19 peptide PTS2 transfer signal small sample prediction system based on incremental learning

Abstract

The invention discloses a 19 peptide PTS2 transport signal small sample prediction system based on incremental learning, and the system comprises a data collection module which is used for collecting a polypeptide sequence, determining a core sequence of PTS2 transport signal peptide, and dynamically intercepting to obtain a 19 peptide fragment; the feature matrix extraction module is used for mapping each amino acid in the 19 peptide fragment into a physicochemical property vector to generate a feature matrix; the PTS2 transport signal prediction model training and prediction module is used for obtaining a PTS2 prediction model based on the feature matrix and through adaptive combination setting of a loss function and incremental training, and carrying out PTS2 prediction; and the PTS2 transfer signal determination module is used for calculating the information entropy based on the prediction result to judge the existence of PTS2. According to the invention, multi-mode characteristics in first 100 amino acids can be effectively identified, accurate positioning and prediction of PTS2 transport signals in a polypeptide sequence are realized, and the method is suitable for protein function analysis and peroxisome targeted drug development.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biomedical big data processing, and particularly relates to a 19-peptide PTS2 transport signal small sample prediction system based on incremental learning. BACKGROUND

[0002] Since the advent of insulin nearly a century ago, more than 80 polypeptide drugs have entered the market. The development of polypeptide drugs and other drugs involves the determination of lead compounds. Because current knowledge is not sufficient to directly synthesize target drugs (polypeptides) without using existing models, effective traditional development of polypeptide drugs relies on screening in potential polypeptide libraries, and therefore, traditional drug development methods have problems of high average time consumption, cost, and clinical failure rate, especially in the development of targeted signal peptides, and a low-cost, rapid screening method is urgently needed.

[0003] Most peroxisomal matrix proteins contain PTS1 signals at the C-terminal, and a small number of proteins contain PTS2 signals near the N-terminal. In the post-genome era, bioinformatics tools can be used to predict protein PTS, but the current PTS signal peptide rules have not been fully understood, limiting the prediction accuracy. Most peroxisomal matrix proteins contain PTS1 type peroxisome targeting signals, and their import mechanism is relatively clear. However, some matrix proteins use PTS2 signals at the amino (N) terminus, and their import mechanism has always been a mystery.

[0004] PTS2 proteins include almost all eukaryotic organisms required for peroxisomal beta-oxidation of thiolase, as well as enzymes necessary for detoxification of branched-chain fatty acids and production of ether lipid precursors for myelin formation. PTS2 protein import defects in humans can cause Refsum disease and life-threatening diseases, so it is urgent to further explore the import mechanism of PTS2 proteins. The core sequence of PTS2 is a 9-peptide conserved motif ((R / K)-(L / V / I)-NNNNN-(H / Q)-(L / A / F), N represents a very low conservative amino acid sequence), but the function is dependent on the N-terminal position (the first 100 sequences) and species-specific variations, making it difficult to accurately identify through sequence alignment technology (AlphaFold technology foundation). Traditional wet experiment verification needs to sequence the first 100 amino acids one by one, with a single cost of more than 1,000 yuan and a time consumption of one week. Traditional drug discovery also relies on computational chemistry, which needs to construct polypeptide structures for sequences and evaluate their binding to receptors, which is time-consuming and labor-intensive.

[0005] Facing the bottleneck of traditional technology, artificial intelligence technology brings new hope for drug discovery, but also encounters multiple barriers in PTS2 screening. As disclosed in reference 1 (Identification of potent antimicrobial peptides via a machine-learning pipeline that mines the entire space of peptide sequences, Nature biomedical engineering, 2023, 7, 797-810), a full library global search technology is used to systematically identify functional peptides, which mines hundreds of billions of sequences in the entire virtual peptide library composed of 6-9 amino acids to identify effective antimicrobial peptides based on jiqixu method. A trainable machine learning module (for performing experience selection, classification, ranking and regression tasks) is composed, which is assembled in order according to the design principle from coarse to fine, to gradually narrow down the search space. However, this method can only handle sequences less than 10 peptides in length, while PTS2 positive sequences not only have a length of up to 19 peptides, but also have extremely low abundance in the proteome (only 0.1%), forming a typical "small sample-long sequence" problem.

