Estimation program, estimation method, and information processing device

Abstract

To accurately estimate the plausible continuous deformation of a three-dimensional atomic model of a polymer. [Solution] The information processing device 100, in the process of training an autoencoder having an encoder and decoder using multiple training data sets in which polymer particle images are explanatory variables and a polymer 3D density map is the target variable, obtains the distribution of latent variables output from the encoder. The information processing device 100 generates a path related to deformation based on the distribution of latent variables. The information processing device 100 selects multiple neighboring latent variables from among the multiple latent variables generated by inputting multiple particle images into the trained encoder, the distance from the path being less than a threshold. The information processing device 100 selects multiple neighboring particle images from among the multiple particle images that correspond to the multiple neighboring latent variables. The information processing device 100 estimates multiple 3D atomic models based on the multiple neighboring particle images.

Description

[Technical Field]

[0001] This invention relates to estimation programs, etc. [Background technology]

[0002] CryoEM (Cryogenic Electron Microscopy) is used to improve the efficiency of drug discovery and other related processes. CryoEM is a device (or technique) for observing samples by irradiating biomolecules such as proteins with an electron beam under liquid nitrogen cooling. For example, prior art 1 and prior art 2 are examples of conventional technologies related to cryoEM.

[0003] Conventional technique 1 is a technique that uses an autoencoder to estimate the continuous deformation of a 3D density map from a set of 2D cryoEM particle images obtained by cryoEM. Conventional technique 2 is a technique that uses MD (Molecular Dynamics) simulations to estimate a plausible 3D atomic model while preserving protein-like characteristics from each 2D cryoEM particle image.

[0004] Figure 15 shows examples of a three-dimensional atomic model and a three-dimensional density map. For example, the three-dimensional atomic model 5A represents the three-dimensional structure of the entire protein by representing the bonds between atoms of each amino acid residue contained in the protein with line segments. On the other hand, the three-dimensional density map 5B is data that represents the distribution of electron density in a protein and is used to visualize the shape and structure of a protein.

[0005] There is also a conventional technique (3) that acquires a 3D density map and a 3D atomic model of a typical structure as a pair, and then moves the typical structure to fit it to the 3D density map.

[0006] Currently, there is no technique to estimate the plausible continuous deformation of a protein's 3D atomic model from a set of 2D cryoEM particle images. However, it is thought that by combining the conventional techniques 1 and 2 (or 3) described above, it may potentially be possible to estimate the plausible continuous deformation of a protein's 3D atomic model. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Trabuco, Leonardo G., et al. "Molecular dynamics flexible fitting: a practical guide to combine cryo-electron microscopy and X-ray crystallography." Methods 49.2, 174-180 (2009) [Overview of the project] [Problems that the invention aims to solve]

[0008] However, simply combining the above conventional techniques 1 and 2 (or 3) presents a problem: it is not possible to accurately estimate the plausible continuous deformation of the three-dimensional atomic model of a protein.

[0009] For example, the 3D density map estimated by conventional technology 1 often contains undefined regions with insufficient accuracy. The presence of such undefined regions makes it difficult to accurately fit a 3D atomic model (typical structure) to the 3D density map.

[0010] In one aspect, the present invention aims to provide an estimation program, estimation method, and information processing device that can accurately estimate plausible continuous deformations of three-dimensional atomic models relating to macromolecules such as proteins. [Means for solving the problem]

[0011] In the first aspect, the computer executes the following processing. In the process of training an autoencoder having an encoder and a decoder using a plurality of learning data in which a polymer particle image is an explanatory variable and a three-dimensional density map of the polymer is an objective variable, the computer obtains the distribution of latent variables output from the encoder. Based on the distribution of the latent variables, the computer generates a path related to deformation. Among the plurality of latent variables generated by inputting a plurality of particle images into the trained encoder, the computer selects a plurality of neighboring latent variables whose distance from the path is less than a threshold value. Among the plurality of particle images, the computer selects a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables. Based on the plurality of neighboring particle images, the computer estimates a plurality of three-dimensional atomic models.

Advantages of the Invention

[0012] It is possible to accurately estimate a plausible continuous deformation of a three-dimensional atomic model for a polymer such as a protein.

