Protein solubility prediction method and device based on deep learning, computer equipment and storage medium

Abstract

The invention relates to a protein solubility prediction method and device based on deep learning, computer equipment and a storage medium, and the method comprises the steps: obtaining a to-be-predicted sample which is a protein amino acid sequence; according to an ESM-2 model and the to-be-predicted sample, a high-dimensional feature vector is obtained, and each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of a protein sequence in a semantic space of the ESM-2 model; obtaining a protein solubility prediction probability according to the high-dimensional feature vector and a depth residual multi-layer perceptron model; and obtaining a protein solubility prediction result according to the protein solubility prediction probability. The accuracy of protein solubility prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedicine, and in particular to a protein solubility prediction method and device based on deep learning, a computer device and a storage medium. BACKGROUND

[0002] Protein solubility is a key prerequisite for efficient heterologous expression, and directly affects the yield, purification efficiency and functional integrity of target proteins in structural biology, enzyme engineering and biopharmaceutical industries.

[0003] Escherichia coli is still the preferred expression host for recombinant proteins worldwide due to its clear genetic background, low culture cost and high transformation efficiency. However, when foreign proteins, especially those of eukaryotic origin, containing multiple domains or rich in disulfide bonds, enter the cytoplasm of Escherichia coli, they are prone to form inactive inclusion bodies (IBs) due to the inability to keep up with the folding rate, the lack of disulfide bond formation in the cytoplasmic reducing environment, and the lack of chaperones / folding enzymes. This has become a technical bottleneck in drug development, structural biology and enzyme engineering. Given that the sequence of amino acids that make up a protein fundamentally determines the folding path and aggregation tendency of the protein, predicting its solubility from the sequence has always been the focus of expression optimization.

[0004] With the explosive growth in the number of sequences measured, existing methods for predicting protein solubility are still insufficient in modeling complex sequence patterns, making it difficult to meet demand. SUMMARY

[0005] To solve the above technical problems or at least partially solve the above technical problems, the present application provides a protein solubility prediction method and device based on deep learning, a computer device and a storage medium.

[0006] In a first aspect, the present application provides a protein solubility prediction method based on deep learning, comprising: obtaining a sample to be predicted, the sample to be predicted being a protein amino acid sequence; obtaining a high-dimensional feature vector according to an ESM-2 model and the sample to be predicted, each feature dimension of the high-dimensional feature vector corresponding to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model; obtaining a protein solubility prediction probability according to the high-dimensional feature vector and a deep residual multi-layer perceptron model; obtaining a protein solubility prediction result according to the protein solubility prediction probability.

[0007] Optionally, obtaining the high-dimensional feature vector based on the ESM-2 model and the sample to be predicted includes: The sample to be predicted is segmented and encoded into words, and then converted into a processable sequence; The processable sequence is sequentially passed through the multi-layer encoder of the ESM-2 model for feature mapping; The output of the residue-level feature output layer in the ESM-2 model is used as the high-dimensional feature vector.

[0008] Optionally, the deep residual multilayer perceptron model includes an input layer, a learning layer, and an output layer; The step of obtaining the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model includes: The high-dimensional feature vector is linearly normalized according to the feature dimension to obtain a standardized feature vector; The standardized feature vector is input into the input layer, and the standardized feature vector is mapped to the preset hidden feature space through the linear mapping method of the residual mapping structure to obtain the initial hidden features; The standardized feature vector is fused with the initial hidden features to obtain the input vector; The input vector is fed into the learning layer and passes through multiple concatenated residual blocks in the learning layer to obtain fused features. The fused features are passed through the output layer to obtain a prediction score; After the predicted score is probabilistically processed, it is converted into the predicted probability of each soluble category to which the protein belongs. The sum of the predicted probabilities corresponding to each soluble category is a preset constant.

