JP7168979B2 - 3D structure determination device, 3D structure determination method, 3D structure discriminator learning device, 3D structure discriminator learning method and program - Google Patents

Description

The present invention relates to a three-dimensional structure determination device, a three-dimensional structure determination method, a three-dimensional structure classifier learning device, a three-dimensional structure classifier learning method, and a program.

In the initial stage of drug discovery, it is necessary to discover ligands such as compounds, peptides, proteins, nucleic acids, etc. that bind to target proteins of drugs. One of screening methods for discovering such ligands is protein-ligand docking simulation. Also, the development of technology for improving the accuracy of docking simulation is underway. For example, Patent Literature 1 discloses a docking scoring method that realizes more accurate screening than conventional docking simulation.

JP 2005-181104 A

In Patent Document 1, after calculating the electronic state of the binding portion based on the three-dimensional structure of the protein, the chemical shift value is analyzed, the binding residue is determined from the chemical shift value, and the binding strength is compared. It is described that high-precision screening can be realized by this. However, considering that the number of ligands to be screened is enormous, the accuracy of Patent Document 1 and conventional docking simulations is still insufficient.

DISCLOSURE OF THE INVENTION The present invention has been made in order to solve the above-mentioned problems, and is capable of significantly improving the accuracy of determination of three-dimensional structures such as protein-ligand binding. It is an object of the present invention to provide a structural classifier learning device, a three-dimensional structure classifier learning method, and a program.

In order to achieve the above object, the three-dimensional structure determination device according to the present invention includes:
image generating means for generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure; ,
a discriminator that, when inputting one image included in the image set, discriminates whether or not the bond associated with the three-dimensional structure projected on the one image has a predetermined property;
A value obtained by inputting each of a plurality of images included in the image set to the discriminator, based on a total value for all images included in the image set, to an image included in the image set Comprehensive judgment means for judging whether or not the bond associated with the projected three-dimensional structure has the property;
Prepare.

According to the present invention, the determination accuracy of three-dimensional structures such as protein-ligand binding can be remarkably improved as compared with conventional methods.

2 is a diagram showing a functional configuration of a determination device according to Embodiment 1; FIG. FIG. 2 is a diagram explaining binding (docking) between a protein and a ligand. FIG. 4 is a diagram for explaining how the image generator according to Embodiment 1 comprehensively generates an image of the docking structure of the protein and the ligand. It is a figure explaining the processing outline of a convolutional neural network (CNN). 4 is a flowchart of learning processing according to the first embodiment; 4 is a flowchart of image generation processing according to the first embodiment; 4 is a flowchart of determination processing according to the first embodiment; It is a figure explaining the process outline|summary of 3D average pooling. FIG. 10 is a diagram for explaining how an image generation unit according to Modification 1 generates an image of a docking structure of a protein and a ligand; 9 is a flowchart of image generation processing according to Modification 1;

Hereinafter, a protein/ligand binding determination device, a protein/ligand binding discriminator learning device, and the like according to embodiments of the present invention will be described with reference to the drawings. The same reference numerals are given to the same or corresponding parts in the drawings.

(Embodiment 1)
The determination device 100 according to the first embodiment uses a large number of images of docking structures for protein/ligand binding with known activity as a discriminator learning device for protein/ligand binding, and trains a discriminator to create a learning model. get. As a protein/ligand binding determination device, the determination device 100 inputs an image of a docking structure of a protein/ligand binding with an unknown activity to a discriminator (learning model) that has been trained so that the activity is unknown. determine the presence or absence of protein-ligand binding activity. Such a determination device 100 will be described below.

The determination device 100 according to the first embodiment includes a control unit 10, a storage unit 20, an output unit 31, a communication unit 32, and an operation input unit 33, as shown in FIG.

The control unit 10 is composed of a CPU (Central Processing Unit) or the like, and by executing a program stored in the storage unit 20, each unit described later (image generation unit 11, activity acquisition unit 12, discriminator 13, discriminator It implements the functions of the learning unit 14 and the comprehensive determination unit 15).

