Method for predicting shape of DNA structure on basis of deep learning, and apparatus therefor
The hybrid learning algorithm for DNA structure prediction models, using graph neural networks and data augmentation, addresses the low accuracy of existing deep learning models by integrating data-based and physics-informed methods, achieving precise and rapid prediction of DNA structure shapes.
Patent Information
- Application Number
- PCT/KR2024/000450
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-19
- Filing Date
- 2024-01-10
- Publication Date
- 2024-11-28
AI Technical Summary
Existing deep learning models, such as AlphaFold and RoseTTAFold, face low prediction accuracy for DNA structure shapes due to the complexity and large number of complementary base pairs in DNA structures, which are challenging to predict accurately compared to proteins.
A hybrid learning algorithm-based method that combines data-based loss and physical information-based loss functions to learn a DNA structure shape prediction model, using a graph neural network to generate and predict DNA structure graphs from blueprints, and applying data augmentation techniques to improve prediction accuracy.
This approach enables accurate and rapid prediction of DNA structure shapes, including complex DNA origami structures, with high precision in real-time, effectively addressing the limitations of existing models by integrating both data-driven and physics-informed learning.
Smart Images

Figure KR2024000450_28112024_PF_FP_ABST
Abstract
Description
Method and device for predicting the shape of a DNA structure based on deep learning
[0001] The present invention relates to a method and a device for predicting the shape of a DNA structure based on deep learning, and more specifically, to a method and a device for predicting the shape of a DNA structure based on learning a neural network model based on a hybrid learning algorithm.
[0002] Artificial intelligence and deep learning technologies are dramatically shifting the paradigm in the analysis and design of molecular structures. Deep learning learns from vast amounts of data, and when new data is input, it uses the learned data to select the most probabilistic answer.
[0003] Deep learning can adaptively operate depending on the image, and because it automatically finds characteristic factors in the process of learning a model based on data, there has been an increasing trend of attempts to utilize it in the field of artificial intelligence.
[0004] AlphaFold and RosettaFold are deep learning models that predict protein sequence, structure, and function, accelerating the process of designing new proteins. Research is also underway to apply these models to accelerate the process of designing DNA structures.
[0005] However, DNA structures are made on a nanoscale by synthesizing tens to hundreds of complementary DNA strands, and have a resolution of several nanometers. Unlike proteins, these DNA structures (especially DNA origami structures) can have complex shapes, and the number of complementary base pairs is very large, in the thousands to tens of thousands, so there is a problem that the prediction accuracy is very low with data-based learning used by existing deep learning models such as AlphaFold and RosettaFold.
[0006] The purpose of the present invention is to provide a DNA structure shape prediction model and an electronic device thereof learned by a hybrid learning algorithm based on a loss function according to data-based loss and physical information-based loss.
[0007] The purpose of the present invention is to provide a method and device for predicting a DNA structure shape corresponding to a DNA structure design drawing using a learned DNA structure shape prediction model.
[0008] According to one embodiment of the present invention, a method for predicting the shape of a DNA structure based on deep learning can be provided, comprising: (a) obtaining a DNA structure graph composed of nodes and edges based on a DNA structure design drawing; and (b) outputting a DNA structure shape corresponding to the DNA structure design drawing based on the DNA structure graph and a DNA structure shape prediction model.
[0009] Here, the DNA structure shape prediction model is configured to include two or more DNA structure shape prediction models, and the step (b) can obtain two or more preliminary DNA structure shapes for the input of the DNA structure graph for each of the two or more DNA structure shape prediction models, and output the preliminary DNA structure shape having the lowest energy among the two or more preliminary DNA structure shapes as the DNA structure shape.
[0010] Here, the DNA structure shape prediction model may be generated through the steps of: (c) obtaining a plurality of learning DNA structure blueprints and generating a learning DNA structure graph corresponding to each of the plurality of acquired DNA structure blueprints; (d) obtaining a correct answer DNA structure shape corresponding to the plurality of learning DNA structure blueprints based on a structure simulation program; and (e) learning a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs and the correct answer DNA structure.
[0011] Here, the step (c) includes: (f) a step of performing data augmentation based on insertion of at least one base pair or deletion of at least one base pair for at least some of the acquired plurality of learning DNA structure designs to obtain a learning augmentation design; and (g) a step of generating a learning augmentation structure graph corresponding to each of the learning augmentation structure designs; and the step (e) can learn the previously stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs, the learning augmentation structure graph, and the correct answer DNA structure.
[0012] Here, the step (e) can learn the previously stored DNA structure shape prediction model by applying different loss functions to learning for the learning DNA structure graph and learning for the learning augmented structure graph.
[0013] Here, the step (e) can perform learning of the pre-stored DNA structure shape prediction model for the learning DNA structure graph by applying a first loss function based on data-based loss and physical information-based loss.
[0014] Here, the step (e) can perform learning of the stored DNA structure shape prediction model for the learning augmented structure graph by applying a second loss function based on physical information-based loss.
[0015] Here, the DNA structure shape prediction model includes one first DNA structure shape prediction model that learns the plurality of learning DNA structure blueprints as one class, and two or more second DNA structure shape prediction models that learn the first DNA structure shape prediction model for each of two or more detailed classes, and each of the two or more detailed classes may have the plurality of learning DNA structure blueprints classified and included according to the two or more detailed classes.
[0016] According to various embodiments of the present invention, a device for predicting the shape of a DNA structure based on deep learning can be provided, including: a storage unit storing a DNA structure shape prediction model; a graph generation unit obtaining a DNA structure graph composed of nodes and edges based on a DNA structure design drawing; and a structure prediction unit outputting a DNA structure shape corresponding to the DNA structure design drawing based on the DNA structure graph and a previously learned DNA structure shape prediction model.
[0017] Here, the DNA structure shape prediction model is configured to include two or more DNA structure shape prediction models, and the structure prediction unit can obtain two or more preliminary DNA structure shapes for the input of the DNA structure graph for each of the two or more DNA structure shape prediction models included in the DNA structure shape prediction model, and determine the preliminary DNA structure shape having the lowest energy among the two or more preliminary DNA structure shapes as the DNA structure shape.
[0018] Here, the device for predicting the shape of a DNA structure based on deep learning further includes a blueprint acquisition unit; a correct answer shape acquisition unit; and a learning processing unit; wherein the DNA structure shape prediction model is generated by the blueprint acquisition unit acquiring a plurality of learning DNA structure blueprints, the graph acquisition unit generating a learning DNA structure graph corresponding to each of the plurality of acquired DNA structure blueprints, the correct answer shape acquisition unit acquiring a correct answer DNA structure shape corresponding to the plurality of learning DNA structure blueprints based on a structure simulation program, and the learning processing unit learning a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs and the correct answer DNA structure.
[0019] Here, the blueprint acquisition unit performs data augmentation based on insertion of at least one base pair or deletion of at least one base pair for at least some of the acquired plurality of learning DNA structure blueprints to acquire a learning augmented blueprint, the graph acquisition unit generates a learning augmented structure graph corresponding to each of the learning augmented structure blueprints, and the learning processing unit can learn a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs, the learning augmented structure graph, and the correct answer DNA structure.
[0020] Here, the learning processing unit can learn the previously stored DNA structure shape prediction model by applying different loss functions to learning for the learning DNA structure graph and learning for the learning augmented structure graph.
[0021] Here, the learning processing unit can perform learning of the previously stored DNA structure shape prediction model for the learning DNA structure graph by applying a first loss function based on data-based loss and physical information-based loss.