Brief Description of the Drawings

[0013] [Figure 1] FIG. 1 is a diagram for explaining CryoTWIN (PaStEL). [Figure 2] FIG. 2 is a diagram for explaining isotropy. [Figure 3] FIG. 3 is a diagram for explaining the process of calculating a plausible continuous deformation of the latent variable z. [Figure 4] FIG. 4 is a diagram for explaining the CryoTM method. [Figure 5] FIG. 5 is a diagram for explaining MDFF. [Figure 6] FIG. 6 is a diagram (1) for explaining the processing of the information processing apparatus according to the present embodiment. [Figure 7] FIG. 7 is a diagram for explaining geodesic distance. [Figure 8] FIG. 8 is a diagram (2) for explaining the processing of the information processing apparatus according to the present embodiment. [Figure 9] Figure 9 is a functional block diagram showing the configuration of the information processing device according to this embodiment. [Figure 10] Figure 10 shows an example of a two-dimensional cryoEM particle image obtained experimentally. [Figure 11] Figure 11 is a diagram that provides supplementary explanation of the processing in the estimation unit. [Figure 12] Figure 12 is a flowchart showing the processing procedure of the information processing device according to this embodiment. [Figure 13] Figure 13 is a diagram illustrating other processes (2) performed by the information processing device. [Figure 14] Figure 14 shows an example of a computer hardware configuration that achieves similar functions to the information processing device described in the embodiment. [Figure 15] Figure 15 shows an example of a three-dimensional atomic model and a three-dimensional density map. [Modes for carrying out the invention]

[0014] The following describes in detail, with reference to the drawings, embodiments of the estimation program, estimation method, and information processing device disclosed in this application. However, this embodiment does not limit the present invention. [Examples]

[0015] Before describing this embodiment, we will explain in more detail CryoTWIN (PaStEL), which corresponds to the above-mentioned prior art 1. CryoTWIN is a combination of spatial-RaDOGAGA ​​(DeepTWIN), etc. PaStEL is an abbreviation for Generator of Pathways with Structural Change on Pseudo Free-Energy Landscape from Cryo-EM Images.

[0016] Figure 1 is a diagram illustrating CryoTWIN(PaStEL). As shown in Figure 1, CryoTWIN10 includes an encoder 11 and a decoder 12. CryoTWIN10 applies cryoEM to DeepTWIN. For the sake of explanation, the device that performs the processing related to CryoTWIN10 will be referred to as "device".

[0017] First, we will explain the process by which the device predicts a three-dimensional density map based on multiple particle images.

[0018] For example, by using cryoEM to image protein 6 from various angles, multiple particle images are generated. The instrument generates a Fourier image X by performing a Fourier Transform (FT) on the particle images 7 obtained from cryoEM. The instrument calculates the latent variable z by inputting the Fourier image X into the encoder 11. The latent variable z is P ψ The method shall follow the (GMM) model, which is an abbreviation for Gaussian Mixture Model. In this example, proteins are used as an example, but macromolecules may also be used, such as nucleic acids, glycans, lipids, etc.

[0019] Next, the device inputs the latent variables z and v into the decoder 12, and then X' z Calculate (v). v represents the 3D position and is defined by equation (1). In equation (1), R' indicates the direction relative to protein 6 when particle image 7 was taken. The v on the right side of equation (1) represents the position (2D position) of Fourier image X.

[0020] v=R'v···(1)

[0021] X' z(v) represents the value of the 3D position v in the 3D Fourier volume. The instrument generates a 3D Fourier volume 8 by repeatedly performing the above process on multiple particle images obtained from the same protein 6. The instrument predicts a 3D density map 9 by performing IFT (Inverse fast Fourier Transform) on the 3D Fourier volume.

[0022] Here, CryoTWIN trains the encoder 11 and decoder 12 using the training dataset. The training dataset contains multiple training data. For example, the explanatory variables (input data) of the training data are the particle images of proteins. The target variables (ground truth data) of the training data are the 3D density maps of proteins (such as the 3D Fourier volumes corresponding to the 3D density maps).

[0023] The device inputs a Fourier image obtained from the particle image of the training data into the encoder 11, and updates the parameters of the encoder 11 and decoder 12 so that the value output from the decoder 12 approaches the ground truth data. For example, the device uses backpropagation. As described above, the device inputs the Fourier image obtained by performing FT on the particle image into the encoder 11, and inputs the latent variable z and the value of the 3D position v into the decoder 12.

[0024] The device repeatedly performs the above process using multiple training data points included in the training dataset, and in the process, obtains the distribution of the latent variable z output from the encoder 11. In the following explanation, the distribution of the latent variable z will be referred to as the "latent distribution".

[0025] Furthermore, the latent distribution obtained during the process of the device training the encoder 11 and decoder 12 using the training dataset is "isometric".

[0026] Figure 2 is a diagram used to explain isometricity. Figure 2 shows graphs G1, G2, and G3. Graph G1 is a graph of the latent distribution obtained from the structures of the original proteins (multiple proteins corresponding to the ground truth data). Graph G2 is a graph of the latent distribution obtained by applying a spatial-VAE (Variational AutoEncoder) to multiple proteins. Graph G3 is a graph of the latent distribution obtained from the encoder 11 described in Figure 1 for multiple proteins.

