System and method for determining artificial intelligence model for generating target protein active molecule information

Abstract

Disclosed is a method for determining an artificial intelligence model for generating target protein active molecule information. The method for determining an artificial intelligence model for generating target protein active molecule information, of the present invention, comprises the steps of: performing pre-training with respect to a prior model that is generative artificial intelligence by using compound data of which general activity is known; performing training with respect to an activity prediction model for target protein; generating an agent model by replicating the prior model; generating N compounds from the prior model and storing the N compounds in an experience memory; and fine-tuning the agent model.

Description

System and method for determining an artificial intelligence model for generating target protein activity molecular information

[0001] The present invention relates to an artificial intelligence model determination system and method for generating target protein activity molecular information.

[0002]

[0003] Conventional new drug development requires enormous effort, time, and money. It's known that traditional drug development takes an average of 15 years or more and costs up to 2 trillion won.

[0004] With the recent advancement of artificial intelligence technology, research and development of various artificial intelligence models for new drug development are actively being conducted to reduce the cost of new drug development and increase efficiency.

[0005] In particular, de novo drug design refers to the process of proposing novel compounds that satisfy target conditions without a basic skeleton. Beginning in 2017, research on de novo drug design using generative artificial intelligence (AI) emerged.

[0006] For an AI model to learn to generate the active molecular structure for a specific protein, a large amount of molecular data on the activity of that protein is required. However, experimental data for unknown proteins, such as the coronavirus, is often scarce, so reinforcement learning methods using separate, independent activity prediction models are often employed.

[0007] If the activity prediction model predicts activity / inactivity for the compound generated by the generative model, that value can be used as a reward to reinforce the generative model.

[0008] Meanwhile, the number of all molecules with drug properties is ~10 60 , and the proportion of molecules active only against specific targets in this vast chemical space is very low.

[0009] In order to effectively achieve the generation of the desired molecular structure, it is necessary to train an optimal generation model by maintaining a balance between exploration and exploitation when using reinforcement learning.

[0010] The left side of Figure 5 represents overexploitation, while the middle side represents overexploitation. The right side of Figure 5 shows that molecules similar to active molecules are generated with balanced diversity and concentration as a result of balanced application of exploration and exploitation.

[0011]

[0012] The technical problem to be solved by the present invention is to provide an artificial intelligence model determination system and method that generates target protein activity molecular information capable of exploring novel active compound structures while simultaneously expecting high activity against the target protein.

[0013]

[0014] In order to solve the above technical problem, an artificial intelligence model determination method for generating target protein activity molecular information according to an embodiment of the present invention comprises the steps of performing pre-training on a prior model, which is a generative artificial intelligence, using data of compounds with generally known activity; performing training on an activity prediction model for a target protein; generating an agent model that replicates the prior model; generating N compounds from the prior model and storing them in an experience memory; and fine-tuning the agent model; wherein the fine-tuning step comprises the steps of extracting and generating a total of k compounds from the experience memory and the agent model to determine M tournament candidate groups; repeating a tournament method of selecting one of two arbitrary compounds to determine n final selection groups; and replacing some of the N compounds with the final selection group and further training the agent model; wherein the tournament can compare at least one indicator of the activity of two compounds for the target protein, a generation probability from the prior model, and a generation probability from the agent model. (Where k is less than N, n is less than or equal to k / 2, and M is less than or equal to k)

[0015] In one embodiment of the present invention, the step of determining n final selection groups by repeating the tournament method may include an activity tournament step of extracting any two compounds from the M tournament candidate groups, comparing the activities of each of the two compounds predicted by the activity prediction model, and repeatedly selecting one having a high activity to determine M / 2 first advancing groups; an exploration tournament step of extracting any two compounds from the M / 2 first advancing groups, comparing the probabilities of the two compounds to be extracted by the agent model, and repeatedly selecting one having a low probability to determine M / 4 second advancing groups; and an adjustment tournament step of extracting any two compounds from the M / 4 second advancing groups, comparing the probabilities of the two compounds to be extracted by the prior model, and repeatedly selecting one having a high probability to determine n=M / 8 final selection groups.

[0016] In one embodiment of the present invention, the fine-tuning step may be performed repeatedly a preset number of times.

