KR20220014798A - Apparatus and method for synthesizing target products using neural networks - Google Patents

Abstract

An apparatus for synthesizing a target product using a neural network predicts candidate reactant combinations for generating a target product by using a reverse synthesis prediction model, predicts a predicted product for each of the candidate reactant combinations by using a response prediction model, and determines experimental priorities of the candidate reactant combinations based on a result of the comparison between the target product and the predicted product.

Description

Apparatus and method for synthesizing target products using neural networks

It relates to an apparatus and method for synthesizing a target product using a neural network.

Neural network refers to a computational architecture that models the biological brain. As neural network technology develops, various types of electronic systems use neural networks to analyze input data and extract valid information.

A technique capable of deriving an effective and valid reactant while using a neural network and providing various candidate reactants is required.

An object of the present invention is to provide an apparatus and method for synthesizing a target product using a neural network. The technical problem to be achieved by the present embodiment is not limited to the above technical problems, and further technical problems may be inferred from the following embodiments.

According to one aspect, a method of synthesizing a target product using a neural network includes predicting candidate reactant combinations for generating a target product using a pre-trained reverse synthesis prediction model, using a pre-trained response prediction model predicting a predicted product for each of the candidate reactant combinations, and determining an experimental priority of the candidate reactant combinations based on a comparison result between the target product and the predicted product.

According to another aspect, an apparatus for generating a target product using a neural network includes a memory storing at least one program and a processor executing the at least one program, wherein the processor generates a pre-trained inverse synthesis prediction model. predicting candidate reactant combinations for generating a target product, predicting a predicted product for each of the candidate reactant combinations using a pre-trained reaction prediction model, and predicting a prediction product for each of the candidate reactant combinations using a comparison result between the target product and the predicted product based on the experimental priority of the candidate reactant combinations.

1 is a block diagram illustrating a hardware configuration of a neural network device according to an embodiment.
FIG. 2 is a diagram for explaining a process of determining an experiment priority of candidate reactant combinations based on a comparison result between a target product and a predicted product, according to an exemplary embodiment;
3 is a diagram for explaining a categorical latent variable.
4 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.
5 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.
6 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.
7 is a flowchart illustrating a method of operating a neural network device according to an embodiment.
8 is a flowchart illustrating a method of learning a reverse synthesis prediction model and a reaction prediction model according to an embodiment.
9 is a flowchart illustrating a method of operating an inverse synthesis prediction model according to an embodiment.
10 is a flowchart illustrating a method of operating a response prediction model according to an exemplary embodiment.
11 is a flowchart illustrating a method of determining an experiment priority of candidate reactant combinations according to an embodiment.

The appearances of the phrases "in some embodiments" or "in one embodiment" in various places in this specification are not necessarily all referring to the same embodiment.

Some embodiments of the present disclosure may be represented by functional block configurations and various processing steps. Some or all of these functional blocks may be implemented in various numbers of hardware and/or software configurations that perform specific functions. For example, the functional blocks of the present disclosure may be implemented by one or more microprocessors, or may be implemented by circuit configurations for a given function. Also, for example, the functional blocks of the present disclosure may be implemented in various programming or scripting languages. The functional blocks may be implemented as an algorithm running on one or more processors. In addition, the present disclosure may employ prior art for electronic configuration, signal processing, and/or data processing, and the like. Terms such as “mechanism”, “element”, “means” and “configuration” may be used broadly and are not limited to mechanical and physical components.

In addition, the connecting lines or connecting members between the components shown in the drawings only exemplify functional connections and/or physical or circuit connections. In an actual device, a connection between components may be represented by various functional connections, physical connections, or circuit connections that are replaceable or added.

Meanwhile, in relation to the terms used in this specification, a structure that is data used in a neural network system may mean an atom-level structure of a material. The structure may be a structural formula based on a bond between atoms. For example, the structure may be in the form of a simple-form string (one-dimensional). As a string format expressing the structure, there are a Simplified Molecular-Input Line-Entry System (SMILES) code, a Smiles Arbitrary Target Specification (SMARTS) code, or an International Chemical Identifier (InChi) code.

In addition, a descriptor is an index value used to express characteristics of a material, and is a value that can be obtained by performing relatively simple arithmetic processing on a given material. In one embodiment, the presenter is a molecular structure fingerprint (eg, Morgan Fingerprint, Extended Connectivity Fingerprint (ECFP)) indicating whether a specific partial structure is included or not, or a molecular weight or partial structure contained within a molecular structure (eg, a molecular weight). For example, it may include a quantitative structure-property relationship (QSPR) descriptor composed of a value that can be calculated immediately, such as the number of rings).

In addition, a property refers to a property of a material, and may be a real value measured through an experiment or calculated through a simulation. For example, if the material is a display material, it may be a transmission wavelength for light, an emission wavelength, etc., or if the material is a battery material, it may be a voltage. Unlike the presenter, complex simulations are required to calculate the physical properties and can be time consuming.

1 is a block diagram illustrating a hardware configuration of a neural network device according to an embodiment.

