Nucleic acid binding protein identification method based on protein graph and protein language model

Abstract

The present invention discloses a method for identifying nucleic acid binding proteins based on protein graphs and protein language models, and belongs to the field of bioinformatics technology. The present invention combines high-dimensional sequence features extracted by the ProtT5 protein language model with three-dimensional structural information predicted by AlphaFold, so that the model can simultaneously capture rich feature information of sequence and structure, significantly improving the accuracy and reliability of NABPs prediction; the representation ability of sequence features is enhanced by both the GAT network and the BiLSTM network, so that the constructed prediction model can more comprehensively understand the important information in the protein sequence, thereby improving the prediction performance of NABPs; the features output by the GAT network and the BiLSTM network are fused, and NABPs are classified by MLP, so that the prediction model can fully utilize feature information from different sources, thereby improving the performance and robustness of NABPs prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for identifying nucleic acid-binding proteins based on protein graphs and protein language models. Background Art

[0002] Nucleic acid-binding proteins (NABPs) play a variety of important roles within cells. They interact with nucleic acids such as DNA and RNA to regulate gene expression, maintain genome stability, and participate in biological processes such as signal transduction. These proteins are crucial for the normal function of cells and homeostasis. Through the precise identification and in-depth study of these proteins, we can decipher their cellular functions and the complex regulatory networks in which they are located. This exploration not only reveals the core principles of cell biology and molecular biology, but also enhances our understanding of the complex mechanisms of gene regulation. Therefore, the efficient and accurate identification of NABPs is of great significance for understanding complex biological responses in organisms and developing new diagnostic and therapeutic methods.

[0003] In the early stages, structural alignment and statistical potential were used to identify NABPs. However, these methods were hampered by the limited availability of three-dimensional (3D) protein structures. With the rise of artificial intelligence, an increasing number of machine learning and deep learning algorithms have been used to construct NABP prediction models based on sequence features. Some researchers have developed the TriPepSVM model based on sequence features and support vector machines. Others have developed the sequence-based StackDPPred using a stacking algorithm. Still others have developed the two-layer model iDRBP-PPCT using a random forest algorithm based on a position-specific scoring matrix (PSSM) and a position-specific frequency matrix (PSFM). The first layer distinguishes NABPs from non-NABPs, while the second layer distinguishes RNA-binding proteins (RBPs), DNA-binding proteins (DBPs), and DNA / RNA-binding proteins (DRBPs). Furthermore, deep learning algorithms have also been widely used to identify NABPs. For example, iDRBP_MMC introduced a multi-label learning model based on convolutional neural networks to address the cross-prediction problem, thereby improving the prediction performance of DNA-binding and RNA-binding proteins. The iDRBP-ECHF model proposed in the paper "iDRBP-ECHF: Identifying DNA- and RNA-binding proteins based onextensible cubic hybrid framework" is built based on five sequence features and a stacked ensemble learning algorithm, using two deep learning algorithms and two machine learning algorithms to establish the base model. The iDRPro-SC proposed in the paper "iDRPro-SC: Identifying DNA-binding proteins and RNA-binding proteins based on subfunction classifiers" is also built using a stacked ensemble learning algorithm based on PSSM features, with a BiLSTM-based base model used to distinguish the subfunctions of NABPs.

[0004] The above-mentioned prior art has the following problems:

[0005] Insufficient Utilization of Structural Information: Due to the scarcity of 3D structural information for NABPs, existing models rarely utilize this structural information. However, structural information plays a crucial role in understanding protein-nucleic acid interactions. Current sequence-based models, such as TriPepSVM, StackDPPred, and iDRBP_MMC, primarily rely on sequence features for prediction and fail to fully utilize 3D structural information, which limits the models' predictive power and accuracy.

[0006] Low data coverage: As described in iDRPro-SC, proteins with available structural information only account for 0.36% of all proteins in UniProtKB. This low coverage limits the wide application of models. Many existing models (such as iDRBP-PPCT and iDRPro-SC) rely primarily on a small amount of 3D structural data, which makes it difficult for these models to generalize to a wider dataset.

