A method for predicting the number of homo-oligomer subunits based on deep learning

Abstract

The present invention discloses a method for predicting the number of homo-oligomer subunits based on deep learning, belonging to the field of protein function prediction technology. Based on the sequences of monomeric and oligomeric proteins collected from the UniProt database and their corresponding subunit numbers, the DeepSub model was developed and constructed using deep learning technology to predict the number of protein subunits. Through a rigorous one-to-one comparison verification of Complex Portal data, the DeepSub model demonstrated high accuracy and demonstrated high value in assisting the AlphaFold2 method in accurately predicting the actual protein structure.

Description

Technical Field

[0001] The present invention relates to the technical field of protein structure prediction, and more specifically, to a method for predicting the number of homo-oligomer subunits based on deep learning. Background Art

[0002] Protein oligomerization is a well-established phenomenon responsible for the function of biological systems, encompassing both eukaryotic and prokaryotic organisms, and involving approximately 30% of total proteins. Protein oligomers have recently attracted significant attention in structural biology, chemical biology, and neurodegeneration. These oligomers typically consist of a limited number of monomers, ranging from two to ten, and possess considerable combinatorial potential, particularly through hetero- and homo-oligomerization. Furthermore, homo-oligomers are known to play key roles in mediating and regulating processes such as gene expression, enzyme function, ion channels, receptors, and intercellular adhesion. Therefore, understanding the structure of homo-oligomers at the molecular level is crucial for understanding the physiological functions of proteins and designing molecular modulators of their regulation.

[0003] The experimental characterization of the structure of homo-oligomers, their dynamic equilibrium between various oligomeric states, and the inference of biologically relevant oligomeric configurations from their crystallized states have been daunting challenges. The main experimental techniques used for this purpose include X-ray and neutron scattering, mass spectrometry, gel filtration, dynamic light scattering, analytical ultracentrifugation, and fluorescence resonance energy transfer (FRET). X-ray crystallography is the most comprehensive method for experimentally determining the structure of protein-protein complexes. However, accurately distinguishing biologically relevant interfaces from crystallographic interfaces is not easy and is prone to errors. In addition, determining the spatial structure of proteins through various experimental methods is time-consuming and expensive. To alleviate the overhead and labor-intensiveness of these methods, computational protocols have also been developed, mainly starting from solved crystal structures.

[0004] Among them, Alphafold2 is one of the most influential neural network-based protein structure prediction methods in recent years. It uses amino acid sequences as input to accurately predict the three-dimensional structure of proteins. Alphafold2 and AlphaFold-multimer methods can predict complex structures by specifying the number of subunits. However, researchers often have difficulty determining the actual number of subunits in a target protein, making it challenging to directly obtain ideal results. This problem also exists when using molecular docking methods to simulate protein structures, such as GalaxyHomomer and SWISS-MODEL.

[0005] Determining the tetrameric state of a protein using structural information is a common and intuitive approach. However, these approaches have drawbacks, particularly when experimentally determined structures are unavailable due to the computationally intensive nature of structure prediction. Notably, significant progress has recently been made in deep learning for predicting protein structure and other properties. Protein language models, which utilize computational natural language models to process proteins, have successfully captured secondary structure, protein cellular localization, and other features from a single sequence. The question arises: can we similarly infer tetrameric structure from protein sequence alone? Recently, Orly et al. published a study titled "QUEEN" on BioRxiv that explored the use of a pretrained model, pLM ESM2, to predict protein tetrameric structure from amino acid sequences. Surprisingly, the QUEEN model's performance did not reach the level of previous machine learning-based studies.

[0006] Therefore, the present invention proposes a method for predicting the number of homo-oligomer subunits based on deep learning. Summary of the Invention

[0007] 1. Technical problems to be solved

[0008] In response to the problems existing in the prior art, the purpose of the present invention is to provide a method for predicting the number of homo-oligomer subunits based on deep learning, which can retrieve the sequences of monomeric proteins and oligomeric proteins and their corresponding subunit numbers from the UniProt database. Based on deep learning technology, a DeepSub model is developed and constructed to predict the number of protein subunits. Through rigorous verification of Complex Portal data, the DeepSub model shows high accuracy. In addition, the DeepSub model has shown a very valuable role in assisting the AlphaFold2 method to accurately predict the actual structure of proteins.

