Method and apparatus for estimating protein aggregation of target protein

Abstract

The present invention relates to a method for estimating protein aggregation and, more particularly, to a technique using a graph convolutional network (GCN)-based machine learning model for evaluating protein aggregation by utilizing a three-dimensional structure of a protein. The method for estimating protein aggregation of a target protein, according to an embodiment of the present specification, comprises the steps of: obtaining a three-dimensional structure of the target protein; generating a graph representation on the basis of the three-dimensional structure; inputting the graph representation to a pre-trained model; and evaluating protein aggregation on the basis of an output from the pre-trained model, wherein nodes of the graph representation may correspond to amino acids of the target protein, and edges of the graph representation may correspond to interactions between the amino acids.

Description

Method and device for estimating protein aggregation of a target protein

[0001] The present invention relates to a method for estimating protein aggregation, and more particularly, to a technique for using a machine learning model based on a graph convolutional network (GCN) that evaluates protein aggregation by utilizing the three-dimensional structure of a protein.

[0002] Protein aggregation (PA) is a phenomenon in which proteins abnormally aggregate in biological systems, disrupting cellular function and causing various diseases. Protein aggregation is particularly prominent in neurodegenerative diseases such as Alzheimer's, Parkinson's, and Huntington's. Abnormal protein aggregates found in these diseases exhibit toxicity within cells, ultimately leading to neuronal cell death. Therefore, understanding and predicting protein aggregation is crucial for identifying the causes of neurological diseases and developing therapeutics.

[0003] Protein aggregation has been measured experimentally, utilizing techniques such as X-ray crystallography, NMR, and cryo-electron microscopy. These techniques provide high-resolution three-dimensional structures of proteins, making them useful for analyzing the causes of aggregation. However, these experimental methods are expensive, time-consuming, and difficult to apply to all proteins.

[0004] Recent studies have revealed that the primary sequence alone is insufficient to accurately predict aggregation. Consequently, there is a growing recognition that three-dimensional protein structure plays a crucial role in predicting aggregation. In particular, consideration of physical properties such as the distance between β-sheets, interaction energy, and structural stability in the three-dimensional structure has been shown to be effective in assessing aggregation. Predictive models based on this structural information can more accurately predict the actual aggregation potential of proteins.

[0005] However, protein 3D structure information is very complex and requires a large amount of data, so new methods are needed to effectively process and analyze the data.

[0006] The background technology described above is technical information that the inventor possessed for the purpose of deriving the present invention or acquired during the process of deriving the present invention, and cannot necessarily be said to be technology disclosed to the general public prior to the application for the present invention.

[0007] The various embodiments described herein are proposed to address the aforementioned issues. The purpose of the present invention is to provide a method and device for efficiently predicting protein aggregation based on three-dimensional structural information of a target protein. Specifically, the accuracy of aggregation assessment is enhanced by converting the protein's amino acid sequence and interactions into a graphical representation and applying it to a learning model.

[0008] According to one embodiment of the present specification for achieving the above-described task, a method for estimating protein aggregation of a target protein comprises: a step of obtaining a three-dimensional structure of the target protein; a step of generating a graph representation based on the three-dimensional structure; a step of inputting the graph representation into a pre-trained model; and a step of evaluating protein aggregation based on an output of the pre-trained model, wherein nodes of the graph representation correspond to amino acids of the target protein, and edges of the graph representation correspond to interactions between the amino acids.

[0009] The step of obtaining the three-dimensional structure may include a step of obtaining three-dimensional structure information of the target protein from the Protein Data Bank (PDB) or the AlphaFold database.

[0010] The step of generating a graph representation based on the above three-dimensional structure may include a step of generating a graph representation using amino acid sequence information of the target protein.

[0011] The step of generating the above graph representation may include a step of connecting nodes corresponding to the two amino acids with an edge when the distance between the alpha carbons of the two amino acids is less than or equal to a preset threshold.

[0012] The above threshold value may be 7 Å.

[0013] The step of generating the above graph representation may include a step of assigning to each node a feature value indicating the type of amino acid corresponding to the node.

[0014] The above pre-trained model may be a graph convolutional network (GCN) model.

[0015] The above pre-trained model may be a model trained using protein aggregation scores calculated using AGGRESCAN3D 2.0 dynamic mode.

[0016] A method for generating a model according to one embodiment of the present specification for achieving the above-described task may include: obtaining a first protein structure and first structural information corresponding to the first protein structure from a protein structure database; calculating a first property value corresponding to the first structural information based on the first structural information; generating a first property estimation model for estimating a property value from the protein structure information based on a first data set including the first protein structure and the first property value; inputting an amino acid sequence into a structure generation model to obtain a second protein structure and second structural information corresponding to the second protein structure; obtaining a second property value corresponding to the second structural information by using the second structural information as an input of the first property estimation model; verifying the second property value to obtain verification data; and generating a second property estimation model using the verification data and the first data set.

[0017] The step of obtaining the first structural information may include the step of obtaining a protein structure composed of multiple chains from the protein structure database; and the step of separating the multiple chains into each single chain.

[0018] A step of comparing amino acid sequence similarity between the above single chains and removing overlapping protein structures may be further included.

[0019] According to one embodiment of the present specification for achieving the above-described task, a computer device comprises a processor for obtaining a three-dimensional structure of a target protein, generating a graph representation based on the three-dimensional structure, inputting the graph representation into a pre-learned model, and evaluating protein cohesion based on an output of the pre-learned model, wherein nodes of the graph representation correspond to amino acids of the target protein, and edges of the graph representation may correspond to interactions between the amino acids.

