Intelligent computing method for gram-negative bacterial secretion system effector protein prediction

Abstract

This invention provides an intelligent computational method for predicting effector proteins in the secretory system of Gram-negative bacteria, comprising: S1: collection and preprocessing of protein sequence datasets; S2: model construction; S3: model training; and S4: model prediction. This invention has advantages such as high prediction accuracy and good anti-interference performance.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and machine learning, specifically involving an intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria. Background Technology

[0002] Gram-negative bacteria are important pathogenic organisms. Their secretory systems inject effector proteins into host cells, directly affecting the host's normal physiological functions and leading to infection and disease. Secretory effector proteins play a crucial role in bacterial pathogenicity; therefore, accurately predicting the sequences and functions of these proteins is essential for understanding bacterial pathogenic mechanisms and developing antimicrobial therapeutic strategies.

[0003] Currently, various methods and tools have been used to predict secretory effector proteins in Gram-negative bacteria. These methods mainly rely on sequence features, evolutionary information, or structural information. For example, tools such as Bastion3 and Bastion6 combine position-specific score matrices (PSSMs) and evolutionary conservation data to predict secretory effector proteins of types T3SEs and T6SEs; while CNN-T4SE improves prediction accuracy by integrating information such as protein secondary structure, solvent accessibility, and one-hot encoding through a multi-source feature fusion strategy. However, existing prediction tools still face many challenges. First, traditional methods often cannot effectively handle high-dimensional features, leading to feature redundancy, which directly affects the accuracy and robustness of the model. Second, the function of many effector proteins depends on long-distance dependencies between amino acid residues, but existing models often struggle to capture these complex long-distance dependencies, limiting their predictive power. Furthermore, secretory effector proteins typically contain signal peptide regions located at the N-terminus or C-terminus, which play a crucial role in protein secretion, but existing methods often struggle to effectively identify and capture these key regions, thus affecting the overall accuracy of the prediction.

[0004] To address the aforementioned issues, this invention proposes an intelligent computational method for predicting Gram-negative bacterial secretory system effector proteins. This method effectively handles high-dimensional features, reduces feature redundancy, captures long-distance dependencies, and accurately identifies key signal peptide regions, thereby improving the accuracy and robustness of Gram-negative bacterial secretory effector protein prediction. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing an intelligent computational method for predicting effector proteins in the secretory system of Gram-negative bacteria with high prediction accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent computing method for predicting effector proteins in the secretory system of Gram-negative bacteria, comprising four processes: collection and preprocessing of protein sequence datasets, model construction, model training, and model prediction, the specific steps of which are as follows: S1: Collection and preprocessing of protein sequence datasets. First, Gram-negative bacterial effector protein sequences from public databases were collected and integrated. Second, the sequences were clustered and screened using the homology redundancy removal tool CD-HIT to remove highly similar sequences. Finally, the data was rigorously cleaned to remove incomplete or low-quality sequences, thereby obtaining a dataset for model training and testing.

[0007] S2: Model construction. The model mainly includes a pre-trained protein language model ESM-1b, a target preprocessing network based on hybrid convolutional experts, and a secretion-specific Transformer module that integrates relative position encoding.

[0008] S3: Model training. The model employs a five-fold cross-validation strategy to ensure training and validation on different data subsets, enhancing the robustness of the results. During training, the Adam optimizer is used to optimize parameters by calculating cumulative momentum, reducing training fluctuations. The loss function uses cross-entropy, and an early stopping mechanism is introduced: training stops when the F1 score on the validation set fails to improve within five consecutive iterations.

[0009] S4: Model prediction. After training and performance evaluation, the trained model is used to predict new protein sequences, identify whether they are secretory effector proteins, and provide corresponding classification results to guide practical applications.

[0010] In the aforementioned intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria, the implementation process of step S1 is as follows: The core of the dataset consists of secretory effector protein sequences (types I-IV and VI) and non-secretory protein sequences from Gram-negative bacteria. Secretory protein sequences were primarily derived from manually selected sets in authoritative public databases such as SecReT4, SecReT6, and BastionHub. Non-secretory protein samples were selected from annotated data from two previous studies, including annotated non-secretory proteins in UniProt and E. coli genomic homologous sequences selected based on BLASTp sequence alignment.

