Molecular structure generation method

Abstract

Provided is a molecular structure generation method in which a large language model is used and which exhibits high versatility and excellent precision. Provided is a molecular structure generation method comprising: a step for inputting a first instruction sentence in which a compound, which serves as a base, and a requirement for characteristics are combined with molecular structure information of a compound corresponding to the compound and a natural sentence; a step for separating a part of the molecular structure information of the compound and a part of the requirement for the characteristics; a step for generating a plurality of similar compounds on the basis of the molecular structure information of the compound which serves as a base; a step for extracting information regarding the characteristics from the requirement for the characteristics; and a step for inputting a second instruction sentence for a large language model. To the second instruction sentence, at least the information regarding the characteristics and, as an example of the molecular, the plurality of similar compounds are inputted.

Description

How to generate molecular structures

[0001] One aspect of the present invention relates to a method and system for generating a chemical molecular structure, and more particularly to a method and system for generating a chemical molecular structure using a language model.

[0002] One embodiment of the present invention is not limited to the above technical field, and examples of the technical field of one embodiment of the present invention include semiconductor devices, display devices, light-emitting devices, power storage devices, memory devices, electronic devices, lighting devices, input devices (e.g., touch sensors), input / output devices (e.g., touch panels), driving methods thereof, and manufacturing methods thereof.

[0003] The chemical space of organic small molecule compounds is approximately 10 ６０ It is said that there are candidate molecules. Finding a compound with desired properties from such a vast search space requires a huge amount of effort. For this reason, in the field of drug discovery, materials screening using computers has been practiced for quite some time, and molecular generation methods using machine learning have also been actively researched (see, for example, Patent Document 1).

[0004] Meanwhile, in the past few years, the development of language models using neural networks has progressed significantly, with large-scale language models (LLMs) attracting particular attention. A large-scale language model is a natural language processing model trained using a large amount of data. A large-scale language model can realize, for example, a dialogue model that responds to user instructions. Non-Patent Document 1 discloses GPT-4 (Generative Pre-trained Transformer 4) as a large-scale language model, and ChatGPT as a dialogue model.

[0005] JP 2022-94334 A

[0006] Summary of ChatGPT / GPT-4 Research and Perspective Towards the Future of Large Language Models, Yiheng Liu et al. (Submitted on 4 Apr 2023, [online], Internet <URL: https: / / arxiv.org / abs / 2304.01852>

[0007] An object of one aspect of the present invention is to provide a molecular structure generation method capable of obtaining appropriate results in a molecular generation system using a large-scale language model.An object of one aspect of the present invention is to provide a molecular structure generation system capable of obtaining appropriate results in a molecular generation system using a large-scale language model.

[0008] The above-listed problems do not preclude the existence of other problems. Note that one embodiment of the present invention does not necessarily solve all of the above-listed problems. Note that problems other than the above-listed problems will become apparent from the description of this specification, drawings, claims, etc., and problems other than the above-listed problems can be extracted from the description of this specification, drawings, claims, etc.

[0009] One aspect of the present invention is a molecular structure generation method comprising the steps of: accepting a first compound and a requirement for a property as a first instruction sentence that combines molecular structure information of the first compound and natural language; generating a plurality of similar compounds based on the molecular structure information of the first compound; extracting information related to the property from the requirement for the property; generating a second instruction sentence; and inputting the second instruction sentence into a large-scale language model to obtain molecular structure information of a second compound that is expected to have the property, wherein the second instruction sentence includes at least the information related to the property and a plurality of similar compounds.

[0010] Alternatively, another aspect of the present invention is a molecular structure generation method comprising the steps of: receiving a first compound and a characteristic requirement as a first instruction sentence combining molecular structure information of the first compound and natural language; generating a plurality of similar compounds based on the molecular structure information of the first compound; extracting information related to the characteristic from the characteristic requirement; generating a second instruction sentence; and acquiring molecular structure information of a second compound expected to have the characteristic by inputting the second instruction sentence to a large-scale language model, wherein input items for the second instruction sentence are determined in advance, and the input items include at least background information, example molecules, and a requirement.

[0011] Alternatively, another aspect of the present invention is a molecular structure generation method having the above-mentioned configuration, which includes a step of separating molecular structure information of the first compound and a request for properties from the first directive.

[0012] Another embodiment of the present invention is a method for generating a molecular structure having the above-described structure, wherein the molecular structure information of the first compound is expressed in SMILES notation.

