Semi-supervised method and system for predicting protein lysine post-translational modification sites

Abstract

The present invention provides a semi-supervised method and system for predicting protein lysine post-translational modification sites. The method includes: collecting and preprocessing data to construct a data set; designing a network structure to construct a multi-scale sequence feature extraction network; training a model using an unlabeled sample self-distillation method; and evaluating the prediction results using indicators such as accuracy and recall rate. The present invention solves the technical problems of ineffective utilization of unlabeled data, difficulty in knowledge transfer between species, and low robustness of species-specific prediction models.

Description

Technical Field

[0001] The present invention relates to the field of biomedical engineering, and particularly to a semi-supervised method and system for predicting protein lysine post-translational modification sites. Background Art

[0002] Lysine has a special molecular structure and is easily modified after translation. For example, various modifications such as ubiquitination, methylation, and acetylation often occur on lysine. Modifications at lysine sites play important roles in cell life, and abnormal changes in modifications are also closely related to various diseases such as cancer. A variety of regulatory enzymes involved in modifications have become effective drug targets [WANG W-H, YUAN T, QIAN M-J, et al. Post-translational modification of KRAS: potential targets for cancer therapy[J]. Acta Pharmacologica Sinica, 2020, 42(8):1201-1211]. For example, VEGF can enhance its binding to BRD4 by stimulating the acetylation of the endothelial cell transcriptional regulator ETS1 to stimulate the pause release of RNA polymerase II (RNAPII), thereby driving angiogenesis [J. Chen et al., "VEGF amplifies transcription through ETS1 acetylation to enable angiogenesis," (in eng), Nat Commun, vol. 8, no. 1, p. 383, Aug 29 2017]. Therefore, the ETS1-BRD4 interaction may be an effective therapeutic target for improving diseases caused by excessive angiogenesis, such as retinopathy of prematurity. For example, the existing patent application document "A microRNA for regulating VEGF gene expression and its application" with the publication number CN114250292A discloses the role of miR-374b-5p in MCP-1-induced VEGF gene expression, clarifying that miR-374b-5p is the bridge connecting the two, and by exogenously supplementing miR-374b-5p, the effect of MCP-1 on VEGF can be relieved, inhibiting the rapid growth and metastasis of tumors caused by MCP-1. In addition, lysine acetylation modification also has an important impact on the growth and development of plants and is involved in regulating the stress resistance response of plants [Zhang Lanjun, Zhang Baocai, and Zhou J. Acta Phytophysiol Sinica, "Acetylation modification of plant cell wall polysaccharides and biological functions," vol. 54, no. 8, p. 7, 2018.]. Therefore, accurately identifying lysine modification sites in different species has important guiding significance in screening disease-related clinical markers and improving agronomic traits. Although current computational methods for predicting lysine sites have achieved certain success.For example, Siraj et al. [A. Siraj, D. Y. Lim, H. Tayara, and K. T. Chong, "UbiComb: A Hybrid Deep Learning Model for Predicting Plant-Specific Protein Ubiquitylation Sites," (in eng), Genes (Basel), vol. 12, no. 5, May 11 2021] proposed a method UbiComb for predicting plant-specific lysine ubiquitylation sites. This method uses a convolutional neural network to extract physicochemical property features, and at the same time uses a long short-term memory neural network to extract protein sequence features. Further, a fully connected neural network is used to fuse the above two types of features for site prediction. For example, in the existing invention patent application document "A Protein Secondary Structure Prediction Method Based on Long Short-Term Memory Network" with publication number CN112365921A, in this scheme, first, according to the protein sequence information of the protein to be predicted for secondary structure with an input residue number of L, the HHblits and PSI-BLAST programs are respectively used to obtain the feature matrices PSFM and PSSM; then, the above two matrices are combined into a feature matrix F and converted into a feature tensor; secondly, the protein sequence is processed into residue samples; thirdly, a long short-term memory network and a conditional random field model are built, a data set is constructed using the protein sequences with known secondary structures and the built network is trained. Finally, the residue samples of the protein sequence to be predicted for secondary structure are input into the trained model to obtain the secondary structure of the sequence. Ning et al. proposed a hybrid learning architecture HybridSucc [W. Ning et al., "HybridSucc: A Hybrid-learning Architecture for General and Species-specific Succinylation Site Prediction," (in eng), Genomics Proteomics Bioinformatics, vol. 18, no. 2, pp. 194-207, Apr 2020., which integrates deep learning and traditional machine learning algorithms, effectively reducing the model's requirement for the amount of training data.Liu et al. [Y. Liu, Q. Wang, and J. Xi, "DeepDA-Ace: A Novel Domain Adaptation Method for Species-Specific Acetylation Site," vol. 10, p. 2364, 2022] proposed a method for predicting species-specific lysine acetylation sites, DeepDA-Ace. This method uses a domain adaptation approach to transfer human acetylation knowledge to assist in site prediction for other species, achieving effective species-specific acetylation site prediction. However, these methods still have the following drawbacks: The UbiComb method requires integrating multiple plant data for training and fails to establish a specific prediction model for different species. In addition, there is little labeled training data for species-specific post-translational modifications, and the existing HybridSucc and DeepDA-Ace methods only build species-specific models based on labeled data, resulting in poor generalization ability.

