Visual purple protein maximum absorption optical wavelength prediction method based on graph neural network

Abstract

The invention discloses a visual purple protein maximum absorption light wavelength prediction method based on a graph neural network. Comprising the following steps: generating a three-dimensional structure model according to an amino acid sequence of a to-be-predicted rhodopsin protein; generating a ligand pocket graph according to the three-dimensional structure model of the visual purple protein to be predicted; and inputting the ligand pocket graph into a pre-trained maximum absorption wavelength prediction model to obtain a maximum absorption wavelength prediction value. The method can make full use of protein three-dimensional structure information, accurately characterizes the ligand binding pocket microenvironment, has high prediction precision and strong generalization ability, and has important scientific significance and application value.

Description

Technical Field

[0001] This invention belongs to the field of technology for predicting the maximum absorption wavelength of microbial rhodopsin proteins, specifically involving a graph neural network deep learning system based on the relationship between protein sequence and three-dimensional structure and its maximum absorption wavelength. Background Technology

[0002] Rhodopsin is one of the most important tool proteins in optogenetics. Since most discovered rhodopsin proteins are optimally excited by low-tissue-penetration blue or green light, requiring the implantation of optical fibers to successfully guide the light to the target tissue, developing novel rhodopsin proteins excited by long-wavelength red or infrared light has been a long-term goal for researchers.

[0003] With the decrease in sequencing costs, the sequencing range and distribution resolution of samples have been greatly improved. In recent years, a large number of rhodopsin channel proteins derived from viruses have even been discovered, greatly expanding the optogenetic toolkit. Faced with such a large number of novel rhodopsin proteins, it would be a difficult task to perform spectral determination for each one after gene synthesis, protein expression, and then analysis. Therefore, it is necessary to develop an algorithmic model that can reliably predict the maximum absorption wavelength of novel rhodopsin proteins, thereby minimizing the workload.

[0004] In existing technologies, researchers have attempted to use machine learning methods to predict the absorption wavelength of rhodopsin. Examples include machine learning methods based on Bayesian experimental design and data-driven machine learning methods. However, these existing methods mainly suffer from the following shortcomings: (1) Most methods are based solely on protein sequence information and fail to fully utilize protein three-dimensional structure information; (2) Traditional machine learning methods have limited ability to characterize interatomic interactions (such as hydrogen bonds, hydrophobic interactions, π-π stacking, etc.), making it difficult to accurately depict the complex interaction network between ligands and proteins; (3) Existing models have insufficient generalization ability for newly discovered rhodopsin proteins (especially novel proteins derived from viruses), and the prediction accuracy needs to be improved. Summary of the Invention

[0005] To address the shortcomings of existing methods for predicting the absorption wavelength of rhodopsin proteins, this invention provides a method for predicting the maximum absorption wavelength of rhodopsin proteins based on a graph neural network (GNN). This method fully utilizes the three-dimensional structural information of proteins, particularly the spatial conformation and interatomic interaction characteristics of the retinal ligand binding pocket, to establish a high-precision, highly generalizable method for predicting the maximum absorption wavelength of rhodopsin proteins. This method can efficiently screen a large number of newly discovered rhodopsin proteins, reduce experimental costs and time, and accelerate the development of redshifted light-sensitive proteins.

[0006] The technical solution adopted in this invention is: I. A method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks The method for predicting the maximum absorption wavelength of rhodopsin includes the following steps: Step S1) Based on the amino acid sequence of the rhodopsin protein to be predicted, use the AlphaFold structure prediction model to generate a three-dimensional structure model containing the retinal ligand.

[0007] Step S2) Generate a ligand pocket diagram based on the three-dimensional structural model of the rhodopsin protein to be predicted.

[0008] Step S2 includes: Step S2.1) Taking the retinaldehyde ligand binding region in the three-dimensional structural model of the rhodopsin protein to be predicted as the center, select all atoms within a preset range, and treat each atom as a graph node, and use the physicochemical properties of the atom as node features.

[0009] In step S2.1, the preset range is set to be within 6 Å of the retinaldehyde ligand.