[0007] Limited Prediction Performance: Due to the inability to fully utilize 3D structural information, the prediction performance of existing models is limited in some cases, especially when dealing with structurally complex proteins. For example, models such as StackDPPred, which is based on sequence features, and iDRBP-ECHF and iDRPro-SC, which are based on stacked ensemble learning, are unable to fully capture the spatial information of proteins when dealing with complex structures, thus affecting the accuracy and performance of predictions.

[0008] The above problems need to be solved urgently. To this end, the present invention proposes a method for identifying nucleic acid-binding proteins based on protein graphs and protein language models. Summary of the Invention

[0009] The technical problem to be solved by the present invention is how to solve the problems existing in the prior art, such as insufficient utilization of structural information, low data coverage, and limited prediction performance. A method for identifying nucleic acid-binding proteins based on protein graphs and protein language models is provided.

[0010] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0011] S1: Dataset processing

[0012] Retrieve protein sequence data from the UniProt database and obtain the predicted 3D structure information of the corresponding proteins from the AlphaFold database. Remove proteins whose sequence similarity exceeds the set value and proteins for which structural information cannot be obtained to obtain a training dataset, and then select a test dataset;

[0013] S2: Feature Extraction

[0014] Use the ProtT5 model to generate embedded features of protein sequences, and represent proteins as graphs based on the predicted 3D structure information to obtain protein graphs that capture their structural and sequence features;

[0015] S3: Model Building

[0016] A prediction model was established using the GAT network, BiLSTM network, and MLP. In the prediction model, the GAT network and BiLSTM network were used to enhance feature representation. The output features of the GAT network and BiLSTM network were then connected, and NABPs were classified and identified using MLP.

[0017] S4: Model training

[0018] Using the training data set to train the prediction model to obtain a trained prediction model;

[0019] S5: NABPs identification

[0020] The trained prediction model is used to perform NABPs recognition on samples in the test dataset to obtain the recognition results.

[0021] Furthermore, in step S1, the specific processing process is as follows:

[0022] S11: Search for protein sequences with DNA-binding and RNA-binding functions from the UniProt database and retrieve their corresponding annotation information;

[0023] S12: The predicted 3D structure of the corresponding protein sequence was retrieved from the AlphaFold database;

[0024] S13: Use the BLASTClust tool to remove protein sequences whose sequence similarity exceeds the set value and protein sequences that lack predicted 3D structure information to obtain a training dataset;

[0025] S14: Datasets PDB255 and TEST474 are selected as test datasets.

[0026] Furthermore, in step S2, the specific processing process is as follows:

[0027] S21: Use the ProtT5 model to perform protein language model embedding, extract the embedding features of each residue from the protein sequence, and generate a feature matrix of size L×1024, where L is the sequence length;

[0028] S22: Search the AlphaFold database to obtain the predicted 3D structure of the protein, and represent each protein as a graph with each residue as a node. The edges between nodes are determined based on their Euclidean distance, thus completing the construction of the protein graph.

[0029] Furthermore, in step S22, each protein is represented as an undirected graph G = (V, H, E, A), where V is the set of nodes in the graph G, V i ∈V,V i represents the i-th residual node, V is the set of all residual nodes in the protein, H is the set of all node features, h∈H |V|×d , |V| is the total number of nodes, d is the dimension of node features, E is the set of all edges in graph G, e ij ∈E represents the edge between node i and node j, A is the adjacency matrix of graph G, if the Euclidean distance between any two nodes is less than or equal to If , the two nodes are considered to be connected, otherwise they are considered to be disconnected.

[0030] Furthermore, in step S3, the GAT network uses an attention mechanism to obtain different weights of adjacent nodes, and then aggregates the information of weighted adjacent nodes to update the node embedding; in the prediction model, multiple stacked GAT layers are used to implement high-level representation of nodes. Each GAT layer dynamically adjusts node features by learning the relationship between nodes and adjacent nodes. In the last layer of the entire GAT network, the features of all nodes are averaged and pooled to obtain the feature representation of the entire protein, namely the graph embedding representation.

[0031] Furthermore, in step S3, the BiLSTM network is able to capture the relationship within the sequence from two directions to maintain the integrity of the sequence semantics. The embedded features generated by the ProtT5 model are processed within the LSTM, and the forget gate determines which information to discard from the cell state, while the new information is merged through the input gate and cell state update. The new information includes the embedded features and the data of the previous time step.

