Predictive control method, information processing device, and predictive control program

Abstract

The challenge is to improve the accuracy of fitting when predicting the target structure. [Solution] The predictive control method acquires first information corresponding to the predicted structure by a machine learning-based structural prediction model and second information which is the fitting target of the first information. Based on the acquired first and second information, the computer generates intermediate information between the first and second information, and then updates the intermediate features of the structural prediction model based on the generated intermediate information.

Description

[Technical Field]

[0001] The present invention relates to a predictive control method, an information processing device, and a predictive control program. [Background technology]

[0002] Techniques for predicting atomic structure from amino acid sequences have been disclosed (see Non-Patent Document 1). For example, machine learning models such as AlphaFold2 and OpenFold (hereinafter referred to as "structure prediction models") output one typical atomic structure of a protein for each input amino acid sequence.

[0003] Figure 1 is a reference diagram illustrating the AlphaFold2 structural prediction model. As shown in Figure 1, the AlphaFold2 structural prediction model, when given an InputSequence representing the amino acid sequence, outputs a single 3D structure as a typical atomic structure of the protein.

[0004] Incidentally, predicting the atomic structure of proteins is a crucial element in drug discovery. For application in drug discovery, it is necessary to predict a variety of atomic structures other than typical ones. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] John Jumper, et al. “Highly accurate protein structure prediction with AlphaFold”, [online], [Retrieved February 6, 2025], Internet<https: / / www.nature.com / articles / s41586-021-03819-2> [Non-Patent Document 2] Arthur Ecoffet, et. al“MorphOT: transport-based interpolation between EM maps with UCSF ChimeraX”,[online],[Searched February 6, Reiwa 7],Internet<https: / / academic.oup.com / bioinformatics / article / 36 / 22-23 / 5528 / <6039110> [Non-Patent Document 3] Kimihiro Yamazaki, et. al“An Auto-Encoder to Reconstruct Structure with Cryo-EM Images via Theoretically Guaranteed Isometric Latent Space, and its Application for Automatically Computing the Conformational Pathway”,[online],[Searched February 6, Reiwa 7],Internet<https: / / conferences.miccai.org / 2023 / papers / 055-Paper2173.html> [Summary of the Invention] [Problem to be Solved by the Invention]

[0006] However, conventional structure prediction models only output one typical atomic structure, and thus cannot predict diverse atomic structures. Furthermore, it is impossible to improve the fitting accuracy when predicting the target structure.

[0007] In one aspect, an object of the present invention is to provide a prediction control method, an information processing apparatus, and a prediction control program that can improve the fitting accuracy when predicting a target structure.

[0008] A predictive control method relating to one aspect involves acquiring first information corresponding to the predicted structure by a machine learning-based structural prediction model and second information which is the fitting target of the first information. Based on the acquired first and second information, the computer generates intermediate information between the first and second information, and then updates the intermediate features of the structural prediction model based on the generated intermediate information. [Effects of the Invention]

[0009] According to one embodiment, the accuracy of fitting during the prediction of the target structure can be improved. [Brief explanation of the drawing]

[0010] [Figure 1] Figure 1 is a reference diagram showing the structural prediction model of AlphaFold2. [Figure 2] Figure 2 is a simplified diagram illustrating the structure prediction model of AlphaFold2. [Figure 3] Figure 3 shows an image of the predictive control according to Example 1. [Figure 4] Figure 4 shows an image of the atomic structure search method according to Example 1. [Figure 5A] Figure 5A is a diagram (1) showing an example of the application results when fitting using the predictive control method according to Example 1 is applied. [Figure 5B] Figure 5B is a diagram (2) showing an example of the application results when fitting using the predictive control method according to Example 1 is applied. [Figure 6] Figure 6 shows an image of the atomic structure search method according to Example 2. [Figure 7] Figure 7 shows an illustration of the differences between the atomic structure search methods in Example 1 and Example 2. [Figure 8] Figure 8 shows an example of the application results obtained by applying the fitting method according to the predictive control method of Example 2. [Figure 9]Figure 9 shows an example of the functional configuration of the information processing device according to Embodiment 1. [Figure 10] Figure 10 shows the distribution of electrons. [Figure 11] Figure 11 shows an example of the output atomic structure display. [Figure 12] Figure 12 shows an example of how the coordinates of the output atomic structure are displayed. [Figure 13] Figure 13 shows an example of a flowchart of the predictive control process according to Example 1. [Figure 14] Figure 14 shows an example of the functional configuration of the information processing device according to Example 2. [Figure 15] Figure 15 shows an example of a flowchart of the predictive control process according to Example 2. [Figure 16] Figure 16 shows an example of a hardware configuration. [Modes for carrying out the invention]

[0011] The following describes embodiments of the predictive control method, information processing device, and predictive control program according to the present invention with reference to the attached drawings. Each embodiment is merely an example or aspect, and such examples do not limit the range of numerical values ​​or functions, or the usage scenarios. Furthermore, each embodiment can be appropriately combined as long as the processing content is not contradictory.

[0012] [1. Introduction] Conventionally, techniques for predicting high-dimensional structural data using molecular structure prediction models are known.

[0013] An example of a conventional technique for predicting high-dimensional structural data using a molecular structure prediction model is the technique disclosed in Non-Patent Document 1 (particularly the schematic diagram of the structure prediction model shown in Figure 1(E)). While details are omitted here, the structure prediction model described in Non-Patent Document 1 is called AlphaFold2, and it outputs 3D structural data when an amino acid sequence is input.

[0014] Figure 2 is a simplified diagram of the AlphaFold2 structure prediction model. The structure prediction model MO11 includes a preprocessing module MD1 for data shaping, a feature extraction module MD2 (corresponding to an encoder called Evoformer) for extracting characteristic features, and a structure prediction module MD3 for outputting 3D structure data.

[0015] The preprocessing module MD1 generates two types of intermediate features from the amino acid sequence IN1. One is a sequence feature of the amino acid sequence IN1, which is generated based on MSA (Multiple Sequence Alignment). This first feature corresponds to the MSA representation (s,r,c) and will be referred to as "first intermediate feature RE11" below.

[0016] The other is a spatial (structural) feature of the amino acid sequence IN1, which is a feature generated based on amino acid pairing. This second feature corresponds to Pair representation(r,r,c) and will be referred to as "First Intermediate Feature RE12" from now on. The two types of intermediate features generated in this way will be combined and referred to as "First Intermediate Feature RE1" from now on. First Intermediate Feature RE1 is the intermediate feature that is input to the feature extraction module MD2 (i.e., the intermediate feature that is the input target for Evoformer).

[0017] The feature extraction module MD2 extracts characteristic features from each of the two types of intermediate features generated by the preprocessing module MD1. In other words, it extracts two types of features. One is the feature extracted from the first intermediate feature RE11. This first feature corresponds to a Single representation (r,c) and will be referred to as the "second intermediate feature RE21" as appropriate below.

[0018] The other is a feature extracted from the first intermediate feature RE12. This second feature corresponds to a Pair representation (r,r,c) and will be referred to as "second intermediate feature RE22" from now on. The two types of intermediate features extracted in this way will be combined and referred to as "second intermediate feature RE2" from now on. Second intermediate feature RE2 is an intermediate feature that is input to the structure prediction module MD3 (i.e., an intermediate feature that is the input for the "Structure module").

[0019] The MD3 structural prediction module generates 3D structural data from two types of intermediate features extracted by the MD2 feature extraction module.

[0020] However, these conventional structural prediction models only output one typical atomic structure, and therefore cannot predict a variety of atomic structures.

[0021] Therefore, in the embodiments described below, we propose a structural prediction model MO11 based on AlphaFold2 that can predict a variety of atomic structures, and further, we propose a prediction control method using the structural prediction model MO11 to improve the accuracy of fitting when predicting the target structure.

[0022] In the embodiment described below, the structure prediction model MO11 is a model that takes an amino acid sequence as input and outputs 3D structural data of an organic compound (such as a protein). The structure prediction model MO11 takes an amino acid sequence as input and outputs 3D structural data of an organic compound as a target structure corresponding to the input amino acid sequence. Furthermore, if the structure prediction model MO11 includes multiple layers, intermediate features may be generated in each of the multiple layers. For example, the first intermediate feature RE1 may be generated in the first layer, and the second intermediate feature RE2 may be generated in the second layer.