The neural network apparatus 100 may be implemented with various types of devices such as a personal computer (PC), a server device, a mobile device, and an embedded device, and as a specific example, voice recognition, image recognition, image classification, etc. using a neural network It may correspond to a smartphone, tablet device, AR (Augmented Reality) device, IoT (Internet of Things) device, autonomous vehicle, robotics, medical device, etc., but is not limited thereto. Furthermore, the neural network apparatus 100 may correspond to a dedicated hardware accelerator mounted on the above device, and the neural network apparatus 100 is a dedicated module for driving a neural network, a neural processing unit (NPU), It may be a hardware accelerator such as a Tensor Processing Unit (TPU), a Neural Engine, or the like, but is not limited thereto.

Referring to FIG. 1 , the neural network device 100 includes a processor 110 , a memory 120 , and a user interface 130 . In the neural network device 100 shown in FIG. 1, only the components related to the present embodiments are shown. Accordingly, it is apparent to those skilled in the art that the neural network device 100 may further include other general-purpose components in addition to the components shown in FIG. 1 .

The processor 110 serves to control overall functions for executing the neural network device 100 . For example, the processor 110 generally controls the neural network device 100 by executing programs stored in the memory 120 in the neural network device 100 . The processor 110 may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), etc. included in the neural network device 100 , but is not limited thereto.

The memory 120 is hardware for storing various data processed in the neural network device 100 . For example, the memory 120 may store data processed by the neural network device 100 and data to be processed. have. Also, the memory 120 may store applications, drivers, and the like to be driven by the neural network device 100 . The memory 120 includes random access memory (RAM), such as dynamic random access memory (DRAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), CD- It may include ROM, Blu-ray or other optical disk storage, a hard disk drive (HDD), a solid state drive (SSD), or flash memory.

The processor 110 may drive at least one of an inverse synthesis prediction model, a synthesis prediction model, a yield prediction model, and a reaction method model.

The reverse synthesis prediction model, the synthesis prediction model, the yield prediction model, and the reaction method model may be implemented as a transformer model. As the neural network device 100 drives the transformer model, parallel processing of data and rapid computation may be possible.

The processor 110 may train the transformer model using a test data set, and in this process, a categorical latent variable may be determined or learned. For example, the processor 110 may learn the conditional probability distribution of the categorical latent variable for the input of each of the test artifacts. In addition, prediction yields and response schemes may be provided for learning conditional probability distributions of latent variables.

The test data set may include test reactants, test products corresponding to each of the test reactant combinations. In addition, the test data set may further include a predicted yield of the test products according to the experimental conditions and reaction modes of each of the test reactant combinations.

Experimental conditions may refer to various conditions set in order to proceed with an experiment for generating a product using a reactant. For example, the experimental conditions may include at least one of a catalyst, a base, a solvent, a reagent, a temperature, and a reaction time.

The predicted yield may refer to an expectation for the yield of a product produced from the reactants when the reactants are tested in a given reaction mode and under given experimental conditions. The predicted yield may be different from the actual yield measured by an actual experiment.

The reaction mode may refer to a chemical reaction method for generating a product using a reactant. For example, when the reactant organoboron (R1-BY2) is used to form the product halide (R2-BY2), the reaction method may be a Suzuki-miyaura method. A plurality of reaction methods may be provided according to structural information of a reactant and structural information of a product.

The processor 110 may predict candidate reactant combinations for generating the target product using the pre-trained inverse synthesis prediction model.

The processor 110 may receive the chemical structure of the target product in the form of a character string (one-dimensional). The structure may be a structural formula based on a bond between atoms. For example, the processor 110 may receive the chemical structure of the target product in the form of a Simplified Molecular-Input Line-Entry System (SMILES) code, a Smiles Arbitrary Target Specification (SMARTS) code, or an International Chemical Identifier (InChi) code format. . According to an embodiment, information on a descriptor of a target product and information on physical properties may be input in a string format together with a chemical structure.

The processor 110 performs an operation using the string of the target product as input data by driving the pre-trained inverse synthesis prediction model, and generates candidate reactant combinations as output data in string format based on the result of the operation. have.

The processor 110 may predict a prediction product for each of the candidate reactant combinations using the pre-trained response prediction model.

The processor 110 may operate the pre-trained reaction prediction model to perform an operation using candidate reactant combinations as input data, and generate prediction products as output data in a string format based on the operation result.

On the other hand, the processor 110 performs an operation using candidate reactant combinations as input data by driving the pre-trained yield prediction model, and a reaction for generating a predicted yield and a target product of the target product based on the result of performing the operation You can create a method as output data. The predicted yield of the target product may be used as data for learning categorical latent variables.

In addition, the processor 110 performs an operation using the target product as input data by driving the pre-trained reaction method prediction model, and generates a reaction method for generating the target product based on the operation result as output data can do. Reaction schemes for generating target products may also be used as data for learning categorical latent variables.

The processor 110 may compare the target product and the predicted product, and may determine an experiment priority of candidate reactant combinations based on the comparison result. Accordingly, optimal candidate reactant combinations for generating the target product may be output according to priority.

The user interface 130 may refer to an input means for receiving a feedback of an experiment result. For example, the user interface includes a key pad, a dome switch, and a touch pad (contact capacitive method, pressure resistive film method, infrared sensing method, surface ultrasonic conduction method, integral tension measurement method) , piezo effect method, etc.), a jog wheel, a jog switch, etc., but is not limited thereto. The processor 110 may update the transformer model by receiving a feedback of the experimental result.

