A secreted protein recognition method fusing statistical prior and bias dynamic calibration
By constructing a multi-branch network architecture and calibrating the category-aware bias term, the problems of hierarchical classification and long-tail distribution in the identification of secretory proteins are solved, improving the identification accuracy and recall of scarce secretory proteins, and making it suitable for biological sequence analysis tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-19
AI Technical Summary
Existing methods for identifying secretory proteins are prone to confusing binary and multi-class classification when dealing with hierarchical classification tasks and long-tailed distributed data. This leads to non-secretory proteins occupying the feature space, and traditional models tend to fit the majority class, resulting in extremely poor ability to identify rare secretory protein categories.
A multi-branch network architecture is constructed, and a category-aware bias term is introduced. The prior probability bias caused by data imbalance is dynamically quantified and offset by Bayes' theorem. A multilayer perceptron and a bias estimation head are used, combined with statistical prior frequency and likelihood difference factor, to dynamically calibrate the model and improve the recognition accuracy and recall rate of rare secretory protein categories.
Effective decoupling of functional identification and type classification improves the accuracy and recall of identifying rare secreted protein categories. The model is more robust when facing unknown data and is suitable for other biological sequence analysis tasks.
Smart Images
Figure CN122245439A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of bioinformatics and artificial intelligence, and in particular to a method for identifying secretory proteins that integrates statistical priors and dynamic bias calibration. Background Technology
[0002] Secretory proteins are key proteins secreted by bacteria into host cells to regulate host function and are the core focus of research on the pathogenic mechanisms of pathogenic bacteria. Current identification methods face two major challenges: The task is hierarchical: the identification process involves two levels: "whether it is secreted (binary classification)" and "what type of secretory protein it belongs to (multi-classification)". Existing methods often confuse this into a single, flat task, resulting in a large number of non-secreting proteins crowding out the feature space.
[0003] Extreme data imbalance (long-tailed distribution): In real biological data, the number of non-secretory proteins (negative samples) far exceeds the number of secretory proteins (positive samples); and within secretory proteins, there may be hundreds of T4 types, while only dozens of T2 types (e.g., 507 T4 types vs. 68 T2 types). Traditional models tend to fit the majority class, resulting in extremely poor ability to identify scarce classes, making them very easy to misclassify as non-secretory proteins. Summary of the Invention
[0004] To address the shortcomings of existing secretory protein prediction methods in handling hierarchical classification tasks and long-tailed distributed data, this invention aims to provide a secretory protein identification method that integrates statistical prior and dynamic bias calibration. By constructing a multi-branch network architecture, a category-aware bias term is introduced to dynamically quantify and offset the prior probability bias caused by data imbalance. This significantly improves the model's identification accuracy and recall rate for scarce secretory protein categories such as T2SE and T1SE without disrupting the data distribution.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying secretory proteins that integrates statistical prior and dynamic bias calibration includes the following steps: A protein sequence training dataset with hierarchical labels was constructed and divided into a positive sample set and a negative sample set. Based on the positive sample set, the prior frequency of each class of secreted proteins was calculated. ; Feature extraction is performed on the protein sequence to obtain global feature representation and end feature representation, which are then concatenated to obtain the final sequence feature vector. ; sequence feature vector The input consists of a multilayer perceptron with three parallel branches. Branch one is a multi-classifier head used to identify specific subtypes of secreted proteins, and the output is a predicted probability distribution. 'k' is the type number of the secretory protein; branch two is a binary classification head used to determine whether a protein is a secretory protein, outputting the original logical value. Branch three is used to learn the likelihood difference factor for each class in the current sample. The bias estimation head, the third branch is used in the model training phase; Based on Bayes' theorem, and combined with the prediction probability of multi-classifier heads... Prior frequency and likelihood difference factor Calculate the category perception bias term for the current sample. Then, the category perception bias item Convert to binary classification head raw logic value Additive corrections in space ; The model's total loss function includes category classification loss. Deviation calibration detection loss and probability constraint loss used to prevent overflow of the probability space after calibration .
