Secreted effector protein recognition method based on representation information amount guidance

Abstract

This invention relates to a method for identifying secretory effector proteins based on representational information content. It assesses the identification difficulty of each protein class by measuring the dispersion of its distribution in the feature space, and estimates the information content using a diagonal approximation method. This generates a dynamic additive margin, transforming abstract feature complexity into a penalty term that can be directly applied to the loss function. This achieves a dynamic correlation between class difficulty and the decision boundary. A novel loss function is constructed, which subtracts the dynamic margin from the cosine similarity of the target class, amplifies it uniformly by a scaling factor to obtain a logical value, and then calculates the cross-entropy. This forces the model to reserve a larger classification margin for high-difficulty classes, effectively alleviating the bias problem of traditional cross-entropy loss on the majority class. The advantages of this invention are: significantly improved model recognition accuracy for high-difficulty, low-sample classes under long-tailed distributions.

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 effector proteins based on characterization information. Background Technology

[0002] The main bottleneck in current technologies for identifying and classifying Gram-negative bacterial secretory effector proteins (T1SE, T2SE, T3SE, T4SE, T6SE) lies in the mismatch between the extremely long-tailed distribution of the data and the feature complexity. For example, the number of T4 secretory protein samples may be in the hundreds, while the number of T2 samples may be only in the tens. Traditional solutions mainly rely on resampling or reweighting based on class frequency (i.e., the total number of samples of a certain class in the entire dataset). These methods implicitly assume that "classes with fewer samples are harder to learn." However, in biological sequence data, the difficulty of class identification is not always consistent with the number of samples. In addition to sample size, the dispersion of class distribution in the feature space and the overlap of inter-class boundaries also affect the classification difficulty. Resampling or reweighting based solely on the number of samples cannot accurately reflect the true discriminative complexity of each class, thus easily leading to unreasonable allocation of decision boundaries, which in turn affects the generalization and identification performance of high feature complexity classes. Summary of the Invention

[0003] To address the limitations of existing technologies that rely solely on sample size for balancing, this invention aims to provide a method for identifying secretory effector proteins based on Information Amount (IA). By guiding dynamic boundary adjustments through information amount, this method achieves accurate identification and classification of secretory effector proteins.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying secretory effector proteins based on representational information content includes a recognition model, the training process of which is as follows: A training dataset of amino acid sequences of secreted proteins was constructed, with each amino acid sequence of a secreted protein labeled to record the true class. By using a pre-trained model and a convolutional neural network, protein sequences of different lengths in the training set are mapped into feature vectors h of fixed dimensions; After each epoch, the updated feature vectors of all training samples are collected. For each category, the variance of the feature vector of that category in each dimension is calculated. Then, the diagonal approximation formula is used to calculate the category variance to obtain the information content of each category. Define the conversion method between information content and dynamic additive margin: map the information content of the category to the preset maximum margin range to obtain the dynamic additive margin of the category; The classifier calculates the logical values ​​for each category: for categories that belong to the category label, the logical value of the category is adjusted by dynamic additive margin; for categories that do not belong to the category label, the logical value is not adjusted. By applying Softmax probability mapping to the logical values ​​of all categories, the probability of a sample belonging to the true category is predicted. The loss value is calculated using the cross-entropy loss function.

[0005] The present invention has the following beneficial effects: 1. Breaking through the static dependence of traditional long-tail learning on the number of samples: Unlike existing methods that rely on sample count weighting, this invention does not rely on sample count weighting, but instead assesses the recognition difficulty by measuring the dispersion of the distribution of each protein class in the feature space. This avoids the erroneous assumption that "few samples equals difficult learning," enabling the model to dynamically allocate decision resources according to the true learning difficulty of the class, significantly improving its generalization ability on long-tailed distributed data.

[0006] 2. A decision boundary based on Logit space is constructed: This invention estimates the information content using a diagonal approximation method, then generates a dynamic additive margin and introduces it into the Logit space. This forces the model to dynamically allocate decision resources based on the feature complexity of the category (rather than the number of samples), thereby directly adjusting the decision boundary position of the classifier. Unlike traditional weighted loss functions that only change the gradient magnitude, the novel loss function constructed in this invention directly adjusts the decision boundary position of the classifier by applying an adaptive margin penalty. This mechanism forces the model to reserve a larger classification margin for high-difficulty categories with highly divergent features, effectively preventing simple categories from excessively squeezing the feature space of complex categories; at the same time, for simple categories with highly concentrated features, even if their sample size is extremely small, they will not be blindly allocated excessive boundary resources. This method significantly improves the recall rate of high feature complexity categories without compromising overall accuracy.