[0011] Independent test sets were taken from the test datasets of tools such as Basetion3, CNN-T4SE, and Basetion6. Independent test data for T1SE and T2SE were randomly extracted from the original BasetionHub dataset. Then, CD-HITv4.8.1 was used to cluster and filter sequences with a sequence consistency threshold of ≥60%, removing highly homologous sequences within the training dataset and between the training and test sets. This resulted in a training dataset containing 1341 secreted proteins (including 128 T1SE, 68 T2SE, 406 T3SE, 507 T4SE, and 232 T6SE) and 1577 non-secreting proteins, as well as an independent test set containing 110 secreted proteins and 150 non-secreting proteins. Benchmark data for comparison with existing binary classification methods directly used the corresponding test datasets of Basetion3, CNN-T4SE, and Basetion6. The benchmark sequences for T1SE and T2SE used relevant samples from the independent test sets of this study.

[0012] In the above-mentioned intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria, the implementation process of step S2 is as follows: The final output of the model is the conditional probability distribution of the non-effect proteins and the five effector proteins (T1SE to T4SE and T6SE) corresponding to the input protein sequence x. This distribution is obtained by the model through multi-stage feature processing and transformation using the softmax activation function: First, the sequence is encoded into 1280-dimensional embedding features using the pre-trained protein language model ESM-1b. This is then refined to 256 dimensions by a target preprocessing network containing 8 cooperating convolutional experts. Next, the relative position encoding Transformer module, containing 4 attention heads and using GELU as the activation function, captures key features. Finally, it is linearly mapped to the classification category dimension through a fully connected layer, ultimately achieving accurate classification of protein types. The specific formula and output logic are as follows:

[0013] Where the input variable x is the original protein sequence, and the output... This represents the conditional probability distribution of a given sequence x belonging to a certain protein class (covering 6 classes: non-effect proteins and 5 effector proteins). The weight matrix maps model features to the classification space, where L is the protein sequence length. PLM represents the embedding vector of the i-th amino acid residue, PLM corresponds to the pre-trained protein language model ESM-1b, TPN is the target preprocessing network, and RPET is the Transformer module that integrates relative position encoding.

[0014] ESM-1b is a Transformer-based pre-trained protein language model used to transform protein sequences into high-dimensional embedding representations. The input protein sequence... It will be converted into a set of embedding vectors using ESM-1b: ;in It is the i-th amino acid in the protein sequence. is the embedding vector corresponding to the amino acid, and L is the length of the protein sequence.

[0015] The Target Preprocessing Network (TPN) is used to extract key information and perform feature compression from the high-dimensional features output by a pre-trained protein language model. The TPN is based on a Hybrid Convolutional Experts (MoCE) architecture, which processes the input data in parallel through multiple convolutional expert modules. Each convolutional expert is responsible for extracting features from different local regions of the protein sequence. The gating network module dynamically assigns weights to each convolutional expert and generates gating weights using a softmax activation function. A probability distribution is constructed to adaptively adjust the contribution of each expert. The output of the gated network module. Defined as: Where x is the original input data feature vector after average pooling and flattening. It is a learnable weight matrix.

[0016] The multi-convolutional expert group module combines convolutional neural networks (CNNs) with the MoE architecture. Each convolutional expert extracts features of different scales and dimensions using an independent parameter set. A gating network assigns weights to each expert and performs weighted fusion to obtain the final output. Its expression is: ;in, These are the weights assigned by the gating network to the i-th convolutional expert. represents the feature extraction result of the i-th convolutional expert, and N is the number of convolutional experts.

[0017] To further enhance the focus on high-value features, TPN introduces a lightweight attention module that optimizes features through dynamic weight adjustments. This module leverages features fused from multi-convolutional expert groups and gating networks. For input, generate element-wise attention masks. The process involves convolution, GELU activation, and sigmoid normalization. Finally, the attention weighting mechanism uses element-wise multiplication to... and Multiplication highlights key task-relevant information, suppresses irrelevant features and noise, and improves network performance. The final output of TPN... Defined as: Through this design, TPN can significantly improve the model's performance during the feature learning process of complex protein sequences by integrating the diverse features of convolutional experts and the selective attention of attention mechanisms.

[0018] The secretion-specific Transformer module incorporating relative position encoding is one of the core components of this invention. Based on the Transformer's multi-head attention mechanism, this module retains the core computational framework of "query (Q) - key (K) - value (V)" and introduces relative position encoding (RPE). First, the module employs four parallel attention heads to capture sequence feature associations from different dimensions. Second, the core computational logic of the introduced relative position encoding strategy is as follows: ; Among them, among them, Characterizing the content correlation between residues, Introduce relative position offset, The feature values ​​are positionally corrected to ensure deep integration of relative position information and content features. Each attention head independently learns relative positional associations at different scales (such as short-range associations between signal peptides and functional domains, and long-range associations between transmembrane regions and intracellular action sites). Finally, multi-dimensional features are integrated through concat and linear mapping, as shown in the formula: ; ; in , , , , and All values ​​are learnable weight matrices for the i-th attention head. This design enables the model to simultaneously capture sequence content similarity and relative displacement relationships between residues, making it particularly suitable for identifying positionally flexible but functionally critical signal regions in secreted proteins. Subsequently, the feedforward network employs the GELU activation function to enhance the model's ability to fit nonlinear secretion features, while suppressing overfitting through a dropout layer (dropout rate 0.4). The multi-head attention outputs are concatenated and mapped through the weight matrix (Wᵒ), ultimately fusing with the features output by the target preprocessing network (TPN) to form a sequence representation that combines local key features with global positional correlations.

