Protein sequence fluorescence intensity prediction method, system and equipment and storage medium

Abstract

The invention belongs to the field of protein fluorescence intensity prediction, and particularly relates to a protein sequence fluorescence intensity prediction method, system and device and a storage medium, and the method comprises the steps: carrying out the feature extraction of a training data set, combining random noise, training a generator and a discriminator, obtaining a generative adversarial model, supplementing data through the generative adversarial model, and obtaining a protein sequence fluorescence intensity prediction result. The method comprises the steps of obtaining a prediction data set, then extracting semantic features of the prediction data set, performing local mode enhancement and fusion on hidden layer features to obtain final features, inputting the final features into a prediction network to train the prediction network to obtain an optimal prediction network, and inputting an actual protein sequence into the optimal prediction network. The fluorescence intensity corresponding to the actual protein sequence is obtained. The method has the effect of improving the accuracy of predicting the fluorescence intensity of the protein sequence.

Description

Technical Field

[0001] This invention belongs to the field of protein fluorescence intensity prediction, specifically relating to a method, system, device, and storage medium for predicting the fluorescence intensity of a protein sequence. Background Technology

[0002] Protein fluorescent labeling technology is an indispensable core tool in fields such as cell imaging, functional studies, and drug screening. It achieves visualization, localization, quantification, and interaction analysis of the dynamic behavior of proteins in living cells or tissues by covalently or non-covalently binding fluorescent groups (such as green fluorescent protein GFP, FITC, Cy3, Cy5, etc.) to target proteins. As life science research advances towards high-throughput, dynamic, and quantitative methods, the demand for precision, stability, and predictability in protein fluorescent labeling is becoming increasingly urgent.

[0003] In recent years, artificial intelligence, especially pre-trained language models, has shown great potential in the field of biological sequence analysis, providing a new paradigm for automatically learning functional and structural features from massive protein sequence data. However, protein fluorescence intensity prediction requires a large amount of labeled data, and the labeling of protein sequences to fluorescence intensity can only be done manually. Data growth is slow and the amount of data is small, which limits the data that the prediction model can learn from, resulting in low prediction accuracy. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, system, device and storage medium for predicting the fluorescence intensity of protein sequences. By generating adversarial networks, the data of protein sequences are enriched, so that the trained model can predict the fluorescence intensity more accurately.

[0005] A method for predicting the fluorescence intensity of a protein sequence, comprising: Obtain a labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value; The protein sequence training dataset is input into the protein language model to obtain a semantic representation vector. The semantic representation vector is then fused with the fluorescence intensity value to obtain a conditional vector. Set random noise, concatenate the random noise with the conditional vector to obtain a concatenated vector, input the concatenated vector into the generator to obtain generated data, and input the generated data and the conditional vector into the discriminator to obtain the generated data score and the conditional vector score. The generator loss is calculated based on the generated data, and the discriminator loss is calculated based on the generated data score and the conditional vector score. The optimal generator parameters and optimal discriminator parameters are obtained when the discriminator loss and generator loss are minimized. The generator and discriminator are updated using the optimal generator parameters and optimal discriminator parameters to obtain the optimal generator and optimal discriminator. The optimal generator and optimal discriminator form a generative adversarial model. Set a target protein sequence set, input the target protein sequence set into the generative adversarial model to obtain a target fluorescence intensity value set, the target protein sequence set and the corresponding target fluorescence intensity value set form a supplementary dataset, and use the supplementary dataset and the training dataset as a prediction dataset; The predicted dataset is input into the protein language model to obtain the hidden layer features of the last layer; CLS features and average pooling features are extracted from the hidden layer features, and the CLS features and average pooling features are concatenated to obtain semantic features; The semantic features are nonlinearly projected to obtain nonlinear projection features. The hidden layer features are enhanced with local patterns to obtain specific features. The nonlinear projection features and specific features are fused to obtain the final features. The final features are input into the prediction network to obtain the prediction result. Based on the prediction result and the fluorescence intensity value, the loss value is obtained. The optimal prediction network parameters with the minimum loss value are obtained. The prediction network is updated using the optimal prediction network parameters to obtain the optimal prediction network. The actual protein sequence is input into the optimal prediction network to obtain the fluorescence intensity corresponding to the protein sequence.