[0009] Optionally, the input vector is fed into a learning layer and sequentially passes through multiple concatenated residual blocks of the learning layer to obtain fused features, including: Each residual block superimposes the features input to the residual block with the features obtained after transformation by the residual block. The last residual block of the learning layer superimposes the features input to the last residual block with the features obtained after transformation by the last residual block to obtain the fused features; Within each residual block, the features input to that residual block are transformed in the following manner: Layer normalization is performed on the features of the input residual block. The features after layer normalization are subjected to nonlinear mapping. The features after nonlinear mapping are activated by an activation function. The features activated by the activation function are randomly deactivated. The activation function is an improved focus loss function, which is as follows: To predict the probability of the true category, For category weights, It is a key regulatory factor.

[0010] Optionally, obtaining the protein solubility prediction result based on the protein solubility prediction probability includes: When the predicted probability of the soluble category to which a protein belongs is greater than or equal to the determination threshold, the protein is determined to be a soluble protein. When the predicted probability of a protein belonging to a soluble category is less than the determination threshold, the protein is determined to be an insoluble protein.

[0011] Optionally, before obtaining the sample to be predicted, the method further includes: Obtain the training set; The first model is trained using the training set to obtain the ESM-2 model; The second model is trained using the training set to obtain the deep residual multilayer perceptron model. The acquisition of the training set includes: Obtain the original sample set; From the original sample set, randomly select any two original samples, generate a training sample by linear interpolation, and use all generated training samples as the training set; The step involves randomly selecting any two original samples from the original sample set and generating a training sample using linear interpolation, as follows: in, and Represents any original sample, The mixing coefficients are those that follow a Beta distribution. Used to control the proportion of linear interpolation.

[0012] Optionally, the method further includes: When training the second model, after the number of training rounds is greater than or equal to the minimum number of rounds threshold, the current model is evaluated using the validation set after each training round, and the validation performance index is obtained. If the current validation performance metric falls below the historical best performance value in multiple consecutive training rounds, the early stopping condition is triggered and training of the model is terminated. The model corresponding to the historical best verification performance index is retained as the deep residual multilayer perceptron model.

[0013] Secondly, a deep learning-based protein solubility prediction device is provided, the device comprising: A sample unit is used to acquire a sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence. The ESM-2 model unit is used to obtain a high-dimensional feature vector based on the ESM-2 model and the sample to be predicted. Each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model. The deep residual multilayer perceptron model unit is used to obtain the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model. The prediction unit is used to obtain the protein solubility prediction result based on the protein solubility prediction probability.

[0014] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any of the preceding claims.

[0015] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0016] This invention provides a method, apparatus, computer device, and storage medium for predicting protein solubility based on deep learning. The method includes: acquiring a sample to be predicted, wherein the sample is a protein amino acid sequence; acquiring a high-dimensional feature vector based on an ESM-2 model and the sample to be predicted, wherein each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model; acquiring a protein solubility prediction probability based on the high-dimensional feature vector and a deep residual multilayer perceptron model; and obtaining a protein solubility prediction result based on the protein solubility prediction probability. The method of this invention, through the ESM-2 model and the deep residual multilayer perceptron model, can optimize imbalanced and difficult-to-classify samples, greatly improving the model's performance. It can fully explore the deep semantic features of the protein sequence, thereby achieving efficient and accurate solubility prediction and improving prediction accuracy. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The diagram shows the application environment of the deep learning-based protein solubility prediction method according to an embodiment of the present invention. Figure 2 The diagram shown is a flowchart of a deep learning-based protein solubility prediction method according to an embodiment of the present invention. Figure 3 The diagram shown is a structural block diagram of a deep learning-based protein solubility prediction device according to an embodiment of the present invention. Figure 4 The diagram shown is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Figure 1 This is a diagram illustrating the application environment of deep learning-based protein solubility prediction in one embodiment. (Refer to...) Figure 1 This deep learning-based protein solubility prediction method is applied to a deep learning-based protein solubility prediction system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers.

