A method for predicting antifreeze protein typing based on transfer learning
Through the BERT-DomainAFP model based on transfer learning, combined with domain annotation and data balancing technology, the accuracy problem of the antifreeze protein typing prediction model under insufficient sample size and diversity complexity was solved, and efficient and accurate antifreeze protein identification and typing was achieved.
Patent Information
- Application Number
- CN202410678242.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-29
- Publication Date
- 2025-10-03
- Estimated Expiration
- 2044-05-29
AI Technical Summary
The performance and accuracy of existing antifreeze protein typing prediction models cannot meet the requirements when the number of training samples is small, and traditional methods cannot effectively handle the diversity and complexity of protein sequences, resulting in limited model generalization ability and prediction accuracy.
The BERT-DomainAFP model based on transfer learning was used to extract features through the ProteinBert model, and domain annotations were used as training labels. Combined with oversampling and undersampling techniques, a deep neural network model was constructed to achieve the identification and typing of antifreeze proteins.
The accuracy, specificity and sensitivity of antifreeze protein identification and prediction have been improved. The model can maintain high predictive ability even on unbalanced data sets, has good scalability and versatility, and is suitable for large-scale bioinformatics data analysis.
Smart Images
Figure CN118447928B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of bioinformatics, and in particular relates to a method for predicting antifreeze protein typing based on transfer learning. Background Art
[0002] Antifreeze proteins (AFPs) are proteins widely found in plants, cold-water fish, insects, microorganisms, and bacteria. In organisms, when water molecules condense into ice crystals under low temperatures, these ice crystals can damage cells and tissues. The primary function of AFPs is to bind to ice crystals in organisms at low temperatures and modify their growth properties, thereby preventing them from damaging cells and tissues. By interacting with ice crystals, these proteins lower the freezing point, inhibit their growth rate, and alter their morphology, thereby protecting the integrity of cells and tissues at low temperatures. In the food and pharmaceutical sectors, due to the widespread use of freezing techniques, the nucleation and recrystallization of ice crystals often cause structural damage to frozen items, such as medicines, foods, and even organs, after thawing. Therefore, in-depth research on AFPs holds the promise of providing a safe and reliable method for the cryopreservation of these organisms. By leveraging the properties of AFPs, it is possible to effectively mitigate or prevent ice crystal damage to biological tissues during freezing, ensuring that frozen items retain their original structure and function after thawing. Currently known antifreeze proteins in fish are primarily classified as types I-IV. Insect antifreeze proteins and antifreeze glycoproteins (AFGPs) also exist. Among these AFPs, type I AFPs are alanine-rich and have an α-helical geometry. Winter flounder AFPs have been shown to interact with the ice plane as a pyramidal plane, while type I binds to the secondary prismatic plane of ice. Type II is the largest AFP known to exist in fish. Type II AFPs contain a large number of globulins and bind to the pyramidal planes of ice crystals to inhibit ice crystal formation. Type III binds to both the pyramidal and prismatic planes of ice crystals. Type IV AFPs, related to the longhorn sturgeon, do not bind to the ice plane. AFGPs are considered highly conserved and include eight classes, from AFGP-I to AFGP-VIII.
[0003] Traditional methods for identifying functional genes rely on sequence alignment to establish homology. However, due to the significant differences in the mechanisms of action and species distribution of various AFP types, AFPs exhibit significant structural and sequence heterogeneity. Comprehensive analysis of AFPs is impractical using traditional bioinformatics methods. Currently, multiple models for identifying AFPs have been proposed. Feature-based machine learning models have achieved good predictive accuracy for antifreeze proteins. These models utilize four predefined features to collect local and global descriptors and evaluate feature vectors using machine learning classifiers based on support vector machines and random forest algorithms. Compared to machine learning models based on features such as homology, rules, and probabilities, another strategy for constructing learning-based AFP prediction models is to utilize deep learning techniques. Deep learning, which constructs deep neural network models to automatically learn and extract useful information from data, has achieved remarkable results in protein structure prediction, enabling accurate prediction of unknown protein structures. However, due to the complexity and diversity of protein sequences, training high-performance deep learning models remains a challenge. This is particularly true when data is scarce or insufficiently labeled, often limiting the generalization and prediction accuracy of these models. To overcome these challenges, transfer learning has been introduced to the field of protein domain prediction, a technique that leverages knowledge learned in a source domain to aid learning in a target domain. In protein domain prediction, deep learning models pre-trained on large-scale datasets can be used as source domain models, and their learned feature extraction and classification capabilities can be transferred to the target domain of protein domain prediction through transfer learning. Specifically, encoded and pre-processed protein sequence data is input into a deep convolutional neural network model for training, enabling the model to automatically learn and extract key features from the sequence. In this way, the model can learn from the data and predict the presence of AFPs without the need for manual feature selection.