[0006] In addition, the existing TransSAFP framework is migrated to the design of other functional polypeptides, and the prediction accuracy on low-abundance data can only reach 86%, which is difficult to meet the strict requirements of drug development for specificity. The biological activity of PTS2 is highly dependent on the secondary structure and dynamic conformation, and the prediction deviation rate of structural prediction models such as AlphaFold for sequences longer than 15 peptides in solution conformation is more than 30%, and the motif of PTS2 appears at random positions within the first 100 amino acids of the protein, which is difficult to predict by "sequence alignment-based" large language models. SUMMARY

[0007] The present application provides a 19-peptide PTS2 transport signal small sample prediction system based on incremental learning, which realizes accurate positioning of PTS2 in polypeptide sequences, solves the problems of PTS2 prediction relying on time-consuming and high-cost experiments, prediction range limited to gene ends, and inability to effectively identify multi-mode features within the first 100 amino acids, and improves the prediction efficiency and accuracy of PTS2 transport signals, providing a low-cost and high-throughput screening tool for protein function analysis, peroxide enzyme body targeting drug development and other scenarios.

[0008] A 19-peptide PTS2 transport signal small sample prediction system based on incremental learning, comprising: A data acquisition module for acquiring polypeptide sequences, determining the core sequence position of the PTS2 transport signal peptide and dynamically intercepting 19-peptide fragments; The feature matrix extraction module is used to map each amino acid in the 19-peptide fragment to a physicochemical property vector to generate a feature matrix. The PTS2 transport signal prediction module is used to obtain an initial PTS2 transport signal prediction model by taking the feature matrix as input and setting the loss function through adaptive combination. Then, incremental training is performed to obtain multiple independently trained PTS2 transport signal prediction models, and the classification results of PTS2 transport signals in 19 peptide fragments are output. The PTS2 transport signal determination module is used to calculate information entropy based on the classification results and set an information entropy threshold to determine the existence of the PTS2 transport signal.

[0009] In one embodiment, the data acquisition module divides the acquired peptide sequences into an initial dataset, a random dataset, and a dataset to be mined. The initial dataset includes peptide data that contains PTS2 transport signals and peptide data that does not contain PTS2 transport signals but whose peptide sequence composition is similar to that of PTS2 transport signals. The random dataset includes peptide sequences that do not contain PTS2 transport signals and are not repeated with the initial dataset; The dataset to be mined includes polypeptide sequences containing potential PTS2 transport signals.

[0010] In one embodiment, the process of determining the core sequence position of the PTS2 transport signal peptide and dynamically extracting the 19-peptide fragment in the data acquisition module includes: For the initial dataset, MEGA is used to achieve global sequence multiple alignment. Then, methionine in the N-terminal region of the peptide sequence is removed and a random amino acid prefix of equal length is added to locate the core 9 peptide position of the PTS2 transport signal. The core 9 peptide position of the PTS2 transport signal and its neighboring peptide sequences are extracted using the context window strategy to form a 19 peptide fragment and obtain the processed initial dataset. For the random dataset and the dataset to be mined, peptide sequences with a length of less than 19 amino acids are deleted, and only the 100 amino acids at the N-terminus are retained. For each peptide sequence, methionine in the N-terminal region is deleted and a random amino acid prefix of equal length is added. Then, a sliding window with a fixed length of 19 amino acids is set to read the fragments until the window can no longer be filled. Repeated sequences are removed to obtain the processed random dataset and the dataset to be mined.

[0011] More preferably, the step of extracting the core 9-peptide position and its neighborhood sequence of the PTS2 transport signal using the context window strategy includes: extracting 3 amino acids forward and 7 amino acids backward from the core 9-peptide, wherein the polypeptide sequence for which 3 amino acids are not extracted forward is filled with non-standard amino acid U.

[0012] In one embodiment, the feature matrix extraction module, which maps each amino acid in the 19-peptide fragment to a physicochemical property vector to generate a feature matrix, includes: Each amino acid in the 19 peptide fragments is converted into a one-hot code using amino acid indexing, resulting in a feature matrix represented by the one-hot code. The feature matrix corresponding to the 19 peptide fragments in the initial dataset is divided into a training dataset and an independent validation set. Among them, amino acid X in multiple sequences is encoded as [1,1,1,1,1]. For non-standard amino acid U, a set of parameters is randomly selected from the standard amino acid physicochemical property vector of the AAindex database as the encoding value.

[0013] More preferably, each amino acid is encoded as a 5-dimensional vector, the feature matrix of each amino acid has a size of 1×5, and the feature matrix of the 19 peptide fragment has a size of 19×5.

[0014] In one embodiment, the PTS2 transport signal prediction module includes a PTS2 transport signal prediction model comprising a bidirectional dilated convolutional unit and a cascaded convolutional unit. The left convolution in the bidirectional dilated convolution unit is set with a dilation coefficient d∈[1,7] to capture the long-range dependence of upstream amino acids, and the right convolution captures the regulatory features of downstream amino acids through sequence flipping. The left and right features are fused through channel splicing. The cascaded convolutional unit contains three convolutional layers, one max pooling layer, two dropout layers, and two fully connected layers, used to classify the input feature matrix and output the classification result of the PTS2 transport signal in the 19 peptide fragments.