[0022] Here, the learning processing unit can perform learning of the previously stored DNA structure shape prediction model for the learning augmented structure graph by applying a second loss function based on physical information-based loss.
[0023] Here, the learning processing unit generates the DNA structure shape prediction model by including one first DNA structure shape prediction model that has learned the plurality of learning DNA structure design drawings as one class, and two or more second DNA structure shape prediction models that have learned the first DNA structure shape prediction model for each of two or more detailed classes, and the two or more detailed classes may be those in which the plurality of learning DNA structure design drawings are distinguished and included according to the two or more detailed classes.
[0024] According to various embodiments of the present invention, there is provided a DNA structure shape prediction model learned by a hybrid learning algorithm based on a loss function according to data-based loss and physical information-based loss, and a device including the same, thereby enabling the shape of a DNA origami structure to be predicted quickly and accurately.
[0025] According to various embodiments of the present invention, there is provided an ensemble model including a plurality of shape prediction models classified and learned according to a DNA structure, and a device including the same, thereby enabling the shape of a monomer DNA origami structure to be predicted with high precision in almost real time.
[0026] According to various embodiments of the present invention, by providing an ensemble model based on a hybrid learning algorithm that combines a data-based learning method and a physical information-based learning method, and a device including the same, it is possible to predict a DNA structure of a complex shape quickly and accurately in almost real time, which has the effect of accurately predicting the shape of bio, nano, and macro structures of various scales where data is insufficient.
[0027] Figure 1 is a conceptual diagram schematically illustrating the configuration of a device according to one embodiment of the present invention.
[0028] FIG. 2 is a block diagram illustrating the configuration of a storage unit in which a shape prediction model is stored and a processing unit and storage unit that perform functions based on the shape prediction model in a device according to one embodiment of the present invention, divided by function.
[0029] FIG. 3 is a conceptual diagram schematically illustrating an operation of generating a DNA structure graph based on a DNA structure design diagram in a device according to one embodiment of the present invention.
[0030] FIG. 4 is a conceptual diagram showing a schematic configuration and operation of a shape prediction model in a device according to one embodiment of the present invention.
[0031] FIG. 5 is a conceptual diagram schematically illustrating an operation of performing learning of a shape prediction model based on a hybrid learning algorithm in a device according to an embodiment of the present invention.
[0032] FIG. 6 is a diagram illustrating a shape of a DNA structure predicted by a shape prediction model learned in a manner different from that learned by a hybrid learning algorithm in a device according to one embodiment of the present invention.
[0033] FIG. 7 is a graph showing shape scores of DNA structure shapes predicted by a shape prediction model learned in a manner different from that learned by a hybrid learning algorithm in a device according to one embodiment of the present invention.
[0034] FIG. 8 is a graph showing the energy levels of DNA structure shapes predicted by a shape prediction model learned in a manner different from that learned by a hybrid learning algorithm in a device according to one embodiment of the present invention.
[0035] FIG. 9 schematically illustrates the configuration of an ensemble model generated based on shape prediction models learned for each of a plurality of classes in a device according to one embodiment of the present invention.
[0036] FIG. 10 is a diagram comparing the structural shapes predicted by the ensemble model and SNUPI corresponding to the supramolecular assembly in a device according to one embodiment of the present invention.
[0037] FIG. 11 is a graph showing shape scores of structure shapes predicted by an ensemble model and SNUPI corresponding to a supramolecular assembly in a device according to one embodiment of the present invention.
[0038] FIG. 12 is a graph showing the energy levels of structural shapes predicted by the ensemble model and SNUPI corresponding to the supramolecular assembly in a device according to one embodiment of the present invention.
[0039] FIG. 13 is a conceptual diagram schematically illustrating the configuration of an unsupervised self-refining function of an ensemble model in a device according to one embodiment of the present invention.
[0040] FIG. 14 is a diagram illustrating structural shapes predicted by an ensemble model including an unsupervised self-refining function in response to a supramolecular assembly in a device according to one embodiment of the present invention.
[0041] FIG. 15 is a graph showing the energy levels of structural shapes predicted by an ensemble model including an unsupervised self-refining function in response to a supramolecular assembly in a device according to one embodiment of the present invention.
[0042] FIG. 16 is a flowchart illustrating the flow of operations for predicting a structure shape using a shape prediction model in a device according to one embodiment of the present invention.
[0043] FIG. 17 is a flowchart illustrating the flow of operations for learning a shape prediction model configured to predict a structure shape in a device according to one embodiment of the present invention.
[0044] Hereinafter, embodiments are described in detail with reference to the attached drawings. However, the embodiments may be modified in various ways, and the scope of the patent application is not limited or restricted by these embodiments. It should be understood that all modifications, equivalents, or alternatives to the embodiments are included within the scope of the patent application.
[0045] The terms used in the examples are for the purpose of description only and should not be construed as limiting. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this specification, the terms "comprises" or "has" and the like are intended to specify the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but should be understood to not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.
[0046] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art to which the embodiments pertain. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and shall not be interpreted in an idealized or overly formal sense unless explicitly defined herein.
[0047] In addition, when describing with reference to the attached drawings, identical components will be assigned the same reference numerals regardless of the drawing numbers, and redundant descriptions thereof will be omitted. When describing embodiments, if a detailed description of a related known technology is judged to unnecessarily obscure the gist of the embodiment, the detailed description will be omitted.
[0048] Additionally, terms such as first, second, A, B, (a), (b), etc. may be used to describe components of the embodiment. These terms are only intended to distinguish the components from other components, and the nature, order, or sequence of the components are not limited by the terms.
[0049] When a component is described as being "connected," "coupled," or "connected" to another component, it should be understood that the component may be directly connected or connected to that other component, but that another component may also be "connected," "coupled," or "connected" between the components.
[0050] Additionally, when it is described that a component is "connected" or "connected" to another component by communication, it should be understood that the component may be connected or connected to that other component by wireless or wired communication, but that other components may be "connected" or "connected" between the components.
[0051] Components included in one embodiment and components with common functions will be described using the same names in other embodiments. Unless otherwise stated, the descriptions given in one embodiment may also apply to other embodiments, and detailed descriptions will be omitted to the extent of overlap.
[0052] The 'data' processed by the device and / or the device can be expressed as the term 'information'. Here, information can be used as a concept that includes data.
[0053] The present invention relates to a method and a device for predicting the shape of a DNA structure based on deep learning. More specifically, the present invention provides a method for predicting the shape of a multidimensional (e.g., two-dimensional, three-dimensional, etc.) DNA structure corresponding to a DNA structure design drawing based on a DNA structure shape prediction model based on deep learning, a method for learning a DNA structure shape prediction model, and a device for the same.
[0054] Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. The drawings attached to this specification, together with the detailed description of the invention, serve to further understand the technical concepts of the present invention. Therefore, the present invention should not be interpreted as being limited to the matters described in the drawings.
[0055] Below, with reference to Fig. 1, the configuration of the device can be roughly explained.
[0056] Figure 1 is a conceptual diagram schematically illustrating the configuration of a device according to one embodiment of the present invention.
[0057] Referring to FIG. 1, the device (100) may include a processing unit (110), a storage unit (120), and a communication unit (130).
[0058] The processing unit (110) is configured to include at least one processor (or controller) and can process control commands related to operations performed through the device (100) through at least one program (app, application, tool, plug-in, software, etc.).
