Method for generating information on new drug candidates using virtual molecularization of protein pocket structures
By employing virtual molecularization and AI to identify compounds with selective ligand properties for target proteins, the method addresses the high costs and time challenges in drug development, optimizing drug candidates with reduced side effects and improved success rates.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- NAMUICT CO LTD
- Filing Date
- 2025-03-31
- Publication Date
- 2026-07-10
Smart Images
Figure 2026116651000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for generating information on new drug candidate substances using virtual molecularization of protein pocket structures. More specifically, using virtual molecularization of protein pocket structures and artificial intelligence, an active substance having selective ligand characteristics only for a target protein that is a causative agent present in cells is identified from existing compounds, new compound information having ligand characteristics similar to those of the active substance for the target protein is generated, and among the new compounds, a new compound whose predicted binding force, In Vitro physical properties, and In Vivo physical properties characteristics meet pre-set criteria is determined as a new drug candidate substance, and relates to a method for generating information on new drug candidate substances.
Background Art
[0002] In modern society, with the development of science and technology and social systems that enable a clean and hygienic life, for example, the construction of social infrastructure such as electricity, communication, sewage, roads, ports, airports, etc., the expansion of social systems such as national health insurance, national pension, compulsory education, etc., and the development of medical technology, people's lives are becoming healthy and comfortable, and the average life expectancy is also increasing significantly.
[0003] However, although various difficult diseases are gradually being conquered by the development of medicine, in some cases of certain diseases, modern medicine still has difficulty in treatment and is fatal, so there are still diseases that lead to death when they occur.
[0004] In particular, various new drugs for the treatment of diseases have been developed, but the new drug development field is a typical high-risk and high-return industrial field, and a huge amount of time and cost are required for the development of new new drugs. Also, it takes an average of about 10 to 12 years from the new drug development process to the commercialization of new drug substances, and the average development cost per new drug is estimated to reach $2.168 billion. The huge amount of time and cost required for new drug development acts as an entry barrier in the growth of relatively small-scale small and medium-sized / venture pharmaceutical companies.
[0005] The new drug development market is dominated by a small number of global pharmaceutical companies known as "Big Pharma," which account for more than 30% of the total sales in the entire new drug development sector. Each of these global pharmaceutical companies maintains a strategy of filing patents at each stage of drug candidate development and monopolizing the economic value of those candidates.
[0006] In the pharmaceutical industry, the absolute scale of R&D spending is crucial for small and medium-sized pharmaceutical companies to gain a competitive edge in new drug development. However, the reality is that R&D spending at small and medium-sized / venture pharmaceutical companies is significantly insufficient compared to global pharmaceutical companies. As a result, the gap between global pharmaceutical companies, which have secured a dominant position in the new drug development market through massive R&D investments, and small and medium-sized / venture pharmaceutical companies is gradually widening.
[0007] On the other hand, with the recent advancements in artificial intelligence, the use of AI in new drug development is increasing. AI can effectively shorten the time and cost required for new drug development, quickly identify candidate groups that may raise patent issues in the early stages of drug development, and rapidly determine whether a drug will fail before reaching the clinical trial stage, thereby contributing to increasing the probability of success in new drug development. For these reasons, the number of cases in recent years where AI is applied to the entire process of new drug development, from candidate substance selection to clinical trials, is increasing.
[0008] The use of artificial intelligence in new drug development is revolutionizing traditional drug development processes by leveraging public databases, predicting drug responsiveness based on genomic data, selecting drug candidates, predicting protein structures, predicting drug toxicity and biological activity, and optimizing clinical trial design.
[0009] Furthermore, while it is crucial to identify target proteins with potential as disease treatments during the new drug development stage and to search for drug candidates that bind to the pocket structure of these target proteins, the large gap between in silico, in vitro, and in vivo experiments makes it extremely difficult for drug candidates that consider only the pocket structure of specific target proteins, as is currently the case, to pass cell experiments, animal tests, and clinical trials and be approved as drugs. Therefore, in order to dramatically increase the success rate of new drug development, which is known to be 1 in 10,000, it is necessary to overcome to the greatest extent possible the complexity of the stages in which the final product of a new drug is created and the inconsistencies at each stage.
[0010] Currently, in the new drug development process, AI models are primarily used to virtually screen for new compounds that have a high binding affinity to target proteins after the target protein has been selected. However, since humans are composed of approximately 36 trillion cells and possess over 25,000 genes and over 100,000 types of proteins, when a person takes a drug, it can bind not only to the target protein but also to countless other proteins, resulting in a variety of side effects and sequelae, as well as drug efficacy and toxicity.