[0017] In one embodiment of the present invention, the M tournament candidates may exclude some of the k compounds in consideration of the effectiveness of the compounds.

[0018] In one embodiment of the present invention, the experience memory extracts k / 2 compounds, and the agent model generates k / 2 compounds so that the M tournament candidates can be determined.

[0019] In one embodiment of the present invention, the experience memory extracts k / 2 compounds, and the agent model generates k / 2 compounds to determine the M tournament candidates. However, as the number of fine-tunings is repeated, the proportion of compounds generated by the agent model among the M tournament candidates may increase.

[0020] In one embodiment of the present invention, some of the N compounds to be replaced with the final selection group may be the n compounds having the lowest activity.

[0021] In order to solve the above technical problem, an artificial intelligence model determination system for generating target protein activity molecular information according to an embodiment of the present invention comprises: a control unit; and a memory including a prior model which is a generative artificial intelligence, an activity prediction model for a target protein, an agent model replicating the prior model, and an experience memory; wherein the control unit performs pre-learning on the prior model using data of compounds having generally known activities, performs learning on the activity prediction model, generates N compounds from the prior model and stores them in an experience memory, and fine-tunes the agent model, wherein the fine-tuning comprises extracting and generating a total of k compounds from the experience memory and the agent model to determine M tournament candidate groups, repeating a tournament method of selecting one of two arbitrary compounds to determine n final selection groups, replacing some of the N compounds with the final selection group, and further training the agent model; and the tournament can compare at least one indicator among the activity of two compounds for the target protein, the generation probability from the prior model, and the generation probability from the agent model.

[0022]

[0023] The present invention has the effect of enabling the exploration of novel active compound structures while simultaneously expecting high activity against target proteins.

[0024]

[0025] Figure 1 illustrates an artificial intelligence model determination system that generates target protein activity molecule information according to one embodiment of the present invention.

[0026] Figure 2 illustrates an artificial intelligence model determination method for generating target protein activity molecular information according to one embodiment of the present invention.

[0027] Figure 3 illustrates in detail a part of a method for determining an artificial intelligence model for generating target protein activity molecular information according to one embodiment of the present invention.

[0028] Figure 4 illustrates in detail a part of a method for determining an artificial intelligence model for generating target protein activity molecular information according to one embodiment of the present invention.

[0029] Figure 5 illustrates the distribution of information that can be generated in cases of overexploitation, overexploration, and a balance of the two.

[0030] Figure 6 illustrates an exemplary compound structure and a string corresponding to the compound structure.

[0031] Figure 7 is a diagram illustrating an artificial intelligence model determination method for generating target protein activity molecular information according to one embodiment of the present invention.

[0032] Figure 8 is a table showing the evaluation results of the performance of an artificial intelligence model determined by one embodiment of the present invention.

[0033] Figure 9 is a two-dimensional distribution visualization of compounds generated by an artificial intelligence model determined by one embodiment of the present invention together with known active molecules for a specific protein.

[0034] Figure 10 is a two-dimensional distribution visualization of compounds generated by an artificial intelligence model determined by one embodiment of the present invention together with known active molecules for a specific protein.

[0035]

[0036] The present invention is susceptible to various modifications and embodiments. Specific embodiments are illustrated and described in detail in the drawings. However, this is not intended to limit the present invention to specific embodiments, but rather to encompass all modifications, equivalents, and alternatives falling within the spirit and technical scope of the present invention.

[0037] The task information that supported the present invention is as follows.

[0038] [Project ID]1711191392

[0039] [Assignment Number] 2020R1A2C2004628 (NN35490)

[0040] [Ministry Name] Ministry of Science and ICT

[0041] [Name of Project Management (Specialist) Institution] National Research Foundation of Korea

[0042] [Research Project Name] Individual Basic Research (Ministry of Science and ICT) - (Type 1-2) Mid-career Research

[0043] [Research Project Title] Development of a System for Predicting Toxicity and Side Effects of Drug Candidates Based on Explainable AI

[0044] [Name of Project Performing Organization] Gwangju Institute of Science and Technology

[0045] [Research Period] March 1, 2023 - February 29, 2024

[0046] In describing the present invention, if it is determined that a detailed description of a related known technology may obscure the gist of the present invention, the detailed description is omitted.