The storage unit 20 includes a ROM (Read Only Memory), a RAM (Random Access Memory), etc., and stores programs executed by the CPU of the control unit 10 and necessary data. The storage unit 20 may also store an activity DB (database) in which the presence or absence of protein-ligand binding activity is recorded.

The output unit 31 is a device for outputting determination results of protein/ligand binding. For example, the output unit 31 is a liquid crystal display or an organic EL (Electro-Luminescence) display. However, the determination device 100 may include these displays as the output unit 31, or may include the output unit 31 as an interface for connecting an external display. If the determination device 100 includes the output unit 31 as an interface, it displays determination results and the like on an external display connected via the output unit 31 .

The communication unit 32 is a device (network interface, etc.) for transmitting/receiving data to/from other external devices (e.g., a server storing an activity DB that records the presence or absence of protein/ligand binding activity). be. The determination device 100 can acquire various data via the communication unit 32 .

The operation input unit 33 is a device that receives a user's operation input to the determination device 100, and is, for example, a keyboard, mouse, touch panel, or the like. The determination device 100 receives an instruction or the like from the user via the operation input unit 33 .

Next, functions of the control unit 10 will be described. The control unit 10 implements the functions of the image generation unit 11 , the activity acquisition unit 12 , the discriminator 13 , the discriminator learning unit 14 , and the comprehensive determination unit 15 by executing the programs stored in the storage unit 20 .

When the type of protein and the type of ligand are given, the image generator 11 obtains a three-dimensional docked structure of the given protein and ligand by docking simulation, and visualizes the docked structure from various angles around it. Generate an image projected onto a two-dimensional plane at the viewpoint. In the docking simulation, as shown in FIG. 2, given a protein 211 and a ligand 212, a three-dimensional structure (docking structure 213) in which they are bound is obtained. Then, as shown in FIG. 3, the image generation unit 11 generates an image as if the docking structure 213 was comprehensively photographed from the surrounding camera 311 at various angles (θ, φ) all around (360°). to generate Actually, the image generation unit 11 creates a 3D image from the docking structure instead of taking an image with the camera 311, and projects the 3D image onto a two-dimensional plane from viewpoints from various directions. Generate.

Specifically, Glide, for example, can be used as software for obtaining a docking structure by docking simulation. Also, PyMOL, for example, can be used as software for creating a 3D image from the docking structure. However, these software are only examples, and any software can be used as long as it can generate an image obtained by projecting onto a two-dimensional plane from viewpoints at various angles around the docking structure. Since hydrogen bonds are considered important in docking, the image generator 11 may generate an image in which hydrogen bonds are highlighted. In addition, the image generation unit 11 may generate an image in which not only hydrogen bonds but also bonds that are considered to play an important role in the properties of bonds determined by the determination device 100 are highlighted. The image generator 11 functions as an image generator.

When the type of protein and the type of ligand are given, the activity acquiring unit 12 refers to an activity DB (Database) in which the presence or absence of protein-ligand binding activity is recorded, and determines the binding of the given protein and ligand. Gets the presence or absence of activity. DUD-E, for example, can be used as such an active DB. The activity acquisition unit 12 functions as property acquisition means.

The discriminator 13 is a discriminator based on a convolutional neural network (CNN) that outputs the presence or absence of activity when an image of protein/ligand binding is given. The control unit 10 functions as a discriminator 13 by executing a program that realizes a discriminator by CNN. As shown in FIG. 4, the discriminator 13 performs convolution processing (scanning of the convolution filters 121 and 123) and pooling processing (scanning of the pooling windows 122 and 124) on the input image given to the input layer 111, and gradually Small-sized feature maps 112, 113, 114, 115, and 116 are calculated in increments, and a two-dimensional vector indicating the discrimination result is finally obtained from the output layer 119 via the fully-connected connection 125 from the feature map 116. FIG. Since the feature map 116 is a one-dimensional vector that directly affects the output (discrimination result) from the output layer 119, it will be called a discrimination vector here.