[0010] The physicochemical properties of the atom include a combination of one or more of the following parameters: atom type, atomic degree, number of bonded hydrogen atoms, valence electron state, hybridization state, formal charge, number of unpaired electrons, aromaticity, whether it is in a ring structure, hydrophobicity, hydrogen bond donor, hydrogen bond acceptor, acidic group, and basic group.

[0011] Step S2.2) Map the atomic pairs that meet the preset conditions (i.e., the spatial distance is less than the preset threshold and they have interactions) to graph edges, and use the inter-atomic interaction parameters of the atomic pairs as edge features.

[0012] In step S2.2), the process of mapping the atomic pairs that meet the preset conditions to graph edges is specifically as follows: Step S2.2.1) Based on the spatial distance between all graph nodes, select all graph node pairs whose spatial distance is less than a preset threshold and generate an inter-atomic distance matrix; in step S2.2.1, the preset threshold is set to 6 Å; Step S2.2.2) Based on the interatomic distance matrix, use cheminformatics tools to obtain all covalently bonded graph node pairs, construct a covalent bond connection matrix, and then obtain all non-covalently bonded graph node pairs to construct a non-covalent bond connection matrix. In practice, the cheminformatics tools can be tools such as Python's Rdkit package and the Open Bable program. Step S2.2.3) Maps the connected graph node pairs in the covalent bond connection matrix and the non-covalent bond connection matrix to graph edges that are covalently bonded and non-covalently bonded, respectively. For graph edges that are covalently bonded, the interatomic interaction parameters include a combination of one or more of the following parameters: chemical bond type, bond stereochemistry, ring structure characteristics, and conjugation characteristics. For graph edges that are non-covalently bonded, the interatomic interaction parameters include a combination of one or more of the following parameters: spatial distance, charge interaction, hydrophobic interaction, hydrogen bonding, acid-base interaction, aromatic stacking, and ring structure interaction.

[0013] Step S2.3) Generate a ligand pocket graph based on all graph nodes and edges.

[0014] Step S3) Input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

[0015] In step S3, the maximum absorption wavelength prediction model includes a graph neural network sub-model and a prediction network sub-model connected in series. The graph neural network sub-model receives the ligand pocket map and outputs the graph-level feature vector to the prediction network sub-model, which outputs the predicted value of the maximum absorption wavelength. The graph neural network sub-model adopts a graph attention network model, which includes several GATv2Conv layers and a double graph pooling layer. The prediction network sub-model adopts a multilayer perceptron model, which includes several fully connected layers connected in series.

[0016] The training process of the maximum absorption wavelength prediction model includes: D1) Obtain multiple rhodopsin protein samples with defined amino acid sequences and experimentally determined maximum absorption wavelengths. Normalize the maximum absorption wavelength of all rhodopsin protein samples. Generate a corresponding ligand pocket map based on the amino acid sequence of each rhodopsin protein sample. Construct a sample pair by combining the ligand pocket map and the normalized maximum absorption wavelength of each rhodopsin protein sample. Combine all sample pairs to obtain the dataset. D2) Construct a prediction model for the maximum absorption wavelength; D3) Based on the dataset, the mean squared error (MSE) is used as the loss function. The Adam optimization method is combined with the learning rate adjustment method to train the maximum absorption wavelength prediction model. After reaching the preset iteration termination condition or the preset number of iterations, the trained maximum absorption wavelength prediction model is obtained.

[0017] II. A Prediction System for the Maximum Absorption Wavelength of Rhodopsin Based on Graph Neural Networks The aforementioned rhodopsin protein maximum absorption wavelength prediction system includes: The structure prediction unit is used to generate a three-dimensional structure model based on the amino acid sequence of the rhodopsin protein to be predicted. The ligand pocket map generation unit is used to generate a ligand pocket map based on the three-dimensional structural model of the rhodopsin protein to be predicted. The maximum absorption wavelength prediction unit is used to input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

[0018] The beneficial effects of this invention are: (1) Make full use of three-dimensional structural information: The present invention constructs ligand pocket diagrams based on the three-dimensional structure of proteins. Compared with traditional methods based solely on sequences, it can accurately capture the spatial conformation and microenvironmental characteristics (interatomic distances and interaction networks) of the retinaldehyde ligand binding pocket, significantly improving the characterization ability of factors affecting absorption wavelength.