[0022] like Figure 2 As shown, in one embodiment, a deep learning-based method for predicting protein solubility is provided. This embodiment mainly applies this method to the aforementioned... Figure 1 Let's take terminal 110 or server 120 as an example for illustration. (Refer to...) Figure 2 The deep learning-based protein solubility prediction method specifically includes the following steps: Step 210: Obtain the sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence; Step 220: Based on the ESM-2 model and the sample to be predicted, obtain a high-dimensional feature vector. Each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model. Step 230: Obtain the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model; Step 240: Obtain the protein solubility prediction result based on the protein solubility prediction probability.

[0023] The method of this invention, through the ESM-2 model (protein structure prediction artificial intelligence model-2) and the deep residual multilayer perceptron model, can optimize imbalanced and difficult-to-determine samples, greatly improving the performance of the model. It can fully explore the deep semantic features of protein sequences, thereby achieving efficient and accurate solubility prediction and improving the accuracy of prediction.

[0024] In this invention, step 220, obtaining the high-dimensional feature vector based on the ESM-2 model and the sample to be predicted, includes: The sample to be predicted is segmented and encoded into words, and then converted into a processable sequence; The processable sequence is sequentially passed through the multi-layer encoder of the ESM-2 model for feature mapping; The output of the residue-level feature output layer in the ESM-2 model is used as the high-dimensional feature vector.

[0025] In this embodiment of the invention, the ESM-2 model is a deep neural network model based on the Transformer architecture. It can learn the contextual semantic information of each amino acid residue in a protein sequence by performing unsupervised pre-training on large-scale protein sequence data.

[0026] In this embodiment of the invention, the 36th layer in the ESM-2 model can be selected as the residue-level feature output layer. This layer can comprehensively characterize the high-order structure, overall sequence features and functional information of the protein sequence, which is beneficial to improving the accuracy of subsequent predictions.

[0027] In this embodiment of the invention, the deep residual multilayer perceptron model includes an input layer, a learning layer, and an output layer; In step 230, obtaining the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model includes: The high-dimensional feature vector is linearly normalized according to the feature dimension to obtain a standardized feature vector; The standardized feature vector is input into the input layer, and the standardized feature vector is mapped to the preset hidden feature space through the linear mapping method of the residual mapping structure to obtain the initial hidden features; The standardized feature vector is fused with the initial hidden features to obtain the input vector; The input vector is fed into the learning layer and passes through multiple concatenated residual blocks in the learning layer to obtain fused features. The fused features are passed through the output layer to obtain a prediction score; After the predicted score is probabilistically processed, it is converted into the predicted probability of each soluble category to which the protein belongs. The sum of the predicted probabilities corresponding to each soluble category is a preset constant.

[0028] In this embodiment of the invention, the input vector is fed into a learning layer and sequentially passes through multiple concatenated residual blocks of the learning layer to obtain fused features, including: Each residual block superimposes the features input to the residual block with the features obtained after transformation by the residual block. The last residual block of the learning layer superimposes the features input to the last residual block with the features obtained after transformation by the last residual block to obtain the fused features; Within each residual block, the features input to that residual block are transformed in the following manner: Layer normalization is performed on the features of the input residual block. The features after layer normalization are subjected to nonlinear mapping. The features after nonlinear mapping are activated by an activation function. The features activated by the activation function are randomly deactivated. The activation function is an improved focus loss function, which is as follows: To predict the probability of the true category, For category weights, It is a key regulatory factor.

[0029] In this embodiment of the invention, standardization preprocessing can eliminate the impact of numerical scale differences between different feature dimensions on model training and prediction results.

[0030] The loss function described above automatically adjusts the class weights and focus control factors, enabling the model to pay more attention to samples that are difficult to classify, thereby improving the model's classification ability and robustness.