[0047] Meanwhile, in the description of the present invention, ‘molecule’ may include the meaning of ‘compound’ depending on the context.

[0048]

[0049] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0050] Figure 1 illustrates an artificial intelligence model determination system (100) (hereinafter, also referred to simply as 'system (100)') that generates target protein activity molecule information according to one embodiment of the present invention.

[0051] Referring to FIG. 1, the system (100) includes a control unit (110), a communication unit (120), and a memory (130).

[0052] The control unit (110) is connected to the communication unit (120) and the memory (130) and controls them.

[0053] The control unit (110) may include at least one processor for controlling the above components and performing operations necessary for control.

[0054] The communication unit (120) can be connected to a user terminal (200) via a wired / wireless network or the Internet and transmit and receive information with them.

[0055] The communication unit (120) may include one or more of various communication modules such as 5G (5th generation communication), LTE-A (long term evolution advanced), LTE (long term evolution), Wi-Fi (wireless fidelity), FM communication, etc., and may transmit and receive information with the above components through the communication module. The communication module and communication method included in the communication unit (120) need not be limited to the above-described embodiment.

[0056] The memory (130) is connected to the control unit (110), and modules related to information analysis and processing and artificial intelligence models can be stored in the form of programs.

[0057] The memory (130) may include at least one of a storage means of non-volatile properties and a storage means of volatile properties. The storage means may include at least one of a flash memory, a hard-disc drive (HDD), a solid-state drive (SSD), a read-only memory (ROM), a buffer memory, and a random access memory (RAM), and is not necessarily limited to the above embodiment.

[0058] The memory (130) includes a prior model (131), an active prediction model (132), an agent model (133), and an experience memory (134).

[0059] The prior model (131) may be a generative artificial intelligence model.

[0060] The prior model (131) can be pre-trained using data on compounds known to be active as training data, regardless of protein type.

[0061] In an embodiment of the present invention, 1,583,425 compounds provided in the CheMBL database were used as pre-training data for the prior model (131).

[0062] Compounds can be represented using SMILES (Simplified Molecular Input Line Entry System) and used as pre-training models.

[0063] The left side of Figure 6 shows the structure of an example compound, and the right side shows a compound indicated by SMILES along with the compound structure.

[0064] Mapping compounds to SMILES is to apply compounds represented by SMILES as training data to a generative LSTM (Long Short-Term Memory) neural network model that learns strings.

[0065] The prior model (131) may include a three-layer LSTM neural network structure.

[0066] The activity prediction model (132) is an artificial intelligence model for predicting the activity of a compound when the compound is given to a specific protein.

[0067] In the embodiment of the present invention, the activity prediction model (132) used 3,230 of 3,881 compounds for the KOR protein provided in the ChEMBL database and 1,215 of 1,462 compounds for the PIK3CA protein provided in the PubChem BioAssay database as training data.

[0068] Fivefold CV (cross-validation) was used as training data.

[0069] Of the 3,881 compounds for the KOR protein, 651 were used as a test set, and of the 1,462 compounds for the PIK3CA protein, 247 were used as a test set.

[0070] The active prediction model (132) can use a random forest regression model. Unlike the prior model (131), compounds are expressed as Morgan Fingerprints in the active prediction model (132).

[0071] In the activity prediction model (132), the plC of the compound for the KOR protein 50 The values ​​can be trained so that the pKx values ​​for the PIK3CA protein can be predicted and output, respectively.

[0072] The active test set (TSA, Test set Actives) can be distinguished based on the active threshold among the test sets of the active prediction model (132). For KOR, the active threshold is 7.0 plC. 50, and the activation threshold for PIK3CA may be 8.0 pKx. Here, pKx may mean pKi or pKd. If the threshold is exceeded, it is classified as an active test set, and 292 of 651 test sets for KOR correspond to TSA, and 115 of 247 test sets for PIK3CA correspond to TSA.

[0073] TSA agent model (133) can be used to evaluate the distance between the TSA compound and the generated compound.