[0059] For example, the processing unit (110) can learn a shape prediction model through a shape prediction program and predict a DNA structure shape corresponding to the acquired DNA structure design drawing. At this time, the shape prediction program can be stored in the storage unit (120) of the device (100) and / or in the storage unit of a device connected to the device (100).
[0060] According to various embodiments of the present invention, a shape prediction program may be configured to include at least one deep learning-based neural network model. More specifically, the shape prediction program may be configured to include a shape prediction model (or a DNA structure shape prediction model) based on a graph neural network (GNN).
[0061] At this time, the shape prediction model may be a neural network model (e.g., a DNA origami graph neural network model) configured to predict the shape of a DNA structure corresponding to a DNA structure design in a DNA origami (or origami) manner.
[0062] The shape prediction program may be configured to include at least a portion of a program (and / or algorithm) for performing operations described through embodiments of the present invention, such as obtaining a DNA structure blueprint, generating a DNA structure graph based on the DNA structure blueprint, learning a shape prediction model based on the DNA structure graph, and predicting a DNA structure shape based on the DNA structure graph.
[0063] The processing unit (110) can share data processing and / or processing results with at least one device (e.g., user device) connected to the device (100) through a shape prediction model.
[0064] Hereinafter, in various embodiments, the device (100) performing an operation according to a control command may be understood as performing a specified operation through at least one program and / or a shape prediction program related to at least one control command processing of the device (100).
[0065] Here, the control command processing is described as being performed through at least one program or shape prediction program installed in the device (100), but is not limited thereto, and may also be provided through another program or a temporarily installed program already installed in the storage unit (120).
[0066] According to one embodiment, the control command processing may be performed by the processing unit (110) based on at least a portion of a database provided free of charge or for a fee from an external device connected to the device (100).
[0067] The operation of the device (100) is performed based on data processing and device control of the processing unit (110), and the processing unit (110) may perform a designated function based on a control command received through the input / output unit or communication unit of the device (100).
[0068] In addition, the processing unit (110) may process data acquired through the communication unit (130) based on the identified user. For example, the processing unit (110) may identify a user based on user information received from a user device connected through the communication unit (130) and / or user information acquired through an input unit (or input / output unit) connected to the device (100), and perform an operation according to a control command input by the identified user.
[0069] The storage unit (120) can store various data processed by at least one component of the device (100) (e.g., the processing unit (110) or the communication unit (130). The data can include, for example, a program for processing a control command, data processed through the program, and / or input data and output data related thereto.
[0070] The storage unit (120) may include an artificial neural network algorithm for processing control commands, a blockchain algorithm, a deep learning algorithm, a regression analysis algorithm, and an artificial intelligence algorithm based on at least some of the mechanisms, operators, language models, and big data related thereto.
[0071] The shape prediction program or shape prediction model stored in the storage unit (120) may be configured to include at least some of the artificial intelligence algorithms described above.
[0072] The shape prediction program may include at least a part of an artificial intelligence algorithm for generating a DNA structure graph composed of nodes and edges based on the acquired DNA structure design drawing, and / or may be connected to an artificial intelligence algorithm for the same.
[0073] Additionally, the artificial intelligence algorithm that constitutes the shape prediction program can be divided into multiple neural network models.
[0074] For example, an artificial intelligence algorithm for a shape prediction program may be configured to include an algorithm based on the attention mechanism of a transformer layer. More preferably, the artificial intelligence algorithm for a shape prediction program may be configured to include an algorithm based on a multi-head attention mechanism.
[0075] Hereinafter, in describing various embodiments of the present invention, the operation of the shape prediction model should be understood as the operation of the shape prediction model or the operation of a shape prediction program including the shape prediction model.
[0076] The storage unit (120) may include at least one DNA structure design diagram for use in predicting the shape of the DNA structure.
[0077] To explain in more detail, the storage unit (120) may store at least a portion of the learning DNA structure design for learning the shape prediction model and the shape of the DNA structure for the correct answer. In this case, the learning DNA structure design and the shape of the DNA structure for the correct answer may be stored as a data set for learning the shape prediction model.
[0078] The storage unit (120) may include data for confirming and processing specified control and operations through signals received through each device included in the input / output unit and / or communication unit (130).
[0079] Operations performed in relation to the storage unit (120) are processed by the processing unit (110), and data for processing the related operations, data being processed, processed data, preset data, etc. can be stored in the storage unit (120) as a database.
[0080] Data stored in the storage unit (120) can be changed, modified, deleted, and / or new data can be created by the processing unit (110) based on user input from an identified user.
[0081] The storage unit (120) can store device setting information of the device (100). The device setting information can be setting information for the device (100) and at least some of the functions and services provided by the device (100).
[0082] The storage unit (120) may store user information (or user account) for at least one user. The user information may include at least some of user identification information (e.g., ID), a password, and user-customized setting information. The user-customized setting information may be setting information for at least some of the control rights and / or functions of the device (100), and may be set and stored based on administrator input.
[0083] Here, at least some of the user information can be stored as data of a shape prediction program.
[0084] The storage unit (120) may be configured to include volatile memory, non-volatile memory, and / or a computer-readable recording medium. In this case, the computer-readable recording medium may store a computer program for causing the device (100) to perform operations according to embodiments of the present invention based on various embodiments of the present invention.
[0085] The communication unit (130) can support the establishment of a wired communication channel between the device (100) and at least one other device (e.g., a user device or an administrator device capable of controlling the device (100), the establishment of a wireless communication channel, and the performance of communication through the established communication channel.
[0086] The communication unit (130) can perform operations such as modulation / demodulation and encryption / decryption during the communication process, which are obvious to those skilled in the art, so further explanation will be omitted.
[0087] The communication unit (130) may be operated dependently or independently of the processing unit (110) and may include one or more communication processors that support wireless communication and / or wired communication.
[0088] According to one embodiment, when the communication unit (130) supports wireless communication, it may include at least some communication modules among wireless communication modules, for example, a cellular communication module, a short-range wireless communication module, and a GNSS (global navigation satellite system) communication module.
[0089] When the communication unit (130) supports wired communication, it may include at least some communication modules among wired communication modules, for example, a LAN (local area network) communication module, a power line communication module, a CAN (controller area network) communication module, and an SPI (serial peripheral interface) communication module.
[0090] To be more specific, the communication unit (130) can communicate with an external device wired and / or wirelessly via a short-range communication network such as Bluetooth, BLE (Bluetooth Low Energy), WiFi, WiFi direct, IrDA (infrared data association), ZigBee, UWB, RF (Radio Frequency) and / or a long-range communication network such as a cellular network, the Internet, or a computer network (e.g., a LAN or WAN).
[0091] The various types of communication modules that make up the communication unit (130) may be integrated into one component (e.g., a single chip) or implemented as multiple separate components (e.g., multiple chips).
[0092] Referring to FIG. 1, the communication unit (130) is illustrated as performing communication with an external device of the device (100), but is not limited thereto, and may perform communication with, for example, the storage unit (120) and / or at least some of the various components of the device (100) that are not configured in FIG. 1.
[0093] In addition, although not shown in FIG. 1, the device (100) may further include an input / output unit (not shown).
[0094] The input / output unit may be configured to include an input unit and an output unit. The input / output unit may be configured to include at least some of an input unit (not shown) for inputting data, such as a keyboard, mouse, or microphone, and an output unit (not shown) for outputting data, such as a driving unit, a display unit (e.g., a display), or a speaker.
