Far homologous protein search method and system based on neural network denoising, terminal and storage medium

Abstract

The invention discloses a far homologous protein search method and system based on neural network denoising, a terminal and a storage medium, and the method comprises the steps: constructing a protein sequence spectrum according to a target data set, determining a preset convolutional neural network, and carrying out the model training of the preset convolutional neural network according to the protein sequence spectrum, and obtaining a superfamily model; obtaining a to-be-queried protein sequence, performing search processing in the target data set according to the to-be-queried protein sequence to obtain an initial candidate list, and constructing a homologous feature matrix according to the initial candidate list and the superfamily model; determining an initial sorting learning model, and performing model training on the initial sorting learning model according to the initial candidate list and the homologous feature matrix to obtain a target sorting learning model; and performing de-noising processing and iterative search processing on the initial candidate list through a target sorting learning model to obtain a target far homologous protein list. The method can effectively improve the accuracy and stability of far homologous protein output.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for searching distant homologous proteins based on neural network denoising. Background Technology

[0002] Distant homologous proteins are proteins with low sequence similarity (usually <25%), but belonging to the same superfamily (a group of genes or proteins that evolved from a common ancestor but have little similarity), and with similar structure or function.

[0003] Existing technologies using protein distant homology search methods have problems with the candidate list of distant homologous proteins being subject to severe noise interference and failing to filter out non-homologous sequences in the candidate list, resulting in low detection accuracy of distant homologous proteins.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for searching distant homologous proteins based on neural network denoising. This invention aims to solve the problem that the candidate lists of distant homologous proteins obtained in the prior art are subject to severe noise interference and cannot filter out non-homologous sequences in the candidate lists, resulting in low detection accuracy of distant homologous proteins.

[0006] To achieve the above objectives, the present invention provides a method for searching distant homologous proteins based on neural network denoising, the method comprising the following steps:

[0007] A target dataset is determined, a protein sequence profile is constructed based on the target dataset, a preset convolutional neural network is determined, and the preset convolutional neural network is trained based on the protein sequence profile to obtain a superfamily model;

[0008] Obtain the protein sequence to be queried, perform search processing on the target dataset based on the protein sequence to be queried to obtain an initial candidate list, and construct a homology feature matrix based on the initial candidate list and the superfamily model;

[0009] An initial ranking learning model is determined, and the initial ranking learning model is trained based on the initial candidate list and the homologous feature matrix to obtain the target ranking learning model;

[0010] The target ranking learning model is used to denoise and iteratively search the initial candidate list to obtain a list of target distant homologous proteins.

[0011] Optionally, in the neural network-based denoising method for searching distant homologous proteins, the protein sequence spectrum includes a hidden Markov model alignment matrix and a position-specific score matrix.

[0012] The process of determining the target dataset and constructing a protein sequence profile based on the target dataset specifically includes:

[0013] The target dataset is determined, and the HHblits-based method is used to perform the first iterative search process on the target dataset to obtain the Hidden Markov Model Alignment Matrix;

[0014] The PSI-BLAST basic method is used to perform a second iterative search process on the target dataset to obtain the location-specific score matrix.

[0015] Optionally, in the neural network-based denoising method for searching distant homologous proteins, the preset convolutional neural network includes a first convolutional neural network and a second convolutional neural network; the superfamily model includes a first superfamily model and a second superfamily model.

[0016] The step of determining a preset convolutional neural network and training the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model specifically includes:

[0017] Obtain the first super family corresponding to the Hidden Markov Model alignment matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the Hidden Markov Model alignment matrix and the first super family to obtain the first super family model.

[0018] Obtain the second superfamily corresponding to the position-specific score matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the position-specific score matrix and the second superfamily respectively to obtain the second superfamily model.

[0019] Optionally, the method for searching distant homologous proteins based on neural network denoising, wherein obtaining the protein sequence to be searched, performing search processing on the target dataset based on the protein sequence to be searched to obtain an initial candidate list, and constructing a homology feature matrix based on the initial candidate list and the superfamily model, specifically includes:

[0020] Obtain the protein sequence to be queried, and use the PSI-BLAST basic method to search the target dataset based on the protein sequence to obtain an initial candidate list;

[0021] Calculate the feature similarity between the protein sequence to be queried and the initial candidate list, and extract the similarity feature matrix from the protein sequence to be queried based on the feature similarity;

[0022] The protein sequence to be queried is input into the superfamily model to obtain the initial homology feature matrix;

[0023] A common origin feature matrix is ​​constructed based on the similarity feature matrix and the initial common origin feature matrix.

[0024] Optionally, the method for searching distant homologous proteins based on neural network denoising, wherein determining an initial ranking learning model and training the initial ranking learning model according to the initial candidate list and the homology feature matrix to obtain a target ranking learning model specifically includes:

[0025] Obtain the initial candidate proteins from the initial candidate list, and obtain the homology features and true labels of the initial candidate proteins and the query protein in the query protein sequence;

[0026] An initial ranking learning model is determined, and the initial ranking learning model is trained using five-fold cross-validation based on the homologous features and the true labels to obtain the target ranking learning model.

[0027] Optionally, the neural network-based denoising method for searching distant homologous proteins, wherein the step of denoising and iteratively searching the initial candidate list using the target ranking learning model to obtain the target distant homologous protein list specifically includes:

[0028] The homology score of each initial candidate protein in the initial candidate list is calculated using the target ranking learning model, and the homology score is compared with a preset score threshold.

[0029] The initial candidate proteins whose homology scores are lower than the preset score threshold in the initial candidate list are deleted to obtain the target protein sequence;

[0030] An iterative search is performed based on the target protein sequence to obtain a list of target distantly homologous proteins.