[0013] Alternatively, another aspect of the present invention is a molecular structure generation method having the above-mentioned configuration, wherein the step of generating a plurality of similar compounds includes the steps of converting molecular structure information of a first compound into a first molecular vector, changing the first molecular vector to generate a second molecular vector, and converting the second molecular vector into molecular structure information of a corresponding compound.

[0014] Alternatively, another aspect of the present invention is a molecular structure generation method in which, in the above configuration, the step of generating a plurality of similar compounds is performed using a variational autoencoder that has been trained using molecular structure information of the compounds.

[0015] Alternatively, another aspect of the present invention is a molecular structure generation method having the above configuration, which includes a step of excluding compounds learned by a trained variational autoencoder from molecules generated by a large-scale language model.

[0016] Alternatively, another aspect of the present invention is a method for generating a molecular structure, in the above configuration, in which calculations regarding desired properties are performed for a plurality of similar compounds, and the results of the calculations are input into a second instruction statement together with an example molecule.

[0017] Alternatively, another aspect of the present invention is a molecular structure generation method in which, in the above configuration, the step of extracting information related to characteristics from requirements for characteristics is performed using a text vector database in which chemical papers are vectorized.

[0018] Other aspects of the present invention will be described in the following embodiments and in the drawings.

[0019] According to one aspect of the present invention, a molecular structure generation method capable of obtaining appropriate results in a molecular structure generation system using a large-scale language model can be provided. Alternatively, according to another aspect of the present invention, a molecular structure generation system capable of obtaining appropriate results in a molecular structure generation system using a large-scale language model can be provided.

[0020] The effects listed above do not preclude the existence of other effects. It is not necessary for one embodiment of the present invention to have all of the effects listed above. Effects other than the effects listed above will become apparent from the description in this specification, drawings, claims, etc., and other effects other than the effects listed above can be extracted from the description in this specification, drawings, claims, etc.

[0021] FIG. 1 is a flowchart showing an example of a molecular structure generation system. FIGS. 2A to 2C are diagrams showing an example of a method for inputting a first instruction statement. FIG. 3 is a diagram showing an example of a second instruction statement. FIG. 4 is a schematic diagram of a molecular structure generation system. FIG. 5 is a block diagram showing an example of a molecular structure generation system. FIG. 6 is a conceptual diagram showing an example of a molecular structure generation system. FIG. 7 is a flowchart showing an example of a molecular structure generation system.

[0022] The embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be readily understood by those skilled in the art that various changes can be made in form and detail without departing from the spirit and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the description of the embodiments shown below.

[0023] In the configuration of the invention described below, the same parts or parts having similar functions are denoted by the same reference numerals in different drawings, and repeated explanations thereof will be omitted. Furthermore, when referring to similar functions, the same hatching pattern may be used and no particular reference numeral may be assigned.

[0024] Furthermore, for ease of understanding, the position, size, range, etc. of each component shown in the drawings may not represent the actual position, size, range, etc. Therefore, the disclosed invention is not necessarily limited to the position, size, range, etc. disclosed in the drawings.

[0025] In this specification, the ordinal numbers "first" and "second" are used for convenience and do not limit the number of components or the order of the components (for example, the order of processes or the order of stacking). Furthermore, the ordinal numbers assigned to components in one part of this specification may not match the ordinal numbers assigned to the same components in other parts of this specification or in the claims.

[0026] (Embodiment 1) Molecular generation methods using machine learning have been actively researched, but it is difficult for machine learning-based molecule generation using a labeled dataset of physical properties to generate molecules with the goal of having properties other than the physical properties assigned as labels, or properties other than those strongly associated with the physical properties assigned as labels.

[0027] However, with the advancement of technology, minor properties or parameters that were previously little known often become important. Such minor and innovative properties or parameters are difficult to label in training datasets, making them difficult to learn. Therefore, it has been difficult to search for candidate compounds with minor properties or parameters using machine learning molecule generation methods that use labeled datasets of physical properties. Furthermore, properties that are difficult to quantify are difficult to label, making learning difficult.

[0028] Here, large-scale language models trained on large amounts of data, including chemical papers, are expected to contain data containing hints to such characteristics, physical properties, or parameters. Therefore, by generating molecules using large-scale language models, it may be possible to generate compounds with minor physical properties and parameters, as well as properties that are difficult to quantify, that cannot be handled by machine learning using labeled datasets of physical properties.

[0029] However, when generating molecules using a large-scale language model, generating molecular structures without using any criteria within the vast chemical space, as mentioned above, may result in the generation of a large number of unnecessary molecules.