[0011] In this context, unless the potential for new drug development is reflected in the virtual screening or optimization process of new compounds with high binding affinity to target proteins—including whether the new drug candidate can penetrate the cell membrane, induce structural and functional changes in specific proteins, induce gene mutations, and affect the activity of signaling molecules present within the cell, the possibilities for new drug development cannot be dramatically improved.
[0012] In other words, in the virtual screening process, it is important to reflect not only the protein level but also the multidimensional properties at the cellular level and the animal and human levels, and virtual screening methods for all proteins used to confirm the selective specificity of new drug candidates to target proteins on which they act as ligands require a large amount of computation.
[0013] Therefore, in order to effectively reduce the time and cost required for new drug development, it is necessary to develop technologies that utilize artificial intelligence and big data to generate optimal information on potential new drug candidates.
[0014] This invention was made out of the need described above, and proposes a method for generating information on new drug candidates. This method involves using virtual molecularization of protein pocket structures and artificial intelligence to identify effective substances from existing compounds that have selective ligand properties only for target proteins that are the cause of disease present in cells, generating information on novel compounds that have ligand properties similar to those of the effective substances for the target protein, and then selecting novel compounds from among the novel compounds whose predicted binding affinity to the target protein, in vitro properties, and in vivo properties conform to pre-set criteria as new drug candidates. The following discloses relevant prior art documents. [Prior art documents] [Patent Documents]
[0015] [Patent Document 1] Korean Patent No. 10-2296188 Specification [Patent Document 2] Korean Patent No. 10-2347108 Specification [Patent Document 3] Korean Patent No. 10-2558546 Specification [Patent Document 4] Korean Patent Application Publication No. 10-2024-0084664 Specification [Overview of the Initiative] [Problems that the invention aims to solve]
[0016] The present invention aims to identify effective substances from existing compounds that have selective ligand properties only for target proteins that cause disease and are present in cells, using virtual molecularization of protein pocket structures and artificial intelligence; generate information on novel compounds that have ligand properties similar to those of the effective substances for the target proteins; and, among the novel compounds, determine new drug candidate substances for which the predicted binding affinity to the target proteins, in vitro properties, and in vitro properties conform to pre-defined criteria, thereby generating information on new drug candidate substances. [Means for solving the problem]
[0017] To solve the aforementioned problems, the method for generating information on new drug candidates using virtual molecularization of protein pocket structures according to the present invention is:
[0018] The process involves an active substance search stage (S100) in which existing compounds that possess selective ligand properties only for target proteins, which are disease-causing proteins present in cells, are identified through virtual molecular modeling of protein pocket structures using artificial intelligence, and the structural morphology (chemical pose) information of the identified existing compounds is generated as active substance information;
[0019] The process includes a similar / leading substance search stage (S200) in which similar substance information is generated for novel compounds having ligand properties similar to the active substance for the target protein, similar substances that are predicted to bind to the target protein with a binding strength greater than or equal to a pre-set value are identified, and the identified similar substance information is generated as leading substance information;
[0020] The method is characterized by including a new drug discovery step (S300) in which the physical properties of lead materials are evaluated and lead materials whose physical properties conform to pre-set criteria are generated as new drug candidate material information. [Effects of the Invention]
[0021] The present invention grasps an active substance having selective ligand characteristics only for a target protein that is a causative agent present in cells from existing compounds by virtual molecularization of a protein pocket structure and artificial intelligence, generates information on a novel compound having ligand characteristics similar to those of the active substance for the target protein, and determines a novel compound among the novel compounds whose predicted binding force, In Vitro physical properties, and In Vivo physical properties characteristics meet pre-set criteria as a new drug candidate substance, thereby generating new drug candidate substance information. Thus, the present invention provides an effect that a new drug can be developed with relatively low R&D costs and a short development time.
Brief Description of the Drawings
[0022] [Figure 1] This is a flowchart of the present invention. [Figure 2] This is an explanatory diagram of the active substance search stage of the present invention. [Figure 3] This is an explanatory diagram of the active substance search stage of the present invention. [Figure 4] This is an explanatory diagram of the active substance search stage of the present invention. [Figure 5] This is an explanatory diagram of the similar / lead substance search stage of the present invention.
Modes for Carrying Out the Invention
[0023] Examples of the present invention will be described in detail with reference to the accompanying drawings.
[0024] The present invention provides a method for generating information on new drug candidates using virtual molecularization of protein pocket structures, which uses virtual molecularization of protein pocket structures and artificial intelligence to identify active substances from existing compounds that have selective ligand properties only for target proteins that are the cause of disease present in cells, generates information on novel compounds that have ligand properties similar to those of the active substances for the target protein, and determines novel compounds from among the novel compounds whose predicted binding affinity to the target protein, in vitro properties, and in vitro properties conform to pre-set criteria as new drug candidate substances. This invention provides the effect of developing new drugs with relatively low R&D costs and short development time, and is characterized by including an active substance search stage (S100), a similar / leading substance search stage (S200), and a new drug substance search stage (S300), as shown in Figure 1.