[0006] The present invention has the following beneficial effects: 1. Fundamentally solves the problem of missed detections caused by long-tail distribution: This invention does not simply solve the imbalance by replicating samples (oversampling), but rather through a mathematically derived bias term. It dynamically provides gradient compensation for scarce categories (such as T2SE), which enables the model to build a sufficiently strong confidence in recognition even if it has only seen a small number of T2 samples.
[0007] 2. Decoupling Functional Identification and Type Classification: By using independent binary and multi-classification heads, the "yes / no problem" and the "selection problem" are decomposed. The binary classification head focuses on using massive negative samples to delineate boundaries, while the multi-classification head focuses on fine-grained features. The two do not interfere with each other and promote each other by sharing features.
[0008] 3. Efficient and robust inference: Complex statistical bias calculations only occur during the training phase. During the inference phase, the model structure reverts to a standard deep neural network, which requires less computation. Furthermore, due to calibration during training, the model exhibits better robustness when facing unknown data and is less prone to bias towards the majority class.
[0009] 4. Wide range of applications: The bias calibration framework of this invention can also be extended to other biological sequence analysis tasks with "imbalance between positive and negative samples" and "long-tailed distribution within positive samples", especially suitable for identification and classification scenarios where the number of samples in each subclass is uneven in protein classification tasks, such as antimicrobial peptide functional classification. Attached Figure Description
[0010] Figure 1This is a framework diagram of the secretory protein prediction model of the present invention; Figure 2 This is a flowchart of the model training process of the present invention. Detailed Implementation
[0011] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Please see Figure 1 and Figure 2 A method for identifying secretory proteins that integrates statistical prior and dynamic bias calibration includes the following steps: S1. Data Construction and Preprocessing: Construct a training dataset with clearly defined hierarchical labels and calculate global statistical prior information. The specific steps are as follows: S1-1. Construct a protein sequence training dataset containing hierarchical labels and divide it into a positive sample set and a negative sample set, for example, The process of constructing the hierarchical labels is as follows: first, construct a binary label indicating whether it is a secretory protein, and then construct a label indicating which category the secretory protein belongs to.
[0012] The dataset partitioning process is as follows: Positive sample set: contains five types of secretory proteins, T1, T2, T3, T4, and T6, which have been experimentally verified.
[0013] Negative sample set: contains non-secreted proteins and is used as auxiliary training data.
[0014] S1-2. Based on the positive sample set, calculate the prior frequency of each type of secreted protein. For example, the number of samples of each type of secreted protein in the positive sample set is counted. and the total number of all positive samples ; Calculate the prior frequency of each type of secreted protein. The vector It will be used as a non-trainable constant parameter in subsequent bias calculations.
[0015] S2. Feature Representation Learning Based on Hybrid Architecture (Shared Feature Extraction Module): Utilizing a pre-trained model and a convolutional neural network, protein features that fuse global semantics and local signals are extracted. The specific steps are as follows: S2-1. To capture evolutionary features and local motif patterns in protein sequences, a pre-trained protein language model, ProtT5-XL-UniRef50 (ProtT5), is used as a feature extractor. The protein sequence is input into the feature extractor one by one. For example, given a protein sequence of length L, ProtT5 will output a residue-level embedding matrix. Based on this embedding matrix, the following constructs the feature representation of the protein from two different perspectives; S2-2. Global Semantic Features: To characterize the overall functional properties of the protein, all residue embeddings are averaged and pooled. Then, a fully connected layer is used to further extract features, resulting in the final global feature representation. ; S2-2. Local Signal Features: Given that bacterial secretion signals are typically enriched in the terminal regions of protein sequences, i.e., the two ends of a protein sequence, a window of length W of 100 can be defined to capture position-specific terminal signal features in the residue embedding matrix. The N-terminal and C-terminal feature representations are defined as follows: Where W is the window length, L is the protein length, the protein sequence is defined by its N-terminus and C-terminus, and e1 represents the characteristic representation of the first amino acid. The following are the definitions of the given values: and The input is fed into a single-layer convolutional neural network with 9 kernels, and after max pooling, the N-terminal feature vector is obtained. and C-end feature vector ; S2-3. Feature Fusion: Global Feature Representation N-terminal feature vector and C-end feature vector The sequences are concatenated to obtain the final sequence feature vector. The sequence feature vector Simultaneously input into the three branches of the multilayer perceptron.