[0020] According to one embodiment of the present specification for achieving the above-described task, a computer-readable recording medium includes instructions for performing the steps of: obtaining a three-dimensional structure of the target protein; generating a graph representation based on the three-dimensional structure; inputting the graph representation into a pre-trained model; and evaluating protein cohesion based on an output of the pre-trained model, wherein nodes of the graph representation may correspond to amino acids of the target protein, and edges of the graph representation may correspond to interactions between the amino acids.

[0021] According to one embodiment of the present specification for achieving the above-described task, a computer-readable recording medium includes instructions for performing the steps of: obtaining a first protein structure and first structural information corresponding to the first protein structure from a protein structure database; calculating a first property value corresponding to the first structural information based on the first structural information; generating a first property estimation model for estimating a property value from the protein structure information based on a first data set including the first protein structure and the first property value; inputting an amino acid sequence into a structure generation model to obtain a second protein structure and second structural information corresponding to the second protein structure; obtaining a second property value corresponding to the second structural information by using the second structural information as an input of the first property estimation model; verifying the second property value to obtain verification data; and generating a second property estimation model using the verification data and the first data set, wherein nodes of the graph representation may correspond to amino acids of the target protein, and edges of the graph representation may correspond to interactions between the amino acids.

[0022] According to the present invention, protein aggregation can be efficiently predicted by generating a graph based on the three-dimensional structure of the protein and the interaction between amino acids.

[0023] In addition, the present invention can provide higher accuracy and lower error rate than existing models by learning a cohesive prediction model by integrating various protein structure data.

[0024] The effects of the present invention are not limited to the effects mentioned above.

[0025] FIG. 1 is a diagram illustrating an operation of estimating protein aggregation of a computer device according to one embodiment of the present invention.

[0026] FIG. 2 is a drawing schematically illustrating a second property estimation model according to one embodiment of the present invention.

[0027] FIG. 3 is a diagram illustrating an operation of generating a learning model of a computer device according to one embodiment of the present invention.

[0028] FIG. 4 is a diagram schematically illustrating an operation of acquiring second structural information and verification data of a computer device according to one embodiment of the present invention.

[0029] Figure 5 is a diagram schematically illustrating a preprocessing process of a protein structure according to one embodiment of the present invention.

[0030] Figure 6 is a diagram schematically illustrating the structure of a learning model according to one embodiment of the present invention.

[0031] FIG. 7 is a block diagram schematically illustrating the structure of a computer device according to one embodiment of the present invention.

[0032] The present invention is capable of various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present invention, as well as methods for achieving them, will become clearer with reference to the embodiments described below in detail, along with the drawings. However, the present invention is not limited to the embodiments disclosed below and can be implemented in various forms.

[0033] Each logical block may represent a module, segment, or portion of code that includes one or more executable instructions for executing a specific logical function. It should be noted that, in one embodiment, the functions described for each block may be executed in a different order than described. For example, even if two blocks are depicted in succession, the functions described for each block may be executed substantially simultaneously, or may be executed in reverse order depending on different execution conditions or environments. In the following embodiments, the singular expression includes the plural expression unless the context clearly indicates otherwise.

[0034] In the examples below, terms such as “include” or “have” mean that a feature or component described herein is present, but do not exclude the possibility that one or more other features or components may be added.

[0035] Instructions executed by a processor of a computer or other programmable data processing device can generate means for performing each function described with reference to a flowchart or block diagram. The instructions can be installed on a computer or the like and generate processes that are executed on the computer or the like to perform a series of operational steps.

[0036] Here, the term "~unit" used in this embodiment refers to a component that performs a specific function performed by software or hardware such as an FPGA (Field Programmable Gate Array) or ASIC (Application Specific Integrated Circuit). However, the "~unit" is not limited to being performed by software or hardware. The "~unit" may exist in the form of data stored in an addressable storage medium, or one or more processors may be configured to execute a specific function.

[0037] In the drawings, the sizes of components may be exaggerated or reduced for convenience of explanation. For example, the sizes and thicknesses of each component shown in the drawings are arbitrarily shown for convenience of explanation, and therefore the present invention is not necessarily limited to what is shown. In addition, in the present disclosure, the expressions "more than" and "less than" are used to determine whether a specific condition is satisfied or fulfilled, but this is only a description to express an example and does not exclude descriptions of more than or less than. A condition described as "more than" may be replaced with "more than," a condition described as "less than" may be replaced with "less than," and a condition described as "more than and less than" may be replaced with "more than and less than." Terms such as "first" and "second" may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. Software may include a computer program, code, instructions, or a combination of one or more of these, and may configure a processing device to perform a desired operation or, independently or collectively, command the processing device. The software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave, for interpretation by the processing device or for providing instructions or data to the processing device. The software may also be distributed over networked computer systems and stored or executed in a distributed manner. The software and data may be stored on one or more computer-readable recording media.

[0038] The "learning model" referred to in the present invention may include any form of algorithm or methodology used to learn or understand specific patterns or structures from data. That is, the learning model may include not only machine learning models such as regression models, decision trees, random forests, support vector machines, K-nearest neighbors, naive Bayes, and clustering algorithms, but also deep learning models such as neural networks, convolutional neural networks, recurrent neural networks, transformer-based neural networks, generative adversarial networks (GANs), and autoencoders. The "learning model" may indicate a set of learned parameters or weights used to estimate or classify outputs for specific inputs, and this model may be trained through methods such as supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Furthermore, this may include various learning methods and structures, such as ensemble models, multi-modal models, and models using transfer learning, in addition to single models. These learning models may be pre-trained on a separate computer device from the computer device that estimates outputs for inputs and then used on another computer device.