[0032] Furthermore, in step S3, the processing formula of MLP is as follows:

[0033]

[0034] Among them, G ′ represents the output of the GAT network, h biRepresents the output of the BiLSTM network, concat represents the Concat function, which is used to connect the output features of the GAT network and the BiLSTM network, and softmax represents the softmax function, which is used to map the result to the range between 0 and 1. Indicates the prediction result.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1. Combining the high-dimensional sequence features extracted by the ProtT5 protein language model with the three-dimensional structural information predicted by AlphaFold enables the model to simultaneously capture rich feature information of sequence and structure, significantly improving the accuracy and reliability of NABPs prediction;

[0037] 2. By representing proteins as graph structures and introducing a graph attention neural network to process these graphs, the GAT network assigns different weights to each node through the attention mechanism, integrating the structural information of the nodes. At the same time, the BiLSTM network is used to process the sequence embedding features extracted from the ProtT5 model to capture long-range dependencies in the sequence. Both the GAT network and the BiLSTM network enhance the representation ability of sequence features, enabling the constructed prediction model to more comprehensively understand the important information in the protein sequence, thereby improving the prediction performance of NABPs.

[0038] 3. The features output by the GAT network and the BiLSTM network are fused, and NABPs are classified through MLP, so that the prediction model can fully utilize the feature information from different sources and improve the performance and robustness of NABPs prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic flow chart of the method for identifying nucleic acid-binding proteins based on protein graphs and protein language models of the present invention;

[0040] Figure 2 This is a comparison of the AUROC values of the GraphNABP model in the embodiment of the present invention and other methods on PDB255, where A: DNA binding protein, B: RNA binding protein;

[0041] Figure 3 is the ROC curve of DNA binding proteins in the PDB255 dataset in the embodiment of the present invention;

[0042] Figure 4 is the ROC curve of RNA binding proteins in the PDB255 dataset in the embodiment of the present invention;

[0043] Figure 5is the ROC curve of DNA binding proteins in the TEST474 dataset in the embodiment of the present invention;

[0044] Figure 6 is the ROC curve of RNA binding proteins in the TEST474 dataset in the embodiment of the present invention;

[0045] Figure 7 is a graph showing the performance of different methods in predicting known nucleic acid binding proteins in the tomato genome according to the embodiments of the present invention;

[0046] Figure 8 : This is a case study result diagram of Q7AKE4 in an embodiment of the present invention, wherein A: a heat map of the attention weight of the Q7AKE4 sequence, B: an adjacency matrix diagram of the Q7AKE4 active pocket predicted by CavityPlus, and C: a visualization diagram of the Q7AKE4 structure with residues E64, A94, G137, L142, and L150. DETAILED DESCRIPTION

[0047] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment provides a technical solution: a method for identifying nucleic acid-binding proteins based on protein graphs and protein language models, which mainly includes the following steps:

[0050] Follow these steps:

[0051] 1. Dataset processing

[0052] Retrieve protein sequence data from the UniProt database and obtain the predicted 3D structure information of the corresponding proteins from the AlphaFold database. Remove proteins whose sequence similarity exceeds the set value and proteins for which structural information cannot be obtained to obtain a training dataset, and then select a test dataset;

[0053] 2. Feature Extraction

[0054] Use the ProtT5 model to generate embedded features of protein sequences, and represent proteins as graphs based on the predicted 3D structure information to obtain protein graphs that capture their structural and sequence features;

[0055] 3. Model Building

[0056] A prediction model was established using the GAT network, BiLSTM network, and MLP. In the prediction model, the GAT network and BiLSTM network were used to enhance feature representation. The output features of the GAT network and BiLSTM network were then connected, and NABPs were classified and identified using MLP.

[0057] 4. Model evaluation and prediction results analysis

[0058] Model selection was performed using 5-fold cross-validation. Evaluation metrics included precision, recall, F1 score, MCC, and AUROC. Ablation experiments were performed to assess the contribution of different modules to model performance. The model was validated on the PDB255 and TEST474 test sets, and NABPs prediction was performed on the tomato genome.