[0023] [2. Explanation of Information Processing] Next, we will explain the information processing involved in the examples. We will explain using two-stage examples: Example 1, which aims to predict diverse atomic structures, and Example 2, which aims to improve the accuracy of fitting when predicting target structures. The contents explained in Example 1 will be omitted as appropriate in the explanation of Example 2.

[0024] (Example 1) First, the concept of predictive control according to Example 1 will be explained with reference to Figure 3. Figure 3 is a diagram illustrating the concept of predictive control according to Example 1. The predictive control process shown in Figure 3 updates the intermediate features of the structure prediction model (such as the first intermediate feature RE1 and the second intermediate feature RE2) to fit the information that limits the target atomic structure, thereby predicting the atomic structure. The information that limits the atomic structure may be referred to as limiting information from now on. Also, in Figure 3, intermediate features such as the first intermediate feature RE1 and the second intermediate feature RE2 will not be distinguished in the explanation.

[0025] The following examples use proteins as merely one example of the target, but the target is not limited to proteins. For example, the target may be organic compounds other than proteins. Also, as an example of information that limits the atomic structure of the target (limiting information), a 3D density map that limits the atomic structure of proteins is given, but it is not limited to this. Note that a 3D density map can be represented, for example, in the form of 3D voxel data. A 3D density map can be represented, for example, with a size of n x ×n y ×n z It is represented in a format in which density values ​​are included within a 3D grid. The auxiliary information, intermediate auxiliary information, and target auxiliary information described below are all 3D density maps.

[0026] As shown in Figure 3, the predictive control process includes a structural prediction model MO11. This structural prediction model MO11 can be any machine learning model, such as AlphaFold2 or OpenFold. In Example 1, we will explain the process assuming that AlphaFold2 is applied to the structural prediction model MO11.

[0027] The prediction control process inputs amino acid sequence information into the structure prediction model MO11 (step S1). The structure prediction model MO11 then extracts intermediate features from the amino acid sequence information, inputs the extracted intermediate features into a layer, and outputs the atomic structure (3D structure) of the protein as the prediction result (step S2). For the sake of explanation, the structure prediction model MO11 is shown as a single layer, but it may also consist of multiple layers.

[0028] The prediction control process converts the atomic structure output as the prediction result into a three-dimensional density map that restricts the said atomic structure in a differentiable form (step S3).

[0029] The prediction control process then calculates the difference between the 3D density map corresponding to the predicted atomic structure and the measured 3D density map (step S4). The measured 3D density map is, for example, an EM (Electron Microscope) image captured by an electron microscope such as a cryo-electron microscope, which is an EM image that limits the atomic structure of the protein, but is not limited to this and can be in voxel data format.

[0030] The predictive control process then calculates the backpropagation of the difference and computes intermediate features that minimize the difference (step S5). The predictive control process then updates the intermediate features of the structure prediction model MO11 with the computed intermediate features (step S6). In other words, the predictive control process updates the intermediate features of the structure prediction model MO11 so that the difference between the 3D density map corresponding to the predicted atomic structure and the measured 3D density map becomes smaller. To put it another way, the predictive control process updates the intermediate features of the structure prediction model MO11 due to constraints from the measured 3D density map.

[0031] Next, the structure prediction model MO11 uses the updated intermediate features to predict the atomic structure of the protein. The prediction control process repeats steps S2 to S6, making it possible to predict a variety of atomic structures.

[0032] Furthermore, if the MO11 structural prediction model has multiple layers, the prediction control process only needs to update the intermediate features input to any of the layers. This also applies to the first intermediate feature RE1 and the second intermediate feature RE2. Backpropagation updates at least one of the multiple intermediate features, including the first intermediate feature RE1 and the second intermediate feature RE2. For example, either the first intermediate feature RE1 or the second intermediate feature RE2 may be updated, or both intermediate features may be updated. For the sake of explanation, the case of updating the second intermediate feature RE2 will be explained below as an example.

[0033] Furthermore, the predictive control process updates the intermediate features of the structural prediction model MO11, but does not update the training parameters of the structural prediction model MO11. This is to avoid the forgetting of the knowledge held by the structural prediction model MO11 by not destroying the existing training parameters through retraining.

[0034] The above is an image of the predictive control according to Example 1. The predictive control method according to Example 1 is a method for searching for the atomic structure of a protein using the structure prediction model MO11, and utilizes a pre-trained structure prediction model MO11. Furthermore, in the predictive control method according to Example 1, intermediate features are updated by backpropagation based on a loss function corresponding to the desired properties, and the updated intermediate features are processed again by the network's structure prediction model MO11 to generate a new structure with the desired properties. The predicted structure is changed to guide the system toward a new structure with the desired properties. That is, the predicted structure is gradually changed by updating the intermediate features, and finally a structure that matches the density map of the target structure is predicted.

[0035] Furthermore, an image of the atomic structure search method according to Example 1 will be explained with reference to Figure 4. Figure 4 is a diagram illustrating an image of the atomic structure search method according to Example 1. In the atomic structure search method shown in Figure 4, first, auxiliary information corresponding to the target structure (target auxiliary information) E Tand auxiliary information E corresponding to the predicted structure S0 ′ 0 updates the intermediate feature amount H0 based on the comparison with . The intermediate feature amount after update is the intermediate feature amount H α . The intermediate feature amount H α is input to a decoder to obtain the predicted structure S α . The predicted structure S α is converted to obtain auxiliary information E ′ α . Next, target auxiliary information E T and auxiliary information E corresponding to the predicted structure S α corresponding to the predicted structure S ′ α updates the intermediate feature amount H based on the comparison with α . The intermediate feature amount after update becomes the intermediate feature amount H1. The intermediate feature amount H1 is input to a decoder to generate the predicted structure S1. The predicted structure S1 is converted to obtain auxiliary information E ′ 1. Then, the intermediate feature amount H1 is updated based on the comparison between the target auxiliary information E T and auxiliary information E ′ 1 corresponding to the predicted structure S1.

[0036] In Example 1, as described above, a prediction control method is proposed that uses the structure prediction model MO11 to predict a target (objective) atomic structure (target structure or objective structure) based on indirect auxiliary information (corresponding to a three-dimensional density map). A prediction control method is proposed that updates intermediate feature amounts (such as the first intermediate feature amount RE1 and the second intermediate feature amount RE2) of the structure prediction model MO11 to perform fitting to the three-dimensional density map of the target structure.

[0037] In the prediction control method according to Example 1, starting from the predicted structure of the structure prediction model MO11, the error between the predicted structure and the target structure is calculated using the target auxiliary information as a hint, and fitting is performed by updating the intermediate feature amount of the structure prediction model MO11 through error backpropagation.

[0038] However, the predictive control method according to Example 1 may not be able to improve the accuracy of fitting when predicting the target structure. For example, the fitting process may follow an unnatural deformation process that passes through intermediate features that deviate from the training range of the structure prediction model MO11, and the fitting may not be completed. The fitting may not be completed if the intermediate state of the fitting deviates from the training range, or if the starting structure is far from the target structure.

[0039] Figures 5A and 5B show examples of application results when fitting using the predictive control method described in Example 1 is applied. The results shown in Figures 5A and 5B refer to known protein structures registered in a database called the Protein Data Bank (PDB), with a protein structure close to the starting structure (1iwo) and a protein structure close to the target structure (1su4) being used as references. 1iwo and 1su4 are proteins with the same amino acid sequence but different shapes.

[0040] In Result EX1 and Result EX2 shown in Figure 5A, the white areas represent the visualization of electron density, and the gray areas represent the visualization of the predicted structure. Result EX1 is the data before fitting, and Result EX2 is the data after fitting. From Result EX1 and Result EX2, it can be seen that the deviation from the white areas is larger before fitting, indicating that the fitting was not sufficient.

[0041] Figure 5B shows data plotting the distance between structures. In Figure 5B, the vertical axis is RMSD, and the horizontal axis is the number of updates. Plot PL1 is the difference from the starting structure, and plot PL2 is the difference from the target structure. The dotted line LN1 is a threshold indicating the success of the fitting. When plot PL2 reaches the dotted line LN1, it indicates that the fitting is sufficient. In Figure 5B, plot PL2 is decreasing, but it has not reached the dotted line LN1, so it can be seen that the fitting is not sufficient.