[0039] "Estimation," as used herein, is defined as the process of calculating or determining an unknown value or outcome based on given data. In other words, "estimation" involves the act of using parameters or weights derived from a model's learning from a specific dataset to produce results for new input data. This process can be applied to a variety of problems, including regression analysis, classification, probability calculation, and time series forecasting, and its purpose is to infer or predict unknown outcome values ​​from data.

[0040] FIG. 1 is a diagram illustrating an operation of estimating protein aggregation of a computer device according to one embodiment of the present invention.

[0041] Referring to FIG. 1, the computer device can obtain a three-dimensional structure of a target protein at step S110.

[0042] Computing devices can first access reliable protein structure databases, such as the Protein Data Bank (PDB) or the AlphaFold database. These databases may contain protein structures experimentally determined through techniques such as X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, or recent cryo-electron microscopy (cryo-EM).

[0043] Computers can retrieve structural data for a target protein using its unique identifier (e.g., PDB ID). If the target protein's structure has not been experimentally determined, the computer can use structure prediction tools such as AlphaFold to predict its three-dimensional structure from its amino acid sequence.

[0044] In this process, at least one of the following information can be extracted: the position of each amino acid residue, the structure of the backbone and side chains, the position of secondary structural elements (alpha helices, beta sheets, etc.), and interactions between residues (hydrogen bonds, salt bridges, etc.).

[0045] A computer device according to one embodiment of the present invention can generate a graph representation based on a three-dimensional structure at step S120.

[0046] Computer devices can establish rules for converting protein structures into graphs. Typically, each amino acid residue is represented as a node in the graph, and interactions between amino acids are represented as edges.

[0047] A computer device can create a node for each amino acid residue. Each node can be assigned attributes such as the amino acid's type, location, and secondary structure. For example, amino acid types can be represented as a 20-dimensional one-hot encoding vector.

[0048] Computing devices can apply criteria to define interactions between amino acids. Typically, an edge is created between two amino acids when the distance between their alpha carbons is below a certain threshold, such as 7 Å. A threshold value between 5 and 9 Å can be used for amino acid embedding, with 7 Å being a more efficient value for computational resources. The computer calculates this distance for every amino acid pair and creates an edge if the threshold is met. After creating the edge, the computer can assign a weight to each edge. This weight can represent the inverse of the distance between the two amino acids or the strength of the interaction. Graphs can express not only simple connection information but also the strength of the interaction.

[0049] Once the graph is generated, the computer can convert the graph representation into a data structure. For example, the graph representation is stored in the form of an adjacency matrix or an adjacency list. In the case of an adjacency matrix, it is expressed as an N x N matrix (N is the number of amino acids), and the elements corresponding to connected amino acid pairs can have a value of 1 (or a weight value), and unconnected pairs can have a value of 0. The computer can analyze the characteristics of the generated graph. This can produce statistics such as the number of nodes, the number of edges, the average connectivity, and the clustering coefficient.

[0050] In step S130, a computer device according to one embodiment of the present invention may input a graph representation into a pre-trained model. For example, the pre-trained model may be a graph-based neural network, such as a graph convolutional network (GCN), which may have been pre-trained using protein aggregation scores calculated through the dynamic mode of AGGRESCAN3D 2.0. This pre-trained model may correspond to the second property estimation model in FIG. 2.

[0051] The computer device can transform the input data format into the format required by the model. The computer device can convert the node feature matrix and adjacency matrix into tensors and provide the tensors to the model's input layer.

[0052] Computer devices can perform forward-pass computations on a model. During this process, input data is transformed as it passes through multiple layers of the model. In graph convolution layers, the characteristics of each node are updated by aggregating information from neighboring nodes, while pooling layers can generate graph-level representations.

[0053] A computer device according to one embodiment of the present invention can evaluate protein aggregation based on the output of a pre-learned model at step S140.

[0054] The computer device can first review the model output obtained in step S130. For example, the output value may be a real number between 0 and 1, with a higher value indicating a higher probability of protein aggregation. The computer device can compare this output value to a predefined threshold. This threshold value can be set differently depending on the characteristics of the dataset or the application, and can be determined through ROC curve analysis.

[0055] Computers can link protein structural characteristics with predicted aggregation potential. For example, the computer can identify whether aggregation potential is associated with a specific region of the protein or is related to structural features.

[0056] FIG. 2 is a drawing schematically illustrating a second property estimation model according to one embodiment of the present invention.

[0057] A protein structure database (210) can store and provide three-dimensional structural information of proteins. This database can include structural data of various proteins obtained experimentally or computationally.

[0058] The protein structure database (210) may include structural information obtained through experimental methods such as X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy, etc. The protein structure database (210) may include at least one of the amino acid sequence of the protein, the three-dimensional coordinates of each amino acid, the secondary structure of the protein, binding positions, information on interacting ligands or metal ions, and positions of water molecules within the protein.

[0059] The first structural information (220) may indicate information related to the three-dimensional structure of the protein. It may be provided based on data extracted from the protein structure database (210). The first structural information (220) is information related to the three-dimensional structure of the protein and the structure, and the three-dimensional structural information is information indicating what shape the protein forms, and the structure-related information may be a concept including a primary structure such as a continuous sequence of amino acids, a secondary structure including patterns such as α-helices and β-sheets indicating how the amino acid sequence is locally folded, a tertiary structure indicating the three-dimensional shape of the entire protein and indicating how the protein folds based on various chemical interactions such as cross-links, hydrogen bonds, ionic bonds, and violent interactions between amino acids, and a quaternary structure which is information on how multiple protein subunits interact with each other to form a complex protein structure, and additional information such as ligands or metal ions connected to the protein structure, the positions of water molecules within the protein, and the thermal characteristics of the protein.