[0003] In summary, the prior art has technical problems such as the inability to effectively utilize unlabeled data, difficulty in knowledge transfer between species, and low robustness of species-specific prediction models. Summary of the Invention

[0004] The technical problem to be solved by the present invention is how to solve the technical problems in the prior art, such as the inability to effectively utilize unlabeled data, difficulty in knowledge transfer between species, and low robustness of species-specific prediction models.

[0005] The present invention solves the above technical problems by adopting the following technical solutions: The semi-supervised protein lysine post-translational modification site prediction method includes:

[0006] S1. Collect acetylation modification information to extract labeled samples from pre-set acetylation site proteins to obtain labeled source domain and target domain data sets. Intercept unlabeled samples from proteins other than pre-set acetylation site proteins, perform data collection and preprocessing, and select samples from the labeled samples to construct an independent test set;

[0007] S2. Use at least two one-hot encoded sequence features of different lengths as inputs, select the appropriate number of convolutional layers and the number of channels in the bottleneck layer, set up a dense connection module, and accordingly construct a multi-scale sequence feature extraction network model;

[0008] S3. Obtain an unlabeled target data set based on the unlabeled samples, and accordingly use the self-distillation method to pseudo-label the unlabeled samples in the unlabeled target data set to obtain pseudo-labeled samples. Perform a reliable evaluation on the pseudo-labeled samples to obtain reliable student samples. Based on the labeled source domain and target domain data sets and the reliable student samples, form sample pairs and calculate the loss of the sample pairs. Accordingly, use the labeled source domain and target domain data sets to pre-train a multi-scale sequence feature extraction network model to obtain an applicable model;

[0009] S4. Use the applicable model to extract and fuse multi-scale high-level sequence features based on the one-hot encoded sequence features to obtain multi-scale fusion features. Predict the modification sites based on the multi-scale fusion features to obtain site prediction results;

[0010] S5. Use an independent test set to test the applicable model, and evaluate the site prediction results using accuracy and recall.

[0011] Compared with the existing lysine modification site prediction methods, the method of the present invention can establish a specific prediction model for small-sample species. In addition, the self-distillation method of unlabeled samples proposed by the present invention can effectively utilize the unlabeled data of other species, thereby enhancing the generalization ability of the model and expanding the application scope. The present invention has a good prediction effect on lysine post-translational modification and is an important basis for studying the biological functions of lysine modification. The present invention realizes the effective utilization of unlabeled data and the knowledge transfer between species, establishes a more robust and efficient species-specific prediction model, fully utilizes a large amount of potential species-specific unlabeled data, and realizes efficient prediction of lysine post-translational modification sites.

[0012] In a more specific technical solution, step S1 includes:

[0013] S11. Collect acetylation modification information from a public database, and extract local sequences with a length of not less than 2 based on the preset modification sites as labeled positive example samples;

[0014] S12. Select lysines without annotated acetylation information from proteins with verified acetylation sites as negative example samples;

[0015] S13. Intercept local sequences around lysine residues from other proteins outside the constructed positive and negative example data sets as unlabeled samples;

[0016] S14. Remove proteins with a similarity exceeding the preset similarity threshold;

[0017] S15. Randomly select a preset proportion of labeled samples from the labeled data as an independent test set.