[0074] The agent model (133) is a model to which fine-tuning has been applied, starting from the trained prior model (131). The specific method of fine-tuning will be described later.

[0075] The experience memory (134) records information about compounds.

[0076] The experience memory (134) can record and update information on compounds, and can provide information on at least some of the compounds stored under the control of the control unit (110).

[0077] The experience memory (134) can be used in the fine-tuning process of the agent model (133).

[0078] Although not shown in the drawing, the user terminal (200) may include a control unit, a communication unit, a storage unit, and an input unit.

[0079] The control unit is connected to the communication unit, storage unit, and input unit and can control them.

[0080] The communications department can transmit and receive information to and from the system.

[0081] The storage unit can store necessary information to facilitate information processing. The storage unit can store applications that can be installed on a user terminal.

[0082] The input unit is an input interface for controlling the user terminal and may be composed of a keyboard or a touch screen.

[0083] The user terminal (200) configured in this manner may be, for example, any one of a smartphone, a tablet PC, a desktop PC, and a notebook, and even if it is not one of these, it is not limited to the above embodiment as long as it can perform the above functions.

[0084] A person who owns and uses a user terminal (200) may be an administrator or user of the system. The administrator or user may input or set various variables and information necessary for system control through the user terminal (200).

[0085]

[0086] Hereinafter, an artificial intelligence model determination method for generating target protein activity molecular information will be described, focusing on an artificial intelligence model determination system (100) that generates target protein activity molecular information. Unless otherwise specified, the artificial intelligence model determination method for generating target protein activity molecular information can be understood as being performed through the collaborative work of the system (100) and its subcomponents.

[0087] Figure 2 illustrates an artificial intelligence model determination method for generating target protein activity molecular information according to one embodiment of the present invention.

[0088] Referring to Fig. 2, learning data is prepared in step S210.

[0089] The training data refers to 3,230 of 3,881 compounds for the KOR protein provided in the ChEMBL database for the above-described prior model (131) and 1,215 of 1,462 compounds for the PIK3CA protein provided in the PubChem BioAssay database.

[0090] Learning data can be prepared by the system (100) accessing the database to download data or by the user terminal (200) accessing the database to download data and transmitting it to the system (100).

[0091] In step S220, the control unit (110) performs pre-learning on the fryer model (131).

[0092] Figure 3 illustrates step S220 in detail.

[0093] Referring to FIG. 3, in step S221, the control unit (110) converts information on 1,583,425 compounds provided from the CheMBL database into string information. The compound data is not for a specific target protein, but is data on compounds generally known to be active.

[0094] The string information can follow the SMILES format described above.

[0095] In step S222, the control unit (110) applies the converted 1,583,425 compound-corresponding SMILES strings to the LSTM neural network structure to perform learning. The LSTM neural network can be used in three layers.

[0096] The learned prior model (131) is Prior or can be displayed as

[0097] The training loss (negative log likelihood) of the prior model (131) can be expressed as follows [Mathematical Formula 1].

[0098]

[0099]

[0100]

[0101] Referring again to FIG. 2, in step S230, the control unit (110) performs learning of the active prediction model (132).

[0102] The training of the active prediction model (132) is performed by training 3,230 of 3,881 compounds for the KOR protein and 1,215 of 1,462 compounds for the PIK3CA protein provided in the PubChem BioAssay database.

[0103] Fivefold CV (cross-validation) was used as training data.

[0104] The active prediction model (132) uses a random forest regression model, and the compound information can be converted into Morgan Fingerprint and applied as learning data.

[0105] The active prediction model (132) is a learning result that predicts the plC of compounds for the KOR protein. 50 The values ​​are trained so that the pKx values ​​for the PIK3CA protein can be predicted and output, respectively.

[0106] In step S240, the control unit (110) generates an agent model (133).

[0107] The agent model (133) may be a direct copy of the learned prior model (131).

[0108] Agent model (133) is Agent or can be displayed as

[0109] At step S250, the control unit (110) initializes the experience memory (134).

[0110] Specifically, the control unit (110) randomly generates N compounds by the agent model (133) and stores them in the experience memory (134). In this step, since the learned prior model (131) and the agent model (133) are the same, the same result is obtained even when the control unit (110) generates N compounds by the prior model (131) and stores them in the experience memory (134).

