Ferroheme binding site prediction method, system, medium and equipment

Abstract

The invention discloses a ferroheme binding site prediction method, system, medium and equipment, and relates to the field of site prediction, and the method comprises the following steps: obtaining protein sequence and structure information; extracting a plurality of physical and chemical attribute features, amino acid embedding features and protein secondary structure features according to the protein sequence and structure information; carrying out fusion operation on the plurality of physicochemical attribute characteristics, determining global evolution characteristics, carrying out time sequence characteristic analysis and fusion on the global evolution characteristics, determining time sequence characteristics, and carrying out nonlinear combination and mapping on the time sequence characteristics to obtain the probability that the unknown protein is the ferroheme binding protein; and splicing the probability that the unknown protein is the ferroheme binding protein, the amino acid embedding feature and the protein secondary structure feature to obtain a spliced feature, and carrying out weighted summation on different parts of the spliced feature to obtain a prediction result of the ferroheme binding site.

Description

Technical Field

[0001] This invention relates to the field of site prediction technology, and in particular to a method, system, medium, and device for predicting heme binding sites. Background Technology

[0002] Proteins are biological macromolecules composed of many amino acids linked together by peptide bonds. As is well known, proteins are the main executors of various life activities. In fact, proteins are widely involved in a variety of important life processes, such as cell structure, signal transduction, immune defense, and substance transport.

[0003] Heme refers to a heme molecule containing ferrous ions. It has a porphyrin ring structure and is an important prosthetic group in many proteins. Heme plays a crucial role in oxygen transport and redox reactions within living organisms. Biologists call proteins that can specifically bind to heme molecules heme-binding proteins. These proteins typically have one or more heme-binding sites, usually composed of specific amino acid residues such as histidine, cysteine, and methionine. Through non-covalent bonding, hydrophobic pockets, and coordinate bonds, they stably insert heme into proteins.

[0004] Existing computation-based prediction methods only focus on the amino acid information around the heme binding site, ignoring the description of the entire protein structure. Some methods use protein big language models to build deep learning models to predict binding sites, but protein big language models are trained on general protein big data, so the prediction accuracy for specific functional sites of proteins may be poor. Summary of the Invention

[0005] This invention provides a method, system, medium, and device for predicting heme binding sites, to solve the aforementioned problems in the prior art, namely, how to improve the accuracy of heme binding site prediction in the prior art. This invention provides a method for predicting heme binding sites, which includes: To obtain the protein sequence and structural information of unknown proteins; Based on protein sequence and structural information, multiple physicochemical property features, amino acid embedding features, and protein secondary structure features are extracted. Multiple physicochemical properties are fused to determine global evolutionary features. Forward and reverse temporal features of the global evolutionary features are analyzed and fused to determine temporal features. By performing nonlinear combination and mapping of temporal features, the probability that the unknown protein is a heme-binding protein is obtained. The probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features are spliced ​​together to obtain spliced ​​features. By introducing a multi-head attention mechanism, different parts of the spliced ​​features are weighted and summed to determine the weighted summed features. The weighted summed features are then nonlinearly combined and mapped to obtain the heme binding site prediction results. The heme binding site prediction results represent the probability that the unknown protein is a heme binding site.

[0006] Optionally, the prediction results of the heme binding sites are obtained through a hierarchical network, which specifically includes: A pre-trained model includes a Bi-LSTM (Bidirectional Long Short-Term Memory) neural network module and a first fully connected neural network FCNN (Finite-CNN) module. The Bi-LSTM module performs forward and reverse temporal feature analysis and fusion on global evolutionary features to determine temporal features. The FCNN module performs nonlinear combination and mapping on the temporal features to obtain the probability that an unknown protein is a heme-binding protein. The prediction model includes a multi-head attention mechanism module and a second fully connected neural network (FCNN) module. The multi-head attention mechanism module is used to concatenate the probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features to obtain concatenated features. By introducing a multi-head attention mechanism, different parts of the concatenated features are weighted and summed to determine the weighted summed features. The second fully connected neural network (FCNN) module is used to perform nonlinear combination and mapping on the weighted summed features to obtain the heme binding site prediction results.

[0007] Optionally, the first fully connected neural network FCNN module specifically includes three linear transformation layers; wherein the activation function between each linear transformation layer is the ReLU function, and the activation function of the final output is the sigmoid function.