[0027] The horizontal axis of graphs G1, G2, and G3 corresponds to the first principal component (PC1) in principal component analysis. The vertical axis of graphs G1, G2, and G3 corresponds to the second principal component (PC2) in principal component analysis. Each plot on graphs G1, G2, and G3 corresponds to the structure of a single protein.

[0028] In graphs G1 and G3, plots of proteins with similar structures are clustered together, indicating isometricity. On the other hand, in graph G2, plots of proteins with dissimilar structures are placed close together, indicating a lack of isometricity. The reason for the lack of isometricity is that the original protein structures are N(z;0,I d This is because it is distorted by ).

[0029] Here, CryoTWIN (PaStEL) calculates the continuous variation of the most likely path for the latent variable z based on the latent distribution obtained using the training dataset. Figure 3 is a diagram illustrating the process of calculating the continuous variation of the most likely path for the latent variable z. The latent space in Figure 3 contains the latent distribution obtained during training. Each latent variable z included in the latent distribution is assigned a probability. In the latent space, darker colors indicate a higher probability for the latent variable z.

[0030] The device is based on the following criteria 1 and 2, μ * i From μ * jGenerate a plausible path z0 up to. For example, the path z0 is shown as in Equation (2).

[0031] Path z0 = μ * i →z1→z2...→z K-1 →z K =μ * j ···(2)

[0032] The first criterion is the criterion of making the total value of the probabilities of the latent variables z on the path z0 as large as possible. For example, the total value of the probabilities on the path z0 is shown as in Equation (3).

[0033]

Number

[0034] The second criterion is the criterion of making the path length as short as possible. For example, the path length is shown as in Equation (4).

[0035]

Number

[0036] The device can obtain a continuous deformation of the three-dimensional density structure as shown in Equation (5) by inputting the path z0 into the trained decoder 12.

[0037]

Number

[0038] For example, by inputting the latent variable z obtained during training into the trained decoder 12, the three-dimensional density structure V’ z can be started, so the latent variable z and the three-dimensional density structure V’ z can be identified with each other.

[0039] Furthermore, the latent distribution is a Gaussian distribution P as shown in Equation (6). ψ'(z)' is isometric, as explained in Figure 2. Therefore, the latent distribution is a three-dimensional density structure V' z This can be interpreted as the distribution of existence and can be defined as shown in equation (7).

[0040]

number

[0041]

number

[0042] Next, we will explain in more detail the CryoTM (Template Matching) method, which corresponds to the conventional technique 2 described above. Figure 4 is a diagram illustrating the CryoTM method. In the CryoTM method, a 2D cryoEM particle image 15a and an initial 3D atomic model (not shown) are used as input, and a 3D atomic model 15b is estimated by performing a multi-stage structural search using MD.

[0043] For example, in the CryoTM method, image matching is performed between various candidate structures obtained by structural sampling of an initial 3D atomic model and a 2D cryoEM particle image 15a, taking into account the degrees of freedom of molecular orientation, and a similarity value is calculated for each. In the CryoTM method, for example, the candidate structure with the highest similarity value is estimated as the most plausible 3D atomic model 15b for the 2D cryoEM particle image 15a.

[0044] Next, we will explain MDFF (Molecular Dynamics Flexible Fitting), which corresponds to the conventional technology 3 described above, in more detail. Figure 5 is a diagram illustrating MDFF. In MDFF, a 3D density map 16a and a 3D atomic model 16b of a typical structure are acquired as a pair, and the 3D atomic model 17 is estimated by changing the structure so that the 3D atomic model 16b fits the 3D density map 16a. In MDFF, when moving the 3D atomic model 16b, MD is used by applying an external force corresponding to the gradient of the 3D density map 16a.

[0045] The above provides a more detailed explanation of the conventional technologies 1, 2, and 3.

[0046] Next, the information processing device according to this embodiment will be described. Figure 6 is Figure (1) for illustrating the processing of the information processing device according to this embodiment. In the following description, the information processing device according to this embodiment will be referred to as "information processing device 100".

[0047] The information processing device 100 utilizes an autoencoder to estimate a three-dimensional density map from a group of two-dimensional cryoEM particle images obtained by cryoEM. This autoencoder corresponds to the CryoTWIN 10 described in Figure 1 and has an encoder 11 and a decoder 12.

[0048] In the process of training the parameters of the autoencoder's encoder 11 and decoder 12 using a training dataset, the information processing device 100 obtains the distribution of the latent variable z output from the encoder 11 (latent distribution Ld).