[0030] In one aspect of the present invention, in order to generate molecules with high accuracy using a large-scale language model, a reference compound and molecules similar to the reference compound are generated, and these molecules are presented to the large-scale language model, thereby enabling the presentation of useful compounds with high accuracy.

[0031] A molecular structure generation method according to one embodiment of the present invention will now be described. Fig. 1 is a flowchart showing a molecular structure generation method according to one embodiment of the present invention.

[0032] First, the molecular structure generation system of the present invention receives a first instruction statement input by the user (step S11).

[0033] This first directive describes a combination of molecular structure information of the original compound and a request for the desired properties. The request for the properties can be input in natural language. Since the request for the properties can be input in natural language, even properties that are difficult to quantify can be accurately expressed, thereby increasing the possibility of obtaining the compound the user desires.

[0034] Figure 2A shows an example of a character string representing molecular structure information of a compound. Figure 2A shows an example using SMILES notation (Simplified Molecular Input Line Entry Specification Syntax). The grammatical rules of SMILES notation are intuitively easy to understand, and the structure can be expressed one-dimensionally (linearly) with a small byte length.

[0035] Another method for expressing molecular structure information of a compound as a character string is molecular fingerprinting. Molecular fingerprinting is the most basic descriptor of chemical structure. For molecular fingerprinting, multiple descriptors using different algorithms have been proposed, and these fingerprint descriptors are implemented in the Python cheminformatics library RDKit and the R language library rcdk, and can be converted from SMILES notation.

[0036] The SMILES notation, or molecular fingerprint, which expresses the molecular structure information of a compound as a character string, is suitable for easy learning using a language model.

[0037] Although a molecular fingerprint has a plurality of descriptors, it is preferable that the descriptors used for learning and the descriptors used in the instruction sentences are of the same type.

[0038] Furthermore, since the expression of the SMILES notation may differ depending on the orientation of the drawn molecule, it is preferable to use the SMILES notation used for learning that has been generated by applying the same rules regarding the orientation of the drawn molecule, and it is also preferable to apply the same rules regarding the orientation of the drawn molecule when generating the SMILES notation to be used in directives.

[0039] 2A, both the molecular structure information of the compound and the desired properties can be entered in the input field 200. Although the SMILES notation is enclosed in double quotation marks, this is not necessary.

[0040] 2B shows an example in which the molecular structure information of a compound is input using a chemical formula drawing tool, a plug-in for the chemical formula drawing tool, or an image. In this case, a field 201 for inputting the molecular structure information of the compound and a field 202 for inputting desired properties are provided. Examples of chemical formula drawing tools include ChemDraw and JMol.

[0041] 2C shows an example in which the molecular structure information of a compound is input as the compound name. In this case, both the compound name and desired properties can be input in the input field 200.

[0042] Next, the molecular structure information portion of the compound and the desired property portion are separated (step S12).

[0043] In the input shown in Figures 2A and 2C, the molecular structure information of the compound and the desired properties are both entered consecutively as text data, so they need to be separated for subsequent processing.

[0044] The separation of the molecular structure information portion of the compound and the desired portion for the properties can be achieved by extracting and judging the characteristics related to the notation method of the molecular structure information of the compound.

[0045] Alternatively, by providing a delimiter such as a double quotation mark as shown in FIG. 2A, it is possible to distinguish and separate the molecular structure information portion from the desired property portion.

[0046] In FIG. 2B, the molecular structure information portion of the compound and the desired property portion are originally input separately, so the step of separating the molecular structure information portion of the compound and the desired property portion may be omitted.

[0047] When the molecular structure information of the compound is input as text, this step can be omitted by separating the input fields for the molecular structure information of the compound from the input fields for the desired properties.

[0048] After separating the molecular structure information portion of the compound from the desired property portion, it is preferable to convert the molecular structure information of the compound into a representation that is easy to process on a computer. Note that, as shown in Figure 2A, if the molecular structure information of the compound is input in a representation method that is suitable for computer processing, such as SMILES notation or molecular fingerprint (hereinafter, sometimes referred to as "SMILES notation, etc."), this step can be omitted.

[0049] As shown in Figures 2B and 2C, when the molecular structure information of a compound is input in a manner other than a representation method suitable for computer processing, such as SMILES notation or molecular fingerprint, it is preferable to convert the input information into the molecular structure information of the compound in SMILES notation or the like.