[0008] Optionally, the plurality of physicochemical properties specifically include: Hydrophilic physicochemical properties, hydrophobic physicochemical properties, polar physicochemical properties, positively charged physicochemical properties, negatively charged physicochemical properties, and aromatic physicochemical properties.

[0009] This invention provides a heme binding site prediction system to realize the above-mentioned heme binding site prediction method.

[0010] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting heme binding sites.

[0011] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting heme binding sites.

[0012] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention provides a method for predicting heme binding sites. This method captures contextual dependencies in protein sequences by analyzing and fusing forward and reverse temporal features of global evolutionary characteristics to obtain temporal features. Furthermore, by performing cross-dimensional fusion and nonlinear transformation of the temporal features, the complex relationship between evolutionary conservation and physicochemical properties is further integrated. In addition, this invention introduces a multi-head attention mechanism, where each attention head performs a weighted summation on different parts of the input sequence, thereby capturing dependencies between different positions in the sequence. By having different attention heads focus on different local or global structural information of the sequence, the sequence can be analyzed from multiple perspectives, leading to a more comprehensive understanding of sequence features and effectively improving the accuracy of heme binding site prediction. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0014] Figure 1 A flowchart illustrating a method for predicting heme binding sites provided in an embodiment of the present invention; Figure 2 This is an overall architecture diagram of a method for predicting heme binding sites provided in an embodiment of the present invention; Figure 3 This is a model architecture diagram of a pre-trained model for heme-binding proteins provided in an embodiment of the present invention. Figure 4 This is a diagram of the prediction model architecture provided in an embodiment of the present invention; Figure 5 A schematic diagram illustrating the sensitivity, specificity, and accuracy of the prediction results of the four models provided in the embodiments of the present invention; Figure 6 A schematic diagram of AUROC values ​​for four models in the ablation experiment provided in this embodiment of the invention; Figure 7 A schematic diagram of a computer device for predicting heme binding sites provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0016] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0017] like Figure 1 and Figure 2 As shown in this embodiment, a method for predicting heme binding sites includes: S1: Obtain the protein sequence and structural information of an unknown protein.

[0018] S2: Based on protein sequence and structural information, extract multiple physicochemical properties, amino acid embedding features, and protein secondary structure features.

[0019] For example, the present invention can introduce six representative physicochemical properties, which may include hydrophilic, hydrophobic, polar, positively charged, negatively charged, and aromatic properties. For amino acids in a protein sequence that conform to these physicochemical properties, the corresponding amino acid set is obtained. ,in Indicates the first Physicochemical properties For example, hydrophilic amino acids include lysine (K), arginine (R), aspartic acid (D), glutamic acid (E), and serine (S). Therefore, The other five physicochemical properties (hydrophobic, polar, positively charged, negatively charged, and aromatic) are similar.

[0020] For example, to extract evolutionary information about proteins, the MMseqs program is used to compute position-specific scoring matrices for these protein sequences. During the calculation process, to enrich the evolutionary information in the generated features as much as possible, the Uniclust30 protein database was used as a comparison reference library. The final result... It is a size of The matrix is ​​expressed as follows:

[0021] in, Represents the position in the position-specific scoring matrix The amino acids were replaced with amino acids The score, , It is a positive integer representing the length of the protein sequence. .here, It represents a set of 20 standard amino acids.

[0022] Next, for each residue position in the protein and each physical and chemical property : 1. Determine the residue type: Identify the residue type Does it belong to physical and chemical properties? The corresponding amino acid set .

[0023] 2. Extract the corresponding position of this residue. Scoring vector: if residues belong Then extract the 20-dimensional data of the residue position. Scoring Vector .

[0024] 3. Accumulation Scoring vector: for physicochemical properties All belonging to residues The scoring vectors are summed element by element to obtain a 20-dimensional vector. ,in Traverse all belonging to The position of the residue.

[0025] 4. Calculation average Scoring vector: Calculates physical and chemical properties average Scoring Vector ,in It is a physicochemical property of proteins The number of residues.