[0095] At this time, the display unit may be configured as a touch screen capable of touch input.
[0096] According to various embodiments of the present invention, the device (100) or user device may include at least a portion of the functions of all information and communication devices, including mobile terminals, multimedia terminals, wired terminals, fixed terminals, and internet protocol (IP) terminals.
[0097] The device (100) may be configured to include at least some functions of a workstation or a large-scale database as a device for processing control commands, or to be connected to at least some of a workstation or a large-scale database through communication.
[0098] As a device (100) or a user device connected to the device (100), examples thereof include a mobile phone, a personal computer (PC), a portable multimedia player (PMP), a mobile internet device (MID), a smartphone, a tablet PC, a phablet PC, a notebook, etc.
[0099] According to various embodiments of the present invention, the user device may be described as a device that connects and communicates with the device (100). For example, although not illustrated in the drawings, the user device may be a user's smartphone that is connected to the device (100) via wireless communication and transmits user input so that the device (100) performs operations according to embodiments of the present invention.
[0100] The user device may be connected to the device (100) through at least one installed program, or may be connected to the device (100) through at least one online webpage provided by the device (100).
[0101] To this end, the device (100) may be configured to include at least some of the functions of a server and / or a terminal for providing a DNA structure shape prediction service.
[0102] FIG. 2 is a block diagram illustrating the configuration of a storage unit in which a shape prediction model is stored and a processing unit and storage unit that perform functions based on the shape prediction model in a device according to one embodiment of the present invention, divided by function.
[0103] According to various embodiments of the present invention, as described above, the device (100) may include a shape prediction program that performs DNA structure shape prediction and learns a shape prediction model for DNA structure shape prediction.
[0104] According to various embodiments of the present invention, the shape prediction program of the device (100) may be configured to include a shape prediction model based on a graph neural network (GNN).
[0105] That is, the device (100) can predict a DNA structure shape corresponding to a DNA structure design drawing obtained through a previously learned shape prediction model, and the accuracy of the DNA structure shape prediction of the shape prediction program of the device (100) can be improved through learning of the shape prediction model.
[0106] First, in the device (100), the storage unit (120) storing data related to the operation of the processing unit (110) that predicts the shape of a DNA structure using a shape prediction model or learns a shape prediction model can be configured to include a data set (221) and a shape prediction model (223) as described above.
[0107] In addition, the processing unit (110) that predicts the shape of a DNA structure or learns a shape prediction model based on the data set (221) and the shape prediction model (223) stored in the storage unit (120) in the device (100) may be configured to include at least one element among a blueprint acquisition unit (201), a graph generation unit (203), a shape acquisition unit for correct answer (205), a learning processing unit (207), and a structure prediction unit (209).
[0108] First, according to one embodiment of a data set (221) for learning a shape prediction model (223), the data set (221) may be configured to include a plurality of DNA structure blueprints. In this case, the plurality of DNA structure blueprints may be divided into a learning data set and a test data set at a specified ratio.
[0109] At least some of the DNA structure designs included in the data set (221) may be labeled. More preferably, each of the DNA structure designs included in the data set (221) may be labeled.
[0110] At this time, each DNA structure design included in the data set (221) may be labeled with DNA structure shape information, for example, three-dimensional coordinates of atoms. Here, the structure shape information may be configured to include information based on molecular dynamics simulation.
[0111] According to various embodiments, the data set (221) may further include an augmented structure design generated based on data augmentation. In this case, the augmented structure design may be stored in an unlabeled state.
[0112] Here, the operation of generating a blueprint for an augmented structure can be performed through a blueprint acquisition unit (201).
[0113] The blueprint acquisition unit (201) can acquire an augmented blueprint for learning by performing data augmentation based on insertion of at least one base pair or deletion of at least one base pair for at least a portion of a DNA structure blueprint stored in a data set.
[0114] Here, the blueprint acquisition unit (201) can perform base pair insertion or base pair deletion for the DNA structure blueprint to satisfy an energy change within a preset value.
[0115] In addition, the data set (221) may be configured to include at least one correct answer DNA structure shape corresponding to a learning DNA structure design or a learning augmentation structure design.
[0116] According to one embodiment, the DNA structure shape for the correct answer may be generated based on a learning DNA structure graph. To this end, the graph generation unit (203) may generate a DNA structure graph or an augmented structure graph based on a DNA structure design drawing or an augmented structure design drawing.
[0117] To explain in more detail, the DNA structure graph processed as input of the shape prediction model (223) can be composed of a graph including nodes and edges corresponding to the base pairs and the binding relationships of the base pairs of the DNA structure design diagram, as illustrated in FIG. 3.
[0118] FIG. 3 is a conceptual diagram schematically illustrating an operation of generating a DNA structure graph based on a DNA structure design diagram in a device according to one embodiment of the present invention.
[0119] Referring to FIG. 3, in constructing a DNA structure graph, the graph generation unit (203) can model base pairs included in a DNA structure design diagram, for example, a learning DNA structure design diagram of a data set (221), as nodes with information about position and direction.
[0120] According to one embodiment, the number of base pairs of the learning DNA structure may be comprised of between 1,000 and 17,000. However, the number of base pairs constituting the learning DNA structure is not limited thereto, and may be designed to vary.
[0121] The graph generation unit (203) can model a structural edge representing a covalent bond between base pairs and an electrostatic edge representing an electrostatic interaction due to the negative charge of DNA, based on the bonding relationship of base pairs included in the DNA structure design diagram.
[0122] Here, the structural edge may be configured to include unique geometric information and mechanical property information according to the base, and the electrostatic edge may be configured to include information on the repulsive force (or a distance function for the repulsive force) formed between base pairs.
[0123] The graph generation unit (203) can generate a DNA structure graph corresponding to the DNA structure design based on modeling of nodes and edges.
[0124] As described above, the graph generation unit (203) can generate a DNA structure graph corresponding to a DNA structure design drawing, and can also generate an augmented structure graph corresponding to an augmented structure design drawing.
[0125] Afterwards, the correct answer shape acquisition unit (205) can acquire the correct answer DNA structure shape based on the generated DNA structure graph.
[0126] Here, the correct answer DNA structure shape may be a DNA structure shape in a three-dimensional equilibrium state (or a final configuration in the equilibrium state, ground-truth, GT)) generated based on at least one simulation program (e.g., Structured NUcleic-acids Programming Interface, SNUPI) that simulates the DNA structure shape based on a multi-scale physics-based model.
[0127] According to one embodiment, at least one program for simulating the shape of a DNA structure may be stored in the storage (120), or may be stored external to the device (100), for example, on a server.
[0128] When at least one program simulating the shape of a DNA structure is stored on the server, the correct answer shape acquisition unit (205) can transmit a learning DNA structure graph (or learning DNA structure design drawing) to the server and receive a correct answer DNA structure shape corresponding to the learning DNA structure graph (or learning DNA structure design drawing) from the server.
[0129] In addition, the shape acquisition unit (205) for correct answer can distinguish detailed classes for a learning DNA structure design drawing or a learning DNA structure graph corresponding to the learning DNA structure design drawing.
[0130] For example, the correct answer shape acquisition unit (205) can distinguish detailed classes for the correct answer DNA structure design drawing or the correct answer DNA structure graph based on labeled information in the correct answer DNA structure design drawing or the correct answer DNA structure shape.
[0131] According to one embodiment, the detailed class may be composed of six detailed classes: block structure, curved structure, twisted structure, hinged structure, two-dimensional wireframe structure, and three-dimensional wireframe structure.