[0050] When structural formula information is input as image data, conversion can be performed using an AI model that has learned image data of the structural formula and a notation in which the molecular structure information of the compound is expressed as a character string. For example, an Image2text model such as BLIP or BLIP-2 can be fine-tuned and used as this AI model. When the molecular structure information of the compound is input using a plug-in for a chemical formula drawing tool, conversion can also be performed by converting the input into image data and then performing similar processing.

[0051] When the molecular structure information of a compound is input using a chemical formula drawing tool or its plug-in, it is also possible to directly convert the drawing information into various notation methods. Some chemical formula drawing tools are equipped with a function for converting to various notation methods, so that the information may be converted into SMILES notation or the like using that function.

[0052] It is preferable that the types of structural notation methods and rules regarding the orientation of drawn molecules are unified in this molecular structure generation system.

[0053] In the case of Figure 2C, it is preferable to convert the substance name into a representation such as SMILES notation, which is a representation method suitable for computer processing. In this case, conversion can be performed using an AI model that has learned the substance name and the SMILES notation. As this AI model, for example, a fine-tuned model such as Seq2Seq, LSTM (Long Short Term Memory), RNN (Recurrent Neural Networks), Transformer, or T5 (Text-To-Text Transfer Transformer) can be used.

[0054] In addition to this method, a two-stage conversion may be performed in which conversion is first performed using an AI model that has learned the substance name and image data of the structural formula, and then conversion is performed using an AI model that has learned the image data of the structural formula and the SMILES notation, etc. This can be done by fine-tuning the Image2text model BLIP, BLIP-2, etc.

[0055] After separating the molecular structure information portion of the compound from the desired portion for the characteristics, the desired portion for the characteristics may be used to extract information relating to the characteristics or physical properties, numerical ranges, etc.

[0056] Methods for extracting information include a method using pre-registered keywords related to characteristics or physical properties, an extraction method using an AI model trained on information extraction tasks related to characteristic or physical property information, and a method of obtaining information by providing a prompt to a large-scale language model to extract characteristic or physical property information and its numerical range.The AI ​​model may be a model fine-tuned for information extraction tasks such as BERT.

[0057] Next, an analogous compound is generated based on the molecular structure information of the compound (step S13_1). The analogous compound can be generated using a molecular generation model. Examples of the molecular generation model that can be used include a Boltzmann machine, a variational autoencoder (VAE), and a generative adversarial network (GAN). These models can be pre-trained using SMILES notation or the like, and the molecular structure information of the compound input to the first directive can be input to output an analogous compound. Datasets such as PubChem and CheEMBL can be used for pre-training. Note that the molecular generation model is not limited to the above models. The number of analogous molecules to be generated is arbitrary, but from the perspective of simplicity, when calculating the physical properties of the analogous molecules in a later step, it is preferable that the number be between 1 and 5.

[0058] In the VAE, the similar molecules output by using the VAE as they are may be used, but an edited molecular vector may be created by subtracting or adding an arbitrary amount to an arbitrary dimension of the molecular structure information of the compound converted into a molecular vector using the encoder, and this edited molecular vector may be converted into a SMILES notation or the like using the decoder of the same VAE, thereby obtaining a SMILES notation or the like of a compound having a structure similar to that of the compound input in step S11.

[0059] In this way, VAE can subtract any amount from any dimension of the vectorized molecular structure information of a compound, making it possible to generate more appropriate similar molecules.

[0060] Among VAEs, JT-VAE (Junction Tree Variational Autoencoder) is preferred because it is easy to create a reasonable structure since it starts with a certain skeleton (or substituent), sequentially constructs a molecule by adding a skeleton (or substituent) to it, and then adding a subsequent skeleton (or substituent) to that, which is a process similar to molecular design by humans.

[0061] The generation of a similar compound may be achieved by defining specific rules and substituting the molecular structure information of the compound by substitution according to the rules.

[0062] When possible, the physical properties or characteristics of the obtained similar molecules may be calculated or predicted. The calculated physical properties can be used as annotations for the similar molecules, and the large-scale language model can improve the accuracy of generating molecules with the desired properties.

[0063] For the calculation or prediction, molecular structure information of the analogous molecule is sent to the physical property calculation unit 13_1_1 as shown in Fig. 4, and the calculation or prediction is performed. When performing the calculation, it is preferable to perform, for example, quantum scientific calculation. Since the analogous compound generated in step S13_1 is expressed in a representation that is easy to process on a computer, such as SMILES notation, when performing quantum scientific calculation, it is preferable to first reproduce the three-dimensional structure from the representation and then perform the calculation.