[0031] In this embodiment of the invention, the learning layer is configured with multiple cascaded residual blocks, which can effectively alleviate the gradient vanishing problem that may occur during network deepening, enabling the network to fully learn the nonlinear mapping relationship between protein sequence features and protein solubility. Through the step-by-step processing of multiple residual blocks, deep nonlinear modeling of protein sequence features can be performed while maintaining stable gradient propagation, thereby fully exploring the complex correlation between protein sequence features and protein solubility.

[0032] The deep residual multilayer perceptron prediction model of this invention introduces a residual connection structure of the input layer during the multilayer nonlinear mapping process of the learning layer, thereby achieving the layer-by-layer fusion of input features and intermediate features, thus improving the stability and feature representation ability of the model while ensuring the network depth.

[0033] In this embodiment, step 240, obtaining the protein solubility prediction result based on the protein solubility prediction probability, includes: When the predicted probability of the soluble category to which a protein belongs is greater than or equal to the determination threshold, the protein is determined to be a soluble protein. When the predicted probability of a protein belonging to a soluble category is less than the determination threshold, the protein is determined to be an insoluble protein.

[0034] In this embodiment of the invention, before obtaining the sample to be predicted, the method further includes: Obtain the training set; The first model is trained using the training set to obtain the ESM-2 model; The second model is trained using the training set to obtain the deep residual multilayer perceptron model. The acquisition of the training set includes: Obtain the original sample set; From the original sample set, randomly select any two original samples, generate a training sample by linear interpolation, and use all generated training samples as the training set; The step involves randomly selecting any two original samples from the original sample set and generating a training sample using linear interpolation, as follows: in, and Represents any original sample, The mixing coefficients are those that follow a Beta distribution. Used to control the proportion of linear interpolation.

[0035] To further improve the model's generalization ability and alleviate overfitting, this embodiment introduces a label smoothing strategy to process the true labels. The calculation method for label smoothing is as follows: in, For label smoothing factor, For the number of categories, The labels are the true labels. The introduction of smooth labeling effectively reduces model overconfidence, avoids overfitting, and thus enhances the model's adaptability to unseen data.

[0036] In this embodiment of the invention, the loss function of the model is constrained by the labels of the original samples, thereby forming a continuous discriminative transition region in the feature space.

[0037] The training set obtained through the above method can effectively expand the distribution range of training data, smooth the decision boundaries between classes, and implicitly constrain the model's classification hyperplane without introducing additional real samples. This method can significantly improve the model's robustness to input feature perturbations and reduce the risk of overfitting, especially when the number of samples is limited or the class distribution is imbalanced, further enhancing the model's generalization performance.

[0038] In this embodiment of the invention, the method further includes: When training the first model, after the number of training rounds is greater than or equal to the number of rounds threshold, the current model is evaluated using the validation set after each training round, and the validation performance index is obtained. If the current validation performance metric falls below the historical best performance value in multiple consecutive training rounds, the early stopping condition is triggered and training of the model is terminated. The model corresponding to the historical best verification performance index is retained as the ESM-2 model.

[0039] In this embodiment of the invention, the method further includes: When training the second model, after the number of training rounds is greater than or equal to the minimum number of rounds threshold, the current model is evaluated using the validation set after each training round, and the validation performance index is obtained. If the current validation performance metric falls below the historical best performance value in multiple consecutive training rounds, the early stopping condition is triggered and training of the model is terminated. The model corresponding to the historical best verification performance index is retained as the deep residual multilayer perceptron model.

[0040] In this embodiment of the invention, a minimum training rounds threshold is preset in the early stage of model training. Before the number of training rounds reaches the threshold, early stopping is not performed to ensure that the model can fully complete the initial feature learning process and avoid underfitting due to insufficient training.

[0041] In this embodiment of the invention, once the number of training rounds reaches the minimum training round threshold, dynamic performance monitoring of the model training process begins based on the validation set. After each training round, the current model is evaluated using the validation set, and the corresponding validation performance metric is calculated. The validation performance metric is an evaluation indicator that comprehensively reflects the classification effect and can be obtained by combining multiple single performance indicators.