Hereinafter, in the description of the present embodiments, methods for preferentially providing optimized experimental conditions using the neural network apparatus 100 as described above will be described in detail. The methods described below may be performed by the processor 110 , the memory 120 , and the user interface 130 of the neural network device 100 .

2 is a diagram for explaining a process of determining an experimental priority of candidate reactant combinations based on a comparison result of a target product and a predicted product according to an embodiment, and FIG. 3 is a diagram for explaining a categorical latent variable .

Referring to FIG. 2 , the processor 110 may receive the chemical structure of the target product 210 in a string format.

The target product 211 in the string format may be input to the inverse synthesis prediction model 220 . In addition, the categorical latent variable 230 may be input to the inverse synthesis prediction model 220 .

The categorical latent variable 230 may refer to a variable that is not directly observed but affects information about the candidate reactant combinations 240 . In an embodiment, the categorical latent variable 230 may indicate a type of response. Types of reactions include reaction forms such as decomposition, combustion, metathesis and displacement, catalysts, bases, solvents, reagents, experimental conditions such as temperature and reaction time, suzuki Miyaura -miyaura) may mean including all of the reaction methods.

The categorical latent variable 230 may include a plurality of classes. A plurality of classes may be generated corresponding to the type of reaction. A categorical latent variable may vary a cluster of candidate reactants or candidate reactant combinations.

The inverse synthesis predictive model may be a probabilistic model that depends on unobserved categorical latent variables. In an embodiment, the inverse synthesis prediction model is composed of a plurality of normal distributions in which information on candidate reactants or candidate reactant combinations has different parameters (eg, expected values, differential phases, etc.) according to categorical latent variables. It may be a Gaussian mixture model, but is not limited thereto.

The processor 110 may output candidate reactant combinations 240 for generating the target product 210 by using the pre-trained inverse synthesis prediction model 220 . The candidate reactant combinations 240 may be output in a string format. In other words, the processor 110 may output the candidate reactant combinations 240 in the string format based on the target product 211 in the string format and the categorical latent variable 230 . In an embodiment, the processor 110 may output, but is not limited to, candidate reactant combinations 240 that maximize both the likelihood of each class for the input of the target product and the likelihood of the backward synthesis prediction result.

The candidate reactant combinations 240 may refer to a set of reactant combinations predicted by the processor 110 to generate the target product 211 . The candidate reactants included in the candidate reactant combination may be plural, but is not limited thereto. In other words, it is possible that only one candidate reactant is included in the candidate reactant combination.

Meanwhile, since the categorical latent variable 230 is input to the inverse synthesis prediction model 220 , diversity of candidate reactant combinations 240 can be secured as shown in FIG. 3 .

Referring to FIG. 3 , when the processor 110 sets any one of the classes included in the categorical latent variable 230 as an initial token, and performs decoding based on the set initial token, one A plurality of candidate reactant combinations 220a, 220b, 220c may be generated for the target product 210 of The candidate reactant combinations 220a , 220b , and 220c may be clustered according to the conditional probability for each class of the candidate reactant combinations 220a , 220b , 220c according to the input of the target product 210 . FIG. 3 illustrates a clustering process of candidate reactant combinations 220a , 220b , and 220c according to the input of the target product 210 when there are three classes.

The processor 110 may calculate a likelihood for each class of the candidate reactant combinations 240 according to the input of the target product 211 . Also, the processor 110 may select a preset number of final candidate reactant combinations from among the candidate reactant combinations 240 based on the likelihood of each class of the candidate reactant combinations 240 . According to an embodiment, it is also possible for the processor 110 to calculate the log likelihood for each class of the candidate reactant combinations 240 and to select a preset number of final candidate reactant combinations based on the log likelihood for each class do. The final candidate reactant combinations may be input to the reaction prediction model 250 and used for prediction of the prediction product 260 .

Referring back to FIG. 2 , candidate reactant combinations 240 may be input to the response prediction model 250 . In addition, the categorical latent variable 230 may also be input to the response prediction model 250 .

The processor 110 may output a prediction product 260 for each of the candidate reactant combinations 240 using the pre-trained response prediction model 250 . The prediction product 260 may be output in a string format. In other words, the processor 110 may output the prediction product 261 in the string form based on the candidate reactant combinations 240 in the string form and the categorical latent variable 230 . In one embodiment, the processor 110 may output a prediction product 260 that maximizes both the likelihood of each class for the input of the candidate reactant combinations 240 and the likelihood of the response prediction result, but is not limited thereto. does not

The predicted product 260 may mean a product predicted from reactants included in the candidate reactant combinations 240 .

On the other hand, since the categorical latent variable 230 is also input to the response prediction model 250 , the diversity of the prediction product 260 may also be secured.

The processor 110 may compare the predicted product 260 and the target product 210 according to the input of each of the candidate reactant combinations 240 . In addition, the processor 110 may determine an experiment priority of the candidate reactant combinations 240 based on the comparison result between the predicted product 260 and the target product 210 .

The processor 110 may output the reordered candidate reactant combinations 270 with an adjusted trial priority. The processor 110 may output the rearranged candidate reactant combinations 280 in the form of a structural formula.