[0132] In addition, the shape acquisition unit (205) for correct answer can set the learning DNA structure design drawings or learning DNA structure graphs included in the data set (221) as one class (or one detailed class).
[0133] The learning processing unit (207) can perform learning of a shape prediction model (233) based on a data set (221) configured as described above.
[0134] The shape prediction model (223), as shown in FIG. 4, may be configured as a neural network model that predicts the predicted DNA structure shape (the shape of the DNA structure in a three-dimensional equilibrium state or the equilibrium configuration) based on the DNA structure graph (or the initial configuration). FIG. 4 is a conceptual diagram schematically illustrating the configuration and operation of the shape prediction model in a device according to an embodiment of the present invention.
[0135] A shape prediction model (223) that predicts the shape of a DNA structure using a DNA structure graph as input can be configured to include a mechanical relaxation block, an electromagnetic refinement block, and a residual block.
[0136] Referring to FIG. 4, it is illustrated as being composed of five mechanical relaxation blocks and one electrostatic purification block, but the number of mechanical relaxation blocks or the number of electrostatic purification blocks may be varied.
[0137] In one embodiment of the mechanical relaxation block, the mechanical relaxation block may be an element designed to find a structural configuration that minimizes mechanical strain energy based on the structure used in ResNet or DenseNet.
[0138] According to one embodiment of the electrostatic refinement block, the structure used in the transformer model includes a convolutional layer and an activation function ReLU (rectified linear unit), and may be an element designed to refine the configuration of a mechanically relaxed structure by reflecting electrostatic repulsion.
[0139] According to one embodiment of the residual block, it may be an element designed to implement skip connections based on the structure used in ResNet, and to solve the vanishing gradient problem occurring in the mechanical relaxation block while taking charge of the connection between the mechanical relaxation block and the electrostatic refinement block.
[0140] For the shape prediction model (223), a DNA structure graph (or values of the DNA structure graph) composed of nodes and edges is processed as input, and the predicted DNA structure shape can be processed as output.
[0141] In mechanical relaxation blocks, the positions and orientations of nodes are updated locally and globally in residual blocks that proceed using graph neural layers, and connections between layers can be configured to prevent over-smoothing.
[0142] The prediction results of local displacement and rotation angle in each of the mechanical relaxation block, electrostatic refinement block, and residual block can be configured to be propagated to the entire graph based on the multi-head attention mechanism of the transformer layer.
[0143] The positions and orientations of nodes constituting the structure between the residual blocks are configured to be incrementally updated in gated recurrent unit cells, which can support capturing long-term dependencies between layers for structures with relatively large deformations.
[0144] As described above, the configuration information of the mechanically relaxed structure can be processed as input to the electrostatic purification block.
[0145] The electrostatic refinement block can be configured to generate electrostatic edges between adjacent nodes within a distance of several nanometers (nm), more specifically, 0 to 3 nm, and more preferably 2.5 nm. The configuration of the final DNA structure shape can be predicted by performing iterative calculations between the electrostatic edges via the electrostatic refinement block.
[0146] The learning processing unit (207) performs learning of the shape prediction model (223) configured as described above, and at this time, as illustrated in FIG. 5, learning of the shape prediction model (223) can be performed based on a hybrid data-driven and physics-informed training algorithm (hereinafter, hybrid learning algorithm).
[0147] FIG. 5 is a conceptual diagram schematically illustrating an operation of performing learning of a shape prediction model (223) based on a hybrid data-based and physical information-based learning algorithm in a device according to one embodiment of the present invention.
[0148] The learning processing unit (207) can perform learning on a shape prediction model (223) based on a learning DNA structure graph, a learning DNA augmentation graph, and a correct answer DNA structure shape.
[0149] At this time, the learning processing unit (207) can perform learning of the shape prediction model (223) using a loss function based on data-based loss (e.g., data driven loss) for the learning DNA structure graph.
[0150] More preferably, the learning processing unit (207) can perform learning of the shape prediction model (223) using a loss function based on data-based loss and physical information-based loss (e.g., physics informed loss) for the learning DNA structure graph.
[0151] In addition, the learning processing unit (207) can perform learning of a shape prediction model (223) using a loss function based on data-based loss or physical information-based loss for a learning DNA structure graph.
[0152] Here, the loss function based on data-based loss (data-based loss function) may be a function of the mean square error of the position and orientation of the nodes between the predicted DNA structure shape (equilibrium configuration) and the correct DNA structure shape based on the shape prediction model (223).
[0153] Additionally, the loss function based on the physical information-based loss (physical information-based loss function) may be a function of the mechanical configuration and energy difference between the predicted DNA structure shape (equilibrium configuration) and the correct DNA structure shape based on the shape prediction model (223).
[0154] The learning processing unit (207), in performing learning of the shape prediction model (223), performs learning to minimize a loss function evaluated for a design diagram (or graph) used in learning, and for this purpose, an optimizer (e.g., Adam or AdamW, etc.) having a step decay learning rate scheduler can be used to optimize the parameters of the shape prediction model (223) (or the GNN constituting the shape prediction model).
[0155] As described above, the performance of the shape prediction model (223) learned based on the hybrid learning algorithm can be explained with reference to FIGS. 6 to 8.
[0156] FIG. 6 is a diagram illustrating a DNA structure shape predicted by a shape prediction model learned in a different manner from a shape prediction model learned in accordance with a hybrid learning algorithm in a device according to an embodiment of the present invention. FIG. 7 is a graph illustrating shape scores of DNA structure shapes predicted by a shape prediction model learned in a different manner from a shape prediction model learned in accordance with a hybrid learning algorithm in a device according to an embodiment of the present invention. FIG. 8 is a graph illustrating energy levels of DNA structure shapes predicted by a shape prediction model learned in a different manner from a shape prediction model learned in accordance with a hybrid learning algorithm in a device according to an embodiment of the present invention.
[0157] According to FIGS. 6 to 8, a shape prediction model (223) according to an embodiment of the present invention uses a data set consisting of a total of 300 DNA structure design drawings including 252 learning DNA structure design drawings and 48 test structure design drawings, and a model learned based on a hybrid learning algorithm, and derives a result by distinguishing between a DNA structure shape (hybrid, w / augmentation) predicted using a shape prediction model (223) (w / augmentation) that has performed data augmentation (w / augmentation) and a DNA structure shape (hybrid, w / o augmentation) predicted using a shape prediction model (223) (w / o augmentation) that has not performed data augmentation (w / o augmentation).
[0158] Here, when performing data augmentation, the design acquisition unit (201) generated two learning augmentation designs for each of the 252 learning DNA structure designs.
[0159] In addition, as described above, in the case of the labeled blueprint (the DNA structure blueprint for learning), both the data-based loss and the physical information-based loss were used for learning in the learning processing unit (207), but in the case of the augmented blueprint for learning, only the physical information-based loss was used for learning.
[0160] The DNA structure shapes compared here include DNA structure shapes predicted using a shape prediction model learned through a data-driven approach (data driven), DNA structure shapes predicted using a shape prediction model learned through a physics-informed approach (physics informed), and DNA structure shapes predicted based on SNUPI (GT).
[0161] The DNA structure shapes shown through FIGS. 6 to 8 are based on the results obtained by processing 48 test DNA structure graphs as input for each shape prediction model of DNA structure shape (GT), DNA structure shape (hybrid, w / augmentation), DNA structure shape (hybrid, w / o augmentation), DNA structure shape (data driven), and DNA structure shape (physics informed).
