Method, device and equipment for identifying acidophilic protein

Abstract

The invention discloses an acidophilic protein recognition method, device and equipment. The method comprises the following steps: constructing a data set which takes acidophilic protein as a positive sample and takes thermophilic, halophilic and basophilic protein as a negative sample, and preprocessing the data set; dividing the data set into a training set and a test set according to a preset proportion; performing feature coding on the protein sequence in the data set by using a pre-trained ESM C model, extracting deep semantic features, and generating high-dimensional embedded representation; learning high-dimensional embedding representation through an adversarial network DCGAN-GP to perform data enhancement, and obtaining a pseudo-negative class sample; combining the pseudo negative sample and the data set to obtain a combined set, and inputting the combined set into a lightweight shared sparse hybrid expert model for training to obtain a trained lightweight shared sparse hybrid expert model; and inputting a to-be-identified protein sequence into the trained lightweight shared sparse hybrid expert model to obtain an acidophilic protein prediction result. According to the method, efficient prediction of the acidophilic protein can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics, and in particular to a method, device and equipment for identifying acidophilic proteins. BACKGROUND

[0002] In modern industrial production, high efficiency and environmental sustainability have become core goals. However, achieving this dual goal still faces significant challenges, especially in enzyme catalysis. Although biological catalysts are widely studied for their specificity and environmental friendliness, traditional enzymes are easily inactivated in extreme industrial environments (such as low pH, high temperature, heavy metal pollution), which seriously affects production efficiency and sustainability. In this context, acidophilic proteins, as a natural adaptation protein derived from extreme acidic environments (such as acid mine, volcanic hot spring), have become an important industrial enzyme candidate due to their excellent stability and activity in strong acid environments. Its application prospects are wide-ranging, covering pollutant degradation, food processing, bioenergy and bio-mining, and many other fields. Therefore, efficient identification and development of acidophilic proteins have important value for promoting green industrial technology.

[0003] Traditionally, the identification of acidophilic proteins relies on a two-stage experimental process: first, through genomic analysis, including extracting DNA from environmental samples, performing whole genome sequencing and functional annotation, and screening potential acidophilic protein coding genes from them; then, through biochemical experiments to verify their enzyme activity and stability in acidic environments. This process has high accuracy, but also faces the limitations of high resource consumption and high time cost. In order to improve efficiency, researchers have proposed computational-based acidophilic protein identification methods. Early methods mostly use manually designed sequence features for classification modeling. For example, Khan et al. used pseudo-amino acid composition (PseAAC) features combined with probabilistic neural network (PNN) for classification; Lin et al. used GGAC features combined with support vector machine (SVM) model; Fan et al. used AAC and acACS features to input SVM classifier; Zhang et al. combined protein secondary structure information and amino acid physicochemical index to construct a random forest model. This method highly depends on manually constructed sequence features, which is not only tedious and time-consuming, but also often requires bioinformatics expertise. In addition, handcrafted features are difficult to fully capture the context and semantic information of protein sequences, resulting in weak model generalization ability and difficulty in adapting to protein sequences of different sources or different backgrounds.

[0004] In recent years, Meredita Susanty et al. introduced the method of combining pre-trained protein language model (PLM) with transfer learning to learn semantic representation from large-scale protein databases, automatically generating embedding vectors of protein sequences for subsequent classification tasks. This method no longer relies on manual feature engineering, and has stronger adaptability and feature expression ability. However, these methods fail to fully utilize the capabilities of modern deep learning in sequence modeling and fail to tap the potential of protein language models in cross-task transfer learning. SUMMARY

[0005] To solve the above technical problems, the present application proposes a method, device and equipment for identifying acidophilic proteins, which integrates protein language model (ESM C), adversarial generative network DCGAN-GP and lightweight shared sparse hybrid expert model to achieve efficient prediction of acidophilic proteins.

[0006] To achieve the above purpose, the technical scheme of the present application is as follows:

[0007] A method for identifying acidophilic proteins, comprising the following steps:

[0008] A data set is constructed with acidophilic proteins as positive samples and thermophilic, halophilic and alkaliphilic proteins as negative samples, and the data set is preprocessed; the data set is divided into a training set and a test set according to a predetermined ratio;

[0009] The pre-trained ESM C model is used to encode the protein sequences in the data set, extract deep semantic features, and generate high-dimensional embedding representations;

[0010] The high-dimensional embedding representations are learned by the adversarial network DCGAN-GP for data augmentation to obtain pseudo-negative samples;

[0011] The pseudo-negative samples and the data set are merged to obtain a merged set and input into the lightweight shared sparse hybrid expert model for training to obtain a trained lightweight shared sparse hybrid expert model;

[0012] The protein sequence to be identified is input into the trained lightweight shared sparse hybrid expert model to obtain the prediction result of the acidophilic protein.

