Information processing system, information processing method, program, and method for manufacturing molecular compound

Abstract

The present disclosure provides an information processing system for identifying a molecule suitable for a drug candidate, the information processing system comprising: a feature prediction unit that predicts a feature of a molecule constituting an element of a sequence information set that is a set of sequence information of a plurality of different molecules; calculating a feature prediction value for each molecule using a prediction model for predicting a feature of the molecule from the sequence information of the molecule, and estimating uncertainty of the prediction; and a candidate molecule identification unit that searches for a candidate of a molecule having a desired feature on the basis of the feature prediction value and the estimated value of the uncertainty of the prediction.

Description

Technical Field

[0001] One aspect of the present invention relates to an information processing system, a molecular design apparatus, an information processing method, a program, and a method for producing molecular compounds. Background Technology

[0002] Recently, in the pharmaceutical field, machine learning information processing technology has been used to alleviate the burden of drug discovery (Patent Document 1). Attempts have been made to utilize predictions made through machine learning to efficiently discover molecules with the properties required for drugs.

[0003] [List of Citations]

[0004] [Patent Literature]

[0005] [Patent Document 1] International Publication No. WO 2018 / 132752 Summary of the Invention

[0006] [Technical Issues]

[0007] Predictions made through machine learning introduce uncertainty, and failing to account for this when utilizing such predictions can lead to results that do not meet expectations. In drug discovery, there are many objects to be analyzed, and it is difficult to apply machine learning predictions to all objects to which machine learning predictions can be performed.

[0008] In view of the above, one aspect of this disclosure is to provide a technique for further reducing the burden of accompanying drug discovery.

[0009] [Solution to the problem]

[0010] One aspect of this disclosure provides the following model.

[0011] [A1]

[0012] An information processing system for identifying molecules suitable for drug candidates, the information processing system comprising:

[0013] A feature predictor, configured to use a prediction model to calculate feature predictions for a molecule that forms an element of a set of structural unit combination information, which is a set of information about combinations of structural units of various molecules; and then estimate the uncertainty of the prediction. The prediction model is used to predict the features of the molecule from information about the structural unit combination of the molecule.

[0014] A candidate molecule discriminator is configured to search for candidates of molecules with desired characteristics based on feature predictions and estimates of the uncertainty of those predictions.

[0015] [A2]

[0016] According to the information processing system described in A1, the structural unit combination information is sequence information about the molecule.

[0017] [A3]

[0018] According to the information processing system described in A1 or A2, the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody, and low molecular weight compound.

[0019] [A4]

[0020] The information processing system according to any one of A1 to A3, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the information of the combination of structural units is information about the amino acid sequence.

[0021] [A5]

[0022] The information processing system according to any one of A1 to A4, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0023] [A6]

[0024] The information processing system according to any one of A1 to A5, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to a target molecule.

[0025] [A7]

[0026] The information processing system according to any one of A1 to A6 further includes a predictive information processor configured to calculate a feature quality value based on a predicted feature value and an estimate of the uncertainty of the prediction.

[0027] The candidate molecule discriminator identifies at least one candidate molecule from the elements of the set of structural unit combination information based on this characteristic quality value.

[0028] [A8]

[0029] According to any one of A1 to A7, in the information processing system, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimate of the uncertainty of the prediction.

[0030] [A9]

[0031] According to the information processing system described in A8, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0032] [A10]

[0033] According to any one of A7 to A9, in the information processing system, the predictive information processor calculates the mean and variance as an objective function to give the feature quality value.

[0034] [A11]

[0035] The information processing system according to any one of A1 to A10, wherein the estimate of the uncertainty of the prediction is the standard deviation of the feature prediction value.

[0036] [A12]

[0037] The information processing system according to any one of A1 to A11, wherein the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimated value of the prediction uncertainty, uses a combinatorial optimization algorithm to obtain a set of structural unit combination information, and updates the extraction parameters of the combinatorial optimization algorithm based on the feature prediction value or feature quality value of the molecule.

[0038] [A13]

[0039] According to the information processing system described in A12, the combinatorial optimization algorithm is a tree-structured Parzen estimator.

[0040] [A14]

[0041] The information processing system according to any one of A1 to A13, wherein the prediction model is a prediction model generated by learning based on information about the combination of structural units of a variety of molecules used for training and the results of evaluating the characteristics of the molecules.

[0042] [B1]

[0043] A molecular design device includes a controller configured to infer the characteristics of a molecule from information about the combination of its structural units using a predetermined predictive model.

[0044] The controller

[0045] Obtain a set of structural unit combination information, which is a collection of structural unit combination information for various different molecules.

[0046] For molecules containing elements that form the set of structural unit combination information, the prediction model is used to calculate the molecule's characteristic prediction values ​​from the structural unit combination information, and the prediction uncertainty is estimated.

[0047] Candidate molecules with the desired characteristics are searched based on the predicted value of this feature and the estimate of the uncertainty of the prediction.

[0048] [B2]

[0049] According to the molecular design device described in B1, the structural unit combination information is sequence information about the molecule.

[0050] [B3]

[0051] According to the molecular design device described in B1 or B2, the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0052] [B4]

[0053] The molecular design apparatus according to any one of B1 to B3, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0054] [B5]

[0055] The molecularly designed device according to any one of B1 to B4, wherein the feature is at least one of binding ability, pharmacological activity, physical properties, kinetics and safety.

[0056] [B6]

[0057] The molecular design apparatus according to any one of B1 to B5, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to the target molecule.

[0058] [B7]

[0059] According to any one of B1 to B6, the molecular design apparatus further calculates a characteristic mass value based on the characteristic prediction value and an estimate of the uncertainty of the prediction, and

[0060] Based on this characteristic quality value, at least one candidate molecule can be identified from the elements of the set of structural unit combination information.

[0061] [B8]

[0062] According to the molecular design apparatus described in B7, the characteristic mass value increases with the increase of the characteristic prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0063] [B9]

[0064] According to the molecular design apparatus described in B8, the characteristic mass value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the predicted characteristic value, and the other variable is an estimate of the uncertainty of the prediction.

[0065] [B10]

[0066] According to any one of B7 to B9, the molecular design device wherein the controller calculates the mean-variance as the objective function to give the characteristic mass value.

[0067] [B11]

[0068] According to any one of B1 to B10, the molecular design apparatus wherein the estimate of the uncertainty of the prediction is the standard deviation of the characteristic prediction value.

[0069] [B12]

[0070] The molecular design apparatus according to any one of B1 to B11, wherein the controller uses a combinatorial optimization algorithm to obtain a set of structural unit combination information and updates the extraction parameters of the combinatorial optimization algorithm based on feature prediction values ​​or feature quality values.