[0111] Here, N can be set to a sufficiently large number, and in the embodiment of the present invention, N=100,000.

[0112] In step S260, the control unit (110) performs fine-tuning on the agent model (133).

[0113] Figure 4 illustrates step S260 in detail.

[0114] Referring to FIG. 4, in step S261, the control unit (110) randomly extracts a first predetermined number of compounds corresponding to a portion of all compounds stored from the experience memory (134). The first predetermined number may be k / 2.

[0115] For example, when k=20,000, the control unit (110) can randomly extract 10,000 compounds from among 100,000 compounds stored in the experience memory (134).

[0116] In step S262, the control unit (110) can generate a second predetermined number of compounds from the agent model (133). The second predetermined number may be k / 2.

[0117] For example, when k=20,000, the control unit (110) can control the agent model (133) to randomly generate 10,000 compounds.

[0118] The first and second predetermined numbers can be set to the same number to balance exploration and utilization, but can be adjusted as needed. For example, the first and second predetermined numbers can be set to 1 / 4k and 3 / 4k, respectively, or conversely, 3 / 4k and 1 / 4k.

[0119] The first predetermined number and the second predetermined number may be determined in accordance with the number of iterations of fine-tuning. For example, at the beginning of fine-tuning, the first predetermined number and the second predetermined number are 1 / 2k and 1 / 2k, respectively, and as the number of fine-tuning increases, the relative proportion of the second predetermined number, i.e., the number of compounds generated by the agent model (133), may increase. For example, the relative proportions of compounds extracted or generated may be adjusted so that the first predetermined number and the second predetermined number become 1 / 4k and 3 / 4k, respectively, at the final fine-tuning.

[0120] In step S263, the control unit (110) determines the tournament candidates.

[0121] The tournament candidates may be k compounds in total, including k / 2 compounds extracted from the experience memory (134) and k / 2 compounds generated from the agent model (133), where k is smaller than N.

[0122] However, invalid compounds may be excluded from the tournament candidate pool. Therefore, the number of tournament candidates may be less than k. Here, validity may refer to the physical and chemical possibility of a compound's existence.

[0123] Hereinafter, the number of tournament candidates is denoted as M. M is less than or equal to k.

[0124] In step S264, the control unit (110) conducts an activity tournament on the tournament candidates.

[0125] The activity tournament selects a random compound from the candidate pool. , Extract, , The target predicted activity is compared to select a compound with higher activity, the selected compound is excluded from the candidate group, and the unselected compound is kept in the candidate group. Here, the indices i and j represent two randomly selected compounds and do not represent compounds arranged in a specific order. The above process can be repeated until M / 2 compounds are selected and the first advancing group is determined.

[0126] The target prediction activity is in the activity prediction model (132). and can be obtained by inputting each of them. At this time, the string in SMILES format can be mapped to Morgan Fingerprint.

[0127] Activity can be calculated for a specific protein. If the activity prediction model (132) is trained for each of multiple proteins, or includes multiple trained models for each protein, activity can be calculated for each of multiple proteins. Meanwhile, fine-tuning can be performed separately for each target protein.

[0128] In one embodiment of the present invention, the activity tournament selects a method that repeatedly compares and selects two random compounds, rather than sorting the compounds in descending order of activity and selecting the top M / 2 compounds. This tournament method does not necessarily guarantee that any of the selected compounds will have higher activity than any of the eliminated compounds. Instead, this tournament method provides a balance that does not excessively favor the selection of compounds with a strong element. This characteristic is also present in other tournaments below.

[0129] The activity tournament results can determine the first group of M / 2 qualifiers.

[0130] In step S265, the control unit (110) performs a search tournament.

[0131] The exploration tournament is a random compound among the M / 2 compounds in the first group. , Extract, , Compounds with a lower probability of being extracted from the agent model (133) may be selected, and compounds with a higher probability of being extracted may be eliminated. The above process is repeated until all compounds in the first group are selected.