[0004] Model building for AFP sequence recognition primarily focuses on binary sequence classification tasks, but no research has yet been conducted on classification models for antifreeze proteins across different AFP types. Furthermore, most models are based on models trained from scratch, resulting in suboptimal performance and accuracy when the number of training samples is small. Multi-label classification models for protein sequences have been extensively studied. Currently, the primary model for multi-label classification of protein sequences uses GO annotation information as training samples. These models use a large number of deduplicated protein sequences as training samples, train on thousands of functional annotations, and are applied to the identification of the HKT gene in Spartina alterniflora. Based on transfer learning using the ProteinBERT model, the present invention introduces annotations of antifreeze protein domains as training labels for the model. Domain annotations are performed on the existing antifreeze protein training set, and a sequence set containing antifreeze protein domains, AntiFreezeDomains, is integrated. Finally, 27 domains commonly found in antifreeze proteins are selected as training labels to establish the BERT-DomainAFP antifreeze protein recognition model, which integrates both antifreeze protein recognition and typing. Summary of the Invention
[0005] To address the above issues, the present invention provides a method for predicting antifreeze protein typing based on transfer learning. By establishing a BERT-DomainAFP model to identify and type antifreeze proteins, the model has high precision, high accuracy, good specificity, high sensitivity, high prediction efficiency and good stability in identifying and predicting antifreeze proteins.
[0006] The present invention is achieved through the following technical solutions:
[0007] A method for predicting antifreeze protein typing based on transfer learning, comprising the following steps:
[0008] S1. Collect domain-annotated antifreeze protein sequences and non-antifreeze protein sequences as positive and negative datasets, and use the domain keyword frequencies of all antifreeze protein sequences as encoding labels to extract all antifreeze protein-related domains with an occurrence frequency greater than 10.
[0009] S2. Based on the balance between the positive and negative datasets and the number of antifreeze protein sequences, oversampling and undersampling methods are used to balance the data, and the domains in the dataset that reach the set domain frequency are used as dataset labels;
[0010] S3. Select the ProteinBert model to extract features from the antifreeze protein sequence dataset to convert them into high-dimensional feature vectors, which are then input into the model for training to establish a correspondence between the antifreeze protein sequences and the domain labels, thereby constructing the BERT-DomainAFP model.
[0011] S4. Input the protein sequence data to be identified or the integrated species protein sequence data into the trained BERT-DomainAFP model for prediction to obtain the corresponding domain prediction value;
[0012] S5. Set the target data label. If the domain prediction value output by the BERT-DomainAFP model is the same as the target data label, it is determined that the antifreeze protein sequence domain exists in the protein sequence to be identified.
[0013] Furthermore, in step S1, the antifreeze protein sequences are divided into five main tags according to the structural domains, namely type I antifreeze protein, type II antifreeze protein, type III antifreeze protein, insect antifreeze protein and other antifreeze protein related domains.
[0014] Furthermore, in step S1, the negative dataset uses reviewed non-antifreeze protein sequences searched in UniProtKB, wherein the ratio of the negative dataset with domain annotations to the negative dataset without domain annotations is 1:19.
[0015] Furthermore, in step S2, the data set is divided into a training set, a validation set, and a test set, with the ratio set to 7:1:2.
