Method and system for predicting dynamic structural changes of protein Aβ42 based on deep learning method

Abstract

The present invention provides a method set system for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method, comprising the following steps: extracting the PDB of each frame from the molecular dynamics trajectory; calculating the 2DIR of each PDB; splitting the PDB according to the functional characteristic region; calculating the α-carbon distance matrix distance map of each PDB; predicting the α-carbon distance matrix distance map of each region based on the trained deep learning model; constructing the three-dimensional structure of the protein; merging the fragments and comparing them with the three-dimensional structure: aligning the optimized three-segment PDB and merging them into a complete three-dimensional structure of the PDB protein; finally, calculating the RMSD error between the predicted PDB and the real PDB. The present invention adopts a divide-and-conquer strategy, using 2DIR to predict three different regions in the protein structure, and finally successfully constructing a three-dimensional structure. This method provides a new method for exploring the structure and dynamic changes of Aβ42, opens up a new perspective for studying Alzheimer's disease, and also provides important information for drug design.

Description

Technical Field

[0001] The present invention relates to the technical field of structural research of the Alzheimer's disease-causing protein Aβ42, and specifically to a method and system for predicting the dynamic structural changes of the protein Aβ42 based on a deep learning method. Background Art

[0002] Alzheimer's disease (AD), also commonly known as senile dementia, is a progressive neurodegenerative disease that primarily affects the elderly. The disease initially manifests as mild memory loss. As the disease progresses, it gradually affects language ability, cognitive function, judgment, and daily living abilities. In the late stages of the disease, patients may completely lose the ability to care for themselves. According to the World Health Organization, approximately 55 million people worldwide suffer from AD, and the number of patients is growing at a rate of nearly 10 million new cases per year. This number is expected to rise to 139 million by 2050. Furthermore, AD is considered a "graveyard" for new drug development within the field, with a clinical failure rate as high as 99.6%. Pharmaceutical giants such as Johnson & Johnson, Pfizer, Eli Lilly, and Roche have all failed in this area. To date, very few AD drugs have been marketed.

[0003] The pathogenesis of Alzheimer's disease is closely related to the misfolding and aggregation of β-amyloid protein (Aβ). Among them, Aβ42 is a particularly important β-amyloid protein fragment that is directly related to the development of AD. Aβ42 is a longer amyloid protein fragment obtained by enzymatic cleavage of a larger precursor protein (Amyloid precursor protein, APP). Compared with Aβ fragments of other lengths, such as Aβ40, Aβ42 is more likely to aggregate in the brain to form so-called "plaques." These plaques are a major pathological feature of AD and are also considered to be the main driving force behind the development of the disease. Therefore, it is very important to study the dynamic structural changes of Aβ42. Understanding the dynamic structural changes of Aβ42 is of great significance for the development of prevention and treatment strategies for Alzheimer's disease. This not only helps researchers identify potential intervention points, but also provides important information for drug design.

[0004] While traditional structural biology techniques such as X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, and cryo-electron microscopy (Cryo-EM) are highly effective in studying fixed structures, they are often technically limited in analyzing dynamic processes and incompletely folded states of proteins. For example, NMR has strict requirements on sample size and state, X-ray crystallography relies on difficult-to-obtain protein crystals, and while Cryo-EM is applicable to large complexes, it has room for improvement in resolving detailed information about the dynamic processes of certain proteins. Therefore, new methods are needed to study the dynamic changes of Aβ42.

[0005] Two-dimensional infrared spectroscopy (2DIR spectroscopy) is an advanced spectroscopic technique, similar to the more widely known two-dimensional nuclear magnetic resonance (2D NMR) technique. It involves applying a series of infrared pulses to the sample and generating a two-dimensional spectrum by measuring the interactions between the pulses and the response of the sample to these pulses. This two-dimensional spectrum shows the coupling relationship between different vibrational modes and can reveal the interactions within and between molecules, especially those that are difficult to distinguish in traditional IR spectroscopy. Therefore, when exploring dynamically changing and complex biomolecules such as Aβ42, two-dimensional infrared spectroscopy (2DIR) provides a powerful alternative with its unique time resolution and sensitivity to intermolecular interactions. 2DIR can capture instantaneous changes in protein structure at the atomic level, enabling this example to gain a deeper understanding of the aggregation process and pathological mechanisms of these biomolecules, thereby providing important structural information for the study of related diseases and drug development.

[0006] However, due to the high dimensionality and complexity of the data generated by two-dimensional infrared spectroscopy (2DIR), its interpretation usually requires extensive expertise and technical experience. Summary of the Invention

[0007] The technical problem to be solved by the present invention is that the interpretation of two-dimensional infrared spectroscopy (2DIR) is currently highly dependent on expert experience, and the data processing efficiency and accuracy need to be improved.