[0019] In the above-mentioned intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria, the implementation process of step S3 is as follows: The model training employs five-fold cross-validation to ensure model stability, while generalization ability is evaluated using an independent test set. The core training configuration is as follows: model weights are initialized using the Xavier method, and gradient updates are performed using the Adam optimizer (initial learning rate). Weight decay Using cross-entropy as the loss function, the specific formula is as follows: ; Where N is the number of samples and C is the number of categories (C=6 here). This indicates whether the i-th sample belongs to the c-th class (1 if it does, 0 otherwise). Indicates the first The probability that a sample is predicted to be of class c.

[0020] During training, the batch size was set to 32, the maximum number of iterations was 30, and an early stopping mechanism (patience value of 5, monitoring F1 score) was introduced to prevent overfitting. Meanwhile, the performance was optimized in the Transformer module through a regularization strategy, with the dropout rates of the attention mechanism, multi-head attention layer, and feedforward network set to 0.05, 0.4, and 0.4, respectively.

[0021] In the above-mentioned intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria, the implementation process of step S4 is as follows: After training and performance evaluation of the model are completed, the trained model can be applied to predict new protein sequences. When a new protein sequence is input, the model calculates the prediction probability for each category to determine whether the protein is a secretory effector protein and assigns it a specific category (such as T1SE, T2SE, T3SE, T4SE, and T6SE).

[0022] Compared with the prior art, the advantages of the present invention are as follows: 1. By combining the pre-trained protein language model ESM-1b, a target preprocessing network based on hybrid convolutional experts, and a secretion-specific Transformer module incorporating relative position encoding, the model can accurately capture complex features in protein sequences, especially the functional regions of secretory effector proteins. Through this multi-level feature extraction and modeling, the model can effectively improve the prediction accuracy of secretory effector proteins and significantly outperform traditional prediction methods based on single features.

[0023] 2. The target preprocessing network reduces the high-dimensional feature vector from 1280 dimensions to 256 dimensions through 8 cooperating convolutional expert modules. While significantly reducing computational complexity, it retains key features directly related to secretion function through a divide-and-conquer strategy, avoiding redundant information from interfering with model decision-making.

[0024] 3. TPN’s multi-convolutional expert group contains multiple parallel convolutional sub-modules. Each sub-module, as an independent expert, has its own set of parameters and can simultaneously extract local features of different scales and dimensions from the input data. It is especially suitable for capturing secretion signal features in effector proteins that are not fixed in location, thus solving the problem of single feature capture in traditional methods.

[0025] 4. The target preprocessing network dynamically assigns weights to each convolutional expert through a gating network module and uses the softmax activation function to construct a probability distribution. It adaptively adjusts the contribution of each expert, making the feature extraction process closely revolve around the core task of distinguishing non-effect proteins and T1SE-T4SE and T6SE. This strengthens the task relevance of features and significantly improves the model's ability to identify different types of effector proteins.

[0026] 5. By incorporating relative position encoding into the Transformer module, the model's ability to model long-range dependencies in protein sequences is enhanced. Relative position encoding can more accurately capture the relative spatial relationships between amino acid residues, which is crucial for the accurate prediction of secretory effector proteins. Therefore, this model can better identify interactions between different regions in the sequence, improving the classification accuracy of various types of secretory effector proteins. Attached Figure Description

[0027] Figure 1 This is a flowchart of the data collection and preprocessing process of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Example 1

[0029] like Figure 1-2 As shown, this embodiment constructs a high-quality, non-redundant dataset. Specifically, it collects Gram-negative bacterial T1SE-T4SE and T6SE secretory effector protein sequences and UniProt-annotated non-secretory protein sequences from authoritative databases such as SecReT4 and BaselineHub. Redundancy is removed using the CD-HIT v4.8.1 tool (sequence consistency ≥60%), resulting in a training set of 2904 samples (1341 secretory proteins, 1563 non-secretory proteins) and an independent test set of 260 samples (110 secretory proteins, 150 non-secretory proteins). The benchmark test data reuses relevant samples from existing tool test sets and the independent test set used in this study. Example 2

[0030] The model constructed in this embodiment is an end-to-end multi-classification model consisting of a "pre-trained protein language model + target preprocessing network + secretion-specific Transformer module", which can achieve direct prediction from the original protein sequence to the effector protein category.