[0025] Specifically, the method for generating information on new drug candidates using the virtual molecularization of protein pocket structures according to the present invention is as shown in Figure 1,
[0026] The process involves an active substance search stage (S100) in which existing compounds that possess selective ligand properties only for target proteins, which are disease-causing proteins present in cells, are identified through virtual molecular modeling of protein pocket structures using artificial intelligence, and the structural and morphological information of the identified existing compounds is generated as active substance information;
[0027] The process includes a similar / leading substance search stage (S200) in which similar substance information is generated for novel compounds having ligand properties similar to the active substance for the target protein, similar substances that are predicted to bind to the target protein with a binding strength greater than or equal to a pre-set value are identified, and the identified similar substance information is generated as leading substance information;
[0028] The method is characterized by including a new drug discovery step (S300) in which the physical properties of lead materials are evaluated and lead materials whose physical properties conform to pre-set criteria are generated as new drug candidate material information.
[0029] The aforementioned active substance search step (S100) involves identifying existing compounds that have selective ligand properties only for target proteins, which are disease-causing proteins present in cells, through virtual molecular modeling of protein pocket structures using artificial intelligence, and generating active substance information from the structural and morphological information of the identified existing compounds.
[0030] In the development of new drug substances, it is necessary to confirm the selective ligand properties of each known compound for each target protein (disease-causing protein) present in cells. This process requires analyzing the ligand properties for each of the numerous (hundreds of millions to billions) protein-compound binding pairs that are binding pairs of known proteins and each known compound.
[0031] However, conventional analytical methods, even when using supercomputers, have the problem of requiring excessive time to analyze the ligand characteristics of each binding pair of numerous protein-compounds. The present invention aims to significantly reduce the time required for analyzing the ligand characteristics of each binding pair of numerous protein-compounds by using virtual molecular modeling of protein pocket structures with artificial intelligence.
[0032] Specifically, the effective substance search step (S100) is as shown in Figures 2, 3, and 4.
[0033] The first step (S110) is to generate binding pair information for protein-ligand compounds, which are binding pairs between each known protein and each known ligand compound, using binding pair information for existing known protein-ligand compounds,
[0034] The second step (S120) involves predicting the binding strength between the protein and the ligand compound for each possible binding pair of protein-ligand compounds, and generating information on possible binding pairs of protein-ligand compounds for which the predicted binding strength is equal to or greater than a predetermined value.
[0035] The third stage (S130) generates first training data regarding the binding relationship characteristics between proteins and ligand compounds that constitute binding pairs of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value,
[0036] The fourth stage (S140) generates second training data for the pocket structure of each protein that constitutes a binding pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value,
[0037] Step 5 (S150) involves generating a binding affinity prediction model for each known protein, which can predict the binding affinity between a protein and a compound when the protein binds to the compound, through virtual molecular modeling of the protein pocket structure of the artificial intelligence 10 that has learned the first and second training data.
[0038] The method is characterized by including a sixth step (S160) in which, using the generated existing known protein-specific binding affinity prediction models and the structural morphological information of each existing compound, existing compounds that act as ligands for target proteins but not for proteins other than the target protein are identified, and the structural morphological information of the identified existing compounds is generated as active substance information.
[0039] As shown in Figure 2A, the first step (S110) is a step in which binding pair information of protein-ligand compounds is generated using binding pair information of protein-ligand compounds that are binding pairs of each protein and each ligand compound that are already known, using binding pair information of protein-ligand compounds that are already known.
[0040] There are approximately 20,000 known protein-ligand compound binding pair combinations. From these existing combinations, we can obtain information on approximately 20,000 types of proteins and 20,000 types of ligand compounds. This allows us to obtain information on possible binding pairs of protein-ligand compounds (20,000 x 20,000 = approximately 400 million binding pairs) between these approximately 20,000 types of proteins and 20,000 types of ligand compounds.
[0041] For example, after obtaining information on proteins 1 through 19,443 and ligand compounds 1 through 19,443 from existing known protein-ligand compound binding pair information, 378,030,249 possible protein-ligand compound binding pair information can be generated by cross-binding these information.
[0042] As shown in Figure 2B, the second step (S120) is a step in which the binding force between the protein and the ligand compound is predicted for each possible binding pair of protein-ligand compounds, and information on possible binding pairs of protein-ligand compounds for which the predicted binding force is equal to or greater than a predetermined value is generated.