[0016] S3. Construction of the multi-branch prediction head architecture: Construct three parallel multilayer perceptron (MLP) branches, each responsible for a different inference task. The specific steps are as follows: S3-1. Classifier Head ): Used to identify specific subtypes of secreted proteins. For example, the output dimension is... After Softmax activation, the conditional probability distribution is obtained. , representing that the sample belongs to the first The probability of secretory proteins.
[0017] S3-2. Binary Classification Head (Detector Head) Function: Used to determine whether a protein is a secretory protein. For example, the output dimension is 1, outputting the original logical value. It should be noted that during the training phase, the original logistic values... The bias term will be corrected; during the inference phase, the probability is directly output via the Sigmoid function.
[0018] S3-3. Bias Estimator Head ): Used to learn the likelihood difference factor for each category in the current sample. For example, the output dimension is After processing by the Softplus activation function, a non-negative vector is output, where ,in This output is only used to calculate bias during the training phase and is not used for inference.
[0019] S4. Statistical bias calibration mechanism: Through mathematical derivation, the statistical bias caused by data imbalance is transformed into a correction quantity that the model can optimize, which is the core of this invention.
[0020] S4-1. Category Perception Bias Item Calculation: Based on Bayes' theorem, combined with the prediction probability of multi-classifier heads. Prior frequency Likelihood difference factor of the output of the bias estimation head Calculate the degree of bias in the current sample: If a sample belongs to the scarce category, such as T2, It is extremely small, and the multi-classification head correctly identified it. If it is larger, then The value will become very large, indicating that the sample is very likely to be misclassified as a negative sample due to data scarcity, and compensation is needed.
[0021] S4-2. Detection Logic Correction Amount Calculate: the deviation term Convert to binary header logical value Additive corrections to space: For scarce positive samples When the value is positive, the predicted score is artificially lowered during loss calculation, forcing the model to output a larger original value during optimization. This will improve the recall rate.
[0022] S5. Model Training and Loss Function: End-to-end joint training is adopted, and the total loss function consists of three parts.
[0023] S5-1. Classification Loss Only for positive samples ( The calculation employs a Logit Adjustment strategy to further mitigate imbalances at the multi-class classification level. ; in For specific subtype labels.
[0024] S5-2. Deviation calibration detection loss Calculate the binary cross-entropy loss using the modified Logit: ; in For authenticity, secreted protein Non-secretory proteins ,pass Loss forced binary classification head original logic value Learn the fair decision boundary after bias calibration.
[0025] S5-3. Probabilistic Constraint Loss : To prevent overflow of the probability space after calibration.
[0026] ; in This represents the Sigmoid function.
[0027] S5-4. Total loss of the model: ;in and This is the balancing coefficient, set as a hyperparameter. It can be set. , .
[0028] S6. Reasoning Stage: The reasoning stage has a simple process and no longer involves complex deviation calculations.
[0029] Protein sequences are input into the model for feature extraction, and sequence feature vectors are output. ; First-level judgment: Predicted probability of secreted proteins output by binary classifier. : ,like It is determined to be a secretory protein. It is determined to be a non-secretory protein, and T is a preset threshold, which can be 0.5.
[0030] Second-level judgment: If it is determined to be a secretory protein, then the maximum value output by the multi-classification head is taken. As the final type of secreted protein.
[0031] S7. Model Evaluation The trained model is evaluated using a separate test set. Common evaluation metrics include: Binary classification task: Area under the ROC curve (AUC-ROC), Area under the PR curve (AUC-PR), accuracy, Matthews Correlation Coefficient (MCC).
[0032] Multi-class classification task: Macro-F1 score, macro-recall, precision and recall per class, and confusion matrix. Special attention is paid to recall metrics for subclasses such as T1, T2, and T6.
[0033] This invention presents a secretory protein identification method that integrates prior and bias dynamic calibration. It designs a hierarchical identification architecture that incorporates prior frequencies, explicitly embedding global statistical information from the training data into the computational graph of the neural network. This allows the deep learning model to incorporate Bayesian statistical priors, thereby exhibiting superior generalization ability on small sample classes. Simultaneously, this invention innovatively designs an unsupervised bias estimation head based on likelihood difference factors, which learns the bias of each class... The factor quantifies the likelihood difference of samples under balanced and imbalanced distributions, providing a basis for subsequent calibration. Finally, this invention also proposes a category-aware bias correction loss function: [The invention then describes a method for correcting class perception bias.] The loss calculation method differs from the traditional weighted loss. This method dynamically corrects the Logit value, which can more accurately correct the gradient for scarce positive samples that are misclassified as negative samples.