[0006] Optionally, the generator loss is calculated based on the generated data, including: Calculate the discriminator reward and generator reward of the generated data, and sum the discriminator reward and generator reward to obtain the expected reward; Calculate the average reward corresponding to the generated dataset; The generator loss is calculated based on the average reward and the expected reward.

[0007] Optionally, the semantic representation vector is fused with the fluorescence intensity value to obtain a conditional vector, including: The semantic representation vector is fused with the fluorescence intensity value to obtain a conditional vector, which is represented as follows:

[0008] in, For conditional vectors, For semantic representation vectors, It is a learnable embedded projection layer. Indicates splicing, For the set of real numbers, To be The dimension after converting to a vector is the dimension of the semantic representation vector.

[0009] Optionally, the step of setting random noise, concatenating the random noise with the conditional vector to obtain a concatenated vector, inputting the concatenated vector into a generator to obtain generated data, and inputting the generated data and the conditional vector into a discriminator to obtain the generated data score and the conditional vector score includes: Set random noise, and concatenate the random noise with the conditional vector to obtain the concatenated vector, which is represented as:

[0010] in, To concatenate vectors, It is random noise. It is a conditional vector; The concatenated vector is input into the generator to obtain a data matrix; Sample each position in the data matrix to obtain the sampling result; Calculate the Gumbel-Softmax distribution of the sampling results to obtain generated data. Input the generated data and the condition vector into the discriminator to obtain the generated data score and the condition vector score.

[0011] Optionally, the step of performing nonlinear projection on the semantic features to obtain nonlinear projection features, performing local pattern enhancement on the hidden layer features to obtain specific features, and fusing the nonlinear projection features and specific features to obtain the final features includes: The semantic features are then subjected to nonlinear projection to obtain nonlinear projection features, which are represented as follows:

[0012] in, It is a nonlinear projection feature. For activation function, , , and These are learnable parameters. Semantic features; The hidden layer features are processed through a lightweight convolutional network for local pattern enhancement to obtain specific features, represented as follows:

[0013]

[0014] in, This is a one-dimensional convolution output. This is a one-dimensional convolution operation. Features of the hidden layer The kernel size is [size]. This represents the number of channels in the output of the convolutional layer. This is a global max pooling operation. It is a specific characteristic; The nonlinear projection features and specific features are fused to obtain the final features.

[0015] Optionally, the step of extracting CLS features and average pooling features from the hidden layer features, and concatenating the CLS features and average pooling features to obtain semantic features, includes: Extract CLS features and average pooling features from the hidden layer features; The CLS features and average pooling features are concatenated to obtain the semantic features, represented as follows:

[0016] in, For semantic features, Features of CLS For average pooling characteristics, For splicing, These are semantic features.

[0017] Optionally, the step of inputting the final features into the prediction network to obtain the prediction result, obtaining the loss value based on the prediction result and the fluorescence intensity value, obtaining the optimal prediction network parameters when the loss value is minimized, and updating the prediction network using the optimal prediction network parameters to obtain the optimal prediction network includes: The final features are input into the prediction network to obtain the prediction result; Based on the prediction results and fluorescence intensity values, the main task loss is obtained; The protein category is obtained based on the final characteristics; The category loss is obtained based on the protein category and the prediction result; The main task loss and category loss are weighted and summed to obtain the loss value; Obtain the optimal prediction network parameters that minimize the loss value, and update the prediction network using the optimal prediction network parameters to obtain the optimal prediction network.