[0016] Furthermore, in step S2, the data balance is performed by oversampling and undersampling methods as follows: oversampling type I antifreeze proteins, type II antifreeze proteins, and insect antifreeze proteins with less than 150 samples, and undersampling type III antifreeze proteins and other antifreeze protein-related domains with more than 150 samples, to ensure a balance of various types of antifreeze proteins.
[0017] Furthermore, in step S2, the dataset label is encoded into a one-dimensional vector with a length threshold of 1024, using integer numbers corresponding to 20 proteins as encoding tags, and for sequences with insufficient length, 0 padding is used for processing.
[0018] Furthermore, in step S3, the architecture of the ProteinBert model adopts a Transformer structure with 6 attention heads and 4 layers of depth.
[0019] Furthermore, in step S3, the ProteinBert model uses Gaussian error linear unit (GELU) as an activation function; introduces a wide convolution layer and a narrow convolution layer with a dilation rate of 5, and a convolution kernel size of 9; and uses a linear layer to output the predicted value of the model.
[0020] Furthermore, step S3 also includes verifying the ProteinBert model by repeated random subsampling. The accuracy, sensitivity, specificity, Matthews correlation coefficient, precision and recall are mainly used for evaluation during the verification process.
[0021] Furthermore, the validation of the ProteinBert model was divided into two parts. One part was to regard the model as a binary classification model, in which the data set ratio of antifreeze proteins and non-antifreeze proteins was 1:1; the other part was to regard the model as a multi-label classification model, in which the data set ratio of non-antifreeze proteins, type I antifreeze proteins, type II antifreeze proteins, type III antifreeze proteins, insect antifreeze proteins and other antifreeze protein-related sequences was 5:1:1:1:1:1:1.
[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are:
[0023] 1. The BERT-DomainAFP model of the present invention has strong feature extraction capabilities, automatically extracting complex feature information from protein sequences, eliminating the tedious manual feature design and selection process required in traditional methods. Furthermore, the model achieved an accuracy of 98.48%, a specificity of 99.80%, and a sensitivity of 99.70% for antifreeze protein prediction, demonstrating high accuracy, specificity, sensitivity, and precision. By constructing a deep neural network structure, the present invention is able to learn and capture high-level, abstract features in protein sequences, thereby improving the efficiency, accuracy, and reliability of antifreeze protein identification and prediction.
[0024] 2. The present invention effectively improves model training speed and prediction accuracy through transfer learning and the use of the BERT model. The BERT-DomainAFP model of the present invention utilizes transfer learning to extend the existing pre-trained ProteinBERT model, significantly reducing the computational resources required for training, accelerating model convergence, and speeding up the model development and deployment process. The integration of the complex deep learning architecture of the ProteinBERT model endows the BERT-DomainAFP model with excellent prediction accuracy, enabling it to expertly capture complex patterns and dependencies within protein sequences, thereby improving the accuracy of antifreeze protein recognition and classification.
[0025] 3. This study applies domain annotation to antifreeze protein typing, introducing a new approach for the BERT-DomainAFP model. Using domain annotations as labels, this approach enables the model to distinguish and classify antifreeze proteins. This approach not only allows the model to predict the presence of antifreeze proteins but also to classify them into specific types based on their unique domain signatures, providing a deeper understanding of their properties and functions.
[0026] 4. This invention enhances its ability to handle unbalanced datasets through oversampling and undersampling. To address data imbalance, the BERT-DomainAFP model undergoes multiple undersampling and oversampling steps during its development process, ensuring that it fully learns the characteristics of various data sets during training. Even when tested on more unbalanced datasets, the model's predictive power remains high, demonstrating that it is not only suitable for balanced datasets but can also effectively address the challenges of unbalanced datasets.
[0027] 5. The BERT-DomainAFP model of the present invention can effectively process and analyze large-scale bioinformatics data, improving prediction efficiency, accuracy, and stability. It also exhibits excellent scalability and versatility, adapting to various protein sequence and domain prediction tasks, and possesses significant application value in the field of antifreeze protein research. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is a flow chart for the prediction of antifreeze protein typing according to the present invention.