[0060] The first property estimation model (230) can designate a model that estimates the physical properties of a protein based on its structural information. The first property estimation model (230) can receive three-dimensional structural data of a protein as input and estimate various property values ​​that the protein may possess. These property values ​​may include protein stability, binding, and cohesion. The protein structure is converted into a graph format, with each amino acid represented as a node and interactions between amino acids represented as edges. This graph structure can be input into the neural network of the first property estimation model (230).

[0061] The amino acid sequence (240) can indicate the linear order of amino acids that constitute the primary structure of a protein. Proteins are composed of long polypeptide chains in which basic units called amino acids are linked together via peptide bonds.

[0062] The structure generation model (250) can estimate the three-dimensional structure of a protein based on the amino acid sequence. The structure generation model (250) calculates how a protein folds and is structured based on the interactions between amino acids and physical principles. For example, the structure generation model (250) can be AlphaFold or Rosetta. This model can estimate at least one of the structural elements, binding sites, and domains of a protein by analyzing the chemical properties of the amino acid sequence and the interactions within the sequence to find a structure that minimizes the energy of the protein. The structure generation model (250) can convert protein sequence information into a three-dimensional structure.

[0063] The physical property value (260) may indicate the physical characteristic data of the protein predicted by the first physical property estimation model (230). This physical property value (260) may correspond to the second physical property value in FIG. 3. This value is a result calculated based on the three-dimensional structural information of the protein and indicates the quantification of various characteristics of the protein. For example, the physical property value may include the stability, binding affinity, and aggregation propensity of the protein. For example, aggregation propensity indicates the tendency of proteins to clump together and form large clumps, which may be associated with diseases such as Alzheimer's. Stability indicates how well the protein structure is maintained in the external environment, and aggregation propensity may indicate how strongly the protein can bind to other molecules.

[0064] The computer device can verify the property values ​​(260) to obtain verification data (270). The verification data (270) can be obtained by a user directly conducting an experiment, calculating a related value, or from protein structure data (210). For example, the computer device can verify the aggregation rate estimated from the second protein structure using the second structural information and A3D to verify the aggregation rate. The computer device can verify the property values ​​based on the Static Mode and the Dynamic Mode. In addition, the computer device can also perform verification based on the aggregation rate to predict a property for stability or binding among the second property values.

[0065] The second property estimation model (280) is a model that estimates properties based on the results and verification data obtained through the first property estimation model (230). The computer device can generate the second property estimation model (280) based on the first property value calculated based on the three-dimensional structural information of the protein extracted from the protein structure database (210) and the protein structure, and verification data (270) that verifies the second property value estimated from the first property prediction model.

[0066] FIG. 3 is a diagram illustrating an operation of generating a learning model of a computer device according to one embodiment of the present invention.

[0067] Referring to FIG. 3, at step S310, the computer device can obtain a first protein structure and first structural information corresponding to the first protein structure from a protein structure database.

[0068] Computers can access public databases, such as the Protein Data Bank (PDB), to obtain 3D structural data on various proteins. This process can obtain protein structural information experimentally determined using X-ray crystallography, NMR spectroscopy, or cryo-electron microscopy techniques.

[0069] Computing devices can parse the acquired protein structure data to extract detailed information such as amino acid sequence, atomic coordinates, and secondary structural elements. Computing devices can collect comprehensive structural information, including the protein's primary structure (amino acid sequence), secondary structure (α-helix, β-sheet, etc.), tertiary structure (overall 3D fold), and, if available, quaternary structure (interactions between multiple subunits). Computing devices can also obtain additional metadata related to the protein structure, such as the protein's function, biological source, and experimental conditions.

[0070] For example, the computer device can also identify complex protein structures composed of multiple chains and perform preprocessing to separate the protein structures into individual single chains. During this process, the computer device can remove duplicate structures and assess the quality of the data to select only reliable structural data.

[0071] In one embodiment, the computer device can, at step S320, calculate a first material property value corresponding to the first structural information based on the first structural information.

[0072] For example, a computer device can first utilize AGGRESCAN3D (A3D) to estimate a protein's aggregation tendency. This process can be performed in two modes: static and dynamic. In static mode, the protein's 3D structure is assumed to be fixed, and the A3D score is calculated based on the position of each amino acid. In dynamic mode, CABS-flex simulations can be used to assess a more realistic aggregation tendency by taking into account protein flexibility.

[0073] The computer can derive aggregation scores for each amino acid residue and the average aggregation score for the entire protein from the A3D calculation results. Furthermore, the computer can use FoldX to calculate free energy to assess protein stability.

[0074] Computing devices can apply algorithms such as CamSol or Protein-Sol to predict protein solubility. These tools can analyze the physicochemical properties of protein surfaces to derive solubility scores. Computing devices can also perform Normal Mode Analysis (NMA) to assess protein structural flexibility.

[0075] In one embodiment, a computer device can generate, at step S330, a first property prediction model that predicts a property value from protein structure information based on a first data set according to a first protein structure and a first property value.

[0076] The computer can first convert the first protein structure data into a graph. In this process, each amino acid can be represented as a node, and interactions between amino acids can be represented as edges. The computer can manage the complexity of the graph by creating edges only when the distance between amino acids is below a certain threshold, such as 7 Å.

[0077] A computer device can assign a feature vector representing the characteristics of the corresponding amino acid to each node. This feature vector can contain various information, such as the amino acid's physicochemical properties, secondary structure, and surface exposure.

[0078] Computer devices can build deep learning models, such as graph convolutional networks (GCNs). These models can consist of multiple graph convolutional layers, pooling layers, and fully connected layers. Convolutional layers aggregate information from surrounding nodes to update the representation of each node, while pooling layers can generate graph-level representations.