[0031] Optionally, the method for searching distant homologous proteins based on neural network denoising, wherein the iterative search process based on the target protein sequence to obtain a list of target distant homologous proteins specifically includes:

[0032] The PSI-BLAST basic method is used to perform a third iteration search based on the target protein sequence to obtain an iterative search list;

[0033] The iterative search list is denoised to obtain a list of target distant homologous proteins.

[0034] Furthermore, to achieve the above objectives, the present invention also provides a distant homology protein search system based on neural network denoising, wherein the distant homology protein search system based on neural network denoising includes:

[0035] The superfamily model construction module is used to determine the target dataset, construct a protein sequence profile based on the target dataset, determine a preset convolutional neural network, and train the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model.

[0036] The homology feature matrix construction module is used to obtain the protein sequence to be queried, perform search processing on the target dataset based on the protein sequence to be queried to obtain an initial candidate list, and construct a homology feature matrix based on the initial candidate list and the superfamily model.

[0037] The ranking learning model training module is used to determine an initial ranking learning model and train the initial ranking learning model according to the initial candidate list and the homologous feature matrix to obtain a target ranking learning model.

[0038] The distant homologous protein output module is used to perform denoising and iterative search processing on the initial candidate list through the target ranking learning model to obtain the target distant homologous protein list.

[0039] In this invention, a target dataset is determined, a protein sequence profile is constructed based on the target dataset, a preset convolutional neural network is determined, and the preset convolutional neural network is trained based on the protein sequence profile to obtain a superfamily model; a query protein sequence is obtained, and a search process is performed on the target dataset based on the query protein sequence to obtain an initial candidate list, and a homology feature matrix is ​​constructed based on the initial candidate list and the superfamily model; an initial ranking learning model is determined, and the initial ranking learning model is trained based on the initial candidate list and the homology feature matrix to obtain a target ranking learning model; the initial candidate list is denoised and iteratively searched using the target ranking learning model to obtain a target list of distantly homologous proteins. This invention, by constructing a protein sequence profile, training a convolutional neural network based on the protein sequence profile to obtain a superfamily model, and constructing a homology feature matrix, effectively removes noise from the candidate list and filters out non-homologous sequences, thus significantly improving the accuracy and stability of distantly homologous protein detection. Attached Figure Description

[0040] Figure 1 This is a flowchart of a preferred embodiment of the distant homology protein search method based on neural network denoising of the present invention;

[0041] Figure 2 This is a schematic diagram of the convolutional neural network predictive protein superfamily framework of a preferred embodiment of the neural network-based denoising method for searching distant homologous proteins of the present invention.

[0042] Figure 3 This is a schematic diagram of the PSI-BLAST denoising framework of a preferred embodiment of the neural network-based denoising method for searching distant homologous proteins of the present invention.

[0043] Figure 4 This is a schematic diagram of the CNN-FR_PSSM framework integrating different methods, representing a preferred embodiment of the neural network-based denoising method for searching distant homologous proteins of the present invention.

[0044] Figure 5 This is a structural diagram of a preferred embodiment of the distant homology protein search system based on neural network denoising of the present invention;

[0045] Figure 6 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0047] Distant homologous proteins are proteins with low sequence similarity (usually <25%), but belonging to the same superfamily (a group of genes or proteins that evolved from a common ancestor but have little similarity), and with similar structure or function.

[0048] Currently, the detection of distantly homologous proteins mainly relies on the following five basic methods: PSI-BLAST (Position-Specific Iterative Basic Local Alignment Search Tool, a tool for protein sequence alignment), DELTA-BLAST (a protein sequence database search tool), PSI-BLASTexB (a tool for protein sequence alignment that improves the sensitivity and specificity of alignment through iterative search), HMMER (a software tool for biological sequence analysis), and HHblits (a tool for protein sequence search).

[0049] The above five methods have the following core problems in distant homology detection:

[0050] 1. Noise Accumulation in Sequence Profile Construction: PSI-BLAST relies on iteratively generated sequence profiles to characterize distant homology relationships, but this method has inherent limitations. First, it can lead to contamination of the initial candidate list: low-similarity sequences lack sufficient feature descriptions and cannot effectively filter non-homologous sequences (false positive rate >30%), resulting in distorted mutation probabilities at sequence profile sites. Second, it can lead to error propagation: noise accumulates continuously during iteration, ultimately reducing the accuracy of distant homology detection.

[0051] 2. Existing technologies exhibit uneven performance across different protein superfamilies, and while methods are complementary, they are not effectively integrated. Superfamily diversity leads to performance imbalances. Protein superfamilies are classified based on evolutionary functional conservation, but significant differences exist within them. Firstly, this results in uneven sequence quantity: some superfamilies have sparse data, leading to training bias. Secondly, it leads to complex evolutionary patterns: different superfamilies exhibit large differences in mutation rates and conserved domain distributions.

[0052] Existing methods (such as PSI-BLAST) employ a uniform modeling strategy, which struggles to adapt to all superfamily characteristics, leading to fluctuations in detection sensitivity. Experiments show that traditional methods perform differently across different superfamilies. For instance, the HHblits-based method performs best for protein superfamily classes A, D, and E, but its performance is only average for classes F and G. Furthermore, existing technologies lack a dynamic fusion mechanism, making it impossible to globally optimize search results.

[0053] In summary, to address the shortcomings of existing protein distant homology search methods, this invention aims to solve the following technical problems:

[0054] 1. Insufficient feature representation under low sequence similarity in PSI-BLAST leads to severe noise interference in the initial candidate list: The traditional method PSI-BLAST relies on manually designed alignment matrices (such as BLOSUM62). When iteratively constructing the sequence spectrum (PSSM), it cannot effectively filter out non-homologous sequences in the initial candidate list, resulting in contamination of sequence spectrum site information and a false positive rate of >30% for distant homology detection.