[0026] Finally, for the six physical and chemical properties, a result was obtained. matrix Each row corresponds to the average of a physical and chemical property. Scoring vector: For example, amino acid embedding features are obtained using the protein language model ESM2 (Evolutionary Scale Modeling 2). ESM2 is a large-scale protein language model based on the Transformer architecture. Through self-supervised learning on massive protein sequences, it can automatically learn high-dimensional semantic features from the original sequence, effectively capturing dependencies and potential structural and functional information between distant residues. This invention uses the esm2_t12_35M_UR50D pre-trained model, which outputs 480-dimensional amino acid embedding features for each residue in the protein sequence.

[0027] Protein secondary structure features were calculated using the PSIPRED tool. Using PSIPRED V3.21 with default parameters, the probability values ​​for α-helix, β-sheet, and random coil of each residue in the protein sequence were obtained. Therefore, each residue includes a 3-dimensional secondary structure probability feature.

[0028] These features characterize potential heme binding sites in protein sequences from different perspectives, thereby improving the accuracy and generalization ability of prediction models.

[0029] S3: Perform fusion operations on multiple physicochemical property features to determine global evolutionary features. Perform forward and reverse temporal feature analysis and fusion on the global evolutionary features to determine temporal features. By performing nonlinear combination and mapping on the temporal features, obtain the probability that the unknown protein is a heme-binding protein.

[0030] For example, to effectively improve the ability to predict heme binding sites, this invention adopts a hierarchical model architecture, constructing a pre-trained model for heme-binding proteins and a prediction model for obtaining heme binding site prediction results. This two-layer model design allows the proposed method to gradually transition from determining the global function of proteins to the fine-grained identification of specific binding sites.

[0031] Generally, the BioLiP2 database is a database focused on biomolecular ligand binding information. It provides annotations of protein-small molecule ligand binding sites and covers detailed information on binding residues and ligand types. The UniProt database comprehensively integrates multi-dimensional data on protein sequence, function, structure, and evolution. These two databases provide important data support for the study of heme binding sites. In terms of sample construction, positive samples were selected from the BioLiP2 database containing heme ligand binding, and their protein structural chains and annotation information of binding residues directly interacting with the heme ligand were extracted. Negative samples were extracted from the BioLiP2 database containing proteins that do not contain heme ligand binding, and their protein structural chains were extracted. The CD-HIT tool (with a similarity threshold set to 30%) was used to remove redundancy from all positive and negative samples to ensure that no two samples were homologous. The ratio of positive to negative samples was set at 1:1. Next, for these protein structural chains, their complete protein sequences will be obtained from the UniProt database, which will help to obtain accurate protein evolution information.

[0032] All positive and negative samples are randomly selected to form a training dataset (70% each) and a test dataset (30% of the samples) to provide data support for the training and testing of the hierarchical network.

[0033] For example, in this invention, the probability that the unknown protein is a heme-binding protein is obtained through a pre-trained model. The pre-trained model specifically includes a bidirectional long short-term memory neural network (Bi-LSTM) module and a first fully connected neural network (FCNN) module. The Bi-LSTM module is used to perform forward and reverse temporal feature analysis and fusion on global evolutionary features to determine temporal features. The FCNN module is used to perform nonlinear combination and mapping on the temporal features to obtain the probability that the unknown protein is a heme-binding protein.

[0034] S4: The probability that the unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features are spliced ​​together to obtain spliced ​​features. By introducing a multi-head attention mechanism, different parts of the spliced ​​features are weighted and summed to determine the weighted summed features. The weighted summed features are then nonlinearly combined and mapped to obtain the heme binding site prediction result; wherein, the heme binding site prediction result is the probability that the unknown protein is a heme binding site.

[0035] Optionally, the prediction results of heme binding sites are obtained through a hierarchical network, which specifically includes: A pre-trained model includes a Bi-LSTM (Bidirectional Long Short-Term Memory) neural network module and a first fully connected neural network FCNN (Finite-CNN) module. The Bi-LSTM module performs forward and reverse temporal feature analysis and fusion on global evolutionary features to determine temporal features. The FCNN module performs nonlinear combination and mapping on the temporal features to obtain the probability that an unknown protein is a heme-binding protein. The prediction model includes a multi-head attention mechanism module and a second fully connected neural network (FCNN) module. The multi-head attention mechanism module is used to concatenate the probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features to obtain concatenated features. By introducing a multi-head attention mechanism, different parts of the concatenated features are weighted and summed to determine the weighted summed features. The second fully connected neural network (FCNN) module is used to perform nonlinear combination and mapping on the weighted summed features to obtain the heme binding site prediction results.