[0043] For example, when 378,030,249 bindable pair information for protein-ligand compounds is generated through the first step (S110), the binding force between the protein and the ligand compound is predicted for each of the 378,030,249 bindable pair of protein-ligand compounds generated using a conventional binding force prediction program (rDock, AutoDock, FlexAID, etc.). Then, each bindable pair of protein-ligand compounds whose predicted binding force is greater than or equal to a preset value (e.g., the absolute value of the predicted binding force points is 4 or more) is classified and identified, and information on bindable pair information for protein-ligand compounds whose predicted binding force is greater than or equal to a preset value is generated.
[0044] As shown in Figure 2C, the third step (S130) is a step in which first training data is generated regarding the binding relationship characteristics between proteins and ligand compounds that constitute a bindingable pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value.
[0045] The first learning data is generated by understanding the binding relationship characteristics between the protein and the ligand compound for each binding possible pair of protein-ligand compounds whose predicted binding force generated through the second step (S120) is greater than or equal to a preset value. Since the binding relationship characteristics between the protein and the ligand compound include the type, number, binding structure, and binding force characteristics of each protein atom in a binding relationship with the ligand compound, the first learning data includes information on the type, number, binding structure, and binding force characteristics of each protein atom in a binding relationship with the ligand compound.
[0046] For example, if the number of bindable pairs with a predicted binding strength greater than or equal to a preset value is 50,000,000, a supercomputer is used to determine the binding relationship characteristics of each protein atom around the protein pocket for each of the 50,000,000 bindable pairs of protein-ligand compounds. Once the binding relationship characteristics of the protein-ligand compounds, such as the type, number, binding structure, and binding strength characteristics of each protein atom around the protein pocket, are determined, first training data is generated regarding the binding relationship characteristics of the protein-ligand compounds that constitute the bindable pairs of protein-ligand compounds with a predicted binding strength greater than or equal to a preset value.
[0047] As shown in Figure 2D, the fourth step (S140) is the step of generating second training data for each pocket structure of each protein that constitutes a binding pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value.
[0048] For each bindable pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a predetermined value, generated through the second step (S120), the pocket structure of each protein constituting the bindable pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a predetermined value (e.g., the absolute value of the number of predicted binding strength points is 4 or more) is determined. Conventional pocket structure prediction programs (e.g., CLAPE, DeepProSite, etc.) can be used. The protein pocket refers to the protein portion to which the ligand compound is bound, and the determined protein pocket structure may relate to the structural shape or pattern of the pocket.
[0049] For example, if the number of bindable pairs with a predicted binding strength of 50,000,000 or more is greater than or equal to a preset value, the pocket structure of the proteins constituting these 50,000,000 bindable pairs of protein-ligand compounds is determined using a pocket structure prediction program (e.g., CLAPE, DeepProSite, etc.). Once the pocket structure of each protein is determined, second training data is generated for the pocket structure of each protein constituting each bindable pair with a predicted binding strength of 50,000,000 or more.
[0050] As shown in Figure 3, the fifth step (S150) is a step in which a binding affinity prediction model is generated for each known protein, which can predict the degree of binding affinity between the protein and the compound when the protein is bound to the compound, through virtual molecular modeling of the protein pocket structure of the artificial intelligence 10 that has learned the first and second training data.
[0051] Specifically, in the fifth stage (S150) described above, the artificial intelligence 10 learns the pocket structure characteristics of the protein pocket, which is the site to which the ligand compound is bound, for each protein using the second training data, and learns the characteristics of the mutual binding relationships between each protein atom around the pocket and each ligand compound atom that is bound to each protein atom, using the type, number, binding structure, and binding strength characteristics of each protein atom that is bound to the ligand compound included in the first training data.
[0052] Using the learning results, we define the properties of each virtual atom that replaces each actual protein atom around the pocket. We then perform a modeling process for each known protein, virtualizing the protein pocket as a virtual molecule composed of these defined virtual atoms. Once the protein pocket is modeled as a virtual molecule, we generate a binding affinity prediction model for each known protein that can predict the binding affinity between the protein and the compound when the protein binds to the compound.
[0053] The characteristics of each virtual atom are such that, when each actual ligand compound atom binds to each actual protein atom around the pocket, it acts by binding to each ligand compound atom in the same way as each actual protein atom.
[0054] The artificial intelligence uses the second training data to learn the pocket structure characteristics of the protein pocket, which is the site to which the ligand compound is bound. The second training data concerns the pocket structure of the protein bound to the ligand compound, for example, the structural shape and pattern of the pocket. The artificial intelligence uses the second training data to learn what structural characteristics the pocket structure of the protein bound to the ligand compound has.