[0055] 2. Uneven performance across superfamilies and lack of multi-method fusion strategies: Existing unified modeling methods struggle to adapt to the evolutionary characteristics of different protein superfamilies (such as mutation rates and differences in conserved domain distribution), resulting in significantly lower detection accuracy for some superfamilies (such as the GPCR family) compared to others. Different basic methods exhibit complementary performance across different superfamilies, but lack a dynamic selection mechanism to globally optimize search results.

[0056] To address the aforementioned problems, this invention proposes a method for searching distant homologous sequences of proteins, belonging to the field of bioinformatics. It efficiently detects distant homologous proteins in a database when only the protein sequence is known but other structural or functional information is lacking.

[0057] This invention achieves superfamily classification and homology feature extraction through convolutional neural networks (CNN), combines a ranking learning model to denoise the PSI-BLAST candidate list, and dynamically fuses search results from multiple methods based on superfamily prediction, significantly improving the sensitivity and robustness of distant homology detection.

[0058] By solving the above problems, the present invention achieves the following technical effects:

[0059] 1. Reduced noise in the initial candidate list of PSI-BLAST: By filtering non-homologous sequences in the initial candidate list of PSI-BLAST using a ranking learning model, the proportion of incorrectly selected sequences is reduced by more than 80%; results with an E-value below 0.002 in PSI-BLAST are used as high-confidence results. After using ranking learning for denoising, this invention performs denoising based on a fixed threshold of 3. This high-confidence result achieves a ROC1 of 0.9754 in benchmark tests. The accuracy of PSI-BLAST detection is significantly improved before and after denoising, and the ROC1 of accuracy detection on the superfamily class G even surpasses the other four existing traditional basic methods, reaching 0.8993.

[0060] 2. Adapting to superfamily diversity and improving sensitivity for distant homology detection: A dynamic selection of the optimal search method based on convolutional neural networks for superfamily prediction addresses the performance imbalance issue, increasing the ROC1 value to over 90%.

[0061] In summary, this invention proposes a neural network-optimized framework for searching distant homologous proteins. It improves the traditional PSI-BLAST search process through deep learning and achieves intelligent fusion of multiple methods, significantly enhancing the accuracy and stability of homology detection for low-similarity proteins. This solution comprises two core technical modules: a PSI-BLAST candidate list denoising module and a superfamily adaptive method fusion module.

[0062] The core innovations include the PSI-BLAST denoising system and the superfamily adaptive fusion framework. The PSI-BLAST denoising system includes: 1. Dual feature input: Simultaneously utilizing the PSSM generated by PSI-BLAST and the HMM matrix generated by HHblits; 2. Neural network feature extraction: Predicting protein superfamily categories through convolutional neural networks (CNNs) and extracting deep homology features to construct a feature matrix; 3. Ranking learning optimization: Fusing sequence similarity features and CNN homology features, using a ranking learning model to rearrange the initial candidate list, and setting a dynamic threshold to filter non-homologous sequences (reducing the error rate by more than 80%); 4. Secondary optimization: Iteratively searching the sequence spectrum after denoising and then denoising again.

[0063] The superfamily adaptive fusion framework includes: 1. Intelligent prediction: predicting the superfamily category of the query protein through a CNN model; 2. Performance database: establishing a database of ROC1 performance of each method in different superfamilies; 3. Dynamic selection: automatically selecting the optimal search method under the superfamily (including HMMER, denoised PSI-BLAST, etc.); 4. Performance improvement: improving the overall ROC1 value to over 90%.

[0064] The technical effects include: 1. Denoising performance: The proportion of incorrectly selected sequences in the initial candidate list of PSI-BLAST is reduced by more than 80%; in the initial candidate list of the denoised PSI-BLAST, results with an E-value lower than 0.002 in PSI-BLAST are used as high-confidence results. After using ranking learning for denoising, this invention performs denoising based on a fixed threshold of 3, and the ROC1 of these high-confidence results reaches more than 95% in benchmark tests; 2. Detection accuracy: The ROC1 value of far homology search reaches more than 90%; 3. Adaptability: Effectively balances the differences in detection performance among different superfamilies.

[0065] The preferred embodiment of the present invention describes a method for searching distant homologous proteins based on neural network denoising, such as... Figure 1 As shown, the method for searching distant homologous proteins based on neural network denoising includes the following steps:

[0066] Step S10: Determine the target dataset, construct a protein sequence profile based on the target dataset, determine a preset convolutional neural network, and train the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model. The protein sequence profile includes a hidden Markov model alignment matrix and a position-specific score matrix; the preset convolutional neural network includes a first convolutional neural network and a second convolutional neural network; the superfamily model includes a first superfamily model and a second superfamily model.

[0067] It is understandable that the training of the prediction model is for predicting the superfamily of the query protein sequence and extracting homology features, and the extracted homology features are used to construct the homology feature matrix of the protein sequence model for noise reduction.

[0068] In order to effectively learn the family information in the protein sequences of the training set, the training set query protein sequence profiles PSSM and HMM are first constructed as input to the neural network. The dataset used in this method is SCOPe 2.06 (i.e. the target dataset in this invention). The SCOPe 2.06 benchmark dataset contains 28,010 protein sequences. The protein pairs have less than 95% identity with each other and are organized into 4,784 families and 2,006 superfamilies.

[0069] Specifically, a target dataset is determined, and the HHblits-based method is used to perform a first iterative search process on the target dataset to obtain the Hidden Markov Model Alignment Matrix; the PSI-BLAST-based method is used to perform a second iterative search process on the target dataset to obtain the Location-Specific Score Matrix.