[0042] From plot PL1, we can see that, before fitting, the predicted structure closely matches the starting structure, although not a perfect match. Furthermore, as fitting progresses and the predicted structure approaches the target structure, it moves away from the starting structure, and in contrast to plot PL2, plot PL1 shows an upward trend.

[0043] When the difference between the starting structure and the target structure is large, the discrepancy in the distance between the structures becomes large, which often makes it difficult to achieve a proper fit.

[0044] Therefore, in order to improve the accuracy of fitting during the prediction of the target structure, we propose a further prediction control method using the structural prediction model MO11.

[0045] (Example 2) In Example 2, we propose a predictive control method that improves fitting accuracy by simulating an intermediate structure between the starting structure and the target structure in an auxiliary information space, and then gradually guiding the model to this intermediate structure, thereby achieving a smooth deformation process.

[0046] In the prediction control method according to Example 1, since the target auxiliary information is directly referenced, the prediction accuracy decreases if it deviates too much from the starting structure when compared. On the other hand, in Example 2, an intermediate structure that is closer to the starting structure than the target structure is provided, and the prediction structure is updated by comparing it with this intermediate structure to gradually guide it. In other words, a density map of intermediate structures (hereinafter referred to as "intermediate auxiliary information") is generated. Note that there is no limit to the number of intermediate auxiliary information; at least one is required.

[0047] Here, the image of the atomic structure search method according to Example 2 will be explained with reference to Figure 6. Figure 6 is a diagram illustrating the image of the atomic structure search method according to Example 2. The inputs are the amino acid sequence to be predicted IN1 and target auxiliary information E T The following are used. The prediction structure S0 is the prediction structure of the initial prediction, and the auxiliary information E ′0 is auxiliary information corresponding to the prediction structure of the initial prediction.

[0048] In the atomic structure search method shown in Figure 6, first, target auxiliary information E T and supplementary information E ′ Between 0 and intermediate auxiliary information E1 ~ intermediate auxiliary information E T-1 It generates intermediate auxiliary information E1 and auxiliary information E ′ The intermediate feature H0 is updated based on a comparison with 0. The updated intermediate feature becomes intermediate feature H1. Then, the intermediate auxiliary information E2 and auxiliary information E ′ The intermediate feature H1 is updated based on a comparison with 1.

[0049] Thus, in Figure 6, the predicted structure S t-1 Prediction structure S with the starting structure as the starting structure t-1 Corresponding auxiliary information E ′ 0 is intermediate auxiliary information E t By fitting it to the predicted structure S t This generates a predictive structure S by repeating this process T times. T This generates the following. It performs fitting to intermediate auxiliary information that is close to the auxiliary information corresponding to the starting structure, and once the fitting is complete, it moves on to the next intermediate auxiliary information. This is repeated in order until the final target auxiliary information E T Fit it to it.

[0050] Thus, Intermediate Auxiliary Information E1 to Intermediate Auxiliary Information E T-1 By setting up and repeating the fitting process, the target auxiliary information E is ultimately obtained. T A predictive structure S that fits to this. T Generates.

[0051] Figure 7 is a diagram illustrating the differences between the atomic structure search methods in Example 1 and Example 2. In the atomic structure search method in Example 1, target auxiliary information E TIn contrast to directly predicting the target structure from the predicted structure S0, the atomic structure search method in Example 2 generates a pseudo-density map of intermediate structures between the predicted structure S0 and the target structure, and then sequentially fits the intermediate structures closest to the predicted structure S0 to the target structure, ultimately guiding the user towards the target structure.

[0052] The following details the conditions for generating intermediate auxiliary information. Intermediate auxiliary information is generated based on predetermined conditions. In the auxiliary information space, intermediate auxiliary information E t From target auxiliary information E T An indicator of ease of reaching is the indicator d(E t ,E T ) Let the index d(E t ,E T ) The smaller the value, the greater the target auxiliary information E T This indicates that it is close to and easily attainable. For example, the predicted structure predicted by the structural prediction model MO11 at this point is predicted structure S. t In this case, the current predicted structure S is used for the final target structure. t This indicates how easily it can be reached. Index d(E t ,E T This is merely to define the conditions for generating intermediate auxiliary information, and it is not necessary to be able to directly compute the target.

[0053] Index d(E t ,E T Let's explain an example of the index d(E). t ,E T ) can be expressed, for example, by equation (1). ρ tag Ng This shows the density map of the target structure, ρ tag Ng This shows the density map of the predicted structure. The cross-correlation between density maps (hereinafter referred to as "cross-correlation C1" as appropriate) indicates how similar the density maps are to each other. The more similar the density map of the target structure and the density map of the predicted structure are, the easier it is to reach the target structure.

number

[0054] In addition, there is the indicator d(E t ,E T ) can be expressed, for example, by the Wasserstein distance. For example, the index d(E t ,E T ) is a density map E represented by density values ​​on a 3D voxel. t and density map E t The Wasserstein distance W between 2 2(E t ,E t It is represented as follows:

[0055] Furthermore, the index d(E t ,E T This is represented by the inter-structure distance corrected by the structure probability estimated by, for example, a structure probability distribution estimation method in the auxiliary information space (for example, the method disclosed in Non-Patent Document 3 above). In this method, when estimating paths to reach a target structure, it calculates whether there is a path with a high probability.

[0056] Index d(E t ,E T In either case, the index d(E t ,E T The generation condition is that the following satisfies equation (2). That is, the intermediate auxiliary information that goes next to fitting is the final target auxiliary information E T The generation condition is that the state is easily reachable.

number

[0057] In this embodiment, an easily reachable state refers to a state in which the target is easier to reach than the current state. Therefore, in this embodiment, ease of reach is not limited to geographical proximity or structural proximity. In this embodiment, ease of reach also includes, for example, paths that move away from the target (paths that eventually approach the target but initially move away) in terms of the cross-correlation between density maps.

[0058] The following describes in detail how to generate intermediate auxiliary information. Based on the starting structure (including the prediction structure of the initial prediction), all intermediate auxiliary information may be generated at once, or it may be generated dynamically during the fitting process in stages. For example, when generating intermediate auxiliary information dynamically, using the example in Figure 6, auxiliary information E ′ After generating 1, target auxiliary information E T and supplementary information E ′ Between 1 and 1, intermediate auxiliary information E2 to intermediate auxiliary information E T-1 You may generate the following, or you may generate only the intermediate auxiliary information E2 that leads to the next fitting.

[0059] When generating all intermediate auxiliary information at once, examples of methods include those based on linear interpolation between density maps (e.g., the method represented by equation (3)) and those based on density interpolation based on optimal transport (e.g., the method disclosed in Non-Patent Document 3 above). The latter method is based on what paths exist when approaching the target auxiliary information, and allows for more natural interpolation (a more natural way of approaching the target auxiliary information) than the former method.

number

[0060] When generating intermediate auxiliary information dynamically, auxiliary information E ′ t-1 and target auxiliary information E T Methods that generate based on linear interpolation with (for example, the method expressed in equation (4)), and auxiliary information E ′ t-1 As intermediate auxiliary information in the vicinity, multiple candidate intermediate auxiliary information are generated and index d(E t ,E T The candidate for the intermediate auxiliary information that minimizes ) is the intermediate auxiliary information E t One example of such a method is to use the following approach.

number

[0061] FIG. 8 is a diagram illustrating an example of an application result obtained by applying fitting according to the predictive control method according to Example 2. FIG. 8 shows a comparison with FIG. 5B. Unlike FIG. 5B, since the plot PL2 reaches the dotted line LN1, it can be confirmed that sufficient fitting has been achieved. It can also be confirmed that fitting is fast. It can be expected that the final fitting can be improved by providing appropriate intermediate auxiliary information.

[0062] As described above, in the predictive control method according to Example 2, the intermediate feature quantities (particularly the second intermediate feature quantity RE2) of the structure prediction model MO11 based on machine learning are updated, and the target auxiliary information E T predicts a target structure based on. Further, in the predictive control method according to Example 2, a predicted structure S predicted from the structure prediction model MO11 at a certain time point t corresponding auxiliary information E ′ t and the target auxiliary information E T is interpolated with pseudo intermediate auxiliary information. Further, in the predictive control method according to Example 2, the quality of fitting is improved by sequentially performing fitting with the interpolated intermediate auxiliary information as a target. Further, in the predictive control method according to Example 2, the predicted structure S by the structure prediction model MO11 t density map (auxiliary information E ′ t ) and the density map of the target structure (target auxiliary information E T ) by performing stepwise fitting toward the density map (intermediate auxiliary information) of the intermediate structure generated based on said (density map of the target structure), thereby improving the prediction accuracy of the final target structure.