[0009] 2. Technical solution

[0010] To solve the above problems, the present invention adopts the following technical solutions.

[0011] A method for predicting the number of homo-oligomer subunits based on deep learning includes the following steps:

[0012] S1. Collect and extract protein sequences and subunit numbers from known databases to create a dataset;

[0013] S2. Developed the DeepSub model based on recurrent neural network (RNN) and trained the model using the created dataset;

[0014] S3. Input the homo-oligomer to be predicted into the trained model to obtain the prediction result of its subunit number.

[0015] Furthermore, the construction of the DeepSub model in step S2 includes the following steps:

[0016] S21. First, we obtain the semantic representation of protein sequences through the unified rational protein engineering (UniR ep) based on sequence-based deep representation learning;

[0017] S22. Subsequently, we use an architecture based on a bidirectional gated recurrent unit (Bi-directional GRU) and an attention layer to learn biological questions related to the downstream subunit prediction task.

[0018] S23. Finally, the output from the attention layer is connected to a fully connected output layer with a Softmax activation function corresponding to the number of categories to obtain the probability of each category.

[0019] Furthermore, the two-layer GRU-RNN model is constructed in step S22, including the following steps:

[0020] S221. Define the input layer: Based on the tensorflow framework, use the keras 'input' function to define the model input, 'inputs = Input(shape = (1,900), name = 'input')';

[0021] S222. Build a bidirectional GRU layer: Create a bidirectional GRU layer using the Bidreactional wrapper and the GRU layer, gru = Bidirectional(GRU(128, dropout = 0.5, return_sequences = True), name = "bi-gru")(inputs);

[0022] S223, add attention mechanism: introduce attention layer, "attention = Attention (32,

[0023] name='attention')(gru)";

[0024] S224. Apply DropOut: Add a Dropout layer after the attention layer, attention_with h_dropout = Dropout(0.1, name="attention_dropout")(attention)";

[0025] S225. Define the output layer: Add a fully connected layer as the output layer and use the softmax activation function, "output = Dense(num_class, activation = 'softmax', name = "dense") (attention_with_dropout)";

[0026] S226. Build and compile the model: Use the "Model" function to build the model, then compile the model, use the "fit" method to train the model, and then use the validation set and test set to evaluate the performance of the model;

[0027] S227. Adjust the hyperparameters based on the model's performance on the validation set, setting the learning rate to lr = 0.01, the epoch to 200, and the batch_size to 1024.

[0028] Furthermore, the input matrix in the model is X∈R^ (1xN) , where N = 1900, X represents the UniRep representation of the input protein, and a bidirectional GRU layer with 128 hidden units is used. The bidirectional structure enables the model to capture sequence dependencies from both the forward and reverse directions. The process is expressed as:

[0029] H gru =f gru ru(X);

[0030] Among them H gru ∈R^ (1xH) is the output of the GRU layer, f gru is the function of the bidirectional GRU layer, and X is the UniRep representation of the input.

[0031] Furthermore, a dropout rate of 20% is set, and each hidden unit has a 20% probability of being randomly selected and temporarily removed from the network.

[0032] Furthermore, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0033] Furthermore, in step S226, when the performance on the validation set cannot be improved, the early stopping method is used to terminate the training.

[0034] Furthermore, DeepSub takes a protein sequence as input and outputs a label representing the number of subunits of the input protein, and uses a loss function that calculates the cross entropy between the true distribution and the predicted distribution. The specific formula is as follows:

[0035] H(p,q)=-∑p(x)logq(x)

[0036] Where p is the true distribution, q is the predicted distribution, Σ represents the sum over all categories, x is a specific category, p(x) is the probability of category x in the true distribution, and q(x) is the probability of category x in the predicted distribution.

[0037] Furthermore, it also includes evaluating model performance, specifically defining four commonly used metrics: mACC (macro average accuracy), mPR (macro average precision), mRecall (macro average recall), and mF1 (macro average F1 score). The specific formulas are as follows:

[0038]

[0039]

[0040]

[0041]

[0042] Where N represents the total number of categories, and ACC_i, PPV_i, and Recall_i represent the accuracy, precision, and recall of the i-th category in a multi-class classification task, respectively.