[0070] like Figure 2As shown, this invention obtains the Hidden Markov Model Alignment Matrix (HMM) of the protein sequences in the entire dataset by inputting each protein sequence in the dataset into the HHblits basic method, without restricting the e_value, and iterating twice in a database constructed using the SCOPe 2.06 dataset (i.e., the first iterative search process in this invention, where the e_value is not restricted, and the sequence profiles are obtained after two iterations). Similarly, by inputting each protein sequence in the dataset into the PSI-BLAST basic method, restricting the e_value to 0.001 (because proteins with a value greater than or equal to 0.001 have a relatively large number of non-homologous proteins), iterates five times in a database constructed using the SCOPe 2.06 dataset (i.e., the second iterative search process in this invention) to obtain the Position Specific Score Matrix (PSSM) of the protein sequences in the entire dataset. Here, the Hidden Markov Model Alignment Matrix (HMM) is a statistical model used to describe the statistical characteristics of a system over time. In long-range homology detection of proteins, Hidden Markov Models (HMMs) are used to represent the generation process of protein sequences. The probability of generating each amino acid depends on the previous amino acid (or a set of previous amino acids) and the current position. By comparing the HMMs of two proteins, the optimal alignment between them can be found. Position-Specific Score Matrix (PSSM) is a statistical model for protein sequence analysis. By quantifying the preference for each amino acid at a specific position, it reflects the evolutionary conservation of protein families and key functional sites. It is the core output of the PSI-BLAST homology search algorithm. PSSM and HMMs combine evolutionary information extracted from protein alignment, making them more expressive than protein amino acid sequences.

[0071] The iterative search process includes: 1. Initial search: The query sequence is aligned to the database using BLAST (based on the BLOSUM62 matrix) to generate a list of candidate homologous proteins; 2. Sequence profile construction: Multiple sequence alignment is performed on the candidate list to generate sequence profiles (such as PSSM for PSI-BLAST, HMM for HHblits); 3. Iterative optimization: The database is searched again using the sequence profiles, low-confidence results (such as E-value exceeding the threshold) are removed, the sequence profiles are updated, and the process is repeated until convergence.

[0072] Obtain the first superfamily corresponding to the Hidden Markov Model alignment matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the Hidden Markov Model alignment matrix and the first superfamily to obtain the first superfamily model; obtain the second superfamily corresponding to the position-specific score matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the position-specific score matrix and the second superfamily to obtain the second superfamily model.

[0073] After obtaining the protein sequence profile of each protein sequence in the target dataset, the protein sequence profile can be used as input, and its corresponding superfamily as label (the dataset contains the protein name, sequence and family information. This invention constructs sequence profile PSSM and HMM through sequence search. They correspond one-to-one with the sequence, so they also correspond to the superfamily of the sequence). The model is trained using a convolutional neural network. The resulting model can be used to predict the superfamily of the query protein through the protein sequence profile and extract protein homology features.

[0074] This invention uses two types of convolutional neural networks: the first type is CNN-FR (i.e., the first convolutional neural network in this invention), and the second type is ResNet50 (i.e., the second convolutional neural network in this invention), which is deeper and more complex, for comparison.

[0075] The CNN-FR network structure is as follows: PSSM and HMM are processed into 400*20 matrices (PSSM and HMM are inherently 20-row matrices with columns representing protein sequence lengths; therefore, if the protein sequence length exceeds 400, it is truncated, and if it is less than 400, it is padded with zeros), with one input channel. The model contains five convolutional layers (conv1 to conv5), each followed by a batch normalization (BN) layer and a ReLU activation function. The ReLU (Rectified LinearUnit) activation function introduces non-linearity, helping the model learn complex patterns. Furthermore, after each convolutional layer (except for the pooling layer after the last convolutional layer conv5), a max pooling layer is followed to reduce the size of the feature maps, thereby reducing computational cost and the risk of overfitting. As the network deepens, the number of output channels (i.e., the number of convolutional kernels) of the convolutional layers gradually increases, starting from 96, then 192, then 384, and then decreasing back to 192, which helps extract increasingly complex features. The model consists of three fully connected layers (fc6, fc7, and fc8). The fc6 layer receives the flattened feature map from the last pooling layer and maps it to a higher-dimensional space (2048-dimensional). The fc7 layer further maps the features to a 1024-dimensional space. The fc8 layer is the output layer; the model's output is a vector with a length equal to the number of 2006 superfamily categories in the dataset. This output vector is used for classification prediction. It is transformed into a probability distribution using a softmax function, and the superfamily with the highest probability is the predicted superfamily.

[0076] The Softmax function is a function that transforms any real vector into a probability distribution, outputting a vector whose elements are all ∈ (0, 1) and sum to 1. The specific calculation process of the Softmax function is as follows:

[0077] Assumptions: Input vector: z = [2.0, 1.0, 0.1] (3 categories);

[0078] Step 1, Calculate the index:

[0079] e2.0≈7.389, e1.0≈2.718, e0.1≈1.105e2.0≈7.389, e1.0≈2.718, e0.1≈1.105;

[0080] Step 2, Summation:

[0081] ∑=7.389+2.718+1.105≈11.212;

[0082] Step 3, Normalization:

[0083] σ(z)=[7.38911.212, 2.71811.212, 1.10511.212]≈[0.659, 0.242, 0.099] σ(z)=[11.2127.389, 11.2122.71811.2121.105]≈[0.659, 0.242, 0.099];

[0084] Results: Probability distribution [65.9%, 24.2%, 9.9%].

[0085] ResNet50 is a deep convolutional neural network model based on the Residual Network (ResNet) architecture. The core of ResNet50 lies in its introduction of "residual connections," also known as "skip connections." These connections allow the network to directly learn the residuals (differences) between the input and output, rather than directly learning the complete mapping from input to output. This invention performs specific initialization of the weights and biases of the convolutional and batch normalization layers in the network: for example, with an input shape of [32, 1, 400, 20], the network first passes through a Conv1 layer (the network's initial convolutional layer), then through four stages, each consisting of multiple residual blocks, with the number of residual blocks within each stage being [3, 4, 6, 3]. The network concludes with a global average pooling layer and a fully connected layer. The global average pooling layer takes the average value of each channel of the feature map as the output, and obtains a fixed-length feature vector. The fully connected layer (FC layer) maps this feature vector to the final number of classes, that is, the number of superfamily classes in the dataset.