[0063] [3. Configuration of Information Processing Apparatus (Example 1)] FIG. 9 is a diagram illustrating an example of a functional configuration of the information processing apparatus according to Example 1. The information processing apparatus 1 illustrated in FIG. 9 is an example of a computer that executes predictive control processing. As illustrated in FIG. 9, the information processing apparatus 1 includes a control unit 10 and a storage unit 20.

[0064] The storage unit 20 contains a Protein Data Bank (PDB) file 21, an Electron Microscope Data Bank (EMDB) file 22, and EM data 23.

[0065] PDB file 21 is a file that stores information on the three-dimensional atomic structure of a protein. For example, PDB file 21 contains information on the atomic structures of the three monomers that make up the protein, as well as information on which chain each atom belongs to. PDB file 21 can be obtained from a PDB available on the web.

[0066] EMDB file 22 is a file containing a collection of EM images acquired by electron microscopes such as cryo-electron microscopes, which specify the three-dimensional atomic structure of proteins. EM images are synonymous with, for example, three-dimensional density maps. EMDB file 22 can be obtained from the EMDB database available on the web.

[0067] EM data 23 is a three-dimensional density map of the atomic structure predicted by predictive control processing. EM data 23 is stored in the storage unit 20 by the output unit 17, which will be described later.

[0068] The control unit 10 includes a plurality of structural prediction models MO11, a preprocessing unit 12, a conversion unit 13, a difference calculation unit 14, a backpropagation unit 15, an update unit 16, and an output unit 17.

[0069] The MO11 structural prediction model predicts the atomic structure of a protein from amino acid sequence information. Each MO11 structural prediction model corresponds to a single monomer. If the amino acid sequence information is for a complex protein, each MO11 structural prediction model predicts the atomic structure of the monomers that make up the protein, chain by chain. In Example 1, the case where the target protein is a trimer is described.

[0070] The preprocessing unit 12 performs preprocessing for predictive control.

[0071] As an example of the first preprocessing step, the preprocessing unit 12 uses the PDB file 21 to find three rigid body transformations that match each chain of the atomic structure of each of the three predicted monomers (trimers) by point cloud alignment. This is to find rigid body transformations that correspond to separate chains. The PDB file 21 contains information on the atomic structures of the three monomers that make up the protein, as well as information on which chain each atom belongs to. Therefore, by using the PDB file 21, the preprocessing unit 12 can easily perform point cloud alignment for each chain of the atomic structure of each of the predicted trimers and find three rigid body transformations that fit each chain. Then, the preprocessing unit 12 applies the found rigid body transformations to each chain of the atomic structure of each of the three predicted monomers and combines them into a single complex.

[0072] As a second example of preprocessing, the preprocessing unit 12 determines the rigid body transformation that best matches the combined complex with the 3D density map of the target in the EMDB file 22. That is, the preprocessing unit 12 determines the rigid body transformation used to align the atomic structure with the 3D density map of the target. The equation for the alignment is expressed, for example, as in equation (5). Rc and tc are rigid body transformations representing rotation and centroid (translation), respectively. xca and x'ca represent the atomic coordinates of the original atom a and the atomic coordinates of atom a after the rigid body transformation, respectively.

number

[0073] For example, in the second preprocessing step, the preprocessing unit 12 performs the first and second steps. In the first step, the preprocessing unit 12 uses the target's three-dimensional density map to perform centroid alignment (translation) and principal component axis alignment (rotation) on the combined complex, determining the approximate tc and Rc. In the second step, based on equation (5), the preprocessing unit 12 fine-tunes the alignment with Rc and tc as variables and xca as a constant (fixed) to determine tc and Rc. In other words, in the second preprocessing step, without using information about which chain each monomer atom originally belonged to, the preprocessing unit 12 determines how to translate and rotate the rigid body of the combined monomer complex to best match the target's three-dimensional density map. This allows the preprocessing unit 12 to align the atomic structure with the target's three-dimensional density map.

[0074] Then, after completing the preprocessing, the preprocessing unit 12 forward-propagates the layers of the structure prediction model MO11 for each chain. The equation for forward-propagating the layers is, for example, expressed by equation (6). Here, rci represents the intermediate features for each chain. i is the index representing each chain. xca represents the atomic coordinates of the predicted atomic structure. When the preprocessing unit 12 forward-propagates the layers, the intermediate features of the structure prediction model MO11 are not updated.

number

[0075] The second preprocessing step was divided into two stages, the first and second stages, to determine the rigid body transformation that best matched the combined complex and the target's 3D density map. However, the second preprocessing step is not limited to this; a genetic algorithm may also be used to determine the rigid body transformation that best matched the combined complex and the target's 3D density map.

[0076] The conversion unit 13 converts the atomic structure output as a prediction result from the structural prediction model MO11 into a three-dimensional density map (EM) that limits the atomic structure in a differentiable form.

[0077] For example, the transformation unit 13 applies the rigid body transformation obtained by the preprocessing unit 12 to the predicted atomic structure to determine the atomic structure of one complex (trimer). That is, the transformation unit 13 uses equation (5) to determine the atomic structure of one complex (trimer) from the predicted atomic structure.

[0078] The conversion unit 13 then uses interpolation equations that satisfy the density conservation law to convert the predicted atomic structure of the trimer into a three-dimensional density map that limits that atomic structure. Interpolation equations that satisfy the density conservation law can be expressed, for example, by equations (7), (8), and (9) below.

[0079] In equation (7), dN represents the length of one side of the voxel. x'Ngk and x'cak represent the k-component of the coordinates of the voxel vertex and the k-component of the atomic coordinates, respectively. uNgcak represents the distance between the normalized atom a and the voxel vertex g.

number

[0080] In equation (8), uNgcak represents the normalized distance between atom a and the voxel vertex g, which is the result of the calculation in equation (7). pNgcak represents the probability of atom a being at the voxel vertex g.

number

[0081] In equation (9), na represents the atomic number of atom a. pNgcak represents the probability of atom a being at vertex g of the voxel, which is the result of the calculation in equation (8). pNgpred represents the calculated 3D density map.

number

[0082] Here, Figure 10 shows the relationship between the distance uNgcak between atom a and the voxel vertex g, as shown in equation (7), and the probability of atom a being at the voxel vertex g, PNGcak, as shown in equation (8). Figure 10 is a diagram showing the spread of electrons. The x-axis of the graph shown in Figure 10 is a set of values ​​obtained from the equation obtained by removing the absolute value of the molecule shown on the right side of equation (7). In other words, the x-axis is a set of values ​​that normalize how close an atom is to the voxel vertex. The y-axis shows the probability of atom a being at the voxel vertex g, PNGcak, as shown in equation (8). From this graph, it can be seen that the closer an atom is to the voxel vertex (the closer the x-axis value is to 0), the higher the probability of electron existence, and the further the atom is from the voxel vertex (the further the x-axis value is from 0), the lower the probability of electron existence. Furthermore, if the value of the x-axis is greater than "2" or less than "-2", the probability of electron existence on the y-axis becomes "0". Therefore, the transformation unit 13 can convert the atom into a smooth, blurry image (3D density map) by using the probability of existence PNgcak of atom a at the vertex g of voxel, which is obtained from equations (7) and (8) (the concept of a filter). In other words, the transformation unit 13 realizes a transformation process that integrates the electron spread and the concept of a filter into a single process.

[0083] Thus, the transformation unit 13 uses equations (7), (8), and (9) to transform the predicted atom a of the trimer into a three-dimensional density map pNgpred that restricts that atom a. These equations (7), (8), and (9) are transformation functions from atomic structure to a three-dimensional density map that restricts that atomic structure, and are differentiable transformation functions.

[0084] The difference calculation unit 14 calculates the difference between the 3D density map corresponding to the atomic structure of the predicted trimer and the 3D density map of the target in the EMDB file 22. For example, the difference calculation unit 14 calculates the difference between the 3D density map that limits the atomic structure of the predicted result and the 3D density map of the target by determining the cross-correlation. The formula for determining the cross-correlation is expressed, for example, by the following formula (10).