[0034] The above description is merely a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for identifying secretory proteins that integrates statistical prior and dynamic bias calibration, characterized in that: Includes the following steps: A protein sequence training dataset with hierarchical labels was constructed and divided into a positive sample set and a negative sample set. Based on the positive sample set, the prior frequency of each class of secreted proteins was calculated. ; Feature extraction is performed on the protein sequence to obtain global feature representation and end feature representation, which are then concatenated to obtain the final sequence feature vector. ; sequence feature vector The input consists of a multilayer perceptron with three parallel branches. Branch one is a multi-classifier head used to identify specific subtypes of secreted proteins, and the output is a predicted probability distribution. 'k' is the type number of the secretory protein; branch two is a binary classification head used to determine whether a protein is a secretory protein, outputting the original logical value. Branch three is used to learn the likelihood difference factor for each class in the current sample. The deviation estimation head; Based on Bayes' theorem, and combined with the prediction probability of multi-classifier heads... Prior frequency and likelihood difference factor Calculate the category perception bias term for the current sample. Then, the category perception bias item Convert to binary classification head raw logic value Additive corrections in space ; The model's total loss function includes category classification loss. Deviation calibration detection loss and probability constraint loss used to prevent overflow of the probability space after calibration .
2. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The positive sample set contains various categories of secreted proteins that have been experimentally verified, and the negative sample set contains non-secreted proteins, serving as auxiliary training data; prior frequencies The statistical steps are as follows: Count the number of samples of each type of secreted protein in the positive sample set. and the total number of all positive samples Calculate the prior frequency of each type of secreted protein. .
3. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The sequence feature vector The generation steps are as follows: The pre-trained protein language model ProtT5 is used as a feature extractor. The protein sequence is input into the feature extractor one by one, and the output is a residue-level embedding matrix. L represents the length of the protein sequence; features are further extracted from this residue embedding matrix to obtain the final global feature representation. The position-specific end signal features in the residue embedding matrix are captured through a window of preset length, and then further features are extracted from the end signal features to obtain the N-terminal feature vector. and C-end feature vector ; will respectively and The input is fed into a single-layer convolutional neural network with 9 kernels, and after max pooling, the N-terminal feature vector is obtained. and C-end feature vector Representing global features N-terminal feature vector and C-end feature vector The features are concatenated to obtain the final sequence feature vector. The sequence feature vector Simultaneously input into the three branches of the multilayer perceptron.
4. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The category perception bias item The calculation formula is: ； If a sample belongs to the scarce category It is extremely small, and the multi-classification head correctly identified it. If it is larger, then The value will become very large, indicating that the model is very likely to misclassify the sample as a negative sample due to data scarcity, and compensation is needed.
5. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The additive correction amount The calculation formula is: ； For scarce positive samples When the value is positive, the predicted score is artificially lowered during loss calculation, forcing the model to output a larger original value during optimization. This will improve the recall rate.
6. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The deviation calibration detection loss It uses the modified Logit: The binary cross-entropy loss is calculated using the following formula: ； in For authenticity, secreted protein Non-secretory proteins ,pass Loss forced binary classification head original logic value Learn the fair decision boundary after bias calibration.
7. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The probability constraint loss The calculation formula is: ； in This represents the Sigmoid function.
8. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The model's total loss function is: ,in and This is the balance coefficient, used as a hyperparameter setting.
9. The method for identifying secretory proteins by fusing statistical prior and dynamic bias calibration according to claim 1, characterized in that: The trained model is then used in the inference phase, following these steps: Protein sequences are input into the model for feature extraction, and sequence feature vectors are output. '; Level 1 Judgment: Calculation ,like It is identified as a secretory protein, and T is a preset threshold. Second-level judgment: If it is determined to be a secretory protein, then the maximum value output by the multi-classification head is taken. As the final type of secreted protein.