[0018] A protein sequence fluorescence intensity prediction system, comprising: The acquisition module is used to acquire a labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value. The fusion module is used to input the protein sequence training dataset into the protein language model to obtain a semantic representation vector, and to fuse the semantic representation vector with the fluorescence intensity value to obtain a conditional vector. The setting module is used to set random noise, concatenate the random noise with the condition vector to obtain a concatenated vector, input the concatenated vector into the generator to obtain generated data, and input the generated data and the condition vector into the discriminator to obtain the generated data score and the condition vector score. The generative adversarial model training module is used to calculate the generator loss based on the generated data, calculate the discriminator loss based on the generated data score and the conditional vector score, obtain the optimal generator parameters and optimal discriminator parameters when the discriminator loss and generator loss are minimized, update the generator and discriminator using the optimal generator parameters and optimal discriminator parameters, and obtain the optimal generator and optimal discriminator, which together form the generative adversarial model; The supplementary module is used to set a target protein sequence set, input the target protein sequence set into the generative adversarial model to obtain a target fluorescence intensity value set, the target protein sequence set and the corresponding target fluorescence intensity value set constitute a supplementary dataset, and the supplementary dataset and the training dataset are used as a prediction dataset; The hidden layer feature extraction module is used to input the prediction dataset into the protein language model to obtain the hidden layer features of the last layer. The concatenation module is used to extract CLS features and average pooling features from the hidden layer features, and concatenate the CLS features and average pooling features to obtain semantic features; An enhancement module is used to perform nonlinear projection on the semantic features to obtain nonlinear projection features, perform local pattern enhancement on the hidden layer features to obtain specific features, and fuse the nonlinear projection features and specific features to obtain the final features; The prediction model training module is used to input the final features into the prediction network to obtain the prediction result, obtain the loss value based on the prediction result and the fluorescence intensity value, obtain the optimal prediction network parameters when the loss value is minimized, update the prediction network using the optimal prediction network parameters, and obtain the optimal prediction network. The prediction module is used to input the actual protein sequence into the optimal prediction network to obtain the fluorescence intensity corresponding to the protein sequence.

[0019] A terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a protein sequence fluorescence intensity prediction method.

[0020] A computer-readable storage medium storing a computer program that, when loaded and executed by a processor, employs a method for predicting the fluorescence intensity of a protein sequence.

[0021] The beneficial effects of this invention are: 1. By setting random noise, the random noise is concatenated with the conditional vector to obtain a concatenated vector. This concatenated vector is input into the generator to obtain generated data. The generated data and the conditional vector are then input into the discriminator to obtain the generated data score and the conditional vector score. The generator loss is calculated based on the generated data, and the discriminator loss is calculated based on the generated data score and the conditional vector score. The optimal generator parameters and optimal discriminator parameters are obtained when the discriminator loss and generator loss are minimized. The generator and discriminator are updated using the optimal generator parameters and optimal discriminator parameters to obtain the optimal generator and optimal discriminator. The optimal generator and optimal discriminator form a generative adversarial model. By setting random noise, the generator and discriminator learn from each other, thereby generating a large amount of realistic training data to provide data for the prediction model training, improving the accuracy of the fluorescence intensity predicted by the prediction model.

[0022] 2. During generator training, an energy evaluation step is introduced. That is, the discriminator loss and generator loss are calculated by expectation and energy. The energy score is used as an additional reward / penalty signal to feed back into the generator's reinforcement learning training loop, so that the generator and discriminator can be trained better and the generated data is more realistic.

[0023] 3. By extracting CLS features and average pooling features from the hidden layer features, and concatenating the CLS features and average pooling features, semantic features are obtained. The semantic features are then subjected to nonlinear projection to obtain nonlinear projection features. Local pattern enhancement is performed on the hidden layer features to obtain specific features. The nonlinear projection features and specific features are then fused to obtain the final features. This step overcomes the shortcomings of traditional feature extraction processes, which mix and dilute information from all positions in the sequence. By generating position-independent local features and complementing and fusing them with global features, it enables the identification and quantification of short sequence motifs that are scattered throughout long sequences but play a decisive role in function. The extracted features are more accurate, allowing the model to learn better and improving the accuracy of fluorescence intensity prediction. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a protein sequence fluorescence intensity prediction method according to the present invention.

[0025] Figure 2 This is a network structure diagram of the predictive network of this invention. Detailed Implementation

[0026] A method for predicting the fluorescence intensity of a protein sequence, such as Figure 1 As shown, the present invention includes: S1. Obtain the labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value. Specifically, retrieve N real data entries from the existing database. ,in, It is an amino acid sequence. This is the normalized fluorescence intensity value.

[0027] S2. Input the protein sequence training dataset into the protein language model to obtain the semantic representation vector, and fuse the semantic representation vector with the fluorescence intensity value to obtain the conditional vector. The semantic representation vector is fused with the fluorescence intensity value to obtain the conditional vector, which includes: The semantic representation vector is fused with the fluorescence intensity value to obtain the conditional vector, which is represented as follows:

[0028] in, For conditional vectors, For semantic representation vectors, It is a learnable embedded projection layer. Indicates splicing, For the set of real numbers, To be The dimension after converting to a vector is the dimension of the semantic representation vector.