Since the neural network apparatus 100 of the present disclosure verifies the candidate reactant combinations 240 derived through the inverse synthesis prediction model 220 through the reaction prediction model 250, the candidate reactant combinations 240 that are inconsistent with the structural grammar ) is output and the case where the target product 210 cannot be synthesized in a string format can be prevented.

4 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.

Referring to FIG. 4 , the processor 110 may receive information on the target product 210 . Information on the target product 210 may be expressed in a string format. The target product 210 in a string format may be input to the inverse synthesis prediction model 220 .

The processor 110 may predict the categorical latent variable 230 based on information about the target product 210 . The categorical latent variable 230 refers to a variable that is not directly observed but affects information about the candidate reactant combinations 240 , and may include a plurality of classes.

The processor 110 drives the retrosynthesis prediction model 220 based on the information about the target product 210 and the categorical latent variable 230 , thereby generating candidate reactant combinations 240 for generating the target product 210 . ) can be predicted.

The desynthesis prediction model 220 includes a word embedding unit 411, a position encoding unit 412a, 412b, 412c, 412d, hereinafter referred to as 412 when there is no need to distinguish), a desynthesis encoder unit 413, and inverse synthesis. It may include a decoder unit 414 and a word generator unit 415 . In FIG. 4 , the word embedding unit 411 , the position encoding unit 412 , the desynthesis encoder unit 413 , the desynthesis decoder unit 414 , and the word generator unit 415 are included in the desynthesis prediction model 220 . Although shown as one unit (unti), the word embedding unit 411, the position encoding unit 412, the desynthesis encoder unit 413, the desynthesis decoder unit 414, and the word generator unit 415 are reversed. It may refer to a layer included in the synthetic prediction model 220 .

The word embedding unit 411 may embed input data in character units. The word embedding unit 411 may map the target product 210 in a string format to a vector having a predetermined dimension. In addition, the word embedding unit 411 may map the categorical latent variable 230 to a vector of a preset dimension.

The positional encoding unit 412 may perform positional encoding to identify the position of each character included in the input data. In an embodiment, the position encoding unit 412 may encode input data using sinusoids of different frequencies.

The first position encoding unit 412a may position-encode the target product 210 in a character string format. The first position encoder 412a may combine the position information with the embedded input data and provide it to the desynthesis encoder 413 .

The second position encoding unit 412b may position-encode the categorical latent variable 230 into an initial token. The second position encoder 412b may combine the position information with the embedded input data and provide it to the desynthesis encoder 413 .

The inverse synthesis encoder unit 413 may include a self-attention sub-layer and a feed-forward sub-layer. For convenience of explanation, one inverse synthesis encoder unit 413 is shown in FIG. 4 , but the inverse synthesis encoder unit 413 may be configured in a stacked form of N encoders.

The desynthesis encoder unit 413 may specify information to be paid attention to among the input sequence of the target product 210 through the self-attention sub-layer. The specified information may be delivered to a feed-forward sub-layer. The feed-forward sub-layer includes a feed-forward neural network, and a transform sequence of an input sequence may be output by the feed-forward neural network. The transform sequence may be provided to the inverse synthesis decoder unit 414 as an encoder output of the inverse synthesis encoder unit 413 .

Like the desynthesis encoder unit 413, one desynthesis decoder unit 414 is shown in FIG. 4 for convenience of explanation, but the desynthesis decoder unit 414 may be configured in a stacked form of N decoders.

The desynthesis decoder unit 414 may include a self-attention sub-layer, an encoder-decoder attention sub-layer, and a feed-forward sub-layer. The encoder-decoder attention sublayer is different from the self-attention sublayer in that a query is a vector of the decoder, and a key and a value are vectors of the encoder.

Meanwhile, the residual connection sub-layer and the normalization sub-layer are individually applied to all lower layers, and the self-attention sub-layer is masked so that the current output position is not used as information about the next output position. ) can be applied. Also, the decoder output may be linearly transformed or a softmax function may be applied.

The desynthesis decoder unit 414 may output an output sequence corresponding to the input sequence of the categorical latent variable 230 and the input sequence of the target product 210 using a beam search procedure. The output sequence may be converted into a string format by the word generator unit 415 and output. Accordingly, the inverse synthesis prediction model 220 may output candidate reactant combinations 240 in a string format corresponding to the input of the target product 210 . In addition, the categorical latent variable 230 may be shared with the response prediction model 250 and used in the calculation of a prediction product.

The processor 110 drives the response prediction model 250 based on the categorical latent variable 230 and the information about the candidate reactant combinations 240 , thereby generating a prediction product 260 for each of the candidate reactant combinations 240 . ) can be predicted.

The reaction prediction model 250 may include a word embedding unit 411 , a position encoding unit 412 , a reaction prediction encoder unit 413 , a reaction prediction decoder unit 416 , and a word generator unit 415 . As will be described later, since the inverse synthesis prediction model 220 and the reaction prediction model 250 share an encoder unit, the inverse synthesis encoder unit 413 and the reaction prediction encoder unit 413 are called a shared encoder unit 413 . can do. In addition, in FIG. 4 , the word embedding unit 411 , the position encoding unit 412 , the reaction prediction encoder unit 413 , the reaction prediction decoder unit 416 , and the word generator unit 415 are included in the reaction prediction model 250 . Although shown as a single unit (unti), the word embedding unit 411, the position encoding unit 412, the reactive prediction encoder unit 413, the reactive prediction decoder unit 416 and the word generator unit 415 are It may also mean a layer included in the response prediction model 250 .