[0086] In the specific implementation process, such as Figure 2 As shown, firstly, the protein sequence profiles (PSSM) and their corresponding superfamily labels from the SCOPe 2.06 dataset are input into the neural networks CNN-FR and ResNet50. Through convolutional training, two models are obtained: CNN-FR_PSSM and ResNet50_PSSM (i.e., the first superfamily model in this invention, which includes CNN-FR_PSSM and ResNet50_PSSM). Similarly, the protein sequence profiles (HMM) and their corresponding superfamily labels from the SCOPe 2.06 dataset are input into the neural networks CNN-FR and ResNet50. Through convolutional training, two models are obtained: CNN-FR_HMM and ResNet50_HMM (i.e., the second superfamily model in this invention, which includes CNN-FR_HMM and ResNet50_HMM). The SCOPe 2.06 dataset contains a total of 2006 superfamilies. Therefore, for a query protein, the trained model will calculate the score of its class based on the learned content and finally output the superfamily with the highest score.

[0087] Step S20: Obtain the protein sequence to be queried, perform search processing on the target dataset according to the protein sequence to be queried to obtain an initial candidate list, and construct a homology feature matrix according to the initial candidate list and the superfamily model.

[0088] This invention employs a denoising system for the PSI-BLAST selected list. For a query protein sequence (i.e., the protein sequence to be queried in this invention), an initial candidate list is first obtained by iterating through PSI-BLAST once without limiting the e_value. Then, a homology feature matrix, including homology features from a convolutional neural network, is constructed and input into a trained ranking learning model. The proteins in the initial candidate list are reordered based on their recalculated homology scores (calculated by the model through learned features during training). Higher scores indicate a greater likelihood of homology. Proteins with scores below 3 are removed from the list, indicating that they are not homologous to the query protein. The denoised protein list is then used to construct a protein sequence profile, and the search continues iteratively in PSI-BLAST four times. The list after the iterative search is then denoised again (as this is the final iteration, a second denoising improves ROC1), resulting in the final search result.

[0089] Specifically, the protein sequence to be queried is obtained, and the PSI-BLAST basic method is used to search the target dataset based on the protein sequence to be queried to obtain an initial candidate list; the feature similarity between the protein sequence to be queried and the initial candidate list is calculated, and the similarity feature matrix in the protein sequence to be queried is extracted based on the feature similarity; the protein sequence to be queried is input into the superfamily model to obtain an initial homology feature matrix; and a homology feature matrix is ​​constructed based on the similarity feature matrix and the initial homology feature matrix.

[0090] For the process of obtaining the initial candidate list: This invention inputs the protein sequences in the SCOPe2.06 dataset one by one into the PSI-BLAST basic method, without limiting the e_value value, and iterates once in the database constructed using the SCOPe2.06 dataset to obtain the search list of protein sequences, which serves as the initial candidate list for optimizing the sequence profile of each protein sequence.

[0091] The construction process of the homology feature matrix is ​​as follows: based on the similarity scores from double sequence alignment, feature similarity between protein sequences is calculated using four types of protein sequence features and five similarity measurement formulas. Furthermore, eight model homology features are added, resulting in a total of six sub-matrices according to feature type. Specifically, this can be represented as:

[0092] M (q) =[M ALI M AAC M ARC M PSEAAC M PROFILE M MOD ];

[0093] Where q represents the query protein sequence, and M (q) To query the homology matrix of proteins. M ALI M represents the sequence similarity feature matrix extracted from the sequence alignment results of the query protein q. ACC M represents a protein sequence similarity feature matrix based on amino acid composition. ARC M represents a protein sequence similarity feature matrix based on protein amino acid association characteristics. PSEAAC M represents a protein sequence similarity feature matrix based on pseudo-amino acid composition. PROFILE M represents the protein sequence similarity feature matrix based on protein sequence profiles. MOD This represents the initial homogeneous feature matrix obtained based on training with CNN-FR_PSSM, CNN-FR_HMM, ResNet50_PSSM, and ResNet50_HMM. Where M... MOD The algorithm includes eight model features and utilizes four models: CNN-FR_PSSM, CNN-FR_HMM, ResNet50_PSSM, and ResNet50_HMM. For each model, two features (Feature 1 and Feature 2) are used. Taking CNN-FR_PSSM as an example: the softmax function outputs the probability distribution of the model's predictions of the query protein superfamily. The model selects the superfamily with the highest probability as the prediction result. For Feature 1: when the predicted probability of the query protein exceeds 0.7 and the predicted probability of the candidate protein also exceeds 0.7, the result is 1 if the predictions are the same, -1 if they are different, and 0 if neither condition is met. For Feature 2: when the predicted results of the query protein model and the candidate protein model are the same, their predicted probabilities are multiplied; otherwise, the result is -1.

[0094] Step S30: Determine the initial ranking learning model, and train the initial ranking learning model according to the initial candidate list and the homologous feature matrix to obtain the target ranking learning model.

[0095] This invention uses a ranking learning model to denoise the initial candidate list of the protein to be queried and construct a sequence profile. The ranking learning model is trained using five-fold cross-validation and denoising PSI-BLAST is implemented on the SCOPe2.06 dataset.

[0096] Specifically, the initial candidate proteins in the initial candidate list are obtained, and the homology features and true labels of the initial candidate proteins and the query protein sequence are obtained; an initial ranking learning model is determined, and the initial ranking learning model is trained using the five-fold cross-validation method based on the homology features and the true labels to obtain the target ranking learning model.