The discriminator learning unit 14 uses a large amount of teacher data consisting of protein/ligand binding images and the presence or absence of protein/ligand binding activity to train the discriminator 13 . However, the presence or absence of activity in the protein/ligand binding samples recorded in the activity DB is usually remarkably imbalanced. For this reason, if learned normally, there is a possibility that the discriminator 13 will all discriminate "no activity" regardless of the presence or absence of true activity. Therefore, the discriminator learning unit 14 trains the discriminator 13 using a weighted error function that increases the penalty when an active protein/ligand bond is discriminated as non-active. As a result, the discriminator learning unit 14 can learn the discriminator 13 with as little influence as possible from the imbalance between the presence and absence of activity. The classifier learning unit 14 functions as learning means.

Ｗ_pos＝（活性無しサンプルの個数）／（活性有りサンプルの個数）
ｎ：サンプルの個数
ｘ⁽ⁱ⁾：ｉ番目のサンプルで作成した画像データを入力した時の判別器１３の出力
ｙ⁽ⁱ⁾：ｉ番目のサンプルの真の活性の有無 Specifically, the classifier learning unit 14 backpropagates an error E calculated by a weighted cross-entropy function using a weighting factor W _pos as represented by the following equation (1): , to learn the classifier 13 .

W _pos = (number of non-active samples)/(number of active samples)
n: Number of samples x ⁽ⁱ⁾ : Output of discriminator 13 when inputting image data created with i-th sample y ⁽ⁱ⁾ : Presence or absence of true activity in i-th sample

Comprehensive determination unit 15 generates an output obtained by inputting each of the plurality of images generated by image generation unit 11 to discriminator 13 for a protein/ligand binding whose activity is unknown, in image generation unit 11 The presence or absence of the protein/ligand binding activity is determined based on the total values for all the images obtained. The comprehensive determination unit 15 functions as comprehensive determination means.

The functional configuration of the determination device 100 has been described above. Next, the learning process performed by the determination device 100 will be described with reference to FIG. The learning process is started when the user instructs the determination device 100 to start the learning process via the operation input unit 33 .

First, from the proteins and ligands registered in the activity DB, the control unit 10 extracts an arbitrary number of data (pairs of proteins and ligands) to be used as learning data, and performs docking simulations for each to extract. A docking structure of protein/ligand binding is acquired for the number of protein-ligand pairs (step S101).

Next, the control unit 10 extracts data desired to be used as learning data from among the docking structures of protein-ligand binding obtained in step S101 (step S102). Here, all the docking structures acquired in step S101 may be used as learning data, or part of them may be used as learning data and the rest may be used as evaluation data.

Next, the image generation unit 11 creates a 3D image from the docking structure extracted as learning data, and captures the entire circumference comprehensively from various angles (projection onto a two-dimensional plane) to create a set of learning images. (image set for learning) is generated (step S103). Step S103 is called an image generation step. Details of the processing (image generation processing) in step S103 will be described later.

Next, the activity acquisition unit 12 refers to the activity DB for the protein and ligand corresponding to each image generated in step S103 (shown in the image) and acquires the presence or absence of protein/ligand binding activity (step S104). Step S104 is called a property acquisition step.

Next, the discriminator learning unit 14 generates teacher data consisting of the image and presence/absence of activity (step S105). At this time, each image generated in step S103 may be randomly rotated by 0°, 90°, 180°, and 270° to generate teacher data. Then, the discriminator learning unit 14 causes the discriminator 13 to learn using the teacher data generated in step S105 (step S106), and ends the learning process. Step S106 is called a learning step.

Next, the image generation processing performed in step S103 will be described with reference to FIG. The image generation process takes arguments N _θ and N _φ . These means dividing the shooting angle (viewpoint angle when projecting onto a two-dimensional plane) into N _θ in the direction of θ and N _φ in the direction of φ shown in FIG. , a total of N _θ ×N _φ images are generated.

First, the image generator 11 initializes a variable i representing an index in the θ direction to 0 (step S201). Then, the angle θ is set to (360°×i)/N _θ (step S202).

Next, the image generator 11 initializes a variable j representing an index in the φ direction to 0 (step S203). Then, the angle φ is set to (360°×j)/N _φ (step S204).