[0013] Preferably, the preprocessing includes removing sequences containing "X", "B", "Z" ambiguous amino acid codes and fragments with a length of less than 100; MMSeqs2 software is used for redundancy processing to control the sequence similarity within 20%.

[0014] Preferably, the adversarial network DCGAN-GP includes a generator and a discriminator introducing CNN, and an adversarial training process of the adversarial network DCGAN-GP includes the following steps:

[0015] The Gaussian random noise and the high-dimensional embedding representation are taken as inputs of the generator, mapped to the latent space representation through a fully connected layer, and then generated pseudo samples through two one-dimensional convolutional layers for stepwise upsampling; the high-dimensional embedding representation or the pseudo sample is taken as an input of the discriminator, features are extracted through a convolutional neural network, a full connection layer is combined to complete binary classification discrimination and calculate a loss function, and parameters of the generator and the discriminator are updated through a back propagation algorithm;

[0016] The generator and the discriminator are subjected to adversarial training, and a gradient penalty mechanism is introduced in the training process to meet a 1-Lipschitz continuity constraint condition, for dynamically constraining a gradient range of the discriminator.

[0017] The trained generator is used to synthesize pseudo negative class samples.

[0018] Preferably, in the adversarial training, a parameter update step frequency of the discriminator and the generator adopts M:1, and the M is an integer not less than 2.

[0019] Preferably, the lightweight shared sparse mixed expert model includes a plurality of stacked TransformerMoEBlock modules, and each TransformerMoEBlock module includes, in sequence, a multi-head attention mechanism, residual connection and layer normalization, a sparse MoE module, and residual connection and layer normalization.

[0020] Preferably, the sparse MoE module includes a group of experts, a shared linear layer, and a gating network, the experts include an activation function and a linear layer, and a processing process of the sparse MoE module includes:

[0021] The gating network introducing Gaussian noise calculates an activation weight of each expert for each Token, and determines a plurality of most relevant experts from the group of experts based on the activation weight by using a Top-k strategy;

[0022] The input feature vector is preliminarily transformed through the shared linear layer, and the preliminarily transformed feature vectors are respectively input into the plurality of most relevant experts for activation function operation and linear layer transformation.

[0023] The outputs of the plurality of most relevant experts are respectively weighted and summed with the corresponding activation weights to form a final feature representation of sparse mixing.

[0024] Preferably, a load balancing loss term is introduced in the pre-training of the sparse MoE module to prevent over-concentration of the experts.

[0025] Preferably, the method further comprises:

[0026] The evaluation index includes accuracy, Matthew correlation coefficient, sensitivity, specificity, precision, recall, F1 score, area under ROC curve and area under PR curve.

[0027] The evaluation index includes accuracy, Matthew correlation coefficient, sensitivity, specificity, precision, recall, F1 score, area under ROC curve and area under PR curve.

[0028] Based on the above, the application further discloses a device for identifying acidophilic proteins, comprising:

[0029] A data set construction module is configured to construct a data set with acidophilic proteins as positive samples and thermophilic, halophilic and alkaliphilic proteins as negative samples, and to preprocess the data set; and the data set is divided into a training set and a test set according to a preset ratio.

[0030] An encoding module is configured to use a pre-trained ESM C model to perform feature encoding on protein sequences in the data set, extract deep semantic features, and generate high-dimensional embedding representations.

[0031] A data enhancement module is configured to learn data enhancement of the high-dimensional embedding representations through an adversarial network DCGAN-GP to obtain pseudo-negative samples.

[0032] A training module is configured to merge the pseudo-negative samples and the data set to obtain a merged set and input the merged set into a lightweight shared sparse hybrid expert model for training to obtain a trained lightweight shared sparse hybrid expert model.

[0033] An identification module is configured to input a protein sequence to be identified into the trained lightweight shared sparse hybrid expert model to obtain a prediction result of the acidophilic protein.

[0034] Based on the above, the application further discloses a computer device, comprising a memory for storing a computer program and a processor for executing the computer program to implement the method according to any one of the above.