[0019] (2) Accurate characterization of interatomic interactions: By encoding detailed information on covalent and non-covalent interactions (hydrogen bonds, hydrophobic interactions, π-π stacking, electrostatic interactions, etc.) in the edge features of the graph, this invention can accurately characterize the complex interaction network of ligand-protein, which is difficult to achieve with traditional machine learning methods.

[0020] (3) High prediction accuracy: In the prediction of 179 non-training rhodopsin proteins, the determination coefficient R of the method of the present invention is high. 2 The accuracy reached 0.781, with a mean absolute error (MAE) of 10.21 nm and a root mean square error (RMSE) of 13.22 nm, demonstrating significantly better prediction accuracy than existing methods.

[0021] (4) Strong generalization ability: By adopting the end-to-end feature learning mechanism of graph neural network, the model can automatically learn the potential correlation between the ligand pocket microenvironment and the absorption wavelength, and has good generalization ability for newly discovered rhodopsin proteins (including novel proteins derived from viruses).

[0022] (5) High computational efficiency: Combining the advantages of AlphaFold3 structure prediction and graph neural network fast reasoning, the method of this invention can quickly screen a large number of candidate proteins without experimental determination, significantly reducing experimental costs and time, and accelerating the development process of redshift light-sensitive proteins.

[0023] (6) Wide range of applications: The method of the present invention is not only applicable to the prediction of the absorption wavelength of natural rhodopsin protein, but also provides important theoretical guidance and tool support for protein engineering modification, rational design and spectral regulation mechanism research, and has a wide range of application prospects. Attached Figure Description

[0024] Figure 1 This is an overall flowchart of the method for predicting the maximum absorption wavelength of rhodopsin protein according to the present invention.

[0025] Figure 2 This is a scatter plot comparing the prediction results of the method of the present invention on the test set with the experimentally measured values. Detailed Implementation

[0026] To thoroughly reveal the core objectives, technical solutions, and significant advantages of this invention, the content of this invention is further elaborated here in conjunction with the accompanying drawings and embodiments. It should be clarified that the embodiments presented herein are only a part of the many examples of this invention, not all of them. The steps shown and described in the drawings can be laid out and designed according to a variety of different schemes. Importantly, the following detailed description of the embodiments of this invention is not intended to limit its scope of protection. Rather, it is intended as an example of certain selected embodiments of this invention. All other embodiments based on this invention proposed by those skilled in the art without inventive effort should be considered to fall within the scope of protection of this invention.

[0027] This invention provides a method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks.

[0028] Figure 1 This is a flowchart illustrating the overall process of the method for predicting the maximum absorption wavelength of rhodopsin protein according to the present invention. It shows the complete flow from protein sequence input, structure prediction, ligand pocket diagram construction, graph neural network feature extraction, to absorption wavelength prediction output. The method of the present invention includes the following steps: Step S1) Based on the amino acid sequence of the rhodopsin protein to be predicted, use the AlphaFold structural prediction model to generate a three-dimensional structural model containing the retinal chorophore.

[0029] Step S2) Generate a ligand pocket diagram based on the three-dimensional structural model of the rhodopsin protein to be predicted.

[0030] Specifically, the following steps are included: Step S2.1) Taking the retinaldehyde ligand binding region in the three-dimensional structural model of the rhodopsin protein to be predicted as the center, select all atoms within a preset range, and treat each atom as a graph node, and use the physicochemical properties of the atom as node features.

[0031] Preferably, the preset range is set to be within 6 Å of the retinaldehyde ligand.

[0032] Optionally, the physicochemical properties of an atom include a combination of one or more of the following parameters: atom type, atomic degree, number of attached hydrogen atoms, valence state, hybridization state, formal charge, number of unpaired electrons, aromaticity, ring membership, hydrophobicity, hydrogen bond donor, hydrogen bond acceptor, acidic group, and basic group.