[0015] In one embodiment, the adaptive combination setting loss function in the PTS2 transport signal prediction module includes: integrating three losses—difficult sample focusing, distribution alignment, and structural consistency—and dynamically adjusting their weights. The difficult sample focusing loss identifies difficult samples by using a confidence threshold and calculates weight coefficients based on the number of difficult samples. The distribution alignment loss is measured by KL divergence to determine the difference between the predicted probability distribution and the true probability distribution of the 19 peptides, and batch mean is used for batch averaging. The structural consistency loss is calculated using mean squared error to determine the error between the amino acid index embedding vector and the physicochemical property vector. The formula for the loss function is as follows: , , , in, This represents the loss function based on three loss terms and dynamically adjusted weights. Focusing on difficult sample items The value is 1.0 / sqrt(hard samples + 1), where 1.0 represents the normalization weight scale, and +1 is used to prevent the denominator from being 0. This represents the number of difficult samples. Cross-entropy loss; The distribution alignment term represents the probability distribution of the presence of the 19 peptides calculated using KL divergence. Difference from the true probability distribution , For the distribution alignment term coefficient; For structural consistency loss term, This represents the coefficient for the structural consistency loss term.

[0016] More preferably, a confidence level of 0.7 is set as the threshold for determining difficult samples.

[0017] In one embodiment, the incremental training to obtain multiple independently trained PTS2 transport signal prediction models includes: The feature matrix of the random dataset is divided into several subsets. One subset and the misclassified subsets output by the initial PTS2 transport signal prediction model are added to the input dataset of the initial PTS2 transport signal prediction model for model training. The classification results predicted by the current model are then evaluated by combining the independent validation set. The above process is iterated to train and update the model until all subsets of the feature matrix of the random dataset are predicted. The final model weights are retained to obtain multiple independently trained PTS2 transport signal prediction models.

[0018] More preferably, the metrics for evaluating the classification results predicted by the current model include: the ratio of positive to negative samples, recall, and F1 score.

[0019] In one embodiment, the step of calculating information entropy based on classification results and setting an information entropy threshold to determine the presence of a PTS2 transport signal in the PTS2 transport signal determination module includes: The classification results of PTS2 transport signals in the 19-peptide fragment include the predicted classification category and the predicted probability value of the corresponding category; The predicted classification category is either containing PTS2 signal or not containing PTS2 signal. The information entropy is calculated based on the predicted classification probability values ​​of the two categories, using the following formula: , in, For information entropy, It is the first The probability predicted by a PTS2 transport signal prediction model. This indicates the number of PTS2 transport signal prediction models. This indicates that it contains two pieces of information: whether it is a PTS2 transfer signal or not. An information entropy threshold containing the PTS2 transport signal was set to determine the presence of the PTS2 transport signal in the 19-peptide fragment.

[0020] More preferably, the information entropy threshold for the PTS2 transport signal is 0.01, which is used to indicate that a polypeptide sequence with a calculated information entropy H lower than 0.01 is considered to contain a PTS2 transport signal.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) Taking the PTS2 transport signal core 9 peptide as the center, the front and back are taken to generate a 19 peptide fragment, and the core is forced to be located at position 4-12, which solves the randomness of the core position and significantly improves the accuracy of residue positioning to 93%, avoiding the core position being missed.

[0022] (2) By fusing bidirectional dilated convolutional units and cascaded convolutional units to construct a PTS2 transport signal prediction model, the dependence of peptide long sequences was overcome, and the prediction accuracy of long sequences reached 99.5%. An adaptive loss function was set to reduce the number of training iterations, which significantly improved the processing speed of 19 peptides and reduced the false positive rate to 0.06%.

[0023] (3) Further combining incremental learning, through the iterative training mechanism of error sample feedback, dynamic expansion of training set and model update, the generalization ability of millions is achieved and the accuracy of difficult samples reaches 85%. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of a small sample prediction system for 19-peptide PTS2 transport signals based on incremental learning, provided in an embodiment of the present invention.

[0025] Figure 2 This is a flowchart illustrating a small-sample prediction method for 19-peptide PTS2 transport signals based on incremental learning, provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the architecture of the PTS2 transport signal prediction model provided in an embodiment of the present invention.

[0027] Figure 4 The ROC curve is shown for the PTS2 transport signal prediction model provided in this embodiment of the invention.

[0028] Figure 5 The PRC curve is the PTS2 transport signal prediction model provided in this embodiment of the invention.