[0085] The pNgpred used in equation (10) represents the 3D density map calculated in equations (7), (8), and (9). pNgtarg represents the 3D density map of the target. LN represents the cross-correlation value of the 3D density map. g and N represent the number of voxel vertices and voxel divisions, respectively.

number

[0086] Thus, the difference calculation unit 14 uses equation (10) to calculate the difference between the 3D density map of atom a of the predicted trimer and the 3D density map of the target. Equation (10) is a differentiable equation. Although the equation for calculating the difference has been described as an equation for finding cross-correlation, it is not limited to this. The equation for calculating the difference may also utilize the L2 norm or the L1 norm.

[0087] Furthermore, the difference calculation unit 14 uses the calculated difference to calculate an objective function that shows constraints on the atomic structure of the predicted trimer. In addition, the difference calculation unit 14 adds constraints that the protein should have to the calculated objective function to calculate the final objective function. For example, the difference calculation unit 14 calculates the final objective function using the following equation (11).

[0088] The LN used in equation (11) is the value calculated in equation (10). Lboundlength and Lboundangle are examples of objective functions that maintain the distance and angle of the peptide bond, respectively. N represents the number of voxel divisions. Ltotal is the final objective function.

number

[0089] The first equation on the right-hand side of equation (11) is the main objective function. This first equation uses the technique of mixing density and coarseness in domain partitioning to smoothly reduce the difference.

[0090] Thus, the difference calculation unit 14 uses equation (11) to calculate an objective function that shows the constraints on atom a of the predicted trimer. Equation (11) is a differentiable equation. The constraints to be added are not limited to, for example, the distance and angle of the peptide bond, but may also be excluded volume effect, disulfide bond distance, hydrogen bond distance, etc. In short, the difference calculation unit 14 only needs to add constraints that the protein must have as options to the main objective function.

[0091] The backpropagation unit 15 calculates the backpropagation of the difference. For example, the backpropagation unit 15 calculates the backpropagation of the difference using the following equations (12) to (17). Equation (12) is the derivative of Ltotal shown in equation (11) with respect to LN. Equation (13) is the derivative of LN shown in equation (10) with respect to PNGpred. Equation (14) is the derivative of PNGpred shown in equation (9) with respect to PNGcak. Equation (15) is the derivative of PNGcak shown in equation (8) with respect to uNgcak. Equation (16) is the derivative of uNgcak shown in equation (7) with respect to x'cak. And equation (17) is the derivative of LN shown in equation (10) with respect to x'cak.

[0092]

number

[0093]

number

[0094]

number

[0095]

number

[0096]

number

[0097]

number

[0098] Furthermore, the backpropagation unit 15 calculates the intermediate features for each chain using the gradient obtained by backpropagation of the difference. For example, the backpropagation unit 15 calculates the intermediate features for each chain using the gradient obtained by backpropagation of the difference (equation (17)) according to the following equation (18). In equation (18), rci represents the intermediate features for each chain. i is the index representing each chain. Also, δxca / δrci is the derivative of the atomic coordinate xca shown in equation (6) with respect to the intermediate features rci of each chain.

number

[0099] The update unit 16 updates the intermediate features of each chain in the structure prediction model MO11 corresponding to each chain. In other words, the update unit 16 updates the intermediate features of the structure prediction model MO11 so that the difference between the predicted atomic structure and the target atomic structure becomes smaller. To put it another way, the update unit 16 updates the intermediate features of the structure prediction model MO11 based on constraints from the target's 3D density map.

[0100] The output unit 17 stores the 3D density map (EM) converted by the conversion unit 13 as EM data 23 in the storage unit 20.

[0101] (Example of output atomic structure display) Figure 11 shows an example of the output atomic structure display. The image in Figure 11 shows the atomic coordinates x'ca (see equation (9)) of the atomic structure output from the structure prediction model MO11 displayed on the screen. These atomic coordinates x'ca of the atomic structure are the atomic coordinates after rigid body transformation has been performed on each chain. The user can use the atomic coordinates of the atomic structure displayed on the screen to fit (adjust) it to a 3D density map.

[0102] In Example 1, the limiting information for defining the atomic structure was described as a three-dimensional density map (EM image). However, the limiting information for defining the atomic structure is not limited to a three-dimensional density map (EM image) captured by an electron microscope, but may also be a three-dimensional density map analyzed by X-rays. Furthermore, the limiting information for defining the atomic structure may also be the coordinates of the electronic structure.

[0103] Here, an example of displaying the output atomic structure when the limiting information for defining the atomic structure is the coordinates of the electronic structure will be explained with reference to Figure 12. Figure 12 is a diagram showing an example of displaying the output atomic structure coordinates. The image in Figure 12 shows the coordinates of the atomic structure output from the structure prediction model MO11 displayed on the screen (solid line). These atomic structure coordinates are the coordinates after rigid body transformation to each chain. In addition, the image in Figure 12 shows the coordinates of the target limiting information displayed on the screen (dashed line). Note that when the limiting information for defining the atomic structure is the coordinates of the electronic structure, the target limiting information can be obtained from PDB file 21. The user can use the displayed atomic structure coordinates (solid line) output from the structure prediction model MO11 to match the target limiting information coordinates (dashed line).

[0104] Furthermore, Figure 12 displays only the carbon atom (Cα) at ​​the center of each amino acid. The number of Cα atoms is about one-tenth of the total number of atoms. The coordinates and line structure of the atomic structure output from the structure prediction model MO11 contain information on all atoms, not just the displayed Cα. On the other hand, the coordinates and line structure of the target's limited information are obtained from the PDB file 21, and although it includes information on atoms other than Cα, only Cα is used as limited information. In other words, even if only about one-tenth of the total number of atoms can be used as limited information, the user can reconstruct the information on all atoms using the atomic structure output from the structure prediction model MO11 as a hint.

[0105] (Flowchart of predictive control processing) Here, the flowchart of the predictive control process performed by the information processing device 1 will be explained with reference to Figure 13. Figure 13 is a diagram showing an example of the flowchart of the predictive control process according to Embodiment 1.

[0106] As shown in Figure 13, the information processing device 1 predicts the monomer structure for each chain from the amino acid sequence (step S11). For example, the structure prediction model MO11 predicts the atomic structure of the monomers that make up the protein for each chain from the amino acid sequence.

[0107] As a preprocessing step, the information processing device 1 uses the PDB file 21 to determine three rigid body transformations that match each chain of the predicted three monomer (trimer) structures (step S12). For example, the preprocessing unit 12 performs the first preprocessing step. Then, the information processing device 1 applies the rigid body transformations obtained for each chain and combines them into a single complex (step S13).

[0108] Then, as a preprocessing step, the information processing device 1 determines the rigid body transformation Rc and tc for the complex that match the target EM (3D density map) (step S14). For example, the preprocessing unit 12 performs a second preprocessing step.

[0109] Then, the information processing device 1 propagates each layer forward for each chain (step S15). For example, the information processing device 1 executes equation (6). Here, the information processing device 1 does not update the intermediate feature rci.

[0110] Then, the information processing device 1 applies the rigid body transformation obtained in the preprocessing steps S12 and S14 to determine the atomic structure of the trimer (step S16). For example, the information processing device 1 executes equation (5).

[0111] Then, the information processing device 1 converts the atomic structure of the trimer into a three-dimensional density map (step S17). For example, the information processing device 1 executes equations (7), (8), and (9).

[0112] Then, the information processing device 1 calculates the difference between the converted 3D density map and the target EM (3D density map) (step S18). For example, the information processing device 1 executes equation (10).

[0113] Then, the information processing device 1 calculates the objective function using the obtained difference (step S19). For example, the information processing device 1 executes equation (11).

[0114] Then, the information processing device 1 calculates backpropagation (S19→S18→S17→S16→S15) and obtains intermediate features (step S20). For example, the information processing device 1 calculates backpropagation and obtains intermediate features by executing equations (12) to (16). Then, the information processing device 1 updates the structure prediction model MO11 with the obtained intermediate features (step S21).

[0115] Then, the information processing device 1 determines whether the structural prediction by the structural prediction model MO11 has converged (step S22). If it determines that the structural prediction has not converged (step S22; No), the information processing device 1 proceeds to step S15 to perform the next structural prediction.

[0116] On the other hand, if it is determined that the structure prediction has converged (step S22; Yes), the information processing device 1 terminates the prediction control process.