[0007] 3. Strong training stability: The Information-based Additive Margin used in this invention performs linear penalty directly in the cosine space, resulting in constant gradients. This avoids the risk of gradient explosion when the model prediction is close to the boundary, making it particularly suitable for high-noise data such as biological sequences.

[0008] 4. Wide range of applications: This invention can 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

[0009] Figure 1 This is a schematic diagram of the model training process proposed in this invention; Figure 2 This is a schematic diagram of the secretion effector protein recognition model prediction of the present invention. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: See Figure 1 This invention discloses a method for identifying secretory effector proteins based on characterization information, comprising a training phase of the identification model and an inference phase of the trained identification model. The implementation process is illustrated by taking the identification of five types of secretory proteins—T1SE, T2SE, T3SE, T4SE, and T6SE—as examples.

[0011] The training phase includes steps S1 to S6.

[0012] S1. Data Construction and Preprocessing.

[0013] Construct a training dataset containing five types of secreted proteins: T1SE, T2SE, T3SE, T4SE, and T6SE. ,in This is the amino acid sequence of a protein. For category labels, .

[0014] S2. Feature representation learning based on hybrid architecture.

[0015] This step utilizes a pre-trained model and a convolutional neural network to map variable-length (inconsistent length) protein sequences into fixed-dimensional feature vectors. .

[0016] S21. To capture evolutionary features and local motif patterns in protein sequences, we employ the pre-trained protein language model ProtT5-XL-UniRef50 (ProtT5) as a feature extractor. Given a protein sequence of length L, ProtT5 generates a residue-level embedding matrix. Based on this embedding matrix, protein representations are constructed from two different perspectives. S22. Global Semantic Features: To characterize the overall functional properties of the protein, average pooling is performed on the embedding matrix of all residues, followed by dimensionality reduction through a fully connected layer to obtain the final global feature representation. .

[0017] S23. Local signal characteristics: The N-terminus (e.g., the first 100 residues) and C-terminus (e.g., the last 100 residues) of the truncated sequence are used as local windows. The embedding windows for the N-terminus and C-terminus are defined as follows: , 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 .

[0018] S24. 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. ; S25. Feature Projection: Projecting the concatenated feature vectors... The feature vector h is obtained by mapping to the final feature space through a fully connected layer. , where 128 is the feature dimension.

[0019] S3. Diagonal approximation and update of information content (IA): This step is the core of the invention and is used to quantify the feature complexity of each category. This embodiment uses the diagonal approximation method.

[0020] S31. Feature Collection: Updates are performed in epochs during training. In the... After each epoch, the feature embeddings of all training samples are collected. .

[0021] Variance calculation: For each category Calculate the variance of its eigenvectors in each dimension: ; in Indicates category All samples in the The values ​​in each feature dimension.

[0022] S32. Information Content (IA) Calculation: Based on the feature independence assumption, the category is calculated using the diagonal approximation formula. : ; This formula is based on the simplified assumption that the feature dimensions are independent (i.e., diagonal approximation). By aggregating the variance of each dimension, it quantifies the dispersion of this type of protein in the feature space. The larger the value, the more drastic the feature variations and the more complex the patterns of this type of protein, thus increasing the difficulty for the model to learn this category. This term is used for numerical stability to prevent logarithmic divergence when the variance is close to 0.

[0023] S4. How to generate dynamic additive margin.

[0024] This step transforms the abstract information into a penalty term that can be used in the loss function.

[0025] A linear normalization strategy is adopted to map the calculated information amount to a preset maximum margin interval [0, 1]. ], for categories Its dynamic additive margin The calculation is as follows: ; in This represents the current information content set for all categories. The maximum interval hyperparameter is set manually; in this embodiment, it is set to 0.5. The more diffuse the characteristics of a certain type of protein, the better. The larger the value, the greater the internal differences within that category, making it more difficult to learn; therefore, a larger value is given. This forces the model to push the decision boundary of that class further outward during training, reserving a larger feature space.

[0026] S5. Classification and Optimization Based on Information-based Additive Margin This embodiment adopts a completely single-head architecture, directly outputs multi-classification results, and uses an improved loss function to handle imbalance.