[0043] Furthermore, when evaluating the performance of the model, we first obtained information about protein complexes from the database, then filtered out data related to heterologous proteins, and then extracted data about homo-oligomers to create an independent dataset for evaluating the performance of DeepSub and traditional machine learning models.

[0044] 3. Beneficial effects

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] This solution can retrieve the sequences of monomeric and oligomeric proteins and their corresponding subunit numbers from the UniProt database. Based on deep learning technology, the DeepSub model is developed and constructed to predict the number of protein subunits. Through rigorous verification of Complex Portal data, the DeepSub model shows high accuracy. In addition, the DeepSub model has shown a valuable role in assisting the AlphaFold2 method in accurately predicting the actual structure of proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is an analysis diagram of the homo-oligomer dataset from UniProt of the present invention;

[0048] Where: (A) Distribution of subunit numbers in the homo-oligomer dataset;

[0049] (B) Distribution of monomers and homo-oligomers in four model organisms as recorded in UniProt;

[0050] (C) Distribution of protein length and protein length multiplied by the number of subunits;

[0051] (D) EC numbers correspond to species of proteins composed of different subunits;

[0052] (E) with EC numbers corresponding to the distribution of proteins with subunit composition greater than 1;

[0053] Figure 2 is a similarity distribution diagram of protein pairs with the same EC number of the present invention;

[0054] Among them: (A) Similarity distribution of protein pairs with the same EC number but different number of subunits;

[0055] (B) Similarity distribution of protein pairs with the same EC number and the same number of subunits;

[0056] Figure 3 This is a diagram showing the construction and result analysis of the deep learning model of the present invention;

[0057] Among them: (A) data source and deep learning model framework;

[0058] (B) Confusion matrix of the test set;

[0059] (C) Visualization of the reduced dimensionality test set;

[0060] (D) Accuracy of different identity intervals in the explosion results of the test set and training set;

[0061] Figure 4 Graphs representing the performance of DeepSub, QUEEN, and four machine learning methods of the present invention on (A) the test set and (B) complex portal data;

[0062] Figure 5 Comparison diagram between the PDB structure (A), Uniprot structure (B) and DeepSub-assisted AlphaFold predicted structure of the present invention

[0063] Figure 6 This is a performance indicator diagram of the accuracy, recall rate and F1 score of different tags in the test set of this invention. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] Example:

[0066] First, a dataset containing the amino acid sequences of proteins with known subunit numbers is collected. The interaction section of the UniProt database uses terms such as "monomer" or "homodimer" to describe the number of subunits of a protein.

[0067] For example, the "interaction" term for the Escherichia coli K12 aspartate aminotransferase, UniProt number "P00509," begins with the keyword "homodimer," indicating that the protein has two subunits. Other keywords with the same mapping relationship include "homodimer" and "forms homodimers." The mapping between keywords and subunit numbers includes 10 labels: "monomer, homodimer, homotrimer, homotetramer, homopentamer, homohexamer, homoheptamer, homooctamer, homodecamer, and homododecamer."

[0068] QUEEN's training dataset comes from the QsBio database, which compiles three-dimensional structural information of biomolecules from the Protein Data Bank (PDB). PDBe-KB is a comprehensive knowledge base that integrates protein structure, annotation, and function information, providing a wealth of protein-related data and visualization tools. Key information from the QsBio data is linked to the PDBe-KB API to establish links with UniProt records. For example, in QsBio, entry 101m corresponds to UniProt entry P02185, and this link facilitates data comparison.

[0069] First, the semantic representation of protein sequences is obtained through the unified rational protein engineering (UniRep) based on sequence-based deep representation learning. Subsequently, the biological problem related to the downstream subunit prediction task is learned using an architecture based on a bidirectional gated recurrent unit (Bi-directional GRU) and an attention layer. Finally, the output from the attention layer is connected to a fully connected output layer with a softmax activation function corresponding to the number of categories to obtain the probability of each category. DeepSub is divided into three core components: UniRep module, bio_gru module and fully connected module.

