Plant lncpeps prediction method based on global pruning and cascaded features

Abstract

The present application relates to a kind of plant lncPEPs prediction method based on global pruning and cascade feature, its technical characteristics are: lncPEPs sequence data are collected and the data set for model training and testing is constructed;Based on undersampling, the extremely unbalanced data set is split, and feature is extracted and selected;Based on multi-angle, global pruning is carried out, and the optimal model and global optimal sub data set are obtained;Cascade feature is obtained by constructing deep integrated structure;Cascade feature is input into the meta-learner composed of machine learning model, and thus the prediction method lncPEPPred is constructed, and the method lncPEPPred is used to predict plant lncPEPs.The present application combines bioinformatics technology and machine learning technology together, can accurately and effectively realize the prediction function of plant lncPEPs, is conducive to revealing the molecular mechanism of life body disease resistance and adversity resistance, and has important value in the field of agricultural production etc..

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and relates to bioinformatics and machine learning methods, especially a method for predicting plant lncPEPs based on global pruning and cascaded features. Background Technology

[0002] Long noncoding RNAs (lncRNAs) are widely distributed in organisms and play a powerful regulatory role, making them a frontier and hot topic in life science research. Traditionally, lncRNAs were thought to lack open reading frames (ORFs), a type of element capable of encoding proteins, and therefore lacked coding ability. However, recent studies have shown that some lncRNAs contain small open reading frames (sORFs), which can encode small peptides involved in regulating growth, development, and stress responses in organisms. The discovery of lncRNA-encoded peptides (lncPEPs) has overturned researchers' understanding of lncRNAs, broadening and deepening researchers' understanding of genomics, making lncPEPs a research hotspot in life sciences.

[0003] Molecular biology experiments are a classic technique for predicting lncPEPs, but they are time-consuming and expensive. Prediction methods developed based on machine learning are more time-efficient and provide more accurate identification data for biological experiments, attracting increasing attention from researchers. However, most of these methods focus on the animal domain, with limited research on the plant domain. Existing methods suffer from the following problems:

[0004] 1. Most existing methods focus on animal and human lncPEPs. The differences between animal and plant data mean that these methods cannot be directly applied to related research in the plant field.

[0005] 2. The number of verified plant lncPEPs is relatively small, resulting in a small number of positive samples when constructing the dataset, which easily leads to extreme imbalanced data problems, thus posing a huge challenge to the training of machine learning models.

[0006] 3. lncPEPs are a class of short sequences with a length of less than 100 amino acid units. Existing artificial features based on domain knowledge cannot fully and effectively characterize lncPEPs.

[0007] 4. The model obtained by integrating traditional machine learning models has a shallow structure and fails to fully explore the relationship between the output features of these machine learning models. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a plant lncPEPs prediction method based on global pruning and cascade features, which can accurately and effectively predict plant lncPEPs.

[0009] The present invention solves the existing technical problems by adopting the following technical solution:

[0010] A method for predicting plant lncPEPs based on global pruning and cascade features includes the following steps:

[0011] Step 1: Collect lncPEPs sequence data, preprocess the collected data, create a knowledge base to filter the data, and build a dataset for model training and testing.

[0012] Step 2: Split the extremely imbalanced dataset based on downsampling to obtain several balanced subsets, extract high-dimensional artificial features from the data, and perform feature dimensionality reduction processing on the high-dimensional features based on the EFISS-ES feature selection method.

[0013] Step 3: Train multiple machine learning models and set filtering conditions to perform global pruning to obtain optimized models and globally optimal subsets of data;

[0014] Step 4: Combine the outputs of multiple machine learning models into a feature vector, and continue to train and filter machine learning models to obtain new feature vectors. Connect the feature vectors and use them as input to the next layer of the model to obtain cascaded features.

[0015] Step 5: Input the cascaded features into the meta-learner composed of machine learning models to construct the prediction method lncPEPPred, and use the prediction method lncPEPPred to predict plant lncPEPs.