The calculation method of the reaction prediction model 250 and the calculation method of the inverse synthesis prediction model 220 may be similar to each other except for the type of the input sequence and the type of the output sequence. In other words, the word embedding unit 411 may receive the candidate reactant combinations 240 and the categorical latent variable 230 as input data, and embed the input data in character units. In addition, the third position encoding unit 412c combines the position information with the embedded input data and provides it to the response prediction encoder 413, and the fourth position encoding unit 412d provides the categorical latent variable 230 as an initial token. It may be provided to the response prediction encoder unit 413 by position-encoding as .

The response prediction encoder unit 413 specifies information to be noted from among the input sequences of candidate reactant combinations 240 through the self-attention sub-layer and delivers it to the feed-forward sub-layer, and the feed-forward sub-layer is a feed-forward neural network can be used to output a transform sequence.

The response prediction decoder unit 416 may include a self-attention sub-layer, an encoder-decoder attention sub-layer, and a feed-forward sub-layer. In addition, the residual connection sublayer and the normalization sublayer are individually applied to all lower layers, and the self-attention sublayer is masked so that the current output position is not used as information about the next output position. ) can be applied. Also, the decoder output may be linearly transformed or a softmax function may be applied.

The response prediction decoder unit 416 may output an output sequence corresponding to the input sequence of the categorical latent variable 230 and the input sequence of the candidate reactant combinations 240 using a beam search procedure. The output sequence may be converted into a string format by the word generator unit 415 and output. Accordingly, the reaction prediction model 250 may output the prediction product 261 in the form of a string corresponding to the input of the candidate reactant combinations 240 .

The processor 110 may compare the target product 211 in the character string format with the prediction product 261 in the character string format, and may determine an experiment priority of the candidate reactant combinations 240 based on the comparison result. The processor 110 determines the experiment priority of the candidate reactant combinations 240 based on whether the target product 211 in the string format matches the predicted product 261 in the string format, and based on the determined experiment priority Reordered candidate reactant combinations 270 may be output.

Meanwhile, in FIG. 4 , since the source language and the target language are the same in a string format (eg, SMILES), the inverse synthesis prediction model 220 and the reaction prediction model 250 are formed by a word embedding unit ( 411), the word generator unit 415 and the encoder unit may be shared. In addition, since the inverse synthesis prediction model 220 and the response prediction model 250 share the same parameters as the categorical latent variable 230 , model complexity is reduced, and regularization of the model is easy.

5 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.

The difference from FIG. 4 is that the response mode is provided as an input value of a categorical latent variable. In FIG. 5 , the same reference numerals are used for the same components as those of FIG. 4 , and duplicate descriptions of the same components are omitted.

Referring to FIG. 5 , the neural network apparatus 100 may further include a response method prediction model 417 . In FIG. 5 , the reaction method prediction model 417 is shown as a part of the retrosynthesis prediction model 310 , but according to an embodiment, the reaction method prediction model 417 is distinguished from the retrosynthesis prediction model 310 . It may be a separate configuration.

The reaction mode prediction model 417 may predict a reaction mode based on information about the encoded target product 210 . In an embodiment, the reaction mode prediction model 417 may obtain at least one of physical property information, presenter information, and structure information of the target product 210 from the encoded representation of the target product 210 . In addition, the reaction method prediction model 417 may predict the reaction methods 510 for generating the target product 210 based on at least one of physical property information, presenter information, and structure information of the target product 210 . have.

The processor 110 may receive an expected response scheme 511 of any one of the predicted response schemes 510 . The expected response method 511 may be input to the processor 110 by a user input through the user interface 130 . The expected response method 511 may be utilized as an input value in the categorical latent variable 230 .

The processor 110 may decode the encoded target product 210 and the expected response scheme 511 to output candidate reactant combinations 240 corresponding to the expected response scheme 511 . For example, if the expected reaction scheme 511 is a Suzuki-miyaura scheme, the processor 110 may generate a target product 210 in a suzuki-miyaura scheme. Candidate reactants Combinations 240 may be output.

As the expected reaction scheme 511 is provided as an input value of the categorical latent variable 230 , optimal candidate reactant combinations 240 may be output even in an environment in which reagent use is limited.

6 is a diagram for explaining operations performed in the inverse synthesis prediction model and the reaction prediction model according to an embodiment.

The difference from FIG. 5 is that the yield according to the experimental conditions is utilized when learning the categorical latent variable 230 . Similarly, in FIG. 6, the same reference numerals are used for the same components as those of FIG. 5, and duplicate descriptions of the same components are omitted.

Referring to FIG. 6 , the neural network apparatus 100 may further include a yield prediction model 610 . In FIG. 6 , the yield prediction model 610 is shown in a configuration distinct from the inverse synthesis prediction model 310 , but according to an embodiment, the yield prediction model 610 may be a part of the inverse synthesis prediction model 310 . have.

The yield prediction model 610 may be implemented as a transformer model. The yield prediction model 610 may be trained by a test data set, and may output a predicted yield of the target product 210 according to experimental conditions.