[0049] The information processing device 100 generates a path 20 relating to the transformation from the starting point S to the ending point E, based on the first and second criteria. The path 20 corresponds to the path z0 shown in equation (2).

[0050] The information processing device 100 calculates the latent variable z (defined by equation (9) below) by inputting the target 2D cryoEM particle image I (defined by equation (8) below) into the encoder 11 of the trained autoencoder. For example, the target 2D cryoEM particle image I is an image of the protein to be analyzed, taken from multiple directions by cryoEM. The information processing device 100 may also use a particle image from the training dataset as the 2D cryoEM particle image I.

[0051]

number

number

[0052] The information processing device 100 searches for nearby latent variables <defined in equation (10) below> for the sequence of points along path 20 whose geodesic distance is less than a threshold, using the latent variable z defined in equation (9). The information processing device 100 obtains nearby cryoEM particle images <defined in equation (11) below> corresponding to the nearby latent variables defined in equation (10).

[0053]

number

number

[0054] Figure 7 is a diagram illustrating geodesic distance. In the example shown in Figure 7, node group 30A and node group 30B are shown. Node group 30A includes nodes 30A-1, 30A-2, 30A-3, 30A-4, 30A-5, 30A-6, and 30A-7. ​​Node group 30B includes nodes 30B-1, 30B-2, 30B-3, 30B-4, 30B-5, 30B-6, and 30B-7. Each node corresponds to a protein molecule, etc.

[0055] For example, since node group 30A and node group 30B are not connected, the geodesic distance between node 30A-2 and node 30B-1 is "infinite". On the other hand, since node 30B-1 and node 30B-5 are connected via nodes 30B-2 to 30B-4, the geodesic distance is the distance of line segment 31 passing through nodes 30B-1 to 30B-5.

[0056] This concludes the explanation of geodesic distance.

[0057] Let's return to the explanation of the processing of the information processing device 100. Figure 8 is Figure (2) for explaining the processing of the information processing device according to this embodiment. The information processing device 100 prepares the initial structure to be used in the CryoTM method in advance. For example, the information processing device 100 obtains a typical three-dimensional atomic model from the structure database 145, etc., and uses it as the initial structure 35. Alternatively, the information processing device 100 may use MDFF to estimate a three-dimensional atomic model from a relatively accurate three-dimensional density map and use that as the initial structure. A relatively accurate three-dimensional density map is the three-dimensional density map 20V (Figure 8) obtained by inputting the latent variable 20z (Figure 6) on the path 20 into the trained decoder 12.

[0058] The information processing device 100 estimates a three-dimensional atomic model sequence (defined in equation (12) below) from a pre-prepared initial structure 35 using the CryoTM method for a nearby cryoEM particle image. The nearby cryoEM particle image may be denoised. In this case, it is possible to start with the nearest nearby latent variable in the latent space that is closest to the initial structure and gradually expand the scope.

[0059]

number

[0060] As described above, the information processing device 100, in the process of training an autoencoder using a training dataset, obtains the latent distribution output from the encoder 11 and generates a path 20 based on the latent distribution. The information processing device 100 selects several neighboring latent variables from among the multiple latent variables generated by inputting a 2D cryoEM particle image of the protein to be analyzed into the trained encoder 11, the neighboring latent variables whose distance from the path 20 is less than a threshold. The information processing device 100 estimates several 3D atomic models based on the multiple neighboring particle images corresponding to the selected multiple neighboring latent variables. This makes it possible to accurately estimate the plausible continuous deformation of the 3D atomic model of the protein.

[0061] For example, the information processing device 100 estimates multiple three-dimensional atomic models based on neighboring particle images corresponding to latent variables in the vicinity of the path 20, which are generated based on the first and second criteria. This avoids the accuracy problem of the three-dimensional density map described in the prior art 1.

[0062] The information processing device 100 selects several neighboring latent variables whose geodesic distance to the path 20 is less than a threshold. This allows for the selection of the most plausible latent variables.

[0063] The information processing device 100 uses the CryoTM method to estimate multiple three-dimensional atomic models corresponding to multiple neighboring cryoEM particle images corresponding to multiple selected neighboring latent variables. This allows for more accurate estimation of continuous deformation.

[0064] Next, an example of the configuration of the information processing device 100 that performs the above processing will be described. Figure 9 is a functional block diagram showing the configuration of the information processing device according to this embodiment. As shown in Figure 9, this information processing device 100 has a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0065] The communication unit 110 performs data communication with external devices, etc., via a network. The communication unit 110 may also receive training datasets 142, etc., from external devices.

[0066] The input unit 120 inputs various types of information to the control unit 150.

[0067] The display unit 130 displays the information output from the control unit 150.