[0070] A self-supervised language model based on a recurrent neural network (RNN), called UniRep, is used to generate a fixed-size representation (1900 dimensions) for the entire protein sequence, effectively encoding proteins of arbitrary length. Furthermore, UniRep can be widely applied to various protein-related bioinformatics tasks, including but not limited to protein function prediction, protein interaction prediction, and protein structure prediction.

[0071] (1) Define the input layer

[0072] First, use the tensorflow framework for development and use the keras 'input' function to define the model input, for example, 'inputs = Input(shape = (1,900), name = 'input')'.

[0073] (2) Constructing a bidirectional GRU layer

[0074] Create a bidirectional GRU layer using the Bidreactional wrapper and the GRU layer, for example: gru = Bidirectional(GRU(128,dropout=0.5,return_sequences=True),name="bi-gru")(inputs).

[0075] (3) Adding attention mechanism

[0076] Introduce the attention layer, for example "attention = Attention(32, name = 'attention')(gru)"

[0077] (4) Apply DropOut

[0078] Add a Dropout layer after the attention layer, for example "attention_with_dropout=Dr opout(0.1,name="attention_dropout")(attention)"

[0079] (5) Define the output layer

[0080] Add a fully connected layer as the output layer and use the softmax activation function, for example "output = Dense(num_class, activation = 'softmax', name = "dense")(attention_with h_dropout)"

[0081] (6) Build and compile the model

[0082] Use the "Model" function to build the model and then compile the model. Use the "fit" method to train the model and then evaluate the performance of the model using the validation set and test set. If the performance on the validation set does not improve, use early stopping to terminate the training.

[0083] (7) Adjust the hyperparameters based on the performance of the model on the validation set, setting the learning rate lr = 0.01, the epoch to 200, and the batch_size to 1024.

[0084] In the model, a recurrent neural network (RNN) structure called a gated recurrent unit (GRU) is first adopted. Compared with the traditional RNN, GRU alleviates the problem of gradient disappearance by introducing "update gate" and "reset gate", thereby more effectively capturing long-distance dependencies. The input matrix is ​​X∈R^ (1xN) , where N = 1900 and X represents the UniRep representation of the input protein. The model uses a bidirectional GRU layer with 128 hidden units. The bidirectional structure enables the model to capture sequence dependencies in both the forward and reverse directions. This complete view of the context usually performs better in sequence tasks. This process can be expressed as:

[0085] H gru =f gru ru(X)

[0086] Among them H gru ∈R^ (1xH) is the output of the GRU layer, fgru is the function of the bidirectional GRU layer, and X is the UniRep representation of the input. A 20% dropout rate is also applied, which is a regularization technique that randomly "drops" neurons to prevent overfitting. When the dropout rate is set to 0.2, it means that at each training step, each hidden unit has a 20% probability of being randomly selected and temporarily removed from the network.

[0087] The dataset is divided into training, validation, and test sets in a ratio of 8:1:1. DeepSub takes a protein sequence as input and outputs a label indicating the number of subunits of the input protein. In this task, a loss function commonly used in multi-classification problems is used, which calculates the cross entropy between the true distribution and the predicted distribution. The specific formula is as follows:

[0088] H(p,q)=-∑p(x)logq(x)

[0089] Where p is the true distribution, q is the predicted distribution, Σ represents the sum over all categories, x is a specific category, p(x) is the probability of category x in the true distribution, and q(x) is the probability of category x in the predicted distribution.

[0090] To evaluate model performance, four common metrics are calculated and defined: mACC (macro-average accuracy), mPR (macro-average precision), mRecall (macro-average recall), and mF1 (macro-average F1 score). mACC is the macro-average accuracy, which is calculated for each class individually and then averaged. Specifically, for each class, the accuracy is calculated by dividing the number of correctly predicted examples by the total number of examples; then, the accuracy for all classes is summed; and finally, the sum is divided by the number of classes.