[0055] Furthermore, the artificial intelligence learns the characteristics of the interbinding relationships between each protein atom around the pocket and each atom of the ligand compound that is bound to the protein, using the first training data. The first training data includes information on the type, number, binding structure, and binding strength characteristics of each protein atom that is bound to the ligand compound, and the artificial intelligence learns using the first training data what kind of binding characteristics each protein atom and each atom of the ligand compound that is bound to the protein are bound to.
[0056] Subsequently, the artificial intelligence uses the learning results to define the properties of each virtual atom that substitutes for each actual protein atom around the pocket. The properties of each virtual atom relate to how each actual ligand compound atom binds to and acts on each ligand compound atom, just like each actual protein atom when each actual ligand compound atom binds to each actual protein atom around the pocket.
[0057] Once the properties of each virtual atom are defined, the artificial intelligence constructs a virtual molecule composed of each virtual atom with defined properties, and then performs modeling to virtualize the protein pocket as this constructed virtual molecule. Modeling to virtualize the protein pocket means replacing the protein pocket with a virtual molecule composed of each virtual atom with defined properties, and this modeling to virtualize the protein pocket is performed for each already known protein.
[0058] Using virtual molecular models of protein pockets created for each known protein, the protein pockets are modeled as virtual molecules, and a binding affinity prediction model is generated for each known protein that can predict the degree of binding affinity between the protein and the compound when the protein binds to the compound.
[0059] A binding affinity prediction model, generated for each known protein, outputs a predicted value of the binding affinity between a specific protein and a specific compound, given specific compound structure and morphology information as input.
[0060] Through the aforementioned fifth step (S150), by inputting the structural morphological information of each existing compound as input values for each existing protein-specific binding affinity prediction model, the binding affinity between existing proteins and existing compounds can be predicted in a short time. This solves the problem in the conventional drug development process where ligand characterization analysis for each binding pair of numerous protein-compounds is excessively time-consuming.
[0061] The sixth step (S160) involves using the generated, existing, known protein-specific binding affinity prediction models and the structural and morphological information of each existing compound to identify existing compounds that act as ligands for target proteins but not for proteins other than the target protein, and generating the structural and morphological information of the identified existing compounds as active substance information.
[0062] New drug substances must possess selective ligand properties, meaning they act as ligands for target proteins but not for other proteins. Therefore, candidate drug substances must also possess selective ligand properties, meaning they do not act as ligands for proteins other than the target protein.
[0063] In this invention, after first identifying similar substances through the similar / leading substance search step (S200) described later, leading substances to be used as new drug candidate substances are identified from among the identified similar substances. However, since the similar substances are derived from the active substance information generated through step (S160), the active substances must also have selective ligand properties, acting as ligands for target proteins but not for proteins other than the target protein. Step 6 (S160) is performed to identify active substances that have selective ligand properties.
[0064] Specifically, the sixth step (S160) is characterized by individually inputting the structural morphological information of each existing compound extracted from the compound database for each of the generated existing known protein-specific binding affinity prediction models, generating binding affinity prediction information between each existing compound and each existing known protein, determining the reference binding affinity (the (-) binding affinity with the maximum absolute value among the binding affinities between the target protein and each existing compound) using the generated binding affinity prediction information between each existing compound and each existing known protein, selecting existing compounds corresponding to the effective substance selection conditions as effective substances, and generating the structural morphological information of the existing compounds selected as effective substances as effective substance information. The effective substance selection conditions are characterized by the fact that the absolute value of the binding affinity with other proteins is smaller than the first set percentage of the absolute value of the reference binding affinity, and the sum of the absolute values of the binding affinities with other proteins within the second set percentage range of the reference binding affinity is 3 times or less the absolute value of the reference binding affinity.
[0065] Referring to Figure 4, as an example, if there are 19,443 types of existing proteins and 220,799,302 types of existing compounds that can be extracted from the compound database, the structural morphological information of the 220,799,302 existing compounds is individually input to each of the 19,443 existing protein-specific binding affinity prediction models that have been generated, and binding affinity prediction information between each existing compound and each existing protein is generated.
[0066] Taking Figure 4 as an example, if the structural morphological information of each existing compound (approximately 220,799,302 compounds) extracted from the compound database is individually input into the binding affinity prediction model for the already known protein #1, the binding affinity prediction model for the protein #1 will generate binding affinity prediction information between the already known protein #1 and each existing compound (approximately 220,799,302 compounds). This binding affinity prediction information between the already known protein #1 and each existing compound (approximately 220,799,302 compounds) includes approximately 220,799,302 predicted binding affinity values ((predicted binding affinity between protein #1 and existing compound #1), predicted binding affinity between protein #1 and existing compound #2), (predicted binding affinity between protein #1 and existing compound #3), and (predicted binding affinity between protein #1 and existing compound #220,799,302)).