[0035] Based on the above technical solution, the application has the following advantages:

[0036] 1) The application combines a pre-trained large-scale protein language model ESM C to enhance the semantic understanding and expression ability of acidophilic protein sequences and significantly improve the performance of the model.

[0037] 2) The application introduces a multi-scale CNN and a gradient penalty mechanism into a high-dimensional WGAN framework, thereby meeting the 1-Lipschitz condition while taking into account the feature extraction ability and training stability. Data enhancement in the embedding space is realized through DCGAN-GP to alleviate the sample imbalance problem in the training process.

[0038] 3) The light weight shared sparse hybrid expert model adopted in the present application combines high-dimensional ESM C protein embedding modeling, noise gating network, load balancing loss and shared expert architecture, efficiently filters and dynamically routes features, balances model expression ability and computing efficiency, and realizes lightweight design of feature optimization. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 Figure 1 is a flowchart of a method for identifying acidophilic proteins in an embodiment;

[0040] Figure 2 Figure 1 is a flowchart of a method for identifying acidophilic proteins in an embodiment;

[0041] Figure 3 Figure 1 is a performance comparison chart of the model of the present application and existing typical models in an embodiment;

[0042] Figure 4 Figure 1 is an experimental comparison chart in an embodiment, wherein, Figure 4 A is a UMAP visualization analysis chart of real data and generated embedding data; Figure 4 B is a comparison chart of the results of the generated embedding on the test set using different training rounds, Figure 4 C is a comparison chart using different data enhancement methods;

[0043] Figure 5 Figure 1 is an ablation experiment comparison chart in an embodiment, wherein, Figure 5 A is a comparison chart using different protein sequence embedding methods; Figure 5 B refers to replacing the sparse MoE module with a normal Transformer, and NoGan refers to not adding the generated data of the adversarial network DCGAN-GP. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0045] As shown in Figure 1 , 2 The present embodiment provides a method for identifying acidophilic proteins, comprising the following steps:

[0046] Step 1, constructing a data set with acidophilic proteins as positive samples and thermophilic, halophilic and alkaliphilic proteins as negative samples, preprocessing the data set; dividing the data set into a training set and a test set according to a predetermined ratio.

[0047] In this example, the dataset constructed by Susanty et al. was used, which was based on protein sequences in the National Center for Biotechnology Information (NCBI) database (Sayers et al., 2021) and was systematically screened and preprocessed. In the initial dataset, the number of acidophilic proteins was 237109, much more than the 7394 non-acidophilic proteins. This difference was due to the large number of entries corresponding to the "acidophilic" keyword in the database, especially in NCBI, PDB and UniProtKB / Swiss-Prot databases. In contrast, non-acidophilic proteins (including thermophilic, halophilic and alkaliphilic) were mainly derived from other research and database resources, such as iThermo (Ahmed et al., 2022) and Sapphire (Charoenkwan et al., 2022) for the collection of thermophilic proteins, and the use of "halophilic" and "alkaliphilic" keywords in NCBI to retrieve halophilic and alkaliphilic proteins. Due to the less annotation of these categories, the total number of non-acidophilic proteins was limited. To ensure the accuracy of the label, the researchers did not simply classify the proteins not labeled as acidophilic as negative, but precisely selected thermophilic, halophilic and alkaliphilic proteins with clear environmental adaptation characteristics as the control group, avoiding the confusion caused by cross-labels (such as thermophilic-acidophilic proteins), thereby improving the accuracy and robustness of the classification task.

[0048] After preprocessing, a high-quality dataset containing 4089 acidophilic proteins and 1654 non-acidophilic proteins was constructed, and was divided into training set and test set according to the ratio of 80 / 20. The training set included 3271 positive samples and 1323 negative samples, and the test set included 818 positive samples and 331 negative samples.

[0049] Step 2, use the pre-trained ESM C model to encode the protein sequences in the dataset, extract deep semantic features, and generate high-dimensional embedding representation.

[0050] In this embodiment, the protein language model ESM C (Evolutionary Scale Modeling Cambrian) is used as a sequence encoder to extract deep semantic features of proteins. ESM C is built based on a Transformer architecture, which includes multi-head attention mechanisms, feed-forward neural networks (FNNs), and residual networks (ResNets), and is pre-trained unsupervised on large-scale protein sequence datasets. By learning the evolutionary patterns of protein sequences and their relationships between structure and function, ESM C can generate corresponding latent vectors for protein sequences to represent their structural and functional characteristics. Compared to the earlier version ESM-2, ESM C significantly improves inference efficiency while maintaining prediction accuracy.