[0008] The present invention solves the above technical problems through the following technical means:

[0009] The method for predicting the dynamic structural changes of protein Aβ42 based on deep learning methods includes the following steps:

[0010] Step 1: Extract the PDB of each frame from the molecular dynamics trajectory;

[0011] Step 2: Calculate the 2DIR of each PDB;

[0012] Step 3: Split the PDB according to functional areas;

[0013] Step 4: Calculate the α-carbon distance map of each PDB;

[0014] Step 5: Predict the α-carbon distance map of each region based on the trained deep learning model;

[0015] Step 6: Constructing the three-dimensional structure of the protein;

[0016] Step 7: Merge the fragments and compare with the three-dimensional structure: Align the three optimized PDB fragments and merge them into a complete PDB protein three-dimensional structure; finally, calculate the RMSD error between the predicted PDB and the true PDB.

[0017] Furthermore, the step 2 specifically includes the following steps:

[0018] Step 2.1: Configure and run a molecular dynamics simulation to prepare the protein dynamics data. First, configure the computing resources by loading the appropriate GROMACS modules and setting up a Slurm job script. Then, create a separate directory for each PDB file and move the files into the appropriate directory. Generate the protein topology using PDB2gmx, set the simulation box using editconf, add solvent using solvate, and add ions to neutralize the system charge using genion. After completing these steps, prepare the input files required for the simulation using grompp and submit them to the computing cluster for execution.

[0019] Step 2.2 Generate Hamiltonian and Dipoles files: Use the NISE software package to process the data obtained from the molecular dynamics simulation to generate Hamiltonian and dipole moment files for 2DIR spectral simulations. These files are copied and adjusted in the directory of each simulation to ensure that they are suitable for spectral calculations. To improve the statistical accuracy and data quality of the 2DIR simulation, the Hamiltonian and Dipoles files are copied and merged multiple times, which can increase the data volume and optimize the results of subsequent spectral analysis.

[0020] Step 2.3: Run 2DIR simulations and convert data: After loading the compiler and MPI library, run the NISE software to perform 1DIR and 2DIR spectral simulations. This process includes reading the adjusted Hamiltonian and Dipoles files, performing the spectral simulation, and converting the results from text format to binary format for subsequent processing.

[0021] Step 2.4 Image processing and data organization: The generated spectral images are processed using Python scripts, including image generation, cropping, and resizing to meet analysis requirements; finally, the processed images and data files are renamed and organized to ensure that they can be efficiently stored and accessed.

[0022] Furthermore, the step 3 specifically includes the following steps:

[0023] The PDB of Aβ42 was split into PDB of residues 1-17, PDB of residues 16-28, and PDB of residues 27-42.

[0024] Furthermore, the step 4 specifically includes the following steps:

[0025] Step 4.1 Atom positioning: First, locate all α-carbon atoms from the PDB file;

[0026] Step 4.2: extracting coordinates: extracting the x, y, and z spatial coordinates of the α-carbon atom;

[0027] Step 4.3 Calculate the distance: For each pair of α-carbon atoms in the coordinates, calculate the distance between the α-carbon atoms using the Euclidean distance formula: The distance formula is

[0028]

[0029] Step 4.4: Build a matrix: Organize all the distance values ​​into a matrix, where the rows and columns of the matrix correspond to the indices of the α-carbon atoms.

[0030] Furthermore, the step 6 specifically includes the following steps:

[0031] Step 6.1 Initial point selection: At the beginning of the algorithm, a set of initial coordinates is randomly generated as the starting point of the optimization process. This random initialization method is often used to ensure that the optimization process can explore the solution space from a variety of possible configurations and reduce the risk of the solution being affected by the initial conditions.

[0032] Step 6.2 Optimization process: Through an iterative method, the positions of these points are continuously adjusted to minimize the objective function; the objective function measures the deviation between the α-carbon distance matrix calculated based on the current coordinates and the constraint condition, that is, the predicted α-carbon distance matrix distance map;

[0033] Interpretation of the results in step 6.3: After the optimization is completed, the obtained coordinates should be as close as possible to the actual three-dimensional structure while satisfying the given distance constraints.

[0034] The present invention also provides a system for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method, comprising:

[0035] PDB data extraction module, used to extract the PDB of each frame from the molecular dynamics trajectory;

[0036] 2DIR calculation module, used to calculate the 2DIR of each PDB;

[0037] The regional splitting module is used to split the PDB according to functional characteristics;

[0038] α-carbon distance matrix calculation module, used to calculate the α-carbon distance matrix distance map of each PDB;

[0039] The prediction module is used to predict the α-carbon distance map of each region based on the trained deep learning model;

[0040] Three-dimensional structure building blocks, used to construct protein three-dimensional structures;

[0041] The error calculation module is used to merge the fragments and compare them with the three-dimensional structure: the three optimized PDB fragments are aligned and merged into a complete PDB protein three-dimensional structure; finally, the RMSD error between the predicted PDB and the true PDB is calculated.