The yield prediction model 610 may include a word embedding unit 411, a fifth position encoding unit 412e, a yield prediction encoder unit 413, a yield prediction unit 611, and an experimental condition prediction unit 612. . The yield prediction model 610 may share the word embedding unit 411 with the inverse synthesis prediction model 220 and the reaction prediction model 250 . Also, the yield prediction model 610 may share an encoder unit with the inverse synthesis prediction model 220 and the response prediction model 250 . The yield prediction unit 611 and the experimental condition prediction unit 612 may perform the same functions as the decoder unit included in the inverse synthesis prediction model 220 and the reaction prediction model 250 .

The experimental condition predictor 612 may predict the experimental conditions of the rearranged candidate reactant combinations 270 for generating the target product 210 . Also, the yield prediction unit 611 may predict the predicted yield of the target product 210 when the experiment is performed under the experimental conditions predicted by the experimental condition prediction unit 612 .

The predicted yield of the target product 210 predicted by the yield prediction unit 611 may be utilized when learning the categorical latent variable 230 . The processor 110 may output the predicted yield of the target product 2100 through a predetermined display means. Accordingly, the user may select the optimal candidate reactant combinations 240 in consideration of the predicted yield.

7 is a flowchart illustrating a method of operating a neural network device according to an embodiment.

Referring to the drawing, in step S710 , the processor 110 may predict candidate reactant combinations 240 for generating the target product 210 using the pre-trained inverse synthesis prediction model 310 .

When information on the target product 210 is input, the processor 110 may predict the categorical latent variable 230 including a plurality of classes based on the information on the target product 210 . In addition, the processor 110 may autoregressively predict the candidate reactant combinations 240 based on the categorical latent variable 230 .

If the information about the target product 210 is x, the categorical latent variable 230 is z, and the candidate reactant combinations 240 are y, then the categorical latent variable 230 according to the input of the target product 210 is y. ) and candidate reactant combinations 240 may be predicted by the likelihood P(z,y|x). The processor 110 may predict the categorical latent variable 230 and the candidate reactant combinations 240 based on the value of the likelihood P(z,y|x).

In operation S720 , the processor 110 may predict the prediction product 260 for each of the candidate reactant combinations 240 using the pre-trained response prediction model 250 .

The processor 110 may obtain information about the predicted product 260 based on the categorical latent variable 230 and the candidate reactant combinations 240 .

In operation S730 , the processor 110 may determine an experiment priority of the candidate reactant combinations 240 based on the comparison result between the target product 210 and the predicted product 260 .

The processor 110 may determine an experiment priority of the candidate reactant combinations 240 based on whether the target product 210 and the predicted product 260 match.

인 경우, 범주형 잠재 변수(230) 및 후보 반응물 조합들(240)의 입력에 따른 표적 생성물(210)과 예측 생성물(260)의 일치 확률은 우도 P(

=x|z,y)에 의해 결정될 수 있다. 프로세서(110)는 우도 P(

=x|z,y)·P(z,y|x)의 값이 큰 순서대로 후보 반응물 조합들(240)의 실험 우선 순위를 결정할 수 있다.The prediction product 260 is

If , the probability of matching the target product 210 and the predicted product 260 according to the input of the categorical latent variable 230 and the candidate reactant combinations 240 is the likelihood P(

=x|z,y). The processor 110 determines the likelihood P(

=x|z,y)·P(z,y|x) may determine the experimental priority of the candidate reactant combinations 240 in the order of the largest values.

8 is a flowchart illustrating a method of learning a reverse synthesis prediction model and a reaction prediction model according to an embodiment.

Referring to FIG. 8 , in step S810 , the processor 110 may receive test reactant combinations and test products corresponding to each of the test reactant combinations.

According to an embodiment, the processor 110 may additionally receive the predicted yield of the test products according to the experimental condition and the reaction method of each of the test reactant combinations.

In step S820 , the processor 110 may learn the categorical latent variable 230 including a plurality of classes based on test reactant combinations and test products.

According to an embodiment, the processor 110 may learn the categorical latent variable 230 further based on the predicted yield provided by the pre-trained yield prediction model.

The processor 110 may learn the conditional probability distribution of the categorical latent variable 230 for each input of the test products.

In an embodiment, when the information on the test products is x _ts and the information on the test reactants is y _ts , the processor 110 calculates the likelihood of the inverse synthesis prediction model 220 as in Equation 1 below. Categorical latent variable 230 such that the cross-entropy loss of P(z,y _ts |x _ts ) and the likelihood P(x _ts |y _ts ,z) of the response prediction model 250 is minimized can be set.

For example, the processor 110 may learn the categorical latent variable 230 using an Expectation-Maximization algorithm (hard EM).

인 경우 다음의 수학식 2와 같이, 교차 엔트로피 손실(L)을 최소화하는 범주형 잠재 변수(230)를 추정하는 단계 및 다음의 수학식 3과 같이, 입력되는 변수들에 대응하여 모수를 업데이트하는 단계를 반복적으로 수행할 수 있다. 프로세서(110)가 모델을 반복적으로 업데이트함에 따라 최적의 범주형 잠재 변수(230)가 도출될 수 있다.The processor 110 determines that the parameters of the current model are

In the case of estimating the categorical latent variable 230 that minimizes the cross entropy loss (L) as shown in Equation 2 below, and as shown in Equation 3 below, updating the parameter in response to the input variables as shown in Equation 3 below. The steps can be performed repeatedly. As the processor 110 iteratively updates the model, an optimal categorical latent variable 230 may be derived.