[0079] The computer device can divide the first data set into a training set and a validation set for model training. It can then set hyperparameters such as batch size, learning rate, and optimizer, and define a loss function. The computer device can iteratively update the model's weights to minimize the difference between predicted and actual property values. To prevent overfitting, techniques such as early stopping, dropout, and L2 regularization can be applied. The computer device can then evaluate the performance of the trained model on a validation set and, if necessary, adjust the model structure or hyperparameters to improve performance. This process can be repeated until a specific performance target is reached.

[0080] In one embodiment, a computer device can input an amino acid sequence into a structure generation model at step S340 to obtain a second protein structure and second structural information corresponding to the second protein structure.

[0081] Computing devices can preprocess input amino acid sequences. This process can validate the sequence, convert it to a standard amino acid code, and, if necessary, normalize the sequence. For example, the computer can generate protein structures using the AlphaFold and RoseTTAFold protein structure prediction algorithms. The computer can identify secondary structural elements (α-helices, β-sheets, loops, etc.) from the generated structures and extract information about them. Furthermore, the computer can analyze additional structural features, such as electrostatic properties of the protein surface, hydrophobic patches, and active sites.

[0082]

[0083] In one embodiment, the computer device can obtain a second property value corresponding to the second structural information by using the second structural information as an input of the first property prediction model at step S350.

[0084] A computer device can process the second protein structure and corresponding second structural information obtained through the structure generation model and convert them into a form suitable for the first property prediction model. During this process, the computer device can represent the three-dimensional structure of the protein in graph form. Each amino acid residue can be represented as a node, and interactions between amino acids can be represented as edges, thereby encoding the topology of the protein structure.

[0085] The computer device can input the transformed graph data into a first property prediction model. The first property prediction model used here may be a graph convolutional network (GCN) trained in the previous step.

[0086] The computer device can obtain a second property value corresponding to the second structural information through the final output layer of the model. This property value may be a numerical value related to the stability, aggregation tendency, or function of the protein. For example, the AGGRESCAN3D aggregation tendency score can be output.

[0087] In one embodiment, a computer device can obtain verification data by verifying the second property value at step S360. The computer device can evaluate and verify the accuracy of the second property value through various methods. First, if there is an experimentally measured property value, the computer device can directly compare it with the predicted second property value. During this process, the computer device can quantify the accuracy of the prediction by calculating statistical indicators such as the mean absolute error (MAE) and the root mean square error (RMSE).

[0088] In the absence of experimental data, the computer can recalculate and compare the second property values ​​using other computational methods or simulation techniques. For example, in the case of protein aggregation prediction, the computer can cross-validate the results using other prediction tools, such as AGGRESCAN3D or FoldAmyloid. The computer can also assess the physical and chemical validity of the second property values. The computer can verify that the predicted values ​​fall within the known range of the corresponding property and do not violate the laws of physics. The computer can analyze the distribution of property values ​​for proteins with similar structures. During this process, the computer can use an outlier detection algorithm to identify statistically significant deviations in the predictions. The computer can also analyze patterns of discrepancies or errors discovered during the validation process.

[0089] In one embodiment, the computer device can generate a second property prediction model using the verification data and the first data set at step S370.

[0090] The computer device can integrate the validation data and the primary data set to construct an expanded training dataset. During this process, the computer device can check the data consistency and, if necessary, perform data normalization or scaling to ensure that all input features have the same scale. The computer device can build a second property prediction model using a graph neural network (GNN) or a neural network architecture. This model can represent the 3D structure of the protein as a graph, with each amino acid as a node and the interactions between amino acids as edges. During the training process, the computer device can optimize hyperparameters such as batch size, learning rate, and number of epochs. The computer device can find the optimal hyperparameter combination using techniques such as grid search or Bayesian optimization. The computer device can also apply regularization techniques such as early stopping, dropout, and L1 / L2 regularization to prevent overfitting. The computer device can store the second property prediction model and use it to estimate properties for new protein structures.

[0091] FIG. 4 is a diagram schematically illustrating the operation of acquiring second structural information and verification data of a computer device according to one embodiment of the present invention. The operation of the computer device in FIG. 4 may correspond to steps S340 to S360 in FIG. 3 . The computer device can select data with high information value, i.e., data that can minimize model uncertainty and have the greatest impact on improving predictive power.

[0092] Referring to FIG. 4, the computer device can determine candidate data to be selected through active learning at step S410. For example, the computer device can determine structural information, excluding proteins with a similarity of 80% or more with the first structural information, as candidate data from the second structural information.

[0093] In one embodiment, a computer device may, at step S420, evaluate the prediction uncertainty for each candidate data item. Through the uncertainty evaluation, the computer device may identify data for which the model has not yet been sufficiently trained and prioritize such data for further learning.

[0094] Computers can assess uncertainty based on the standard deviation of predicted values. After making multiple predictions for a given protein structure, the computer can calculate the variance of each prediction. A larger standard deviation indicates greater uncertainty in the data. Conversely, a smaller standard deviation indicates that the model is confident in its predictions.

[0095] Computers can assess uncertainty through the Expected Improvement (EI) value. EI is a method for predicting how much newly selected data will improve model performance compared to the maximum prediction performance to date. The computer calculates an EI value for each candidate data set and can preferentially select data with higher values ​​for further training. EI can be used to select data with a high probability of improving the model's predictive performance by combining the mean and standard deviation of the predicted values. This can be expressed by the following mathematical equation (1).