[0042] Furthermore, the 2DIR calculation module specifically includes the following steps:

[0043] Step 2.1: Configure and run a molecular dynamics simulation to prepare the protein dynamics data. First, configure the computing resources by loading the appropriate GROMACS modules and setting up a Slurm job script. Then, create a separate directory for each PDB file and move the files into the appropriate directory. Generate the protein topology using PDB2gmx, set the simulation box using editconf, add solvent using solvate, and add ions to neutralize the system charge using genion. After completing these steps, prepare the input files required for the simulation using grompp and submit them to the computing cluster for execution.

[0044] Step 2.2 Generate Hamiltonian and Dipoles files: Use the NISE software package to process the data obtained from the molecular dynamics simulation to generate Hamiltonian and dipole moment files for 2DIR spectral simulations. These files are copied and adjusted in the directory of each simulation to ensure that they are suitable for spectral calculations. To improve the statistical accuracy and data quality of the 2DIR simulation, the Hamiltonian and Dipoles files are copied and merged multiple times, which can increase the data volume and optimize the results of subsequent spectral analysis.

[0045] Step 2.3: Run 2DIR simulations and convert data: After loading the compiler and MPI library, run the NISE software to perform 1DIR and 2DIR spectral simulations. This process includes reading the adjusted Hamiltonian and Dipoles files, performing the spectral simulation, and converting the results from text format to binary format for subsequent processing.

[0046] Step 2.4 Image processing and data organization: The generated spectral images are processed using Python scripts, including image generation, cropping, and resizing to meet analysis requirements; finally, the processed images and data files are renamed and organized to ensure that they can be efficiently stored and accessed.

[0047] Furthermore, the region splitting module specifically includes the following steps:

[0048] The PDB of Aβ42 was split into PDB of residues 1-17, PDB of residues 16-28, and PDB of residues 27-42.

[0049] Furthermore, the α-carbon distance matrix calculation module specifically includes the following steps:

[0050] Step 4.1 Atom positioning: First, locate all α-carbon atoms from the PDB file;

[0051] Step 4.2: extracting coordinates: extracting the x, y, and z spatial coordinates of the α-carbon atom;

[0052] Step 4.3 Calculate the distance: For each pair of α-carbon atoms in the coordinates, calculate the distance between the α-carbon atoms using the Euclidean distance formula: The distance formula is

[0053]

[0054] Step 4.4: Build a matrix: Organize all the distance values ​​into a matrix, where the rows and columns of the matrix correspond to the indices of the α-carbon atoms.

[0055] Furthermore, the three-dimensional structure building module specifically includes the following steps:

[0056] Step 6.1 Initial point selection: At the beginning of the algorithm, a set of initial coordinates is randomly generated as the starting point of the optimization process. This random initialization method is often used to ensure that the optimization process can explore the solution space from a variety of possible configurations and reduce the risk of the solution being affected by the initial conditions.

[0057] Step 6.2 Optimization process: Through an iterative method, the positions of these points are continuously adjusted to minimize the objective function; the objective function measures the deviation between the α-carbon distance matrix calculated based on the current coordinates and the constraint condition, that is, the predicted α-carbon distance matrix distance map;

[0058] Interpretation of the results in step 6.3: After the optimization is completed, the obtained coordinates should be as close as possible to the actual three-dimensional structure while satisfying the given distance constraints.

[0059] The advantages of the present invention are:

[0060] This study uses deep learning to predict the protein's α-carbon distance matrix, employing a divide-and-conquer strategy. Using 2DIR, the researchers predict three distinct regions within the protein structure, ultimately successfully constructing a three-dimensional structure. This method provides a novel approach for exploring the structure and dynamics of Aβ42, opening up new perspectives for studying Alzheimer's disease and providing important information for drug design. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a method for predicting the three-dimensional structure of Aβ42 based on a deep learning method in an embodiment of the present invention;

[0062] Figure 2 This is a deep learning workflow diagram for a method for predicting the three-dimensional structure of Aβ42 based on a deep learning method in an embodiment of the present invention;

[0063] Figure 3 The results of the method in the embodiment of the present invention are shown in part (RMSD of the three figures from top to bottom are );

[0064] Figure 4 This is a network structure diagram in an embodiment of the present invention;

[0065] Figure 5 FIG. 4 is a logic principle diagram of MAE in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] This embodiment uses a method for predicting the three-dimensional structure of Aβ42 based on a deep learning method, such as Figure 1As shown, proceed as follows:

[0068] Step 1: Extract the PDB of each frame from the molecular dynamics trajectory.