[0027] S51. Cosine Classifier and Construction of Logical Values: Traditional fully connected classification layers typically compute the input feature vector. The original logical value is obtained by taking the dot product of vectors. According to the classical geometric definition of the vector dot product, this calculation process can be expressed as: ; in, The magnitude of the class weight vector. The magnitude of the eigenvector. Let the cosine angle between the two vectors in the feature space represent their directional similarity. In traditional methods, a larger magnitude often dominates the dot product result, causing the model to favor the majority class with a larger sample size. To eliminate the interference of vector magnitude on classification scores and allow the model to focus on the essence of the features, this invention modifies the input feature vector... and the weight vector of each category in the classification head. All vectors are subjected to L2 normalization. After normalization, the magnitude of all vectors is forcibly scaled to 1, i.e. and At this point, the above dot product formula is strictly equivalent to the cosine similarity: ; After obtaining strict cosine similarity for each category Subsequently, before feeding it into the Softmax function, this scheme introduces dynamic additive margin and performs feature scaling operations at the logit level. The specific process is as follows: For the target category with the true label y, subtract the additive residual dynamically generated in step S4 from its target cosine similarity. Then multiply by the scaling factor s to obtain the improved logical value of the target category. : ; For non-target classes Multiplying the cosine similarity by the scaling factor s directly yields the logistic value for the non-target category. : ; The scaling factor s is a hyperparameter, set to 30 in this scheme. Through the above process, this scheme calculates the logistic values ​​of all categories after dynamic interval adjustment, which will serve as the direct input for subsequent probability mapping.

[0028] S52. IAM Loss Function Calculation: Construct an Information-based Additive Margin (IAM) loss function.

[0029] First, perform Softmax probability mapping: for a target class with a true label of y, the network predicts the probability that the sample belongs to class y. The calculation is as follows: ; Where C is the total number of categories, and in this embodiment C=5. As can be seen from the formula, the dynamic additive margin... It acts directly on the index term of the target category.

[0030] Then, perform the cross-entropy loss calculation: after obtaining the predicted probability of the target class. Then, the standard cross-entropy loss formula is used, and the negative logarithm of this probability is taken as the final loss value: .

[0031] This loss function mathematically constitutes a rigorous penalty loop: when the features of category y are extremely divergent, a larger penalty will be assigned through the calculation in step S4. This will lead to the probability calculated in step S51 Significantly reduced; thus leading to the loss value generated in step S52. The loss value increases. To minimize this loss, the network is forced to optimize the feature representation during backpropagation, increasing the classification margin between the target and non-target classes in the feature space.

[0032] An optimal training process includes: Warm-up phase: In the first 3 epochs, set the dynamic additive margin for all categories. At this point, the loss function constructed in this invention... Mathematically, it degenerates into the standard multi-class cross-entropy loss based on pure cosine similarity to obtain a preliminary feature distribution.

[0033] Dynamic calibration phase: Starting from the 4th epoch, a dynamic boundary adjustment mechanism is introduced. After each epoch, steps S3 and S4 are executed to calculate and update the category-specific dynamic additive margin. And in the next epoch of training, the updated... Substitution Gradient backpropagation is performed during the logical value calculation step.

[0034] S6. Model Evaluation.

[0035] The trained model is evaluated using an independent test set. Commonly used metrics include 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.

[0036] S7. Reasoning stage.

[0037] Please see Figure 2 This stage no longer involves calculating information content or deducting dynamic additive residuals. The model uses the trained and fixed class weight vector as a reference. The specific steps are as follows: Input the protein amino acid sequence; The feature vector is obtained by step S2. ; For vector features and the weight vector for each category Perform L2 normalization on each, then calculate the dot product of the two and multiply by the scaling factor. This yields the original logical values ​​for each category corresponding to the sample: ; The probability distribution is obtained by performing a Softmax transformation on the logistic values ​​of all categories, and the category with the highest probability is taken as the final prediction result.

[0038] This invention proposes a classification difficulty measurement method based on feature information content: it quantifies feature complexity by utilizing the variance of feature vectors of each category across various dimensions, and measures the dispersion of category distribution in the feature space with low computational cost through a diagonal approximation formula, thus improving upon the limitations of existing methods that only assess learning difficulty based on sample size. Furthermore, this invention proposes a dynamic additive residual generation strategy guided by information content: a linear normalization strategy that maps information content to additive residuals, transforming abstract feature complexity into a penalty term that can be directly applied to the loss function, achieving a dynamic correlation between category difficulty and decision boundary. Finally, this invention constructs a novel loss function that subtracts the dynamic residual from the cosine similarity of the target category, uniformly amplifies it through a scaling factor to obtain a logical value, and then calculates the cross-entropy. This forces the model to reserve a larger classification margin for high-difficulty categories, effectively alleviating the bias problem of traditional cross-entropy loss on the majority class.