Then, the image generator 11 generates an image by projecting the 3D image of the docking structure onto a two-dimensional plane from the viewpoint from the direction (θ, φ) as shown in FIG. 3 (step S205). The image generator 11 then increments the variable j (step S206) and determines whether the variable j is less than _Nφ (step S207).

If the variable j is less than _Nφ (step S207; Yes), return to step S204. If the variable j is greater than or equal to _Nφ (step S207; No), the image generator 11 increments the variable i (step S208) and determines whether the variable i is less than _Nθ (step S209). .

If the variable i is less than _Nθ (step S209; Yes), the process returns to step S202. If the variable i is greater than or equal to _Nθ (step S209; No), the image generation process is terminated.

Through the learning process (FIG. 5) and the image generation process (FIG. 6) described above, the discriminator 13 is learned. It will output whether it is active or not. As described above, the determination device 100 comprehensively generates images from various viewpoints over the entire circumference (360°) of the docking structure of the protein/ligand binding, and uses each of these comprehensive images to determine the classifier 13 is learned. Therefore, the CNN of the discriminator 13 after learning becomes a learning model from which the three-dimensional features of the docking structure are extracted.

Next, determination processing for determining protein-ligand binding whose activity is unknown using the discriminator 13 thus obtained will be described with reference to FIG. The determination process is started when the user instructs the determination apparatus 100 to start the determination process via the operation input unit 33 . When instructing the start of the determination process, the user inputs the type of protein and the type of ligand to be determined to determination device 100 .

First, the control unit 10 performs a docking simulation for a protein and a ligand input by a user, and acquires a docked structure of protein/ligand binding (step S301).

Next, the image generating unit 11 creates a 3D image from the docking structure acquired in step S301, and captures N _θ ×N _φ images comprehensively from various angles (projecting onto a two-dimensional plane). A set of determination images (image set for determination) is generated (step S302). This process is the same as the above-described image generation process (FIG. 6), and step S302 is also called an image generation step.

Next, the comprehensive determination unit 15 inputs each of the determination images generated in step S302 to the discriminator 13, and acquires the feature map 115 immediately before the final average pooling layer of the CNN for the number of determination images. (step S303). Step S303 is called a determination step.

Then, as shown in FIG. 8, the comprehensive determination unit 15 generates a comprehensive feature map 117 using all the feature maps 115 obtained in step 303, and performs average pooling on the comprehensive feature map 117 (step S304). . This process performs normal (two-dimensional) average pooling in the direction of N _θ ×N _φ images captured (projected onto a two-dimensional plane) covering the entire periphery (one dimension is added). 3D average pooling processing).

Then, based on the output from the output layer 119 after the 3D average pooling process, the comprehensive determination unit 15 determines the presence or absence of protein-ligand binding activity (step S305), and ends the determination process. Step S305 is called a comprehensive determination step.

A supplementary description of the 3D average pooling process will be given with reference to FIG. First, in step S302, N _{θ ×} N _φ judgment images are obtained _. Enter 111. Then, n feature maps 115 immediately before the final average pooling layer are obtained inside the CNN of the discriminator 13 . However, this does not have to be done in parallel, and one discriminator 13 (CNN) may be sequentially used n times to obtain n feature maps 115 .

Each feature map 115 has a plurality of channels (2048 in FIG. 8). A feature map 117 is calculated. A feature map 116 is obtained by performing an average pooling process on the integrated feature map 117 thus obtained. This is the 3D average pooling process. Then, from the feature map 116 , the output of the output layer 119 is obtained via the fully connected connection 125 .

By performing such processing, it is possible to obtain determination results with significantly higher determination accuracy than the output obtained by inputting a single image to the discriminator 13 (discrimination results by the discriminator 13). The feature map 116 obtained by the 3D average pooling process is based on the comprehensive feature map 117 and is a one-dimensional vector that directly affects the output (discrimination result) from the output layer 119, so it is called a comprehensive discriminant vector. to

The results of actual experiments are shown below. In this experiment, Glide was used for the docking simulation, PyMol was used for image generation of the docking structure, and DUD-E was used for the active DB. Then, in the learning data extraction (step S102) of the learning process (FIG. 5), 70% of the docking structures acquired in step S101 are extracted as learning data, and the remaining 30% are used as evaluation data. . The image input size was 224, and ResNet-50 was used as the CNN of the discriminator 13 . The number of batches during learning was set to 128, and learning was performed by randomly rotating the image to the right by 0°, 90°, 180°, and 270° for each batch.