[0068] The memory unit 140 includes an autoencoder 141, a training dataset 142, latent distribution data 143, neighboring particle image data 144, and a structure database 145. The memory unit 140 is a memory, etc.

[0069] The autoencoder 141 corresponds to the CryoTWIN 10 described in Figure 1. The autoencoder 141 has an encoder 11 and a decoder 12.

[0070] The training dataset 142 is used to train the autoencoder 141. The training dataset contains multiple training data. For example, the explanatory variables (input data) of the training data are the particle images of proteins. The target variables (ground truth data) of the training data are the 3D density maps of proteins (such as the 3D Fourier volumes corresponding to the 3D density maps).

[0071] The particle images in the training data are experimentally obtained 2D cryoEM particle images. Figure 10 shows an example of an experimentally obtained 2D cryoEM particle image. In the example shown in Figure 10, the 2D cryoEM particle image I n ,I n-1 ,···,I2,I1 is shown. For example, the information processing device 100 displays a 2D cryoEM particle image I n ,I n-1 By inputting the Fourier images of I2 and I1 into the encoder 11 of the autoencoder 141, the latent variable z is obtained from the encoder 11. n ,z n-1 ,···,z2,z1 are output.

[0072] The latent distribution data 143 is the distribution (latent distribution) of the latent variable z output from the encoder 11 during the training process of the autoencoder 141 using the training dataset 142. The latent distribution is a Gaussian distribution P as shown in equation (6). ψ It is (z).

[0073] The neighboring particle image data 144 is data of particle images corresponding to neighboring latent variables whose geodesic distance is less than a threshold for the sequence of points along path 20 described in Figure 6. For example, the neighboring particle image data 144 is the neighboring cryoEM particle image shown in Figure 8.

[0074] Structural database 145 contains typical three-dimensional atomic models used as initial structures for the CryoTM method.

[0075] Next, we will move on to the explanation of the control unit 150. The control unit 150 includes a training unit 151, a generation unit 152, a selection unit 153, and an estimation unit 154. The control unit 150 is a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc.

[0076] The training unit 151 trains the autoencoder 141 (encoder 11 and decoder 12) using the training dataset 142. The process by which the training unit 151 trains the autoencoder 141 is the same as the process by which the device described in Figure 1 trains the encoder 11 and decoder 12. In the process of training the autoencoder 141, the training unit 151 acquires the distribution of the latent variable z output from the encoder 11 (latent distribution) and registers the acquired latent distribution as latent distribution data 143 in the storage unit 140.

[0077] The generation unit 152 generates a path z0 as shown in equation (2) based on the latent distribution data 143. For example, the generation unit 152 generates a path z0 based on the first criterion and the second criterion. Path z0 corresponds to path 20 shown in Figure 6. The generation unit 152 outputs the path z0 data to the selection unit 153.

[0078] The selection unit 153 calculates the latent variable z defined in equation (9) by inputting the target 2D cryoEM particle image I to the encoder 11 of the trained autoencoder 141. The selection unit 153 uses the particle image from the training dataset 142 as the target 2D cryoEM particle image I.

[0079] The selection unit 153 selects, for the sequence of points along path z0, the nearest latent variables whose geodesic distance is less than a threshold from the latent variables z defined in equation (9). The nearest latent variables are defined in equation (10).

[0080] The selection unit 153 selects neighboring cryoEM particle images corresponding to neighboring latent variables from the particle images of the training dataset 142. The neighboring cryoEM particle image is defined as shown in equation (11). The selection unit 153 outputs the selected neighboring cryoEM particle image to the estimation unit 154.

[0081] Further explanation regarding the selection unit 153 is the same as that described in Figures 6 and 8.

[0082] The estimation unit 154 obtains a typical three-dimensional atomic model, which will be the initial structure 35, from the structure database 145. The estimation unit 154 estimates a sequence of three-dimensional atomic models from the initial structure 35 using the CryoTM method for the neighboring cryoEM particle image. The sequence of three-dimensional atomic models is defined by equation (12).

[0083] Figure 11 is a diagram to supplement the processing of the estimation unit. For example, the estimation unit 154 generates a neighborhood cryoEM particle image I' n By applying the CryoTM method to the 3D atomic model B' n The estimation unit 154 estimates the nearby CryoEM particle image I' n-1 By applying the CryoTM method to the 3D atomic model B' n-1The estimation unit 154 estimates the 3D atomic model B'2 by applying the CryoTM method to the nearby cryoEM particle image I'2. The estimation unit 154 estimates the 3D atomic model B'1 by applying the CryoTM method to the nearby cryoEM particle image I'1.

[0084] The estimation unit 154 outputs the estimated 3D atomic model sequence to the display unit 130 for display. Further details regarding the estimation unit 154 are the same as those described in Figure 8.