[0117] As described above, the information processing device 1 according to Example 1 updates the intermediate features of the structure prediction model MO11 so that the difference between the 3D density map obtained by transforming the predicted structure output as a prediction result from the structure prediction model MO11 using an interpolation formula that satisfies the density conservation law and a different 3D density map becomes small. Therefore, the information processing device 1 according to Example 1 can predict a variety of atomic structures. Furthermore, by updating the intermediate features of the structure prediction model MO11 so that the difference between the 3D density map obtained by transforming the predicted structure and the correct 3D density map becomes small, the information processing device 1 can predict a variety of atomic structures in a realistic amount of time.

[0118] (Variation of Example 1) Incidentally, in the information processing device 1 according to Example 1, the conversion unit 13 was described as converting the atomic structure output as a prediction result from the structure prediction model MO11 into a three-dimensional density map (EM) that limits the atomic structure using an interpolation formula that satisfies the density conservation law in a differentiable form. However, the conversion unit 13 is not limited to an interpolation formula that satisfies the density conservation law, and may, for example, use an approximation formula based on a spherically symmetric Gaussian distribution to convert into a three-dimensional density map (EM) that limits the atomic structure.

[0119] When molecular shape is used as limiting information to restrict the atomic structure, a molecular electron density map (3D density map) can be used. A molecular 3D density map can be accurately calculated based on physics by using given atomic positions and atomic scattering factors. However, the 3D density map obtained from atomic scattering factors has many undulations, and when used as limiting information for searching, it can hinder the search, for example, by getting stuck in local optima.

[0120] Therefore, by approximating the functional form of the atomic scattering factor and pre-smoothing the 3D density map converted from the approximated functional form, it is possible to improve the smoothness of the search while maintaining the accuracy of the molecular shape.

[0121] The atomic scattering factor f(q) is expressed by four Gaussian functions and a constant term, as shown in equation (19). Furthermore, the atomic scattering factor f(q) can be approximated by four or fewer Gaussian functions, as shown in equation (20). Note that N is approximated by 1, 2, 3, and 4, respectively. a, b, and c are fitting parameters and may be newly fitted.

number

number

[0122] Furthermore, as an example of a method for smoothing a 3D density map (making it a smooth image), a low-pass filter in wavenumber space (cutoff wavenumber f c The following can be used: When using a low-pass filter in wavenumber space, the number of voxels N per dimension can be given by the following equation (21), where R[A] is the target resolution (degree of smoothing) and L[A] is the molecular size parameter. The molecular size parameter L can be predicted with high accuracy by using an accurately calculated 3D density map and the molecular surface defined therefrom. N=2L / R=Lf c ...Equation (21)

[0123] Thus, the transformation unit 13 uses equations (20) and (21) to transform the predicted atomic structure of the trimer into a three-dimensional density map that limits that atomic structure. Equation (20) is a transformation function from atomic structure to a three-dimensional density map that limits that atomic structure, and it is a differentiable transformation function.

[0124] As a result, the information processing device 1 in this variation updates the intermediate features of the structure prediction model MO11 so that the difference between the 3D density map obtained by transforming the predicted structure output as a prediction result from the structure prediction model MO11 using an approximation formula based on a spherically symmetric Gaussian distribution and a different 3D density map becomes small. Therefore, the information processing device 1 in this variation becomes capable of predicting a variety of atomic structures. Furthermore, by updating the intermediate features of the structure prediction model MO11 so that the difference between the 3D density map obtained by transforming the predicted structure and the ground truth 3D density map becomes small, the information processing device 1 becomes capable of predicting a variety of atomic structures in a realistic amount of time.

[0125] [4. Configuration of the Information Processing Device (Example 2)] FIG. 14 is a diagram illustrating an example of a functional configuration of an information processing apparatus according to the second embodiment. The information processing apparatus 1A illustrated in FIG. 14 is an example of a computer that executes prediction control processing. As illustrated in FIG. 14, the information processing apparatus 1A includes a control unit 10A and a storage unit 20A. Descriptions that are the same as those in the first embodiment will be omitted as appropriate.

[0126] The storage unit 20A includes a PDB file 21, an EMDB file 22, EM data 23, and intermediate auxiliary information 24. Descriptions of the PDB file 21, the EMDB file 22, and the EM data 23 are the same as those in the first embodiment, and thus will be omitted.

[0127] The intermediate auxiliary information 24 is intermediate auxiliary information generated by prediction control processing. The intermediate auxiliary information 24 may be intermediate auxiliary information collectively generated based on a starting structure, or may be intermediate auxiliary information dynamically generated in a stepwise fitting process. The intermediate auxiliary information 24 is stored in the storage unit 20A by the output unit 17A.

[0128] The control unit 10A includes a plurality of structure prediction models MO11, a preprocessing unit 12A, a conversion unit 13, a difference calculation unit 14A, a back propagation unit 15, an update unit 16A, and an output unit 17A. Descriptions of the conversion unit 13 and the back propagation unit 15 are the same as those in the first embodiment, and thus will be omitted.

[0129] The preprocessing unit 12A generates intermediate auxiliary information, for example, in addition to the processing performed by the preprocessing unit 12. For example, the preprocessing unit 12A generates a predicted structure S by the structure prediction model MO11 t corresponding auxiliary information E ′ t (corresponding to first information) and target auxiliary information E that is a fitting target of the auxiliary information E T (corresponding to second information), based on which, the auxiliary information E ′ t and the target auxiliary information E T as intermediate auxiliary information, intermediate auxiliary information E t+1 to intermediate auxiliary information E T-1 (corresponding to intermediate information) is generated. For example, the preprocessing unit 12A generates the intermediate auxiliary information based on a current predicted structure S tStarting with this as the initial action, intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 Generates the predicted structure S. t If the initial prediction structure is (i.e., prediction structure S0), the preprocessor 12A processes intermediate auxiliary information E1 to intermediate auxiliary information E T-1 Generates.

[0130] The preprocessing unit 12A acquires, for example, first information and second information which is the fitting target of the first information. For example, the preprocessing unit 12A acquires a prediction structure S t The system obtains two pieces of information: the first piece of information, which is a converted 3D density map, and the second piece of information, which is a 3D density map input by the user along with the amino acid sequence as the correct 3D density map.

[0131] The preprocessing unit 12A processes, for example, intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 These may be generated all at once or dynamically. If generated all at once, intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 This is determined first. On the other hand, when generated dynamically, intermediate auxiliary information E t+1 This is determined first, and after the initial fitting, intermediate auxiliary information E t+2 Once determined, intermediate auxiliary information E is provided after the second fitting. t+3 This is determined, and by repeating this T times, intermediate auxiliary information E is obtained. t+1 ~Intermediate Supplemental Information E T-1 It is determined.

[0132] The preprocessing unit 12A performs, for example, the index d(E t ,E T Intermediate auxiliary information E such that ) satisfies equation (2) t+1 ~Intermediate Supplemental Information E T-1 The preprocessor 12A generates the following: For example, the intermediate auxiliary information that will be used for fitting is the final target auxiliary information E. T Intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 Generates.

[0133] The preprocessing unit 12A processes intermediate auxiliary information E based on, for example, a method based on linear interpolation between density maps or a method based on density interpolation based on optimal transport. t+1 ~Intermediate Supplemental Information E T-1 It generates all of the above at once. In addition, the preprocessing unit 12A generates, for example, auxiliary information E ′ t-1 and target auxiliary information E T Methods based on linear interpolation, and auxiliary information E ′ t-1 As intermediate auxiliary information in the vicinity, multiple candidate intermediate auxiliary information are generated and index d(E t ,E T The candidate for the intermediate auxiliary information that minimizes ) is the intermediate auxiliary information E t Based on methods such as the above, intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 It generates dynamically.

[0134] The preprocessing unit 12A, for example, processes the current predicted structure S t Based on linear interpolation between the 3D density maps of the first information (including the prediction structure S0 of the initial prediction) and the second information, one or more 3D density maps are generated at once as intermediate auxiliary information. The preprocessing unit 12A generates, for example, the current prediction structure S t Based on density interpolation between 3D density maps that correspond to the first information (including the prediction structure S0 of the initial prediction) and the second information, one or more 3D density maps are generated simultaneously as intermediate auxiliary information.