[0016] Furthermore, the specific implementation method of step 1 includes the following steps:

[0017] Step 1.1: Collect cross-species data of plant lncPEPs obtained by mass spectrometry and ribosome imprinting, mine sORFs in plants that are still considered to lack coding ability, obtain the amino acid sequences corresponding to sORFs based on the correspondence between nucleotide codons and amino acids, and combine the above data into a training and validation set.

[0018] Step 1.2: Download lncPEPs validated by molecular biology experiments from the ncEP database as positive set data; randomly select amino acid sequences corresponding to sORFs in non-coding RNAs, and construct multiple independent test sets according to different ratios of positive and negative data to verify the generalization ability of the prediction method and its performance under imbalanced data conditions.

[0019] Furthermore, steps 1.1 and 1.2 also include a process step of removing sequence redundancy using the CD-HIT tool with a threshold of 0.8.

[0020] Furthermore, the specific implementation method of step 2 includes the following steps:

[0021] Step 2.1: First, the negative set samples in the training and validation sets are downsampled and split into multiple subsets according to the sequence length distribution; then, samples are randomly selected from each subset to obtain 10 datasets with the same number of samples as the positive set; finally, the positive and negative samples are combined to obtain 10 balanced subsets.

[0022] Step 2.2: Extract 15 high-dimensional features from the above data, including amino acid composition transformation distribution, amino acid pair composition, k-spaced amino acid pair composition, amino acid composition, paired amino acid composition, amino acid pair composition appearing in an adaptive skipping manner, composition of grouped amino acid types, amino acid triplet group composition, and k-spaced amino acid group pair composition.

[0023] Step 2.3: Use the EFISS-ES feature selection method to perform feature dimensionality reduction processing on the above 15 types of high-dimensional features respectively.

[0024] Furthermore, the EFISS-ES feature selection method is as follows: First, the feature selection is performed on the feature selection based on the importance score E obtained through the embedding method. imp Importance score of statistical methods S imp Perform weighted operations to obtain the feature ranking result imp. f :

[0025] imp f =w×E imp +(1-w)×S imp

[0026] Where w is the weight, set to 0.5;

[0027] Subsequently, a sequential forward search strategy is used to obtain the final feature selection results.

[0028] Furthermore, the specific implementation method of step 3 includes the following steps:

[0029] Step 3.1: Train 10 machine learning models using the 15 features after dimensionality reduction, use 10-fold cross-validation to obtain the average value of the evaluation index, and conduct a comprehensive evaluation of the above models;

[0030] Step 3.2: Set multiple filtering conditions for the evaluation indicators and perform global pruning from the perspectives of data, features, and models.

[0031] Furthermore, the 10 machine learning models in step 3.1 include: decision tree, Naive Bayes, k-nearest neighbors, linear discriminant analysis, support vector machine, logistic regression, gradient boosting decision tree, random forest, lightweight gradient boosting machine, and extreme gradient boosting tree.

[0032] Furthermore, the filtering conditions in step 3.2 include:

[0033] The filtering criteria for features and models are obtained by calculating the difference between the sensitivity and specificity of different models:

[0034] Rule 1: Abs|SN-SP|≤3%

[0035] The filtering criteria for the data are based on the average accuracy (avg) of the model. ACC and variance std ACC This yields the final balanced dataset index:

[0036] rule 2:

[0037] Furthermore, the specific implementation method of step 4 includes the following steps:

[0038] Step 4.1: The ensemble structure used to obtain cascaded features consists of multiple identical structural units. In this ensemble structure, each layer corresponds to one structural unit, and each unit contains several models. In this ensemble structure, the outputs of the machine learning models selected in Step 3 are first used to form a feature vector.

[0039] Step 4.2: In each structural unit, input the above feature vector into 10 machine learning models to obtain the model outputs. Filter the models according to the filtering conditions of features and models, and connect the outputs of the models that meet the conditions to obtain a new output vector. Connect the output vector with the input vector to obtain the input of the next layer structural unit.

[0040] Step 4.3: Evaluate the output of each layer in the integrated structure. When the number of layers reaches the set number or the accuracy reaches the maximum, the structure stops the operation of the next layer unit and obtains the cascaded feature; otherwise, continue to execute step 4.2.