[0042] Optionally, the validation performance metric is obtained by combining at least two single evaluation metrics that can reflect classification performance from different perspectives. One type of evaluation metric measures the model's ability to distinguish between positive and negative samples, while the other type measures the model's classification stability under imbalanced sample distribution conditions. By combining multiple evaluation metrics, the limitations of a single evaluation metric under specific data distribution conditions can be effectively avoided, thereby improving the reliability of early stopping determination.

[0043] If the overall performance of the current model on the validation set is lower than the historical best performance value in several consecutive training rounds, it is determined that the model training process has converged, and the early stopping condition is automatically triggered to terminate the subsequent training process.

[0044] Optionally, a duration can be set. If the overall performance of the current model on the validation set is lower than the historical best performance value within several consecutive training epochs and the duration exceeds a preset patience window, the model training process is determined to have converged, and the early stopping condition is automatically triggered to terminate the subsequent training process.

[0045] After the early stopping condition is triggered, the model that achieved the best historical performance on the validation set is retained as the final model, while models that do not achieve better performance in subsequent training are discarded. In this way, the early stopping mechanism can effectively reduce unnecessary training rounds while ensuring model prediction performance, improving model training efficiency and reducing the risk of overfitting.

[0046] In this embodiment of the invention, multiple optimization strategies are integrated and introduced in terms of model training and optimization strategies, including loss function design based on class weight and focus control, label smoothing mechanism, and progressive early stopping mechanism based on comprehensive performance index of validation set. Through multi-dimensional control of the training process, the robustness and generalization ability of the model under complex data distribution conditions are improved.

[0047] The method of this invention can achieve efficient and accurate prediction of protein solubility while ensuring model stability.

[0048] The aforementioned protein solubility prediction method utilizes unique technical features to achieve the beneficial effect of solving the technical problems raised in the background art.

[0049] In one embodiment, such as Figure 3 As shown, a deep learning-based protein solubility prediction device is provided, comprising: Sample unit 310 is used to acquire a sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence; ESM-2 model unit 320 is used to obtain a high-dimensional feature vector based on the ESM-2 model and the sample to be predicted. Each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model. The deep residual multilayer perceptron model unit 330 is used to obtain the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model. The prediction unit 340 is used to obtain a protein solubility prediction result based on the protein solubility prediction probability.

[0050] In this embodiment of the invention, the ESM-2 model is also used for: The sample to be predicted is segmented and encoded into words, and then converted into a processable sequence; The processable sequence is sequentially passed through the multi-layer encoder of the ESM-2 model for feature mapping; The output of the residue-level feature output layer in the ESM-2 model is used as the high-dimensional feature vector.

[0051] The deep residual multilayer perceptron model includes an input layer, a learning layer, and an output layer; The deep residual multilayer perceptron model unit 330 is also used for: The high-dimensional feature vector is linearly normalized according to the feature dimension to obtain a standardized feature vector; The standardized feature vector is input into the input layer, and the standardized feature vector is mapped to the preset hidden feature space through the linear mapping method of the residual mapping structure to obtain the initial hidden features; The standardized feature vector is fused with the initial hidden features to obtain the input vector; The input vector is fed into the learning layer and passes through multiple concatenated residual blocks in the learning layer to obtain fused features. The fused features are passed through the output layer to obtain a prediction score; After the predicted score is probabilistically processed, it is converted into the predicted probability of each soluble category to which the protein belongs. The sum of the predicted probabilities corresponding to each soluble category is a preset constant.