[0029] Among them, protein language models are pre-trained deep learning models that have been pre-trained on large-scale, unlabeled protein sequence databases (such as UniRef). These models (such as ESM-2 and ProtBERT) have learned to extract implicit evolutionary, structural, and functional semantic information from the amino acid sequences of proteins through self-supervised learning tasks (such as masked language modeling). They can transform any protein sequence into a high-dimensional, information-rich mathematical vector (i.e., a representation vector).

[0030] Specifically, the goal of the generator G is to take a random noise vector z sampled from a standard normal distribution and a conditional vector c as input, and output a soft sequence probability distribution. G is typically a multilayer perceptron (MLP). The discriminator D is a conditional discriminator that takes a sequence (or a soft sequence representation) and a conditional vector c as input, and outputs a scalar that measures the confidence that the input sequence comes from the true distribution given the condition c.

[0031] S3. Set random noise, concatenate the random noise with the conditional vector to obtain the concatenated vector, input the concatenated vector into the generator to obtain the generated data, input the generated data and the conditional vector into the discriminator to obtain the generated data score and the conditional vector score. Add random noise, concatenate the random noise with the conditional vector to obtain the concatenated vector, which is represented as follows:

[0032] in, To concatenate vectors, It is random noise. This is the condition vector.

[0033] Input the concatenated vector into the generator to obtain the data matrix; Sample each position in the data matrix to obtain the sampling result; The Gumbel-Softmax distribution of the sampling results is calculated to obtain the generated data. The generated data and the condition vector are then input into the discriminator to obtain the generated data score and the condition vector score.

[0034] In this data matrix, each row represents the unnormalized score of all amino acids at the j-th position of the sequence.

[0035] Sample Gumbel noise at each location:

[0036]

[0037] in, Let be a random variable uniformly distributed in the interval [0,1], used in the Gumbel-Softmax reparameterization technique.

[0038] Calculate the Gumbel-Softmax distribution of the sampling results to obtain the generated data, represented as:

[0039] in, For temperature parameters, For the distribution results, It is noise.

[0040] S4. Calculate the generator loss based on the generated data, calculate the discriminator loss based on the generated data score and the conditional vector score, obtain the optimal generator parameters and optimal discriminator parameters when the total discriminator loss and generator loss are minimized, update the generator and discriminator using the optimal generator parameters and optimal discriminator parameters to obtain the optimal generator and optimal discriminator, and the optimal generator and optimal discriminator form a generative adversarial model. The generator loss is calculated based on the generated data, including: Calculate the discriminator reward and generator reward for the generated data, and sum the discriminator reward and generator reward to obtain the expected reward; Calculate the average reward corresponding to the generated dataset; The generator loss is calculated based on the average reward and the expected reward.

[0041] Specifically, for each protein sequence, the input to the discriminator yields the following reward:

[0042] in, This represents the true probability.

[0043] The generator reward is:

[0044] Expected reward: + .

[0045] in, The physical energy function E(S) is a function that evaluates the physical plausibility (stability) of a protein sequence, mapping the sequence to an energy value (the lower the energy, the better the sequence). In generative models, we guide the generator to produce sequences that both fool the discriminator and conform to physical laws through physical rewards (negatively correlated with energy). The physical energy function can be a molecular mechanical potential function, a statistical potential function, etc.

[0046] Calculate the average expected reward for all protein sequences in the training dataset. For each sequence, calculate its advantage:

[0047] in, As an advantage, The sum of the expected rewards for all protein sequences. b This represents the average of the expected reward.

[0048] The generator's goal is to maximize the expected reward. We update it via the policy gradient, and the loss function of the generator G (the loss to be minimized) can be defined as the negative expected advantage:

[0049] in, Generate a sequence for generator G given (z, c). The probability is given by z, where z is noise and c is the condition vector.

[0050] The loss function of the discriminator is typically the binary cross-entropy:

[0051] in, The amino acid sequences were sampled from a real protein dataset (training dataset). To generate data, To generate data scores, Score the conditional vector.

[0052] S5. Set the target protein sequence set, input the target protein sequence set into the generative adversarial model to obtain the target fluorescence intensity value set. The target protein sequence set and the corresponding target fluorescence intensity value set form a supplementary dataset. Use the supplementary dataset and the training dataset as the prediction dataset. Specifically, the target protein sequence set consists of protein sequences with known amino acid sequences obtained from a database, but whose fluorescence intensity is unknown.