[0091] mPR is the macro-average precision, which is calculated by calculating the precision for each class individually and then averaging these precisions. Specifically, for each class, calculate the precision, divide the number of correctly predicted examples for that class by the total number of examples predicted for that class, then sum the precision for all classes, and finally divide the sum by the number of classes. mRecall is the macro-average recall, which is calculated by calculating the recall for each class individually and then averaging these recalls. Specifically, for each class, calculate the recall, divide the number of correctly predicted examples for that class by the total number of examples actually predicted for that class, then sum the recall for all classes, and finally divide the sum by the number of classes. mF1 is the macro-average F1 score, which is calculated by calculating the F1 score for each class individually and then averaging these F1 scores. Specifically, for each class, calculate the F1 score: 2 x ((precision x recall) / (precision + recall)), then sum the F1 scores for all classes, and finally divide the sum by the number of classes.

[0092]

[0093]

[0094]

[0095]

[0096] Where N represents the total number of categories, and ACC_i, PPV_i, and Recall_i represent the accuracy, precision, and recall of the i-th category in a multi-class classification task, respectively.

[0097] We obtained information on protein complexes from the Complex Portal database and filtered out data related to heterologous proteins, such as the glycine cleavage system complex (composed of gcvT and gcvP and gcvH and lpdA). We then extracted data on homo-oligomers, such as the complement factor H complex (composed of CFH), to create an independent dataset that was used to evaluate the performance of DeepSub, QUEEN, and four traditional machine learning models.

[0098] AlphaFold2, developed by the DeepMind team, represents an advanced deep learning model that can accurately predict protein structures. Using the AlphaFold2 model deployed on the Google Colab platform, we predict the three-dimensional structure of proteins. The target protein sequence is used as input data, using a format similar to "PI...SK:PI...SK for ahomodimer" to specify the number of subunits in the homo-oligomer, while keeping all other parameters at their default settings.

[0099] First, protein sequences and their corresponding tags were selected from the UniProt database. Based on manually summarized keyword-to-tag mappings, a dataset of 110,000 entries was obtained. Analysis of the distribution of tags in the dataset showed that the proportion of homodimers and monomers was significantly higher than that of other homo-oligomers. Overall, homo-oligomers with even subunits were more common than those with odd subunits, and the size of the dataset decreased as the number of subunits increased ( Figure 1 A). Furthermore, species-specific analyses of four widely studied organisms revealed similar ratios of protein types between humans and mice. However, yeast exhibited a higher proportion of homotrimers, while Escherichia coli showed a higher proportion of homotetramers and a lower proportion of homodimers ( Figure 1 B).

[0100] The QUEEN training dataset consists of 69,773 entries. When these 69,773 structures were linked to records in the UniProt database, they corresponded to 22,888 different proteins. This discrepancy is due to the inclusion of incomplete chain crystal structures and structural data for protein-nucleic acid complexes in the PDB. This means that many sequences in the QUEEN training dataset should not be considered valid proteins, resulting in a significant amount of noise in the dataset.

[0101] The relationship between protein length and number of subunits, e.g. Figure 1 C. Proteins with varying numbers of subunits were observed to exhibit a distribution of lengths between 200 and 400 amino acids. The product of the number of subunits and the length of the protein sequence can be considered the total number of amino acids involved in functional protein folding. Figure 1 Figure C shows that as the number of subunits increases, the length of the functional protein also increases. The formation of protein oligomers, such as homo-oligomers, may be attributed to the tendency of overly long amino acid sequences in functional proteins to fold incorrectly. Through evolution, these proteins gradually evolved into multimeric forms with multiple subunits.

[0102] The dataset also contains a total of 2,575 EC numbers and their corresponding subunit counts. Of these, 80.7% of EC numbers are associated with only one homo-oligomer in the database, while 15.3% have two types of homo-oligomers. The EC numbers with the most homo-oligomer types are 1.15.1.1 and 4.2.1.1, both of which include monomers, homodimers, homotrimers, homotetramers, and homohexamers.

[0103] Proteins with the same EC number were paired according to UniProt information and Blast comparison was performed. The similarity score distribution when the number of subunits was the same and when the number of subunits was different was as follows: Figure 2 As shown, it can be observed that when the number of subunits is the same ( Figure 2 A), the overall similarity is high, while the similarity is low when the number of subunits is different, indicating that the alignment-based method is effective. However, it is worth noting that for the same number of subunits, the similarity is distributed in the range of 20-60, but for different numbers of subunits, there are also a considerable number of results in the range of 20-40 ( Figure 2 B) This suggests that relying solely on alignment-based methods to summarize prior knowledge about proteins may not always be reliable.