Table 1 compares the determination results of the determination apparatus 100 and Glide when 49 images are generated with N _θ =7 and N _φ =7 in the image generation process. Table 2 compares the determination results of the determination apparatus 100 and Glide when 81 images are generated with N _θ =9 and N _φ =9.

The evaluation index “AUC” is the value of AUC (Area Under the Receiver Operator Curve), which is the area under the ROC (Receiver Operating Characteristic) curve. Also, "EF1%" is an index EF (Enrichment Factor) represented by the following formula (2). This indicates how much active ligands can be concentrated in the top 1%, and is an index that is emphasized in actual drug discovery.
EF=na/(NA×0.01) (2)
na: number of active protein-ligand bonds ranked in the top 1% NA: number of active protein-ligand bonds out of all experimental subjects

From Tables 1 and 2, it can be confirmed that the determination accuracy of the determination device 100 is significantly superior to that of the conventional technology (Glide). This is because it was possible to learn with a large amount of images that cover the entire circumference of the three-dimensional structure of protein / ligand binding, and it is possible to learn evenly using a weighted cross-entropy function for biased learning data. By using 3D Average Pooling using multiple images captured comprehensively (projected on a two-dimensional plane) at the time of judgment, each image learning result can be integrated and grasped as a three-dimensional object. This is thought to be due to

(Modification 1)
In the above-described Embodiment 1, when a 3D image created from the docking structure of protein-ligand binding is comprehensively photographed (projected onto a two-dimensional plane), the surface of the protein at the time of 3D image generation is eliminated, and which The ligand can be confirmed even if it is photographed from an angular viewpoint (projected onto a two-dimensional plane). However, in reality, proteins have surfaces, and ligands can only be identified from the parts without surfaces. Therefore, a modified example 1 will be described in which a plurality of images are generated from the side without a surface, assuming that the surface of the protein is present when the 3D image is generated.

In the determination device 100 of Modification 1, the image generation unit 11 views the ligand 212 in the surface-free portion of the protein 211 from the front, and photographs the docking structure 213 (projected onto a two-dimensional plane), as shown in FIG. Then, an image is generated by photographing (projecting onto a two-dimensional plane) from points such as x and o in FIG. The learning process and the determination process in the determination device 100 of Modification 1 are basically the same as the learning process (FIG. 5) and the determination process (FIG. 7) in the determination device 100 of the first embodiment. Since the image generation process called is different, this process will be described with reference to FIG.

The image generation process of Modification 1 takes N as an argument. N means that the angle for photographing (projection onto a two-dimensional plane) is divided by N in the direction of φ shown in FIG. 9 . In the image generation process of Modification 1, an image from the front, N images around the front at an angle of θ, and N images at an angle of 2θ are generated. An image is generated.

First, as shown in FIG. 9, the image generator 11 generates an image by projecting a 3D image of the docking structure onto a two-dimensional plane from the front where the ligand 212 is not hidden by the surface of the protein 211 (step S221). The image generator 11 then initializes the variable i representing the index in the φ direction to 0 (step S222). Then, the angle φ is set to (360°×i)/N (step S223).

Next, as indicated by x in FIG. 9, the image generating unit 11 generates an image by projecting the 3D image of the docking structure onto a two-dimensional plane from the viewpoint from the direction of φ on the circumference that is θ shifted from the front. (step S224). Then, as indicated by o in FIG. 9, the image generator 11 generates an image by projecting the 3D image of the docking structure onto a two-dimensional plane from the viewpoint from the direction of φ on the circumference that is shifted by 2θ from the front. (Step S225).

