represents the different weight matrices for learning. t It is the input, h t-1 This is the previous hidden state. And the output of the multi-scale convolutional network layer is used as input.

[0068] At time t, the reset gate r of the j-th hidden unit is similarly calculated by the following formula:

[0069]

[0070] Then, the j-th hidden unit is activated at time t. Calculated by the following formula:

[0071]

[0072] The candidate activations are calculated accordingly using the following formula.

[0073]

[0074] Where tanh is the hyperbolic tangent function, ⊙, and r t These are element-wise multiplication and a set of reset gates, respectively.

[0075] S2-5, Classification Module: This application uses a fully connected layer as the classification module. This facilitates the conversion of the convolutional layer output into a vector pattern, and each output neuron can be connected to an input neuron. A rectified linear unit (RCU) is used as the rectified linear activation unit. Furthermore, the vector from the fully connected layer is used as the input to the output layer. In the multi-label problem, the probabilities of each node are independent, with binary cross-entropy used as the loss function. A sigmoid function is used as the activation function to obtain the score of each node between 0 and 1. Finally, this application uses a 0.7 threshold to obtain the predicted label for each class, and the elements in the multidimensional prediction vector correspond to the labels of ACP, ADP, AHP, AIP, AMP, AOP, IMP, and NUP.

[0076] S2. Evaluation Indicators:

[0077] Previous studies have proposed several evaluation metrics to assess the performance of multi-label models. To evaluate the performance of the method in this application, these measures were used, including precision, coverage, accuracy, true value, and false value. These metrics are defined as follows:

[0078]

[0079]

[0080]

[0081]

[0082] Where N is the total number of multifunctional bioactive peptides, ∪ represents the union in set theory, ∩ represents the intersection in set theory, |||| represents the operation of counting the number of elements, and L i Let represent the subset of the i-th sample that has true labels. Let represent a subset with the predicted labels for the i-th sample, and

[0083]

[0084] Furthermore, within specific categories of functional peptides, peptides possessing that function are considered positive samples, while other peptides lacking that function are defined as negative samples. To evaluate the predictive performance of different multi-labeling methods for different peptide functions, this application applied sensitivity (SEN) and specificity (SPE), calculated as follows:

[0085]

[0086]

[0087] TP, TN, FP, and FN correspond to the number of true positives, true negatives, false positives, and false negatives, respectively. For example, on the test set, when calculating the SEN and SPE of ACP in this application, ACP is a positive sample, and other peptides in this group without anti-cancer function are considered negative samples. Using the prediction results of these positive and negative samples, SEN and SPE are obtained through a confusion matrix. Detailed information about the prediction model is shown in Table 2.

[0088] Table 2. Detailed information on the prediction model

[0089]

[0090]

[0091] S3. Implementation process:

[0092] TensorFlow's high-level Keras API is used to build and train the predictive model of this application. This API can be used for rapid prototyping, cutting-edge research, and real-world production, and has three major advantages: ease of use, modularity and composability, and easy extensibility. In addition, this application uses GPUs for acceleration.

[0093] Training Process. In the multi-label prediction model of this application, a grid search procedure was used to adjust the hyperparameters on the training set through 5x cross-validation. As shown in Table 2, these parameters were adjusted, including embedding size, pooling size, GRU units, fully connected dimension, and learning rate. For the CNN layer, it was constructed using the Conv1D function in Keras. To extract various convolutional features, three convolutional kernel sizes (size, number of channels, depth) ks∈{2,3,8} were selected. Then, the model was trained using the Adam (adaptive moment estimation) optimizer, which combines two popular algorithms, "Adagrad" (for handling sparse gradients) and "RMSPro" (for handling non-stationary data), to address the optimization problem of large data volumes and high feature dimensions in machine learning. Finally, a model with a batch size of 64 and 30 epochs was trained. Due to the random initialization in deep learning frameworks, these nine identical deep framework models were trained in parallel with different random initializations. The final prediction score for the test samples was obtained by averaging the scores of all models.

[0094] S4. Results and Discussion:

[0095] S4-1 Performance Comparison of Different Multi-Label Prediction Models: The multi-label prediction model employs embedding mechanisms, CNNs, and RNNs to build the model. For the CNN layer, this application utilizes multi-scale CNNs to extract multiple convolutional features. Previous research reports indicate that CNNs with small convolutional kernels exhibit more economical and better performance. For the RNN layer, gated recurrent units (GRUs) and long short-term memory (LSTMs) are commonly used in classification tasks. To select the best-performing model, this application created five deep learning models for comparison, including the base model CNN, CNN with BiGRU (CNN-BiGRU), CNN with bidirectional LSTM (CNN-BiLSTM), CNN with small convolutional kernels (CNN-sk), and a combination of CNN-sk and BIGR (CNN-sk-BIGRU).

[0096] This application evaluates their performance on training and testing datasets. Among the evaluation metrics listed in the methods, the most important metrics are accuracy and true value. The two metrics for various models on the training and testing datasets are as follows: Figure 1 As shown in Table 3, CNN-BiGRU achieved the best performance compared to the other four multi-label prediction models on the training dataset. CNN-BiGRU achieved the best performance with an accuracy of 0.749 and a ground truth value of 0.740. Table 3 lists other performance metrics for these different multi-label prediction models.