[0162] In Fig. 6, the values in parentheses for each of the illustrated DNA structure shape (GT), DNA structure shape (hybrid, w / augmentation), DNA structure shape (hybrid, w / o augmentation), DNA structure shape (data driven), and DNA structure shape (physics informed) represent the root mean square deviation (RMSD) and the ratio of total energy for the DNA structure shape (GT). Here, the scale bar is defined as 10 nm.
[0163] Here, RMSD can represent the root mean square distance between the nodes of the DNA structure shape and the DNA structure shape (GT).
[0164] Figure 7 is a graph showing the shape scores, for example, RMSD and OS (orientation score), for each of DNA structure shape (GT), DNA structure shape (hybrid, w / augmentation), DNA structure shape (hybrid, w / o augmentation), DNA structure shape (data driven), and DNA structure shape (physics informed).
[0165] Here, OS can represent the direction score defined as the inner product of the direction vectors of each node of the DNA structure shape.
[0166] According to Fig. 7, it can be confirmed that the hybrid learning algorithm of the shape prediction model (223) (w / augmentation) that predicted the DNA structure shape (hybrid, w / augmentation) has the highest performance.
[0167] For example, it can be confirmed that the shape prediction model (223) (w / augmentation) that predicted the DNA structure shape (hybrid, w / augmentation) predicted the DNA structure shape with the lowest RMSD and the highest OS value, thereby predicting the DNA structure shape closest to the DNA structure shape (GT) value.
[0168] Figure 8 is a graph showing the values of energy levels, for example, mechanical energy and electrostatic energy, for each of DNA structure shape (GT), DNA structure shape (hybrid, w / augmentation), DNA structure shape (hybrid, w / o augmentation), DNA structure shape (data driven), and DNA structure shape (physics informed).
[0169] According to Fig. 8, it can be confirmed that the hybrid learning algorithm of the shape prediction model (223) (w / augmentation) that predicted the DNA structure shape (hybrid, w / augmentation) shows the energy level closest to the DNA structure shape (GT).
[0170] On the other hand, it may not be easy to accurately predict the shape of a DNA structure even with a shape prediction model based on data augmentation and hybrid learning algorithms, given that the structural features of labeled DNA structure blueprints (or DNA structure graphs) are very diverse and distinct, and the number of blueprints with similar features is limited.
[0171] Accordingly, the learning processing unit (207) can apply an ensemble approach that combines multiple shape prediction models. At this time, the learning processing unit (207) can learn a separate, independent shape prediction model for each class by using a data set (221) in which the learning DNA structure design diagram (or learning DNA structure graph) is divided according to the class.
[0172] To explain in more detail, as described above, when the learning DNA structure design included in the data set (221) is divided into six detailed classes, the learning processing unit (207) can independently learn the shape prediction model (223) for each of the six detailed classes and learn the shape prediction model (223) for the entire data set (221) as one class (or detailed class), thereby generating a total of seven learned shape prediction models.
[0173] The learning processing unit (207) can configure an ensemble shape prediction model (hereinafter, ensemble model) including a total of seven learned shape prediction models.
[0174] At this time, the ensemble model can be configured to obtain DNA structure shapes (equilibrium configurations) predicted by each of the seven learned shape prediction models corresponding to the input initial configuration, as illustrated in FIG. 9, and determine the structure shape with the lowest energy as the final DNA structure shape.
[0175] As described above, the ensemble model is described as being composed of seven shape prediction models, but is not limited thereto, and the number of shape prediction models constituting the ensemble model may be changed based on the number of detailed classes.
[0176] FIG. 9 schematically illustrates the configuration of an ensemble model generated based on shape prediction models learned for each of a plurality of classes in a device according to one embodiment of the present invention.
[0177] In the operation of the structure prediction unit (209) predicting the DNA structure shape for a specific input DNA structure design using an ensemble model, the DNA structure shape with the lowest energy can be output as the final result.
[0178] Referring to FIG. 9, the structure prediction unit (209) can determine and output the DNA structure shape of the CS shape with the lowest energy among the plurality of preliminary structure shapes obtained as a result of processing the ensemble model in response to a specific DNA structure design drawing as the final shape.
[0179] Here, the learning processing unit (207) can perform learning of each shape prediction model based on a hybrid learning algorithm when learning multiple shape prediction models to configure an ensemble model.
[0180] As described above, the learning processing unit (207) can perform prediction of a supramolecular structure, such as a polyhedron, composed of a plurality (e.g., tens to hundreds) of DNA origami monomers using an ensemble model.
[0181] At this time, the graph generation unit (203) can generate graph information for a hierarchically assembled large assembly by introducing interface edges for interaction at the interface between single structures.
[0182] Through FIGS. 10 to 12, a single DNA structure predicted using an ensemble model according to an embodiment of the present invention can be compared with a single DBA structure (GT) predicted using SNUPI.
[0183] FIG. 10 is a diagram comparing structure shapes predicted by an ensemble model and SNUPI for a single DNA structure in a device according to an embodiment of the present invention. FIG. 11 is a graph showing shape scores of structure shapes predicted by an ensemble model and SNUPI for a single DNA structure in a device according to an embodiment of the present invention. FIG. 12 is a graph showing energy levels of structure shapes predicted by an ensemble model and SNUPI for a single DNA structure in a device according to an embodiment of the present invention.
[0184] Referring to Fig. 10, the single DNA structure shape (colored) predicted by the ensemble model was in good agreement with the shape (gray) predicted by SNUPI, and it can be confirmed that all types of deformations, including block structure, curved structure, twisted structure, hinged structure, two-dimensional wireframe structure, and three-dimensional wireframe structure distortion, were successfully predicted.
[0185] Referring to Fig. 11, the inference time of the ensemble model for predicting the shape of a DNA structure based on a design diagram of a DNA structure for testing was measured to be in the range of 0.55 to 1.47 seconds, with an average of 0.92 seconds, which was a near real-time prediction.
[0186] The RMSD and OS values were found to have an average of 2.97 nm and 0.97, and the standard deviations were 2.59 and 0.04, respectively. These results indicate that the ensemble model can accurately estimate both the overall DNA structure shape and the local DNA configuration from the input structural information.
[0187] Furthermore, as illustrated in Figure 12, the energy levels predicted by the ensemble model are shown to be very close to the energy levels of the GT structure. The ratio of the total energy of the initial configuration to the GT structure averaged 103.5, and significantly decreased to 1.4 for the predicted equilibrium configuration.
[0188] This significant energy reduction was observed for all types of deformations, including block, curved, twisted, hinged, two-dimensional wireframe, and three-dimensional wireframe structures. In summary, the ensemble model according to the present invention can provide more accurate predictions of shape and energy levels than the pre-trained model, even for highly deformed structures.
[0189] In addition, the ensemble model can be configured to perform learning of the ensemble model or shape prediction models constituting the ensemble model based on the unsupervised learning function of the GNN model. For example, the ensemble model can be configured to evaluate the physical information of the graph based on the GNN model. In this regard, referring to FIGS. 13 to 15, FIG. 13 is a conceptual diagram schematically illustrating the configuration of the unsupervised self-refinement function of the ensemble model in a device according to an embodiment of the present invention. FIG. 14 is a diagram illustrating structure shapes predicted in response to a supramolecular assembly by an ensemble model including an unsupervised self-refinement function in a device according to an embodiment of the present invention. FIG. 15 is a graph showing the energy levels of structure shapes predicted in response to a supramolecular assembly by an ensemble model including an unsupervised self-refinement function in a device according to an embodiment of the present invention.