[0085] Next, an example of the processing procedure of the information processing device 100 according to this embodiment will be described. Figure 12 is a flowchart of the processing procedure of the information processing device according to this embodiment. As shown in Figure 12, the training unit 151 of the information processing device 100 trains the autoencoder 141 using the training dataset 142 (step S101). The training unit 151 registers the latent distribution data 143 output from the encoder 11 during training in the storage unit 140 (step S102).

[0086] The generation unit 152 of the information processing device 100 generates a path z0 based on the latent distribution data 143 (step S103). The selection unit 153 of the information processing device 100 calculates the latent variable z by inputting the target 2D cryoEM particle image I to the encoder 11 of the trained autoencoder 141 (step S104).

[0087] The selection unit 153 selects neighboring latent variables from the calculated latent variables z such that the geodesic distance is less than a threshold for the sequence of points along path z0 (step S105). The selection unit 153 selects neighboring cryoEM particle images corresponding to the neighboring latent variables from the particle images of the training dataset 142 (step S106).

[0088] The estimation unit 154 of the information processing device 100 obtains a typical three-dimensional atomic model that will become the initial structure 35 from the structure database 145 (step S107). The estimation unit 154 estimates a sequence of three-dimensional atomic models by applying the CryoTM method to the neighboring cryoEM particle image (step S108). The estimation unit 154 outputs the sequence of three-dimensional atomic models to the display unit 130 for display (step S109).

[0089] Next, the effects of the information processing device 100 according to this embodiment will be described. In the process of training an autoencoder using a training dataset, the information processing device 100 acquires the latent distribution output from the encoder and generates a path based on the latent distribution. The information processing device 100 selects several neighboring latent variables from among the multiple latent variables generated by inputting a 2D cryoEM particle image of the protein to be analyzed into the trained encoder, the neighboring latent variables whose distance from the path is less than a threshold. The information processing device 100 estimates several 3D atomic model sequences based on the multiple neighboring particle images corresponding to the selected multiple neighboring latent variables. This makes it possible to accurately estimate the plausible continuous deformation of the 3D atomic model of the protein.

[0090] By the way, the processing content of the information processing device 100 described above is just one example, and the information processing device 100 may perform other processing. Below, other processing (1) and other processing (2) of the information processing device 100 will be explained in order.

[0091] The "other processing (1)" performed by the information processing device 100 will now be explained. In the above explanation, the information processing device 100 selected neighboring latent variables from the latent variable z defined in equation (9) for the sequence of points in path z0 such that the geodesic distance is less than a threshold. However, in other processing (1), the latent variables included in path z0 are used as they are.

[0092] For example, the estimation unit 154 of the information processing device 100 generates a cryoEM particle image corresponding to each latent variable by inputting each latent variable included in the path z0 into the decoder 12 of the trained autoencoder 141.

[0093] The estimation unit 154 estimates a three-dimensional atomic model sequence from the initial structure 35 using the CryoTM method for the generated cryoEM particle image.

[0094] As described above, in the other process (1), the information processing device 100 can obtain a sequence of three-dimensional atomic models corresponding to the sequence of points by associating a three-dimensional atomic model with the cryoEM particle image obtained by the decoder 12 from the sequence of points along the path using the CryoTM method.

[0095] The "other processing (2)" performed by the information processing device 100 will now be explained. In the above explanation, the information processing device 100 used the distribution of the latent variable z output from the encoder 11 as latent distribution data 143 in the process of training the autoencoder 141 using the training dataset 142, but it is not limited to this.

[0096] Figure 13 is a diagram illustrating other processes (2) performed by the information processing device. The information processing device 100 performs MD on the whole-atom model 50 of the protein and the whole-atom model {B} whose structure has changed due to structural sampling. n The information processing device 100 generates the all-atom model {B}. n} is an MD image that simulates the cryoEM particle image, and is an MD image of various orientations {I n Obtain}.

[0097] The information processing device 100 displays the MD image {I} as explained in Figure 13. n The correct data corresponding to} is obtained to generate a training dataset 242, and the autoencoder 141 is trained using the training dataset 242. During the training process, the information processing device 100 obtains the distribution of latent variables output from the encoder 11 of the autoencoder 141 as the first latent distribution.

[0098] Meanwhile, the information processing device 100 trains the autoencoder 141 using a pre-prepared training dataset 142, in the same manner as in the above embodiment. During the training process, the information processing device 100 obtains the distribution of latent variables output from the encoder 11 of the autoencoder 141 as the second latent distribution.

[0099] Furthermore, the information processing device 100 may, after training with the training dataset 242, further train the autoencoder 141 using the training dataset 142, or it may, after resetting the parameters of the autoencoder 141, train the autoencoder 141 using the training dataset 142.