[0018] In a more specific technical solution, in step S2: in the k-th dense connection module, the output of the i-th convolutional layer is:

[0019]

[0020] In the formula represents the output of the i-th convolutional layer in the k-th dense connection module, represents the convolutional kernel weight of the i-th convolutional layer, represents the activation function, and ∪ represents the channel splicing operation between feature maps.

[0021] In a more specific technical solution, in step S2, in the dense connection convolutional neural network, by using the connection between modules, the high-level sequence features extracted from no less than 2 dense connection modules are fused, and according to the following logic, the connection is performed according to the channels of multiple feature maps:

[0022]

[0023] In the formula, h C is the output of the last convolutional layer of the dense connection module, and h f is the feature map obtained by splicing the outputs of multiple dense connection modules according to channels.

[0024] The scale sequence feature extraction network in the present invention uses a dense connection convolutional neural network module to replace the traditional multi-layer convolutional neural network. The network includes 3 dense connection modules, which respectively correspond to 3 different scales of protein sequence lengths. Each dense connection module is composed of multiple interconnected one-dimensional convolutional layers. This dense connection enables each convolutional layer to receive the feature maps of all previous convolutional layers, thereby enhancing the propagation of sequence features in the network.

[0025] In a more specific technical solution, step S3 includes:

[0026] S31. Pseudo-label the unlabeled samples, and select reliable student samples through reliability evaluation to construct a reliable student sample set;

[0027] S33. Select applicable samples from the labeled source domain D s and the target domain D TL data sets and the reliable student sample set to form sample pairs;

[0028] S34. Construct an auxiliary generation module to generate auxiliary features to confuse the distribution of the sample pairs;

[0029] S35. Use the labeled source domain and target domain data sets The cross-entropy loss is obtained through processing, and the sample pairs are processed by sample self-distillation to obtain the sample pair loss;

[0030] S36. Use the reliable student samples with pseudo-labels to obtain the additional loss, and based on this, combine the cross-entropy loss and the sample pair loss to obtain the total loss, and based on this, pre-train the multi-scale sequence feature extraction network model;

[0031] In a more specific technical solution, in step S31, the following logic is used to perform pseudo-labeling on the unlabeled samples:

[0032]

[0033] where x′ ∈ D TU , D TU is the unlabeled data set in the target domain, is the pseudo-label of x′.

[0034] The present invention adopts the sample pair self-distillation method, makes full use of the unlabeled translated and modified data, and enables the model to adapt to the target domain. Each sample pair consists of a teacher sample from the labeled data set and a student sample from the unlabeled data set. This method first generates assistant features by fusing the intermediate layers of the teacher and the student, and then trains the model by minimizing the output difference between the student and the assistant. During the training process, the assistant gradually bridges the difference between the two, so that the student can easily learn from the teacher.

[0035] In a more specific technical solution, in step S32, the following logic is used to construct a reliable student sample set

[0036]

[0037] where g represents the feature extractor, h represents the classifier, π j represents the selection operator for selecting the j-th maximum value, and δ is a hyperparameter.

[0038] In a more specific technical solution, step S35 includes:

[0039] S351. Use the following logic to obtain the cross-entropy loss:

[0040] L l = E (x,y) [-log(p(y|x))] (6)

[0041] where E (x,y) represents the mathematical expectation.

[0042] S352. Use the following logic to obtain the sample pair loss:

[0043]

[0044] where D KL represents the KL divergence, is calculated when the teacher sample x passes through the feature extractor.

[0045] The sample pair loss of the present invention effectively reduces the inter-domain difference by using the source domain sample and the student sample pair, and effectively suppresses the intra-domain difference by using the target domain sample and the student sample pair.

[0046] In a more specific technical solution, in step S352, according to the following logic, the weighted cross-entropy loss is calculated based on the sample pair loss technique:

[0047]

[0048] where α represents the weight of the student sample.

[0049] The present invention introduces an additional loss by using the student sample with pseudo-label. By using the latest prediction of the student sample to determine the reliability of the pseudo-label and multiplying it by the cross-entropy loss of each student sample, the student sample is trained using the weighted cross-entropy loss, which can reduce the impact of less reliable student samples on the model update.