[0041] Furthermore, step 5 encapsulates the code involved in steps 3 to 4 to obtain the method lncPEPPred for predicting plant lncPEPs.

[0042] The advantages and positive effects of this invention are:

[0043] 1. This invention combines bioinformatics and machine learning techniques. Through methods such as acquiring plant lncPEPs data, segmenting and selecting extremely imbalanced datasets, constructing cascaded features, and developing lncPEPs prediction methods, it accurately and effectively achieves the prediction function of plant lncPEPs. In-depth research on the prediction of plant lncPEPs is beneficial for revealing the molecular mechanisms of disease resistance and stress tolerance in organisms, and has significant value in agricultural production and other fields.

[0044] 2. This invention is based on downsampling and splitting extremely imbalanced datasets. It performs global pruning from multiple dimensions such as data, features, and models to obtain a balanced dataset with information similar to the original dataset. This solves the problem that machine learning model training under extremely imbalanced data is prone to bias towards majority class samples.

[0045] 3. Based on the adaptive selection of features and models, this invention obtains cascaded features through a deep integration structure, which can more effectively characterize short sequences such as lncPEPs.

[0046] 4. This invention employs a special integration structure to continuously integrate and iterate the machine learning model, achieving deep integration of the model. Attached Figure Description

[0047] Figure 1 This is a flowchart of the overall prediction method of the present invention;

[0048] Figure 2 This is a flowchart illustrating the construction, partitioning, feature extraction, and selection of the training and validation sets according to the present invention.

[0049] Figure 3 This is a flowchart of the global pruning process of the present invention;

[0050] Figure 4 This is a schematic diagram illustrating the performance of the machine learning model of the present invention under filter condition 1;

[0051] Figure 5 This is a structural unit construction diagram of the present invention;

[0052] Figure 6 This is a schematic diagram illustrating the performance of the present invention on an independent test set. Detailed Implementation

[0053] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] The design concept of this invention is to predict unlabeled data by modeling the data on whether already labeled lncPEPs are plant lncPEPs. Since predicting high-confidence lncPEPs is fundamental to in-depth analysis of their biological functions, accurately predicting these unique amino acid sequences is crucial. Therefore, the prediction method developed based on machine learning technology can provide a reference for subsequent biological experiments.

[0055] The innovations of this invention are mainly twofold: firstly, global pruning of extremely imbalanced datasets to obtain balanced subsets; and secondly, cascaded features obtained through deep ensemble of machine learning models. The combination of these two approaches helps improve the classification accuracy and generalization ability of the final prediction method.

[0056] Based on the above ideas, this invention proposes a method for predicting plant lncPEPs based on global pruning and cascade features, such as... Figure 1 As shown, it includes the following steps:

[0057] Step 1: Collect lncPEPs sequence data from public datasets and published literature, preprocess the data, create a knowledge base to filter the data and improve the credibility of the experimental data, thereby constructing a dataset for model training and testing.

[0058] In this embodiment, the specific implementation method of this step includes the following steps:

[0059] Step 1.1: Collect cross-species data of plant lncPEPs obtained by mass spectrometry and ribosome blotting. Download plant sORFs that are still considered non-coding from the Ensemble Plants database (http: / / plants.ensembl.org / index.html). Based on the correspondence between nucleotide codons and amino acids, obtain the amino acid sequences corresponding to the sORFs. The knowledge base includes some manual screening rules, such as: length between 3-100 amino acid units and the use of the CD-HIT tool. In particular, the CD-HIT tool uses 0.8 as a threshold to remove sequence redundancy. The above data constitutes the training and validation set D1 of the experiment (positive to negative sample imbalance ratio of 1:10).

[0060] Step 1.2: Download the molecularly validated lncPEPs from the ncEP database (http: / / www.jianglab.cn / ncEP / ) as the positive set data. Randomly select the amino acid sequences corresponding to sORFs in non-coding RNAs, and use the CD-HIT tool with a threshold of 0.8 to remove sequence redundancy. Construct independent test sets D2 and D3 according to the different ratios of positive and negative data to verify the generalization ability of the prediction method and its performance on imbalanced data (positive to negative sample ratios of 1:1 and 1:4, respectively). Detailed information on the training and validation sets and independent test sets is shown in Table 1.