[0132] The exploration tournament is a random compound among the M / 2 compounds in the first group. , Extract, , In the agent model (133), a compound with a lower probability of being extracted is selected, the selected compound is excluded from the first advance group, and the unselected compound is kept in the first advance group. Here, the indices i and j represent two randomly selected compounds and do not represent compounds arranged in a specific order. The above process can be repeated until M / 4 compounds are selected and the second advance group is determined.

[0133] The control unit (110) can input a compound into the agent model (133) and calculate the probability of extraction from the agent model (133).

[0134] The purpose of the exploration tournament is to maintain or increase the exploration potential of various compounds of the agent model (133) by having the agent model (133) select compounds with low extractability.

[0135] The results of the Exploration Tournament may determine the second group of M / 4.

[0136] In step S266, the control unit (110) performs a coordination tournament.

[0137] The adjustment tournament is a random compound among the M / 4 compounds of the second group. , Extract, , In the prior model (131), compounds with a higher probability of being extracted are selected, the selected compounds are excluded from the second group of advances, and the unselected compounds are kept in the second group of advances. Here, the indices i and j represent two randomly selected compounds and do not represent compounds arranged in a specific order. The above process can be repeated until M / 8 compounds are selected and the final selection group is determined.

[0138] The control unit (110) can input the compound into the prior model (131) and calculate the probability of extraction from the prior model (131).

[0139] The aim of the steering tournament is to select compounds with a high extractability of the prior model (131) and prevent the agent model (133) from deviating too much from compounds with high chemical efficacy.

[0140] The final selection group can be determined by the results of the adjustment tournament. n can be equal to or less than k / 2. In one embodiment of the present invention, n = M / 8. Depending on the embodiment, n can be M / 2 or M / 4, depending on the number of tournaments.

[0141] Steps S264 to S266 are preferably performed sequentially, but are not necessarily limited thereto, and may be performed in a different order. However, if the order is changed, the results of the fine-tuning may vary.

[0142] Steps S264 to S266 may be performed as a single step by combining functions from two or more steps. For example, the control unit (110) may perform any compound , Extract, , The weights are the probability of being extracted from the agent model (133). Multiply by , , Weights the probability of being extracted from the Prior model (131) Compare the value obtained by multiplying the former by the latter and subtracting the former. , You can select any one of them and exclude it from the candidate pool.

[0143] In step S267, the control unit (110) updates the experience memory (134).

[0144] Updating the experience memory (134) may be done by selecting n compounds in the order of the lowest predicted activity among the N compounds stored in the experience memory (134) and replacing them with n compounds from the final selection group.

[0145] In step S268, the control unit (110) additionally learns (trains) the agent model (133) with n compounds of the final selection group.

[0146] As step S268 is repeated multiple times, the agent model (133) may have a higher probability of generating diverse and effective compounds with high predicted activity.

[0147] In step S269, the control unit (110) determines whether steps S261 to S268 have been repeated a set number of times. If they have been repeated a set number of times, the process proceeds to step S270, and if they have not been repeated, the process returns to step S260.

[0148] In step S270, the control unit (110) determines the agent model (133) that has undergone fine-tuning for the set number of times as an active molecular information generation model.

[0149] Figure 7 is a diagram illustrating an artificial intelligence model determination method for generating target protein activity molecular information according to one embodiment of the present invention.

[0150] Figure 8 is a table showing the evaluation results of the performance of an artificial intelligence model determined by one embodiment of the present invention.

[0151] Prior is a learned prior model (131), and the example is a performance evaluation result for a determined active molecular information generation model that has completed fine tuning.

[0152] “Segler” refers to Segler MHS, Kogej T, Tyrchan C, Waller MP (2018) Generating focused molecule libraries for drug discovery with recurrent neural networks. Information published in ACS Cent Sci 4(1):120–131. “REINVENT” Olivecrona M, Blaschke T, Engkvist O, Chen H (2017) Molecular de-novo design through deep reinforcement learning. Information published in J Cheminform 9(1):48. “DrugEX” is Liu J Cheminform 11(1):35. Information published in, “AHC” represents the results of the model according to information published in Thomas M, O'Boyle NM, Bender A, de GraafC (2022) Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation. J Cheminform14(1):68.

[0153] The results of each performance evaluation are presented for KOR and PIK3CA, respectively.