[0033] Step S2.2) Map the atomic pairs that meet the preset conditions (i.e., the spatial distance is less than the preset threshold and they have interactions) to graph edges, and use the inter-atomic interaction parameters of the atomic pairs as edge features.

[0034] The process of mapping atomic pairs that meet the preset conditions to graph edges is as follows: Step S2.2.1) Based on the spatial distance between all graph nodes, select all graph node pairs whose spatial distance is less than a preset threshold and generate an inter-atomic distance matrix.

[0035] Preferably, the preset threshold is set to 6 Å.

[0036] Step S2.2.2) Based on the interatomic distance matrix, using cheminformatics tools, first obtain all graph node pairs connected by covalent bonds and construct a covalent bond connection matrix, then obtain all graph node pairs without covalent bonds and construct a non-covalent bond connection matrix, thereby distinguishing different types of atomic interaction relationships.

[0037] Step S2.2.3) Map the connected graph node pairs in the covalent bond connection matrix and the non-covalent bond connection matrix to covalent bond connected graph edges and non-covalent bond connected graph edges, respectively.

[0038] Optionally, for graph edges connected by covalent bonds, the interatomic interaction parameters include a combination of one or more of the following parameters: bond type, stereochemical configuration, ring structure characteristics, and conjugation characteristics.

[0039] Optionally, for graph edges connected by non-covalent bonds, the interatomic interaction parameters include a combination of one or more of the following parameters: spatial distance, electrostatic interaction, hydrophobic interaction, hydrogen bonding, acid-base interaction, aromatic stacking and cyclic interaction.

[0040] Step S2.3) Generate a ligand pocket graph based on all graph nodes and edges.

[0041] Step S3) Input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

[0042] Specifically, the maximum absorption wavelength prediction model includes a graph neural network sub-model and a prediction network sub-model connected in series. The graph neural network sub-model receives the ligand pocket map and outputs the graph-level feature vector to the prediction network sub-model, which outputs the predicted value of the maximum absorption wavelength.

[0043] Specifically, the graph neural network sub-model employs a graph attention network model, comprising several GATv2Conv layers and a dual graph pooling layer. Each GATv2Conv layer is configured to perform weighted aggregation of node features through attention weight allocation. The input of the first GATv2Conv layer serves as the input of the graph neural network sub-model. The GATv2Conv layers are sequentially connected to obtain high-dimensional node embedding vectors through multi-layer feature extraction. The output of the last GATv2Conv layer is connected to the input of the dual graph pooling layer, which performs full-graph-level feature aggregation using a pooling mechanism. The output of the dual graph pooling layer serves as the output of the graph neural network sub-model, outputting a graph-level feature vector, which is then used for downstream spectral property prediction tasks.

[0044] Preferably, in the graph neural network sub-model, the number of GATv2Conv layers is five.

[0045] Specifically, the prediction network sub-model adopts a multilayer perceptron model, which includes several cascaded fully connected layers, with adjacent fully connected layers connected by nonlinear activation function layers.

[0046] Preferably, in the prediction network sub-model, there are three fully connected layers. The output of the first fully connected layer is fed into the second connected layer through the first nonlinear activation function layer and the first dropout layer. The output of the second fully connected layer is fed into the third fully connected layer through the second nonlinear activation function layer and the second dropout layer. Both the first nonlinear activation function layer and the second nonlinear activation function layer use the ReLU activation function.

[0047] In the preferred embodiment described above, in the prediction network sub-model, the first fully connected layer is used to map the dimension of the graph-level feature vector to the hidden dimension, the second fully connected layer is used to reduce the dimension of the output of the first Dropout layer, and the third connected layer is used to obtain the predicted value of the maximum absorption wavelength based on the output of the second Dropout layer.

[0048] In practice, the graph-level feature vector has a dimension of 512, the hidden dimension is 1024, and the second fully connected layer is used to halve the dimensions.