[0052] In this embodiment of the invention, the deep residual multilayer perceptron model unit 330 is further used for: Each residual block superimposes the features input to the residual block with the features obtained after transformation by the residual block. The last residual block of the learning layer superimposes the features input to the last residual block with the features obtained after transformation by the last residual block to obtain the fused features; Each residual block is used for: Layer normalization is performed on the features of the input residual block. The features after layer normalization are subjected to nonlinear mapping. The features after nonlinear mapping are activated by an activation function. The features activated by the activation function are randomly deactivated. The activation function is an improved focus loss function, which is as follows: To predict the probability of the true category, For category weights, It is a key regulatory factor.

[0053] In this embodiment of the invention, the prediction unit 340 is further configured to: When the predicted probability of the soluble category to which a protein belongs is greater than or equal to the determination threshold, the protein is determined to be a soluble protein. When the predicted probability of a protein belonging to a soluble category is less than the determination threshold, the protein is determined to be an insoluble protein.

[0054] In this embodiment of the invention, the device further includes a training unit, used for: Obtain the training set; The first model is trained using the training set to obtain the ESM-2 model; The second model is trained using the training set to obtain the deep residual multilayer perceptron model. In this embodiment of the invention, the training unit is further configured to: Obtain the original sample set; From the original sample set, randomly select any two original samples, generate a training sample by linear interpolation, and use all generated training samples as the training set; The step involves randomly selecting any two original samples from the original sample set and generating a training sample using linear interpolation, as follows: in, and Represents any original sample, The mixing coefficients are those that follow a Beta distribution. Used to control the proportion of linear interpolation.

[0055] In this embodiment of the invention, the device further includes an early stop unit, used for: When training the second model, after the number of training rounds is greater than or equal to the minimum number of rounds threshold, the current model is evaluated using the validation set after each training round, and the validation performance index is obtained. If the current validation performance metric falls below the historical best performance value in multiple consecutive training rounds, the early stopping condition is triggered and training of the model is terminated. The model corresponding to the historical best verification performance index is retained as the deep residual multilayer perceptron model.

[0056] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the following method: acquiring a sample to be predicted, the sample being a protein amino acid sequence; acquiring a high-dimensional feature vector based on an ESM-2 model and the sample to be predicted, each feature dimension of the high-dimensional feature vector corresponding to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model; acquiring a protein solubility prediction probability based on the high-dimensional feature vector and a deep residual multilayer perceptron model; and obtaining a protein solubility prediction result based on the protein solubility prediction probability.

[0057] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the following method: acquiring a sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence; acquiring a high-dimensional feature vector based on the ESM-2 model and the sample to be predicted, wherein each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model; acquiring a protein solubility prediction probability based on the high-dimensional feature vector and a deep residual multilayer perceptron model; and obtaining a protein solubility prediction result based on the protein solubility prediction probability.

[0058] The aforementioned method for predicting protein solubility achieves the beneficial effect of solving the technical problems mentioned in the background art.

[0059] Figure 2 This is a flowchart illustrating a protein solubility prediction method in one embodiment. It should be understood that, although... Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0060] Figure 4 An internal structural diagram of a computer device in one embodiment is shown. Specifically, this computer device may be... Figure 1 Terminal 110 or server 120 in the middle. For example... Figure 4 As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a protein solubility prediction method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the protein solubility prediction method. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0061] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0062] In one embodiment, the four devices provided by the present invention can be implemented as a computer program, and the computer program can be implemented as follows: Figure 4 It runs on the computer device shown.

[0063] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A deep learning-based method for predicting protein solubility, characterized in that, The method includes: Obtain the sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence; Based on the ESM-2 model and the sample to be predicted, a high-dimensional feature vector is obtained. Each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model. Based on the high-dimensional feature vector and the deep residual multilayer perceptron model, the protein solubility prediction probability is obtained; Based on the predicted protein solubility probability, the protein solubility prediction result is obtained.

2. The method according to claim 1, characterized in that, The step of obtaining a high-dimensional feature vector based on the ESM-2 model and the sample to be predicted includes: The sample to be predicted is segmented and encoded into words, and then converted into a processable sequence; The processable sequence is sequentially passed through the multi-layer encoder of the ESM-2 model for feature mapping; The output of the residue-level feature output layer in the ESM-2 model is used as the high-dimensional feature vector.