[0029] Figure 6This is the loss curve of the PTS2 transport signal prediction model provided in this embodiment of the invention. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0031] This invention aims to address the technical problems in existing technologies, such as PTS2 prediction relying on time-consuming and costly experiments, prediction range being limited to gene ends, and inability to effectively identify multimodal features within the first 100 amino acids. It provides a small-sample prediction method for 19-peptide PTS2 transport signals based on incremental learning, achieving precise localization of PTS2 in peptide sequences, thereby significantly improving prediction efficiency and accuracy, reducing experimental dependence, and providing a low-cost, high-throughput screening tool for protein function analysis, peroxisome-targeted drug development, and other scenarios.

[0032] PTS2 is a peroxisome-targeting signal peptide, typically located in the first 100 amino acids of the N-terminus of a protein. It is a core sequence motif consisting of 9 amino acids: "-arginine-phenylalanine / tyrosine-NNNNN-glutamine-leucine-" (where N represents a random amino acid).

[0033] like Figure 1 As shown in the embodiment, a small-sample prediction system for 19-peptide PTS2 transport signals based on incremental learning is provided, comprising: The data acquisition module is used to acquire peptide sequences, determine the core sequence position of the PTS2 transport signal peptide, and dynamically extract 19 peptide fragments.

[0034] like Figure 2 The diagram shown is a flowchart of a small-sample prediction method for 19-peptide PTS2 transport signals based on incremental learning, provided in an embodiment. The details are explained in conjunction with the flowchart.

[0035] In the embodiments, such as Figure 2 As shown, the peptide sequence dataset comprises three subsets: 1. The initial dataset, which consists of 2763 peptides containing PTS2 transport signals and 4106 peptides whose sequence composition is very similar to PTS2 transport signals but do not contain them. Each peptide consists of 100 amino acids and is used for model training and performance evaluation. 2. The random dataset, which contains 293362 peptide sequences that do not contain PTS2 transport signals and are not repeated in the initial dataset, used to expand the model's generalizability. 3. The dataset to be mined, which is a collection of peptide sequences used to mine potential PTS2 transport signals, containing a plant genome library of 65,354,210 genes.

[0036] Data processing aims to transform raw sequence data with known classifications into a normalized format suitable for machine learning training and prediction. In this example, the MUSCLE algorithm in the molecular evolutionary genetics analysis (MEGA) package is first used to perform multiple alignment of the global sequences. After alignment, considering that the PTS2 motif may start from the first amino acid of the protein or from the middle of the first 100 amino acids, to ensure consistent sequence length and improve generalization, the N-terminal methionine (M) is first removed. Then, an equal-length prefix "XXX" (where X represents a random amino acid) is added to the N-terminus of all sequences. Subsequently, a context window strategy is used to dynamically extract the 9 key amino acids (S1-S2-N1-N2-N3-N4-N5-S3-S4) and their neighboring sequences in the PTS2 core region. As shown in Table 1, dynamic convolution analysis reveals that the 19-peptide sequence (including the core 9 peptides) composed of the first 3 amino acids and the last 7 amino acids of the core region exhibits the best performance in feature expression.

[0037] Table 1

[0038] Next, a fixed-length subsequence of 19 amino acids was extracted from the sequence, consisting of the first 3 conserved core amino acids, the core sequence (2+5+2), and the last 7 core amino acids. For sequences with fewer than 3 amino acids before the core, "X" was used to pad the length. To further clarify sequences containing non-standard letters, the "X" was replaced with the non-standard amino acid "U". Through these steps, an initial dataset for preliminary training or evaluation of the fixed-length model was obtained.

[0039] Subsequently, the initial dataset was randomly divided into a training dataset (90%) and an independent validation set (10%). The training dataset contained 2483 positive samples and 3700 negative samples, while the independent validation set contained 280 positive samples and 406 negative samples.

[0040] For both the random dataset and the dataset to be mined, peptide sequences shorter than 19 amino acids were removed, and only the N-terminal 100 amino acids were retained. Next, methionine residues were removed from the N-terminal region of each peptide sequence, and a uniform prefix XXX was added to all peptide sequences. Furthermore, it was confirmed that the peptide sequences contained only 20 standard amino acid letters, with other letters replaced by the uncertain letter U.

[0041] Next, a sliding window of fixed length 19 amino acids was set for each polypeptide sequence. Fragments were read starting from the beginning of the polypeptide sequence, according to the window size (19 amino acids). After reading, the window was moved down one position to read the next fragment, repeating this process until the window could no longer be filled. These fragments for each polypeptide sequence constituted a subset of the polypeptide's data. Finally, duplicate sequences were removed to obtain a unique subset sequence set, resulting in a total of 23,472,577 canonical random datasets and several canonical datasets to be mined.

[0042] The feature matrix extraction module is used to map each amino acid in the 19-peptide fragment to a physicochemical property vector to generate a feature matrix.