[0097] Table 3. Performance of various multi-label prediction models

[0098]

[0099] As shown in Table 3, on the training set, CNN-BiGRU achieves an accuracy of 0.749, a coverage of 0.752, an absolute error of 0.108, and a cross-validation multiplier of 5, outperforming the other four models. On the test set, CNN-BiGRU also achieves the best performance in these four metrics. This is because, compared to models based solely on CNN, the model incorporating the BiGRU mechanism can not only acquire convolutional features but also dynamically remember or forget information flow. On these evaluation metrics, CNN-BiLSTM and CNN-BiGRU have similar performance, but CNN-BiLSTM requires a longer convergence time. GRU is an RNN algorithm proposed to address problems such as long-term memory and gradients in backpropagation, similar to LSTM. On one hand, GRU has fewer parameters than LSTM. LSTM has three gate mechanisms (forget gate, input gate, and output gate), while GRU only has two gate mechanisms: update gate and reset gate. Therefore, BiGRU has a shorter convergence time. On the other hand, GRU achieves comparable results and is easier to train, significantly improving training efficiency. GRU also demonstrates better problem-solving capabilities than traditional RNNs. Finally, this application selects the best-performing model, CNN-BiGRU, to construct a multi-label bioactive peptide predictor.

[0100] S4-2. Impact of Different Sequence Lengths on Model Performance: In the benchmark dataset, the lengths of various bioactive peptides range from 5 to 500 residues. Considering the wide length range of bioactive peptides, this application used vectors with different fixed lengths as input to test whether they affected model performance. First, to ensure that the amount of data in each experiment was the same and as much as possible, this application extracted samples with lengths ≤100 from the benchmark dataset. Subsequently, 2001 sequences were selected for further testing, then converted into numeric vectors and padded with zeros to different lengths, including 100, 200, 230, 400, and 500. As shown in Table 4, the accuracy values ​​for different fixed lengths were similar to each other. The highest value (0.751) was only 1.3% higher than the lowest value (0.738). Meanwhile, in terms of accuracy alone, the values ​​of the other four evaluation metrics were similar in each experiment. Overall, sequence length has a slight impact on the model.

[0101] Table 4. Prediction accuracy for various lengths

[0102]

[0103] S4-3, Case Prediction: As mentioned above, this application constructs a multi-label bioactive peptide prediction model. Performance comparisons of different multi-label methods show that this prediction model outperforms other methods in predicting multifunctional bioactive peptides. To further illustrate the superior performance of this prediction model, this application conducts a case study of these multi-label predictors to predict the functions of given negative and positive samples, covering all functions of eight bioactive peptide classes. The prediction results using different algorithms are shown in Table 5. This prediction model demonstrates good accuracy in predicting all eight active functions, while other algorithms only detected two to three true positives.

[0104] Table 5. Accuracy compared with other algorithms

[0105]

[0106] S4-4, Test Set Prediction Results: This application tested the method using a new test set (7100 AMPs, 2800 ACPs, 3834 AIPs, 2462 AHPs, 1956 ADPs, 2824 AOPs, 2473 IMPs, and 2821 NUPs). The results are shown in Table 6 and... Figure 2 As shown, the accuracies are: AHP: 78.4%, AMP: 82.7%, AIP: 80.3%, ACP: 79.5%, ADP: 80.4%, AOP: 80.1%, IMP: 79.9%, NUP: 80.2%, and the overall accuracy is 80.4%.

[0107] Table 6. Prediction Results for the Test Set

[0108]

[0109] Where there is no conflict, the above embodiments and features described herein can be combined with each other.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting peptide function, characterized in that, The specific steps include the following: S1. Dataset Construction and Data Preprocessing Eight bioactive peptides were extracted from the database: antimicrobial peptides AMPs, anticancer peptides ACPs, anti-inflammatory peptides AIPs, antihypertensive peptides AHPs, antidiabetic peptides ADPs, antioxidant peptides AOPs, immunomodulatory peptides IMPs, and neuroactive peptides NUPs, to construct a benchmark dataset. Then, the benchmark dataset was preprocessed by analyzing redundancy and homology bias. 80% of the sequences in the preprocessed data were used to construct the training set, and 20% of the sequences were used to construct the test set. S2. Establish a multi-label prediction model, which consists of a feature embedding module, a multi-scale convolutional neural network (CNN) module, a BiGRU module for processing context-related sequences, and a classification module. The feature embedding module transforms the polypeptide biological sequence vector into a dense vector of fixed size; the amino acids {A, C, D, E, F, G, H, I, M, N, P, Q, R, S, T, V, W, Y} are assigned to natural numbers {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20}, and a zero-filling method is used to prepare a dense vector of fixed size of 500×500 dimensions; the embedding size of the feature embedding module is any one of 50, 100, and 150; The multi-scale convolutional neural network (CNN) module consists of an encoding module and a decoding module, and is composed of multiple parallel convolutional blocks stacked together. The input of each node in the convolutional layer is a small piece of the previous neural network layer. Multiple convolutional pooling layers are used to encode the sequence and extract the global features of the sequence. After obtaining the convolutional feature matrix, deconvolutional layers are used to extract feature maps at different scales. By reducing the number of features, a max pooling mechanism is used to prevent excessive overlap. The pooling size of the CNN module is 3 or 5. The BiGRU module for processing context-related sequences receives forward input and learns inverse input to extract sequence context information through two gated recurrent units (GRUs) – forward and backward; the GRU unit value of the BiGRU module is any one of 50, 100, and 150. The classification module uses a fully connected layer as the classification module, which converts the output of the convolutional layer into a vector pattern, and each output neuron is connected to the input neuron; the fully connected dimension of the classification module is 128 or 256. S3. Train the constructed multi-label prediction model using the acquired dataset, and use optimization algorithms to improve the model's feature extraction and parameter prediction performance; the training learning rate is 0.01 or 0.