[0067] Furthermore, when the structural morphological information of each existing compound (approximately 220,799,302 compounds) extracted from the compound database is individually input into the existing known #2 protein binding affinity prediction model, the #2 protein binding affinity prediction model will generate binding affinity prediction information between the existing known #2 protein and each existing compound (approximately 220,799,302 compounds). This binding affinity prediction information between the existing known #2 protein and each existing compound (approximately 220,799,302 compounds) will include approximately 220,799,302 predicted binding affinity values ((predicted binding affinity between #2 protein and existing #1 compound), (predicted binding affinity between #2 protein and existing #2 compound), (predicted binding affinity between #2 protein and existing #3 compound), and (predicted binding affinity between #2 protein and existing #220,799,302 compounds)).
[0068] By individually inputting the structural morphological information of each existing compound (approximately 220,799,302 compounds) extracted from the compound database into the binding affinity prediction model for the last known #19,443 proteins using the method described above, binding affinity prediction information between each existing compound and each known protein is generated.
[0069] Once prediction information on the binding affinity between each existing compound and each known protein is generated, the reference binding affinity is determined using the generated prediction information for each existing compound and each known protein.
[0070] The binding affinity between each existing compound and each known protein can be either (+) or (-) binding affinity. A (+) binding affinity indicates a weak binding affinity between the compound and the protein, with a larger absolute value indicating less proportionality. A (-) binding affinity indicates a strong binding affinity between the compound and the protein, with a larger absolute value indicating more proportionality.
[0071] The reference binding affinity refers to the (-) binding affinity with the highest absolute value among the binding affinities between the target protein and each existing compound. For example, using the binding affinity prediction information between each existing compound and each known protein, if the absolute value of the binding affinity between the target protein and each existing compound is 18, then -18 is determined as the reference binding affinity.
[0072] After determining the reference binding affinity, an existing compound corresponding to the effective substance selection criteria is selected as an effective substance. The effective substance selection criteria are such that each absolute value of the binding affinity to each protein other than the target protein is less than a first set percentage of the absolute value of the reference binding affinity, while each binding affinity to each protein other than the target protein is within a second set percentage range of the reference binding affinity, and the sum of the absolute values of the binding affinity to each protein is 3 times or less the absolute value of the reference binding affinity, and the upper limit of the second set percentage range is smaller than the first set percentage, and the first and second set percentage ranges can be adjusted arbitrarily.
[0073] Once the reference binding affinity is determined, it is judged whether all existing compounds satisfy the criteria for selecting effective substances.
[0074] For example, we will explain the selection of an effective substance using the case where the determined standard binding affinity is -18, the first set percentage is 90%, and the second set percentage range is 30% to 20%.
[0075] Existing compounds whose absolute value of binding affinity to proteins other than the target protein is 16.2 or higher (90% of the absolute value of the reference binding affinity of -18, which is 18) are excluded from the list of active substance candidates. In other words, existing compounds whose binding affinity to at least one other protein is 90% or higher of the absolute value of the reference binding affinity are excluded from the list of active substance candidates.
[0076] Furthermore, existing compounds are selected as active ingredients if each absolute value of the binding affinity to each protein other than the target protein is less than 16.2, which is 90% of the absolute value of the reference binding affinity of 18 (i.e., they have a negative binding affinity and the absolute value of their binding affinity to other proteins is less than 16.2), while the binding affinity to each protein other than the target protein is in the range of -5.4 to -3.6, which is 30% to 20% of the reference binding affinity of -18, and the sum of the absolute values of the binding affinity to each protein is 54 or less, which is three times the absolute value of the reference binding affinity of 18.
[0077] For example, if there are existing compounds A and B, where each absolute value of the binding affinity to each protein other than the target protein is less than 16.2 (90% of the absolute value of the reference binding affinity), but each binding affinity to each protein other than the target protein is in the range of -5.4 to -3.6 (30% to 20% of the reference binding affinity), then if the sum of the absolute values of the binding affinity to each protein of existing compound A is 40 (less than or equal to 54, which is three times the absolute value of the reference binding affinity of 18), then existing compound A is selected as the active substance. If the sum of the absolute values of the binding affinity to each protein of existing compound B is 60 (greater than or equal to 54, which is three times the absolute value of the reference binding affinity of 18), then existing compound B is not selected as the active substance.
[0078] Through the process described above, effective substances are selected from existing compounds, and structural and morphological information of the selected existing compounds is generated as effective substance information (54 or less, which is three times the absolute value of the reference binding affinity of 18).
[0079] The aforementioned similar / leading substance search step (S200) involves generating similar substance information for novel compounds having ligand properties similar to the active substance for the target protein, identifying similar substances among the similar substances that are predicted to bind to the target protein with a binding strength equal to or greater than a pre-set value, and generating leading substance information from the identified similar substances.