[0049] Preferably, the training process of the maximum absorption wavelength prediction model includes: Dataset Construction: Obtain multiple rhodopsin protein samples with defined amino acid sequences and experimentally determined maximum absorption wavelengths. Normalize the maximum absorption wavelengths of all rhodopsin protein samples. Generate corresponding ligand pocket maps based on the amino acid sequences of each rhodopsin protein sample. Construct a sample pair by combining the ligand pocket map and the normalized maximum absorption wavelength of each rhodopsin protein sample. Combine all sample pairs to obtain the dataset and divide it into training and test sets. Construct a model for predicting the maximum absorption wavelength; Training the maximum absorption wavelength prediction model: Based on the training set in the dataset, the mean squared error (MSE) is used as the loss function. The Adam optimization method is used in combination with the learning rate adjustment method to train the maximum absorption wavelength prediction model. After reaching the preset iteration termination condition or the preset number of iterations, the trained maximum absorption wavelength prediction model is obtained.

[0050] Furthermore, during training, a five-fold cross-validation method is used to evaluate model performance in order to avoid overfitting and improve model generalization.

[0051] This invention also provides a system for predicting the maximum absorption wavelength of rhodopsin protein, used to execute the above-described graph neural network-based method for predicting the maximum absorption wavelength of rhodopsin protein. It includes: The structure prediction unit is used to generate a three-dimensional structure model based on the amino acid sequence of the rhodopsin protein to be predicted. The ligand pocket map generation unit is used to generate a ligand pocket map based on the three-dimensional structural model of the rhodopsin protein to be predicted. The maximum absorption wavelength prediction unit is used to input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

[0052] Example 1 This embodiment specifically illustrates the method for constructing the dataset required for the maximum absorption wavelength prediction model. The dataset consists of the name, amino acid sequence, and experimentally determined maximum absorption wavelength of each rhodopsin protein sample.

[0053] In this embodiment, the dataset mainly comes from the following publicly published literature: (1) Exploration of Natural Red-Shifted Rhodopsins Using a MachineLearning-Based Bayesian Experimental Design; (2) Understanding Color Tuning Rules and Predicting AbsorptionWavelengths of Microbial Rhodopsins by Data-Driven Machine-Learning Approach; (3) Structural basis for ion selectivity in potassium-selective channelrhodopsins; (4) Structural basis for channel conduction in the pump-like channelrhodopsin ChRmine; (5) The high-light-sensitivity mechanism and optogenetic properties of the bacteriorhodopsin-like channelrhodopsin GtCCR4; (6) Structural basis for color tuning and passive ion conductance inred-shifted pump-fold channelrhodopsin ChR024.

[0054] After data processing and quality control, this embodiment collected 912 rhodopsin protein samples with clearly defined amino acid sequences and experimentally determined maximum absorption wavelengths. The dataset covers various types of rhodopsin proteins from different microbial sources (including bacteria, archaea, and viruses), with absorption wavelengths ranging from approximately 470 nm to approximately 650 nm, effectively representing the spectral diversity of the rhodopsin protein family. The dataset was randomly divided into a training set (733 samples) and a test set (179 samples) in an 8:2 ratio. The training set was used for model training and five-fold cross-validation, while the test set was used to evaluate the model's generalization performance.

[0055] Example 2 This embodiment provides a training method for a maximum absorption wavelength prediction model. This embodiment includes: converting the three-dimensional structure generated from the rhodopsin protein sequence into a ligand pocket diagram with atoms as nodes, and then using a graph neural network to predict the absorption wavelength of the rhodopsin protein using a deep learning model. Specifically, it includes the following steps: Step 1: Protein 3D Structure Prediction Using the AlphaFold3 structure prediction model, the amino acid sequence of a rhodopsin protein sample is input to generate a three-dimensional protein structure containing the retinal ligand. AlphaFold3 can simultaneously predict the binding conformation of the protein and ligand, providing accurate atomic coordinate information for subsequent ligand pocket diagram construction.

[0056] Step 2: Ligand pocket map construction (1) Taking the retinaldehyde ligand as the center, extract all amino acid residues and their atoms within 6 Å of the ligand, together with all atoms of the retinaldehyde ligand, as the node candidate set of the ligand pocket diagram.