[0100] The information processing device 100 generates a path z0 based on a latent distribution obtained by superimposing a first latent distribution and a second latent distribution. The processing after the information processing device 100 generates the path z0 is the same as the processing described in the above embodiment.

[0101] In other words, for the sequence of points along path z0, the information processing device 100 selects nearby latent variables whose geodesic distance is less than a threshold from the latent variable z defined by equation (9). The information processing device 100 selects nearby cryoEM particle images corresponding to the nearby latent variables from the particle images of the training dataset 142,242. The information processing device 100 estimates a 3D atomic model sequence from the initial structure 35 using the CryoTM method for the nearby cryoEM particle images.

[0102] As described above, in the other process (2), the information processing device 100 performs MD on the whole-atom model 50 of the protein and the whole-atom model {B} whose structure has changed due to structural sampling. n} generates MD images {I n} obtain the relevant MD image {I n The first latent distribution is obtained through training using {B}. The information processing device 100 generates a path z0 generated from a latent distribution obtained by superimposing the first latent distribution and the second latent distribution. This allows for the acquisition of not only the particle image of the training dataset 142, but also the whole atom model {B}. nMD images of various orientations obtained from {I n A latent distribution that takes} into account can be used to generate the path z0.

[0103] Next, an example of a computer hardware configuration that achieves the same functions as the information processing device 100 described above will be explained. Figure 14 is a diagram showing an example of a computer hardware configuration that achieves the same functions as the information processing device according to the embodiment.

[0104] As shown in the figure, the computer 200 includes a CPU 201 that performs various calculations, an input device 202 that receives data input from the user, and a display 203. The computer 200 also includes a communication device 204 and an interface device 205 that exchange data with external devices via a wired or wireless network. Furthermore, the computer 200 includes a RAM 206 for temporarily storing various information and a hard disk drive 207. Each of these devices 201 to 207 is connected to a bus 208.

[0105] The hard disk drive 207 contains a training program 207a, a generation program 207b, a selection program 207c, and an estimation program 207d. The CPU 201 reads each of the programs 207a to 207d and loads them into the RAM 206.

[0106] The training program 207a functions as the training process 206a. The generation program 207b functions as the generation process 206b. The selection program 207c functions as the selection process 206c. The estimation program 207d functions as the estimation process 206d.

[0107] The processing of training process 206a corresponds to the processing of training unit 151. The processing of generation process 206b corresponds to the processing of generation unit 152. The processing of selection process 206c corresponds to the processing of selection unit 153. The processing of estimation process 206d corresponds to the processing of estimation unit 154.

[0108] Furthermore, programs 207a to 207d do not necessarily have to be stored on the hard disk drive 207 from the beginning. For example, each program could be stored on a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card inserted into the computer 200. Then, the computer 200 could read and execute each program 207a to 207d.

[0109] With regard to embodiments including each of the above examples, the following additional information is disclosed.

[0110] (Note 1) In the process of training an autoencoder having an encoder and a decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An estimation program characterized by having a computer perform the processing.

[0111] (Note 2) The estimation process is characterized by estimating the plurality of three-dimensional atomic models from the plurality of neighboring particle images using the cryoTM (Template Matching) method, as described in Note 1.

[0112] (Note 3) The estimation program described in Note 1, characterized in that the process of generating the path generates the path based on a first criterion that the sum of the probabilities of the latent variables included in the path is higher, and a second criterion that the length of the path is as short as possible.

[0113] (Note 4) The estimation program described in Note 1, characterized in that the computer further performs a process to obtain the distribution of the latent variables by acquiring multiple images based on multiple whole-atom models obtained by structurally changing the whole-atom model of the polymer, and using the acquired multiple images as explanatory variables of the training data.

[0114] (Note 5) The estimation program described in Note 1, characterized in that the process of selecting the plurality of neighboring latent variables is to select a plurality of neighboring latent variables from among the plurality of latent variables generated by inputting a plurality of images into the trained encoder, such that the geodesic distance to the path is less than a threshold.

[0115] (Appendix 6) The estimation program according to Appendix 1, characterized in that the computer further performs a process of estimating the plurality of three-dimensional atomic models based on the plurality of particle images corresponding to the latent variables included in the path.

[0116] (Note 7) In the process of training an autoencoder having an encoder and a decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An estimation method characterized by the processing being performed by a computer.

[0117] (Note 8) The estimation method described in Note 7, characterized in that the estimation process involves estimating the plurality of three-dimensional atomic models from the plurality of neighboring particle images using the cryoTM (Template Matching) method.