[0071] [B13]

[0072] According to the molecular design device described in B12, the combinatorial optimization algorithm is a tree-structured Parzen estimator.

[0073] [B14]

[0074] The molecular design apparatus according to any one of B1 to B13 further includes an output terminal configured to output information about the combination of structural units on the candidate molecule.

[0075] [B15]

[0076] The molecular design apparatus according to any one of B1 to B14, wherein the prediction model is a prediction model generated by learning based on information about the combination of structural units of a variety of molecules used for training and the results of evaluating the characteristics of the molecules.

[0077] [C1]

[0078] An information processing system for identifying molecules suitable for drug candidates, the information processing system comprising:

[0079] A structural unit combination information processor is configured to obtain a set of structural unit combination information according to a combination optimization algorithm. The set of structural unit combination information is a collection of structural unit combination information about multiple different molecules.

[0080] A feature predictor, configured to use a prediction model to compute feature prediction values ​​for molecules with respect to the elements forming the set of structural unit combination information, the prediction model being used to predict the features of the molecule from the information about the structural unit combination of the molecule; and

[0081] A candidate molecule discriminator is configured to search for candidates of molecules possessing the desired characteristics based on the feature prediction value.

[0082] The structural unit combines an information processor.

[0083] The extraction parameters of the combined optimization algorithm are updated based on the predicted feature values ​​of the molecule to include molecules with more desirable features.

[0084] [C2]

[0085] According to the information processing system described in C1, the candidate molecule discriminator searches for candidates of molecules with desired features based on a first feature prediction value calculated about molecules of elements forming a first set of structural unit combination information and a second feature prediction value calculated about molecules of elements forming a second set of structural unit combination information. The first set of structural unit combination information is a set of structural unit combination information obtained before updating the extraction parameters, and the second set of structural unit combination information is a set of structural unit combination information obtained after updating the extraction parameters.

[0086] [C3]

[0087] According to the information processing system described in C2, the first structural unit combination information set and the second structural unit combination information set are sets that include information about structural unit combinations of different molecules.

[0088] [C4]

[0089] The information processing system according to any one of C1 to C3, wherein the structural unit combination information is sequence information about the molecule.

[0090] [C5]

[0091] The information processing system according to any one of C1 to C4, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0092] [C6]

[0093] The information processing system according to any one of C1 to C5, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the information of the combination of structural units is information of an amino acid sequence.

[0094] [C7]

[0095] The information processing system according to any one of C1 to C6, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0096] [C8]

[0097] The information processing system according to any one of C1 to C7, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to a target molecule.

[0098] [C9]

[0099] The information processing system according to any one of C1 to C8, wherein the feature predictor calculates the feature prediction value and calculates an estimate of the uncertainty of the prediction.

[0100] [C10]

[0101] The information processing system according to any one of C1 to C9, wherein the structural unit combines an information processor.

[0102] The extraction parameters of the combinatorial optimization algorithm are updated based on the predicted feature values ​​of the molecule and the estimated uncertainty of the prediction, so as to include molecules with more desirable features.

[0103] [C11]

[0104] According to the C9 or C10 information processing system, it further includes a predictive information processor configured to process predictive information, which includes feature prediction values ​​and / or estimates of the uncertainty of the prediction, and outputs from the feature predictor.

[0105] The predictive information processor calculates the feature quality value based on the predicted feature value and the estimated uncertainty of the prediction.

[0106] The candidate molecule discriminator identifies at least one molecule from the set of structural unit combination information based on this characteristic quality value.

[0107] [C12]

[0108] According to the information processing system described in C11, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0109] [C13]

[0110] According to the information processing system described in C12, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0111] [C14]

[0112] The information processing system according to any one of C11 to C13, wherein the predictive information processor calculates the mean-variance as an objective function to give characteristic quality values.

[0113] [C15]

[0114] The information processing system according to any one of C9 to C14, wherein the estimate of the uncertainty of the prediction is the standard deviation of the characteristic prediction value.

[0115] [C16]

[0116] The information processing system according to any one of C1 to C15, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm.

[0117] [D1]

[0118] A molecular design device includes a controller configured to infer the characteristics of a molecule from information about the combination of its structural units using a predetermined predictive model.

[0119] The controller

[0120] A set of structural unit combination information is obtained using a combinatorial optimization algorithm. This set of structural unit combination information is a collection of information about the combinations of structural units in various different molecules.

[0121] For the molecule of the elements that form the set of information about this structural unit combination, the prediction model is used to calculate the predicted feature values.

[0122] The extraction parameters of the combined optimization algorithm are updated using this prediction model based on the feature prediction values ​​of each molecule, in order to include molecules with more desired features.

[0123] Based on the predicted value of this feature, search for candidates of molecules with the desired features.

[0124] [D2]

[0125] According to the molecular design device described in D1, wherein the controller

[0126] Before updating the extraction parameters, obtain the first set of structural unit combination information, which is a collection of information about the combination of structural units of various molecules.

[0127] A prediction model is used to calculate the predicted value of the first feature relative to the molecule of the elements that form the information set of the first structural unit combination.

[0128] After updating the extraction parameters, a second set of structural unit combination information is further obtained, which serves as a collection of information about the combination of structural units in various different molecules.

[0129] The predicted value of the second feature is calculated using a prediction model relative to the molecule of the elements that form the information set of the second structural unit combination.

[0130] Candidate molecules with the desired characteristics are searched based on the first feature prediction value and the second feature prediction value.

[0131] [D3]

[0132] According to the molecular design device described in D2, the first structural unit combination information set and the second structural unit combination information set are sets including sequence information about different molecules.

[0133] [D4]

[0134] The molecular design apparatus according to any one of D1 to D3, wherein the structural unit combination information is sequence information about the molecule.

[0135] [D5]

[0136] The molecular design apparatus according to any one of D1 to D5, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0137] [D6]

[0138] The molecular design apparatus according to any one of D1 to D5, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0139] [D7]

[0140] The molecularly designed device according to any one of D1 to D6, wherein the feature is at least one of binding ability, pharmacological activity, physical properties, kinetics and safety.

[0141] [D8]

[0142] The molecular design apparatus according to any one of D1 to D7, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to the target molecule.

[0143] [D9]

[0144] According to any one of D1 to D8, the molecular design apparatus wherein the controller calculates a characteristic prediction value and calculates an estimate of the uncertainty of the prediction.

[0145] [D10]

[0146] According to any one of D1 to D9, the molecular design apparatus wherein the controller acquires a set of structural unit combination information based on the molecule’s characteristic predictions and estimates of the uncertainty of the predictions, so as to include molecules with more desirable characteristics.