[0096] [Mathematical Formula 1]

[0097]

[0098] In mathematical formula 1 is the average of the model's predictions for data x, Standard deviation of predictions for data x, The highest performance value predicted by the model, probability density function, Indicates the cumulative distribution function.

[0099] If the predicted standard deviation (σ) is large, the data has high uncertainty and the model's performance is likely to improve through additional learning.

[0100] In one embodiment, the computer device may, at step S430, select data with high uncertainty and data likely to contribute to performance improvement through an exploration strategy that aims to improve the overall performance of the model by selecting data with high uncertainty and an utilization strategy that maximizes performance by learning more data with high performance.

[0101] In one embodiment, the computer device may retrain the first property prediction model using the selected data at step S440, or the computer device may generate a second property prediction model by integrating the selected data and the first data set to form an extended training data set.

[0102] Figure 5 is a diagram schematically illustrating a preprocessing process of a protein structure according to one embodiment of the present invention.

[0103] The operation of the computer device in FIG. 5 can correspond to step S310 in FIG. 3.

[0104] Referring to Figure 5, at step S510, the computer device can analyze and classify the initial PDB (Protein Data Bank) dataset. Specifically, the computer device can divide a total of 144,768 protein structure data into two main categories. The computer device can identify 52,845 single-chain proteins. These can indicate proteins with structures composed of a single polypeptide chain.

[0105] The computer device can classify 91,923 multi-chain proteins. These can indicate proteins whose structures are formed by the combination of two or more polypeptide chains.

[0106] A computer device can analyze the number of chains in each protein structure and assign them to categories based on the results.

[0107] In one embodiment, a computer device can separate a multi-chain protein structure into individual single chains at step S520. For example, the following operations can be performed on each of the 91,923 multi-chain proteins.

[0108] The computer device can analyze the structure file of each multi-chain protein to identify individual polypeptide chains. For each identified chain, the computer device can extract the amino acid sequence and three-dimensional coordinate information of the chain. Based on the extracted information, the computer device can reconstruct each chain into an independent single-chain protein structure. For each reconstructed single chain, the computer device can assign a unique identifier and maintain association information with the multi-chain structure. The computer device can generate a total of 270,605 new single-chain structures through this process. The generated single-chain structures can be combined with the existing 52,845 single-chain structures to form a total of 323,450 extended datasets.

[0109] According to one embodiment, a computer device can perform a task of removing redundancy in the multi-single chain data generated in step S530. Specifically, the computer device can identify and remove duplicate chains from a multi-single chain dataset of 270,605 items, thereby refining the data into 176,152 unique chains. The computer device can compare and analyze the amino acid sequences of each chain. For example, if the amino acid sequence similarity between two chains exceeds a predefined threshold, such as 80% or more, the chains can be considered duplicates and removed except for one.

[0110] Computers can use secondary or tertiary structure information to determine redundancy. By selecting only one representative chain among chains with high structural similarity, they can maintain dataset diversity.

[0111] Through the deduplication process, computational devices can effectively reduce the size of the dataset while preserving the diversity of protein structures. Consequently, after step S530, the entire dataset can be reduced to 228,997 (52,845 original single chains + 176,152 refined multi-single chains).

[0112] In one embodiment, a computer device can remove duplication between single chain data and newly generated multi-single chain data at step S540. The computer device can compare and analyze 52,845 existing single chain data and 176,152 multi-single chain data. During this process, the amino acid sequences (Seq a, Seq b, Seq c, etc.) of each protein chain can be compared in detail to evaluate similarity. The computer device can calculate the similarity between sequences using an advanced algorithm, and if the similarity exceeds a preset threshold, the data can be determined as duplicate. This duplication determination criterion can be adjusted according to the research purpose and the characteristics of the data. For data identified as duplicate, the computer device can select and preserve data with better quality or more complete information and remove the rest. During this process, the computer device can make an optimal selection by considering various factors such as data resolution, experimental method, and data reliability. As a result, the computer device can refine the multi-single chain data from 176,152 to 175,371.

[0113] In one embodiment, a computer device can calculate the aggregation tendency of each protein structure by applying the A3D algorithm to the 228,216 PDB Single chain dataset refined in step S550.

[0114] Specifically, the computer device can perform the following operations on each protein structure: The computer device can input the three-dimensional structure of each protein into the A3D algorithm and analyze it.

[0115] Using A3D's dynamic mode, the computer can simulate the flexibility and possible structural changes of each protein structure. The computer can then identify areas vulnerable to aggregation, or aggregation hotspots, within each protein structure.

[0116] The A3D algorithm allows the computer to calculate an overall aggregation score for each protein. The computer can then normalize the calculated aggregation scores to enable comparisons between proteins of different sizes and structures.

[0117] The computer device can store the calculated aggregation propensity data for each protein in a structured format. This data may include information such as protein ID, aggregation score, and aggregation hotspot location.

[0118] FIG. 6 is a diagram schematically illustrating the structure of a learning model according to one embodiment of the present invention.

[0119] The first block (510) may include a GCN (Graph Convolutional Network) layer and a batch normalization (BN) layer.

[0120] The GCN layer, which extracts 32-dimensional features, can receive protein data in a graph structure as input and generate new node representations by comprehensively considering the features of each node and information from neighboring nodes. The BN layer can normalize the output of the GCN layer to stabilize the learning process. In the first block (510), the computer device can also apply the ReLU activation function.

[0121] The second block (620) and the third block (630) may have the same structure as the first block (610).

[0122] The fourth block (640) may include a global mean pooling layer. The global mean pooling layer can transform node-level features extracted from previous GCN layers into graph-level features. This layer can calculate the average of the feature vectors of all nodes in the graph and create a single fixed-size vector.