[0043] In this embodiment, the data from the initial dataset, random dataset, and dataset to be mined are transformed into feature matrices for use as training, testing, and prediction sets for model training and evaluation. Specifically, the PTS2 sequence depends on its three-dimensional structure, and the physicochemical properties of amino acids play a decisive role in the formation of secondary structures; therefore, the physicochemical properties of amino acids need to be efficiently encoded. To handle the 544 physicochemical properties of 20 amino acids, this study uses the AAindex encoding method to reduce their dimensionality and map them into low-dimensional vectors, with each amino acid represented by 5 parameters. Each amino acid in all data is converted into a one-hot code through the amino acid index, resulting in a 1×5 feature matrix represented by the one-hot code. Here, X is encoded as [1,1,1,1,1], and non-standard amino acid U is randomly selected from the position where AA is U in the table. Thus, the set of 19 amino acid polypeptide sequences is converted into a set of one-hot encoded feature matrices of size 19×5.

[0044] The PTS2 transport signal prediction module takes the feature matrix as input and obtains an initial PTS2 transport signal prediction model by adaptively combining and setting the loss function. Then, it performs incremental training to obtain multiple independently trained PTS2 transport signal prediction models and outputs the classification results of PTS2 transport signals in 19 peptide fragments.

[0045] In this embodiment, the matrix feature set in the processed initial dataset is randomly arranged into two parts: a training dataset (90% of the data) and an independent validation dataset (10% of the data). The training dataset is used to train the initial PTS2 transport signal prediction model, while the independent validation dataset is used to evaluate the classification results predicted by the current model and output images of the subset of incorrect samples. During model training, the training dataset images (2483 positive samples and 3700 negative samples) are randomly arranged and then five-fold cross-validation is performed to obtain five disjoint training and validation subsets. These five training subsets serve as the input data for the initial training of the model, thus training the initial PTS2 transport signal prediction model.

[0046] Given the randomness of the N-terminal position of the PTS2 transport signal peptide, a bidirectional dilated convolutional unit was designed to capture the long-range spatial associations between the 9-peptide core and neighboring amino acids, achieving long-range dependency modeling. Specifically, the left convolution in the bidirectional dilated convolutional unit, by setting the dilation coefficient d∈[1,7] to cover the context of the 9 residues, is used to capture the long-range dependency of upstream amino acids. The right convolution, through sequence flipping, is used to capture the regulatory features of downstream amino acids, and the left and right features are fused through channel splicing. In this embodiment, the dilation coefficient d=3.

[0047] Furthermore, such as Figure 3 The diagram shows the architecture of the PTS2 transport signal prediction model provided in this embodiment. The PTS2 transport signal prediction model uses a ConvNet network structure, which can accurately capture amino acid spatial correlation patterns and is more suitable for extracting PTS2 transport signal sequence features. The ConvNet network contains three convolutional layers, one max-pooling layer, two dropout layers, and two fully connected layers, used to classify the input feature matrix and output the classification results of the PTS2 transport signal in the 19 peptide fragments. Specifically: The first layer has 1 input channel, 64 output channels, a kernel size of (3,5) and an activation function of ReLU. The kernel covers the vertical space of 3 amino acids, and the 5 horizontal parameters ensure that the physicochemical properties of amino acids are not confused, capturing the combination of continuous amino acid physicochemical properties. The second layer has 64 input channels, 128 output channels, a kernel size of (2,1), and the activation function is ReLU. The (2,1) kernel can focus on the local structure of adjacent amino acids. The third layer has 128 input channels, 256 output channels, a kernel size of (3,1), and the activation function is ReLU. The vertical receptive field is expanded by the (3,1) kernel to integrate complex spatial dependencies, and the horizontal dimension is fixed at 1 to avoid feature confusion, thus achieving hierarchical feature extraction.

[0048] Subsequently, a max pooling layer with a pooling window of (2,1) is applied, followed by a Dropout layer with a dropout probability of 0.25. By combining the Dropout layer with pooling technology, parameter redundancy is reduced, overfitting is prevented, and feature representativeness is preserved in a limited dataset.

[0049] The data is then converted into a one-dimensional vector and fed into the first fully connected layer. The input is dynamically adjusted based on the length of the one-dimensional vector, with ReLU as the activation function, resulting in a 512-dimensional output. This data then passes through a Dropout layer before entering the second fully connected layer, with an input dimension of 512 and an output of 2 (binary classification). The ConvNet network is implemented using the PyTorch framework. The model classifies each data point ("containing PTS2" and "not containing PTS2") and provides the predicted probability value for the corresponding classification.