[0190] Referring to Figure 13, the initial configuration for the supramolecular assembly having a tripod structure was generated using several DNA structure geometries predicted by the ensemble model, and the ensemble model can be configured to find the configuration with the lowest energy among various variations of the initial configuration.
[0191] According to one embodiment, the ensemble model can be configured to predict the supramolecular structure shape corresponding to the supramolecular assembly, by predicting DNA structure shapes corresponding to DNA structure designs constituting the supramolecular assembly, and determining the configuration with the minimized energy of the predicted DNA structure shapes as the supramolecular structure shape.
[0192] As described above, the ensemble model can learn to minimize energy and exclude data-based loss in the operation of predicting the shape of a supramolecular structure through the operation of predicting the shape of a DNA structure.
[0193] That is, such unsupervised self-refinement functions can be configured to learn the parameters of the ensemble model to find the energy-minimizing configuration based on the structural shapes predicted by the ensemble model.
[0194] Figure 14 depicts dozens to hundreds of DNA origami structures (V-brick, connector, and triangle structures) predicted by an ensemble model with self-purification capabilities, along with the larger structures connecting them. The radii of the predicted polyhedra (tetrahedron: 131 nm, hexahedron: 151 nm, dodecahedron: 217 nm) are in good agreement with those of the experiments (tetrahedron: 125 nm, hexahedron: 150 nm, dodecahedron: 210 nm). The protein GroEL is shown for size comparison.
[0195] Referring to Figure 15, it can be confirmed that the ensemble model including the unsupervised self-refinement function can reduce the total energy of the supramolecular assembly and predict the exact shape (or geometry) of the supramolecular structure.
[0196] Hereinafter, the flow of operations for predicting the shape of a DNA structure using an ensemble model and for learning the ensemble model in a device according to an embodiment of the present invention can be described.
[0197] FIG. 16 is a flowchart illustrating the flow of operations for predicting a structure shape using a shape prediction model in a device according to one embodiment of the present invention.
[0198] At step 1601, the graph generation unit (203) can obtain a DNA structure graph composed of nodes and edges based on the DNA structure design diagram.
[0199] Here, the DNA structure blueprint may be a DNA structure blueprint obtained through a blueprint acquisition unit (201). In addition, the structure blueprint may be a structure blueprint for a supramolecular assembly.
[0200] The graph generation unit (203) models nodes and edges based on the base pairs and the bonding relationship of the base pairs included in the DNA structure design drawing, and can generate a DNA structure graph including nodes and edges.
[0201] Here, nodes are configured to include information about the location and orientation of base pairs, and edges can be configured to include structural edges representing covalent bonds between base pairs and electrostatic edges representing electrostatic interactions due to the negative charge of DNA.
[0202] At step 1603, the structure prediction unit (209) can output a DNA structure shape corresponding to the DNA structure design drawing based on the DNA structure graph and the learned DNA structure shape prediction model (shape prediction model).
[0203] Here, the DNA structure shape prediction model may be configured as an ensemble model including two or more DNA structure shape prediction models.
[0204] To be more specific, the ensemble model may include one first DNA structure shape prediction model that has learned the plurality of learning DNA structure blueprints as one class, and two or more second DNA structure shape prediction models that have learned the first DNA structure shape prediction model for each of two or more detailed classes. Here, each of the two or more detailed classes may have the plurality of learning DNA structure blueprints classified and included according to the two or more detailed classes.
[0205] According to one embodiment, subclasses may be distinguished based on the structure of the DNA structure. For example, subclasses may be distinguished based on at least one structure among a block structure, a curved structure, a twisted structure, a hinged structure, a two-dimensional wireframe structure, and a three-dimensional wireframe structure.
[0206] The structure prediction unit (209) can obtain two or more preliminary DNA structure shapes for the input of the DNA structure graph for each of two or more DNA structure shape prediction models included in the DNA structure shape prediction model. The structure prediction unit (209) can determine the preliminary DNA structure shape with the lowest energy among the two or more preliminary DNA structure shapes as the DNA structure shape.
[0207] FIG. 17 is a flowchart illustrating the flow of operations for learning a shape prediction model configured to predict a structure shape in a device according to one embodiment of the present invention.
[0208] In step 1701, the blueprint acquisition unit (201) acquires a plurality of learning DNA structure blueprints, and the graph generation unit (203) can generate a learning DNA structure graph corresponding to each of the plurality of acquired DNA structure blueprints.
[0209] Here, the blueprint acquisition unit (201) can acquire a learning augmented blueprint by performing data augmentation based on at least one base pair insertion or at least one base pair deletion for at least some of the acquired plurality of learning DNA structure blueprints.
[0210] Thereafter, the graph generation unit (203) can generate a learning augmented structure graph corresponding to each of the learning augmented structure design drawings.
[0211] At step 1703, the correct answer shape acquisition unit (205) can acquire the correct answer DNA structure shape corresponding to the plurality of learning DNA structure design drawings based on a structure simulation program (e.g., SNUIPI).
[0212] At this time, if the structure simulation program is configured outside of the device (100) (e.g., server), the correct answer shape acquisition unit (205) can receive the correct answer DNA structure shape corresponding to the plurality of learning DNA structure design drawings from the server.
[0213] At step 1705, the learning processing unit (207) can learn a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs and the correct answer DNA structure.
[0214] Here, as described above, when a learning augmented structure graph is generated, the learning processing unit (207) can learn the previously stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs, the learning augmented structure graph, and the correct answer DNA structure.
[0215] The learning processing unit (207) can learn the previously stored DNA structure shape prediction model by applying different loss functions to learning the learning DNA structure graph and learning the learning augmented structure graph.
[0216] For example, the learning processing unit (207) can learn a DNA structure shape prediction model by applying a loss function based on at least some of a data-based loss function and a physical information-based loss function to learning a DNA structure graph for learning or learning an augmented structure graph for learning.
[0217] To explain in more detail, the learning processing unit (207) can perform learning of the previously stored DNA structure shape prediction model for the learning DNA structure graph by applying a first loss function based on data-based loss and physical information-based loss.
[0218] In addition, the learning processing unit (207) can perform learning of the previously stored DNA structure shape prediction model for the learning augmented structure graph by applying at least some of the first loss function and the second loss function based on physical information-based loss.
[0219] As described above, according to various embodiments of the present invention, there is provided a DNA structure shape prediction model learned by a hybrid learning algorithm based on a loss function according to data-based loss and physical information-based loss, and a device including the same, thereby having the effect of quickly and accurately predicting the shape of a DNA origami structure.
[0220] According to various embodiments of the present invention, there is provided an ensemble model including a plurality of shape prediction models classified and learned according to a DNA structure, and a device including the same, thereby enabling the shape of a monomer DNA origami structure to be predicted with high precision in almost real time.
[0221] According to various embodiments of the present invention, by providing an ensemble model based on a hybrid learning algorithm that combines a data-based learning method and a physical information-based learning method, and a device including the same, it is possible to predict a DNA structure of a complex shape quickly and accurately in almost real time, which has the effect of accurately predicting the shape of bio, nano, and macro structures of various scales where data is insufficient.