[0154] Evaluation metrics can be divided into standard metrics and optimization metrics.

[0155] Standard metrics that assess the quality of general generative models, regardless of purpose, include validity, uniqueness, novelty, and diversity.

[0156] Validity is evaluated as the ratio of generated compounds to valid compounds, and the higher the ratio, the more preferred the model.

[0157] Uniqueness is evaluated as the proportion of unique compounds among the generated compounds, with a higher value indicating a preferred model.

[0158] Novelty is evaluated as the ratio of compounds that are not in the pre-training data among the generated compounds, and the higher the novelty, the more preferred the model.

[0159] Diversity refers to the average pairwise distance between the generated compounds, and a higher value is considered a preferred model.

[0160] Optimization indices include PredAct, PwSim, FCD, and OTD, which evaluate the model's active optimization performance and whether the generated compounds are close to the actual active distribution.

[0161] PredAct is the average of the predicted activities of the products, and a higher value is considered a preferred model.

[0162] PwSim represents the average similarity between products and TSA, and a higher value is considered a preferred model.

[0163] FCD represents the distance between the distributions formed by two groups in chemical space. Compounds can be converted into vectors in chemical space. Gaussian distributions are established for the product vector set and the TSA vector set, and the distance between them is calculated. A lower value is considered a preferred model.

[0164] Based on optimal transport theory, OTD defines the distance between the product distribution and the TSA distribution as the optimal mapping distance that minimizes the total distance when each molecule is transported 1:1. The intermolecular distance is measured using Tanimoto similarity. A lower value is considered a more desirable model.

[0165] Each model shown in Fig. 8 is a comparison of models that have undergone the same pre-training as the embodiment of the present invention and use the same prediction model, but have undergone different fine-tuning than the embodiment of the present invention.

[0166] The fine-tuned agent model (133) according to the embodiment of the present invention exhibited high optimization performance for both KOR and PIK3CA. In particular, the FCD and OTD indices were very good, indicating that the products appeared in a distribution very similar to the actual target active molecules.

[0167] In contrast, some comparison models (e.g., VGPC, DrugEx) exhibit inferior FCD and OTD results compared to the Prior model, indicating that they deviate from the actual target activation distribution even further than the pretrained model. This suggests that solving this optimization problem requires an appropriate fine-tuning strategy, as described in the embodiments of the present invention.

[0168] FIG. 9 is a two-dimensional distribution visualization of compounds generated by an artificial intelligence model determined by one embodiment of the present invention together with known active molecules for the KOR protein, and FIG. 10 is a two-dimensional distribution visualization of compounds generated by an artificial intelligence model determined by one embodiment of the present invention together with known active molecules for the PIK3CA protein.

[0169] The comparative model illustrated in Fig. 8 was also visualized with the same distribution in Figs. 9 and 10.

[0170] For each model, 20,000 compounds were generated and converted into Morgan fingerprint vectors, including both TSA and some pretrained compounds. The pretrained compounds were 50,000 randomly selected from the 1,583,425 compounds in the pretrained data. Although not shown in Figures 9 and 10, these compounds were also used to calculate the two-dimensional axes in t-SNE dimensionality reduction.

[0171] After transforming the vectors into a two-dimensional space using the t-SNE dimensionality reduction method, 1,000 randomly selected vectors (blue dots in the figure) and TSA (red dots in the figure) for each model are displayed in the figure.

[0172] The model that formed the product distribution covering the TSA of KOR and PIK3CA (with identical or close coordinates) was the fine-tuned agent model (133) of the present invention, which was the best. In other words, the fine-tuned agent model (133) of the present invention is evaluated to have transformed the distribution of the existing prior model through sufficient exploration into a model that extracts products closer to the actual active compound distribution.

[0173] In contrast, the product distributions of some comparative models exhibit weak search, resulting in a product distribution that fails to find most TSAs. For example, VGPC for the KOR protein and DrugEx for the PIK3CA protein were examples of this.