[0059] Step 1: Dataset processing

[0060] The benchmark dataset was obtained from the UniProt database and filtered based on protein function. Specifically, proteins with DNA- and RNA-binding functions were searched for, and their corresponding annotation information (UniProt ID, functional description, protein name, and sequence information) was retrieved. This annotation information was used to ensure data accuracy and provide detailed context for subsequent processing. To ensure data quality and reliability, proteins with sequence similarity exceeding 35% were excluded using the BLASTClust tool. Following this processing pipeline, the final benchmark dataset included 10,188 non-NABPs and 11,875 NABPs. These NABPs were further classified into 7,594 DNA-binding proteins (DBPs), 3,948 RNA-binding proteins (RBPs), and 333 proteins that bind both DNA and RNA (DRBPs). In constructing the GraphNABP model, we utilized protein structure information predicted by AlphaFold. By querying the UniProt ID, we could retrieve the predicted 3D structure of the corresponding protein from the AlphaFold database. It is worth noting that due to the lack of predicted 3D structures of several proteins, we removed these proteins from the dataset. After the above processing, the training dataset contains 7070 DBPs, 3591 RBPs, and 9732 non-NABPs.

[0061] To evaluate and fairly compare our model with other methods, we used two widely used test datasets for NABP prediction tasks: PDB255 and TEST474. PDB255 contains 93 DBPs, 70 RBPs, and 92 non-NABPs. TEST474 includes 183 DBPs, 76 RBPs, and 223 non-NABPs. The predicted structures of the proteins in the test datasets were also obtained from AlphaFold.

[0062] Step 2: Feature Extraction

[0063] The feature extraction method of the present invention comprehensively utilizes protein language model embedding and three-dimensional (3D) structural information. First, the ProtT5 model is used to embed the protein language model, and the embedding features of each residue are extracted from the protein sequence to generate a feature matrix of size L×1024, where L is the sequence length. These embedded features capture the semantics and contextual information of the residues. Next, a protein structure search is performed through the AlphaFold database to obtain the predicted 3D structure of the protein, and each protein is represented as a graph with each residue as a node. The edges between nodes are determined according to their Euclidean distance (less than or equal to 14 angstroms), thereby completing the construction of the protein graph. In this way, GraphNABP can comprehensively capture and integrate the sequence and structural information of proteins, providing rich feature representations for subsequent NABPs predictions.

[0064] 2.1 Protein Language Model Embedding

[0065] Protein language models have been widely used in current bioinformatics research. These models utilize the Transformer architecture from the field of natural language processing and are adapted and optimized for protein sequence analysis to better understand the semantic information of protein sequences. Among them, the ProtT5 model is widely used, which is pre-trained on a large number of unlabeled protein sequences. Previous studies have shown that protein language models have excellent performance in predicting hotspot residues on the protein-DNA binding interface and may have great potential in predicting nucleic acid-binding proteins. In the present invention, we use ProtT5 to extract the embedded features of each residue, which are then used for model construction. Each protein sequence is encoded as a feature matrix of size L×1024, where L is the sequence length.

[0066] 2.2 Protein structure retrieval and protein graph construction

[0067] The protein structures in our dataset are from the AlphaFold database, and all the atomic coordinates of the proteins are from the PDB files provided by the AlphaFold database. Based on the predicted 3D structure, each protein is represented as a graph with each residue as a node. The distance matrix is calculated based on the coordinates of the alpha carbon atom of each residue. Formally, each protein is represented as an undirected graph G = (V, H, E, A), where V is the set of nodes in the graph G and V is the distance matrix of each residue. i ∈V(V i represents the i-th residual node, V is the set of all residual nodes in the protein), H is the set of all node features, h∈H |V|×d , where |V| is the total number of nodes and d is the dimension of the node features. E is the set of all edges in the protein graph, where e ij ∈E represents the edge between node i and node j. A is the adjacency matrix of the graph. If the Euclidean distance between any two nodes is less than or equal to 14 angstroms, The two nodes are considered to be connected, and the judgment method is as follows:

[0068]

[0069] Among them, 1 means that the two nodes are connected, and 0 means that the two nodes are disconnected.