[0061] Table 1. Dataset details used for plant lncPEPs prediction

[0062]

[0063] Step 2: Split the extremely imbalanced dataset based on downsampling to obtain several balanced subsets, and extract high-dimensional artificial features from the data. Then, perform feature dimensionality reduction on the high-dimensional artificial features based on the classic EFISS-ES feature selection method.

[0064] In this embodiment, the specific implementation method of this step includes the following steps:

[0065] Step 2.1: First, split the negative samples in the training and validation sets into multiple subsets based on sequence length distribution. Then, randomly select samples from each subset, thus obtaining 10 datasets with the same number of samples as the positive set. Finally, combine the positive and negative samples respectively to obtain 10 balanced subsets.

[0066] Step 2.2: Extract 15 types of artificial features containing domain knowledge from the above data, including amino acid composition transformation distribution (CTDC, CTDT, and CTDD), amino acid pair composition (DPC), k-spaced amino acid pair composition (C1SAAP, C2SAAP, and C3SAAP), amino acid composition (AAC), paired amino acid composition (GDPC), amino acid pair composition appearing in an adaptive skipping manner (ASDC), grouped amino acid type composition (GAAC), amino acid triplet grouped composition (GTPC), and k-spaced amino acid grouped pairing composition (C1SAAGP, C2SAAGP, and C3SAAGP).

[0067] Step 2.3: Use the EFISS-ES feature selection method to perform feature dimensionality reduction on the above 15 categories of high-dimensional features. EFISS-ES first assigns importance scores (E) to the features selected by the embedded feature ranking algorithm. imp Importance scores of statistical feature ranking methods (S) impThe weighted operation is performed to obtain the feature ranking result imp. f :

[0068] imp f =w×E imp +(1-w)×S imp

[0069] Where w is the weight, and in this invention, w is set to 0.5. Subsequently, a sequential forward search strategy is used to obtain the final feature selection result. The overall process of constructing the training and validation set D1 and extracting and selecting relevant artificial features in steps 1 and 2 is as follows: Figure 2 As shown.

[0070] Step 3: Train multiple machine learning models and set multiple filtering conditions to perform global pruning from multiple perspectives, including data, features, and models, to obtain high-performing models and globally optimal subsets of data. The global pruning process is as follows: Figure 3 As shown.

[0071] In this embodiment, the specific implementation method of this step includes the following steps:

[0072] Step 3.1: Train 10 machine learning models using the 15 features after dimensionality reduction. The models include: Decision Tree (DT), Naive Bayes (NB), k-Nearest Neighbors (kNN), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), Lightweight Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting Tree (Xgboost). In this invention, 10-fold cross-validation is used to obtain the average value of the evaluation metrics, and the above models are comprehensively evaluated. The evaluation metrics selected are: Accuracy (ACC), Recall (REC), Precision (PRE), F1 score (F1_score), Minority Accuracy (MinACC), Majority Accuracy (MajACC), and the geometric mean (G-mean) of the minor and majority accuracy.

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] Where TP is the number of correctly predicted lncPEPs, TN is the number of correctly predicted non-lncPEPs, FP is the number of incorrectly predicted non-lncPEPs, and FN is the number of incorrectly predicted lncPEPs. Furthermore, in this invention, the minority class corresponds to the positive set of samples, and the majority class corresponds to the negative set of samples.

[0081] Step 3.2: Set multiple filtering conditions for the evaluation metrics, and perform global pruning from multiple perspectives such as data, features, and models. The purpose of global pruning is to select high-performing models and suitable datasets, and to reduce the time consumption in subsequent steps.