[0069] To extract a PDB file for each frame from a molecular dynamics simulation trajectory, use a Python library such as MDAnalysis. First, you need to load the trajectory and topology files. Then, by iterating through each frame in the trajectory, addressing periodic boundary conditions, and saving the structure of each frame in PDB format. This process involves initializing the simulation system, selecting an appropriate force field, setting the time step, and saving the data for each frame in an appropriate output directory to facilitate further structural analysis and visualization. This approach not only automates the data extraction process but also ensures that the data is organized and stored in a format that is easy to process later.

[0070] Step 2: Calculate the 2DIR of each PDB.

[0071] Use supercomputing resources for step-by-step calculations.

[0072] Step 2.1 Molecular dynamics simulation preparation and execution. In the first step, this example configures and runs a molecular dynamics simulation to prepare dynamic data for the protein. First, configure the computing resources by loading the appropriate GROMACS module and setting up a Slurm job script. Then, create a separate directory for each PDB file and move the files to the corresponding directory. Use PDB2gmx to generate the protein topology, use editconf to set the simulation box type, solvate to add solvent, and use genion to add ions to neutralize the system charge. After completing these steps, use grompp to prepare the input files required for the simulation and submit them to the computing cluster to execute the simulation.

[0073] Step 2. Optimize the results of subsequent spectral analysis.

[0074] Step 2.3 runs 2DIR simulation and data conversion. In the 3rd step, after loading necessary compiler and MPI library, run NISE software to carry out 1DIR and 2DIR spectrum simulation. This process comprises reading the Hamiltonian and Dipoles files after adjustment, performs spectrum simulation, and the result is generated Hamiltonian and Dipoles files from text 2. In this step, present embodiment uses NISE (Nonlinear Infrared Spectroscopic Simulations) software package to process the data obtained from molecular dynamics simulation, to generate Hamiltonian (Hamiltonian) and dipole moment (Dipoles) files for two-dimensional infrared (2DIR) spectrum simulation. By copying and adjusting these files in the directory of each simulation, present embodiment guarantees that they are suitable for carrying out spectrum calculation. In order to improve statistical accuracy and the data quality of 2DIR simulation, these Hamiltonian and Dipoles files can be copied and merged many times, so that the amount of data can be increased and the format is converted into binary format for convenience of subsequent processing. This step is the key link that generates the spectral data that can be used for further analysis.

[0075] Step 2.4: Image Processing and Data Organization. The final step focuses on generating and processing the 2DIR spectral image data. A Python script is used to process the generated spectral images, including image generation, cropping, and resizing to meet analysis requirements. Finally, the processed image and data files are renamed and organized to ensure efficient storage and access, providing researchers with accurate spectral analysis results.

[0076] Step 3: Split the PDB according to functional feature areas.

[0077] Using a divide-and-conquer strategy, the PDB for Aβ42 was split into residues 1-17, 16-28, and 27-42. This is because Aβ42 can be divided into three regions: residues 1-16, 17-28, and 29-42, each with distinct structural and functional properties. The residues 1-16 region is the relatively polar N-terminus, often involved in interactions with cell membranes or water-soluble molecules, and possesses structural flexibility. The residues 17-28 region contains a critical central hydrophobic region, the hydrophobicity of which is the primary driver of Aβ42 aggregation and fibril formation, favoring the formation of a β-sheet structure. Residues 29-42 are the C-terminal region, the addition of which significantly enhances the aggregation propensity of Aβ42, making it more likely to form stable fibrils than Aβ40. Overall, the synergistic effects of these three regions contribute to the high aggregation and neurotoxicity of Aβ42 associated with Alzheimer's disease.

[0078] Step 4: Calculate the α-carbon distance map for each PDB.

[0079] The α-carbon distance matrix is ​​an N×N matrix, where N is the number of amino acid residues in a protein. Each element of the matrix, dij, represents the Euclidean distance between the α-carbon atoms of the i-th and j-th amino acid residues. The α-carbon atom is part of every amino acid residue in a protein and is located along the amino acid backbone. Calculating the distance matrix between α-carbon atoms is a common method used in protein structural analysis to assess a protein's three-dimensional structure and relative spatial arrangement.

[0080] Step 4.1 Atom Positioning. First, locate all α-carbon atoms from the PDB file. This typically involves parsing the file and searching for lines with specific atom names (e.g., "CA").

[0081] Step 4.2 Extract coordinates: Extract the x, y, and z spatial coordinates of these α-carbon atoms.

[0082] Step 4.3 Calculate the distance. For each pair of α-carbon atoms in the coordinates, use the Euclidean distance formula to calculate the distance between them. The distance formula is

[0083]

[0084] Step 4.4: Build a matrix. Organize all distance values ​​into a matrix whose rows and columns correspond to the indices of the α-carbon atoms. This distance matrix is ​​very helpful for understanding the three-dimensional structure of proteins, especially when performing structural comparisons, identifying folding patterns, or analyzing dynamic simulations.

[0085] Step 5: Predict the α-carbon distance map of each fragment based on the trained deep learning model.