[0036] The input to the pre-trained model is a global evolutionary feature that integrates physicochemical properties, such as... Figure 3 As shown, the model consists of two parts: a Bi-LSTM module and a fully connected neural network (FCNN) module. Bi-LSTM is used to capture contextual dependencies in protein sequences; the FCNN module performs cross-dimensional fusion and nonlinear transformation on the temporal features output by Bi-LSTM, further integrating the complex relationships between evolutionary conservation and physicochemical properties. Finally, the pre-trained model outputs the probability of whether a protein possesses heme-producing ability, serving as a preliminary judgment at the protein's functional level.

[0037] Long Short-Term Memory (LSTM) networks utilize gating mechanisms such as input gates, forget gates, and output gates to control the transmission and retention of information, thereby effectively capturing contextual dependencies in longer sequences. Bi-LSTM, an extension of LSTM, introduces two directional LSTM layers: a forward LSTM that processes data from the beginning of the sequence backward, and a backward LSTM that processes data from the end of the sequence forward. This bi-directional structure allows the model to simultaneously utilize information from both the current position and the preceding / following positions to enhance its understanding of the global context of the sequence. In other words, by bi-directionally scanning the global evolutionary features of protein fusion physicochemical properties, it effectively captures information rich in distant residues within the protein sequence.

[0038] For example, the fully connected neural network FCNN module includes three linear transformation layers with 60, 30, and 10 nodes respectively. The activation function between each linear transformation layer is the ReLU function, and the final output activation function is the sigmoid function. The sigmoid function output is a real number between 0 and 1, representing the probability that the unknown protein is a heme-binding protein. This invention further nonlinearly combines and maps the temporal features extracted by Bi-LSTM through the first fully connected neural network FCNN module, fusing global evolutionary information. This allows it to learn a higher-level abstract representation between features, ultimately outputting the probability that the unknown protein is a heme-binding protein as the discrimination result. Figure 2 A diagram of the pre-trained model architecture is provided.

[0039] This invention uses a training dataset to train and optimize a pre-trained model, and a test set to evaluate the model's predictive performance.

[0040] like Figure 4 As shown, the input features of the heme prediction model include three parts: (1) the output features of the first-layer pre-trained model; (2) the secondary structure features of the protein; and (3) the amino acid embedding features of ESM2. The output features of the first-layer pre-trained model are 10-dimensional, the secondary structure features of the protein are 3-dimensional, and the amino acid embedding features of ESM2 use a 3-layer fully connected neural network FCNN architecture, including 3 linear transformation layers with 120, 60, and 10 nodes respectively. The activation function between each linear transformation layer is the ReLU function, and the final output features are 10-dimensional. Among them, the pre-trained model is the first-layer pre-trained model, that is, the parameter weights of the first layer that have been trained are frozen to retain the effective feature representations learned by it in protein function discrimination. The core purpose of this design is to achieve precise localization at the residue level, that is, to accurately predict the heme binding site, based on the screening of potential heme-binding proteins in the pre-trained model.

[0041] For example, after splicing the probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features, a multi-head attention mechanism is introduced to effectively mine the dynamic dependencies between the heme binding site and its adjacent and non-adjacent residues. In the multi-head attention mechanism, each attention head performs a weighted summation on different parts of the input sequence, thereby capturing the dependencies between different positions in the sequence. By focusing on different local or global structural information of the sequence with different attention heads, the sequence can be analyzed from multiple perspectives, resulting in a more comprehensive understanding of sequence features and a more accurate prediction of the heme binding site. The multi-head attention mechanism has 23 input features and 10 output features. The probability that each residue is a heme binding site is output through a second fully connected neural network (FCNN) module. The second fully connected neural network FCNN module uses a 3-layer fully connected neural network FCNN architecture, including 3 linear transformation layers. The activation function between each linear transformation layer is the ReLU function, and the final output activation function is the sigmoid function. The output value of the sigmoid function is a real number between 0 and 1, representing the probability that the unknown protein is a heme binding site. Figure 3 The diagram shown is a prediction model architecture diagram for the second layer of heme.