[0222] In the present invention, when a function performed by a device (100) is described through various embodiments, but a component performing the function is not described as a component of the device (100), it should be understood that the device (100) includes and / or is connected to a commonly known component performing the function.
[0223] Although the embodiments have been described with limited drawings as described above, those skilled in the art can apply various technical modifications and variations based on the various embodiments.
[0224] For example, suitable results may 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 by other components or equivalents.
[0225] In particular, when explaining with reference to a flow chart, it is described as comprising multiple steps and executing the steps sequentially in a specified order, but it is not necessarily limited to the described order.
[0226] In other words, it is also possible to apply the embodiment by changing or deleting at least some of the steps described in the flowchart and executing them, or by adding at least one step, and it is also possible to apply the embodiment by executing one or more steps in parallel. In other words, the steps are not necessarily limited to operating in a chronological order, and should be included in the embodiment of the present invention.
[0227] Therefore, other implementations, other embodiments, and equivalents to the claims should also fall within the scope of the claims described below.
Claims
1. A method for predicting the shape of a DNA structure based on deep learning, (a) a step of obtaining a DNA structure graph composed of nodes and edges based on a DNA structure design drawing; (b) a method for predicting the shape of a DNA structure based on deep learning, comprising a step of outputting a DNA structure shape corresponding to the DNA structure design drawing based on the DNA structure graph and the DNA structure shape prediction model.
2. In paragraph 1, The above DNA structure shape prediction model is configured to include two or more DNA structure shape prediction models, The step (b) above obtains two or more preliminary DNA structure shapes for the input of the DNA structure graph for each of the two or more DNA structure shape prediction models, A method for predicting the shape of a DNA structure based on deep learning, wherein the preliminary DNA structure shape having the lowest energy among the two or more preliminary DNA structure shapes is output as the DNA structure shape.
3. In paragraph 1, The above DNA structure shape prediction model is, (c) a step of obtaining a plurality of learning DNA structure blueprints and generating a learning DNA structure graph corresponding to each of the obtained plurality of DNA structure blueprints; (d) a step of obtaining a correct answer DNA structure shape corresponding to the plurality of learning DNA structure designs based on a structure simulation program; and (e) A method for predicting the shape of a DNA structure based on deep learning, wherein the method comprises: a step of learning a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs and the correct answer DNA structure; 4. In paragraph 3, Step (c) above, (f) a step of performing data augmentation based on at least one base pair insertion or at least one base pair deletion for at least some of the acquired plurality of learning DNA structure designs to obtain a learning augmentation design; and (g) a step of generating a learning augmented structure graph corresponding to each of the learning augmented structure design drawings; The above step (e) is a method for predicting the shape of a DNA structure based on deep learning, wherein the previously stored DNA structure shape prediction model is learned based on the plurality of learning DNA structure graphs, the learning augmented structure graphs, and the correct answer DNA structure.
5. In paragraph 4, The above step (e) is a method for predicting the shape of a DNA structure based on deep learning, wherein different loss functions are applied to learning the learning DNA structure graph and learning the learning augmented structure graph to learn the previously stored DNA structure shape prediction model.
6. In paragraph 4, A method for predicting the shape of a DNA structure based on deep learning, wherein the step (e) performs learning of the pre-stored DNA structure shape prediction model for the learning DNA structure graph by applying a first loss function based on data-based loss and physical information-based loss.
7. In paragraph 4, A method for predicting the shape of a DNA structure based on deep learning, wherein the step (e) performs learning of the previously stored DNA structure shape prediction model for the learning augmented structure graph by applying a second loss function based on a physical information-based loss.
8. In paragraph 3, The above DNA structure shape prediction model is, It includes one first DNA structure shape prediction model that has learned the plurality of learning DNA structure design drawings as one class, and two or more second DNA structure shape prediction models that have learned the first DNA structure shape prediction model for each of two or more detailed classes, A method for predicting the shape of a DNA structure based on deep learning, wherein each of the two or more detailed classes is divided and included according to the two or more detailed classes.
9. In a device for predicting the shape of a DNA structure based on deep learning, A storage unit in which a DNA structure shape prediction model is stored; A graph generation unit that obtains a DNA structure graph composed of nodes and edges based on a DNA structure blueprint; A device for predicting the shape of a DNA structure based on deep learning, comprising a structure prediction unit that outputs a DNA structure shape corresponding to the DNA structure design drawing based on the DNA structure graph and a previously learned DNA structure shape prediction model.
10. In paragraph 9, The above DNA structure shape prediction model is configured to include two or more DNA structure shape prediction models, The above structure prediction unit obtains two or more preliminary DNA structure shapes for the input of the DNA structure graph for each of two or more DNA structure shape prediction models included in the DNA structure shape prediction model, A device for predicting the shape of a DNA structure based on deep learning, wherein the preliminary DNA structure shape having the lowest energy among the two or more preliminary DNA structure shapes is determined as the DNA structure shape.
11. In paragraph 10, Blueprint acquisition department; Correct answer shape acquisition unit; and further comprising a learning processing unit; The above DNA structure shape prediction model is, The above blueprint acquisition unit acquires a plurality of learning DNA structure blueprints, The above graph acquisition unit generates a learning DNA structure graph corresponding to each of the acquired plurality of DNA structure blueprints, The correct answer shape acquisition unit acquires the correct answer DNA structure shape corresponding to the plurality of learning DNA structure designs based on the structure simulation program, and The above learning processing unit is a device for predicting the shape of a DNA structure based on deep learning, wherein the device is created by learning a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs and the correct answer DNA structure.
12. In paragraph 11, The above blueprint acquisition unit performs data augmentation based on insertion of at least one base pair or deletion of at least one base pair for at least some of the acquired plurality of learning DNA structure blueprints to acquire a learning augmentation blueprint, The above graph acquisition unit generates a learning augmented structure graph corresponding to each of the learning augmented structure design drawings, and The above learning processing unit is a device for predicting the shape of a DNA structure based on deep learning, which learns a pre-stored DNA structure shape prediction model based on the plurality of learning DNA structure graphs, the learning augmented structure graph, and the correct answer DNA structure.
13. In paragraph 12, The above learning processing unit is a device for predicting the shape of a DNA structure based on deep learning, which learns the previously stored DNA structure shape prediction model by applying different loss functions to learning the learning DNA structure graph and learning the learning augmented structure graph.
14. In paragraph 12, A device for predicting the shape of a DNA structure based on deep learning, wherein the learning processing unit performs learning of the previously stored DNA structure shape prediction model for the learning DNA structure graph by applying a first loss function based on data-based loss and physical information-based loss.
15. In paragraph 12, A device for predicting the shape of a DNA structure based on deep learning, wherein the learning processing unit performs learning of the previously stored DNA structure shape prediction model for the learning augmented structure graph by applying a second loss function based on physical information-based loss.
16. In paragraph 12, The learning processing unit generates the DNA structure shape prediction model by including one first DNA structure shape prediction model that has learned the plurality of learning DNA structure design drawings as one class, and two or more second DNA structure shape prediction models that have learned the first DNA structure shape prediction model for each of two or more detailed classes, A device for predicting the shape of a DNA structure based on deep learning, wherein the two or more detailed classes are divided and included in the plurality of learning DNA structure blueprints according to the two or more detailed classes.
Citation Information
Patent Citations
Machine learning for determining protein structures
WO2020058174A1
Predicting protein structures using protein graphs
WO2022112255A1
Cyclic peptide structure prediction via structural ensembles achieved by molecular dynamics and machine learning
WO2022266626A1