[0123] The fourth block (640) may include a dropout layer. For example, the dropout layer ratio may be set to 0.2. Dropout may randomly disable 20% of neurons during the learning process. This mechanism can prevent overfitting of the model and improve generalization performance.

[0124] The fifth block (650) may include a first fully connected (Dense) layer. This layer can receive graph-level features extracted from the previous block as input and learn higher-level abstracted features. The fully connected layer can model complex nonlinear relationships by connecting all input nodes to all output nodes.

[0125] The fifth block (650) may include an activation function.

[0126] The fifth block (650) may include a dropout layer. For example, the dropout layer ratio may be set to 0.2. This layer is positioned after the fully connected layer to prevent overfitting and improve the model's generalization ability. By randomly disabling 20% ​​of the neurons during training, excessive reliance on specific features can be prevented.

[0127] The fifth block (650) may include a second fully connected (Dense) layer. This layer can perform final feature extraction and transformation to prepare data in a format suitable for the output layer. This layer can combine high-level features learned in previous layers to extract information necessary for the final prediction.

[0128] The sixth block (660) may include the final output layer of the learning model. The output layer may calculate a final prediction value based on features extracted and processed in previous layers.

[0129] The sixth block (660) may be composed of a fully connected (Dense) layer, and its structure and activation function may vary depending on the characteristics of the property to be predicted. For example, when predicting a continuous value such as the cohesiveness of a protein, the output layer of the sixth block (660) may be composed of a single neuron. In this case, a linear activation function can be used to directly output a continuous value. In the case of a problem of classifying a specific property of a protein, the output layer of the sixth block (660) may include as many neurons as the number of classes to be classified. In this case, the softmax activation function can be used to output the probability of belonging to each class.

[0130] A loss function is defined that calculates the difference between the output value of the 6th block (660) and the actual value, and the parameters of the model can be updated based on this.

[0131] Figure 7 is a block diagram schematically illustrating the structure of a computer device according to one embodiment of the present invention. The computer device may include a memory (710) and a processor (720). The computer device may execute one or more sets of instructions that cause any one or more of the methodologies described herein to be performed.

[0132] The memory (710) may store a set of instructions including system-related instructions and user interface-related instructions that perform any one or more of the methodologies described herein. The memory (710) temporarily or permanently stores data such as basic programs, application programs, and configuration information for the operation of the device. The memory (710) may include, but is not limited to, a non-volatile mass storage device such as a random access memory (RAM), a read only memory (ROM), and a disk drive. These software components may be loaded from a computer-readable storage medium separate from the memory (710) using a drive mechanism. The separate computer-readable storage medium may include a computer-readable storage medium such as a floppy drive, a disk, a tape, a DVD / CD-ROM drive, or a memory card. Depending on the embodiment, the software components may also be loaded into the memory (710) through a communication unit rather than a computer-readable storage medium. Additionally, the memory (710) can provide stored data upon request of the processor (720).

[0133] The processor (720) controls the overall operations of the computer device. Furthermore, the processor (720) may be configured to process computer program instructions by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (720) by the memory (710). For example, the processor (720) may be configured to execute instructions received according to program code stored in a storage device such as the memory (710). For example, the processor (720) may be controlled to perform operations according to the various embodiments described above.

[0134] According to one embodiment, a processor (720) can obtain a three-dimensional structure of a target protein, generate a graph representation based on the three-dimensional structure, input the graph representation into a pre-trained model, and evaluate protein aggregation based on an output of the pre-trained model.

[0135] According to one embodiment, a processor (720) can obtain a three-dimensional structure of a target protein. This structure can be obtained from a database such as PDB or AlphaFold. Based on the obtained three-dimensional structure, the processor can convert it into a graph. In this graph, nodes can represent individual amino acids, and edges can represent interactions between amino acids. During this process, the processor can calculate distances between amino acids and create edges between amino acids within a specific threshold (e.g., 7 Å). Furthermore, each node can be assigned an attribute representing the characteristics of the corresponding amino acid.

[0136] In one embodiment, a processor (720) may input the transformed graph into a pre-trained model. This model may primarily be a graph convolutional network (GCN) and may consist of multiple graph convolution layers and pooling layers. The processor can extract and analyze graph features through this model. Based on the model output, the processor can evaluate the cohesion of the corresponding protein. During this process, the processor can interpret the output value and compare it with a predefined threshold value to determine the cohesion potential of the protein.

[0137] According to one embodiment, a processor (720) can obtain a first protein structure and corresponding first structural information from a protein structure database. During this process, the processor can efficiently process and manage large amounts of data. The obtained structural information can be used to calculate a first property value. This property value can be calculated using AGGRESCAN3D 2.0, and the processor can perform automated calculations in conjunction with this tool. Furthermore, the processor can perform preprocessing tasks, such as separating multi-chain protein structures into single chains and removing duplicate structures.

[0138] In one embodiment, a processor (720) may generate a data set including a first protein structure and a first property value, and train a first property estimation model. This model may perform a function of estimating the property value from protein structural information. The processor may optimize the model by applying various machine learning algorithms, and may evaluate and improve the model's performance through techniques such as cross-validation. The processor (720) may receive an amino acid sequence as input and generate a second protein structure and corresponding second structural information through a structure generation model. During this process, the processor may manage linkage with external models and evaluate the quality of the generated structure. The generated information may be input into the first property estimation model to obtain a second property value. The processor may verify the obtained second property value to generate verification data. During this process, the accuracy of the property value may be evaluated through comparison with experimental methods or other computational tools. The verification data may be combined with an existing first data set to generate a more sophisticated second property estimation model. During this process, the processor (720) may adjust the model's architecture and optimize hyperparameters.