[0118] (Note 9) The estimation method described in Note 7, characterized in that the process of generating the path generates the path based on a first criterion that the sum of the probabilities of the latent variables included in the path is higher, and a second criterion that the length of the path is as short as possible.

[0119] (Note 10) The estimation method according to Note 7, characterized in that the computer further performs a process to obtain the distribution of the latent variables by acquiring multiple images based on multiple whole-atom models obtained by structurally changing the whole-atom model of the polymer, and using the acquired multiple images as explanatory variables of the training data.

[0120] (Note 11) The estimation method according to Note 7, characterized in that the process of selecting the plurality of neighboring latent variables is to select a plurality of neighboring latent variables from among the plurality of latent variables generated by inputting a plurality of images into the trained encoder, such that the geodesic distance to the path is less than a threshold.

[0121] (Note 12) The estimation method according to Note 7, characterized in that a computer further performs a process of estimating the plurality of three-dimensional atomic models based on the plurality of particle images corresponding to the latent variables included in the path.

[0122] (Note 13) In the process of training an autoencoder having an encoder and a decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An information processing device having a control unit that performs processing.

[0123] (Note 14) The information processing apparatus according to Note 13, characterized in that the estimation process uses the TM (Template Matching) method to estimate the plurality of three-dimensional atomic models from the plurality of neighboring particle images.

[0124] (Note 15) The information processing apparatus according to Note 13, characterized in that the process of generating the path generates the path based on a first criterion that the sum of the probabilities of the latent variables included in the path is higher, and a second criterion that the length of the path is as short as possible.

[0125] (Note 16) The information processing apparatus according to Note 13, characterized in that the control unit further performs a process to obtain the distribution of the latent variables by acquiring multiple images based on multiple whole-atom models obtained by structurally changing the whole-atom model of the polymer, and using the acquired multiple images as explanatory variables of the training data.

[0126] (Note 17) The information processing device according to Note 13, characterized in that the process of selecting a plurality of neighboring latent variables is to select a plurality of neighboring latent variables from among a plurality of latent variables generated by inputting a plurality of images into the trained encoder, such that the geodesic distance to the path is less than a threshold.

[0127] (Note 18) The information processing apparatus according to Note 13, characterized in that the control unit further performs a process of estimating the plurality of three-dimensional atomic models based on the plurality of particle images corresponding to the latent variables included in the path. [Explanation of Symbols]

[0128] 100 Information Processing Devices 110 Communications Department 120 Input section 130 Display section 140 Storage section 141 Autoencoder 142 training datasets 143 Latent Distribution Data 144 Nearby Particle Image Data 145 Structural Databases 150 Control Unit 151 Training Department 152 Generation part 153 Selection Section 154 Estimation Department

Claims

1. In the process of training an autoencoder having an encoder and decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An estimation program characterized by having a computer perform the processing.

2. The estimation program according to claim 1, characterized in that the estimation process estimates the plurality of three-dimensional atomic models from the plurality of neighboring particle images using the cryoTM (Template Matching) method.

3. The estimation program according to claim 1, characterized in that the process of generating the path generates the path based on a first criterion that the sum of the probabilities of the latent variables included in the path is higher, and a second criterion that the length of the path is as short as possible.

4. The estimation program according to claim 1, characterized in that it causes a computer to perform a process to obtain the distribution of the latent variables by acquiring multiple images based on multiple whole-atom models obtained by structurally changing the whole-atom model of the polymer, and using the acquired multiple images as explanatory variables of the training data.

5. The estimation program according to claim 1, characterized in that the process of selecting a plurality of neighboring latent variables is to select a plurality of neighboring latent variables from among a plurality of latent variables generated by inputting a plurality of images into the trained encoder, such that the geodesic distance to the path is less than a threshold.

6. The estimation program according to claim 1, characterized in that it causes a computer to further perform a process of estimating the plurality of three-dimensional atomic models based on the plurality of particle images corresponding to the latent variables included in the aforementioned path.

7. In the process of training an autoencoder having an encoder and decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An estimation method characterized by the processing being performed by a computer.

8. In the process of training an autoencoder having an encoder and decoder using multiple training data sets in which the particle image of the polymer is the explanatory variable and the three-dimensional density map of the polymer is the objective variable, the distribution of the latent variables output from the encoder is obtained. Based on the distribution of the aforementioned latent variables, a path for deformation is generated. From among the multiple latent variables generated by inputting multiple particle images into the trained encoder, a number of neighboring latent variables whose distance from the path is less than a threshold are selected. From the plurality of particle images, select a plurality of neighboring particle images corresponding to the plurality of neighboring latent variables, Based on the aforementioned images of multiple neighboring particles, multiple three-dimensional atomic models are estimated. An information processing device having a control unit that performs processing.

Images