[0147] [D11]

[0148] According to the molecular design apparatus described in D10, the controller further calculates the characteristic mass value based on the characteristic prediction value and an estimate of the uncertainty of the prediction, and

[0149] The candidate molecule discriminator identifies at least one molecule from the set of structural unit combination information based on this characteristic quality value.

[0150] [D12]

[0151] According to the molecular design apparatus described in D11, the characteristic mass value increases with the increase of the characteristic prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0152] [D13]

[0153] According to the molecular design apparatus described in D12, the characteristic mass value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the predicted characteristic value, and the other variable is an estimate of the uncertainty of the prediction.

[0154] [D14]

[0155] The molecular design apparatus according to any one of D11 to D13, wherein the controller calculates the mean-variance as the objective function to give characteristic mass values.

[0156] [D15]

[0157] The molecular design apparatus according to any one of D9 to D14, wherein the estimate of the uncertainty of the prediction is the standard deviation of the characteristic prediction value.

[0158] [D16]

[0159] Molecular design apparatus according to any one of D1 to D15, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm.

[0160] [D17]

[0161] The molecular design apparatus according to any one of D1 to D16, wherein the prediction model is a model learned based on training data, which represents information about the combination of structural units of the molecule and the characteristics of the molecule.

[0162] [E1]

[0163] An information processing method in an information processing system for identifying molecules suitable for drug candidates includes the following steps: for a molecule whose elements form a set of structural unit combination information, using a prediction model to calculate a predicted feature value for the molecule, and then estimating the uncertainty of the prediction, wherein the set of structural unit combination information is a set of structural unit combination information for a variety of different molecules, and the prediction model is used to predict the features of the molecule from the structural unit combination information of the molecule; and

[0164] Candidates for molecules with more desirable characteristics are searched based on the predicted value of this feature and the uncertainty of the prediction.

[0165] [E2]

[0166] According to the information processing method described in E1, the structural unit combination information is information about the sequence of the molecule.

[0167] [E3]

[0168] According to the information processing method described in E1 or E2, the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody, and low molecular weight compound.

[0169] [E4]

[0170] The information processing method according to any one of E1 to E3, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0171] [E5]

[0172] The information processing method according to any one of E1 to E4, wherein the feature is at least one of binding ability, pharmacological activity, physical properties, kinetics and safety.

[0173] [E6]

[0174] The information processing method according to any one of E1 to E5, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to the target molecule.

[0175] [E7]

[0176] The information processing method according to any one of E1 to E6 further includes the step of calculating a feature quality value based on the feature prediction value and an estimate of the uncertainty of the prediction.

[0177] The search steps include identifying at least one molecule from the set of structural unit combination information based on the feature quality value.

[0178] [E8]

[0179] According to the information processing method described in E7, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0180] [E9]

[0181] According to the information processing method described in E8, the feature quality value is the output value of an arbitrary function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0182] [E10]

[0183] According to any one of E7 to E9, the step of calculating the feature quality value includes calculating the mean and variance as the objective function to give the feature quality value.

[0184] [E11]

[0185] The information processing method according to any one of E1 to E10, wherein the estimate of the uncertainty of the prediction is the standard deviation of the prediction value.

[0186] [E12]

[0187] The information processing method according to any one of E1 to E11 includes the step of outputting information about the combination of structural units of the identified candidate molecules.

[0188] [E13]

[0189] According to the information processing method described in E12, it includes further outputting information about the predicted characteristic values ​​of the identified molecules.

[0190] [E14]

[0191] The information processing method according to any one of E7 to E13, wherein the search step includes identifying molecules from the set of structural unit combination information whose characteristic mass values ​​are equal to or greater than a predetermined value.

[0192] [E15]

[0193] The information processing method according to any one of E7 to E13, wherein the search step includes identifying the ranking of the characteristic quality values ​​of each molecule and identifying molecules at or above a predetermined ranking.

[0194] [E16]

[0195] The information processing method according to any one of E1 to E15, wherein the search step includes the step of selecting at least one molecule from the set of structural unit combination information wherein the feature prediction value and the prediction uncertainty respectively satisfy predetermined conditions.

[0196] [E17]

[0197] The information processing method according to any one of E1 to E16 includes using a combinatorial optimization algorithm to obtain a set of structural unit combination information, and updating the extraction parameters of the combinatorial optimization algorithm based on feature prediction values ​​or feature quality values.

[0198] [E18]

[0199] According to the information processing method described in E17, the combinatorial optimization algorithm is a tree-structured Parzen estimator.

[0200] [E19]

[0201] The information processing method according to any one of E1 to E18, wherein the prediction model is a model learned based on training data, which represents information about the combination of structural units of a molecule and the characteristics of the molecule.

[0202] [F1]

[0203] A program configured to cause a computer to execute:

[0204] Regarding a molecule that forms an element of a set of structural unit combination information that is a set of information about the combination of structural units of various different molecules, a predictive model is used to calculate the predicted characteristic value of the molecule, followed by an estimation of the uncertainty of the prediction. This predictive model is used to predict the characteristics of the molecule from the information about the combination of structural units of the molecule; and

[0205] The process of searching for candidates of molecules with more desirable features based on the predicted value of this feature and the uncertainty of the prediction.

[0206] [F2]

[0207] According to the procedure described in F1, the structural unit combination information is the sequence information about the molecule.

[0208] [F3]

[0209] The procedure is as described in F1 or F2, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody, and low molecular weight compound.

[0210] [F4]

[0211] The procedure according to any one of F1 to F3, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0212] [F5]

[0213] The procedure according to any one of F1 to F4, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0214] [F6]

[0215] The procedure according to any one of F1 to F5, wherein the molecule is a molecule that binds to the target molecule, and the feature is the ability to bind to the target molecule.

[0216] [F7]

[0217] The procedure according to any one of F1 to F6 further includes a process for calculating a feature quality value based on the feature prediction value and an estimate of the uncertainty of the prediction.

[0218] The search process includes identifying at least one molecule from the set of structural unit combination information based on the feature quality value.

[0219] [F8]

[0220] According to the procedure described in F7, the feature quality value increases as the feature prediction value increases and decreases as the estimated value of the prediction uncertainty decreases.

[0221] [F9]

[0222] According to the procedure described in F8, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the predicted feature value, and the other variable is an estimate of the uncertainty of the prediction.

[0223] [F10]

[0224] According to any one of F7 to F9, the process of calculating the characteristic quality value includes calculating the mean and variance as the objective function to give the characteristic quality value.