[0174]

[0175] The terminology used in this application is only used to describe specific embodiments and is not intended to limit the present invention. In this application, terms such as "comprise" or "have" are intended to indicate the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but should be understood to not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

Claims

1. A method for determining an artificial intelligence model that generates target protein activity molecular information executed by one or more control units of a computer device, A step of performing pre-training on the generative artificial intelligence, the Prior model, using data of compounds with known general activity; A step of performing learning on an activity prediction model for a target protein; A step of creating an agent model replicating the above-mentioned prior model; A step of generating N compounds from the above fryer model and storing them in the experience memory; and a step of fine-tuning the agent model; The above fine-tuning step is, A step of extracting and generating a total of k compounds from the above experience memory and the agent model to determine M tournament candidates; A step of determining n final selection groups by repeating a tournament method of selecting one of two random compounds; and A step of replacing some of the N compounds with the final selection group and further learning the agent model; A method for determining an artificial intelligence model for generating target protein activity molecular information, characterized in that the above tournament is conducted including at least the following three types. Type 1: Comparing the activity of two compounds against the target protein and selecting the one with higher activity. Type 2: Comparing the generation probabilities from the above agent models of two compounds and selecting the one with the lower probability. Type 3: Comparing the formation probabilities from the above-mentioned prior model of two compounds and selecting the one with the higher probability. (Where k is less than N, n is less than or equal to k / 2, and M is less than or equal to k) 2. In paragraph 1, The step of determining n final selection groups by repeating the above tournament method is: An activity tournament step for determining M / 2 first advancing groups by repeatedly comparing the activity predicted by the activity prediction model for two random compounds extracted from among the M tournament candidates and selecting one with higher activity; An exploration tournament step for determining M / 4 second advancing groups by repeatedly comparing the probability of the agent model extracting any two compounds extracted from the M / 2 first advancing groups and selecting one with a lower probability; and An artificial intelligence model determination method for generating target protein activity molecular information, characterized in that it comprises an adjustment tournament step for determining the final selection group of n=M / 8 by repeatedly comparing the probability that the above-mentioned prior model will extract and selecting one with a higher probability.

3. In paragraph 1, An artificial intelligence model determination method for generating target protein activity molecular information, characterized in that the above fine-tuning step is performed repeatedly a preset number of times.

4. In paragraph 1, A method for determining an artificial intelligence model for generating target protein activity molecular information, wherein the M tournament candidates are characterized in that some of the k compounds are excluded in consideration of the effectiveness of the compounds.

5. In paragraph 1, An artificial intelligence model determination method for generating target protein activity molecule information, characterized in that the above experience memory extracts k / 2 compounds and the agent model generates k / 2 compounds, thereby determining the M tournament candidates.

6. In paragraph 3, The above experience memory extracts k / 2 compounds, and the agent model generates k / 2 compounds to determine the M tournament candidates. An artificial intelligence model determination method for generating target protein activity molecular information, characterized in that as the number of repetitions of the above fine-tuning step increases, the proportion of compounds generated by the agent model among the M tournament candidates increases.

7. In paragraph 1, An artificial intelligence model determination method for generating target protein activity molecular information, characterized in that some of the N compounds to be replaced by the final selection group are n compounds selected in descending order of activity.

8. Control unit; and A generative artificial intelligence, a priori model, an activity prediction model for a target protein, an agent model replicating the priori model, and a memory including an experience memory; The above control unit, Pre-training is performed on the above-mentioned prior model using data on compounds with known general activity, Perform learning on the above active prediction model, Generate N compounds from the above fryer model and store them in the experience memory, Fine-tune the above agent model, The above fine-tuning is, Extract and generate a total of k compounds from the above experience memory and the agent model to determine M tournament candidates, A tournament method is repeated to select one of two random compounds, and a final selection set of n is determined. Replace some of the N compounds with the final selection group, and further train the agent model, An artificial intelligence model decision system for generating target protein activity molecular information, characterized in that the above tournament is conducted including at least the following three types. Type 1: Comparing the activity of two compounds against the target protein and selecting the one with higher activity. Type 2: Comparing the generation probabilities from the above agent models of two compounds and selecting the one with the lower probability. Type 3: Comparing the formation probabilities from the above-mentioned prior model of two compounds and selecting the one with the higher probability. (Where k is less than N, n is less than or equal to k / 2, and M is less than or equal to k)

Images