[0097] The process of constructing the model training and test sets is as follows: A PSI-BLAST search is used to generate an initial candidate list of query proteins. Each protein in the list is called an initial candidate protein. If a query protein is homologous to a candidate protein, the true homology label is 1; otherwise, it is 0. A one-to-one correspondence between the homology features and true labels of candidate proteins and query proteins is used as a sample data point for the query protein. Each protein in the dataset is used sequentially as a query protein, and its homology features and true labels constitute the dataset for training the ranking learning model. This dataset is divided into five equal parts, and one part is selected in turn as the test set for each fold, with the remaining four parts used as the training set.

[0098] For the model training process: This invention uses the LambdaMART algorithm tool within the RankLib package to train the model. In each training fold, RankLib reads the training dataset, learns homology features, recalculates the query protein and each list protein, effectively distinguishing scores for non-homologous proteins, and reorders the proteins in the list from high to low scores. This training process is repeated five times, and the model is retained. By evaluating the ROC1 and ROC50 on the test set, the model with the best performance is selected as the final model.

[0099] Step S40: The initial candidate list is denoised and iteratively searched using the target ranking learning model to obtain a list of target distant homologous proteins.

[0100] Specifically, the homology score of each initial candidate protein in the initial candidate list is calculated by the target ranking learning model, and the homology score is compared with a preset score threshold; the initial candidate proteins in the initial candidate list whose homology scores are lower than the preset score threshold are deleted to obtain the target protein sequence.

[0101] The reordered list is then filtered: For proteins in the reordered list, a higher score indicates a higher probability of homology with the query protein. A threshold can be set to filter out non-homologous proteins. This invention evaluates the ROC of the filtered list and sets the threshold at 3 (i.e., the preset score threshold in this invention). Candidate proteins with scores below 3 are considered non-homologous to the query protein, and thus deleted from the list. The filtered list is then used to construct the protein sequence profile.

[0102] To better compare the effects of PSI-BLAST before and after denoising the selected list, proteins in the scope2.06 dataset were divided into five categories based on the initial candidate list generated by PSI-BLAST: (i) Correctly selected proteins: The proteins used to construct the sequence profile in the initial candidate list are all homologous to the query protein. The constructed sequence profile contains only correct homology information, which is most effective in finding more homologous proteins in the iterative search. (ii) Mixed incorrectly selected proteins: The proteins used to construct the sequence profile in the initial candidate list include both homologous and non-homologous proteins. The constructed sequence profile contains both correct homology information and incorrect non-homologous information. (iii) Completely empty incorrectly selected proteins: There are no proteins used to construct the sequence profile in the initial candidate list, i.e., there are neither homologous nor non-homologous proteins, but homologous proteins are among the proteins not selected for constructing the sequence profile in the initial candidate list. (iv) Completely incorrectly selected proteins: All non-homologous proteins are among the proteins selected for constructing the sequence profile in the initial candidate list, but homologous proteins are among the proteins not selected for constructing the sequence profile in the initial candidate list. (v) Error: No homologous proteins were found in the entire initial candidate list, meaning there was no valid information.

[0103] For the four types of errors mentioned above, the first three—incorrectly selected proteins (i.e., a mixture of incorrectly selected proteins, completely empty incorrectly selected proteins, and completely wrong incorrectly selected proteins)—can be corrected by reordering the initial candidate list, placing homologous proteins higher and non-homologous proteins lower, thus filtering out non-homologous proteins and making them correctly selected. However, those that cannot be selected cannot be improved because there is no valid information in the initial list.

[0104] like Figure 3 As shown, for the initial candidate list generated by PSI-BLAST, proteins in the candidate list are often selected by limiting a threshold before denoising to construct the sequence profile. This invention compares the changes in the number of incorrectly selected proteins before and after denoising for the selected list with e_value <= 0.002 under the scope2.06 dataset (as shown in Table 1). It can be seen that: before denoising, there are 245 mixed incorrectly selected proteins, of which 211 are transformed into correct selections; there are 894 completely empty incorrectly selected proteins, of which 735 are transformed into correct selections; and there are 7 completely wrong incorrectly selected proteins, of which 5 are transformed into correct selections. After denoising, the total proportion of incorrectly selected sequences is reduced by 82.98%.

[0105] The PSI-BLAST basic method is used to perform a third iteration search based on the target protein sequence to obtain an iterative search list; the iterative search list is then denoised to obtain a list of target distant homologous proteins.

[0106] After obtaining the target protein sequence, the optimized protein sequence profile (i.e., the target protein sequence in this invention) is used for iterative search, and the list after iterative search is denoised again.

[0107] Furthermore, the optimized sequence spectrum is input into the PSI-BLAST base method, with e_value set to the default. The search is iterated four times, outputting a search list. A homology feature matrix is ​​then reconstructed, with a size of L2*111, where L2 is the length of the candidate homology protein list generated after the iterative search. Each protein in the list has 111 features. The homology feature matrix is ​​then used to retrain the ranking learning model, and the list is reordered to obtain the final far homology protein search list.

[0108] Comparing the ROC1 performance of the search list after 5 iterations in various protein superfamilies before and after using the denoising system in PSI-BLAST, as shown in Tables 2 and 3: the accuracy of PSI-BLAST detection is significantly improved before and after denoising, and the accuracy ROC1 of detection in superfamily G even exceeds that of the other four existing traditional basic methods, reaching 0.8993.

[0109] Furthermore, this invention proposes a dynamic fusion framework for various methods based on superfamily prediction, which uses a neural network to dynamically select traditional methods to complete the search task, thereby achieving complementary advantages of different search methods.

[0110] This invention uses a neural network fusion with traditional basic search methods: This invention calculates the ROC1 of different protein superfamily classes after five iterations on the scope2.06 dataset, including PSI-BLASTexB, DELTA-BLAST, HHblits, HMMER, and denoised PSI-BLAST.