[0225] [F11]

[0226] The procedure is according to any one of F1 to F10, wherein the estimate of the uncertainty is the standard deviation of the predicted value.

[0227] [F12]

[0228] The procedure according to any one of F1 to F11 includes a process of outputting information about the combination of structural units of a candidate molecule, which is identified by searching for candidates of molecules with more desirable characteristics.

[0229] [F13]

[0230] According to the procedure described in F12, it includes further outputting information about the predicted feature values ​​of candidate molecules, which are identified by searching for candidates of molecules with more desired features.

[0231] [F14]

[0232] The procedure according to any one of F7 to F13, wherein the search process includes identifying molecules from the set of structural unit combination information whose characteristic mass values ​​are equal to or greater than a predetermined value.

[0233] [F15]

[0234] The procedure according to any one of F7 to F13, wherein the search process includes identifying the ranking of the characteristic quality values ​​of each molecule and identifying molecules at or above a predetermined ranking.

[0235] [F16]

[0236] According to any one of F1 to F15, the search process includes selecting from the set of structural unit combination information at least one molecule in which the predicted value of the feature and the estimate of the uncertainty of the prediction respectively satisfy predetermined conditions.

[0237] [F17]

[0238] The procedure according to any one of F1 to F16 further includes a process of using a combinatorial optimization algorithm to update the extraction parameters for obtaining the set of structural unit combinatorial information based on feature prediction values ​​or feature quality values.

[0239] [F18]

[0240] According to the procedure described in F17, the combinatorial optimization algorithm is a tree-structured Parzen estimator.

[0241] [F19]

[0242] The procedure according to any one of F1 to F18, wherein the prediction model is a model learned based on training data, which represents information about the combination of structural units of the molecule and the characteristics of the molecule.

[0243] [G1]

[0244] A method for producing a molecular compound, the method comprising:

[0245] The input step involves accessing a set of structural unit combination information, which is a collection of structural unit combination information for various different molecules.

[0246] The information set of the structural unit combination is input into the prediction model;

[0247] The inference step involves searching for molecules with more desirable characteristics from the set of structural unit combinations based on the predicted feature values ​​and the estimates of the uncertainty of the predictions for each molecule included in the set of structural unit combination information, and output from the prediction model, and then identifying them as candidate molecules.

[0248] The output step outputs information about the structural unit combination of the candidate molecule; and

[0249] The generation step is to generate a molecular compound having a molecular sequence represented by the combination information of the structural units.

[0250] [G2]

[0251] According to the method described in G1, the candidate molecule has a biological sequence, and the structural unit combination information is sequence information.

[0252] [G3]

[0253] The method according to G1 or G2, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody, and low molecular weight compound.

[0254] [G4]

[0255] The method according to any one of G1 to G3, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0256] [G5]

[0257] The method according to any one of G1 to G4, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0258] [G6]

[0259] The method according to any one of G1 to G5, wherein the molecule is a molecule that binds to the target molecule, and the feature is the ability to bind to the target molecule.

[0260] [G7]

[0261] The method according to any one of G1 to G6 includes the following steps: calculating a feature quality value based on a feature prediction value and an estimate of the uncertainty of the prediction, and identifying at least one candidate molecule from a set of structural unit combination information based on the feature quality value.

[0262] [G8]

[0263] According to the method described in G7, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimate of the prediction uncertainty.

[0264] [G9]

[0265] According to the method described in G8, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0266] [G10]

[0267] According to any one of G7 to G9, the step of calculating the characteristic quality value includes calculating the mean and variance as the objective function to give the characteristic quality value.

[0268] [G11]

[0269] According to any one of G1 to G10, the estimate of the uncertainty is the standard deviation of the predicted value.

[0270] [G12]

[0271] According to any one of G1 to G11, the information about the candidate molecule includes information about the combination of structural units of the candidate molecule.

[0272] [G13]

[0273] The method according to any one of G1 to G12, wherein the information about the candidate molecule includes the predicted characteristic values ​​of the candidate molecule.

[0274] [G14]

[0275] According to any one of G7 to G13, the step of searching for candidate molecules includes identifying molecules from the set of structural unit combination information where the characteristic mass value is equal to or greater than a predetermined value as the candidate molecule.

[0276] [G15]

[0277] According to any one of G7 to G13, the step of searching for candidate molecules includes identifying the ranking of the characteristic quality values ​​of each molecule and identifying molecules at or above a predetermined ranking as the candidate molecule.

[0278] [G16]

[0279] According to any one of G1 to G15, the step of searching for candidate molecules includes selecting at least one molecule from the set of structural unit combination information in which the predicted value of features and the estimate of the uncertainty of the prediction respectively satisfy predetermined conditions.

[0280] [G17]

[0281] According to any one of G1 to G16, the method uses a combinatorial optimization algorithm to obtain a set of structural unit combinatorial information, and

[0282] The method further includes a process of updating the extraction parameters of the combined optimization algorithm based on the molecule's feature prediction value and feature quality value.

[0283] [G18]

[0284] According to the method described in G17, the combinatorial optimization algorithm is a tree-structured Parzen estimator.

[0285] [G19]

[0286] The method according to any one of G1 to G18 includes a selection step, which is to select a candidate molecule with the best experimental value among experimental values ​​having the characteristics of a variety of candidate molecules.

[0287] [H1]

[0288] An information processing method in an information processing system for identifying molecules suitable for drug candidates, the method comprising the following steps:

[0289] The combinatorial optimization algorithm is used to obtain the set of structural unit combination information, which is a set of structural unit combination information about various molecules.

[0290] For molecules of elements that form the set of structural unit combination information, a prediction model is used to calculate the feature prediction value of the molecule. This prediction model is used to predict the features of the molecule from the information about the combination of structural unit elements of the molecule.

[0291] Based on this feature prediction value, search for molecules with more desired features; and

[0292] The extraction parameters of the combined optimization algorithm are updated based on the predicted feature values ​​of the molecule to include molecules with more desirable features.

[0293] [H2]

[0294] According to the information processing method described in H1, it includes the following steps: searching for candidates of molecules with desired features based on a first feature prediction value calculated about molecules of elements forming a first structural unit combination information set and a second feature prediction value calculated about molecules of elements forming a second structural unit combination information set, wherein the first structural unit combination information set is a structural unit combination information set obtained before updating the extraction parameters, and the second structural unit combination information set is a structural unit combination information set obtained after updating the extraction parameters.

[0295] [H3]

[0296] According to the information processing method described in H2, the first structural unit combination information set and the second structural unit combination information set are sets that include information about structural unit combinations of different molecules.

[0297] [H4]

[0298] The information processing method according to any one of H1 to H3, wherein the structural unit combination information is information about the sequence of the molecule.