[0135] The preprocessing unit 12A, for example, processes the current predicted structure S t In the process of stepwise fitting based on linear interpolation between the 3D density maps of the first information (including the prediction structure S0 of the initial prediction) and the second information, one or more 3D density maps are dynamically generated as intermediate auxiliary information. The preprocessing unit 12A generates, for example, the current prediction structure S t Multiple candidates for intermediate auxiliary information are generated around the first information (including the prediction structure S0 of the initial prediction), and an index d(E) is used to indicate the ease of reaching the second information. t ,E TIn the process of fitting step by step, using the candidate that minimizes ) as intermediate auxiliary information, one or more 3D density maps are dynamically generated as intermediate auxiliary information.

[0136] The preprocessing unit 12A, for example, uses an index d(E) that indicates the ease of accessing the second information. t ,E T Intermediate auxiliary information is generated so that ) satisfies predetermined conditions. The preprocessor 12A generates, for example, an index d(E) based on the cross-correlation between 3D density maps. t ,E T Intermediate auxiliary information is generated based on the following: For example, the preprocessor 12A generates an index d(E) based on the Wasserstein distance. t ,E T Intermediate auxiliary information is generated based on the following. The preprocessor 12A generates an index d(E) based on the inter-structure distance, which is based on the structure probability estimated by a structure probability distribution estimation method in a 3D density map space. t ,E T Intermediate auxiliary information is generated based on ). The preprocessor 12A generates intermediate auxiliary information such that, for example, under predetermined conditions, the intermediate auxiliary information leading to the next fitting is more likely to reach the second information.

[0137] The difference calculation unit 14A performs the same processing as the difference calculation unit 14, but unlike the difference calculation unit 14, it calculates the difference between the 3D density map corresponding to the atomic structure of the predicted trimer and the 3D density map of the intermediate structure, rather than the difference between the 3D density map corresponding to the atomic structure of the predicted trimer and the target 3D density map in the EMDB file 22. In other words, in Example 2, unlike Example 1, the 3D density map that is the target of fitting is the target auxiliary information E T Rather, it serves as intermediate auxiliary information. Prediction structure S t If the initial prediction structure is (i.e., prediction structure S0), the difference calculation unit 14A calculates the auxiliary information E ′ The difference between 0 and the intermediate auxiliary information E1 is calculated. Prediction structure S t If the prediction structure is S1, the difference calculation unit 14A calculates the auxiliary information E ′ Calculate the difference between 1 and the intermediate auxiliary information E2.

[0138] The update unit 16A, for example, in addition to the processing of the update unit 16, also processes the intermediate auxiliary information E generated by the preprocessing unit 12A. t+1 ~Intermediate Supplemental Information E T-1 Based on this, the intermediate features of the structural prediction model MO11 are updated after processing by the difference calculation unit 14A and the backpropagation unit 15. The update unit 16A updates intermediate features such as the first intermediate feature RE1 and the second intermediate feature RE2. Predicted structure S t If the initial prediction structure is (i.e., prediction structure S0), the update unit 16A will update the auxiliary information E ′ The intermediate feature H0 is updated based on 0 and the intermediate auxiliary information E1. Prediction structure S t If the predictive structure is S1, the update unit 16A will provide auxiliary information E ′ The intermediate feature H1 is updated based on 1 and the intermediate auxiliary information E2. The update unit 16A updates the intermediate feature so that the difference is minimized by backpropagation based on the difference between the intermediate auxiliary information to be fitted next and the first information.

[0139] The update unit 16A, for example, displays target auxiliary information E T When fitting to the target, the current predicted structure S is evaluated at each fixed number of iterations (m times). t Corresponding auxiliary information E ′ t and target auxiliary information E T The cross-correlation C1 with the function is calculated. The update unit 16A, for example, calculates the updated intermediate feature H when the calculated cross-correlation C1 exceeds a predetermined threshold TH1, or when a certain number of iterations (n ​​times) have been completed. t And, the predicted structure S t Corresponding auxiliary information E ′ t Save the data and perform sequential fitting, setting t to t+1.

[0140] For example, in addition to the processing performed by the output unit 17, the output unit 17A stores the 3D density map converted by the conversion unit 13 as EM data 23 in the storage unit 20A.

[0141] (Flowchart of predictive control processing) Here, the flowchart of the predictive control process performed by the information processing device 1A will be explained with reference to Figure 15. Figure 15 is a diagram showing an example of the flowchart of the predictive control process according to Embodiment 2.

[0142] As shown in Figure 15, the information processing device 1A receives the amino acid sequence IN1 and target auxiliary information E T Obtain the following (step S101).

[0143] The information processing device 1A receives the acquired amino acid sequence IN1 and target auxiliary information E T This is input into the structural prediction model MO11 to output the predicted structure S0 of the initial prediction (step S102).

[0144] The information processing device 1A sets t=0 (step S103).

[0145] The information processing device 1A updates the intermediate features n times. ′ to n ′ Assuming = 0, Intermediate auxiliary information E t+1 ~Intermediate Supplemental Information E T-1 Generate (step S104).

[0146] The information processing device 1A performs fitting a certain number of times (m times) and generates intermediate feature quantities H m Prediction structure S m and supplementary information E ′ m Obtain (step S105).

[0147] The information processing device 1A determines whether t=T and whether the cross-correlation C1 exceeds the threshold TH1 (step S106).

[0148] If step S106 is No, then the information processing device 1A ′ Whether or not a certain number of times (n times) has been reached (n ′ Determine if =n?) (Step S107).

[0149] Furthermore, if step S107 is No, the information processing device 1A will n ′ n ′ Substitute +1 (n ′ ←n ′ +1) (step S108), and return to step S105.

[0150] On the other hand, if step S106 is Yes, or if step S107 is Yes, the information processing device 1A substitutes t+1 for t (t←t+1) (step S109).

[0151] Then, the information processing device 1A determines whether t has reached T+1 (t=T+1?) (step S110).

[0152] If step S110 is No, the information processing device 1A will H t H m Substitute (H t ←H m ), S t S m Substitute (S t ←S m ) and E ′ t E ′ m Substitute (E ′ t ←E ′ m ) then (step S111), and return to step S104.

[0153] On the other hand, if step S110 is Yes, the information processing device 1A predicts the structure S T The output is (step S112). In this way, the information processing device 1A outputs a prediction structure S based on the intermediate features after a predetermined number of updates. T This is output as a target structure corresponding to the input amino acid sequence.

[0154] Now, while embodiments of the disclosed apparatus have been described, the present invention may be implemented in various other forms besides those described above. Therefore, other embodiments included in the present invention will be described below.

[0155] The processing procedures, control procedures, specific names, and various data and parameters shown in the documents and drawings of Examples 1 and 2 above may be changed at will unless otherwise specified.

[0156] Furthermore, the specific forms of distribution and integration of the components of each device are not limited to those shown in the diagram. In other words, all or part of the components may be functionally or physically distributed and integrated in any unit depending on various loads and usage conditions. Moreover, all or any part of the processing functions of each device may be implemented by a CPU and the program that is analyzed and executed by the CPU, or by hardware using wired logic.

[0157] The various processes described in Examples 1 and 2 above can be implemented by executing a pre-prepared program on a computer such as a personal computer or workstation. Therefore, below, an example of a computer that executes a control program having the same functions as in Examples 1 and 2 will be described using Figure 16.

[0158] Figure 16 shows an example of a hardware configuration. As shown in Figure 16, the computer 100 has an operating unit 110a, a speaker 110b, a camera 110c, a display 120, and a communication unit 130. Furthermore, the computer 100 has a CPU 150, a ROM 160, an HDD 170, and RAM 180. These parts 110 to 180 are connected via a bus 140.

[0159] As shown in Figure 16, HDD170 stores a predictive control program 170a that performs the same functions as the structural prediction model MO11, pre-processing unit 12, pre-processing unit 12A, conversion unit 13, difference calculation unit 14, difference calculation unit 14A, backpropagation unit 15, update unit 16, update unit 16A, output unit 17, and output unit 17A shown in Examples 1 and 2 above. This predictive control program 170a may be integrated or separated, similar to the components of the structural prediction model MO11, pre-processing unit 12, pre-processing unit 12A, conversion unit 13, difference calculation unit 14, difference calculation unit 14A, backpropagation unit 15, update unit 16, update unit 16A, output unit 17, and output unit 17A shown in Figures 9 and 14. In other words, HDD170 does not necessarily have to store all the data shown in Examples 1 and 2 above; it is sufficient that the data used for processing is stored in HDD170.