[0070] Step 3: Model building

[0071] The model building part of the present invention enhances feature representation by combining a graph attention (GAT) network and a bidirectional long short-term memory (BiLSTM) network. First, the GAT network is used to process the protein graph, and the attention mechanism is used to assign different weights to each node, thereby integrating the structural information of the node. Then, the BiLSTM network is used to process the sequence embedding features extracted from the ProtT5 model to capture the long-distance dependencies in the sequence. Finally, the output features of the GAT network and the BiLSTM network are connected, and NABPs classification is performed through a multi-layer perceptron (MLP) to achieve efficient prediction. In this way, the GraphNABP model (prediction model) can make full use of the sequence and structural information of the protein to improve the accuracy and reliability of the prediction.

[0072] 3.1 Graph Attention (GAT) Network

[0073] The GAT network uses an attention mechanism to capture the different weights of adjacent nodes, then aggregates the information of these weighted adjacent nodes to update the node embedding. In our prediction model (GraphNABP model), multiple stacked GAT layers are used to achieve a high-level representation of nodes. Each GAT layer dynamically adjusts node features by learning the relationship between a node and its adjacent nodes. In the final layer of the entire GAT network, the features of all nodes are average pooled to obtain a feature representation of the entire protein. The principle of the GAT network can be described by the following formula:

[0074]

[0075] In the first formula above, h i and h j denote the embedding of node i and its adjacent node j, respectively, where ( is the set of adjacent nodes of node i). W represents the learnable parameter matrix of different layers, a is the activation function, e ij is the attention coefficient. The third formula is an extension of the second formula. The softmax operation is performed on the attention coefficient of each adjacent node to obtain the attention coefficient between node i and its adjacent nodes. Where ∥ represents the concatenation operation of the feature vector. In the fourth formula, the feature vector of node i is updated using the obtained attention weight. In the fifth formula, all updated nodes are average pooled to obtain the updated feature vector of the entire graph, where Represents the total number of nodes in the graph.

[0076] 3.2 Bidirectional Long Short-Term Memory (BiLSTM) Network

[0077] A bidirectional long short-term memory (BiLSTM) network is used to enhance the representation of sequence information generated by the protein language model. The BiLSTM network can capture relationships within the sequence from both directions to maintain the integrity of the sequence semantics. Specifically, the embedding representation generated by the protein language model is processed within the LSTM. The "forget gate" determines which information is discarded from the cell state, while new information, including the embedding representation and data from the previous time step, is incorporated through the "input gate" and "cell state update."

[0078] For the forward LSTM, given input X = (x1, x2, ..., x T ), where T is the sequence length. The calculation formula is as follows:

[0079] i t =σ(W ix x t +W ih h t-1 +b i )

[0080] f t =σ(W fx x t +w fh h t-1 +b f )

[0081] o t =σ(W ox x t +W oh h t-1 +b o )

[0082] C t =f t C t-1 +i t tanth(W cx x t +W ch h t-1 +b c )

[0083] h t =o t tanh(C t )

[0084] Among them, σ represents the sigmoid activation function, i t 、f t 、C t and h t They represent the input gate, forget gate, cell state, and output gate at time step t, respectively. W and b are the weight matrix and bias vector, respectively.

[0085] For the backward LSTM, the input is represented as X rev =(x T ,x T-1 ,…,x1). In the LSTM module, the final backward hidden state is represented as The final output is obtained by concatenating the forward and backward hidden states

[0086] 3.3 Multilayer Perceptron (MLP)

[0087] The outputs of GAT and BiLSTM are connected, and then a multi-layer perceptron (MLP) is used to classify NABPs. The specific formula is as follows:

[0088]

[0089] The Concat function concatenates the graph embedding representation with the hidden state of the BiLSTM and passes it to the multi-layer perceptron for classification. The softmax function maps the result to a range between 0 and 1. Represents the prediction result. (See Table 1 for the hyperparameters used for prediction model training)

[0090] Table 1 Hyperparameters used in GraphNABP model training

[0091]

[0092]

[0093] Step 4: Model evaluation and prediction results analysis

[0094] 4.1 Cross-validation and evaluation metrics

[0095] During model training, we use 5-fold cross validation for model selection. To measure the performance of the model, we use the following metrics:

[0096]

[0097]

[0098] Where TP is true positive, FP is false positive, TN is true negative, and FN is false negative.