[0299] [H5]

[0300] The information processing method according to any one of H1 to H4, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0301] [H6]

[0302] The information processing method according to any one of H1 to H5, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0303] [H7]

[0304] According to the information processing method described in H6, the feature is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety.

[0305] [H8]

[0306] The information processing method according to any one of H1 to H7, wherein the molecule is a molecule that binds to a target molecule, and the feature is the ability to bind to the target molecule.

[0307] [H9]

[0308] The information processing method according to any one of H1 to H8 includes the steps of calculating the predicted value of the feature and the estimate of the uncertainty of the prediction.

[0309] [H10]

[0310] According to the information processing method described in H9, it includes the step of calculating a feature quality value based on the feature prediction value and an estimate of the uncertainty of the prediction.

[0311] The search steps include identifying at least one molecule from the set of structural unit combination information based on the feature quality value.

[0312] [H11]

[0313] According to the information processing method described in H10, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0314] [H12]

[0315] According to the information processing method described in H11, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0316] [H13]

[0317] According to any one of H10 to H12, the step of calculating the feature quality value includes the step of calculating the mean and variance as the objective function to give the feature quality value.

[0318] [H14]

[0319] The information processing method according to any one of H9 to H13, wherein the estimate of the uncertainty of the prediction is the standard deviation of the predicted feature value.

[0320] [H15]

[0321] The information processing method according to any one of H1 to H14, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm associated with the update of the extracted parameters.

[0322] [I1]

[0323] A program configured to cause a computer to execute:

[0324] The acquisition process involves obtaining a set of structural unit combination information based on a combinatorial optimization algorithm. This set of structural unit combination information is a collection of information about the combination of structural units of various molecules.

[0325] The prediction process involves using a prediction model to calculate the predicted features of the molecules that form the set of information about the combination of structural units. This prediction model is used to predict the features of the molecule from information about the combination of structural units.

[0326] The update process involves updating the extraction parameters of the combined optimization algorithm based on the feature prediction values ​​of each molecule to include molecules with more desired features; and

[0327] The search process involves searching for molecules with the desired characteristics based on the predicted feature value.

[0328] [I2]

[0329] According to the procedure described in I1, the acquisition process includes:

[0330] The process of obtaining a first set of structural unit combination information, which is a collection of information about the combination of structural units of various molecules, before updating the extraction parameters; and

[0331] After updating the extraction parameters, the process of further obtaining a second set of structural unit combination information, which is a collection of information about the combination of structural units of various different molecules, is described.

[0332] The prediction process includes:

[0333] The process of using a prediction model to calculate the predicted value of a first feature relative to the molecule of the elements forming the information set of the first structural unit combination; and

[0334] The process of using a prediction model to calculate the predicted value of the second feature, relative to the molecule of the elements that form the information set of the second structural unit combination,

[0335] The search process includes

[0336] The process of searching for candidates of molecules with desired features from the first and second sets of structural unit combination information based on the first and second feature prediction values.

[0337] [I3]

[0338] According to the procedure described in I2, the first set of structural unit combination information and the second set of structural unit combination information are sets that include information about the combination of structural units of different molecules.

[0339] [I4]

[0340] The procedure according to any one of I1 to I3, wherein the structural unit combination information is information about the sequence of the molecule.

[0341] [I5]

[0342] The procedure according to any one of I1 to I4, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0343] [I6]

[0344] The procedure according to any one of I1 to I5, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0345] [I7]

[0346] The procedure according to any one of I1 to I6, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0347] [I8]

[0348] The procedure according to any one of I1 to I7, wherein the molecule is a molecule that binds to the target molecule, and the feature is the ability to bind to the target molecule.

[0349] [I9]

[0350] The procedure according to any one of I1 to I8 includes the process of calculating the predicted value of the feature and the process of calculating an estimate of the uncertainty of the prediction.

[0351] [I10]

[0352] According to the procedure described in I9, it further includes a process for calculating a feature quality value based on the feature prediction value and an estimate of the uncertainty of the prediction.

[0353] The search process includes

[0354] The process of identifying at least one molecule from a set of structural unit combination information based on this characteristic quality value.

[0355] [I11]

[0356] According to the procedure described in I10, the feature quality value increases as the feature prediction value increases and decreases as the estimated value of the prediction uncertainty decreases.

[0357] [I12]

[0358] According to the procedure described in I11, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0359] [I13]

[0360] The procedure according to any one of I10 to I12, wherein the process of calculating the characteristic quality value includes calculating the mean and variance as the objective function to give the characteristic quality value.

[0361] [I14]

[0362] The procedure according to any one of I9 to I13, wherein the estimate of the uncertainty of the prediction is the standard deviation of the characteristic prediction value.

[0363] [I15]

[0364] The procedure according to any one of I1 to I14, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm associated with the update of the set of structural unit combination information.

[0365] [I16]

[0366] The procedure according to any one of I1 to I15, wherein the prediction model is a model learned based on training data, which represents information about the combination of structural units of the molecule and the characteristics of the molecule.

[0367] [J1]

[0368] A method for producing a molecular compound, the method comprising:

[0369] The acquisition step involves obtaining a set of structural unit combination information based on a combinatorial optimization algorithm. This set of structural unit combination information is a collection of structural unit combination information for various different molecules.

[0370] The input step involves inputting the set of information about the structural unit combination into the prediction model.

[0371] The update step involves updating the extraction parameters of the combinatorial optimization algorithm based on the feature prediction values ​​of each molecule included in the set of structural unit combination information and output from the prediction model, so as to include molecules with more desirable features.

[0372] Based on the predicted value of this feature, the steps of searching for molecules with more desired features from the set of structural unit combination information and then identifying them as candidate molecules are performed.

[0373] The output step outputs information about the structural unit combination of the candidate molecule; and

[0374] The generation step is to generate the molecular compound having a molecular sequence represented by the combination information of the structural units.

[0375] [J2]

[0376] According to the method described in J1, it includes the following steps: searching for candidate molecules based on a first feature prediction value calculated about molecules of elements forming a first set of structural unit combination information and a second feature prediction value calculated about molecules of elements forming a second set of structural unit combination information, wherein the first set of structural unit combination information is a set of structural unit combination information obtained before updating the extraction parameters, and the second set of structural unit combination information is a set of structural unit combination information obtained after updating the extraction parameters.

[0377] [J3]

[0378] According to the method described in J2, the first set of structural unit combination information and the second set of structural unit combination information are sets that include information about the combination of structural units of different molecules.

[0379] [J4]

[0380] The method according to any one of J1 to J3, wherein the candidate molecule has a biological sequence.

