A machine learning-based viral rna silencing suppressor identification method
By using machine learning-based methods, an identification model for viral RNA silencing inhibitors was constructed using ESM-2 and XGBoost. This solved the problem of identifying viral RNA silencing inhibitors in traditional methods, achieving high-precision and high-efficiency identification, and is suitable for large-scale screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional methods are difficult to effectively identify viral RNA silencing repressors, especially since repressors from different viral sources lack conserved homologous sequences or domains in terms of sequence and structure, making it difficult for bioinformatics methods to extrapolate between different viruses.
A machine learning-based approach was adopted, using the ESM-2 model to extract deep features of protein sequences and then identifying them using the XGBoost classifier. Parameter optimization was performed by combining class weight adjustment and 10-fold cross-validation to construct an identification model for viral RNA silencing inhibitors.
It improves the accuracy and reliability of identifying viral RNA silencing inhibitors, shortens the research cycle, reduces R&D costs, is suitable for large-scale screening, and has strong model stability and generalization ability.
Smart Images
Figure CN122337342A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of bioinformatics and protein function identification, specifically relating to a method for identifying viral RNA silencing inhibitors based on machine learning. Background Technology
[0002] RNA silencing is a crucial defense mechanism for plants and animals against viral infection. To counteract host defense, viruses encode and express protein factors in their genomes that specifically inhibit the host's RNA silencing mechanism; these protein factors are called viral RNA silencing repressors. These repressors interfere with or block transgene-induced post-transcriptional gene silencing through diverse mechanisms at different stages of the RNA silencing pathway, thereby removing host restrictions on transgene expression and serving as a tool to enhance transgene expression efficiency in plants.
[0003] Current identification of viral RNA silencing repressors primarily relies on traditional experimental methods. However, these repressors exhibit high diversity in both sequence and structure, and repressors from different viral origins often lack conserved homologous sequences or domains. Therefore, traditional bioinformatics methods based on sequence similarity or conserved domains are difficult to apply effectively to different viruses. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides a method for identifying viral RNA silencing inhibitors based on machine learning, comprising:
[0005] Step 1: Obtain training data from the Uniprot database. The training data includes viral protein sequences with RNA silencing and inhibition functions and viral protein sequences without RNA silencing and inhibition functions.
[0006] Step 2: Label viral protein sequences with RNA silencing inhibition function as positive samples and viral protein sequences without this function as negative samples.
[0007] Step 3: Perform redundancy removal on the protein sequences in the training data;
[0008] Step 4: Input the deredundant protein sequence into the pre-trained ESM-2 model and extract the feature vector corresponding to the [CLS] label output by the model as the initial feature representation of the protein sequence.
[0009] Step 5: Divide the dataset corresponding to the initial feature representation into a training set, a validation set, and a test set. Use the training set to train multiple candidate classifiers, and optimize the parameters during the training process by adjusting the class weights and combining 10-fold cross-validation with grid search.
[0010] Step 6: Use the validation set to evaluate the performance metrics of each candidate classifier, and use the test set to verify the generalization ability of the classifier with the best performance. Select the XGBoost classifier that performs best on the test set as the final recognition model.
[0011] Step 7: Input the amino acid sequence of the viral protein to be tested into the functional prediction model, and output the identification result of whether the protein is a viral RNA silencing inhibitor.
[0012] The beneficial effects of this invention are:
[0013] (1) High prediction accuracy and strong generalization ability. This invention cleverly combines the powerful feature extraction capability of the ESM-2 large-scale protein language model with the excellent tabular data processing performance of the XGBoost classifier. ESM-2 can capture deep evolutionary and structural information in protein sequences through self-supervised learning and output it in the form of high-dimensional vectors; XGBoost can efficiently process these features and perform accurate classification. By comparing multiple classic classifiers (such as AdaBoost, RF, DNN, etc.) and selecting XGBoost, the best performance of the model on the test set is ensured, which significantly improves the accuracy and reliability of identifying viral RNA silencing inhibitors.
[0014] (2) High prediction efficiency, suitable for large-scale screening. Compared with traditional biological identification methods that rely on in vivo or in vitro experiments, this invention is based entirely on a computational model. Once the model is trained, only the amino acid sequence of the viral protein (e.g., FASTA format) needs to be input, and the prediction results can be output in a very short time, which greatly shortens the research cycle, reduces the R&D cost, and is suitable for high-throughput functional annotation and screening of proteins at the genome scale.