[0039] 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 effector proteins based on characterization information, characterized in that: It includes a recognition model, and the training steps of the recognition model are as follows: A training dataset of amino acid sequences of secreted proteins was constructed, with each amino acid sequence of a secreted protein labeled to record the true class. By using a pre-trained model and a convolutional neural network, protein sequences of different lengths in the training set are mapped into feature vectors h of fixed dimensions; After each epoch, the updated feature vectors of all training samples are collected. For each category, the variance of the feature vector of that category in each dimension is calculated. Then, the diagonal approximation formula is used to calculate the category variance to obtain the information content of each category. Define the conversion method between information content and dynamic additive margin: map the information content of the category to the preset maximum margin range to obtain the dynamic additive margin of the category; The classifier calculates the logical values ​​for each category: for categories that belong to the category label, the logical value of the category is adjusted by dynamic additive margin; for categories that do not belong to the category label, the logical value is not adjusted. By applying Softmax probability mapping to the logical values ​​of all categories, the probability of a sample belonging to the true category is predicted. The loss value is calculated using the cross-entropy loss function.

2. The method for identifying secretory effector proteins based on characterization information as described in claim 1, characterized in that: The training dataset is ,in This is the amino acid sequence of a protein. C represents the category label; C represents the total number of categories; the variance is calculated as follows: for each category... Calculate the variance of its eigenvectors in each dimension to generate a variance set. : ,in Indicates category All samples in the The numerical values ​​in each feature dimension; the calculation method for the information content: calculating the information content of category c using the diagonal approximation formula. : .

3. The method for identifying secretory effector proteins based on characterization information as described in claim 2, characterized in that: The calculation method for the dynamic additive margin is as follows: a linear normalization strategy is adopted to map the calculated information amount to a preset maximum margin interval [0, ...]. For categories Its dynamic additive margin The calculation is as follows: ； in This represents the current information content set for all categories. The maximum interval hyperparameter is set manually.

4. The method for identifying secretory effector proteins based on characterization information as described in claim 3, characterized in that: The classifier uses a cosine classifier, setting the original logical value to equal... , where j represents the category, Let be the cosine angle between the class weight vector and the feature vector in the feature space. For the target class with label y, j=y, the output logical value is... The calculation method is as follows: For non-target categories, The output logical value The calculation method is as follows , where the scaling factor s is a hyperparameter.

5. The method for identifying secretory effector proteins based on characterization information as described in claim 4, characterized in that: The steps of the Softmax probability mapping are as follows: For a target category with a true label of y, the network predicts the probability that the sample belongs to category y. The calculation is as follows: ； Where C represents the total number of categories; The loss value is calculated as follows: .

6. The method for identifying secretory effector proteins based on characterization information as described in claim 5, characterized in that: Model training includes a warm-up phase and a dynamic calibration phase. The warm-up phase is the first t epochs, during which the dynamic additive margin for all categories is set to 0. The dynamic calibration phase starts from the (t+1)th epoch and introduces a dynamic boundary adjustment mechanism: after each epoch, the information content and dynamic additive margin for each category are calculated, and in the training of the next epoch, the updated dynamic additive margin for each category is substituted into the logical value calculation step of the loss function for gradient backpropagation.

7. The method for identifying secretory effector proteins based on characterization information as described in claim 2, characterized in that: The pre-trained model is a protein language model. Given a protein sequence of length L, the protein language model generates a residue-level embedding matrix. Further feature extraction is performed on the residue embedding matrix to obtain a global feature representation. The system captures position-specific end signal features in the residue embedding matrix through a window of preset length, and then further extracts features from the end signal features to obtain N-terminal embedding window representations and C-terminal embedding window representations. These N-terminal and C-terminal embedding window representations are then input into the convolutional neural network, and max pooling is performed to obtain the N-terminal feature vector. and C-end feature vector Representing global features N-terminal feature vector and C-end feature vector The sequences are concatenated to obtain the final sequence feature vector. The spliced ​​features The feature vector h is obtained by mapping to the final feature space through a fully connected layer.

8. The method for identifying secretory effector proteins based on characterization information as described in claim 2, characterized in that: Once the model is trained and a fixed class weight vector is obtained, the following steps are performed when using the model for inference: The input protein amino acid sequence is processed for feature extraction, and the output feature vector h is generated. L2 normalization is performed on the feature vector h and the weight vector Wj for each class, then the dot product of the two is calculated and multiplied by the scaling factor s to obtain the original logistic values ​​of the sample for each class. The calculation formula is as follows: : The probability distribution is obtained by performing a Softmax transformation on the logistic values ​​of all categories, and the category with the highest probability is taken as the final prediction result.

Images