[0082] (1) The filtering criteria for features and models are obtained by calculating the difference between the model's sensitivity (SN) and specificity (SP):

[0083] Rule 1: Abs|SN-SP|≤3%

[0084]

[0085]

[0086] (2) The filtering criteria for the data are based on the mean average of the model's accuracy (ACC). ACC and variance std ACC This yields the final balanced dataset number:

[0087] rule 2:

[0088] Global pruning based on feature and model filtering conditions can yield models with better performance. Data filtering conditions can identify subsets that best reflect the training and validation sets, replacing the original extremely imbalanced datasets and reducing the negative impact of imbalanced data on model training. The performance of 10 types of machine learning models under different features is shown below. Figure 4 As shown.

[0089] Step 4: Combine the outputs of multiple machine learning models into a feature vector, and continue to train and filter machine learning models to obtain new feature vectors. Connect the feature vectors and use them as input to the next layer of the model to obtain cascaded features.

[0090] In this embodiment, the specific implementation method of this step includes the following steps:

[0091] Step 4.1: The ensemble structure used to obtain cascaded features consists of multiple identical structural units. In this special structure, each layer corresponds to a structural unit, and each structural unit contains several models. This is a deep structure for ensemble learning. Furthermore, to reflect diversity, the classification principles of the models are not entirely the same. In this structure, the outputs of the machine learning models selected in Step 3 are used to form a feature vector.

[0092] Step 4.2: In each structural unit, input the aforementioned feature vector into 10 machine learning models to obtain the model outputs. Filter the models according to filtering condition 1. Concatenate the outputs of the models that meet the condition to obtain a new output vector. Concatenate the output vector with the input vector to obtain the input for the next layer of structural units.

[0093] Step 4.3: Evaluate the output of each layer in the integrated structure, focusing on accuracy. When the set number of layers is reached, or the accuracy reaches its maximum, the structure stops processing the next layer, resulting in cascaded features; otherwise, continue with step 4.2. Figure 5 As shown. In this invention, the maximum number of layers in the deep integration structure is set to 3 to obtain cascaded features for more effective characterization of lncPEPs.

[0094] Step 5: Input the cascaded features into a meta-learner composed of machine learning models to construct the prediction method lncPEPPred, and use this method lncPEPPred to predict plant lncPEPs.

[0095] To verify the effectiveness of this invention, the prediction method lncPEPPred was tested on multiple independent test sets. During testing, a confusion matrix was constructed based on the predicted labels of the test data to achieve a systematic evaluation of the lncPEPPred prediction method. The method of this invention was compared with existing methods, and the classic prediction methods CPCred, DeepCPP, and MiPepid were selected as control groups. The experimental results are as follows: Figure 6 As shown.

[0096] Experimental results show that lncPEPPred outperforms existing prediction methods on multiple evaluation metrics, indicating that lncPEPPred can more effectively predict plant lncPEPs. This is mainly due to lncPEPPred's use of global pruning to obtain a balanced subset and the mining of cascaded features. Obtaining a balanced subset containing rich original information can replace the original extremely imbalanced dataset, mitigating the impact of imbalanced data on model training. Furthermore, the acquisition of cascaded features references deep neural network operations, achieving more effective feature representations of lncPEPs through deep integration with machine learning models.