[0139] According to one embodiment, a processor (720) may obtain a first protein structure and first structural information corresponding to the first protein structure from a protein structure database, calculate a first property value corresponding to the first structural information based on the first structural information, generate a first property estimation model that estimates the property value from the protein structure information based on a first data set including the first protein structure and the first property value, input an amino acid sequence into a structure generation model to obtain a second protein structure and second structural information corresponding to the second protein structure, obtain a second property value corresponding to the second structural information by using the second structural information as an input of the first property estimation model, obtain verification data by verifying the second property value, and generate a second property estimation model using the verification data and the first data set.

[0140] The processor (720) may be implemented as a single CPU or multiple CPUs (or DSPs, SoCs). The processor (720) may be implemented as a digital signal processor (DSP), a microprocessor, or a time controller (TCON) that processes digital signals. However, the processor is not limited thereto, and may include one or more of a central processing unit (CPU), a micro controller unit (MCU), a micro processing unit (MPU), a controller, an application processor (AP), a communication processor (CP), or an ARM processor, or may be defined by the terms thereof.

[0141] Although the embodiments have been described above with limited examples and drawings, those skilled in the art will appreciate that various modifications and variations can be made based on the above teachings. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents. In other embodiments consistent with the principles of the present invention, the order of these operations may be changed. In addition, non-dependent operations may be executed in parallel.

[0142] Therefore, other implementations, embodiments, and equivalents of the claims also fall within the scope of the claims described below. That is, it will be apparent to those skilled in the art that the present invention can be extended to a method for preparing a pharmaceutical composition using a method for predicting protein aggregation.

[0143] For example, a method for preparing a pharmaceutical composition utilizing the protein aggregation estimation method of the present invention may be considered. Such a method may include the following steps: obtaining a three-dimensional structure of a target protein, generating a graph representation based on the three-dimensional structure, inputting the graph representation into a pre-trained model, evaluating protein aggregation based on the output of the pre-trained model, and identifying a pharmaceutical composition that inhibits a protein with high aggregation potential associated with neurodegenerative diseases such as Alzheimer's disease, Parkinson's disease, and Huntington's disease.

Claims

1. In a method for estimating protein aggregation of a target protein, A step of obtaining a three-dimensional structure of the target protein; A step of generating a graph representation based on the above three-dimensional structure; A step of inputting the above graph representation into a pre-trained model; and A step of evaluating protein aggregation based on the output of the above pre-learned model, A method for estimating protein cohesion, characterized in that the nodes of the graph representation correspond to amino acids of the target protein, and the edges of the graph representation correspond to interactions between the amino acids.

2. In paragraph 1, The step of obtaining the above three-dimensional structure is: A method for estimating protein aggregation, characterized by comprising a step of obtaining three-dimensional structural information of the target protein from the Protein Data Bank (PDB) or the AlphaFold database.

3. In paragraph 1, The step of generating a graph representation based on the above 3D structure is A method for estimating protein cohesion, characterized by comprising a step of generating a graph representation using amino acid sequence information of the target protein.

4. In paragraph 1, The steps of generating the above graph representation are: A method for estimating protein cohesion, characterized in that it comprises a step of connecting nodes corresponding to two amino acids with edges when the distance between alpha carbons of the two amino acids is less than a preset threshold value.

5. In paragraph 4, A method for estimating protein aggregation, characterized in that the threshold value is 7 Å.

6. In paragraph 1, The steps of generating the above graph representation are: A method for estimating protein cohesion, characterized in that it includes a step of assigning to each node a feature value indicating the type of amino acid corresponding to the node.

7. In paragraph 1, A method for estimating protein aggregation, characterized in that the above pre-learned model is a graph convolutional network (GCN) model.

8. In paragraph 1, A method for estimating protein aggregation, characterized in that the above pre-learned model is a model learned using a protein aggregation score calculated using AGGRESCAN3D 2.0 dynamic mode.

9. A method for generating a model for estimating the properties of a protein, A step of obtaining a first protein structure and first structural information corresponding to the first protein structure from a protein structure database; A step of calculating a first property value corresponding to the first structural information based on the first structural information; A step of generating a first property estimation model that estimates a property value from protein structure information based on a first data set including the first protein structure and the first property value; A step of inputting an amino acid sequence into a structure generation model to obtain a second protein structure and second structural information corresponding to the second protein structure; A step of obtaining a second property value corresponding to the second structural information by using the second structural information as an input to the first property estimation model; A step of obtaining verification data by verifying the second property value; and A method for generating a protein property estimation model, characterized in that it comprises a step of generating a second property estimation model using the verification data and the first data set.

10. In paragraph 9, The step of obtaining the above first structural information is: A step of obtaining a protein structure composed of multiple chains from the above protein structure database; and A method for generating a protein property estimation model, characterized in that it comprises a step of separating the above multi-chains into each single chain.

11. In paragraph 10, A method for generating a protein property estimation model, characterized in that it further comprises a step of removing overlapping protein structures by comparing amino acid sequence similarity between the above single chains.

12. In computer devices, A processor comprising: a processor for obtaining a three-dimensional structure of a target protein, generating a graph representation based on the three-dimensional structure, inputting the graph representation into a pre-trained model, and evaluating protein aggregation based on the output of the pre-trained model; A computer device characterized in that the nodes of the graph representation correspond to amino acids of the target protein, and the edges of the graph representation correspond to interactions between the amino acids.

13. In a computer-readable recording medium, A computer-readable recording medium recording a program for executing the method of claim 1.

14. In a computer-readable recording medium, A computer-readable recording medium recording a program for executing the method of Article 9.

Images