[0064] When making predictions, predictions can be made using labeled trained regression models, etc.

[0065] Furthermore, after separating the molecular structure information of the compound and the requirements for the characteristics from the first directive sentence, one or more sentences (context sentences) serving as background information are obtained from the requirements for the separated characteristics or information on the characteristics or physical properties extracted based on the requirements for the characteristics, numerical ranges, etc. (step S13_2).

[0066] The context sentence may be obtained by searching a text-based database using information extracted based on the characteristic request as a keyword, or by converting the characteristic request portion into a sentence vector and evaluating the similarity with a database that has been converted into a sentence vector in advance. SentenceBERT or the like may be used to convert a sentence from which the characteristic request portion has been separated into a sentence vector. The similarity may be evaluated by calculating the cosine similarity of the vectors. The distance between vectors may be used instead of the cosine similarity.

[0067] Since step S13_1 and step S13_2 are independent from each other, they may be executed sequentially or in parallel. Furthermore, step S13_1 and step S13_2 may be operated by the same server, or may be independent, such as an analogous compound generating server 13_1 and a background information acquiring server 13_2, as shown in FIG. 4.

[0068] Next, a second directive containing information on the molecular structure and properties of the analogous compound is generated (step S14).

[0069] The second instruction corresponds to a prompt that is input to a large language model to obtain an answer.

[0070] The second set of instructions is preferably generated using a template that includes instructions requesting the generation of a molecule on demand, based on background information and example molecules.

[0071] Furthermore, the template is input with the similar molecule generated in step S13_1 as the molecular example, the context statement obtained in step S13_2 as the background information, and the first directive statement as the request. At this time, if the molecular structure information of the compound in the first directive statement is image data, a compound name, or the like, and is used by converting the representation method into SMILES notation or the like when obtaining the similar structure, it is preferable to change the molecular structure information portion in the first directive statement input as the request to the representation using the converted SMILES notation or the like and input it.

[0072] The second instruction sentence generated in step S14 is input as a prompt to a model (sentence generation model, dialogue model, etc.) that uses a large-scale language model, as shown in FIG. 3, and the generated result is obtained (S15).

[0073] The results can be obtained as the molecular structure of the compound, expressed in SMILES notation or the like.

[0074] Examples of large-scale language models include GPT-3, GPT-3.5, GPT-4, LaMDA (Language Model for Dialogue Applications), PaLM (Pathways Language Model), and PaLM2, with GPT-4 being preferred. In this case, the large-scale language model is preferably fine-tuned using a representation of molecular structure information of compounds used in molecular example and property requests. For this fine-tuning, datasets such as PubChem and CheEMBL can be used.

[0075] The molecular examples may be accompanied by calculated or predicted values ​​of physical properties related to the desired characteristics, which allows for the generation of molecular structures with even greater accuracy.

[0076] Finally, the obtained results are displayed to the user (step S16).

[0077] The display may be in the form of SMILES notation as output from the large-scale language model, or may be in the form of an image in which the SMILES notation is converted into a structural formula, or a plug-in for a chemical formula drawing tool, etc. By going through the above steps, candidate compounds having desired properties can be obtained using a large-scale language model.

[0078] If you want to obtain a new compound that matches the characteristics, you can add this to the "### Instructions" section of the second instruction. If you want to obtain a compound that matches the characteristics from existing substances, there is no need to specify this, but you can add a note to the "### Instructions" section to select from learned molecules. You may be asked which compound you want to obtain when inputting the first instruction or when generating the second instruction. Furthermore, since the first instruction can be input in natural language, you may include this information in the request section.

[0079] Alternatively, as shown in FIG. 7, after molecules generated by a large-scale language model are obtained, an operation (filtering) to remove known compounds (step S17) may be performed.

[0080] By displaying the compounds remaining after executing this step to the user (step S16_2), compounds that are more likely to be novel compounds can be presented. Filtering can be performed using datasets such as PubChem and CheEMBL. Furthermore, if a trained AI model is used to acquire similar compounds, filtering may be performed using the compounds used to train the AI ​​model.

[0081] In molecule generation based on learning using a typical labeled dataset, it is difficult to generate molecules by specifying properties other than those assigned as labels. However, by generating molecular structures using a large-scale language model, it becomes possible to extract relevant information from a vast amount of descriptions in papers and other sources, thereby obtaining compounds that may have those properties.