[0099] 4.2 Ablation Experiment

[0100] To evaluate the contributions of different modules to the GraphNABP model, we performed ablation studies on the training dataset using 5-fold cross-validation. Given the significant structural and functional differences between DNA-binding and RNA-binding proteins, we conducted ablation experiments separately for these two tasks. Based on the embeddings obtained from the ProtT5 model, our model uses GAT and BiLSTM networks in parallel to enhance feature representations. Therefore, each of the two networks was removed for ablation analysis. As shown in Table 2, the cross-validation performance of both the BiLSTM-only and GAT-only models is inferior to that of GraphNABP in terms of AUROC values, which demonstrates the effectiveness of the proposed framework. In addition, we note that the prediction performance of RBPs is slightly worse than that of DBPs, which may be due to the less training data available for RBPs compared to DBPs.

[0101] Table 2 Ablation results of GraphNABP model

[0102]

[0103] 4.3 Analysis of prediction results

[0104] 4.3.1 Prediction results for PDB255

[0105] To evaluate the generalization performance of the GraphNABP model, we conducted a fair performance comparison with other state-of-the-art methods on the PDB255 dataset, which is widely used to evaluate NABPs prediction models. It should be noted that most nucleic acid binding protein prediction models lack public source code, so we collected prediction performance reports of other models from the literature. Figure 2-4 As shown, on the PDB255 dataset, GraphNABP achieved an AUROC value of 0.8889 in the DNA-binding protein (DBP) prediction task, outperforming the second-best method, iDRBP-ECHF, by 7.29 percentage points and significantly exceeding the other three methods. In the RNA-binding protein (RBP) prediction task, the GraphNABP model achieved an AUROC value of 0.8297, slightly higher than iDRBP-ECHF (0.828). These results demonstrate that the GraphNABP model performs well in predicting NABPs.

[0106] 4.3.2 Prediction Performance on TEST474

[0107] Compared to PDB255, TEST474 contains a larger dataset with an unbalanced distribution of positive and negative samples, reflecting the true distribution of NABPs. As shown in Table 3, for DBP prediction, we compared the prediction performance of GraphNABP with eight state-of-the-art methods (DNAbinder, stackDPPred, DPP-PseAAC, DeepDRBP-2L, iDRBP_MMC, IDRBP-PPCT, iDRBP-ECHF, and iDRPro-SC). The GraphNABP model achieved an AUROC value of 0.9732, nearly 2.3 percentage points higher than the second-best method, iDRBP-ECHF (AUROC value of 0.950). Furthermore, for metrics such as ACC, PRE, REC, F1, and MCC, the GraphNABP model achieved values of 0.9367, 0.9408, 9344, 0.9194, and 0.8676, respectively, significantly outperforming other prediction methods. This significant performance improvement highlights the effectiveness of our model in both model design and feature utilization. Furthermore, for RBP prediction, we performed a detailed comparison with 11 state-of-the-art methods, as shown in Table 3. On the TEST474 dataset, we achieved an AUROC of 0.9201, very close to that of iDRBP-ECHF (AUROC: 0.925) and outperforming other methods reporting AUROC values. Furthermore, the GraphNABP model achieved an MCC of 0.7409, 2.09 percentage points higher than the second-best model (0.72). Combined with the results for DNA-binding protein prediction on TEST474 and nucleic acid-binding protein prediction on PDB255, the GraphNABP model demonstrated superior ability in predicting nucleic acid-binding proteins. Figure 5-6 The ROC curve of the GraphNABP model on the TEST474 dataset is given.

[0108] Table 3 Comparison of GraphNABP model with other methods on TEST474

[0109]

[0110]

[0111] 4.3.3 Prediction of NABPs in the Tomato Genome

[0112] In order to verify the practical application of this prediction model, we predicted NABPs of the tomato genome. Specifically, GO terms were used for prediction, and 577 known DNA-binding proteins and 171 RNA-binding proteins were obtained from the tomato genome dataset (ITAG4.1). The GraphNABP model identified 557 NBPs (96.53%) and 167 RBPs (97.66%) from the known NABPs. To compare with other methods, we used iDRBP-ECHF, iDRPro-SC and iDRBP_MMC available web servers to predict the corresponding proteins. Figure 7 As shown in the results, the GraphNABP model significantly outperformed other prediction methods, demonstrating its versatility and superior performance in predicting NABPs in real-world scenarios.