[0086] According to the trained model, the 2DIR of the test set is input to predict the α-carbon distance matrix distance map of the corresponding three regions.

[0087] Step 6: Construct the three-dimensional structure of the protein.

[0088] The predicted α-carbon distance map can be considered a constraint that guides the generation of a three-dimensional structure. Specifically, this constraint defines the distances between atoms or molecules, and the optimization algorithm's task is to adjust the spatial coordinates of these atoms so that the distance matrix generated by these coordinates conforms to these constraints as closely as possible.

[0089] Step 6.1: Initial point selection. The algorithm begins by randomly generating a set of initial coordinates, which serve as the starting point for the optimization process. This random initialization method is often used to ensure that the optimization process explores the solution space from a wide range of possible configurations, reducing the risk that the solution will be affected by the initial conditions.

[0090] Step 6.2 Optimization process. The positions of these points are adjusted iteratively to minimize the objective function. The objective function here measures the deviation between the α-carbon distance matrix calculated based on the current coordinates and the constraints (i.e., the predicted α-carbon distance matrix distance map). Optimization method: The L-BFGS-B algorithm used is an efficient local search algorithm that updates coordinate positions by approximating the Hessian matrix to seek the configuration that minimizes the objective function.

[0091] Interpretation of the results in step 6.3. After the optimization is complete, the resulting coordinates should be as close as possible to the true 3D structure while satisfying the given distance constraints. However, it should be noted that due to the existence of local minima and the limitations of the optimization method, the resulting structure may not be the global optimal solution, but rather an approximate solution.

[0092] Step 7: Merge the fragments and compare with the three-dimensional structure.

[0093] The three optimized PDB segments are aligned and merged into a complete PDB (protein three-dimensional structure). Finally, the RMSD error between the predicted PDB and the real PDB is calculated. Figure 3 As shown, Figure 3 The three figures in the middle show the effects of different RMSD errors.

[0094] The working principle of this embodiment is as follows:

[0095] We used a pre-trained Swin Transformer model and fine-tuned it for this task to establish a spectrum-structure relationship between the two-dimensional infrared (2DIR) spectroscopic descriptor and the α-carbon distance map. We then constructed the three-dimensional structure using the predicted α-carbon distance map. The following is the deep learning workflow.

[0096] Step 1: Data collection and analysis.

[0097] To build a robust and transferable deep learning (DL) model, the first task is to create a comprehensive and high-quality dataset. The molecular dynamics trajectory data used in this paper are derived from published research, see the following reference for details: Heller GT, Aprile FA, Michaels TCT, et al. Small-molecule sequestration of amyloid-β as a drug discovery strategy for Alzheimer's disease [J]. Science advances, 2020, 6(45): eabb5924.

[0098] As described in the aforementioned references, an advanced integration approach was employed, combining nuclear magnetic resonance (NMR) chemical shift data with molecular dynamics simulations to precisely investigate the dynamic behavior of the monomeric state of Aβ42. Metadynamics meta-inference techniques were employed to compensate for force field inaccuracies while also incorporating errors in the experimental data and significantly enhancing the system's sampling efficiency. Simulations were performed using GROMACS 2018.3 software with the PLUMED library version 2.6.0-dev, employing the CHARMM22* force field and the TIP3P water model to ensure robust and accurate simulations.

[0099] The initial simulation step involved generating a linear conformation of Aβ42 using PyMOL software, followed by a 1 ns dynamics simulation in vacuum to collapse the extended conformation of the peptide chain. This structure was then dissolved in a volume of 362 nm. 3 The system was then minimized using the steepest descent method to achieve stability. A preliminary 2 ns simulation was then performed in the NVT ensemble at 600 K, from which 48 initial conformations were extracted for subsequent analysis. Subsequent simulations included 500 ps of equilibration in the NVT and NPT ensembles at 278 K, using a Bussi-Donadio-Parrinello thermostat and Berendsen pressure coupling, with positional constraints imposed on heavy atoms to ensure simulation accuracy.

[0100] During the production run, long simulations were performed in the NPT ensemble using a Parrinello-Rahman manometer with a fixed time step of 2 fs. LINCS constraints were applied to all bonds, while the van der Waals forces were cut off at 1.2 nm. The electron interactions were treated using the particle mesh Ewald method.

[0101] In addition, this example used CamShift software to back-calculate chemical shifts at each time step to verify the accuracy of the simulation data. Because the error of the CamShift predictor can be larger than the chemical shift perturbation after compound addition, this example used the same chemical shift constraints for the Aβ42 monomer simulation. During meta-inference, a Gaussian noise model was used with an error parameter assigned for each nucleus type, and an uninformative Jeffreys prior was used to handle the error parameter.