[0042] This invention employs a hierarchical model design and transfer learning strategy. The model fully learns the global evolutionary features of heme-binding proteins in the first-layer pre-training model, and then performs transfer learning in the second-layer heme prediction model. The global evolutionary features of the first layer are complementary and integrated with the local secondary structure features and amino acid embedding features of the second layer, thereby significantly improving the model's prediction accuracy and generalization ability.

[0043] To verify the effectiveness of the method of this invention, an ablation experiment was used to assess the contribution of each important part of the model to the prediction level. Specifically, by progressively removing core modules or features of the model, changes in model performance were observed to verify the contribution of each module to the final prediction result. This experiment mainly verifies the following questions: 1) Can the transfer learning strategy effectively utilize the output information of the already optimized first-layer pre-trained model? 2) Can the hierarchical strategy effectively improve the prediction performance of heme binding sites? 3) Can the global evolutionary features that integrate physicochemical properties promote the accurate identification of heme binding sites?

[0044] Based on this, the following four models were constructed: (1) Complete model: using the complete feature construction method and model framework design proposed in this invention; (2) No transfer learning model: the first-layer pre-trained model is not trained or optimized, and the network parameters are not frozen; (3) No hierarchical strategy model: the first-layer pre-trained model is removed, and only the second-layer heme prediction model is retained; (4) No global evolution feature model: both the first and second layers use only protein secondary structure features and amino acid embedding features. Next, the models were built and optimized on the constructed training dataset, and the prediction effects of the four models on the test dataset were tested respectively, such as Figure 5 As shown.

[0045] Generally, sensitivity (SN) measures the model's ability to correctly identify positive samples. A value closer to 1 indicates a stronger ability to identify positive samples, meaning the model in this invention can better predict heme binding sites. Specificity (SP) measures the model's ability to correctly identify negative samples. A value closer to 1 indicates a stronger ability to identify negative samples, meaning the model can better distinguish non-heme binding sites. Accuracy (ACC) measures the proportion of correct predictions among all model predictions; it is an overall measure of the model's predictive accuracy. Figure 4 The results show that the complete model can better identify heme-binding sites and non-heme-binding sites. The sensitivity, specificity, and accuracy of the complete model are 73%, 71%, and 72%, respectively, while those of the model without transfer learning are 67%, 64%, and 63%. Compared with the complete model, the sensitivity, specificity, and accuracy of the model without transfer learning decreased by 6%, 7%, and 9%, respectively. The sensitivity, specificity, and accuracy of the model without stratification strategy are 64%, 63%, and 64%, respectively. Compared with the complete model, the sensitivity, specificity, and accuracy of the model without stratification strategy decreased by 9%, 8%, and 8%, respectively. The model without global evolutionary features has a sensitivity, specificity, and accuracy of 65%, 67%, and 62% on the test set, respectively. Compared with the complete model, its sensitivity, specificity, and accuracy decreased by 8%, 4%, and 10%, respectively.

[0046] like Figure 6 The figure shows the Area Under the Receiver Operating Characteristic Curve (AUROC) values ​​of four models in the ablation experiment. AUROC is an important indicator of the performance of a classification model; the closer the AUROC value is to 1, the better the model's performance in distinguishing between positive and negative samples. Figure 6As can be seen, the AUROC value of the complete model can reach 0.75, while the AUROC values ​​of the other three models are 0.66, 0.65, and 0.68, respectively. This indicates that the three models are not as good as the complete model in distinguishing between positive and negative samples. The complete model can more accurately predict the heme binding site.

[0047] Depend on Figure 5 and Figure 6 The experimental results clearly show that the complete model outperforms the other three models in all performance metrics. This also demonstrates that the transfer learning strategy significantly improves the model's understanding of features, the hierarchical strategy helps the model transition from global judgment to local residue recognition, enhancing the hierarchical nature of feature learning, and the global evolutionary features incorporating physicochemical properties are crucial for predicting heme binding sites; their absence degrades model performance. The ablation experiments above effectively validate the effectiveness of the key designs in this invention, especially the introduction of transfer learning and the hierarchical strategy, which play a vital role in improving the accuracy of heme binding site prediction.