[0057] Specifically, this step involves extracting all amino acid residues within 6 Å (angstroms) of the retinaldehyde ligand, along with the retinaldehyde ligand, to generate a ligand pocket PDB file. The Open Bable program is then used to convert the PDB file into a mol2 file containing chemical information.

[0058] (2) Using the Rdkit package in Python, based on the chemical bond information of the protein, extract all covalently linked atom pairs and construct a covalent bond connection matrix.

[0059] (3) Calculate the Euclidean distance between all nodes, extract all atom pairs with a distance within 6 Å, and construct the interatomic interaction matrix.

[0060] Step 3: Encoding of node and edge features (1) Node features: Each atom is a graph node in the ligand pocket diagram, and its feature vector is represented by one-hot encoding and boolean encoding. The total dimension of the node feature vector is 56, specifically including: (1) Atom type one-hot encoding (6-dimensional); (2) Unithermal encoding of atomic degree (7-dimensional); (3) Connecting hydrogen atoms in a unique thermal encoding (5-dimensional); (4) Implicit valence one-heat encoding (6-dimensional); (5) Explicit valence one-heat encoding (6-dimensional); (6) Implicit hydrogen atom one-heat encoding (2-dimensional); (7) Explicit hydrogen atom one-heat encoding (5-dimensional); (8) Hybrid state one-hot encoding (5-dimensional); (9) Formal charge one-heat encoding (7-dimensional); (10) Whether it is Boolean encoded in the aromatic ring (1-dimensional); (11) Whether it is Boolean encoded in the ring (1-dimensional); (12) Whether it is hydrophobic Boolean encoding (1-dimensional); (13) Whether it is a hydrogen bond donor Boolean code (1D); (14) Whether it is a hydrogen bond acceptor Boolean encoding (1-dimensional); (15) Whether it is Boolean encoded in acidic groups (1-dimensional); (16) Whether it is Boolean encoded in basic groups (1-dimensional).

[0061] (2)~(10), (11), (13)~(16) are generated based on the Rdkit package of Python, and (12) is generated based on the provided amino acid atom hydrophobic table.

[0062] (2) Edge features: The edges of the ligand pocket diagram are constructed based on the interatomic interaction matrix. The edge features are divided into two categories based on whether they are included in the covalent bond matrix: (2.1) Edge features connected by covalent bonds, the total feature vector has 26 dimensions, including: ①Key type identifier one-hot encoding (2-dimensional); ② Covalent bond type one-hot encoding (4-dimensional); ③ Covalent bond stereoconfiguration one-hot encoding (6-dimensional); ④ Whether it is Boolean encoded in a ring (1-dimensional); ⑤ Whether conjugate Boolean encoding (1-dimensional); ⑥ Distance features (4-dimensional); ⑦ Placeholder (8 dimensions).

[0063] Among them, ②③④⑤ are generated based on the Rdkit package in Python.

[0064] (2.2) Edge features of non-covalent connections, the total feature vector has 26 dimensions, including: ①Key type identifier one-hot encoding (2-dimensional); ② Placeholder (12 dimensions); ③ Distance features (4-dimensional); ④Charge and valence state characteristics (2-dimensional, including total charge and charge product); ⑤ Hydrophobic interactions (1-dimensional); ⑥ Hydrogen bond characteristics (2D); ⑦ Acid-base characteristics (1-dimensional); ⑧ Biaromatic characteristics (1-dimensional); ⑨ Double-ring feature (1-dimensional).

[0065] Among them, ④⑤⑥⑦⑧ are generated based on the node feature vectors of the two atoms.

[0066] Step 4: Extract features from the ligand pocket map using a graph neural network sub-model. This invention employs a graph attention network as the core feature extraction module. Specifically: (1) A five-layer stacked GATv2Conv layer is used to extract features from the ligand pocket map layer by layer. Each layer assigns attention weights to each atom node through a graph attention mechanism, enabling the model to adaptively focus on the atoms and their interactions that are most important for predicting absorption wavelengths.

[0067] (2) After iterative computation of the 5-layer graph attention network, each node obtains a high-dimensional embedding vector that incorporates information from its neighborhood.