[0113] In this paper, the GraphNABP model is proposed, a nucleic acid binding protein (NABPs) prediction framework that combines a graph attention neural network (GAT) and a protein language model. The GraphNABP model significantly improves the prediction performance of NABPs by integrating protein sequence and predicted three-dimensional structure information. Specifically, sequence embeddings are first obtained from the ProtT5 protein language model, and the 3D structure of the protein is predicted by AlphaFold. Then, the feature representation is enhanced using a graph attention network (GAT) and a bidirectional long short-term memory network (BiLSTM), and finally NABPs are predicted through a fully connected layer. To the best of our knowledge, this is the first time that AlphaFold and a protein language model have been combined for the prediction of NABPs.

[0114] Most existing methods predict NABPs solely based on protein sequence information, ignoring three-dimensional structural information. This limits the model's predictive power and accuracy. GraphNABP, on the other hand, achieves more efficient feature extraction and accurate prediction by comprehensively leveraging sequence and structural information. The GraphNABP model offers the following advantages: First, the GraphNABP model represents proteins as a graph structure by combining high-dimensional sequence features extracted by the ProtT5 model with the three-dimensional protein structure information predicted by AlphaFold. This approach enables the model to capture more diverse and rich feature information, significantly improving the accuracy and reliability of NABP prediction. Second, GraphNABP employs a GAT network to process the protein graph structure, leveraging an attention mechanism to assign different weights to each node, thereby more effectively integrating node structural information. The GAT network better captures protein topology and enhances the accuracy of feature representation. Compared to traditional methods, the GraphNABP model performs better when handling complex protein structures and can more accurately predict NABPs. Furthermore, a bidirectional long short-term memory (BiLSTM) network is used to process the sequence embedding features extracted from the ProtT5 model, capturing long-range dependencies within the sequence. The BiLSTM network enhances the representation of sequence features, enabling the model to more comprehensively understand important information in protein sequences, thereby improving the prediction performance of NABPs. By combining the GAT and BiLSTM neural networks, the GraphNABP model fully leverages the complementarity between protein structure and sequence information, improving the performance and robustness of NABP prediction. Experimental results demonstrate that the GraphNABP model outperforms existing state-of-the-art methods on multiple independent test sets and the tomato genome. Furthermore, the GraphNABP model enhances model transparency and credibility through interpretability analysis. Visualizing the attention weight matrix of the GAT network allows us to understand which residues play a key role in NABP prediction. These advantages and effects make the GraphNABP model significantly innovative and practical compared to existing technologies.

[0115] Example 2

[0116] In this example, we further discussed the effects of embeddings obtained from different protein language models. To this end, we compared the performance of embeddings obtained from three widely used protein language models (ProtBert, esb-1b, and ProtT5). As shown in Table 4, from the cross-validation performance, ESM-1b and ProtT5 have the highest AUROC of 0.9715 for DBPs prediction, and ProtT5 has the highest AUROC of 0.9630 for RBPs prediction. Considering the performance of the two prediction tasks, we chose the ProtT5 model as the model for embedding feature extraction.

[0117] Table 4 5-fold cross-validation performance based on different protein language models

[0118]

[0119] Interpretability analysis: The GraphNABP model adopts a graph attention (GAT) network to capture the topological structural information of proteins. In this paper, we visualized the attention weights learned by the GAT network and analyzed the potential factors that contribute to the model's accurate prediction of NABPs. Specifically, we extracted the attention weight matrix generated by the last GAT layer of the GraphNABP model and summed the weights of each sequential position to represent the contribution of each residue to the prediction of NABPs. By quantifying the contribution of residues, the interpretability of the GraphNABP model is significantly enhanced, providing insights into the interaction between nucleic acids and proteins. The following is a case study of NABP Q7AKE4 from TEST474. Q7AKE4 is a transcription factor that is essential for the formation of aerial hyphae on rich culture medium and binds to the ramC promoter. The attention weights of the Q7AKE4 sequence are visualized as follows. Figure 8 As shown in Figure A, the attention weights of some residues are significantly higher than those of other positions, indicating that these high-weight residues play a crucial role in the entire sequence. These residues (E64, A94, G137, L142, and L150) are mapped to the three-dimensional structure of the protein, as shown in Figure 3. Figure 8 As shown in C in , these residues are located in a potential protein pocket. To confirm this, we used CavityPlus to predict a high-confidence protein pocket (e.g. Figure 8 The GraphNABP model successfully identified a potential nucleic acid binding region, further validating the model's robust predictive performance.