001. S4. The performance of the multi-label prediction model is evaluated by calculating five indicators: accuracy, coverage, precision, absolute true value, and absolute false value.

2. The polypeptide function prediction method as described in claim 1, characterized in that, In step S1, the database includes DRAMP, AntiCP, PreAIP, MAHTPRED, BioDADpep, and Uniprot; the preprocessing process involves using the protein sequence or nucleic acid sequence clustering tool CD-HIT to delete sequences with a similarity greater than 90%, avoiding redundancy and homology bias within each functional peptide dataset.

3. The polypeptide function prediction method as described in claim 2, characterized in that, In step S2, before implementing the embedding mechanism, the peptide is converted into a filling vector of equal length and then input into the multi-tag prediction model. The peptide has a length of 5 to 500 amino acids, and the filling vector of equal length is processed as an input file.

4. The polypeptide function prediction method as described in claim 3, characterized in that, In step S2, in the BiGRU module that processes context-dependent sequences, the formula for calculating the activation of the j-th hidden unit in the GRU layer is as follows: Equation (1) Where σ is the logistic sigmoid function, Wz and Uz represent different weight matrices learned, Xt is the input, and ht is the input. 1 represents the previous hidden state, and the output of the multi-scale convolutional network layer is used as the input; At time t, the reset gate r of the j-th hidden unit is similarly calculated by the following formula: Equation (2) Then, the activation of the j-th hidden unit at time t is calculated by the following formula: Equation (3) The candidate activations are calculated accordingly using the following formula: Equation (4) Where tanh is the hyperbolic tangent function, and rt is the element-wise multiplication and a set of reset gates, respectively.

5. The polypeptide function prediction method as described in claim 4, characterized in that, In step S4, the five indicators—precision, coverage, accuracy, absolute true value, and absolute false value—are defined as follows: Where N is the total number of multifunctional bioactive peptides, ∪ represents the union in set theory, and ∩ represents the intersection in set theory. This represents the operation of counting the number of elements, where Li represents the subset of the i-th sample that has a true label. This represents a subset of samples with predicted labels for the i-th sample.

6. The polypeptide function prediction method as described in claim 5, characterized in that, In step S4, the sensitivity (SEN) and specificity (SPE) were also used to evaluate the predictive performance of different multi-labeling methods for different peptide functions. The calculation formulas are as follows: ， TP, TN, FP, and FN correspond to the number of true positives, true negatives, false positives, and false negatives, respectively.

7. A polypeptide function prediction device, characterized in that, It includes a dataset construction and data preprocessing module, a feature embedding module, a multi-scale convolutional neural network (CNN) module, a BiGRU module for processing context-dependent sequences, and a classification module; The dataset construction and data preprocessing module constructs a benchmark dataset by extracting eight bioactive peptides (AMPs, ACPs, AIPs, AHPs, ADPs, AOPs, IMPs, and NUPs) from the database. Then, redundancy and homology bias are processed using the protein sequence or nucleic acid sequence clustering tool CD-HIT, and sequences with similarity greater than 90% are deleted. In the preprocessed data, 80% of the sequences are used to construct the training set, and 20% of the sequences are used to construct the test set. The feature embedding module converts polypeptides with a length of 5 to 500 amino acids into filling vectors of equal length, and the filling vectors of equal length are processed as input files. The multi-scale convolutional neural network (CNN) module consists of an encoding module and a decoding module, and is composed of multiple parallel convolutional blocks. The input of each node in the convolutional layer is a small piece of the previous neural network layer. Multiple convolutional pooling layers are used to encode the sequence and extract the global features of the sequence. After obtaining the convolutional feature matrix, deconvolutional layers are used to extract feature maps at different scales. The BiGRU module for processing context-related sequences uses two gated recurrent units (GRUs) – one forward and one backward – to receive forward input and learn backward input to extract sequence context information. The classification module uses a fully connected layer as the classification module, which converts the output of the convolutional layer into a vector pattern, and each output neuron is connected to the input neuron.

8. A polypeptide function prediction device, characterized in that, It includes a processor and a memory, wherein the memory stores an application program that can be executed by the processor to cause the processor to perform the peptide function prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It contains computer-readable instructions for performing the polypeptide function prediction method as described in any one of claims 1 to 6.

Images