[0051] Specifically, ESM C-600m is used for encoding, representing each amino acid as a 1152-dimensional vector, and the entire sequence is encoded as a high-dimensional embedding representation. This high-dimensional embedding representation not only preserves the composition information of the sequence, but also contains evolutionary background and functional characteristics, serving as the basis for subsequent feature optimization and classification. By introducing the ESM C model, the system can eliminate manual feature engineering and directly use the embedding generated by the pre-trained language model for end-to-end learning, effectively improving the accuracy and efficiency of protein function recognition.

[0052] Step 3: Learn data augmentation for high-dimensional embedding representation through adversarial network DCGAN-GP to obtain pseudo-negative class samples.

[0053] In this embodiment, to alleviate the class imbalance problem in protein sequence embedding data, a DCGAN-GP-based adversarial network framework is introduced, and the discriminator and generator mainly use convolutional architecture to learn the embedding features of protein sequences.

[0054] Step 4: Merge the pseudo-negative samples and the dataset to obtain a merged set and input it into the lightweight shared sparse hybrid expert model for training to obtain the trained lightweight shared sparse hybrid expert model.

[0055] In this embodiment, the obtained high-dimensional embedding representation of protein sequences is input into the lightweight shared sparse hybrid expert model. This model mainly consists of four layers of stacked TransformerMoEBloc modules, each of which includes multi-head attention mechanisms, sparse MoE modules, residual connections, and layer normalization. Each sparse MoE module includes thirty experts, and each time a token is processed, the three highest scoring experts are obtained using a gating network for fusion to obtain the final features. Noise gating networks and load balancing losses are also introduced to stabilize the performance of the sparse MoE module.

[0056] Step 5, input the protein sequence to be identified into the trained lightweight shared sparse hybrid expert model, and obtain the prediction result of acidophilic protein.

[0057] In one embodiment of the acidophilic protein identification method, the preprocessing includes removing sequences containing "X", "B", "Z" ambiguous amino acid codes and fragments with a length of less than 100 to improve data quality; MMSeqs2 software (Olenyi et al., 2022) is used for redundancy processing, and the sequence similarity between the training set and the test set is controlled within 20% to ensure the generalization ability of the model on low homology data.

[0058] In one embodiment of the acidophilic protein identification method, the adversarial network DCGAN-GP includes a generator and a discriminator that introduce CNN, and the adversarial training process includes the following steps: Gaussian random noise and high-dimensional embedding representation are taken as the input of the generator, which is mapped to the hidden space representation through the fully connected layer, and then two one-dimensional convolution layers are used for stepwise upsampling to generate pseudo samples; the high-dimensional embedding representation or the pseudo samples are taken as the input of the discriminator, which extracts features through the convolutional neural network, combines the fully connected layer to complete the binary classification and calculate the loss function, and updates the parameters of the generator and the discriminator through the back propagation algorithm; the generator and the discriminator are trained adversarially, and a gradient penalty mechanism is introduced in the training process to meet the 1-Lipschitz continuity constraint condition, which is used to dynamically constrain the gradient range of the discriminator; the trained generator is used to synthesize pseudo negative class samples.

[0059] In this embodiment, to solve the technical problems of serious imbalance between positive and negative samples in protein sequence embedding data, complex distribution of negative class and high dimension, a high-dimensional negative sample enhancement method based on DCGAN-GP is proposed. Unlike traditional methods of directly disturbing in the original sequence space or generating samples through simple oversampling, this embodiment first models the negative class distribution in the high-dimensional embedding space generated by ESM C, which can generate high-fidelity and high-diversity pseudo-negative samples while maintaining the semantic consistency of protein sequences. Specifically, the generator takes 128-dimensional Gaussian random noise as input, first maps it to the hidden space representation through the fully connected layer, and then gradually upsamples through two layers of one-dimensional convolution (Conv1d) to generate pseudo samples with the same size as the real embedding tensor (dimension (batch_size, 300, 1152)), which fully captures local patterns and global dependencies through a multi-scale convolutional structure; the discriminator receives real or generated embedding data, extracts features through a convolutional neural network, and combines a fully connected layer to complete binary classification discrimination, thereby effectively guiding the generator to approximate the real distribution. At the same time, a gradient penalty mechanism is introduced during training to satisfy the 1-Lipschitz continuity constraint, combined with a 5:1 discriminator-generator update step frequency strategy, which significantly improves the stability of adversarial training and reduces the risk of mode collapse. Finally, a large number of high-quality pseudo-negative samples are synthesized using the trained generator, and the imbalance of sample distribution is significantly alleviated after fusion with the original dataset, thereby improving the overall performance of the classification model. Compared with the method of directly constructing negative samples in the original sequence space, this strategy effectively avoids the semantic drift problem caused by disturbing the sequence. After training, a large number of high-quality pseudo-negative embedding samples are synthesized using the generator, and the imbalance of sample distribution is significantly alleviated after fusion with the original dataset. This strategy not only improves the sample diversity of the minority class, but also significantly enhances the recognition ability of the downstream classification model on the minority class, improving the overall generalization performance and robustness.