는 업데이트된 모수를 의미한다.Meanwhile, in Equations 2 and 3, L _h means that the processor 110 used a hard expected value maximization algorithm, and in Equation 3

is the updated parameter.

9 is a flowchart illustrating a method of operating an inverse synthesis prediction model according to an embodiment.

Referring to FIG. 9 , in step S910 , the processor 110 may receive information on the target product 210 .

The processor 110 may receive the chemical structure of the target product 210 in a string format. For example, the processor 110 inputs the chemical structure of the target product 210 in the form of a Simplified Molecular-Input Line-Entry System (SMILES) code, Smiles Arbitrary Target Specification (SMARTS) code, or International Chemical Identifier (InChi) code format. can receive

In step S920 , the processor 110 may predict the categorical latent variable 230 including a plurality of classes.

The processor 110 may drive the inverse synthesis prediction model 220 to predict the categorical latent variable 230 as a first token. According to an embodiment, the expected response manner of the target product 210 may be provided as an input value of the categorical latent variable 230 . When the expected reaction scheme 511 is provided as an input value of the categorical latent variable 230 , optimal candidate reactant combinations 240 may be output even in an environment in which reagent use is restricted.

In operation S930 , the processor 110 may obtain information on the candidate reactant combinations 240 based on the information on the target product 210 and the categorical latent variable 230 .

The processor 110 may autoregressively predict tokens for the candidate reactant combinations 240 based on the information about the target product 210 and the predicted categorical latent variable 230 . have. The likelihood of the categorical latent variable 230 and the candidate reactant combinations 240 according to the input of the target product 210 may be determined by Equation 4 below.

In Equation 4, T is the length of the token sequence, and y _{< t} means a target token before y _t .

The processor 110 selects candidate reactant combinations that maximize both the likelihood of each class for the input of the target product 210 and the likelihood of the reverse synthesis prediction result for the input of the target product 210 and the categorical latent variable 230 . (240) can be output.

In an embodiment, the processor 110 may select a preset number of final candidate reactant combinations based on the likelihood of each class and the likelihood of the inverse synthesis prediction result. For example, the processor 110 may select a preset number of sets in the order of increasing the likelihood of the categorical latent variable 230 and the candidate reactant combinations 240 according to the input of the target product 210 . The set may include pairs of categorical latent variables 230 and final candidate reactant combinations 240 . The pair of categorical latent variable 230 and final candidate reactant combinations 240 may be input to response prediction model 250 and used in calculation of prediction product 260 .

10 is a flowchart illustrating a method of operating a response prediction model according to an exemplary embodiment.

Referring to FIG. 10 , in step S1010 , the processor 110 may receive information on candidate reactant combinations 240 .

The processor 110 may receive the structure of the candidate reactant combinations 240 in the form of a character string. In this case, the string format of the candidate reactant combinations 240 may be the same as the string format of the target product 210 . For example, if the string format of the target product 210 is a SMILES code, the string format of the candidate reactant combinations 240 may also be a SMILES code.

In step S1020 , the processor 110 may receive a categorical latent variable 230 including a plurality of classes.

In other words, the processor 110 may receive a pair of the categorical latent variable 230 and the final candidate reactant combinations 240 .

In step S1030 , the processor 110 may obtain information on the predicted product for each of the candidate reactant combinations based on the information on the candidate reactant combinations 240 and the categorical latent variable 230 .

The processor 110 may generate a token of the prediction product 260 based on the likelihood P(x|z,y). The prediction product 260 may be used for validation of the inverse synthesis prediction model 220 .

11 is a flowchart illustrating a method of determining an experiment priority of candidate reactant combinations according to an embodiment.

Referring to FIG. 11 , in step S1110 , the processor 110 may determine whether the prediction product 260 according to the input of each of the candidate reactant combinations 240 matches the target product 210 .

=x|z,y)에 의해 결정할 수 있다.The processor 110 determines whether the target product 210 and the predicted product 260 according to the input of the categorical latent variable 230 and the candidate reactant combinations 240 match the likelihood P(

=x|z,y).

In operation S1120 , the processor 110 may determine an experiment priority of the candidate reactant combinations 240 based on whether the predicted product 260 and the target product 210 match.

=x|z,y)·P(z,y|x)의 값이 큰 순서대로 후보 반응물 조합들(240)의 실험 우선 순위를 결정할 수 있다.The processor 110 determines the likelihood P(

=x|z,y)·P(z,y|x) may determine the experimental priority of the candidate reactant combinations 240 in the order of the largest values.

Meanwhile, the above-described embodiments can be written as a program that can be executed on a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium. In addition, the structure of data used in the above-described embodiments may be recorded in a computer-readable recording medium through various means. The computer-readable recording medium includes a storage medium such as a magnetic storage medium (eg, a ROM, a floppy disk, a hard disk, etc.) and an optically readable medium (eg, a CD-ROM, a DVD, etc.).