[0080] Specifically, the similar / leading substance search step (S200) is characterized by generating structural morphological information of novel compounds that do not exist in the past and have ligand-binding characteristics similar to the active substances derived through the active substance search step (S100) for the target protein, for each active substance derived through the active substance search step (S100), generating the structural morphological information of each novel compound obtained as a result as similar substance information, predicting the predicted binding force between each novel compound generated as a similar substance and the target protein using a binding force prediction program, and generating the structural morphological information of each novel compound whose predicted binding force is equal to or greater than a preset value as leading substance information, wherein the ligand-binding characteristics include hydrogen bonding characteristics, ionic bonding characteristics, and pi-pi interaction characteristics.
[0081] For example, among the active substances derived through the active substance search stage (S100), the structural morphological information of a novel compound #1 that does not exist and has ligand-binding characteristics (including hydrogen bonding characteristics, ionic bonding characteristics, and pi-pi interaction characteristics) similar to those of active substance #1 with respect to the target protein is generated. As shown in Figure 5 as an example, each atom (substitutable atom) of active substance #1 that maintains the ligand-binding characteristics (including hydrogen bonding characteristics, ionic bonding characteristics, and pi-pi interaction characteristics) of the active substance with respect to the target protein when substituted with other atoms is determined, and each determined atom is substituted with other substitute atoms to generate structural morphological information of a novel compound #1 that does not exist and has ligand-binding characteristics similar to those of active substance #1 with respect to the target protein.
[0082] Using the same method, for each of the #2 active substances, structural morphological information of a novel #2 compound that does not exist and has ligand-binding characteristics (including hydrogen bonding characteristics, ionic bonding characteristics, and pi-pi interaction characteristics) similar to the #2 active substance with respect to the target protein is generated until the last active substance is found, and structural morphological information of each of the n novel compounds corresponding to all the active substances (n active substances) derived through the active substance search stage (S100) is generated as similar substance information.
[0083] Once similar substance information is generated, the predicted binding force between each new compound corresponding to the similar substance information and the target protein is predicted using a binding force prediction program (rDock, AutoDock, FlexAID, etc.). For each new compound whose predicted binding force is greater than or equal to a pre-set value (e.g., the absolute value of the predicted binding force points is 6 or greater), structural morphological information is generated as leading substance information.
[0084] The aforementioned new drug substance discovery step (S300) is a step in which the physical properties of lead substances are evaluated and lead substances whose physical properties conform to pre-set criteria are generated as new drug candidate substance information, characterized in that the physical properties of the lead substances include in vitro physical properties and in vivo physical properties.
[0085] The aforementioned in vitro physical properties include fat / water solubility, cell membrane permeability (Caco-2 and MDCK cell permeability), degree of inhibition of CYP450 enzyme activity, metabolic stability, plasma stability, drug-induced CYP expression, cytotoxicity, cellular activity, efficacy evaluated by the IC50 value, which is a quantitative measurement indicating the amount of lead substance required to inhibit a biological process or biological component by 50% in a test tube, drug tolerance, side effects, and drug stability in the body. The aforementioned in vivo physical properties also include the median lethal dose of experimental animals, change in lead substance concentration over time, time to reach peak blood concentration (Tmax), peak blood concentration (Tmax) and half-life (half-life, T1 / 2), whether or not the lead substance crosses the blood-brain barrier (BBB), and BBB permeability.
[0086] Leading substances whose in vitro and in vivo physical properties conform to pre-defined criteria are generated as new drug candidate substance information.