[0050] In a more specific technical solution, the semi-supervised protein lysine post-translational modification site prediction system includes:

[0051] A sample set establishment module for collecting acetylation modification information, extracting labeled samples from pre-set acetylated site proteins to obtain labeled source domain and target domain data sets, intercepting unlabeled samples from proteins other than pre-set acetylated site proteins, performing data collection and preprocessing, and selecting samples from the labeled samples to construct an independent test set;

[0052] A model construction module for using one-hot encoded sequence features of at least two different lengths as inputs, selecting appropriate numbers of convolutional layers and bottleneck layer channels, setting up a dense connection module, and constructing a multi-scale sequence feature extraction network model accordingly;

[0053] A model training module for obtaining an unlabeled target data set according to the unlabeled samples, using the self-distillation method to pseudo-label the unlabeled samples in the unlabeled target data set to obtain pseudo-labeled samples, performing a reliable evaluation on the pseudo-labeled samples to obtain reliable student samples, forming sample pairs according to the labeled source domain and target domain data sets and the reliable student samples, calculating the sample pair loss, and pre-training the multi-scale sequence feature extraction network model according to the labeled source domain and target domain data sets to obtain an applicable model. The model training module is connected to the model construction module and the sample set establishment module;

[0054] The site prediction module is used to utilize an applicable model to extract and fuse multi-scale high-level sequence features based on one-hot encoded sequence features to obtain multi-scale fusion features, and predict modification sites according to the multi-scale fusion features to obtain site prediction results. The site prediction module is connected to the model training module;

[0055] The evaluation and testing module is used to test the applicable model using an independent test set and evaluate the site prediction results using accuracy and recall. The evaluation and testing module is connected to the sample set establishment module and the site prediction module.

[0056] The present invention has the following advantages compared with the prior art: Compared with the existing methods for predicting lysine modification sites, the method of the present invention can establish a specific prediction model for small-sample species. In addition, the label-free sample self-distillation method proposed by the present invention can effectively utilize label-free data of other species, thereby enhancing the generalization ability of the model and expanding the application scope. The present invention has a good prediction effect on post-translational modification of lysine and is an important basis for studying the biological functions of lysine modification. The present invention realizes the effective utilization of label-free data and knowledge transfer between species, establishes a more robust and efficient species-specific prediction model, fully utilizes a large amount of potential species-specific label-free data, and realizes efficient prediction of post-translational modification sites of lysine.

[0057] The scale sequence feature extraction network in the present invention uses a densely connected convolutional neural network module to replace the traditional multi-layer convolutional neural network. The network includes 3 densely connected modules, corresponding to 3 different scales of protein sequence lengths respectively. Each densely connected module consists of multiple interconnected one-dimensional convolutional layers. This dense connection enables each convolutional layer to receive the feature maps of all previous convolutional layers, thereby enhancing the propagation of sequence features in the network.

[0058] The present invention adopts the sample pair self-distillation method to fully utilize label-free post-translational modification data and make the model adapt to the target domain. Each sample pair consists of a teacher sample from the labeled data set and a student sample from the unlabeled data set. This method first generates assistant features by fusing the intermediate layers of the teacher and the student, and then trains the model by minimizing the output difference between the student and the assistant. During the training process, the assistant gradually bridges the difference between the two, enabling the student to easily learn from the teacher.

[0059] The sample pair loss of the present invention effectively reduces the inter-domain difference using source domain samples and student sample pairs, and effectively suppresses the intra-domain difference using target domain samples and student sample pairs.

[0060] The present invention introduces an additional loss by using student samples with pseudo-labels. By utilizing the latest predictions of the student samples to determine the reliability of the pseudo-labels and multiplying it by the cross-entropy loss of each student sample, the student samples are trained using weighted cross-entropy loss, which can reduce the impact of less reliable student samples on model updates. The present invention solves the technical problems existing in the prior art, such as the ineffective utilization of unlabeled data, the difficulty in knowledge transfer between species, and the low robustness of species-specific prediction models. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the basic steps of the semi-supervised protein lysine post-translational modification site prediction method in Embodiment 1 of the present invention;

[0062] Figure 2 Schematic diagram of the Semi-ssPTM of the semi-supervised protein lysine post-translational modification site prediction method in Embodiment 1 of the present invention;

[0063] Figure 3 Schematic diagram of the specific steps for constructing the dataset in Embodiment 1 of the present invention;

[0064] Figure 4 Schematic diagram of the network training steps in Embodiment 1 of the present invention;