[0102] To enhance the conformational sampling of Aβ42, this example uses parallel biased metadynamics with moderately mild and multi-walker protocols, setting multiple collective variables (CVs). Six CVs were used in the unbound simulation, while the bound simulation included 14 CVs to enhance the contact sampling between the compound and the peptide, and four CVs to enhance the sampling of soft dihedral angles in the small molecule. These simulations were performed for 27.8 microseconds until the data converged, resulting in a total of 252,816 frames of three-dimensional structure, which were stored in the PDB. Overall, this complex series of simulation and analysis steps provides this example with detailed dynamic information about the Aβ42 monomer, which is important for understanding its biological function and possible disease mechanisms.

[0103] Finally, all PDB sequences were converted into 2DIRs and the corresponding α-carbon distance maps were calculated, resulting in a comprehensive and high-quality dataset.

[0104] Step 2: Data preprocessing.

[0105] Before the descriptors are fed into the deep neural network learning training process, the 2DIR data are preprocessed.

[0106] Step 2.1 uses the Python Pillow library to automatically crop transparent boundaries within PNG images. It first loads the image and converts it to RGBA mode. It then assumes the pixel color in the top-left corner of the image is the boundary color. It then iterates through each pixel to determine the actual boundaries of the image's main content. Finally, it crops the image based on these boundaries and saves it. By cropping transparent boundaries within the image, it reduces interference from non-critical areas, allowing the model to focus on learning the valid image content. This improves the model's ability to identify key image features and reduces noise and the risk of overfitting during training and inference.

[0107] Step 2.2 scales the large image to a smaller size. Use the resize method to resize the image to 224x224 pixels. This is the standard input size for many common deep learning models, such as convolutional neural networks. This standardization ensures that all input data has a consistent shape and size, which improves model training efficiency and performance.

[0108] Together, these processing steps optimize the quality and consistency of training data, enabling deep learning models to be trained on cleaner, more standardized datasets, thereby improving the model's generalization and prediction accuracy. In short, these image preprocessing techniques help improve model performance in real-world applications, reduce computational costs, speed up training, and improve the final model's performance.

[0109] Step 3: Model selection and construction.

[0110] The Swin Transformer model can be found as a pre-trained model in deep learning frameworks like PyTorch or TensorFlow, and can be used through libraries like timm (PyTorch Image Models).

[0111] In order to select a deep learning model suitable for this task, this example compares a group of common deep learning pre-trained models, including ResNet, U-Net, VGG, DenseNet, Inception, etc. However, these pre-trained models do not perform well.

[0112] The Swin Transformer, through its multi-level architecture, effectively extracts features at different scales, which is particularly important for spectral data such as 2DIR, which exhibits molecular vibrational information at multiple scales. This hierarchical processing helps capture spectral features from local to global perspectives, each of which may correspond to distinct interactions and functional domains in the molecular structure. By limiting the scope of self-attention to a local window, the Swin Transformer improves computational efficiency, enabling the model to process large-scale 2DIR datasets without sacrificing performance. The Swin Transformer's shifted window mechanism enhances the model's ability to integrate information across windows, allowing it to better understand local features within a global context. In the conversion from 2DIR to α-carbon distance matrices, this means the model can better correlate spectral features with spatial structural features represented in the distance matrix. Furthermore, the Swin Transformer's architecture supports inputs of various sizes, making it flexible for 2DIR spectral data of varying resolutions. Furthermore, the Swin Transformer can be pre-trained on large amounts of data to learn universal features, which can then be fine-tuned for specific tasks. This makes it particularly effective in predicting the structure of biomacromolecules with complex patterns.

[0113] Step 4: Model training and evaluation.

[0114] like Figure 4 As shown, the model first loads weights from the pre-trained Swin Transformer and uses this powerful visual Transformer model to extract low-level features from images (or other forms of input data). The Swin Transformer was chosen because it can handle large-scale visual data and is particularly suitable for complex feature extraction tasks.

[0115] The core of the model is the segmentation_head, which is responsible for further transforming the output features of SwinTransformer to meet the ultimate task of predicting the α-carbon distance matrix. This part gradually increases the size of the feature map through multiple levels of upsampling and convolution layers, ultimately reaching the required dimensions to match the protein distance matrix. The segmentation_head specifically includes:

[0116] Upsampling and convolutional layers: The upsampling method using bilinear interpolation gradually increases the size of the feature map, and convolutional layers are used after each upsampling stage to further process the features.

[0117] Batch normalization and activation function: Batch normalization and Mish activation function are used after some convolutional layers, which helps the training stability and nonlinear feature expression of the model.

[0118] Self-attention layer: Introducing the self-attention layer into the feature processing process enhances the model's ability to capture global dependencies, which is particularly important for understanding complex protein structures.

[0119] Training-related methods

[0120] Weight initialization: The init_weights method is used to initialize the network weights, especially the Kaiming initialization is used for the convolutional layer, which helps prevent the gradient from disappearing or exploding in the early stages of training.