[0381] [J5]

[0382] According to any one of J1 to J4, the structural unit combination information is sequence information about the molecule.

[0383] [J6]

[0384] The method according to any one of J1 to J5, wherein the molecule is at least one of nucleic acid, peptide, cyclic peptide, protein, antibody and low molecular weight compound.

[0385] [J7]

[0386] The method according to any one of J1 to J6, wherein the molecule is a protein, antibody, peptide or cyclic peptide, and the structural unit combination information is information about the amino acid sequence.

[0387] [J8]

[0388] The method according to any one of J1 to J7, wherein the feature is at least one of binding capacity, pharmacological activity, physical properties, kinetics and safety.

[0389] [J9]

[0390] The method according to any one of J1 to J8, wherein the molecule is a molecule that binds to the target molecule, and the feature is the ability to bind to the target molecule.

[0391] [J10]

[0392] The method according to any one of J1 to J9 includes the steps of calculating the predicted value of the feature and the estimate of the uncertainty of the prediction.

[0393] [J11]

[0394] According to the method described in J10, it includes the step of calculating a feature quality value based on the feature prediction value and an estimate of the uncertainty of the prediction.

[0395] The inference step of searching for candidate molecules includes the step of identifying at least one molecule from the set of structural unit combination information based on the characteristic quality value.

[0396] [J12]

[0397] According to the method described in J11, the feature quality value increases with the increase of the feature prediction value and decreases with the decrease of the estimated value of the prediction uncertainty.

[0398] [J13]

[0399] According to the method described in J12, the feature quality value is the output value of a predetermined function, in which two variables are used as inputs: one variable is the feature prediction value, and the other variable is an estimate of the uncertainty of the prediction.

[0400] [J14]

[0401] According to any one of J11 to J13, the step of calculating the characteristic quality value includes calculating the mean and variance as the objective function to give the characteristic quality value.

[0402] [J15]

[0403] The method according to any one of J10 to J14, wherein the estimate of the uncertainty of the prediction is the standard deviation of the characteristic prediction value.

[0404] [J16]

[0405] The method according to any one of J1 to J15, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm associated with the update of the set of structural unit combination information.

[0406] [J17]

[0407] The method according to any one of J1 to J16 includes the step of selecting a candidate molecule with the best experimental value among experimental values ​​having the characteristics of a variety of candidate molecules.

[0408] [K]

[0409] A computer system including a processor and memory, wherein the memory is configured to store one or more commands, and

[0410] This command causes the processor to:

[0411] For a molecule that forms an element of a set of structural unit combination information as a set of information about the combination of structural units of various different molecules, a prediction model is used to calculate the predicted characteristic value of the molecule, and then the uncertainty of the prediction is estimated. This prediction model is used to predict the characteristics of the molecule from the information about the combination of structural units of the molecule; and

[0412] Candidates for molecules with more desirable characteristics are searched based on the predicted value of this feature and the uncertainty of the prediction.

[0413] [L]

[0414] A non-transitory computer-readable storage medium for storing one or more commands.

[0415] This command causes the computer to:

[0416] For a molecule that forms an element of a set of structural unit combination information as a set of information about the combination of structural units of various different molecules, a prediction model is used to calculate the predicted characteristic value of the molecule, and then the uncertainty of the prediction is estimated. This prediction model is used to predict the characteristics of the molecule from the information about the combination of structural units of the molecule; and

[0417] Based on the predicted value of this feature and the uncertainty of the prediction, candidate molecules with more desirable features are searched.

[0418] [M]

[0419] A computer system including a processor and memory, wherein the memory is configured to store one or more commands, and

[0420] This command causes the processor to:

[0421] The combinatorial optimization algorithm is used to obtain the set of structural unit combination information, which is a set of structural unit combination information about various molecules.

[0422] A predictive model is used to calculate the predicted features of the molecules that form the set of information about the combination of structural units. This predictive model is used to predict the features of the molecule from information about the combination of structural units of the molecule.

[0423] The extraction parameters of this combined optimization algorithm are updated based on the feature prediction values ​​of each molecule to include molecules with more desired features; and

[0424] Based on the predicted value of this feature, search for molecules with the desired feature.

[0425] [N]

[0426] A non-transitory computer-readable storage medium for storing one or more commands.

[0427] This command causes the computer to:

[0428] The combinatorial optimization algorithm is used to obtain the set of structural unit combination information, which is a set of structural unit combination information about various molecules.

[0429] A predictive model is used to calculate the predicted features of the molecules that form the set of information about the combination of structural units. This predictive model is used to predict the features of the molecule from information about the combination of structural units of the molecule.

[0430] The extraction parameters of this combined optimization algorithm are updated based on the feature prediction values ​​of each molecule to include molecules with more desired features; and

[0431] Based on the predicted value of this feature, search for molecules with the desired feature.

[0432] [Beneficial effects of the invention]

[0433] According to one aspect of this disclosure, the burden of accompanying drug discovery can be further reduced. Attached Figure Description

[0434] [Figure 1] Figure 1 shows a schematic diagram illustrating an example of a drug discovery system including a molecular design device according to a first embodiment.

[0435] [Figure 2] Figure 2 shows a block diagram illustrating an example of a drug discovery system including a molecular design device according to a second embodiment.

[0436] [Figure 3] Figure 3 illustrates examples of sequence information for various molecules used to describe embodiments of this application.

[0437] [Figure 4] Figure 4 is a schematic diagram illustrating an example of the hardware configuration of the molecular design device in an embodiment of this application.

[0438] [Figure 5] Figure 5 is a schematic diagram illustrating an example of the configuration of the controller in an embodiment of this application.

[0439] [Figure 6] Figure 6 shows a flowchart illustrating an example of the processing flow performed by the molecular design device in the first embodiment.

[0440] [Figure 7] Figure 7 shows a flowchart illustrating an example of the processing flow performed by the molecular design device in the second embodiment.

[0441] [Figure 8] Figure 8 shows a flowchart illustrating an example of combined optimization processing in the second embodiment.

[0442] [Figure 9] Figure 9 shows a flowchart illustrating another example of the combined optimization process in the second embodiment.

[0443] [Figure 10] Figure 10 is a schematic diagram showing the configuration of the optimization process for an application example according to an embodiment of the present application.

[0444] [Figure 11] Figure 11 shows a schematic diagram illustrating an instance of calculating the objective function value by optimization using TPE based on the first verification instance.

[0445] [Figure 12] Figure 12 shows a schematic diagram illustrating the relationship between the predicted mean and the predicted standard deviation of the sequence obtained by sampling with TPE according to the first validation example.