Next, the image generator 11 increments the variable i (step S226) and determines whether the variable i is less than N (step S227). Then, if the variable i is less than N (step S227; Yes), the process returns to step S223. If the variable i is greater than or equal to N (step S227; No), the image generation process is terminated.

As described above, in Modification 1, an image of the docking structure 213 including the ligand 212 is captured (projected onto a two-dimensional plane) from a portion of the protein 211 without a surface. It is possible to generate an image from which the three-dimensional structure of the bond can be grasped.

(Modification 2)
In Embodiment 1 and Modification 1 described above, the determination device 100 performs both the learning process and the determination process, but the determination device 100 is not limited to this. For example, the determination device 100 may be a discriminator learning device that does not perform determination processing but performs learning processing to train the discriminator 13 . Further, the determination device 100 may be a determination device that does not perform learning processing, but performs determination processing using the discriminator 13 that has been learned by another determination device 100 . The learning process is difficult to implement without a supercomputer, such as the need to create a large amount of image data for learning and perform deep learning. However, if only the determination process is performed, it is possible to perform the determination simply by creating image data for determination using the discriminator 13 that has already been trained. can be done.

(Modification 3)
In the first embodiment and the modification described above, the 3D average pooling process is performed in step S304 in the determination process (FIG. 7), but this process is not essential. For example, in step S303, the comprehensive determination unit 15 inputs each of the determination images generated in step S302 to the discriminator 13, generates the output of the discriminator 13 for the number of determination images, and skips step S304. You may In this case, in step S305, the comprehensive determination unit 15 determines the presence or absence of protein/ligand binding activity based on the average output of the discriminator 13 acquired in step S303 (existing for the number of images for determination). Judge.

(Other modifications)
In the first embodiment and the modification described above, the angle of the viewpoint is changed at regular angular intervals when generating the image of the docking structure, but the present invention is not limited to this. For example, in Embodiment 1, the number of divisions in the θ direction is increased near φ of 0° or 180° (the portion corresponding to the equator on the earth), and ), the number of divisions in the θ direction may be reduced.

Also, in the above-described embodiment and modification, the image input to the CNN of the discriminator 13 has an input size of 224×224 pixels, and each pixel is a color image with RGB 3 channels, but this is just an example. The input size may be set to a larger value (eg, 448×448 pixels) or a smaller value (eg, 112×112 pixels). Also, the number of pixels vertically and horizontally need not be the same (for example, 1920×1080 pixels). Also, the image need not be a color image and may be a black and white image. In the case of a black-and-white image, each pixel is one-channel information, so the convolution filter 121 shown in FIG. 4 is a one-channel filter (for example, 7×7×1ch).

Moreover, the first embodiment and the modification described above can be combined as appropriate. For example, when Embodiment 1 and Modification 1 are combined, the image generation unit 11 generates an image of the docking structure 213 of the protein 211 and the ligand 212 from a comprehensive viewpoint without a surface (360°) and a surface An image generated from the viewpoint from the side with and without the surface is generated, respectively. Then, the classifier learning unit 14 learns the classifier 13 using both the image without the surface and the image with the surface, and the comprehensive judgment unit 15 judges using both the image without the surface and the image with the surface . By doing so, it becomes possible to make a determination using both the features in the presence of a surface and the features in the absence of a surface.

Also, in Modification 1, the viewpoint is from an angle that has a two-fold relationship, such as θ and 2θ, but these two angles may be completely unrelated angles. Also, the number of angles is not limited to two, and viewpoints from three or more angles around the front may be used. For example, for θ=20°, 30°, 55°, and 70°, the 3D image of the docking structure is projected onto a two-dimensional plane from the viewpoint from the direction of φ on the circumference that is shifted by θ from the front. An image may be generated. Also, the division number N in the φ direction may be set to a different value for each θ.