[0048] To further verify the advantages of the method of this invention compared with traditional machine learning algorithms, the following experiment was designed: The model of the method of this invention was compared with four comparative methods: K-nearest neighbors, support vector machines, linear regression, and a stochastic model (the model outputs random values, representing the baseline prediction level). All the above methods used the same training and testing datasets and the same input features. This experiment was used to verify the effectiveness of the deep learning framework proposed in this invention.

[0049] To comprehensively compare the prediction performance of various methods, each method was first learned and optimized on the same training set, and then the optimized model was tested on the test dataset. Next, 50% of the proteins were randomly selected from the test dataset to calculate various evaluation metrics. This process was repeated 10 times. Table 1 shows the performance comparison (mean ± standard deviation) between the method of this invention and traditional machine learning methods in these 10 experiments.

[0050] Table 1. Performance comparison between the method of this invention and traditional machine learning methods As shown in Table 1, the method of this invention significantly outperforms traditional machine learning methods. This demonstrates that the deep learning framework proposed in this invention is more suitable for handling long-range contextual dependencies in protein sequences and for analyzing protein sequence features, thus better predicting heme binding sites. Compared to traditional machine learning methods, the method of this invention exhibits stronger accuracy across various metrics.

[0051] The above are one or more embodiments of the method for predicting heme binding sites provided in this specification. Based on the same idea, this specification also provides a corresponding heme binding site prediction system to implement the heme binding site prediction method provided in the above embodiments.

[0052] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for predicting heme binding sites.

[0053] The present invention also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the heme binding site prediction method provided in the above embodiments.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for predicting heme binding sites, characterized in that, include: To obtain the protein sequence and structural information of unknown proteins; Based on protein sequence and structural information, multiple physicochemical property features, amino acid embedding features, and protein secondary structure features are extracted. Multiple physicochemical properties are fused to determine global evolutionary features. Forward and reverse temporal features of the global evolutionary features are analyzed and fused to determine temporal features. By performing nonlinear combination and mapping of temporal features, the probability that the unknown protein is a heme-binding protein is obtained. The probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features are spliced ​​together to obtain spliced ​​features. By introducing a multi-head attention mechanism, different parts of the spliced ​​features are weighted and summed to determine the weighted summed features. The weighted summed features are then nonlinearly combined and mapped to obtain the heme binding site prediction results. The heme binding site prediction results represent the probability that the unknown protein is a heme binding site.

2. The method for predicting heme binding sites as described in claim 1, characterized in that, The prediction results of the heme binding sites were obtained through a hierarchical network, which specifically includes: A pre-trained model includes a Bi-LSTM (Bidirectional Long Short-Term Memory) neural network module and a first fully connected neural network FCNN (Finite-CNN) module. The Bi-LSTM module performs forward and reverse temporal feature analysis and fusion on global evolutionary features to determine temporal features. The FCNN module performs nonlinear combination and mapping on the temporal features to obtain the probability that an unknown protein is a heme-binding protein. The prediction model includes a multi-head attention mechanism module and a second fully connected neural network (FCNN) module. The multi-head attention mechanism module is used to concatenate the probability that an unknown protein is a heme-binding protein, amino acid embedding features, and protein secondary structure features to obtain concatenated features. By introducing a multi-head attention mechanism, different parts of the concatenated features are weighted and summed to determine the weighted summed features. The second fully connected neural network (FCNN) module is used to perform nonlinear combination and mapping on the weighted summed features to obtain the heme binding site prediction results.

3. The method for predicting heme binding sites as described in claim 2, characterized in that, The first fully connected neural network FCNN module specifically includes three linear transformation layers; wherein the activation function between each linear transformation layer is the ReLU function, and the final output activation function is the sigmoid function.

4. The method for predicting heme binding sites as described in claim 1, characterized in that, The aforementioned multiple physicochemical properties specifically include: Hydrophilic physicochemical properties, hydrophobic physicochemical properties, polar physicochemical properties, positively charged physicochemical properties, negatively charged physicochemical properties, and aromatic physicochemical properties.

5. The method for predicting heme binding sites as described in claim 1, characterized in that, The amino acid embedding features are obtained from the protein sequence using the protein language model ESM2.

6. A system for predicting heme binding sites, characterized in that, Implement the heme binding site prediction method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the heme binding site prediction method according to any one of claims 1-5.

8. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the heme binding site prediction method according to any one of claims 1-5.

Images