[0082] Furthermore, by specifying a base molecule, generating similar molecules, and then specifying multiple reference molecules, it becomes possible to generate a molecular structure with higher accuracy. Furthermore, by calculating or predicting physical properties related to the required characteristics using similar molecules and adding the values ​​to the second directive, it becomes possible to generate a molecular structure with higher accuracy.

[0083] Furthermore, when a user inputs information, the user can input desired characteristics in natural language, making the method for generating molecular structures more versatile.

[0084] Embodiment 2 In this embodiment, a molecular structure generation system 10 according to one embodiment of the present invention will be described.

[0085] 5 includes a reception unit 110, a storage unit 120, a processing unit 130, an output unit 140, and a transmission unit 150. Data is transmitted between the molecular structure generation system 10 and an information processing device 40 on the external service side (generation AI side) that provides the large-scale language model via a network 30. Data input and output between the reception unit 110, the storage unit 120, the processing unit 130, and the output unit 140 is transmitted via the transmission unit 150.

[0086] In the drawings accompanying this specification, the components are classified by function and shown as independent blocks in the block diagrams. However, in reality, it is difficult to completely separate the components by function, and one component may be involved in multiple functions. For example, part of the processing unit 130 may function as the reception unit 110. Also, one function may be involved in multiple components. For example, the processing performed by the processing unit 130 may be executed by different servers depending on the processing.

[0087] The receiving unit 110 receives an input sentence containing a plurality of words from outside the molecular structure generation system 10 .

[0088] The input statement received by the receiving unit 110 is a first directive statement, which includes information including molecular structure information of the original compound and a request for the desired properties.

[0089] The storage unit 120 preferably stores a pre-trained molecular generation model or specific rules for generating similar molecules, a trained language model for acquiring sentences that serve as background information, or a database of chemical papers, etc. Note that these may also be models or databases provided by external businesses. In this case, the storage unit 120 can be considered a conceptual collection of multiple storage media.

[0090] The storage unit 120 also sequentially stores user inputs and results of processing performed by the processing unit. For example, the storage unit 120 appropriately stores and reads out the first instruction statement input by the user, the molecular structure information portion and desired portion of the compound separated from the first instruction statement, the generated similar molecule, the physical properties or characteristics of the similar molecule obtained by calculation or prediction, the vectorized desired portion, background information, molecular vectors converted by the encoder, edited molecular vectors, SMILES notation of the generated similar molecule, a list of similar molecules, a list of similar molecule physical properties, calculation (prediction) results of the physical properties of the similar molecule, the second instruction statement, the generated molecule, and the like.

[0091] The storage unit 120 includes at least one of a volatile memory and a non-volatile memory. Examples of the volatile memory include a dynamic random access memory (DRAM) and a static random access memory (SRAM). Examples of the non-volatile memory include a resistive random access memory (ReRAM), a phase change random access memory (PRAM), a ferroelectric random access memory (FeRAM), a magnetoresistive random access memory (MRAM), and a flash memory. The storage unit 120 may also include a recording media drive, such as a hard disk drive (HDD) or a solid state drive (SSD).

[0092] The processing unit 130 separates the molecular structure information of the compound from the desired portion in the input first instruction sentence, converts the molecular structure information of the compound into SMILES notation or the like, extracts background information from the desired portion, generates similar molecules (converts the input compound's SMILES notation or the like into a molecular vector, edits the molecular vector, and converts the edited molecular vector into SMILES notation or the like), obtains physical properties of similar molecules (generates three-dimensional structures from SMILES notation or the like of similar molecules, performs quantum chemical calculations, and regression calculations), generates second instruction sentences, filters the generated compounds, etc. Furthermore, if these processes are performed by different servers, the processing unit 130 also transmits information to each server.

[0093] Furthermore, the processing unit 130 transmits the second instruction sentence as a prompt to the large-scale language model of the information processing device 40 via the network 30. The information processing device 40 generates a compound structure based on the data and transmits it to the molecular structure generation system 10 via the network 30.

[0094] The processing unit 130 may include, for example, an arithmetic circuit, a central processing unit (CPU), and a graphics processing unit (GPU).

[0095] The processing unit 130 may have a microprocessor such as a DSP (Digital Signal Processor). The microprocessor may be implemented by a PLD (Programmable Logic Device) such as an FPGA (Field Programmable Gate Array) or an FPAA (Field Programmable Analog Array). The processing unit 130 may also have a quantum processor. The processing unit 130 can perform various data processing and program control by interpreting and executing instructions from various programs using the processor. Programs that can be executed by the processor are stored in at least one of the memory area of ​​the processor and the storage unit 120.