[0060] The multi-scale CNN and gradient penalty mechanism are combined into the high-dimensional WGAN framework, so that the 1-Lipschitz condition is satisfied while the feature extraction ability and training stability are considered. Specifically, the multi-scale convolutional structure of CNN can efficiently capture local patterns and global dependencies in protein sequences, significantly improving the modeling ability of the generator for high-dimensional features; while the gradient penalty dynamically constrains the gradient range of the discriminator during adversarial training, effectively suppressing the problems of gradient explosion and gradient disappearance. Compared with traditional GAN, this dual optimization strategy not only improves the convergence speed and generation quality of the model, but also significantly reduces the risk of mode collapse, so that the model has stronger robustness and generalization ability on high-dimensional and complex protein data distribution. The formula is as follows:

[0061]

[0062] where denotes the gradient penalty loss for discriminator training, denotes the gradient penalty weight coefficient, denotes the expectation of the interpolated sample , i.e., averaging within a batch, denotes the output of the discriminator.

[0063] In a method for identifying acidophilic proteins in an embodiment, a light shared sparse mixed expert model (LS-MoE) includes a plurality of stacked Transformer MoE Block modules, the Transformer MoE Block module including, in sequence, a multi-head attention mechanism, a residual connection and layer normalization, a sparse MoE module, a residual connection and layer normalization. The sparse MoE module includes a group of experts, a shared linear layer, and a gating network, the experts including an activation function and a linear layer, and the processing process of the sparse MoE module including: the gating network introducing Gaussian noise to calculate the activation weight of each Token expert, and determining the most relevant multiple experts from the group of experts based on the activation weight using a Top-k strategy; the input feature vector is preliminarily transformed by the shared linear layer, and the preliminarily transformed feature vectors are respectively input into the most relevant multiple experts for activation function operation and linear layer transformation; the outputs of the most relevant multiple experts are respectively weighted and summed with the corresponding activation weights to form the final feature representation of sparse mixing.

[0064] In the embodiment, to achieve more efficient feature extraction and optimization, the present application proposes a light shared sparse mixed expert model based on Transformer as a main classification network, and the backbone structure includes 4 Transformer MoE Block modules, each module containing a standard multi-head attention mechanism, a residual connection, layer normalization, and a sparse MoE module.

[0065]

[0066]

[0067]

[0068] In the case where the input sequence is represented as , the query vector , the key vector , and the value vector are obtained by projection matrix . In attention calculation, denotes the dimension of the key vector, used to scale the dot product result. In multi-head attention, denotes the output of the attention head, denotes the number of attention heads, and the output transformation matrix is finally obtained by linear mapping to obtain the result.

[0069] The sparse MoE module selects multiple sub-experts for activation through a gating mechanism, effectively improving network capacity and avoiding redundant calculations. Each sparse MoE module contains 30 MLP experts, and the NoisyTop-k strategy is used to select 3 experts at each position to participate in forward propagation.

[0070]

[0071] wherein, is the final weighted output result of the MoE model, is the total number of experts, denotes the index of the selected expert, denotes the input vector, denotes the prediction result of the expert numbered on the input , denotes the score value of the expert , denotes the sum of the score values of all selected experts, denotes the normalized weight of the expert.

[0072] 1) Shared linear layer expert segmentation structure

[0073] Traditional sparse MoE module: Each expert is composed of an independent two-layer MLP (linear layer → activation function → linear layer), and the experts are completely independent, with high computational overhead.