[0029] Figure 2 This is a diagram of the antifreeze protein typing prediction model of the present invention.
[0030] Figure 3 Accuracy plot of multiple models built for repeated random subsampling in this invention.
[0031] Figure 4 This is a diagram showing the classification evaluation of multiple types of antifreeze proteins by the BERT-DomainAFP model of the present invention. DETAILED DESCRIPTION
[0032] The present invention is further described in detail below through examples. These examples are only used to illustrate the present invention and do not limit the scope of protection of the present invention.
[0033] Example
[0034] The flow chart and model diagram of the antifreeze protein typing prediction method based on transfer learning of the present invention are respectively as follows: Figure 1 and 2 shown.
[0035] The method specifically comprises the following steps:
[0036] S1. This study combined 487 AFP sequences from DatasetR with 771 sequences found in the Interpro database with a sequence identity of <30% using CD-HIT, using fields such as "Antifreeze." To construct a dataset, AntiFreezeDomains, containing 1094 antifreeze protein-positive sequences, was constructed. The dataset was then input into the Interpro database in batches of 100 sequences for domain annotation. Domain keyword frequencies across all antifreeze protein sequences were used as encoding labels, and all antifreeze protein-related domains with an occurrence frequency greater than 10 were extracted. The antifreeze sequences were categorized into five primary domain labels: type I antifreeze protein (AFP_I), type II antifreeze protein (AFP_II), type III antifreeze protein (AFP_III), insect antifreeze protein (AFP_Insect), and other antifreeze protein-associated (AFP_Associate). Twenty-two domain terms that frequently appear in antifreeze protein sequences were also added:
[0037] Type antifreeze protein signature, SignalP, terminal region of a signal peptide, Hydrophobic region of a signal peptide, Extramembrane Region, Extramembrane Region_Ⅱ, Membrane-embedded Region_Ⅱ, C-type, C-type lectin 2 (CLECT_2), consensus disorder prediction, MANNOSE, PHOSPHOLIPASE, LECTINRECEPTOR RELATED, Antifreeze_Ⅲ, Antifreeze protein-like domain profile, Safety Assessment Factor (SAF), NeuB family, Aldolase, Sialic acid synthase The 22 domain terms were used as secondary labels and finally encoded using a dictionary. The resulting encoded labels were a one-dimensional vector of length 27, which was used for subsequent model prediction.
[0038] The negative dataset consisted of 9,637 reviewed non-antifreeze protein sequences searched from UniProtKB. Each sequence in the negative dataset was similarly divided into batches of 100 for domain prediction and encoded as labels. For 27 domains, most of the negative data failed to obtain any domain annotations, while the small amount of negative data that did have domain annotations did not contain direct antifreeze protein domains. The ratio of the negative dataset with domain annotations to the negative dataset without domain annotations was 1:19. Finally, the training and test sets were divided for testing, with a training set, validation set, and test set ratio of 7:1:2.
[0039] S2. For protein sequence data, encode it into a one-dimensional vector with a length threshold of 1024, where the encoding form of each vector is a 1-20-bit integer number corresponding to 20 proteins as a tag sequence. Amino acid sequences with a length of less than 1024 are padded with 0s. For amino acid sequences with a length of more than 1024, they are split into multiple sequences with a length of less than 1024 according to the domain range, and their domain labels are updated for use as multiple data.