[0096] The output section 140 presents the transmitted compound structures to the user.

[0097] FIG. 6 is a conceptual diagram of the molecular structure generation system 10 shown in FIG. 5, peripheral information terminals, and an information processing device 40 connected via a network 30.

[0098] 6, a molecular structure generation system 10 is connected to an information processing device 40 via a network 30. Furthermore, the molecular structure generation system 10 is connected to a plurality of information terminals 20 via a network 31. In FIG. 6, information terminal 20a, information terminal 20b, information terminal 20c, and information terminal 20d are shown as examples of the information terminals 20.

[0099] A user of the molecular structure generation system 10 can access the molecular structure generation system 10 from the information terminals 20a to 20d, etc. The reception unit 110 of the molecular structure generation system 10 can receive input from the information terminals 20a to 20d, etc. The user can then receive services using the molecular structure generation method of one aspect of the present invention.

[0100] A user can use the molecular structure generation system 10, for example, by using dedicated software or an application installed on the information terminal 20a to the information terminal 20d. Alternatively, a user can use the molecular structure generation system 10 from a web browser on the information terminal 20, for example.

[0101] The molecular structure generation system 10 can generate a molecular structure using an input statement (first instruction statement) input from the information terminal 20a via the network 31. The molecular structure generation system 10 can transmit the result of the generation of the molecular structure to the information terminal 20a via the network 31. This can reduce the computational load on the information terminal 20a. Here, the information terminal 20a has been described as an example, but the same can be said for the information terminals 20b to 20d.

[0102] The molecular structure generation system 10 can transmit the second instruction sentence as a prompt to the information processing device 40 via the network 30. The information processing device 40 can generate a molecule using the transmitted data. The molecular structure generation system 10 can receive molecular structure information of the generated compound from the information processing device 40 via the network 30. Then, the molecular structure generation system 10 can transmit data generated based on the molecular structure information of the compound received from the information processing device 40 to the information terminal 20a via the network 31.

[0103] The information terminals 20a to 20d are each an information terminal device such as a computer used by a user, and may also be referred to as a client computer. As an example, FIG. 6 illustrates the information terminal 20a, which is a desktop computer, the information terminal 20b, which is a notebook computer, the information terminal 20c, which is a smartphone, and the information terminal 20d, which is a tablet computer. The number of information terminals connected to the molecular structure generation system 10 is not particularly limited. While FIG. 6 illustrates four information terminals, the number of information terminals may be one, two, three, five or more. Examples of such information terminals include desktop information terminals, notebook information terminals, tablet information terminals, and mobile information terminals such as smartphones.

[0104] The molecular structure generation system 10 is a large computer such as a server computer or a supercomputer. The molecular structure generation system 10 is a computer with higher processing power than the information terminal 20. The molecular structure generation system 10 can function as a parallel computer. By using the molecular structure generation system 10 as a parallel computer, it is possible to perform large-scale calculations required for learning and inference of artificial intelligence (AI), for example. Furthermore, the molecular structure generation system 10 may be capable of performing processing using a natural language processing model using AI.

[0105] The information processing device 40 is a large computer such as a server computer or a supercomputer. The information processing device 40 is a computer with higher processing power than the information terminal 20. The information processing device 40 is also a computer with higher processing power than the molecular structure generation system 10. It is preferable that the information processing device 40 has a function as a parallel computer. By using the information processing device 40 as a parallel computer, it is possible to perform large-scale calculations required for AI learning and inference, for example. Note that when both the molecular structure generation system 10 and the information processing device 40 have a function as a parallel computer, the information processing device 40 has a higher processing power and can perform large-scale calculations than the molecular structure generation system 10.

[0106] The information processing device 40 can perform processing using a natural language processing model that uses AI. Examples of the natural language processing model that uses AI include BERT (Bidirectional Encoder Representations from Transformers) and T5 (Text-to-Text Transfer Transformer). The information processing device 40 can also perform processing using a model that utilizes a large-scale language model (such as a sentence generation model or a dialogue model). Examples of large-scale language models include GPT-3, GPT-3.5, GPT-4, LaMDA (Language Model for Dialogue Applications), PaLM (Pathways Language Model), and PaLM2, and it is preferable to use GPT-4.