[0065] Figure 5a Schematic diagram of the comparison results of the performance indicators of the present solution and other existing methods on yeast in Embodiment 2 of the present invention;

[0066] Figure 5b Schematic diagram of the comparison results of the performance indicators of the present solution and other existing methods on rats in Embodiment 2 of the present invention;

[0067] Figure 5c Schematic diagram of the comparison results of the performance indicators of the present solution and other existing methods on Bacillus in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] Embodiment 1

[0070] As Figure 1 and Figure 2 shown, the semi-supervised protein lysine post-translational modification site prediction method provided by the present invention includes:

[0071] S1. Perform data collection and preprocessing to construct a dataset;

[0072] In this embodiment, acetylation is taken as an example.

[0073] As Figure 3 shown, in this embodiment, step S1 of constructing the dataset further includes the following specific steps:

[0074] S11. Collect a large amount of acetylation modification information from a public database, and extract local sequences of multiple lengths according to the modification sites as labeled positive example samples;

[0075] In this embodiment, a large amount of acetylation modification information is collected from a public database, and local sequences with lengths of 21, 33, and 51 are extracted according to the modification sites as labeled positive example samples.

[0076] S12. Select lysines without any labeled acetylation information from proteins containing at least one experimentally verified acetylation site as negative example samples;

[0077] S13. Intercept the local sequences around lysine residues from other proteins outside the constructed positive and negative example datasets as unlabeled samples;

[0078] S14. For all data, use the CD-HIT tool to remove proteins with a similarity exceeding 40% to avoid performance optimization caused by homology;

[0079] S15. Randomly select 10% of the samples from the labeled data as an independent test set.

[0080] S2. Design a network structure to construct a multi-scale sequence feature extraction network;

[0081] In this embodiment, the multi-scale sequence feature extraction network is constructed as follows: Use sequence features encoded in one-hot format with three different lengths as inputs. In this embodiment, the scale sequence feature extraction network uses a densely connected convolutional neural network module instead of a traditional multi-layer convolutional neural network. In this embodiment, the network in the densely connected convolutional neural network module includes: 3 densely connected modules, corresponding to protein sequence lengths of 3 different scales respectively.

[0082] In this embodiment, for the output of the i-th convolutional layer in the k-th densely connected module:

[0083]

[0084] where ∪ represents the channel concatenation operation between feature maps.

[0085] In this embodiment, since the feature maps of each layer within the module are concatenated with the feature maps of all the previous layers, if there are too many network layers in the module, it will lead to an excessive number of channels in the feature maps, thus becoming a burden on the network. Therefore, in this project, an appropriate number of convolutional layers will be selected when designing the dense connection module, and a bottleneck layer will be introduced to reduce the number of channels in the feature maps.

[0086] In this embodiment, each dense connection module consists of multiple interconnected one-dimensional convolutional layers. This dense connection enables each convolutional layer to receive the feature maps of all the previous convolutional layers, thereby enhancing the propagation of sequence features in the network. The connections between different modules will fuse the PTM features of different scales and finally be used for site prediction.

[0087] In this embodiment, the connections between the modules in the dense connection convolutional neural network fuse the high-level sequence features extracted by multiple dense connection modules and are connected according to the channels of multiple feature maps:

[0088]

[0089] where h C is the output of the last convolutional layer of the dense connection module. Finally, this feature map is converted into a one-dimensional tensor as the input to the classifier. The classifier consists of two layers of fully connected neural networks.

[0090] S3. Train the model using the unlabeled sample self-distillation method;

[0091] As Figure 4 shown, in this embodiment, the network training steps in step S3 include the following specific steps:

[0092] S31. Perform pseudo-labeling on the samples in the unlabeled target dataset D tu and select reliable samples through reliability evaluation;

[0093] In this embodiment, the self-distillation method realizes semi-supervised domain adaptation by simultaneously minimizing the inter-domain and intra-domain differences, and it is mainly achieved through two iterative steps of student sample selection and training. In the student sample set generation stage, perform pseudo-labeling on the samples in the unlabeled target dataset D tu and select reliable samples through reliability evaluation. The method of pseudo-labeling is as follows:

[0094]

[0095] where x′ ∈ D TU , is the pseudo-label of x′.