[0053] S6. Input the predicted dataset into the protein language model to obtain the hidden layer features of the last layer; S7. Extract CLS features and average pooling features from the hidden layer features, and concatenate the CLS features and average pooling features to obtain semantic features; CLS features and average pooling features are extracted from the hidden layer features. The CLS features and average pooling features are concatenated to obtain semantic features, including: Extract CLS features and average pooling features from the hidden layer features; The semantic features are obtained by concatenating the CLS features and the average pooling features, as follows:

[0054] in, For semantic features, Features of CLS For average pooling characteristics, For splicing.

[0055] S8. Perform nonlinear projection on the semantic features to obtain nonlinear projection features, perform local pattern enhancement on the hidden layer features to obtain specific features, and fuse the nonlinear projection features and specific features to obtain the final features. The semantic features are nonlinearly projected to obtain nonlinear projection features. Local pattern enhancement is applied to the hidden layer features to obtain specific features. The nonlinear projection features and specific features are then fused to obtain the final features, including: The semantic features are then subjected to a nonlinear projection to obtain the nonlinear projection features, which are represented as follows:

[0056] in, It is a nonlinear projection feature. For activation function, , , and These are learnable parameters. Semantic features; The hidden layer features are processed through a lightweight convolutional network for local pattern enhancement to obtain specific features, represented as follows:

[0057]

[0058] in, This is a one-dimensional convolution output. This is a one-dimensional convolution operation. Features of the hidden layer The kernel size is [size]. This represents the number of channels in the output of the convolutional layer. This is a global max pooling operation. It is a specific characteristic; The final feature is obtained by fusing nonlinear projection features and specific features.

[0059] S9. Input the final features into the prediction network to obtain the prediction results. Based on the prediction results and fluorescence intensity values, obtain the loss value. Obtain the optimal prediction network parameters when the loss value is minimized. Update the prediction network using the optimal prediction network parameters to obtain the optimal prediction network. The final features are input into the prediction network to obtain the prediction results. Based on the prediction results and fluorescence intensity values, the loss value is obtained. The optimal prediction network parameters, which minimize the loss value, are then used to update the prediction network, resulting in the optimal prediction network, including: The final features are input into the prediction network to obtain the prediction results; Based on the prediction results and fluorescence intensity values, the main task loss is obtained; Specifically, the main task loss is represented as follows:

[0060] in, The loss is the primary task loss, and N is the number of samples in the current training batch. The Hube loss function, For the first The true fluorescence intensity value of each sample for The prediction result, i.e., the predicted result of the first prediction. Predicted fluorescence intensity values ​​for each sample.

[0061] The Hube loss function is defined as:

[0062] in, This is the threshold hyperparameter.

[0063] The protein category is determined based on the final characteristics; Specifically, fluorescence intensity is not randomly distributed; it is strongly correlated with the family to which the protein belongs. Different families of fluorescent proteins have different chromophore chemical structures, as well as different structures and folding patterns. Once the model can accurately determine the family, it means that it has identified a series of decisive features associated with that family. These features are also the core basis for predicting fluorescence intensity, thus improving the accuracy of the model's predictions. The protein category is preset in advance.

[0064] The category loss is obtained based on the protein category and the prediction results; Specifically, the category loss is represented as:

[0065] in, For category loss, The total number of protein categories, where N is the number of samples in the current training batch. For the first The weight coefficients of each category For the first The sample at the th Real labels in each category For the first The sample at the th Predicted probabilities for each category.

[0066] The main task loss and category loss are weighted and summed to obtain the loss value; Specifically, the loss value is expressed as:

[0067] in, These are the weighting coefficients.

[0068] Obtain the optimal prediction network parameters that minimize the loss value, and update the prediction network using the optimal prediction network parameters to obtain the optimal prediction network.

[0069] Prediction network training can also be done in the following ways: Based on the requirements of the fluorescence intensity prediction task, a prediction network structure was designed, as shown in the figure below. Figure 2 As shown, Figure 2 In this model, which consists of four fully connected (Linear) layers and three ReLU activation functions, it is a feedforward neural network. The data passes through the input layer sequentially through the first fully connected layer → activation function → the second fully connected layer → activation function → the third fully connected layer → activation function → the fourth fully connected layer, and finally outputs the prediction result. The first fully connected layer has an input feature dimension of 4760 and an output feature dimension of 1024, with a bias term enabled. The second fully connected layer has an input feature dimension of 1024 and an output feature dimension of 512, with a bias term enabled. The third fully connected layer has an input feature dimension of 512 and an output feature dimension of 128, with a bias term enabled. The fourth fully connected layer has an input feature dimension of 128 and an output feature dimension of 1, with a bias term enabled.