[0111] As shown in Table 3, it can be seen that the search performance of distant homology sequences for different protein superfamily categories exhibits a complementary relationship. The HHblits-based method performs best for protein superfamily categories A, D, and E, but its performance is average for categories F and G. The DELTA-BLAST-based method performs best for protein superfamily categories B and F, but its performance is average for categories A and G. The HMMER-based method performs best for protein superfamily category C, while the denoised PSI-BLAST method performs best for protein category G.

[0112] This invention integrates five search methods—PSI-BLASTexB, DELTA-BLAST, HHblits, HMMER, and denoised PSI-BLAST—using the CNN-FR_PSSM convolutional neural network framework.

[0113] like Figure 4 As shown, for a query protein sequence, it is first input into the PSI-BLAST basic method, with e_value constrained to 0.001. The search is iteratively performed five times in a database constructed using the SCOPe 2.06 dataset to obtain the position-specific score matrix (PSSM) of the query protein sequence. The PSSM is then padded and cropped to a 400*20 matrix and input into the CNN-FR_PSSM model to predict its superfamily. The best-performing method within this superfamily is selected from the five methods mentioned above for the search task. As shown in Table 3, after five iterations, the ROC1 of the search list is 0.9066.

[0114] The application value of this invention (which can be widely applied to): 1. Target protein identification in new drug development; 2. Functional annotation of unknown proteins; 3. Discovery of biomarkers for precision medicine.

[0115] Technical effects of the present invention:

[0116] 1. Regarding the noise reduction effect of PSI-BLAST:

[0117] (1) As shown in Table 1, it can be seen from the last column that before denoising the selected list used by PSI-BLAST to construct the sequence profile, the total number of incorrectly selected proteins that can be improved is 1146; compared with the number corrected in the first column, 951 incorrectly selected proteins became correctly selected after denoising, and the error rate was reduced by 82.98%.

[0118] (2) As shown in Table 2, results with an E-value lower than 0.002 in PSI-BLAST are considered as high-confidence results. After using ranking learning for denoising, this invention performs denoising based on a fixed threshold of 3. The ROC1 of these high-confidence results reaches 0.9754 in the benchmark test.

[0119] (3) As shown in Tables 3 and 4, the accuracy of PSI-BLAST increased significantly before and after using the denoising system of the present invention. The accuracy ROC1 of the superfamily class G even exceeded that of the other four traditional basic methods, reaching 0.8993.

[0120] 2. Regarding detection accuracy: As shown in Table 4, after dynamic fusion via neural network, the ROC1 value of the far-homogeneous search reaches 90.66%.

[0121] 3. Regarding adaptability: It can effectively balance the differences in detection performance among different superfamilies. As shown in Table 2, after dynamic fusion, the performance of each protein superfamily is relatively balanced.

[0122] The following are the details of Tables 1, 2, 3, and 4:

[0123] Table 1: Changes in the number of correctly selected proteins and the number of incorrectly selected proteins in the selection list for PSI-BLAST sequence profiling before and after noise reduction.

[0124]

[0125]

[0126] Table 2: ROC1 results at different thresholds after PSI-BLAST denoising

[0127]

[0128] Table 3: ROC1 performance of 5 iterations of PSI-BLAST before and after denoising in different categories of protein superfamilies.

[0129]

[0130] Table 4: ROC1 performance of PSI-BLASTexB, DELTA-BLAST, HHblits, HMMER, denoised PSI-BLAST, and neural network dynamic fusion method after five iterations.

[0131]

[0132]

[0133] Furthermore, such as Figure 5 As shown, based on the above-described neural network-based denoising method for searching distant homologous proteins, this invention also provides a neural network-based denoising system for searching distant homologous proteins, wherein the neural network-based denoising system for searching distant homologous proteins includes:

[0134] The superfamily model construction module 51 is used to determine the target dataset, construct a protein sequence profile based on the target dataset, determine a preset convolutional neural network, and train the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model.

[0135] The homology feature matrix construction module 52 is used to obtain the protein sequence to be queried, perform search processing on the target dataset according to the protein sequence to be queried, obtain an initial candidate list, and construct a homology feature matrix according to the initial candidate list and the superfamily model.

[0136] The ranking learning model training module 53 is used to determine an initial ranking learning model and train the initial ranking learning model according to the initial candidate list and the homologous feature matrix to obtain a target ranking learning model.

[0137] The distant homologous protein output module 54 is used to perform denoising and iterative search processing on the initial candidate list through the target ranking learning model to obtain the target distant homologous protein list.

[0138] Furthermore, such as Figure 6 As shown, based on the above-mentioned method and system for searching distant homologous proteins using neural network denoising, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 6 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0139] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a neural network-based denoising distant homologous protein search program 40, which can be executed by the processor 10 to implement the neural network-based denoising distant homologous protein search method of this application.

[0140] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the neural network-based denoising method for searching distant homologous proteins.

[0141] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0142] In one embodiment, the steps of the neural network-based denoising distant homologous protein search method are implemented when the processor 10 executes the neural network-based denoising distant homologous protein search program 40 in the memory 20.

[0143] In summary, this invention provides a method, system, and terminal for searching distant homologous proteins based on neural network denoising. The method includes: determining a target dataset; constructing a protein sequence profile based on the target dataset; determining a preset convolutional neural network; and training the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model; acquiring a query protein sequence; performing search processing on the query protein sequence in the target dataset to obtain an initial candidate list; and constructing a homology feature matrix based on the initial candidate list and the superfamily model; determining an initial ranking learning model; and training the initial ranking learning model based on the initial candidate list and the homology feature matrix to obtain a target ranking learning model; and performing denoising processing and iterative search processing on the initial candidate list using the target ranking learning model to obtain a target list of distant homologous proteins. This invention constructs a protein sequence profile, trains a convolutional neural network based on the protein sequence profile to obtain a superfamily model, and constructs a homology feature matrix. By using a ranking learning model and the homology feature matrix to perform denoising and iterative search on the candidate list, it can effectively remove noise from the candidate list and filter out non-homologous sequences in the candidate list, thus effectively improving the accuracy and stability of distant homologous protein detection.