[0040] Due to the severe imbalance in the number of positive and negative datasets and antifreeze protein sequences, the transfer learning on the training model adopts an oversampling and undersampling strategy. The steps are to oversample the minority antifreeze protein samples and undersample the majority AFP categories to ensure a balance of various types of antifreeze proteins. Specifically:
[0041] The training dataset contains 750 positive antifreeze protein sequences, so the ideal number of sequences for each antifreeze protein class is 150. We oversampled antifreeze protein classes with fewer than 150 samples (type I antifreeze proteins, type II antifreeze proteins, and insect antifreeze proteins), and undersampled antifreeze protein classes with more than 150 samples (type III antifreeze proteins and other antifreeze protein-related domains). The small-batch random oversampling method follows these steps:
[0042] (1) Determine the mini-batch size: Choose an appropriate mini-batch size B, which is the number of data samples used in each training iteration of the model. (2) Create mini-batches: For each mini-batch: Randomly select samples from the minority class until the mini-batch size reaches half, B / 2. And randomly select the remaining B / 2 samples from all classes. (3) Repeat the process: Repeat the above process for each mini-batch in the training dataset. (4) Train the model: Train the model using the created mini-batches.
[0043] Finally, balanced antifreeze protein data and new randomly sampled non-antifreeze protein data are used for training in each training round to ensure that the model is constantly exposed to balanced antifreeze protein samples during training, and to allow the model to extract more features from non-antifreeze protein sequences during training.
[0044] S3. The present invention selects ProteinBert, a protein sequence pre-training model based on the Transformer architecture, which is pre-trained on 106M proteins derived from UniProtKB / UniRef90. The model labels are as many as 8943 annotation labels, and the protein sequences cover the entire tree of life. The deep learning model architecture of the present invention is trained on a pre-trained model with 6 attention heads and 4 layers, and during the training process, the original labels are randomly modified with a probability of 5%, annotations are randomly added with a probability of 1%, and annotation information is randomly added with a probability of 25% to enhance the model's ability to handle incompletely annotated data and prevent the model from overfitting. All hidden layers and linear layers in the model use GELU (Gaussian error linear unit), and the convolution kernel size used for the wide convolution layer (expansion rate = 5) and the narrow convolution layer is 9. Finally, the linear layer is used directly to output the prediction result of the model, and the classification threshold is manually set. During the training process, the default parameters of the pre-training model are used for most of the other parameters. Due to the difference between the training task and the training data set, the model reduces the complexity and number of layers. The model uses multi-label classification cross entropy as the loss function, and its formula is:
[0045] Where y is the true label value of the sample, and p is the predicted label value of the sample.
[0046] Due to the imbalanced nature of data between different types of antifreeze proteins and between antifreeze proteins and non-antifreeze protein data, it is difficult to uniformly perform K-fold cross-validation on all types of antifreeze protein data during training. The present invention uses a repeated random subsampling method to validate the model. The validation first randomly divides the original dataset into training, validation, and test sets multiple times. Antifreeze protein prediction models are then simultaneously established for each of the multiple, differently divided datasets, and each model is evaluated to eliminate model contingency. The Dataset-domain is randomly sampled and divided ten times, each dataset containing an equal number of training, test, and validation sets. The model with the highest prediction index among the ten models is selected as the BERT-DomainAFP result model. Each dataset is named train_1-10, and a model is then established and validated on each dataset. Model validation primarily uses accuracy, sensitivity, specificity, Matthews correlation coefficient (MCC), precision, and recall to evaluate the model. The calculation formulas for these indicators are:
[0047]
[0048] Among the above indicators, TN, TP, FN and FP are true negative, true positive, false positive and false negative, respectively.
[0049] The validation of the model is divided into two parts. One part is to regard the model as a binary classification model and calculate the prediction accuracy of antifreeze protein sequences and non-antifreeze protein sequences. The ratio of antifreeze protein and non-antifreeze protein in the test set is 1:1. The second part is to regard the model as a multi-label classification model. The main classification categories are non-antifreeze protein, antifreeze protein type I, antifreeze protein type II, antifreeze protein type III, insect antifreeze protein and antifreeze protein-related sequences. The ratio of the above data sets is 5:1:1:1:1:1:1.