[0031] The model input is the original protein sequence. (L is the sequence length, For the i-th amino acid residue, the output is the conditional probability distribution of the input sequence for the six classes (non-effect protein, T1SE, T2SE, T3SE, T4SE, T6SE). It is obtained by transformation through the softmax activation function, and the core formula is: ; in Here, PLM is the classification weight matrix, PLM is the pre-trained protein language model ESM-1b, TPN is the target preprocessing network, and RPET is the Transformer module that fuses relative position encodings. This represents the embedding vector of the i-th amino acid residue.

[0032] The pre-trained protein language model ESM-1b module is based on the Transformer architecture, containing a 33-layer Transformer encoder with a hidden layer dimension of 1280, processing each amino acid residue in the input sequence x. Output the corresponding 1280-dimensional embedding vector This forms a sequence feature matrix with dimension L×1280.

[0033] The target preprocessing network (TPN) module is based on a hybrid convolutional expert (MoCE) architecture, where the multi-convolutional expert group consists of 8 parallel convolutional expert modules. The gated network module receives the feature vector after average pooling and flattening, and generates 8 gated weights through a fully connected layer (256 hidden units) and a softmax activation function. (The formula for weight calculation is) , The weight matrix is ​​a learnable matrix (256×8 dimensions). Expert feature fusion is achieved by weighting and calculating the fused features. ( (This refers to the feature extraction result of the i-th convolutional expert). The lightweight attention module uses... As input, an element-wise attention mask α is generated through 1D convolution, GELU activation function, and sigmoid normalization, and then processed through element-wise multiplication. By highlighting key features and suppressing noise, the 1280-dimensional high-dimensional features are compressed to 256 dimensions, resulting in a refined feature matrix with a dimension of L×256.

[0034] The secretion-specific Transformer module, which incorporates relative position encoding, sets up four parallel attention heads and introduces a position-dependent bias matrix. and The core calculation logic is as follows: ; in is the dimension of the key vector. Multi-head attention fusion integrates the features of each attention head through a concat operation, followed by a linear mapping to obtain the output, as shown in the formula: ; The output weight matrix is ​​determined. The feedforward network consists of two fully connected layers, using the GELU activation function and a dropout rate of 0.4. Finally, the multi-head attention output is compared with the output of the TPN module. Perform residual connections to form a sequence representation with a dimension of L×256.

[0035] The fully connected classification layer uses two fully connected layers. The first layer has 128 hidden units (GELU activation function, dropout rate 0.4), and the second layer has 6 output units (corresponding to 6 classification categories). Finally, the conditional probability distribution is output through the softmax activation function. . Example 3

[0036] This embodiment employs a five-fold cross-validation strategy to ensure model stability and optimizes the training strategy to improve the model's generalization ability. The training hyperparameters are set as follows: using the Adam optimizer, with an initial learning rate of... Weight decay The loss function used is the cross-entropy loss function (calculated using the following formula). ), where N is the batch size and C=6 is the number of classification categories. Let be the true label (one-hot encoded) of the i-th sample. (where is the probability that the i-th sample is predicted to be of class c). The batch size is set to 32, and the maximum number of iterations is 30. The early stopping mechanism uses the validation set F1 score as the monitoring metric, and sets the patience value to 5, meaning that if the validation set F1 score does not improve in 5 consecutive iterations, training is stopped and the current optimal model parameters are saved. Example 4

[0037] In this embodiment, the protein sequence is input into the trained and saved model to make predictions. AUC (Area Under the Receiver Operating Characteristic) is the core performance evaluation indicator. The ROC curve is plotted with the true positive rate (sensitivity) on the ordinate and the false positive rate (1-specificity) on the abscissa. The higher the AUC value, the stronger the model's ability to distinguish between positive and negative samples.

[0038] In five-fold cross-validation, the method proposed in this invention achieved AUC values ​​between 0.930 and 0.992 for non-effect proteins and five types of secretory effector proteins (T1SE, T2SE, T3SE, T4SE, and T6SE). Compared to the existing popular multi-classification model DeepSecE (AUC values ​​between 0.926 and 0.989), the method of this invention has a better AUC range, fully demonstrating superior classification and discrimination capabilities.