[0015] (3) The design of the scheme is scientific and reasonable, and has good robustness. In the data processing stage, the present invention sets a 70% sequence similarity threshold for redundancy removal, avoiding data leakage and model overfitting caused by high sequence similarity; in the model training stage, a class weight adjustment strategy is introduced to effectively deal with the actual situation of imbalance between positive and negative samples; at the same time, 10-fold cross-validation combined with grid search is used for hyperparameter optimization, which further improves the stability and generalization ability of the model. Attached Figure Description
[0016] Figure 1 This is a flowchart of the viral RNA silencing inhibitor identification method according to an embodiment of the present invention;
[0017] Figure 2 This is a schematic diagram of the recognition model architecture based on ESM-2 and XGBoost according to an embodiment of the present invention;
[0018] Figure 3 This is a bar chart comparing the F1 scores of multiple candidate classifiers on the test set in an embodiment of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] This embodiment provides a method for identifying viral RNA silencing inhibitors based on machine learning. The specific process is as follows: Figure 1 As shown, it includes the following steps:
[0021] Step 1: Dataset construction and annotation.
[0022] In this embodiment, all protein sequence data were obtained from the Uniprot database (Universal Protein Resource). To construct a representative and targeted training dataset, the data collection scope was limited to viral protein sequences, and the host sources were further limited to plants, insects, and vertebrates.
[0023] From this data source, viral protein sequences that have been experimentally verified to have RNA silencing and inhibition functions were selected as the positive sample set; simultaneously, viral protein sequences that have not been experimentally verified to have this function were selected as the negative sample set. Subsequently, the data were binary labeled: viral protein sequences with RNA silencing and inhibition functions were labeled as "1" (positive samples), and viral protein sequences without this function were labeled as "0" (negative samples).
[0024] Step 2: Data redundancy removal.
[0025] To reduce the risk of model overfitting due to high sequence homology and to ensure the diversity of training data and the model's generalization ability, this embodiment performs redundancy removal on the dataset obtained in step 1. Specifically, the CD-HIT (Cluster Database at High Identity with Tolerance) software tool is used, and the sequence similarity threshold is set to 70%. That is, in sequence clusters with similarity greater than 70%, only one representative sequence is retained, and the remaining sequences are removed, thereby obtaining a non-redundant negative sample dataset.
[0026] Step 3: Feature extraction of protein sequence.
[0027] This step utilizes a large-scale protein language model to automatically extract deep features from the sequences. Specifically, all protein sequences (including positive samples and negative samples after redundancy removal) processed in step 2 are input into a pre-trained ESM-2 (Evolutionary Scale Modeling 2) model. ESM-2 is a protein language model based on the Transformer architecture, which, through pre-training with massive amounts of protein sequence data, can capture the evolutionary, structural, and functional information of protein sequences.
[0028] For each input protein sequence, the ESM-2 model outputs a feature matrix. In this embodiment, the feature vector corresponding to the position of the special classification marker CLS (Class Token) is extracted as the initial feature representation of the protein sequence. The dimension of this feature vector is exactly the same as the feature dimension of the output layer of the ESM-2 model.
[0029] Step 4: Divide the dataset.
[0030] To objectively evaluate the model's performance, the dataset needs to be divided into independent training, validation, and test sets. This embodiment employs a partitioning strategy that balances class balance and source diversity. First, from all sample feature representations obtained in step 3, a test set is constructed according to the following rules: 10 sequences are randomly selected from the positive samples; from the negative samples, sequences are stratified according to host source, specifically 10 from plants, 5 from insects, and 5 from vertebrates. These 30 samples constitute an independent test set, used to ultimately evaluate the model's generalization ability. The remaining sample data is then randomly divided into training and validation sets in an 8:2 ratio. The training set is used for model parameter learning, while the validation set is used for model selection and hyperparameter tuning during training.
[0031] Step 5: Multi-model training and parameter optimization.
[0032] Using the training set partitioned in step 4, multiple candidate classifiers are trained. The classifiers selected in this embodiment include: AdaBoost, XGBoost, CatBoost, Random Forest (RF), Multilayer Perceptron (MLP), and Deep Neural Network (DNN).
[0033] Considering the potential imbalance between positive and negative samples, a class weight adjustment strategy was introduced during training to assign higher weights to the minority class (positive samples) to reduce the model's bias towards the majority class. Simultaneously, to achieve optimal model performance, a combination of 10-fold cross-validation and grid search was used to optimize the hyperparameters of each model. For example, for the ultimately selected XGBoost model, the hyperparameters used in the grid search included, but were not limited to, the learning rate, maximum tree depth, and subsample ratio.