[0446] [Figure 13] Figure 13 shows a schematic diagram illustrating the density distribution of edit distances of sequences obtained by sampling with TPE according to the first verification instance.

[0447] [Figure 14] Figure 14 shows a schematic diagram illustrating the distribution of pseudo-correct solution model scores for the proposed sequence based on the first verification instance.

[0448] [Figure 15] Figure 15 shows a schematic diagram illustrating the edit distance of the proposed sequence based on the first verification instance.

[0449] [Figure 16] Figure 16 shows a schematic diagram illustrating the predicted standard deviation values ​​of the proposed sequence based on the first validation example.

[0450] [Figure 17] Figure 17 shows a table illustrating examples of candidates for amino acids at the corresponding mutant candidate positions in the search space according to the first validation example.

[0451] [Figure 18] Figure 18 shows a table illustrating an example of parameter configuration for the TPE based on the first verification instance.

[0452] [Figure 19] Figure 19 shows a schematic diagram illustrating the distribution of the predicted mean and predicted standard deviation of the sequence obtained by sampling with TPE according to the second verification example.

[0453] [Figure 20] Figure 20 shows a schematic diagram of t-SNE visualization of sequences sampled using TPE according to the second verification instance.

[0454] [Figure 21] Figure 21 shows a schematic diagram illustrating the predicted mean of the proposed sequence based on the second verification instance.

[0455] [Figure 22] Figure 22 shows a schematic diagram illustrating the predicted variance of the proposed sequence based on the second validation example.

[0456] [Figure 23] Figure 23 shows a schematic diagram illustrating the distribution of expression levels of the proposed sequence according to the second verification instance.

[0457] [Figure 24] Figure 24 shows a schematic diagram illustrating the distribution of binary octet values ​​according to the corresponding sequence of the second verification instance. Detailed Implementation

[0458] I. Definition

[0459] In this specification, the term “and / or” is used to indicate the items described before and after “and / or”, or any combination thereof. For example, “A, B and / or C” includes the subjects “A”, “B” and “C”, as well as the combinations “A and B”, “A and C”, “B and C” and “A and B and C”.

[0460] - amino acids

[0461] This specification may include both natural and non-natural amino acids. In the case of natural amino acids, the amino acid is indicated by a single-letter code, a three-letter code, or both, such as Ala / A, Leu / L, Arg / R, Lys / K, Asn / N, Met / M, Asp / D, Phe / F, Cys / C, Pro / P, Gln / Q, Ser / S, Glu / E, Thr / T, Gly / G, Trp / W, His / H, Tyr / Y, Ile / I, or Val / V.

[0462] - Changes in amino acids

[0463] Known methods, such as site-specific mutagenesis (Kunkel et al. (Proc. Natl. Acad. Sci. USA (1985) 82, 488-492)) and overlap-amplified PCR, can be suitably used to alter the amino acid sequence of antigen-binding molecules. Various methods known in the art can also be employed as alteration methods to replace amino acids with those other than natural amino acids (Annu. Rev. Biophys. Biomol. Struct. (2006) 35, 225-249; and Proc. Natl. Acad. Sci. USA (2003) 100 (11), 6353-6357). For example, cell-free translation systems including tRNA (Clover Direct (ProteinExpress Co., Ltd.)) can also be used, in which non-natural amino acids are bound to a succinate inhibitor tRNA complementary to one of the UAG codons (succinate codons) in the stop codon.

[0464] - Antigen

[0465] In this article, the structure of "antigen" is not limited to a specific structure, as long as the structure contains an epitope to which the antigen-binding domain binds. In some models, the antigen is a peptide, polypeptide, or protein having four or more amino acids.

[0466] Examples of antigens include membrane-associated molecules expressed on the cell membrane and soluble molecules secreted by the cell extracellularly.

[0467] - Antigen-binding domain

[0468] In this paper, any structural domain that binds to the antigen of interest can be used as an "antibody-binding domain". Examples of such domains include variable regions of the heavy and light chains of antibodies; modules called A domains contained within affinity multimers (Avimers), which are cell membrane proteins present in living organisms and composed of approximately 35 amino acids (International Publication Nos. WO 2004 / 044011, WO 2005 / 040229); Adnectin, which contains a 10Fn3 domain that binds to proteins in fibronectin, glycoproteins expressed on the cell membrane (International Publication No. WO 2002 / 032925); Affibody, which uses an IgG-binding domain as a scaffold, forming three helical bundles, each consisting of protein A composed of 58 amino acids (International Publication No. WO 1995 / 001937); and DARPin, a region in which ankyrin repeat sequences (ARs) are exposed on the molecular surface, each containing 33 amino acids. The structure comprises an amino acid residue having a structure including a repetitively stacked rotating subunit, two antiparallel helices, and a loop (International Publication No. WO 2002 / 020565); anticalin, which is a four-loop region on one side of a supporting barrel structure in which eight highly conserved antiparallel chains of lipocalin (such as neutrophil gelatinase-associated lipocalin (NGAL)) twist towards the center (International Publication No. WO2003 / 029462); and a concave region of parallel sheet-like structures within the horseshoe-shaped structure, including repetitively stacked leucine-rich repeat sequence (LRR) modules of variable lymphocyte receptors (VLRs) that do not have the immunoglobulin structure of the acquired immune system of jawless animals (such as lampreys and hagfish) (International Publication No. WO 2008 / 016854).

[0469] Examples of antigen-binding domains in this disclosure include antigen-binding domains containing variable regions of both the heavy and light chains of an antibody. Examples of such antigen-binding domains include “scFv (single-chain Fv)”, “single-chain antibody”, “Fv”, “scFv2 (single-chain Fv2)”, “Fab”, or “F(ab')2”.

[0470] - Antigen-binding molecules

[0471] In this disclosure, the term "antigen-binding molecule" is used in the broadest sense, including various types of molecules containing an antigen-binding domain. An antigen-binding molecule can be a molecule consisting solely of an antigen-binding domain, or a molecule containing an antigen-binding domain and another domain. If the antigen-binding molecule is a molecule formed by an antigen-binding domain and an Fc region bound together, examples of such molecules include complete antibodies and antibody fragments. Antibodies can include monoclonal antibodies (including agonist and antagonist antibodies), human antibodies, humanized antibodies, and chimeric antibodies. The term "antigen-binding molecule" in this disclosure includes scaffold molecules formed by using only a portion of a three-dimensional structure serving as a scaffold (such as existing stable αβ / barrel protein structures) as a library for constructing antigen-binding domains.