[0104] The dataset in this example consists of protein amino acid sequences and their corresponding subunit numbers obtained from UniProt. To handle variable sequence lengths, the amino acid sequences were embedded to convert them into vectors of the same dimension. The DeepSub model was trained using a GRU module and achieved an accuracy of 96% on the test set.

[0105] A confusion matrix is ​​a visual tool for comparing the differences between a model's predictions and the actual labels. Figure 3 B shows the confusion matrix of the DeepSub model on the test set. The numbers along the diagonal indicate the number of examples correctly predicted by the model, while the numbers outside the diagonal indicate the number of examples incorrectly predicted by the model. DeepSub performs well on some labels, such as "homodimer" and "monomer," but performs poorly on other labels, such as "homotetramer" and "homopentamer."

[0106] In addition, the performance of the DeepSub model was evaluated using common parameter evaluation metrics, including precision, recall, and F1 score, which provide a comprehensive understanding of the model's performance in each category. Figure 6As can be seen from the figure, the DeepSub model shows high precision and recall in most categories, with an F1 score exceeding 0.85 in most categories. This shows that the DeepSub model has high accuracy and reliability in predicting the number of subunits.

[0107] A BLAST comparison was performed between the test set and the training set, and the prediction accuracy was calculated at different similarity ranges in the test set. Figure 3 As shown in the figure, the DeepSub model performs well in the range of similarity above 80%, achieving a remarkable 95% accuracy at 100% similarity. In the range below 60%, the accuracy is 75%. This shows that the DeepSub model can effectively predict highly similar sequences and has strong generalization ability.

[0108] To evaluate DeepSub, four classic machine learning models and the QUEEN model were trained using the same data, resulting in six models being compared on the same test dataset. Figure 4 As can be seen in Figure A, DeepSub achieved the highest accuracy of 0.9745, followed by KNN (0.9598). Furthermore, DeepSub outperformed other methods in terms of mACC, mRecall, and mF1 scores. Among them, Random Forest (RF) achieved the highest mPrecision (0.9896), 0.0348 higher than DeepSub. Overall, DeepSub outperformed other machine learning models.

[0109] For machine learning models, generalization ability is a key indicator. The prediction accuracy of different methods was tested using homo-oligomer information collected from Complex Portal, such as Figure 4 As shown in Figure B, the six methods have varying degrees of performance on the test dataset. DeepSub maintains the highest prediction accuracy of 0.8660, demonstrating its excellent generalization ability. Notably, QUEEN performs poorly in both evaluations, scoring 0.7087 and 0.4131, respectively, significantly lower than the other methods.

[0110] AlphaFold2 is one of the most influential neural network-based protein structure prediction methods in recent years. It accurately predicts the three-dimensional structure of proteins. Using amino acid sequences as input, AlphaFold2 can predict the structure of a complex by specifying the number of subunits. However, in practical applications, researchers often struggle to determine the actual number of subunits in a target protein, making it difficult to achieve ideal results. The DeepSub model, on the other hand, can predict the number of subunits given a given amino acid sequence, thereby obtaining the true protein structure.

[0111] Taking P00448 as an example, the protein structures recorded in UniProt and AlphaFold protein structure databases are consistent, both indicating a monomeric structure. However, the crystal structure of this protein in the PDB database is a homodimer. Using the DeepSub model, the prediction result for this protein is a homodimer. Combined with the prediction of AlphaFold2, its structure is similar to that in the PDB, as shown in Figure 2. Figure 5 As shown in c.

[0112] This approach retrieved the sequences of monomeric and oligomeric proteins and their corresponding subunit numbers from the UniProt database. Using deep learning techniques, the DeepSub model was developed to predict the number of protein subunits. Through rigorous validation against Complex Portal data, DeepSub demonstrated high accuracy. Furthermore, DeepSub proved valuable in assisting the AlphaFold2 method in accurately predicting actual protein structures.

[0113] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any person skilled in the art who, within the technical scope disclosed by the present invention, makes equivalent substitutions or modifications based on the technical solutions and improved concepts of the present invention shall be covered by the scope of protection of the present invention.