Further, in the above embodiments and modifications, the determination device 100 for determining the presence or absence of protein/ligand binding activity has been described as an example. It is not limited to presence or absence. The determination device 100 can also perform other determinations according to the learning data by using other learning data. For example, by using data on the presence or absence of protein-protein binding activity and data on the docking structure of the protein-protein binding (generated by a molecular graphics tool such as PyMOL) as learning data, the determination device 100 can: The presence or absence of protein-protein binding activity can be determined. In addition, if data on the presence or absence of binding activity between an arbitrary biomolecule and a substance that binds to the biomolecule, and data on the docking structure of the binding between the biomolecule and the substance, can be prepared as learning data. The device 100 can determine the presence or absence of such any biomolecule-substance binding activity.

Further, the object to be judged by the judging device 100 is not limited to the presence or absence of binding activity. For example, data on some properties (predetermined properties) relating to binding between an arbitrary first substance and a second substance that binds to the first substance, and binding between the first substance and the second substance docking structure data (generated by a molecular graphics tool such as PyMOL) and the like can be prepared as learning data, the determination device 100 can determine the property related to the binding of such arbitrary first substance and second substance A determination of presence or absence can be made.

In the first embodiment and the modification described above, the control unit 10 executes a program that implements the discriminator 13 by CNN, so that the control unit 10 also functions as the discriminator 13. However, this is not the only option. can't The determination device 100 may include a device (for example, a GPU (Graphics Processing Unit), a dedicated IC (Integrated Circuit), etc.) that implements the function of the discriminator 13 , separately from the control unit 10 .

Further, the discriminator 13 may be a discriminator using a neural network other than CNN (for example, RNN (Recurrent Neural Network), etc.). If 3D average pooling processing is not performed in the determination processing (FIG. 7), the discriminator 13 may be a discriminator other than a neural network, such as an SVM (Support Vector Machine).

Note that the determination processing of the determination device 100 can also be performed by a computer such as a normal PC. Also, in the future, it is conceivable that the learning process can be performed not by a supercomputer but also by a computer such as a normal PC. Specifically, in the above embodiment, the programs for the learning process and the determination process performed by the determination device 100 are pre-stored in the ROM of the storage unit 20 . However, the program may be stored in a computer-readable storage medium such as a flexible disk, CD-ROM (Compact Disc Read Only Memory), DVD (Digital Versatile Disc), MO (Magneto-Optical Disc), memory card, USB (Universal Serial Bus) memory, etc. By storing and distributing the program in a recording medium, and reading and installing the program in the computer, a computer capable of realizing each of the functions described above may be configured.

Although the preferred embodiments of the present invention have been described above, the present invention is not limited to such specific embodiments, and the present invention includes the invention described in the claims and their equivalents. be

DESCRIPTION OF SYMBOLS 10... Control part, 11... Image generation part, 12... Activity acquisition part, 13... Discriminator, 14... Discriminator learning part, 15... Comprehensive determination part, 20... Storage part, 31... Output part, 32... Communication part, 33 Operation input unit 100 Determination device 111 Input layer 112, 113, 114, 115, 116 Feature map 117 Comprehensive feature map 119 Output layer 121, 123 Convolution filter 122, 124 ... pooling window, 125 ... all binding connections, 211 ... protein, 212 ... ligand, 213 docking structure, 311 ... camera

Claims (14)