[0068] (3) The dual pooling mechanism is used to aggregate the features of all nodes into a full-graph-level representation vector, which captures the overall microenvironment features of the ligand pocket.

[0069] Step 5: Predict absorption wavelength using a multilayer sensing machine sub-model The full-image-level representation vector obtained in step 4 is input into a multilayer perceptron for final absorption wavelength prediction. This multilayer perceptron employs a three-layer fully connected structure: (1) The first layer maps the pooling feature dimension (512 dimensions) to the hidden dimension (1024 dimensions) and introduces a non-linear transformation through the ReLU activation function; (2) The second layer performs feature compression, halves the dimension and uses ReLU activation again; (3) The third layer outputs a single predicted value (1-dimensional output), namely the maximum absorption wavelength of rhodopsin (unit: nm).

[0070] Dropout is used between each layer to prevent overfitting and ensure that the model has good generalization ability.

[0071] Step 6: Model Training and Optimization (1) The absorption wavelength values ​​of the training data are normalized using the target value standardization technique to optimize the training effect; (2) The mean squared error (MSE) is used as the loss function; (3) Enable early stopper and learning rate adjuster. When the loss function fails to improve multiple times during the training loop, the learning rate will be automatically adjusted. When the performance on the validation set no longer improves, training will be terminated and the best model will be restored. (4) Perform five-fold cross-validation on the training data and use the average coefficient of determination R of the five training results. 2 To guide model architecture adjustments and hyperparameter optimization; (5) After selecting the optimal combination of hyperparameters, train the final model using the entire training set and verify its generalization performance on an independent test set.

[0072] Example 3 This embodiment uses a training set to evaluate the performance of the model built and trained according to Embodiment 2.

[0073] This embodiment uses the following four metrics to evaluate the model's predictive performance on the test set: (1) Mean Squared Error (MSE): Reflects the average squared error between the predicted value and the actual value; (2) Mean Absolute Error (MAE): reflects the average absolute deviation between the predicted value and the actual value; (3) Root Mean Squared Error (RMSE): The square root of MSE, which has the same dimension (nm) as wavelength. (4) Coefficient of determination (R) 2 ): Reflects the goodness of fit of the model to the data, with a value ranging from 0 to 1. The closer to 1, the better the fit.

[0074] On an independent test set containing 179 rhodopsin protein samples, the method of this invention achieved the following performance metrics: MSE=174.76, MAE=10.28 nm, RMSE=13.22 nm, R 2 =0.781.

[0075] The above results demonstrate that the method of the present invention can predict the maximum absorption wavelength of rhodopsin with high accuracy, and the coefficient of determination R0 is [value missing]. 2 The accuracy reached 0.801, with a mean absolute error of only 9.71 nm, far below the typical error range measured in experiments. This performance significantly outperforms existing sequence-based prediction methods, validating the effectiveness of using three-dimensional structural information and graph neural networks for absorption wavelength prediction.

[0076] Figure 2 This scatter plot compares the prediction results of the method of the present invention on the test set with the experimentally measured values, demonstrating the prediction accuracy and fitting effect of the model. The plot clearly shows a high linear correlation between the predicted and actual values, further proving the reliability and practicality of the method of the present invention.

[0077] The above specific embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

[0078] The above description is only a preferred embodiment of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of this patent application are included in the scope of this patent application.

Claims

1. A method for predicting the maximum absorption wavelength of rhodopsin protein based on graph neural networks, characterized in that, Includes the following steps: Step S1) Generate a three-dimensional structural model based on the amino acid sequence of the rhodopsin protein to be predicted; Step S2) Generate a ligand pocket diagram based on the three-dimensional structural model of the rhodopsin protein to be predicted; Step S3) Input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

2. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 1, characterized in that: Step S2 includes: Step S2.1) Taking the retinaldehyde ligand binding region in the three-dimensional structural model of the rhodopsin protein to be predicted as the center, select all atoms within a preset range, and treat each atom as a graph node, and use the physicochemical properties of the atom as node features. Step S2.2) Map the atomic pairs that meet the preset conditions to graph edges, and use the inter-atomic interaction parameters of the atomic pairs as edge features; Step S2.3) Generate a ligand pocket graph based on all graph nodes and edges.

3. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 2, characterized in that: In step S2.2), the process of mapping the atomic pairs that meet the preset conditions to graph edges is specifically as follows: Step S2.2.1) Based on the spatial distance between all graph nodes, select all graph node pairs whose spatial distance is less than a preset threshold and generate an inter-atomic distance matrix; Step S2.2.2) Based on the interatomic distance matrix, obtain all graph node pairs connected by covalent bonds, construct a covalent bond connection matrix, and then obtain all graph node pairs without covalent bonds, construct a non-covalent bond connection matrix; Step S2.2.3) Map the connected graph node pairs in the covalent bond connection matrix and the non-covalent bond connection matrix to covalent bond connected graph edges and non-covalent bond connected graph edges, respectively.

4. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 3, characterized in that: In step S2.1, the preset range is set to be within 6 Å of the retinaldehyde ligand; in step S2.2.1, the preset threshold is set to 6 Å.

5. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 2, characterized in that: The physicochemical properties of the atom include a combination of one or more of the following parameters: atom type, atomic degree, number of bonded hydrogen atoms, valence electron state, hybridization state, formal charge, number of unpaired electrons, aromaticity, whether it is in a ring structure, hydrophobicity, hydrogen bond donor, hydrogen bond acceptor, acidic group, and basic group. For graph edges connected by covalent bonds, the interatomic interaction parameters include a combination of one or more of the following parameters: chemical bond type, bond stereoconfiguration, ring structure characteristics, and conjugation characteristics; For graph edges connected by non-covalent bonds, the interatomic interaction parameters include a combination of one or more of the following parameters: spatial distance, charge interaction, hydrophobic interaction, hydrogen bonding, acid-base interaction, aromatic stacking and ring structure interaction.

6. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 1, characterized in that: In step S3, the maximum absorption wavelength prediction model includes a graph neural network sub-model; the graph neural network sub-model receives the ligand pocket map, outputs the graph-level feature vector to the prediction network sub-model, and the prediction network sub-model outputs the predicted value of the maximum absorption wavelength; the graph neural network sub-model adopts a graph attention network model, including several GATv2Conv layers and a double graph pooling layer; the prediction network sub-model adopts a multilayer perceptron model, including several cascaded fully connected layers.

7. The method for predicting the maximum absorption wavelength of rhodopsin based on graph neural networks according to claim 6, characterized in that: The training process of the maximum absorption wavelength prediction model includes: Step D1) Obtain multiple rhodopsin protein samples with defined amino acid sequences and experimentally determined maximum absorption wavelengths. Normalize the maximum absorption wavelengths of all rhodopsin protein samples. Generate a corresponding ligand pocket map based on the amino acid sequence of each rhodopsin protein sample. Construct a sample pair by combining the ligand pocket map and the normalized maximum absorption wavelength of each rhodopsin protein sample. Combine all sample pairs to obtain the dataset. Step D2) Construct a prediction model for the maximum absorption wavelength; Step D3) Based on the dataset, the mean squared error is used as the loss function. The Adam optimization method is combined with the learning rate adjustment method to train the maximum absorption wavelength prediction model. After reaching the preset iteration termination condition or the preset number of iterations, the trained maximum absorption wavelength prediction model is obtained.

8. A system for predicting the maximum absorption wavelength of rhodopsin protein using the graph neural network-based method for predicting the maximum absorption wavelength of rhodopsin protein as described in any one of claims 1 to 7, characterized in that, include: The structure prediction unit is used to generate a three-dimensional structure model based on the amino acid sequence of the rhodopsin protein to be predicted. The ligand pocket map generation unit is used to generate a ligand pocket map based on the three-dimensional structural model of the rhodopsin protein to be predicted. The maximum absorption wavelength prediction unit is used to input the ligand pocket diagram into the pre-trained maximum absorption wavelength prediction model to obtain the predicted value of the maximum absorption wavelength.

Images