[0050] Figure 3 The multi-model accuracy graph for repeated random subsampling shows that when the training on the pre-trained model reaches 10 rounds, the average model accuracy reaches 97%. Each model is validated on the test set using the model with the highest accuracy saved during the training process. The final average accuracy on the test set is 98.48%. Since the accuracy comes from the test results of multiple different training and test sets, the high accuracy of the model caused by the accidental data set is eliminated. The main evaluation indicators of the model are shown in Table 1:
[0051] Table 1 Main evaluation indicators of the model in training for 10 batches of randomly divided datasets
[0052]
[0053]
[0054] The prediction accuracy of each antifreeze protein type was evaluated, and the ROC curve of each antifreeze protein type prediction was drawn, which is a comprehensive indicator of the sensitivity and specificity continuous variables of the response model for each antifreeze protein type ( Figure 4 a), the ROC curves of AFP_Ⅲ, AFP_Insect, AFP_Ⅱ, AFP_Ⅰ and AFP_Associate are highly overlapped, and the area under the curve is almost close to the maximum area 1. The higher the value of the area under the ROC curve (AUC), the better the classification ability of the prediction model. Figure 4 b shows the average accuracy of the model's ability to identify each type of antifreeze protein, with the identification accuracy rates for AFP_Ⅲ, AFP_Insect, AFP_Ⅱ, AFP_Ⅰ and AFP_Associate being 99.75, 90.2, 99.5, 97.5 and 98.5%, respectively.
[0055] Based on the training results of the above models, the train_1 training batch with the highest accuracy was selected as the model predictor, and the antifreeze protein recognition model BERT-DomainAFP, which integrates antifreeze protein recognition and typing, was established. The BERT-DomainAFP model was compared with the existing models in terms of accuracy, specificity, sensitivity, and Matthews correlation coefficient (MCC). The comparison results are shown in Table 2:
[0056] Table 2 Comparison of prediction indicators of BERT-DomainAFP model with other models
[0057] Model Classifier Accuracy (%) Specificity (%) Sensitivity (%) MCC (%) iAFP Support Vector Machine 95.60 97.23 9.91 8.00 AFP-Pred RF 77.34 77.04 91.16 23.00 AFP-PSSM Support Vector Machine 93.01 93.20 75.80 34.00 AFP-PseAAC Support Vector Machine 84.75 84.74 85.08 27.00 afpCOOL Support Vector Machine 96.00 98.00 0.72 -- AFP-CKSAAP DNN 88.00 87.80 0.94 -- AFP-CMBPred Support Vector Machine 91.65 91.98 75.14 32.00 Wang et al. DT 91.33 -- -- 82.00 AFP-LSE Autoencoder 88.17 91.52 84.83 46.00 TargetFreeze Support Vector Machine 90.95 91.78 90.11 81.00 BERT-DomainAFP BERT 98.48 99.80 99.79 95.49
[0058] As shown in Table 2, the model of the present invention achieved a prediction accuracy of 98.48%, a specificity of 99.80%, a sensitivity of 99.70%, and an MCC of 95.4%, all of which are superior to existing models. This indicates that the BERT-DomainAFP model constructed in the present invention exhibits higher accuracy, specificity, and sensitivity for antifreeze protein identification and typing, demonstrating superior performance in both the identification and typing of antifreeze proteins.
[0059] S4. Encode the protein sequence data to be identified into the same data format as in step S2 and input it into the BERT-DomainAFP model for prediction to obtain its corresponding domain prediction value. The sequence in Table 4 is the input data. After obtaining the domain prediction value corresponding to the sequence, the domain prediction value is input into Table 3 for query to obtain the entry corresponding to the domain prediction value, and the prediction results are classified according to the entry;
[0060] Table 3 Query table of protein sequence domain terms to be identified
[0061]
[0062]
[0063] Table 4 Prediction values of structural domains corresponding to the protein sequences to be identified
[0064]
[0065] S5. Set the output positive terms as target data labels, including the related terms of AFP_Ⅲ, AFP_Insect, AFP_Ⅱ, AFP_Ⅰ and AFP_Associate, as the main antifreeze protein categories, as shown in Table 5:
[0066] Table 5: Positive terms for the final prediction results of the protein sequence to be identified
[0067]
[0068]
[0069] The main categories of positive entries corresponding to the domain prediction values output by the BERT-DomainAFP model are classified into five antifreeze protein categories: AFP_Ⅲ, AFP_Insect, AFP_Ⅱ, AFP_Ⅰ, and AFP_Associate. The annotation results of the domain prediction values finally output by the protein sequence to be identified are shown in Table 6:
[0070] Table 6 Final prediction results of protein sequences to be identified
[0071]
[0072] Table 6 shows the final prediction results for the protein sequence to be identified. The output is divided into three parts: gene number, gene sequence, and target data label. The gene number and gene sequence are the initial data input to the model. The target data label is the prediction result of the BERT-DomainAFP model and is a structural domain related to antifreeze protein. This indicates that the protein sequence to be identified contains an antifreeze protein sequence domain.