[0144] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal 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 terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0145] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0146] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for searching distant homologous proteins based on neural network denoising, characterized in that, The neural network-based denoising method for searching distant homologous proteins includes: A target dataset is determined, a protein sequence profile is constructed based on the target dataset, a preset convolutional neural network is determined, and the preset convolutional neural network is trained based on the protein sequence profile to obtain a superfamily model; Obtain the protein sequence to be queried, perform search processing on the target dataset based on the protein sequence to be queried to obtain an initial candidate list, and construct a homology feature matrix based on the initial candidate list and the superfamily model; An initial ranking learning model is determined, and the initial ranking learning model is trained based on the initial candidate list and the homologous feature matrix to obtain the target ranking learning model; The target ranking learning model is used to denoise and iteratively search the initial candidate list to obtain a list of target distant homologous proteins.

2. The method for searching distant homologous proteins based on neural network denoising according to claim 1, characterized in that, The protein sequence profile includes a hidden Markov model alignment matrix and a position-specific score matrix; The process of determining the target dataset and constructing a protein sequence profile based on the target dataset specifically includes: The target dataset is determined, and the HHblits-based method is used to perform the first iterative search process on the target dataset to obtain the Hidden Markov Model Alignment Matrix; The PSI-BLAST basic method is used to perform a second iterative search process on the target dataset to obtain the location-specific score matrix.

3. The method for searching distant homologous proteins based on neural network denoising according to claim 2, characterized in that, The preset convolutional neural network includes a first convolutional neural network and a second convolutional neural network; the superfamily model includes a first superfamily model and a second superfamily model; The step of determining a preset convolutional neural network and training the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model specifically includes: Obtain the first super family corresponding to the Hidden Markov Model alignment matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the Hidden Markov Model alignment matrix and the first super family to obtain the first super family model. Obtain the second superfamily corresponding to the position-specific score matrix, and perform convolution training on the first convolutional neural network and the second convolutional neural network according to the position-specific score matrix and the second superfamily respectively to obtain the second superfamily model.

4. The method for searching distant homologous proteins based on neural network denoising according to claim 2, characterized in that, The process of obtaining the protein sequence to be queried, performing a search on the target dataset based on the protein sequence to obtain an initial candidate list, and constructing a homology feature matrix based on the initial candidate list and the superfamily model specifically includes: Obtain the protein sequence to be queried, and use the PSI-BLAST basic method to search the target dataset based on the protein sequence to obtain an initial candidate list; Calculate the feature similarity between the protein sequence to be queried and the initial candidate list, and extract the similarity feature matrix from the protein sequence to be queried based on the feature similarity; The protein sequence to be queried is input into the superfamily model to obtain the initial homology feature matrix; A common origin feature matrix is ​​constructed based on the similarity feature matrix and the initial common origin feature matrix.

5. The method for searching distant homologous proteins based on neural network denoising according to claim 1, characterized in that, The process of determining an initial ranking learning model and training the initial ranking learning model based on the initial candidate list and the homogeneous feature matrix to obtain a target ranking learning model specifically includes: Obtain the initial candidate proteins from the initial candidate list, and obtain the homology features and true labels of the initial candidate proteins and the query protein in the query protein sequence; An initial ranking learning model is determined, and the initial ranking learning model is trained using five-fold cross-validation based on the homologous features and the true labels to obtain the target ranking learning model.

6. The method for searching distant homologous proteins based on neural network denoising according to claim 2, characterized in that, The step of performing noise reduction and iterative search processing on the initial candidate list using the target ranking learning model to obtain the target far homologous protein list specifically includes: The homology score of each initial candidate protein in the initial candidate list is calculated using the target ranking learning model, and the homology score is compared with a preset score threshold. The initial candidate proteins whose homology scores are lower than the preset score threshold in the initial candidate list are deleted to obtain the target protein sequence; An iterative search is performed based on the target protein sequence to obtain a list of target distantly homologous proteins.

7. The method for searching distant homologous proteins based on neural network denoising according to claim 6, characterized in that, The iterative search process based on the target protein sequence to obtain a list of target distantly homologous proteins specifically includes: The PSI-BLAST basic method is used to perform a third iteration search based on the target protein sequence to obtain an iterative search list; The iterative search list is denoised to obtain a list of target distant homologous proteins.

8. A distant homology protein search system based on neural network denoising, characterized in that, The neural network-based denoising system for searching distant homologous proteins includes: The superfamily model construction module is used to determine the target dataset, construct a protein sequence profile based on the target dataset, determine a preset convolutional neural network, and train the preset convolutional neural network based on the protein sequence profile to obtain a superfamily model. The homology feature matrix construction module is used to obtain the protein sequence to be queried, perform search processing on the target dataset based on the protein sequence to be queried to obtain an initial candidate list, and construct a homology feature matrix based on the initial candidate list and the superfamily model. The ranking learning model training module is used to determine an initial ranking learning model and train the initial ranking learning model according to the initial candidate list and the homologous feature matrix to obtain a target ranking learning model. The distant homologous protein output module is used to perform denoising and iterative search processing on the initial candidate list through the target ranking learning model to obtain the target distant homologous protein list.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a neural network-based denoising distant homologous protein search program stored in the memory and executable on the processor. When the neural network-based denoising distant homologous protein search program is executed by the processor, it implements the steps of the neural network-based denoising distant homologous protein search method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a neural network-based denoising distant homologous protein search program, which, when executed by a processor, implements the steps of the neural network-based denoising distant homologous protein search method as described in any one of claims 1-7.

Images