[0121] Freezing and unfreezing of layers: The methods freeze_swin_layers and unfreeze_swin_layers are provided to allow dynamic freezing or unfreezing of specific layers of the Swin Transformer during training, which helps to improve the flexibility of model training and optimize the effect.

[0122] During the training process, the data of the features and target variables are divided into training and validation sets. These data are strictly screened and cleaned to ensure data quality and accuracy. Cross-validation is used to train the model to ensure its generalization ability to unknown data.

[0123] MAE can quantitatively determine the agreement between the predicted α-carbon distance matrix and the true α-carbon distance matrix. When it comes to predicted and true matrices, such as the contact map (distance matrix) in a predicted protein structure, MAE can be applied to each element in the matrix, such as Figure 5 shown.

[0124] Step 5: Deployment and use of the model.

[0125] Save the trained model and load it directly when you need it. Use the trained model to input the 2DIR of the protein and predict the α-carbon distance matrix of the three segments of the protein.

[0126] The key to this example is the development of a deep learning-based model for predicting the three-dimensional structure of the Aβ42 protein. This model uses 2DIR (two-dimensional infrared spectroscopy) data as input and, by learning from the structural information and spectroscopic features of a large dataset, achieves high-precision prediction of the Aβ42 protein structure.

[0127] The technology of the present invention is not limited to the structure prediction of Aβ42 protein, but can also be applied to the structure prediction of other similar proteins, including but not limited to the conformation prediction of other proteins and the simulation of protein-small molecule interactions.

[0128] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method, characterized in that: The following steps are involved: Step 1: Extract the PDB of each frame from the molecular dynamics trajectory; Step 2: Calculate the 2DIR of each PDB; Step 3: Split the PDB according to functional areas; Step 4: Calculate the α-carbon distance map of each PDB; Step 5: Predicting the α-carbon distance map for each region based on a trained deep learning model; the deep learning model uses a pre-trained Swin Transformer, which includes upsampling and convolution layers. After each upsampling stage, a convolution layer is used to further process features; batch normalization and Mish activation function are used after the convolution layer; and a self-attention layer is introduced into the feature processing flow; Step 6: Constructing the three-dimensional structure of the protein, including the following steps: Step 6.1 Selection of initial points: At the beginning of the algorithm, a set of initial coordinates is randomly generated as the starting point of the optimization process; Step 6.2 Optimization process: Through an iterative method, the starting point position is continuously adjusted to minimize the objective function; the objective function is measured as the deviation between the α-carbon distance matrix calculated based on the current coordinates and the constraint condition, that is, the predicted α-carbon distance matrix distancemap; Interpretation of the results of step 6.3: After the optimization is completed, the obtained coordinates are as close as possible to the real three-dimensional structure while satisfying the given distance constraints; Step 7: Merge the fragments and compare with the three-dimensional structure: Align the three optimized PDB fragments and merge them into a complete PDB protein three-dimensional structure; finally, calculate the RMSD error between the predicted PDB and the true PDB.

2. The method for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1: Configure and run a molecular dynamics simulation to prepare the protein dynamics data. First, configure the computing resources by loading the appropriate GROMACS modules and setting up a Slurm job script. Then, create a separate directory for each PDB file and move the files into the appropriate directory. Generate the protein topology using PDB2gmx, set the simulation box using editconf, add solvent using solvate, and add ions to neutralize the system charge using genion. After completing these steps, prepare the input files required for the simulation using grompp and submit them to the computing cluster for execution. Step 2.2 Generate Hamiltonian and Dipoles files: Use the NISE software package to process the data obtained from the molecular dynamics simulation to generate Hamiltonian and dipole moment files for 2DIR spectral simulations. These files are copied and adjusted in the directory of each simulation to ensure that they are suitable for spectral calculations. To improve the statistical accuracy and data quality of the 2DIR simulation, the Hamiltonian and Dipoles files are copied and merged multiple times to increase the data volume and optimize the results of subsequent spectral analysis. Step 2.3: Run 2DIR simulations and convert data: After loading the compiler and MPI library, run the NISE software to perform 1DIR and 2DIR spectral simulations. This process includes reading the adjusted Hamiltonian and Dipoles files, performing the spectral simulation, and converting the results from text format to binary format for subsequent processing. Step 2.4 Image processing and data organization: The generated spectral images are processed using Python scripts, including image generation, cropping, and resizing to meet analysis requirements; finally, the processed images and data files are renamed and organized to ensure that they can be efficiently stored and accessed.

3. The method for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method according to claim 1, characterized in that: The step 3 specifically includes the following steps: The PDB of Aβ42 was split into PDB of residues 1-17, PDB of residues 16-28, and PDB of residues 27-42.