[0160] Under these conditions, the CPU 150 reads the predictive control program 170a from the HDD 170 and then loads it into the RAM 180. As a result, the predictive control program 170a functions as a predictive control process 180a, as shown in Figure 16. This predictive control process 180a loads various data read from the HDD 170 into the memory area of ​​the RAM 180 allocated to the predictive control process 180a, and then executes various processes using the loaded data. For example, the processes that the predictive control process 180a executes may include those shown in Figures 13 and 15. It should be noted that the CPU 150 does not necessarily need to operate all the processing units shown in the above-described examples 1 and 2; it is sufficient if the processing units corresponding to the processes to be executed are virtually implemented.

[0161] The predictive control program 170a described above does not necessarily have to be stored in the HDD 170 or ROM 160 from the beginning. For example, the predictive control program 170a could be stored on a "portable physical medium" such as a flexible disk, CD-ROM, DVD disk, magneto-optical disk, or IC card inserted into the computer 100. The computer 100 could then retrieve and execute the predictive control program 170a from these portable physical media. Alternatively, the predictive control program 170a could be stored on another computer or server device connected to the computer 100 via a public network, the internet, LAN, WAN, etc. The computer 100 could then download and execute the predictive control program 170a stored in this manner.

[0162] The following additional information is disclosed regarding the above embodiments. (1) We obtain first information corresponding to the predicted structure by a machine learning-based structural prediction model, and second information which is the target for fitting the first information. Based on the acquired first information and second information, intermediate information between the first information and the second information is generated. Based on the generated intermediate information, update the intermediate features of the structure prediction model. A predictive control method in which a computer performs a process. (2) The first information is a three-dimensional density map obtained by transforming the predicted structure, The aforementioned second piece of information is a correct 3D density map entered by the user. The predictive control method described in (1) above. (3) The structure prediction model outputs the predicted structure based on the intermediate features after a predetermined number of updates as a target structure corresponding to the input amino acid sequence. The predictive control method described in (1) or (2) above. (4) The aforementioned update process is: Next, based on the difference between the intermediate information going to fitting and the first information, the intermediate features are updated by backpropagation to minimize the difference. A predictive control method described in any one of (1) to (3) above. (5) The aforementioned generation process is, The intermediate information is generated such that an indicator showing the ease of accessing the second information satisfies predetermined conditions. A predictive control method described in any one of (1) to (4) above. (6) The aforementioned generation process is, As a predetermined condition, the intermediate information is generated such that the intermediate information leading to the next fitting process can more easily reach the second information. The predictive control method described in (5) above. (7) The aforementioned generation process is, The intermediate information is generated based on the index based on the cross-correlation between 3D density maps, the index based on the Wasserstein distance, or the index based on the inter-structure distance estimated by a structure probability distribution estimation method in the 3D density map space. The predictive control method described in (5) or (6) above. (8) The aforementioned generation process is, Based on linear interpolation between the 3D density maps of the first information corresponding to the prediction structure of the initial prediction and the second information, or density interpolation between the 3D density maps based on optimal transport between the first information corresponding to the prediction structure of the initial prediction and the second information, one or more 3D density maps are generated collectively as intermediate information. A predictive control method according to any one of (1) to (7) above. (9) The aforementioned generation process is, Based on linear interpolation between three-dimensional density maps of the first information corresponding to the current predicted structure and the second information, or by generating a plurality of candidates of the intermediate information around the first information corresponding to the current predicted structure, and setting the candidate with the smallest index indicating ease of reaching the second information as said intermediate information, in the stepwise fitting process, one or more three-dimensional density maps are dynamically generated as said intermediate information The prediction control method according to any one of (1) to (7) above. (10) The structure prediction model is a model that, when an amino acid sequence is input, outputs the predicted structure which is three-dimensional structure data of an organic compound The prediction control method according to any one of (1) to (9) above. (11) The structure prediction model includes a plurality of layers, and an intermediate feature amount is generated in each layer of the plurality of layers The prediction control method according to any one of (1) to (10) above. (12) Acquiring first information corresponding to a structure predicted by a structure prediction model based on machine learning, and second information that is a fitting target of the first information, Generating intermediate information between the first information and the second information based on the acquired first information and the second information, Updating an intermediate feature amount of the structure prediction model based on the generated intermediate information An information processing apparatus comprising a control unit that executes the processing. (13) Acquiring first information corresponding to a structure predicted by a structure prediction model based on machine learning, and second information that is a fitting target of the first information, Generating intermediate information between the first information and the second information based on the acquired first information and the second information, Updating an intermediate feature amount of the structure prediction model based on the generated intermediate information A prediction control program that causes a computer to execute the processing.

Description of Reference Numerals

[0163] 1. 1A Information Processing Device 10, 10A Control Unit 12, 12A Pre-processing section 13 Conversion section 14, 14A difference calculation section 15 Backpropagation section 16, 16A update section 17, 17A Output Section 20, 20A storage section 21 PDB files 22 EMDB files 23 EM Data 24. Intermediate Supplemental Information

Claims

1. First information corresponding to the predicted structure by a machine learning-based structural prediction model, and second information which is the target for fitting the first information, are obtained. Based on the acquired first information and second information, intermediate information between the first information and the second information is generated. Based on the generated intermediate information, update the intermediate features of the structure prediction model. A predictive control method in which a computer performs a process.

2. The first information is a three-dimensional density map obtained by transforming the prediction structure, The aforementioned second piece of information is a correct 3D density map entered by the user. The predictive control method according to claim 1, wherein the processing is performed by a computer.

3. The structure prediction model outputs the predicted structure based on the intermediate features after a predetermined number of updates as a target structure corresponding to the input amino acid sequence. The predictive control method according to claim 1, wherein the processing is performed by a computer.

4. The aforementioned update process is: Next, based on the difference between the intermediate information going to fitting and the first information, the intermediate features are updated by backpropagation to minimize the difference. The predictive control method according to claim 1, wherein the processing is performed by a computer.

5. The aforementioned generation process is, The intermediate information is generated such that an indicator showing the ease of accessing the second information satisfies predetermined conditions. The predictive control method according to claim 1, wherein the processing is performed by a computer.

6. The aforementioned generation process is, As a predetermined condition, the intermediate information is generated such that the intermediate information leading to the next fitting process can more easily reach the second information. The predictive control method according to claim 5, wherein the processing is performed by a computer.

7. The aforementioned generation process is, The intermediate information is generated based on the index based on the cross-correlation between three-dimensional density maps, the index based on the Wasserstein distance, or the index based on the inter-structure distance based on the structure probability estimated by a structure probability distribution estimation method in the three-dimensional density map space. The predictive control method according to claim 5, wherein the processing is performed by a computer.

8. The aforementioned generation process is, Based on linear interpolation between the three-dimensional density maps of the first information corresponding to the prediction structure of the initial prediction and the second information, or density interpolation between the three-dimensional density maps based on optimal transport between the first information corresponding to the prediction structure of the initial prediction and the second information, one or more three-dimensional density maps are generated at once as intermediate information. The predictive control method according to claim 1, wherein the processing is performed by a computer.

9. The aforementioned generation process is, Based on linear interpolation between the three-dimensional density maps of the first information corresponding to the current prediction structure and the second information, or by generating multiple candidates for the intermediate information around the first information corresponding to the current prediction structure, and using the candidate with the smallest index indicating the ease of reaching the second information as the intermediate information, one or more three-dimensional density maps are dynamically generated as the intermediate information during the fitting process. The predictive control method according to claim 1, wherein the processing is performed by a computer.

10. First information corresponding to the predicted structure by a machine learning-based structural prediction model, and second information which is the target for fitting the first information, are obtained. Based on the acquired first information and second information, intermediate information between the first information and the second information is generated. Based on the generated intermediate information, update the intermediate features of the structure prediction model. An information processing device including a control unit that performs processing.

11. First information corresponding to the predicted structure by a machine learning-based structural prediction model, and second information which is the target for fitting the first information, are obtained. Based on the acquired first information and second information, intermediate information between the first information and the second information is generated. Based on the generated intermediate information, update the intermediate features of the structure prediction model. A predictive control program that directs a computer to perform a process.

Images