[0074] Sparse MoE module in this embodiment: 30 experts share the output of a high-dimensional linear layer, and it is divided into 30 sub-blocks; each sub-block is then transformed and learned through its own independent activation function and linear layer.

[0075] In the design of expert networks, this embodiment adopts an efficient shared linear layer + expert segmentation (Shared-Linear Split Expert) architecture, which significantly optimizes computational efficiency and model generalization performance. The beneficial effects achieved are: reducing computational resource consumption: sharing a high-dimensional linear layer significantly reduces parameter quantity and memory occupancy; adapting to high-dimensional ESM C embedding: improving the modeling ability of the model for high-dimensional features of protein sequences; alleviating overfitting problem: sharing parameters realizes implicit regularization while preserving expert independence.

[0076] 2) Noise gating network

[0077] Gaussian noise is introduced into traditional gating networks to enhance the diversity of expert selection. Specifically, the original scoring formula for each expert in the corresponding traditional gating network when selecting an expert to process this token is as follows:

[0078]

[0079] in This represents the representation of the i-th token. For the linear layer weights of the gated network, This indicates gating bias.

[0080] After introducing Gaussian noise, the gated network becomes:

[0081]

[0082] in, The Gaussian noise is represented by this improvement, which allows the model to explore more expert combinations in each training iteration, effectively avoiding the "expert collapse" phenomenon in traditional MoE, where the model relies on a few experts for a long time, thereby improving the generalization ability of the overall model. Let represent a multidimensional normal distribution with a mean vector of 0 and a covariance matrix of . ,in It is an identity matrix, meaning that the dimensions are independent and uncorrelated. This indicates that the variance of each dimension is the same, and the noise intensity is determined by... control, The larger the disturbance, the stronger it is; the closer it is to zero, the smaller the noise. This is isotropic Gaussian noise, with each component being independent and identically distributed. In other words, the added noise is uniform in all directions and is not biased towards any particular direction.

[0083] 3) Load balancing losses

[0084] To further promote balanced training among experts, a load balancing loss is introduced to encourage greater diversity among experts when processing tokens. The load balancing loss is defined as:

[0085]

[0086] in, Represents the balanced loss function. Represents the total number of samples. The number of output components (e.g., number of categories or output dimensions) for each sample. This represents the loss value or loss contribution of the i-th sample to the b-th output component. Through joint optimization... We avoid the problem of overusing a single expert and ensure that each expert is fully updated during training, thereby improving the overall expressive power of the model.

[0087] This embodiment combines high-dimensional ESM C protein embedding modeling, Gaussian noise gating network, load balancing loss and shared expert architecture for the first time, solving problems such as low computational efficiency, poor generalization and unstable generation quality that have not been addressed in the prior art.

[0088] One embodiment of the method for identifying eosinophilic proteins further includes: determining evaluation metrics, performing 10-fold cross-validation on a lightweight shared sparse hybrid expert model, and testing it on an independent test set; the evaluation metrics include accuracy, Matthews correlation coefficient, sensitivity, specificity, precision, recall, F1 score, area under the ROC curve, and area under the PR curve.

[0089] In this embodiment, to comprehensively evaluate the model's performance in the eosinophil protein recognition task, several classic classification metrics were employed, including accuracy (ACC), Matthews correlation coefficient (MCC), sensitivity (Sn), specificity (Sp), precision, recall, F1 score, area under the ROC curve (AUC), and area under the PR curve (AUPRC). These metrics evaluate the model's classification ability from different dimensions, ensuring strong generalization ability and stable and reliable prediction results. The calculation formulas for the evaluation metrics are as follows:

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] In this model, TP, TN, FN, and FP represent the number of true positives, true negatives, false negatives, and false positives, respectively. Sn represents the proportion of correctly identified positive samples. Sp represents the proportion of correctly identified negative samples. Accuracy (ACC) represents the proportion of correctly classified samples out of the total number of samples, measuring the overall classification accuracy of the model. The Matthews correlation coefficient (MCC) measures the correlation between the predicted results and the true values, with a value range of [-1,1][-1,1][-1,1], where 1 represents perfect agreement, 0 represents random prediction, and -1 represents complete disagreement. Sensitivity (Sn) represents the model's ability to identify all positive samples, i.e., the proportion of true positives out of all actual positives. Specificity (Sp) reflects the model's ability to identify negative samples, i.e., the proportion of true negatives out of all actual negatives. Precision represents the proportion of samples predicted as positive that are actually positive. Furthermore, to comprehensively compare the performance of different models, the area under the receiver operating characteristic (ROC) curve (AUC) also needs to be calculated. The ROC curve depicts the relationship between the true positive rate (TPR) and the false positive rate (FPR) at different classification thresholds. The AUC value ranges from 0 to 1, with higher values ​​indicating stronger overall classification ability of the model. Generally speaking, higher values ​​for these metrics indicate better model performance and stronger generalization ability in the eosinophil protein identification task.