[0050] To avoid difficult samples—sequences whose peptide sequences are very similar to but are not the PTS2 transport signal peptide (with a confidence level of 0.7 set as the threshold for difficult sample identification)—the model training is disrupted in the early stages. A dynamically weighted adaptive combination of loss functions is used, specifically including: integrating three losses—difficult sample focusing, distribution alignment, and structural consistency—and dynamically adjusting their weights. The difficult sample focusing loss identifies difficult samples based on the confidence threshold, calculating weight coefficients based on the number of difficult samples. The distribution alignment loss measures the difference between the predicted probability distribution and the true probability distribution of the 19 peptides using KL divergence, and uses batchmeans for batch averaging. The structural consistency loss calculates the error between the amino acid index embedding vector and the physicochemical property vector using mean squared error. The formula for the loss function is as follows: , , , in, This represents the loss function based on three loss terms and dynamically adjusted weights. Focusing on difficult sample items The value is 1.0 / sqrt(hard samples + 1), where 1.0 represents the normalization weight scale, and +1 is used to prevent the denominator from being 0. This represents the number of difficult samples. Cross-entropy loss; The distribution alignment term represents the probability distribution of the presence of the 19 peptides calculated using KL divergence. Difference from the true probability distribution , For the distribution alignment term coefficient; For structural consistency loss term, This represents the coefficient for the structural consistency loss term.

[0051] like Figures 4-6As shown, to quantitatively evaluate the model performance during training, ROC curves, PRC curves, and loss curves were calculated for each cross-validation subset, and the AUC values ​​corresponding to each curve were recorded to monitor the changes in the model's classification power at different training stages in real time. The results of the ROC, PRC, and loss curves indicate that the model exhibits extremely strong classification ability, stable generalization, and efficient convergence in the small-sample 19-peptide PTS2 transport signal prediction task.

[0052] Among them, such as Figure 4 As shown, the AUC of the 5-fold cross-validation of the ROC curve is in the range of 0.98-0.99, and the curve is close to the upper left corner, which proves that the model has strong feature extraction ability and stable ability to distinguish between two types of samples. like Figure 5 As shown, the average precision (AP) of the PR curve for 5-fold validation is mostly between 0.99 and 1.0, and the curve is close to the upper right corner, indicating that the model has a good ability to identify minority class samples and the classification results are both comprehensive and accurate. like Figure 6 As shown, the loss curves indicate that the loss of all folds decreases rapidly within 20 iterations, and then tends to flatten and overlap in the later stages, reflecting the high optimization efficiency and stable generalization of the model.

[0053] In summary, the three curves, from the three dimensions of "classification discrimination", "precision and recall balance" and "training stability", verify the strong adaptability of the PTS2 transport signal prediction module of the system constructed in this invention in the small sample scenario. Among them, the feature extraction of bidirectional dilated convolution and the optimization logic of dynamic weighted loss of the PTS2 transport signal prediction module solve the pain points of "difficult feature capture, difficult loss balance and poor generalization" in the small sample scenario.

[0054] In addition, to improve the generalizability of the PTS2 transport signal prediction model in identifying erroneous data while retaining its accuracy in identifying correct data, incremental training was performed as follows: First, to shorten the training time and accelerate model iteration, the unique sequences in the processed random dataset are randomly partitioned. Based on the AAindex encoding method, the 19 peptide fragments are converted into a set of one-hot encoded feature matrices of size 19×5. These matrices are then divided into subsets of 1000 fragments each. Fragments are read through a reading frame, and the feature matrices of each 1000 random subsets are combined with the misclassified subset images (excluding independent validation set images) output by the initial PTS2 transport signal prediction model to form new training data images. The PTS2 transport signal prediction model is then trained again. The classification results predicted by the current model are evaluated by combining and randomizing the independent validation set images with the random dataset subset images. The proportion of correctly predicted positive and negative samples (accuracy), recall, and F1 score are recorded to quantitatively evaluate the initial classification effect of the model. For the misclassified sample subsets in the random data subset images, they are further incrementally added to the input dataset for updating.

[0055] A three-round "train-prediction-update" cyclical approach is employed: the above process is repeated three times with different random data, resulting in an updated model in each round. More specifically, a new model is trained based on the updated input dataset and used to predict combinations of independent validation sets and random datasets, repeating this process until all subsets of the feature matrices of the random datasets are used for prediction. Ultimately, three independently trained PTS2 transport signal prediction models are obtained, and the models and their weights are saved (in .pth file format).

[0056] The PTS2 transport signal determination module is used to calculate information entropy based on the classification results and set an information entropy threshold to determine the existence of the PTS2 transport signal.

[0057] In this embodiment, to fully explore polypeptide sequences similar to the PTS2 transport signal, the feature matrix of the dataset to be explored is input into three independently trained PTS2 transport signal prediction models for prediction. Finally, the classification and confidence level are obtained based on the information entropy integration results. The prediction results include the predicted category (containing or not containing the PTS2 signal) and the predicted probability value of the corresponding category.