[0096] S32. Discard the unreliable samples and construct a reliable student sample set;

[0097] In this embodiment, although the pseudo-labels can be used for supervised training of unlabeled samples, in the early stage of training, the pseudo-labels are usually incorrect, so unreliable samples are discarded. Let π j represent the selection operator for selecting the j-th maximum value. This study constructs a reliable student sample set through the following method:

[0098]

[0099] where δ is a hyperparameter. In this embodiment, when the difference between the maximum value and the second maximum value of the prediction result is large enough, the sample is selected. By this method, the model only assigns pseudo-labels to samples with high confidence and generates a reliable student sample set.

[0100] S33. Randomly select samples from the labeled source domain and target domain sample sets and the student sample set to form sample pairs;

[0101] S34. Construct an auxiliary generation module to generate auxiliary features, which transfer the intermediate features of the teacher samples and the student samples and confuse the distribution of the sample pairs;

[0102] S35. Pre-train the model using the labeled data from the source domain and the target domain;

[0103] In this embodiment, to improve the credibility of the pseudo-labels, before introducing the unlabeled data, the model is pre-trained using the labeled data from the source domain and the target domain, and the cross-entropy loss function is used:

[0104] L l = E (x,y) [-log(p(y|x))] (6) Therefore, the total loss function is:

[0105] L = L l + L u + λL p (7)

[0106] where λ is a weighted hyperparameter used to control the impact of the sample pair loss on the overall training.

[0107] In this embodiment, unlabeled samples are introduced through sample self-distillation;

[0108] In this embodiment, a sample pair loss is designed to effectively bridge the difference between the student samples and the auxiliary features, and the loss function is as follows:

[0109]

[0110] Among them It is calculated when the teacher sample x passes through the feature extractor. This loss effectively reduces the inter-domain difference by using the source domain sample and the student sample pair, and effectively suppresses the intra-domain difference by using the target domain sample and the student sample pair.

[0111] In this embodiment, introducing the student sample with pseudo-label introduces an additional loss;

[0112] In this embodiment, the reliability of the pseudo-label is determined by using the latest prediction of the student sample, and it is multiplied by the cross-entropy loss of each student sample, assuming its pseudo-label is its true label, that is, using the weighted cross-entropy loss (WCE) for training the student sample:

[0113]

[0114] where α is the weight of the student sample, which can reduce the influence of the student sample with lower credibility on the model update.

[0115] S4. Use metrics such as accuracy and recall to evaluate the prediction results.

[0116] In this embodiment, the student sample set generation step is iterated alternately, and the model is updated by minimizing the total loss. By alternately iterating the student sample set generation step and the sample pairing, auxiliary generation and self-distillation steps until the model converges on the validation set. Exemplarily, the evaluation metrics are defined as follows:

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] where TP represents the total number of positive example samples correctly predicted, TN represents the total number of negative example samples correctly predicted, FP represents the total number of samples mispredicted as positive examples, and FN represents the total number of samples mispredicted as negative examples.

[0123] Example 2

[0124] As Figures 5a to 5c shown, we comprehensively evaluated the performance of the method of the present invention. The embodiment of the present invention calculated its evaluation metrics on the lysine acetylation test sets of three species and compared them with other existing prediction tools and methods. The results are as Figure 2As shown. It can be seen from the figure that the Semi-ssPTM of the present invention is not only superior to the existing general acetylation site prediction methods CapsNet-PTM and PAIL, but also higher than the species-specific acetylation site prediction method DeepDA-Ace. For example, compared with the CapsNet-PTM and PAIL methods, the present invention has greatly improved various performance indicators such as ACC value, Pre value, and F1 value on rats, yeast, and Bacillus, indicating that establishing a species-specific model is a necessary condition for accurately predicting acetylation sites in different species. In addition, the performance indicators of the present invention on multiple species have also been improved to a certain extent compared with the species-specific acetylation prediction method DeepDA-Ace, indicating that using a large amount of potential acetylation data can enhance the generalization ability of the model and further improve the prediction performance. The above results show that the semi-supervised learning method based on the sample self-distillation strategy proposed by the present invention can effectively utilize unlabeled post-translational modification data, thereby improving the prediction performance of the species-specific model on multiple species.