[0070] At the same time, initialize the weights of the prediction network. and bias Parameters are typically initialized using a random initialization method. For the prediction network... Layer, its weight matrix The dimension is ,in and They are the first Layer and first Number of neurons in a layer, bias vector The dimension is .

[0071] As the amount of data increases, the performance of neural networks typically improves. ReLU activation function is an effective choice to alleviate the vanishing gradient problem, as detailed below:

[0072] The optimization algorithm selected is Adam, which can adaptively adjust the learning rate. The parameter update formula for the Adam algorithm is:

[0073]

[0074]

[0075]

[0076]

[0077] in, and These are the first-order moment estimates and second-order moment estimates of the corrected gradient, respectively. and It is the attenuation rate parameter. It is the learning rate ， It is a smooth term. It is a loss function. It is the loss function with respect to the parameters The gradient. Furthermore, the optimal architecture and parameters typically need to be determined through experimentation and tuning. For the first t First-order moment estimation at the next iteration For the first t-1 First-order moment estimation at the next iteration The decay rate of the first moment, The decay rate of the second moment. For the first t Second-order moment estimation at the next iteration.

[0078] During model training, the output of each layer of the neural network is first calculated. Layer output ,have:

[0079] in, It is an activation function. It is the first The layer's output. Then, using mean squared error (MSE) as the loss function, the error between the model's predictions and the true labels is calculated:

[0080] in, For real labels, For predicting labels, n This represents the number of samples.

[0081] During training, it is also necessary to check whether the current loss meets the preset requirements or threshold. If it does, the model prediction phase begins; otherwise, the following steps are performed: calculating the error of each layer to prepare for backpropagation.

[0082] in, It is the output layer. For output layer error, It is the gradient of the loss function with respect to the output of the output layer. This indicates element-wise multiplication. This is the derivative of the activation function with respect to the input of the output layer. The error is backpropagated from the output layer to the hidden layers, updating the error of each layer. For the ... layer:

[0083] Calculate the gradient of the loss function with respect to the model parameters:

[0084] in, loss function J Regarding the first l Layer weights The gradient.

[0085]

[0086] in, For the loss function J with respect to the th l Layer bias The gradient.

[0087] Finally, the Adam optimization algorithm is used to update the model's weights and bias parameters based on the calculated gradients, and then the trained model is used for prediction.

[0088] S10. Input the actual protein sequence into the optimal prediction network to obtain the fluorescence intensity corresponding to the protein sequence.

[0089] A protein sequence fluorescence intensity prediction system, comprising: The acquisition module is used to acquire a labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value. The fusion module is used to input the protein sequence training dataset into the protein language model to obtain a semantic representation vector, and then fuse the semantic representation vector with the fluorescence intensity value to obtain a conditional vector. The setting module is used to set random noise, concatenate the random noise with the conditional vector to obtain the concatenated vector, input the concatenated vector into the generator to obtain the generated data, and input the generated data and the conditional vector into the discriminator to obtain the generated data score and the conditional vector score. The generative adversarial model training module is used to calculate the generator loss based on the generated data, calculate the discriminator loss based on the generated data score and the conditional vector score, obtain the optimal generator parameters and optimal discriminator parameters when the discriminator loss and generator loss are minimized, update the generator and discriminator using the optimal generator parameters and optimal discriminator parameters, and obtain the optimal generator and optimal discriminator. The optimal generator and optimal discriminator form the generative adversarial model. The supplementary module is used to set the target protein sequence set. The target protein sequence set is input into the generative adversarial model to obtain the target fluorescence intensity value set. The target protein sequence set and the corresponding target fluorescence intensity value set form the supplementary dataset. The supplementary dataset and the training dataset are used as the prediction dataset. The hidden layer feature extraction module is used to input the prediction dataset into the protein language model to obtain the hidden layer features of the last layer. The concatenation module is used to extract CLS features and average pooling features from the hidden layer features, and concatenate the CLS features and average pooling features to obtain semantic features; The enhancement module is used to perform nonlinear projection on semantic features to obtain nonlinear projection features, perform local pattern enhancement on hidden layer features to obtain specific features, and fuse nonlinear projection features and specific features to obtain the final features. The prediction model training module is used to input the final features into the prediction network to obtain the prediction results. Based on the prediction results and fluorescence intensity values, the loss value is obtained. The optimal prediction network parameters with the minimum loss value are obtained. The prediction network is updated using the optimal prediction network parameters to obtain the optimal prediction network. The prediction module is used to input the actual protein sequence into the optimal prediction network to obtain the fluorescence intensity corresponding to the protein sequence.