[0058] Based on the classification probability, the information entropy is calculated using the following formula: , in, For information entropy, It is the first The probability predicted by a PTS2 transport signal prediction model. This indicates the number of PTS2 transport signal prediction models. This indicates that it contains two pieces of information: whether it is a PTS2 transfer signal or not.

[0059] The information entropy threshold for PTS2 transport signals is set to 0.01, which is used to indicate that a polypeptide sequence with a calculated information entropy H lower than 0.01 is considered to contain PTS2 transport signals.

[0060] To further verify the feasibility and effectiveness of the incremental learning-based small-sample prediction system for 19-peptide PTS2 transport signals provided by this invention, training and verification were performed on a previously obtained set of millions of peptide sequence fragments that had undergone rigorous standardization, including 23,472,577 standardized random sequences and several sequences to be mined (million-level dataset). The results are shown in Table 2.

[0061] Table 2

[0062] As shown in Table 2, compared with existing traditional methods (global search technology), this invention achieves an overall accuracy of 99.5% (±0.25%) on a dataset of millions of samples, while ensuring a positive sample recall rate of 90.93%±1.37% and a negative sample recall rate of 99.51%±0.27%.

[0063] In summary, the present invention provides a small-sample prediction system for 19-peptide PTS2 transport signals based on incremental learning. This system overcomes the problems of existing technologies, such as reliance on time-consuming and costly experiments for PTS2 transport signal prediction, limitation of prediction range to gene ends, and inability to effectively identify multi-modal features within the first 100 amino acids. By constructing a 19-peptide fragment based on the core of PTS2 transport signals and combining it with machine learning, the system achieves precise localization of PTS2 transport signals in peptide sequences, thereby significantly improving prediction efficiency and accuracy, reducing experimental dependence, and providing a low-cost, high-throughput screening tool for scenarios such as protein function analysis and peroxisome-targeted drug development.

[0064] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A 19-peptide PTS2 transport signal small sample prediction system based on incremental learning, characterized by, The application relates to a PTS2 transport signal prediction method and device. The application comprises the following steps: A data collection module is used to collect polypeptide sequences, determine the position of a PTS2 transport signal peptide core sequence and dynamically cut 19 peptide fragments; A feature matrix extraction module is used to map each amino acid in the 19 peptide fragments into a physicochemical property vector and generate a feature matrix; A PTS2 transport signal prediction module is used to take the feature matrix as input, set a loss function through adaptive combination, obtain an initial PTS2 transport signal prediction model, perform incremental training to obtain a plurality of independently trained PTS2 transport signal prediction models, and output the classification results of PTS2 transport signals in the 19 peptide fragments; 2. The 19-peptide PTS2 transit signal small sample prediction system according to claim 1, characterized in that, A PTS2 transport signal determination module is used to calculate information entropy based on the classification results and set an information entropy threshold to determine the existence of the PTS2 transport signal. In the data collection module, the collected polypeptide sequences are divided into an initial data set, a random data set and a data set to be mined; The initial data set comprises polypeptide data containing PTS2 transport signals and polypeptide data not containing PTS2 transport signals but having similar polypeptide sequences to PTS2 transport signals; The random data set comprises polypeptide sequences not containing PTS2 transport signals and not repeated with the initial data set; 3. The 19-peptide PTS2 transit signal small sample prediction system according to claim 2, characterized in that, The data set to be mined comprises polypeptide sequences containing potential PTS2 transport signals. In the data collection module, the determination of the position of the PTS2 transport signal peptide core sequence and the dynamic cutting of the 19 peptide fragments comprises the following steps: For the initial data set, MEGA is used to realize multiple alignment of global sequences, methionine in the N-terminal region of the polypeptide sequence is deleted and an equal-length random amino acid prefix is added, the core 9-peptide position of the PTS2 transport signal is located, the core 9-peptide position of the PTS2 transport signal and the polypeptide sequence in the neighborhood thereof are cut through a context window strategy, 19 peptide fragments are formed, and a processed initial data set is obtained; 4. The 19-peptide PTS2 transit signal small sample prediction system according to claim 3, characterized in that, For the random data set and the data set to be mined, polypeptide sequences with a length less than 19 amino acids are deleted, only the N-terminal 100 amino acids are retained, methionine in the N-terminal region of each polypeptide sequence is deleted and an equal-length random amino acid prefix is added, a sliding window with a fixed length of 19 amino acids is used to read fragments until the window cannot be filled any more, and repeated sequences are removed, so that a processed random data set and a processed data set to be mined are obtained.