[0125] In summary, compared with the existing lysine modification site prediction methods, the method of the present invention can establish a specific prediction model for small-sample species. In addition, the proposed unlabeled sample self-distillation method of the present invention can effectively utilize the unlabeled data of other species, thereby enhancing the generalization ability of the model and expanding the application scope. The present invention has a good prediction effect on lysine post-translational modification and is an important basis for studying the biological functions of lysine modification. The present invention realizes the effective utilization of unlabeled data and the knowledge transfer between species, establishes a more robust and efficient species-specific prediction model, and fully utilizes a large amount of potential species-specific unlabeled data to achieve efficient prediction of lysine post-translational modification sites.

[0126] The scale sequence feature extraction network in the present invention uses a densely connected convolutional neural network module to replace the traditional multi-layer convolutional neural network. The network contains 3 densely connected modules, corresponding to 3 different scales of protein sequence lengths respectively. Each densely connected module consists of multiple interconnected one-dimensional convolutional layers. This dense connection enables each convolutional layer to receive the feature maps of all previous convolutional layers, thereby enhancing the propagation of sequence features in the network.

[0127] The present invention adopts the sample pair self-distillation method to fully utilize unlabeled post-translational modification data and make the model adapt to the target domain. Each sample pair consists of a teacher sample from the labeled dataset and a student sample from the unlabeled dataset. This method first generates assistant features by fusing the intermediate layers of the teacher and the student, and then trains the model by minimizing the output difference between the student and the assistant. During the training process, the assistant gradually bridges the difference between the two, enabling the student to easily learn from the teacher.

[0128] The sample pair loss of the present invention effectively reduces the inter-domain difference by using the source domain sample and the student sample pair, and effectively suppresses the intra-domain difference by using the target domain sample and the student sample pair.

[0129] The present invention introduces an additional loss by using the student sample with a pseudo-label. By using the latest prediction of the student sample to determine the reliability of the pseudo-label and multiplying it by the cross-entropy loss of each student sample, the student sample is trained with a weighted cross-entropy loss, which can reduce the impact of less reliable student samples on model update. The present invention solves the technical problems existing in the prior art, such as the ineffective utilization of unlabeled data, the difficulty of knowledge transfer between species, and the low robustness of species-specific prediction models.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semi-supervised method for predicting protein lysine post-translational modification sites, characterized in that, The method includes: S1. Collect acetylation modification information, extract labeled samples from pre-set acetylated site proteins to obtain labeled source domain and target domain datasets, intercept unlabeled samples from proteins other than the pre-set acetylated site proteins, conduct data collection and preprocessing, and select samples from the labeled samples to construct an independent test set; S2. Use one-hot encoded sequence features of no less than two different lengths as inputs, select applicable numbers of convolutional layers and bottleneck layer channels, set up dense connection modules, and accordingly construct a multi-scale sequence feature extraction network model; S3. According to the unlabeled samples, obtain an unlabeled target dataset, and accordingly use the self-distillation method to pseudo-label the unlabeled samples in the unlabeled target dataset to obtain pseudo-labeled samples, conduct reliable evaluation on the pseudo-labeled samples to obtain reliable student samples, form sample pairs based on the labeled source domain and target domain datasets and the reliable student samples, and calculate the loss of the sample pairs, and accordingly use the labeled source domain and target domain datasets to pre-train the multi-scale sequence feature extraction network model to obtain an applicable model; S4. Use the applicable model to extract and fuse multi-scale high-level sequence features according to the one-hot encoded sequence features to obtain multi-scale fusion features, and predict modification sites based on the multi-scale fusion features to obtain site prediction results; S5. Use the independent test set to test the applicable model, and evaluate the site prediction results using accuracy and recall rate.

2. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 1, wherein The step S1 includes: S11. Collect the acetylation modification information from a public database, and extract local sequences of no less than two lengths according to the pre-set modification sites as labeled positive example samples; S12. Select lysines without labeled acetylation information from the proteins with verified acetylated sites as negative example samples; S13. Intercept local sequences around lysine residues from other proteins outside the constructed positive and negative example datasets as the unlabeled samples; S14. Remove the proteins with similarity exceeding the pre-set similarity threshold; S15. Randomly select a pre-set proportion of the labeled samples from the labeled data as the independent test set.

3. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 1, wherein In the step S2: In the k-th dense connection module, the output of the i-th convolutional layer is: where represents the output of the i-th convolutional layer in the k-th densely connected module, and W i k represents the convolutional kernel weights of the i-th convolutional layer, represents the activation function, and ∪ represents the channel concatenation operation between feature maps.

4. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 1, wherein In the step S2, in the dense connection convolutional neural network, use inter-module connections to fuse the high-level sequence features extracted by no less than two dense connection modules, and use the following logic to connect according to the channels of multiple feature maps: where h C is the output of the last convolutional layer of the dense connection module, and h f is the feature map obtained by concatenating the outputs of multiple dense connection modules by channel.

5. The semi-supervised method for predicting post-translational modification sites of protein lysine according to claim 1, wherein The step S3 includes: S31. Pseudo-label the unlabeled samples and select the reliable student samples through reliability evaluation to construct a reliable student sample set; S33. Select applicable samples from the labeled source domain D s and the target domain D TL data set and the reliable student sample set to form the sample pair; S34. Construct an auxiliary generation module to generate auxiliary features to confuse the distribution of the sample pairs; S35. Utilize the labeled source domain and target domain datasets Process to obtain the cross-entropy loss, and process the sample pairs through sample self-distillation to obtain the sample pair loss; S36. Use the reliable student samples with pseudo-labels to calculate an additional loss, and accordingly calculate the total loss by combining the cross-entropy loss and the sample pair loss, and accordingly pre-train the multi-scale sequence feature extraction network model.

6. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 5, characterized in that, In the step S31, the following logic is used to perform pseudo-labeling on the unlabeled samples: where \(x\) ′ \(\in D\) TU and \(D\) TU is an unlabeled dataset in the target domain, is the pseudo-label of \(x\) ′ .

7. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 5, wherein In the step S33, the reliable student sample set is constructed by using the following logic where \(g\) represents the feature extractor, \(h\) represents the classifier, and \(\pi\) j represents the selection operator for selecting the \(j\)-th maximum value, and \(\delta\) is a hyperparameter.

8. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 5, wherein The step S35 includes: S351. Using the following logic to obtain the cross-entropy loss: L l = E (x,y) [-log(p(y|x))] (6) In the formula, E (x,y) represents the mathematical expectation; S352. Using the following logic to obtain the sample pair loss: where D KL represents the KL divergence, which is calculated when the teacher sample x passes through the feature extractor.

9. The semi-supervised method for predicting protein lysine post-translational modification sites according to claim 8, characterized in that, In the step S352, the following logic is used to calculate the weighted cross-entropy loss according to the sample pair loss technology: In the formula, α represents the weight of the student sample.

10. Semi-supervised protein lysine post-translational modification site prediction system, characterized in that, The system includes: A sample set establishment module, which is used to collect acetylation modification information, extract labeled samples from pre-set acetylation site proteins to obtain labeled source domain and target domain data sets, intercept unlabeled samples from proteins other than the pre-set acetylation site proteins, perform data collection and preprocessing, and select samples from the labeled samples to construct an independent test set; A model construction module, which is used to use one-hot encoded sequence features of not less than 2 different lengths as inputs, select the applicable number of convolutional layers and the number of channels in the bottleneck layer, set up a densely connected module, and accordingly construct a multi-scale sequence feature extraction network model; A model training module, which is used to obtain an unlabeled target data set according to the unlabeled samples, and accordingly use the self-distillation method to pseudo-label the unlabeled samples in the unlabeled target data set to obtain pseudo-labeled samples, perform a reliable evaluation on the pseudo-labeled samples to obtain reliable student samples, form sample pairs according to the labeled source domain and target domain data sets and the reliable student samples, and calculate the sample pair loss, and accordingly use the labeled source domain and target domain data sets to pre-train the multi-scale sequence feature extraction network model to obtain an applicable model. The model training module is connected to the model construction module and the sample set establishment module; A site prediction module, which is used to use the applicable model to extract and fuse multi-scale high-level sequence features according to the one-hot encoded sequence features to obtain multi-scale fusion features, and predict the modification site according to the multi-scale fusion features to obtain a site prediction result. The site prediction module is connected to the model training module; An evaluation and testing module, which is used to test the applicable model using the independent test set, and evaluate the site prediction result using accuracy and recall. The evaluation and testing module is connected to the sample set establishment module and the site prediction module.

Images