[0120] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for identifying nucleic acid-binding proteins based on protein graphs and protein language models, characterized in that: The following steps are involved: S1: Dataset processing Retrieve protein sequence data from the UniProt database and obtain the predicted 3D structure information of the corresponding proteins from the AlphaFold database. Remove proteins whose sequence similarity exceeds the set value and proteins for which structural information cannot be obtained to obtain a training dataset, and then select a test dataset; S2: Feature Extraction Use the ProtT5 model to generate embedded features of protein sequences, and represent proteins as graphs based on the predicted 3D structure information to obtain protein graphs that capture their structural and sequence features; S3: Model Building A prediction model was established using the GAT network, BiLSTM network, and MLP. In the prediction model, the GAT network and BiLSTM network were used to enhance feature representation. The output features of the GAT network and BiLSTM network were then connected, and NABPs were classified and identified using MLP. S4: Model training Using the training data set to train the prediction model to obtain a trained prediction model; S5: NABPs identification The trained prediction model is used to perform NABPs recognition on samples in the test dataset to obtain the recognition results.

2. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 1, characterized in that: In step S1, the specific processing process is as follows: S11: Search for protein sequences with DNA-binding and RNA-binding functions from the UniProt database and retrieve their corresponding annotation information; S12: The predicted 3D structure of the corresponding protein sequence was retrieved from the AlphaFold database; S13: Use the BLASTClust tool to remove protein sequences whose sequence similarity exceeds the set value and protein sequences that lack predicted 3D structure information to obtain a training dataset; S14: Datasets PDB255 and TEST474 are selected as test datasets.

3. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 2, characterized in that: In step S2, the specific processing process is as follows: S21: Use the ProtT5 model to perform protein language model embedding, extract the embedding features of each residue from the protein sequence, and generate a feature matrix of size L×1024, where L is the sequence length; S22: Search the AlphaFold database to obtain the predicted 3D structure of the protein, and represent each protein as a graph with each residue as a node. The edges between nodes are determined based on their Euclidean distance, thus completing the construction of the protein graph.

4. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 3, characterized in that: In step S22, each protein is represented as an undirected graph G = (V, H, E, A), where V is the set of nodes in the graph G, V i ∈V,V i represents the i-th residual node, V is the set of all residual nodes in the protein, H is the set of all node features, h∈H |V|×d , |V| is the total number of nodes, d is the dimension of node features, E is the set of all edges in graph G, e ij ∈E represents the edge between node i and node j, A is the adjacency matrix of graph G, if the Euclidean distance between any two nodes is less than or equal to If , the two nodes are considered to be connected, otherwise they are considered to be disconnected.

5. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 4, characterized in that: In step S3, the GAT network uses an attention mechanism to obtain different weights of adjacent nodes, and then aggregates the information of weighted adjacent nodes to update the node embedding; in the prediction model, multiple stacked GAT layers are used to implement high-level representation of nodes. Each GAT layer dynamically adjusts node features by learning the relationship between nodes and adjacent nodes. In the last layer of the entire GAT network, the features of all nodes are averaged and pooled to obtain the feature representation of the entire protein, namely the graph embedding representation.

6. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 5, characterized in that: In step S3, the BiLSTM network is able to capture the relationship within the sequence from two directions to maintain the integrity of the sequence semantics. The embedded features generated by the ProtT5 model are processed within the LSTM, and the forget gate determines which information to discard from the cell state, while the new information is merged through the input gate and cell state update. The new information includes the embedded features and the data of the previous time step.

7. The method for identifying nucleic acid binding proteins based on protein graphs and protein language models according to claim 6, characterized in that: In step S3, the processing formula of MLP is as follows: Among them, G ′ represents the output of the GAT network, h bi Represents the output of the BiLSTM network, concat represents the Concat function, which is used to connect the output features of the GAT network and the BiLSTM network, and softmax represents the softmax function, which is used to map the result to the range between 0 and 1. Indicates the prediction result.

Images