4. The method for predicting the dynamic structural changes of protein Aβ42 based on a deep learning method according to claim 1, characterized in that: The step 4 specifically includes the following steps: Step 4.1 Atom positioning: First, locate all α-carbon atoms from the PDB file; Step 4.2: Extracting coordinates: extracting the x, y, and z spatial coordinates of the α-carbon atom; Step 4.3 Calculate the distance: For each pair of α-carbon atoms in the coordinates, calculate the distance between the α-carbon atoms using the Euclidean distance formula: The distance formula is Step 4.4: Build a matrix: Organize all the distance values ​​into a matrix, where the rows and columns of the matrix correspond to the indices of the α-carbon atoms.

5. A system for predicting the dynamic structural changes of protein Aβ42 based on deep learning methods, characterized by: include: PDB data extraction module, used to extract the PDB of each frame from the molecular dynamics trajectory; 2DIR calculation module, used to calculate the 2DIR of each PDB; The regional splitting module is used to split the PDB according to functional characteristics; α-carbon distance matrix calculation module, used to calculate the α-carbon distance matrix distance map of each PDB; A prediction module is used to predict the α-carbon distance matrix distancemap for each region based on a trained deep learning model; the deep learning model uses a pre-trained Swin Transformer, which includes upsampling and convolution layers. The convolution layer is used after each upsampling stage to further process the features; batch normalization and Mish activation function are used after the convolution layer; and a self-attention layer is introduced in the feature processing flow; The three-dimensional structure building module is used to construct a three-dimensional protein structure; the steps include: Step 6.1 Selection of initial points: At the beginning of the algorithm, a set of initial coordinates is randomly generated as the starting point of the optimization process; Step 6.2 Optimization process: Through an iterative method, the starting point position is continuously adjusted to minimize the objective function; the objective function is measured as the deviation between the α-carbon distance matrix calculated based on the current coordinates and the constraint condition, that is, the predicted α-carbon distance matrix distancemap; Interpretation of the results of step 6.3: After the optimization is completed, the obtained coordinates are as close as possible to the real three-dimensional structure while satisfying the given distance constraints; The error calculation module is used to merge the fragments and compare them with the three-dimensional structure: the three optimized PDB fragments are aligned and merged into a complete PDB protein three-dimensional structure; finally, the RMSD error between the predicted PDB and the true PDB is calculated.

6. The system for predicting the dynamic structural changes of protein Aβ42 based on deep learning method according to claim 5, characterized in that: The 2DIR calculation module specifically includes the following steps: Step 2.1: Configure and run a molecular dynamics simulation to prepare the protein dynamics data. First, configure the computing resources by loading the appropriate GROMACS modules and setting up a Slurm job script. Then, create a separate directory for each PDB file and move the files into the appropriate directory. Generate the protein topology using PDB2gmx, set the simulation box using editconf, add solvent using solvate, and add ions to neutralize the system charge using genion. After completing these steps, prepare the input files required for the simulation using grompp and submit them to the computing cluster for execution. Step 2.2 Generate Hamiltonian and Dipoles files: Use the NISE software package to process the data obtained from the molecular dynamics simulation to generate Hamiltonian and dipole moment files for 2DIR spectral simulations. These files are copied and adjusted in the directory of each simulation to ensure that they are suitable for spectral calculations. To improve the statistical accuracy and data quality of the 2DIR simulation, the Hamiltonian and Dipoles files are copied and merged multiple times to increase the data volume and optimize the results of subsequent spectral analysis. Step 2.3: Run 2DIR simulations and convert data: After loading the compiler and MPI library, run the NISE software to perform 1DIR and 2DIR spectral simulations. This process includes reading the adjusted Hamiltonian and Dipoles files, performing the spectral simulation, and converting the results from text format to binary format for subsequent processing. Step 2.4 Image processing and data organization: The generated spectral images are processed using Python scripts, including image generation, cropping, and resizing to meet analysis requirements; finally, the processed images and data files are renamed and organized to ensure that they can be efficiently stored and accessed.

7. The system for predicting the dynamic structural changes of protein Aβ42 based on deep learning method according to claim 5, characterized in that: The region splitting module specifically includes the following steps: The PDB of Aβ42 was split into PDB of residues 1-17, PDB of residues 16-28, and PDB of residues 27-42.

8. The system for predicting the dynamic structural changes of protein Aβ42 based on deep learning method according to claim 5, characterized in that: The α-carbon distance matrix calculation module specifically includes the following steps: Step 4.1 Atom positioning: First, locate all α-carbon atoms from the PDB file; Step 4.2: Extracting coordinates: extracting the x, y, and z spatial coordinates of the α-carbon atom; Step 4.3 Calculate the distance: For each pair of α-carbon atoms in the coordinates, calculate the distance between the α-carbon atoms using the Euclidean distance formula: The distance formula is Step 4.4: Build a matrix: Organize all the distance values ​​into a matrix, where the rows and columns of the matrix correspond to the indices of the α-carbon atoms.

Images