[0107] The information processing device 40 can execute processing using an AI model that can perform various natural language processing. When the molecular structure generation system 10 and the information processing device 40 can each execute processing using a model that utilizes a large-scale language model, the information processing device 40 can execute processing using a larger-scale AI model than the molecular structure generation system 10.

[0108] Network 31 is a computer network smaller than network 30. Typically, network 30 is a global network, and network 31 is a local network. It is preferable to use the Internet, which is the foundation of the World Wide Web (WWW), as network 30. It is preferable to use an intranet or an extranet as network 31. In this embodiment, an example will be mainly described in which the Internet is used as network 30 and an intranet is used as network 31.

[0109] Other computer networks that can be used as the network 31 include a PAN (Personal Area Network), a LAN (Local Area Network), a CAN (Campus Area Network), a MAN (Metropolitan Area Network), a WAN (Wide Area Network), and a GAN (Global Area Network).

[0110] Furthermore, when wireless communication is performed, communication standards such as the fourth generation mobile communication system (4G), fifth generation mobile communication system (5G), and sixth generation mobile communication system (6G), or specifications standardized by the IEEE such as Wi-Fi (registered trademark) and Bluetooth (registered trademark), can be used as communication protocols or communication technologies.

[0111] The molecular structure generation system 10 shown in FIG. 6 is suitable for a case where the user who uses the information terminal 20 and the service provider belong to the same organization such as a company.

[0112] For example, it is preferable that data be transmitted and received between the information terminal 20 and the molecular structure generation system 10 using a network 31 established within an organization such as a company. It is also preferable that data be transmitted and received between the molecular structure generation system 10 and the information processing device 40 using the Internet. This allows applications, tools, services, etc. provided outside the organization to be used more safely than when the information terminal 20 is directly connected to the information processing device 40 via the Internet.

[0113] The above-described configuration can prevent leakage of confidential information and realize a molecular structure generation system that can be used safely by users. Note that data transmission and reception between the information terminal 20 and the molecular structure generation system 10 may be performed using a network 30 (such as the Internet).

[0114] A plurality of configuration examples shown in this embodiment mode can be combined as appropriate.

[0115] 10: molecular structure generation system, 20: information terminal, 20a: information terminal, 20b: information terminal, 20c: information terminal, 20d: information terminal, 30: network, 31: network, 40: information processing device, 110: reception unit, 120: storage unit, 130: processing unit, 140: output unit, 150: transmission unit

Claims

receiving a request for a first compound and a property as a first instruction sentence combining molecular structure information of the first compound and natural language; generating a plurality of similar compounds based on molecular structure information of the first compound; extracting information related to the characteristic from the request for the characteristic; generating a second directive; and acquiring molecular structure information of a second compound expected to have the property by inputting the second instruction sentence into a large-scale language model, A method for generating a molecular structure, wherein the second instruction sentence includes at least information related to the characteristics and the plurality of similar compounds. receiving a request for a first compound and a property as a first instruction sentence combining molecular structure information of the first compound and natural language; generating a plurality of similar compounds based on molecular structure information of the first compound; extracting information related to the characteristic from the request for the characteristic; generating a second directive; and acquiring molecular structure information of a second compound expected to have the property by inputting the second instruction sentence into a large-scale language model, The second instruction sentence has input items determined in advance, The input items include at least background information, molecular examples, and requirements; information relating to said characteristics as said background information; the plurality of similar compounds as the molecular examples; A method for generating a molecular structure, comprising inputting the first instruction text as the request. In claim 1 or 2, A method for generating a molecular structure comprising the step of separating the molecular structure information of the first compound and the request for the property from the first instruction statement. In claim 1 or 2, The method for generating a molecular structure, wherein the molecular structure information of the first compound is in SMILES notation. In claim 1 or 2, generating said plurality of similar compounds, converting the molecular structure information of the first compound into a first molecular vector; varying the first molecular vector to generate a second molecular vector; and converting the second molecular vector into molecular structure information of a corresponding compound. In claim 5, generating a plurality of similar compounds, A method for generating molecular structures using a variational autoencoder trained on molecular structure information of compounds. In claim 6, A method for generating molecular structures comprising the step of removing compounds trained by the trained variational autoencoder from molecules generated by the large-scale language model. In claim 1 or 2, performing calculations regarding the property demands for the plurality of similar compounds; A method for generating a molecular structure, comprising inputting the result of the calculation into the second instruction statement together with an example molecule. In claim 1 or 2, A step of extracting information related to the characteristic from the request for the characteristic, A method for generating molecular structures using a text vector database that vectorizes chemical papers.

Images