[0039] Furthermore, in benchmark tests, this method demonstrated excellent overall performance in predicting five classes of secretory effector proteins: In T1SE prediction, it achieved an accuracy of 98.2% and an F1 score of 0.942, matching DeepSecE and significantly outperforming the traditional method T1SEstacker; in T2SE prediction, it achieved an accuracy of 94.4% and an F1 score of 0.667, surpassing DeepSecE across the board; in T6SE prediction, it achieved an accuracy of 98.6% and an F1 score of 0.957, outperforming tools such as Basement6 and DeepSecE; in T3SE prediction, it achieved an F1 score of 0.918 and an MCC score of 0.835, and in T4SE prediction, it achieved an accuracy of 98.3% and an F1 score of 0.949, ranking among the top in both tasks and exhibiting superior classification stability. All evaluation metrics met or exceeded the levels of existing mainstream prediction methods. These experimental results fully demonstrate the effectiveness and robustness of the proposed method in predicting secretory effector proteins.

[0040] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0041] Although this paper uses terms such as pre-trained protein language model and target preprocessing network extensively, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. An intelligent computational method for predicting effector proteins in the secretion system of Gram-negative bacteria, characterized in that, Includes the following steps: S1: Collection and preprocessing of protein sequence datasets: Collection of secretory effector protein sequences and non-secretory protein sequences of Gram-negative bacteria, followed by redundancy removal and cleaning. S2: Model building, constructing an end-to-end deep learning model, which includes: A pre-trained protein language model is used to encode the input raw protein sequence into high-dimensional embedding features; The target preprocessing network, connected after the pre-trained protein language model, is based on a hybrid convolutional expert architecture and includes multiple parallel convolutional expert modules and a gated network module, used to perform feature compression and task-specific refinement on the high-dimensional embedded features. A secretion-specific Transformer module that incorporates relative position encoding is connected after the target preprocessing network to capture long-range dependencies and relative position information between amino acid residues in the refined sequence features; S3: Model training, using a preprocessed protein sequence dataset to train a deep learning model; S4: Model prediction. Using a trained model, new protein sequences are predicted, and the classification result is output as either non-effect proteins or different types of secretory effector proteins.

2. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, The pre-trained protein language model is ESM-1b, which takes the input protein sequence as input. Each amino acid residue Convert to a 1280-dimensional embedding vector The output is a sequence feature matrix with dimension L×1280, where L is the sequence length.

3. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, The target preprocessing network achieves feature compression and refinement in the following ways: S21: Multiple parallel convolutional expert modules independently extract local features at different scales; S22: The gated network module dynamically assigns weights to each convolutional expert module through the softmax function, and fuses the output features of each expert through weighted summation; S23: Generate an element-wise attention mask through a lightweight attention module, and perform weighted enhancement on the fused features, finally outputting a refined feature matrix with a dimension of L×256.

4. The intelligent computational method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 3, characterized in that, The number of convolutional expert modules is 8; the gated network module contains a fully connected layer with 256 hidden units.

5. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, The core computational logic of the secretion-specific Transformer module that incorporates relative position encoding is as follows: ； Where Q, K, and V are the query, key, and value matrices, respectively. and This is a learnable relative position bias matrix. denoted as the dimension of the key vector; this module employs a multi-head attention mechanism, which uses multiple parallel attention heads to capture relative positional relationships at different scales, and then concatenates and linearly maps the outputs of each head.

6. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 5, characterized in that, The secretion-specific Transformer module, which incorporates relative position encoding, contains four parallel attention heads; its feedforward network employs the GELU activation function and has a dropout rate of 0.

4.

7. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, In step S3, the model training employs a five-fold cross-validation strategy; the Adam optimizer is used for parameter updates, with an initial learning rate set to... Weight decay is set to ; The loss function uses cross-entropy loss; and an early stopping mechanism is introduced, which stops training when the F1 score on the validation set does not improve in 5 consecutive iterations.

8. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, The secretory effector proteins are classified into five types: T1SE, T2SE, T3SE, T4SE, and T6SE. The model ultimately outputs the conditional probability distribution of the input sequence belonging to non-effector proteins and the above five types of effector proteins through the softmax activation function.

9. The intelligent computing method for predicting effector proteins in the secretion system of Gram-negative bacteria according to claim 1, characterized in that, In step S1, the CD-HIT tool is used to cluster and remove redundancy from the sequences with a sequence consistency of ≥60% as the threshold.

10. An intelligent computing system for predicting effector proteins in the secretory system of Gram-negative bacteria, used to run the intelligent computing method for predicting effector proteins in the secretory system of Gram-negative bacteria as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the protein sequence to be predicted; Memory, which stores computer-executable instructions; Processor, for executing the computer-executable instructions; The output module is used to output the predicted classification results.

Images