[0097] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A method for predicting plant lncPEPs based on global pruning and cascade features, characterized in that: Includes the following steps: Step 1: Collect lncPEPs sequence data, preprocess the collected data, create a knowledge base to filter the data, and build a dataset for model training and testing. Step 2: Split the extremely imbalanced dataset based on downsampling to obtain several balanced subsets, extract high-dimensional artificial features from the data, and perform feature dimensionality reduction processing on the high-dimensional features based on the EFISS-ES feature selection method. Step 3: Train multiple machine learning models and set filtering conditions to perform global pruning to obtain optimized models and globally optimal subsets of data; Step 4: Combine the outputs of multiple machine learning models into a feature vector, and continue to train and filter machine learning models to obtain new feature vectors. Connect the feature vectors and use them as input to the next layer of the model to obtain cascaded features. Step 5: Input the cascaded features into the meta-learner composed of machine learning models to construct the prediction method lncPEPPred, and use the prediction method lncPEPPred to predict plant lncPEPs. The specific implementation method of step 2 includes the following steps: Step 2.1: First, the negative set samples in the training and validation sets are downsampled and split into multiple subsets according to the sequence length distribution; then, samples are randomly selected from each subset to obtain 10 datasets with the same number of samples as the positive set; finally, the positive and negative samples are combined to obtain 10 balanced subsets. Step 2.2: Extract the high-dimensional features of the amino acid composition transformation distribution, amino acid pair composition, k-spacer amino acid pair composition, amino acid composition, paired amino acid composition, amino acid pair composition appearing in an adaptive skipping manner, composition of grouped amino acid types, amino acid triplet group composition, and k-spacer amino acid group pair composition from the above data. Step 2.3: Use the EFISS-ES feature selection method to perform feature dimensionality reduction on the above high-dimensional features respectively; The EFISS-ES feature selection method is as follows: First, the importance score obtained through the embedding method is calculated. Importance score of statistical methods Perform weighted operations to obtain the feature ranking results. imp f : in, w The weight is set to 0.5; Subsequently, a sequential forward search strategy is used to obtain the final feature selection results.

2. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 1, characterized in that: The specific implementation method of step 1 includes the following steps: Step 1.1: Collect cross-species data of plant lncPEPs obtained by mass spectrometry and ribosome imprinting, mine sORFs in plants that are still considered to lack coding ability, obtain the amino acid sequences corresponding to sORFs based on the correspondence between nucleotide codons and amino acids, and combine the above data into a training and validation set. Step 1.2: Download lncPEPs validated by molecular biology experiments from the ncEP database as positive set data; randomly select amino acid sequences corresponding to sORFs in non-coding RNAs, and construct multiple independent test sets according to different ratios of positive and negative data to verify the generalization ability of the prediction method and its performance under imbalanced data conditions.

3. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 2, characterized in that: Steps 1.1 and 1.2 also include a process step of removing sequence redundancy using the CD-HIT tool with a threshold of 0.

8.

4. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 1, characterized in that: The specific implementation method of step 3 includes the following steps: Step 3.1: Train 10 machine learning models using the dimensionality-reduced features, use 10-fold cross-validation to obtain the average value of the evaluation metrics, and then comprehensively evaluate the above models. Step 3.2: Set multiple filtering conditions for the evaluation indicators and perform global pruning from the perspectives of data, features, and models.

5. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 4, characterized in that: The 10 machine learning models in step 3.1 include: decision tree, Naive Bayes, k-nearest neighbors, linear discriminant analysis, support vector machine, logistic regression, gradient boosting decision tree, random forest, lightweight gradient boosting machine, and extreme gradient boosting tree.

6. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 4, characterized in that: The filtering conditions in step 3.2 include: The filtering criteria for features and models are obtained by calculating the difference between the sensitivity and specificity of different models: The filtering criteria for the data are based on the mean of the model's accuracy. avg ACC and variance std ACC This yields the final balanced dataset number. index : 。 7. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 1, characterized in that: The specific implementation method of step 4 includes the following steps: Step 4.1: The ensemble structure used to obtain cascaded features consists of multiple identical structural units. In this ensemble structure, each layer corresponds to one structural unit, and each unit contains several models. In this ensemble structure, the outputs of the machine learning models selected in Step 3 are first used to form a feature vector. Step 4.2: In each structural unit, input the above feature vector into 10 machine learning models to obtain the model outputs. Filter the models according to the filtering conditions of features and models, and connect the outputs of the models that meet the conditions to obtain a new output vector. Connect the output vector with the input vector to obtain the input of the next layer structural unit. Step 4.3: Evaluate the output of each layer in the integrated structure. When the number of layers reaches the set number or the accuracy reaches the maximum, the structure stops the operation of the next layer unit and obtains the cascaded feature; otherwise, continue to execute step 4.

2.

8. The plant lncPEPs prediction method based on global pruning and cascade features according to claim 1, characterized in that: Step 5 involves encapsulating the code involved in steps 3 to 4 to obtain the method lncPEPPred for predicting plant lncPEPs.

Images