[0034] Step 6: Model Evaluation and Selection.
[0035] Using the validation set defined in step 4, evaluate the performance metrics of each candidate model after parameter optimization, including accuracy, precision, recall, F1 score, and AUC. Based on the evaluation results of the validation set, select the top-performing models.
[0036] Subsequently, the selected models were validated for their generalization ability using a pre-reserved test set that had never been used in the training and validation process. Test results showed that the XGBoost classifier achieved the best overall performance among all candidate classifiers. Figure 3 The bar chart showing the F1 scores of multiple candidate classifiers on the test set in this embodiment of the invention clearly demonstrates that XGBoost's F1 score is significantly higher than other models. Therefore, this embodiment selects the XGBoost classifier as the final recognition model and saves its model parameters.
[0037] Step 7: Perform functional prediction on the viral proteins to be tested.
[0038] like Figure 2 The diagram shown illustrates the identification model architecture of this invention based on ESM-2 and XGBoost. In practical applications, for a viral protein amino acid sequence (in FASTA format) with an unknown function, it is first input into the trained ESM-2 model to extract the feature vector corresponding to its CLS marker; then the feature vector is input into the XGBoost function prediction model saved in step 6, and the model can quickly output the identification result of whether the viral protein is a viral RNA silencing inhibitor ("pisitive" indicates a viral RNA silencing inhibitor, and "negative" indicates that it is not a viral RNA silencing inhibitor).
[0039] Optimal criteria for parameter settings:
[0040] To further demonstrate the rationality of the technical solution of this invention, this embodiment explores the basis for setting the redundancy removal threshold in step 2. Comparative experiments were conducted with sequence similarity thresholds set to 40%, 50%, 60%, 70%, 80%, and 90%, respectively. The results show that setting the threshold too high (e.g., ≥80%) leads to excessive redundant sequences, causing overfitting on the training set and decreased generalization ability on the test set; setting the threshold too low (e.g., ≤60%) results in excessive data loss and the loss of some rare features. Considering both data diversity and information integrity, setting the threshold to 70% ensures the model's generalization ability while retaining sufficient training samples, achieving the best recognition effect.
[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for identifying viral RNA silencing suppressors based on machine learning, characterized by, include: Step 1: Obtain training data from the Uniprot database. The training data includes viral protein sequences with RNA silencing and inhibition functions and viral protein sequences without RNA silencing and inhibition functions. Step 2: Label viral protein sequences with RNA silencing inhibition function as positive samples and viral protein sequences without this function as negative samples. Step 3: Perform redundancy removal on the protein sequences in the training data; Step 4: Input the deredundant protein sequence into the pre-trained ESM-2 model and extract the feature vector corresponding to the [CLS] label output by the model as the initial feature representation of the protein sequence. Step 5: Divide the dataset corresponding to the initial feature representation into a training set, a validation set, and a test set. Use the training set to train multiple candidate classifiers, and optimize the parameters during the training process by adjusting the class weights and combining 10-fold cross-validation with grid search. Step 6: Use the validation set to evaluate the performance metrics of each candidate classifier, and use the test set to verify the generalization ability of the classifier with the best performance. Select the XGBoost classifier that performs best on the test set as the final recognition model. Step 7: Input the amino acid sequence of the viral protein to be tested into the functional prediction model, and output the identification result of whether the protein is a viral RNA silencing inhibitor.
2. The method of claim 1, wherein the method is based on machine learning. Redundancy removal processing of viral protein sequences in the training data includes: Using the CD-HIT software tool, the sequence similarity threshold was set to 70%. In the sequence clusters with similarity greater than 70%, only one representative sequence was retained, and the rest of the sequences were removed, thus obtaining non-redundant negative sample data. 3.The method of claim 1, wherein, The ESM-2 model is a large-scale Transformer model pre-trained with protein language. 4.The method of claim 1, wherein, The dimension of the feature vector is consistent with the dimension of the output layer of the ESM-2 model used.
5. The method of claim 1, wherein the method is based on machine learning. The amino acid sequence of the viral protein to be tested is in FASTA format.
6. The method of claim 1, wherein the method is based on machine learning. When training various classifiers, cross-validation is used to optimize hyperparameters, which include at least one of the following: learning rate, maximum tree depth, and subsampling ratio.
7. The method for identifying viral RNA silencing inhibitors based on machine learning according to claim 1, characterized in that, The alternative classifiers include: AdaBoost, XGBoost, CatBoost, Random Forest (RF), Multilayer Perceptron (MLP), and Deep Neural Network (DNN).