Those of ordinary skill in the art related to the present embodiment will understand that the embodiment may be implemented in a modified form without departing from the essential characteristics of the above description. Therefore, the disclosed embodiments are to be considered in an illustrative rather than a restrictive sense. The scope of the rights is indicated in the claims rather than the above description, and all differences within the scope equivalent thereto should be construed as being included in the present embodiment.

100: neural network device
110: processor
220: retrosynthesis prediction model
250: response prediction model

Claims (20)

In the method of synthesizing a target product using a neural network,
predicting candidate reactant combinations to produce a target product using the pre-trained retrosynthesis prediction model;
predicting a predicted product for each of the candidate reactant combinations using a pre-trained response prediction model; and
determining an experiment priority of the candidate reactant combinations based on a comparison result of the target product and the predicted product. According to claim 1,
The pre-trained inverse synthesis prediction model and the pre-trained reaction prediction model are
receiving test reactant combinations and test products corresponding to each of the test reactant combinations;
learning categorical latent variables comprising a plurality of classes based on the test reactant combinations and the test products. 3. The method of claim 2,
The step of learning the categorical latent variable is
and learning the conditional probability distribution of the categorical latent variable for each input of the test products. 3. The method of claim 2,
The step of learning the categorical latent variable is
A method for learning the categorical latent variable based on a predicted yield provided by a pre-trained yield prediction model. 3. The method of claim 2,
wherein the learned categorical latent variable is used to predict the candidate reactant combinations and predict the predicted product. According to claim 1,
Predicting the candidate reactant combinations comprises:
receiving information on the target product;
predicting a categorical latent variable including a plurality of classes based on the information on the target product; and
obtaining information on the candidate reactant combinations based on the information on the target product and the categorical latent variable. 7. The method of claim 6,
Obtaining information on the candidate reactant combinations comprises:
calculating a likelihood for each class of the candidate reactant combinations according to the input of the target product;
calculating a likelihood of a reverse synthesis prediction result for the target product and the input of the categorical latent variable; and
Selecting a preset number of final candidate reactant combinations based on the likelihood for each class and the likelihood of the reverse synthesis prediction result. 7. The method of claim 6,
The step of predicting the categorical latent variable is
wherein an expected response mode for producing the target product is provided as an input value of the categorical latent variable. According to claim 1,
Predicting the predicted product comprises:
receiving information on the candidate reactant combinations;
receiving a categorical latent variable including a plurality of classes; and
obtaining information on the predicted product for each of the candidate reactant combinations based on the information on the candidate reactant combinations and the categorical latent variable. According to claim 1,
determining the experimental priority of the candidate reactant combinations;
determining whether the predicted product matches the target product according to the input of each of the candidate reactant combinations; and
determining an experiment priority of the candidate reactant combinations based on whether the predicted product matches the target product. An apparatus for generating a target product using a neural network, comprising:
a memory in which at least one program is stored; and
a processor executing the at least one program;
the processor is
predicting candidate reactant combinations to produce a target product using the pre-trained retrosynthesis prediction model;
Predict the predicted product for each of the candidate reactant combinations using a pre-trained response prediction model,
An apparatus for determining an experimental priority of the candidate reactant combinations based on a comparison result of the target product and the predicted product. 12. The method of claim 11,
The pre-trained inverse synthesis prediction model and the pre-trained reaction prediction model are
An apparatus for learning categorical latent variables including a plurality of classes based on test reactant combinations and test products corresponding to each of the test reactant combinations. 13. The method of claim 12,
The pre-trained inverse synthesis prediction model and the pre-trained reaction prediction model are
An apparatus for learning the conditional probability distribution of the categorical latent variable for each input of the test products. 13. The method of claim 12,
The pre-trained inverse synthesis prediction model and the pre-trained reaction prediction model are
An apparatus for learning the categorical latent variable based on a predicted yield provided by a pre-trained yield prediction model. 13. The method of claim 12,
wherein the learned categorical latent variable is used for prediction of the candidate reactant combinations and prediction of the prediction product. 12. The method of claim 11,
the processor is
receiving information about the target product,
Predict a categorical latent variable comprising a plurality of classes based on the information about the target product,
An apparatus for obtaining information about the candidate reactant combinations based on the information on the target product and the categorical latent variable. 17. The method of claim 16,
the processor is
calculating the likelihood of each class of the candidate reactant combinations according to the input of the target product,
calculating the likelihood of a retrosynthesis prediction result for the input of the target product and the categorical latent variable,
An apparatus for selecting a preset number of final candidate reactant combinations based on the likelihood for each class and the likelihood of the inverse synthesis prediction result. 17. The method of claim 16,
The expected reaction scheme to produce the target product is
A device provided as an input value of the categorical latent variable. 12. The method of claim 11,
the processor is
receiving information about the candidate reactant combinations;
Receive a categorical latent variable containing a plurality of classes,
An apparatus for obtaining information about the predicted product for each of the candidate reactant combinations based on the information about the candidate reactant combinations and the categorical latent variable. 12. The method of claim 11,
the processor is
Determine whether the predicted product matches the target product according to the input of each of the candidate reactant combinations,
An apparatus for determining an experimental priority of the candidate reactant combinations based on whether the predicted product and the target product match.

Images