3. The method according to claim 1, characterized in that, The deep residual multilayer perceptron model includes an input layer, a learning layer, and an output layer; The step of obtaining the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model includes: The high-dimensional feature vector is linearly normalized according to the feature dimension to obtain a standardized feature vector; The standardized feature vector is input into the input layer, and the standardized feature vector is mapped to the preset hidden feature space through the linear mapping method of the residual mapping structure to obtain the initial hidden features; The standardized feature vector is fused with the initial hidden features to obtain the input vector; The input vector is fed into the learning layer and passes through multiple concatenated residual blocks in the learning layer to obtain fused features. The fused features are passed through the output layer to obtain a prediction score; After the predicted score is probabilistically processed, it is converted into the predicted probability of each soluble category to which the protein belongs. The sum of the predicted probabilities corresponding to each soluble category is a preset constant.

4. The method according to claim 3, characterized in that, The input vector is fed into the learning layer and passed through multiple concatenated residual blocks in the learning layer to obtain fused features, including: Each residual block superimposes the features input to the residual block with the features obtained after transformation by the residual block. The last residual block of the learning layer superimposes the features input to the last residual block with the features obtained after transformation by the last residual block to obtain the fused features; Within each residual block, the features input to that residual block are transformed in the following manner: Layer normalization is performed on the features of the input residual block. The features after layer normalization are subjected to nonlinear mapping. The features after nonlinear mapping are activated by an activation function. The features activated by the activation function are randomly deactivated. The activation function is an improved focus loss function, which is as follows: To predict the probability of the true category, For category weights, It is a key regulatory factor.

5. The method according to claim 1, characterized in that, The step of obtaining the protein solubility prediction result based on the protein solubility prediction probability includes: When the predicted probability of the soluble category to which a protein belongs is greater than or equal to the determination threshold, the protein is determined to be a soluble protein. When the predicted probability of a protein belonging to a soluble category is less than the determination threshold, the protein is determined to be an insoluble protein.

6. The method according to claim 1, characterized in that, Before obtaining the sample to be predicted, the method further includes: Obtain the training set; The first model is trained using the training set to obtain the ESM-2 model; The second model is trained using the training set to obtain the deep residual multilayer perceptron model. The acquisition of the training set includes: Obtain the original sample set; From the original sample set, randomly select any two original samples, generate a training sample by linear interpolation, and use all generated training samples as the training set; The step involves randomly selecting any two original samples from the original sample set and generating a training sample using linear interpolation, as follows: in, and Represents any original sample, The mixing coefficients are those that follow a Beta distribution. Used to control the proportion of linear interpolation.

7. The method according to claim 6, characterized in that, The method further includes: When training the second model, after the number of training rounds is greater than or equal to the minimum number of rounds threshold, the current model is evaluated using the validation set after each training round, and the validation performance index is obtained. If the current validation performance metric falls below the historical best performance value in multiple consecutive training rounds, the early stopping condition is triggered and training of the model is terminated. The model corresponding to the historical best verification performance index is retained as the deep residual multilayer perceptron model.

8. A deep learning-based protein solubility prediction device, characterized in that, The device includes: A sample unit is used to acquire a sample to be predicted, wherein the sample to be predicted is a protein amino acid sequence. The ESM-2 model unit is used to obtain a high-dimensional feature vector based on the ESM-2 model and the sample to be predicted. Each feature dimension of the high-dimensional feature vector corresponds to an abstract representation component of the protein sequence in the semantic space of the ESM-2 model. The deep residual multilayer perceptron model unit is used to obtain the protein solubility prediction probability based on the high-dimensional feature vector and the deep residual multilayer perceptron model. The prediction unit is used to obtain the protein solubility prediction result based on the protein solubility prediction probability.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images