[0087] The technical concept of the present invention has been described above with accompanying drawings, but this is merely an illustrative description of preferred embodiments of the present invention and does not limit the invention. Furthermore, it should be obvious that the scope of the rights of the present invention is not limited to the embodiments, but also includes modifications made by persons with ordinary skill in this art within the scope of the technical concept of the present invention. [Explanation of symbols]
[0088] 10 Artificial Intelligence
Claims
1. In a method for generating information on new drug candidates using virtual molecularization of protein pockets, The process involves an active substance search step (S100) in which existing compounds that possess selective ligand properties only for target proteins, which are disease-causing proteins present in cells, are identified through virtual molecular modeling of protein pocket structures using artificial intelligence, and the structural and morphological information of the identified existing compounds is generated as active substance information; The process includes a similar / leading substance search step (S200) in which similar substance information is generated for novel compounds having ligand properties similar to the active substance for the target protein, similar substances are identified that are predicted to bind to the target protein with a binding strength greater than or equal to a pre-set value, and the identified similar substance information is generated as leading substance information; A method for generating new drug candidate substance information using virtual molecularization of protein pocket structures, characterized by comprising: a new drug discovery step (S300) in which the physical properties of lead materials are evaluated and lead materials whose physical properties conform to pre-set criteria are generated as new drug candidate substance information; and
2. The aforementioned effective substance search step (S100) is, The first step (S110) is to generate binding pair information for protein-ligand compounds, which are binding pairs between each known protein and each known ligand compound, using binding pair information for existing known protein-ligand compounds, The second step (S120) involves predicting the binding strength between the protein and the ligand compound for each possible binding pair of protein-ligand compounds, and generating information on possible binding pairs of protein-ligand compounds for which the predicted binding strength is equal to or greater than a predetermined value. The third step (S130) generates first learning data regarding the binding relationship characteristics between proteins and ligand compounds that constitute a bindingable pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value, The fourth step (S140) generates second learning data for each of the proteins that constitute a binding pair of protein-ligand compounds whose predicted binding strength is greater than or equal to a preset value, Step 5 (S150) involves generating a binding affinity prediction model for each known protein, which can predict the binding affinity between a protein and a compound when the protein and compound are bound, through virtual molecular modeling of the protein pocket structure of the artificial intelligence (10) that has learned the first and second training data. A method for generating new drug candidate substance information using virtual molecularization of protein pocket structures, as described in claim 1, comprising: a sixth step (S160) in which existing compounds that act as ligands for target proteins but not for proteins other than target proteins are identified using a generated existing known protein-specific binding affinity prediction model and structural morphological information of each existing compound, and the structural morphological information of the identified existing compounds is generated as active substance information.
3. The first training data includes information on the type, number, binding structure, and binding strength characteristics of each protein atom that is in a binding relationship with the ligand compound. In the fifth stage (S150), the artificial intelligence (10) Using the second training data, the structural characteristics of the protein pocket, which is the site where the ligand compound is bound, are learned for each protein. Using the type, number, binding structure, and binding strength characteristics of each protein atom that is in a binding relationship with the ligand compound included in the first training data, the characteristics of the mutual binding relationships between each protein atom around the pocket and each ligand compound atom that is in a binding relationship with that protein atom are learned. Using the learning results, we define the properties of each virtual atom that replaces each actual protein atom around the pocket, and perform a modeling of the protein pocket as a virtual molecule composed of each virtual atom with defined properties for each known protein. Once the protein pocket is modeled as a virtual molecule, we generate a binding affinity prediction model for each known protein that can predict the binding affinity between the protein and the compound when the protein binds to the compound. The method for generating information on new drug candidates using the virtualization of a protein pocket structure, as described in claim 2, characterized in that the characteristics of each virtual atom relate to the properties of binding to and acting on each ligand compound atom in the same way as actual protein atoms when actual ligand compound atoms bind to each actual protein atom around the pocket.
4. The sixth step (S160) is, The structural morphological information of each existing compound extracted from the compound database is individually input for each of the generated existing known protein-specific binding affinity prediction models. Binding affinity prediction information between each existing compound and each existing known protein is then generated. Using this generated binding affinity prediction information, the baseline binding affinity (the (-) binding affinity with the maximum absolute value among the binding affinities between the target protein and each existing compound) is determined. Then, existing compounds corresponding to the effective substance selection criteria are selected as effective substances, and the structural morphological information of the selected existing compounds is generated as effective substance information. The above-mentioned criteria for selecting the effective substance are: The conditions are as follows: Each absolute value of the binding affinity to each protein other than the target protein is less than the first set percentage of the absolute value of the reference binding affinity, while each binding affinity to each protein other than the target protein is within the second set percentage range of the reference binding affinity, and the sum of the absolute values of the binding affinity to each protein is 3 times or less the absolute value of the reference binding affinity, A method for generating information on a new drug candidate substance using virtual molecularization of a protein pocket structure, as described in claim 2, characterized in that the upper limit of the second set percentage range is smaller than the first set percentage.
5. The aforementioned similar / leading material search step (S200) is, Structural morphological information of novel compounds that do not exist and have ligand-binding characteristics similar to the active substances derived through the active substance search step (S100) for the target protein is generated for each active substance derived through the active substance search step (S100), and the structural morphological information of each novel compound obtained as a result is generated as similar substance information. The predicted binding force between each newly generated compound and the target protein is predicted using a binding force prediction program, and structural morphological information of each newly generated compound whose predicted binding force is greater than or equal to a pre-set value is generated as lead material information. A method for generating information on a new drug candidate substance using virtual molecularization of a protein pocket structure according to claim 1, characterized in that the ligand binding features include hydrogen bonding features, ionic bonding features, and pi-pi interaction features.
6. A method for generating information on new drug candidates using virtual molecularization of protein pocket structures, as described in claim 1, characterized in that the physical properties of the lead substance in the new drug discovery stage (S300) include in vitro physical properties and in vivo physical properties.