Claims

1. A method for predicting the number of homo-oligomer subunits based on deep learning, characterized by: The following steps are involved: S1. Collect and extract protein sequences and subunit numbers from known databases to create a dataset; S2. Developed the DeepSub model based on recurrent neural network (RNN) and trained the model using the created dataset; S3. Input the homo-oligomer to be predicted into the trained model to obtain the prediction result of its subunit number; The construction of the DeepSub model in step S2 includes the following steps: S21. First, we obtain the semantic representation of protein sequences through the unified rational protein engineering (UniR ep) based on sequence-based deep representation learning; S22. Subsequently, we use an architecture based on a bidirectional gated recurrent unit (Bi-directional GRU) and an attention layer to learn biological questions related to the downstream subunit prediction task. S23. Finally, the output from the attention layer is connected to a fully connected output layer with a Softmax activation function corresponding to the number of categories to obtain the probability of each category.

2. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 1, characterized in that: The bidirectional GRU-RNN model is constructed in step S22, including the following steps: S221. Define the input layer: Based on the tensorflow framework, use the keras 'input' function to define the model input, 'inputs = Input(shape = (1,900), name = 'input')'; S222. Build a bidirectional GRU layer: Create a bidirectional GRU layer using the Bidreactional wrapper and the GRU layer, gru = Bidirectional(GRU(128, dropout = 0.5, return_sequences = True), name = "bi-gru")(inputs); S223, add attention mechanism: introduce attention layer, "attention=Attention(32, name='attention')(gru)"; S224. Apply DropOut: Add a Dropout layer after the attention layer, attention_with h_dropout = Dropout(0.1, name="attention_dropout")(attention)"; S225. Define the output layer: Add a fully connected layer as the output layer and use the softmax activation function, "output = Dense(num_class, activation = 'softmax', name = "dense")(attention_with_dropout)"; S226. Build and compile the model: Use the "Model" function to build the model, then compile the model, use the "fit" method to train the model, and then use the validation set and test set to evaluate the performance of the model; S227. Adjust the hyperparameters based on the model's performance on the validation set, setting the learning rate to lr = 0.01, the epoch to 200, and the batch_size to 1024.

3. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 2, characterized in that: The input matrix in the model is X∈R^ (1xN) , where N = 1900, X represents the UniRep representation of the input protein, and a bidirectional GRU layer with 128 hidden units is used. The bidirectional structure enables the model to capture sequence dependencies from both the forward and reverse directions. The process is expressed as: H gru =f gru ru(X); Among them H gru ∈R^ (1xH) is the output of the GRU layer, f gru is the function of the bidirectional GRU layer, and X is the UniRep representation of the input.

4. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 3, characterized in that: With a dropout rate of 20%, each hidden unit has a 20% probability of being randomly selected and temporarily removed from the network.

5. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 1, characterized in that: The dataset is divided into training set, validation set and test set in a ratio of 8:1:

1.

6. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 2, characterized in that: In step S226, when the performance on the validation set cannot be improved, the early stopping method is used to terminate the training.

7. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 1, characterized in that: DeepSub takes a protein sequence as input and outputs a label representing the number of subunits of the input protein. It uses a loss function that calculates the cross entropy between the true distribution and the predicted distribution. The specific formula is as follows: H(p,q)=-∑p(x)logq(x) Where p is the true distribution, q is the predicted distribution, Σ represents the sum over all categories, x is a specific category, p(x) is the probability of category x in the true distribution, and q(x) is the probability of category x in the predicted distribution.

8. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 1, characterized in that: It also includes evaluating model performance, specifically defining four commonly used metrics: mACC (macro average accuracy), mPR (macro average precision), mRecall (macro average recall), and mF1 (macro average F1 score). The specific formulas are as follows: Where N represents the total number of categories, and ACC_i, PPV_i, and Recall_i represent the accuracy, precision, and recall of the i-th category in a multi-class classification task, respectively.

9. The method for predicting the number of homo-oligomer subunits based on deep learning according to claim 8, characterized in that: When evaluating model performance, we first obtain information about protein complexes from the database, then filter out data related to heterologous proteins, and then extract data about homologous oligomers to create an independent dataset for evaluating the performance of DeepSub and traditional machine learning models.

Images