[0090] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a protein sequence fluorescence intensity prediction method.

[0091] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0092] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0093] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0094] In this terminal device, a protein sequence fluorescence intensity prediction method from the above embodiments is stored in the terminal device's memory and loaded and executed on the terminal device's processor for convenient use.

[0095] This application also discloses a computer-readable storage medium, which stores a computer program, wherein when the computer program is executed by a processor, it employs a protein sequence fluorescence intensity prediction method as described in the above embodiments.

[0096] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0097] The protein sequence fluorescence intensity prediction method described in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0098] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0099] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for predicting the fluorescence intensity of a protein sequence, characterized in that, include: Obtain a labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value; The protein sequence training dataset is input into the protein language model to obtain a semantic representation vector. The semantic representation vector is then fused with the fluorescence intensity value to obtain a conditional vector. Set random noise, concatenate the random noise with the conditional vector to obtain a concatenated vector, input the concatenated vector into the generator to obtain generated data, and input the generated data and the conditional vector into the discriminator to obtain the generated data score and the conditional vector score. The generator loss is calculated based on the generated data, and the discriminator loss is calculated based on the generated data score and the conditional vector score. The optimal generator parameters and optimal discriminator parameters are obtained when the discriminator loss and generator loss are minimized. The generator and discriminator are updated using the optimal generator parameters and optimal discriminator parameters to obtain the optimal generator and optimal discriminator. The optimal generator and optimal discriminator form a generative adversarial model. Set a target protein sequence set, input the target protein sequence set into the generative adversarial model to obtain a target fluorescence intensity value set, the target protein sequence set and the corresponding target fluorescence intensity value set form a supplementary dataset, and use the supplementary dataset and the training dataset as a prediction dataset; The predicted dataset is input into the protein language model to obtain the hidden layer features of the last layer; CLS features and average pooling features are extracted from the hidden layer features, and the CLS features and average pooling features are concatenated to obtain semantic features; The semantic features are nonlinearly projected to obtain nonlinear projection features. The hidden layer features are enhanced with local patterns to obtain specific features. The nonlinear projection features and specific features are fused to obtain the final features. The final features are input into the prediction network to obtain the prediction result. Based on the prediction result and the fluorescence intensity value, the loss value is obtained. The optimal prediction network parameters with the minimum loss value are obtained. The prediction network is updated using the optimal prediction network parameters to obtain the optimal prediction network. The actual protein sequence is input into the optimal prediction network to obtain the fluorescence intensity corresponding to the actual protein sequence.

2. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The generator loss is calculated based on the generated data, including: Calculate the discriminator reward and generator reward of the generated data, and sum the discriminator reward and generator reward to obtain the expected reward; Calculate the average reward corresponding to the generated data; The generator loss is calculated based on the average reward and the expected reward.

3. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The semantic representation vector is fused with the fluorescence intensity value to obtain a conditional vector, including: The semantic representation vector is fused with the fluorescence intensity value to obtain a conditional vector, which is represented as follows: in, For conditional vectors, For semantic representation vectors, It is a learnable embedded projection layer. Indicates splicing, For the set of real numbers, To be The dimension after converting to a vector is the dimension of the semantic representation vector.

4. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The process involves setting random noise, concatenating the random noise with the conditional vector to obtain a concatenated vector, inputting the concatenated vector into a generator to obtain generated data, and inputting the generated data and the conditional vector into a discriminator to obtain the generated data score and the conditional vector score, including: Set random noise, and concatenate the random noise with the conditional vector to obtain the concatenated vector, which is represented as: in, To concatenate vectors, It is random noise. It is a conditional vector; The concatenated vector is input into the generator to obtain a data matrix; Sample each position in the data matrix to obtain the sampling result; Calculate the Gumbel-Softmax distribution of the sampling results to obtain generated data. Input the generated data and the condition vector into the discriminator to obtain the generated data score and the condition vector score.

5. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The process of performing nonlinear projection on the semantic features to obtain nonlinear projection features, performing local pattern enhancement on the hidden layer features to obtain specific features, and fusing the nonlinear projection features and specific features to obtain the final features includes: The semantic features are then subjected to nonlinear projection to obtain nonlinear projection features, which are represented as follows: in, It is a nonlinear projection feature. For activation function, , , and These are learnable parameters. Semantic features; The hidden layer features are processed through a lightweight convolutional network for local pattern enhancement to obtain specific features, represented as follows: in, This is a one-dimensional convolution output. This is a one-dimensional convolution operation. Features of the hidden layer The kernel size is [size]. This represents the number of channels in the output of the convolutional layer. This is a global max pooling operation. It is a specific characteristic; The nonlinear projection features and specific features are fused to obtain the final features.

6. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The step of extracting CLS features and average pooling features from the hidden layer features, and concatenating the CLS features and average pooling features to obtain semantic features includes: Extract CLS features and average pooling features from the hidden layer features; The CLS features and average pooling features are concatenated to obtain the semantic features, represented as follows: in, For semantic features, Features of CLS For average pooling characteristics, For splicing.

7. The protein sequence fluorescence intensity prediction method according to claim 1, characterized in that, The process of inputting the final features into the prediction network to obtain a prediction result, obtaining a loss value based on the prediction result and the fluorescence intensity value, acquiring the optimal prediction network parameters that minimize the loss value, and updating the prediction network using the optimal prediction network parameters to obtain the optimal prediction network includes: The final features are input into the prediction network to obtain the prediction result; Based on the prediction results and fluorescence intensity values, the main task loss is obtained; The protein category is obtained based on the final characteristics; The category loss is obtained based on the protein category and the prediction result; The main task loss and category loss are weighted and summed to obtain the loss value; Obtain the optimal prediction network parameters that minimize the loss value, and update the prediction network using the optimal prediction network parameters to obtain the optimal prediction network.

8. A protein sequence fluorescence intensity prediction system, characterized in that, include: The acquisition module is used to acquire a labeled protein sequence training dataset, which includes the amino acid sequence of the protein and the corresponding fluorescence intensity value. The fusion module is used to input the protein sequence training dataset into the protein language model to obtain a semantic representation vector, and to fuse the semantic representation vector with the fluorescence intensity value to obtain a conditional vector. The setting module is used to set random noise, concatenate the random noise with the condition vector to obtain a concatenated vector, input the concatenated vector into the generator to obtain generated data, and input the generated data and the condition vector into the discriminator to obtain the generated data score and the condition vector score. The generative adversarial model training module is used to calculate the generator loss based on the generated data, calculate the discriminator loss based on the generated data score and the conditional vector score, obtain the optimal generator parameters and optimal discriminator parameters when the discriminator loss and generator loss are minimized, update the generator and discriminator using the optimal generator parameters and optimal discriminator parameters, and obtain the optimal generator and optimal discriminator, which together form the generative adversarial model; The supplementary module is used to set a target protein sequence set, input the target protein sequence set into the generative adversarial model to obtain a target fluorescence intensity value set, the target protein sequence set and the corresponding target fluorescence intensity value set constitute a supplementary dataset, and the supplementary dataset and the training dataset are used as a prediction dataset; The hidden layer feature extraction module is used to input the prediction dataset into the protein language model to obtain the hidden layer features of the last layer. The concatenation module is used to extract CLS features and average pooling features from the hidden layer features, and concatenate the CLS features and average pooling features to obtain semantic features; An enhancement module is used to perform nonlinear projection on the semantic features to obtain nonlinear projection features, perform local pattern enhancement on the hidden layer features to obtain specific features, and fuse the nonlinear projection features and specific features to obtain the final features; The prediction model training module is used to input the final features into the prediction network to obtain the prediction result, obtain the loss value based on the prediction result and the fluorescence intensity value, obtain the optimal prediction network parameters when the loss value is minimized, update the prediction network using the optimal prediction network parameters, and obtain the optimal prediction network. The prediction module is used to input the actual protein sequence into the optimal prediction network to obtain the fluorescence intensity corresponding to the protein sequence.

9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.

Images