[0097] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0098] experiment

[0099] 1. Comparison of experimental results with typical deep learning models on benchmark datasets

[0100] To demonstrate the superiority of the model in this application (AcidNetPro), it was compared with several typical models, including ACE (currently the best-performing model), SVM (Support Vector Machine, the baseline model), and LR (Logistic Regression, the baseline model), such as... Figure 3 As shown in Figure 1A, this is the result of the 10-fold cross-validation. B represents the performance comparison on the independent test set. In identifying eosinophils, this application outperforms all typical pre-trained models and deep learning methods in terms of Sn (sensitivity), Sp (specificity), ACC (accuracy), AUC (area under the receiver operating characteristic curve), and MCC (Matthews correlation coefficient). This indicates that this application effectively mines both local and global information from the protein sequence. Compared to typical pre-trained models and deep learning methods, it demonstrates superior capabilities in eosinophil identification and prediction.

[0101] 2. Ablation experiment

[0102] To further verify the reliability of the model, different protein sequence embedding methods, such as... Figure 5 A, different sampling methods, such as Figure 4 C, ablation experiments were conducted with different training rounds, such as... Figure 4 B, and removing the embedded data generated by the adversarial network, using a standard Transformer architecture for feature engineering, and comparing various metrics, it can be seen that using adversarial networks for data augmentation to address the data imbalance problem and using ESM C-600m for data augmentation are correct choices in this application. Figure 4 A UMAP visualization analysis of the real data and the generated embedded data shows the effectiveness of the embedded data generated by the adversarial network DCGAN-GP. Figure 4 B. Comparison of embeddings generated by GANs with different training epochs on the test set. We selected 300 epochs of training. It can be seen that the embeddings are only slightly lower in the Sp and Precision metrics, while the other metrics are the best. Figure 4 C represents different data augmentation methods, using DCGAN_GP, which is compared with GAN methods that remove convolutions (where the generator and discriminator use only simple linear layers, removing the convolutional network), weighted sampling, and upsampling. Figure 4 In DeconvolutionalGAN, the discriminator and generator use a simple MLP network instead of convolutions. WeightedSampling means weighted sampling, and Upsampling means upsampling. Figure 5 In B, Transformer refers to replacing the sparse MoE module with a regular Transformer, and NoGan refers to the data generated without the adversarial network DCGAN-GP.

[0103] In one embodiment, an eosinophil protein recognition device is provided, comprising:

[0104] The dataset construction module is used to construct a dataset with eosinophilic proteins as positive samples and thermophilic, halophilic, and basophilic proteins as negative samples, and to preprocess the dataset.

[0105] The encoding module is used to encode the protein sequences in the dataset using a pre-trained ESM C model, extract deep semantic features, and generate high-dimensional embedding representations.

[0106] The data augmentation module is used to augment data by learning high-dimensional embedding representations through the adversarial network DCGAN-GP, thereby obtaining pseudo-negative class samples;

[0107] The training module is used to merge pseudo-negative samples and the dataset to obtain a merged set, which is then input into the lightweight shared sparse hybrid expert model for training, resulting in a trained lightweight shared sparse hybrid expert model.

[0108] The identification module is used to input the protein sequence to be identified into a trained lightweight shared sparse mixture expert model to obtain the prediction results of eosinophilic proteins.

[0109] In the above embodiments, each module of the eosinophil protein recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0110] In one embodiment, a computer device is provided, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps as described in all the above method embodiments.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for identifying eosinophilic proteins, characterized in that, Includes the following steps: A dataset is constructed with eosinophilic proteins as positive samples and thermophilic, halophilic, and basophilic proteins as negative samples. The dataset is preprocessed and then divided into training and testing sets according to a preset ratio. The pre-trained ESM C model is used to encode the protein sequences in the dataset, extract deep semantic features, and generate high-dimensional embedding representations. Data augmentation is performed by learning high-dimensional embedding representations through the adversarial network DCGAN-GP to obtain pseudo-negative class samples; The pseudo-negative samples and the dataset are merged to obtain the merged set, which is then input into the lightweight shared sparse hybrid expert model for training, resulting in a trained lightweight shared sparse hybrid expert model. The protein sequence to be identified is input into a trained lightweight shared sparse mixture expert model to obtain the prediction results of eosinophilic proteins.