[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for predicting antifreeze protein typing based on transfer learning, characterized in that: The following steps are involved: S1. Collect domain-annotated antifreeze protein sequences and non-antifreeze protein sequences as positive and negative datasets, and use the domain keyword frequencies of all antifreeze protein sequences as encoding labels to extract all antifreeze protein-related domains with an occurrence frequency greater than 10. S2. Based on the balance between the positive and negative datasets and the number of antifreeze protein sequences, oversampling and undersampling methods are used to balance the data, and the domains in the dataset that reach the set domain frequency are used as dataset labels; S3. Select the ProteinBert model to extract features from the antifreeze protein sequence dataset to convert them into high-dimensional feature vectors, which are then input into the model for training to establish a correspondence between the antifreeze protein sequences and the domain labels, thereby constructing the BERT-DomainAFP model. The architecture of the ProteinBert model adopts a Transformer structure with 6 attention heads and 4 layers of depth; The ProteinBert model uses Gaussian error linear unit (GELU) as the activation function, and introduces wide convolutional layers and narrow convolutional layers with a dilation rate of 5 and a convolution kernel size of 9. And use the linear layer to output the model's predicted value; The validation of the ProteinBert model was divided into two parts. One part was to treat the model as a binary classification model, in which the ratio of the antifreeze protein and non-antifreeze protein datasets was 1:1; the other part was to treat the model as a multi-label classification model, in which the ratio of the datasets of non-antifreeze protein, type I antifreeze protein, type II antifreeze protein, type III antifreeze protein, insect antifreeze protein and other antifreeze protein-related sequences was 5:1:1:1:1:
1. S4. Input the protein sequence data to be identified or the integrated species protein sequence data into the trained BERT-DomainAFP model for prediction to obtain the corresponding domain prediction value; S5. Set the target data label. If the domain prediction value output by the BERT-DomainAFP model is the same as the target data label, it is determined that the antifreeze protein sequence domain exists in the protein sequence to be identified.
2. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: In step S1, the antifreeze protein sequences are divided into five main tags according to the structural domains: type I antifreeze protein, type II antifreeze protein, type III antifreeze protein, insect antifreeze protein and other antifreeze protein related domains.
3. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: In step S1, the negative dataset uses reviewed non-antifreeze protein sequences searched in UniProtKB, wherein the ratio of negative datasets with domain annotations to negative datasets without domain annotations is 1:
19.
4. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: In step S1, the data set is divided into a training set, a validation set, and a test set, with the ratio set to 7:1:
2.
5. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: In step S2, the data balance is performed by oversampling and undersampling methods, specifically: oversampling type I antifreeze proteins, type II antifreeze proteins, and insect antifreeze proteins with less than 150 samples, and undersampling type III antifreeze proteins and other antifreeze protein-related domains with more than 150 samples, to ensure a balance of various types of antifreeze proteins.
6. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: In step S2, the dataset label is encoded into a one-dimensional vector with a length threshold of 1024, using integer numbers corresponding to 20 proteins as encoding tags, and for sequences with insufficient length, zero padding is used for processing.
7. The antifreeze protein typing prediction method based on transfer learning according to claim 1, characterized in that: Step S3 also includes verifying the ProteinBert model by repeated random subsampling. The accuracy, sensitivity, specificity, Matthews correlation coefficient, precision and recall are mainly used for evaluation during the verification process.