5. The 19-peptide PTS2 transit signal small sample prediction system of claim 3, wherein, The core 9-peptide position of the PTS2 transport signal and the sequence in the neighborhood thereof are cut through a context window strategy, which comprises cutting 3 amino acids forward and 7 amino acids backward from the core 9-peptide, wherein non-standard amino acid U is used to fill the polypeptide sequence with less than 3 amino acids cut forward. In the feature matrix extraction module, the mapping of each amino acid in the 19 peptide fragments into a physicochemical property vector and the generation of a feature matrix comprise the following steps: Each amino acid in the 19 peptide fragment is converted into a one-hot code through an amino acid index to obtain a feature matrix represented by a one-hot code, and the feature matrix corresponding to the 19 peptide fragment in the initial data set is divided into a training data set and an independent validation set; wherein the amino acid X in the plurality of sequences is coded as [1, 1, 1, 1, 1]; for the non-standard amino acid U, a set of parameters is randomly selected from the standard amino acid physicochemical property vectors in the AAindex database as the coding value.

6. The 19-peptide PTS2 transit signal small sample prediction system according to claim 5, characterized by, In the PTS2 transport signal prediction module, the PTS2 transport signal prediction model comprises a bidirectional hollow convolution unit and a cascaded convolution unit. The left convolution in the bidirectional hollow convolution unit sets a hollow coefficient d [1, 7], which is used to capture the long-range dependence relationship of the upstream amino acid, and the right convolution captures the regulatory characteristics of the downstream amino acid by sequence flipping, and the left and right characteristics are fused through channel splicing. The cascaded convolution unit includes three convolution layers, one maximum pooling layer, two Dropout layers and two fully connected layers, which are used to classify the input feature matrix and output the classification result of the PTS2 transport signal in the 19 peptide fragment.

7. The 19-peptide PTS2 transit signal small sample prediction system according to claim 6, characterized in that, In the PTS2 transport signal prediction module, the adaptive combination loss function comprises: integrating three loss terms of difficult sample focus, distribution alignment and structural consistency and dynamically adjusting the weight; The difficult sample focus loss identifies difficult samples through a confidence threshold, and calculates a weight coefficient based on the number of difficult samples; The distribution alignment loss measures the difference between the predicted 19 peptide probability distribution and the real probability distribution through KL divergence, and uses batchmean for batch averaging; The structural consistency loss calculates the error between the amino acid index embedding vector and the physicochemical property vector using mean square error, and the calculation formula of the loss function is as follows: ， ， ， wherein, represents a loss function based on three losses and dynamically adjusted weights, is a difficult sample focusing item, is 1.0 / sqrt(difficult sample+1), 1.0 represents a normalization weight scale, and +1 is used to prevent the denominator from being 0, represents the number of difficult samples, is a cross-entropy loss; is a distribution alignment item, used to represent the 19 peptide existing probability distribution calculated by KL divergence difference from the real probability distribution , is a distribution alignment item coefficient; is a structure consistency loss item, is a structure consistency loss item coefficient.

8. The 19-peptide PTS2 transit signal small sample prediction system according to claim 7, characterized by, The plurality of independently trained PTS2 transport signal prediction models obtained by further incremental training, comprising: Divide the random data set feature matrix into several subsets, and incrementally input one of the subsets and the error classification subset output by the initial PTS2 transport signal prediction model into the initial PTS2 transport signal prediction model input data set to train the model, and then evaluate the classification result predicted by the current model in combination with the independent validation set; Iterate the above process to train and update the model until all subsets of the random data set feature matrix are predicted, retain the final model weight, and obtain a plurality of independently trained PTS2 transport signal prediction models.

9. The 19-peptide PTS2 transit signal small sample prediction system according to claim 8, characterized by, In the PTS2 transport signal determination module, the information entropy is calculated based on the classification result, and the information entropy threshold is set to determine the existence of the PTS2 transport signal, comprising: The classification result of the PTS2 transport signal in the 19 peptide fragment includes the predicted classification category and the predicted probability value corresponding to the classification; The predicted classification category is whether it contains a PTS2 signal or not, and the information entropy is calculated according to the classification prediction probability value of the two, and the calculation formula is as follows: ， wherein, is the information entropy, is the probability predicted by the th PTS2 transport signal prediction model, denotes the number of PTS2 transport signal prediction models, denotes the information comprising 2, being a PTS2 transport signal or not a PTS2 transport signal; Set the information entropy threshold of the PTS2 transport signal to determine the existence of the PTS2 transport signal in the 19 peptide fragment.

10. The 19-peptide PTS2 transit signal small sample prediction system of claim 9, wherein, The information entropy threshold of the PTS2 transport signal-containing is 0.01, which is used to indicate that the polypeptide sequence with a calculated information entropy H lower than 0.01 is considered to contain the PTS2 transport signal. The information entropy threshold of the PTS2 transport signal-containing is 0.01, which is used to indicate that the polypeptide sequence with a calculated information entropy H lower than 0.01 is considered to contain the PTS2 transport signal.

Images