2. The method for identifying eosinophilic proteins according to claim 1, characterized in that, The preprocessing includes removing sequences containing ambiguous amino acid codes for "X", "B", and "Z" as well as fragments shorter than 100; redundancy is handled using MMSeqs2 software to keep the sequence similarity below 20%.

3. The method for identifying eosinophilic proteins according to claim 1, characterized in that, The adversarial network DCGAN-GP includes a generator and a discriminator incorporated from a CNN. The adversarial training process of DCGAN-GP includes the following steps: Gaussian random noise and high-dimensional embedding representation are used as input to the generator, which is mapped to the latent space representation through a fully connected layer, and then gradually upsampled through two layers of one-dimensional convolution to generate pseudo-samples. The high-dimensional embedding representation or pseudo-samples are used as input to the discriminator, which extracts features through a convolutional neural network and combines them with a fully connected layer to complete binary classification and calculate the loss function. The parameters of the generator and discriminator are updated through the backpropagation algorithm. Adversarial training is performed on the generator and discriminator. A gradient penalty mechanism is introduced during the training process to satisfy the 1-Lipschitz continuity constraint, which is used to dynamically constrain the gradient range of the discriminator. Pseudo-negative class samples are synthesized using the trained generator.

4. The method for identifying eosinophilic proteins according to claim 3, characterized in that, In adversarial training, the parameter update step frequency of the discriminator and the generator adopts M:1, where M is an integer not less than 2.

5. The method for identifying eosinophilic proteins according to claim 1, characterized in that, The lightweight shared sparse hybrid expert model includes several stacked TransformerMoEBlock modules, each of which includes a multi-head attention mechanism, residual connections and layer normalization, a sparse MoE module, and residual connections and layer normalization connected in sequence.

6. The method for identifying eosinophilic proteins according to claim 5, characterized in that, The sparse MoE module includes a set of experts, a shared linear layer, and a gated network. The experts include activation functions and linear layers. The processing procedure of the sparse MoE module includes: A gating network with Gaussian noise is introduced to calculate the activation weight of an expert for each token. Based on the activation weight, a Top-k strategy is used to determine the most relevant experts from a set of experts. The input feature vector undergoes a preliminary transformation through a shared linear layer. The pre-transformed feature vector is then input into the most relevant experts for activation function operations and linear layer transformations. The outputs of the most relevant experts are weighted and summed with their corresponding activation weights to form a sparse hybrid final feature representation.

7. The method for identifying eosinophilic proteins according to claim 6, characterized in that, The sparse MoE module incorporates a load balancing loss term during pre-training to prevent excessive concentration of experts.

8. The method for identifying eosinophilic proteins according to claim 1, characterized in that, Also includes: We determined the evaluation metrics, performed 10-fold cross-validation on the lightweight shared sparse hybrid expert model, and tested it on an independent test set. The evaluation metrics include accuracy, Matthews correlation coefficient, sensitivity, specificity, precision, recall, F1 score, area under the ROC curve, and area under the PR curve.

9. A device for recognizing eosinophilic proteins, characterized in that, include: The dataset construction module is used to construct a dataset with eosinophilic proteins as positive samples and thermophilic, halophilic, and basophilic proteins as negative samples, and to preprocess the dataset. The dataset is divided into training and testing sets according to a preset ratio; The encoding module is used to encode the protein sequences in the dataset using a pre-trained ESM C model, extract deep semantic features, and generate high-dimensional embedding representations. The data augmentation module is used to augment data by learning high-dimensional embedding representations through the adversarial network DCGAN-GP, thereby obtaining pseudo-negative class samples; The training module is used to merge pseudo-negative samples and the dataset to obtain a merged set, which is then input into the lightweight shared sparse hybrid expert model for training, resulting in a trained lightweight shared sparse hybrid expert model. The identification module is used to input the protein sequence to be identified into a trained lightweight shared sparse mixture expert model to obtain the prediction results of eosinophilic proteins.

10. A computer device, characterized in that, Includes: memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 8.

Images