image generating means for generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure; ,
a discriminator that, when inputting one image included in the image set, discriminates whether or not the bond associated with the three-dimensional structure projected on the one image has a predetermined property;
A value obtained by inputting each of a plurality of images included in the image set to the discriminator, based on a total value for all images included in the image set, to an image included in the image set Comprehensive judgment means for judging whether or not the bond associated with the projected three-dimensional structure has the property;
A three-dimensional structure determination device. The image generating means generates an image set including a plurality of images obtained by projecting the three-dimensional structure of the binding of the protein and the ligand onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure,
The discriminator, when inputting one image included in the image set, discriminates the presence or absence of activity of the binding related to the three-dimensional structure projected on the one image,
The overall determination means determines the image based on the values obtained by inputting each of the plurality of images included in the image set into the discriminator, based on the values obtained by totaling all the images included in the image set. Determining the presence or absence of activity of the bond associated with the three-dimensional structure projected on the image included in the set;
The three-dimensional structure determination device according to claim 1. The image generating means generates an image set including a plurality of images obtained by projecting the three-dimensional structure of the binding of the protein and the ligand onto a two-dimensional plane from a plurality of viewpoints covering the entire periphery.
The three-dimensional structure determination device according to claim 2. The image generation means generates a three-dimensional structure of the binding between the protein and the ligand with the surface of the protein present, and projects a plurality of images obtained by projecting onto a two-dimensional plane from a plurality of viewpoints from the side without the surface. generate an image set containing the images,
The three-dimensional structure determination device according to claim 2 or 3. The comprehensive determination means uses 3D Average Pooling to determine the presence or absence of the property of the bond related to the three-dimensional structure projected on the images included in the image set.
The three-dimensional structure determination device according to any one of claims 1 to 4. an image generating step of generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure; ,
a determination step of determining whether or not the bond associated with the three-dimensional structure projected onto one image included in the image set has a predetermined property;
In the determination step, the results of determining whether or not the property of the bond related to the three-dimensional structure projected on each of the plurality of images included in the image set exists are totaled for all the images included in the image set. a comprehensive determination step of determining the presence or absence of the property of the bond associated with the three-dimensional structure projected onto the images included in the image set based on the values;
A three-dimensional structure determination method comprising: image generating means for generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure; ,
a property acquiring means for acquiring the presence or absence of a predetermined property of said bond;
a discriminator for discriminating whether or not each of the combinations has the properties, using images included in the image set generated by the image generation means and the properties acquired by the property acquisition means as training data for each of the combinations; a learning means for learning;
A three-dimensional structure discriminator learning device comprising: The image generating means generates an image set including a plurality of images obtained by projecting the three-dimensional structure of the binding of the protein and the ligand onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure,
The property acquisition means acquires the presence or absence of activity of the bond,
For each of the bonds, the learning means determines whether or not the bond is active, using images included in the image set generated by the image generating means and the presence or absence of activity obtained by the property obtaining means as teacher data. train the discriminator,
8. The three-dimensional structure discriminator learning device according to claim 7. The image generating means generates an image set including a plurality of images obtained by projecting the three-dimensional structure of the binding of the protein and the ligand onto a two-dimensional plane from a plurality of viewpoints covering the entire periphery.
9. The three-dimensional structure discriminator learning device according to claim 8. The image generation means generates a three-dimensional structure of the binding between the protein and the ligand with the surface of the protein present, and projects a plurality of images obtained by projecting onto a two-dimensional plane from a plurality of viewpoints from the side without the surface. generate an image set containing the images,
10. The three-dimensional structure discriminator learning device according to claim 8 or 9. The learning means learns the discriminator using a weighted cross-entropy function,
11. The three-dimensional structure discriminator learning device according to any one of claims 7 to 10. an image generating step of generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure; ,
a property obtaining step of obtaining the presence or absence of a predetermined property of the bond;
a discriminator for discriminating whether or not the combination has the property, using images included in the image set generated in the image generation step and the presence or absence of the property acquired in the property acquisition step for each of the combinations as training data; a learning step for learning;
A three-dimensional structure classifier training method comprising: to the computer,
An image generation step of generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure;
a determination step of determining whether or not the bond associated with the three-dimensional structure projected onto one image included in the image set has a predetermined property;
In the determination step, the results of determining whether or not the property of the bond related to the three-dimensional structure projected on each of the plurality of images included in the image set exists are totaled for all the images included in the image set. a comprehensive determination step of determining the presence or absence of the property of the bond associated with the three-dimensional structure projected onto the images included in the image set, based on the values;
program to run the to the computer,
An image generation step of generating an image set including a plurality of images obtained by projecting a three-dimensional structure of a bond of a first substance and a second substance onto a two-dimensional plane from a plurality of viewpoints around the three-dimensional structure;
a property acquisition step of acquiring the presence or absence of a predetermined property of the bond; and
a discriminator for discriminating whether or not the combination has the property, using images included in the image set generated in the image generation step and the presence or absence of the property acquired in the property acquisition step for each of the combinations as training data; learning steps to learn,
program to run the

Images