[0472] In this specification, the term "antibody" is used in the broadest sense and covers a variety of antibody structures, including but not limited to monoclonal antibodies, polyclonal antibodies, multispecific antibodies (e.g., bispecific antibodies), and antibody fragments, as long as the antibody exhibits the desired antigen-binding activity.

[0473] Antibodies can be isolated from natural resources such as plasma and serum in which the antibody is naturally present, or from the culture supernatant of hybridoma cells that produce the antibody, or can be synthesized partially or completely using techniques such as gene recombination. Examples of antibodies include isotypes of immunoglobulins and subtypes of these isotypes. There are nine classes (isotypes) of human immunoglobulins: IgG1, IgG2, IgG3, IgG4, IgA1, IgA2, IgD, IgE, and IgM. Among these isotypes, the term "antibody" in this disclosure includes IgG1, IgG2, IgG3, and IgG4. Although multiple allotype sequences due to genetic polymorphism are described in the constant regions of human IgG1, human IgG2, human IgG3, and human IgG4 in Sequences of proteins of immunological interest, NIH Publication No. 91-3242, any of them are acceptable in this disclosure. In particular, in the case of human IgG1 sequences, the amino acid sequence specified by the EU number as ranging from position 356 to position 358 can be DEL or EEM. Although multiple allomorphic sequences due to gene polymorphism are described in the human Igκ (Kappa) constant region and the human Igλ (Lambda) constant region in Sequences of proteins of immunological interest, NIH Publication No. 91-3242, either of them is acceptable in this disclosure.

[0474] "Antibody fragment" refers to a molecule other than the complete antibody that contains a portion of the complete antibody and binds to the antigen to which the complete antibody binds. Examples of antibody fragments include, but are not limited to, Fv, Fab, Fab', Fab'-SH, F(ab')2, bisomatic antibodies, linear antibodies, single-chain antibody molecules (e.g., scFv), and multispecific antibodies formed from antibody fragments.

[0475] The terms “full-length antibody,” “intact antibody,” and “all antibody” are used interchangeably in this specification and refer to antibodies having a structure substantially similar to that of natural antibodies or having a heavy chain containing an Fc region as defined in this specification.

[0476] The term "variable region" or "variable domain" refers to a domain of the heavy or light chain of an antibody that participates in antibody-antigen binding. The variable domains (VH and VL) of the heavy and light chains of natural antibodies normally have similar structures, including a frame region (FR) comprising four conserved domains and three hypervariable regions (HVR) (e.g., see Kindt et al., Kuby Immunology, 6th ed., WH Freeman and Co., p. 91 (2007)). One VH or VL domain is sufficient to confer antigen-binding specificity. Furthermore, antibodies binding to certain antigens can be isolated using the VH or VL domain of the antibody bound to that antigen by screening for libraries complementary to that VL or VH domain. For example, see Portolano et al., J. Immunol. 150:880-887 (1993); Clarkson et al., Nature 352:624-628 (1991).

[0477] As used herein, the term "molecular weight" refers to the sum of the atomic weights of the atoms that make up a compound molecule (unit: "g / mol"), and is obtained by calculating the sum of the atomic weights of the atoms contained in the molecular structure. The unit of molecular weight is sometimes omitted in this text. Molecular weight can be measured, for example, by liquid chromatography-mass spectrometry (LC / MS).

[0478] As used herein, a medium-molecular-weight compound or a medium-molecular-weight compound is a compound having a molecular weight of 500 g / mol or greater and less than 30,000 g / mol.

[0479] Medium-molecular-weight compounds are, for example, compounds having a molecular weight of 500 g / mol or greater and less than 6,000 g / mol, and may also be compounds having a molecular weight of 500 g / mol or greater and less than 4,000 g / mol, compounds having a molecular weight of 600 g / mol or greater and less than 4,000 g / mol, or compounds having a molecular weight of 700 g / mol or greater and less than 3,000 g / mol. Medium-molecular-weight compounds are peptide compounds containing peptide chains, nucleic acid chains, or sugar chains, and may be peptide compounds containing 5 to 30 amino acid residues, peptide compounds containing 7 to 25 amino acid residues, or peptide compounds containing 9 to 20 amino acid residues. The medium molecular weight compound is, for example, a peptide compound having a molecular weight of 500 g / mol or greater and less than 6,000 g / mol, and may be a peptide compound having a molecular weight of 500 g / mol or greater and less than 4,000 g / mol, may be a peptide compound having a molecular weight of 600 g / mol or greater and less than 4,000 g / mol, or may be a peptide compound having a molecular weight of 700 g / mol or greater and less than 3,000 g / mol.

[0480] As used herein, low molecular weight compounds are compounds with a molecular weight of less than 500 g / mol. As used herein, high molecular weight compounds are compounds with a molecular weight of 30,000 g / mol or greater.

[0481] (Definition of experimental techniques)

[0482] [Characteristics]

[0483] As a model disclosed herein, a predictive model is generated through machine learning based on sequence information about the molecule and assessment results of molecular characterization. Examples of non-limiting models for molecular characterization include, but are not limited to, assessments of binding capacity, pharmacological activity, physical properties, kinetics, and molecular safety.

[0484] - Combined with competency assessment

[0485] Techniques for assessing the ability of a target molecule-binding molecule to bind to a target molecule are not limited in any way and can be assessed by quantitatively evaluating the binding of the target molecule-binding molecule to the target molecule. The target molecule is, for example, a target protein. The target molecule-binding molecule is, for example, an antigen-binding molecule, and the target molecule is, for example, an antigen. For example, when the target molecule is an antigen, the assessment can be made by measuring the binding activity of the antigen-binding molecule to the antigen. "Binding activity" refers to the total strength of non-covalent interactions between one or more binding sites of a molecule (e.g., an antibody) and its binding partner (e.g., an antigen). Here, "binding activity" is not strictly limited to 1:1 interactions between members of a binding pair (e.g., antibody and antigen). For example, when the members of a binding pair reflect a monovalent 1:1 interaction, binding activity implies intrinsic binding affinity (sometimes simply referred to as "affinity"). If the members of a binding pair are capable of both monovalent and multivalent binding, then binding activity is the sum of their affinities. The binding activity of molecule X to its partner Y is typically expressed as a dissociation constant (KD) or "the amount of analyte bound per unit amount of ligand". Binding activity can be measured using any of the conventional methods known in the art, including those described herein. Those skilled in the art can suitably determine conditions other than the concentration of compounds specific to target tissues. In a particular mode, the antigen-binding molecule provided herein is an antibody, and the binding activity of this antibody is ≤1 μM, ≤100 nM, ≤10 nM, ≤1 nM, ≤0.1 nM, ≤0.01 nM, or ≤0.001 nM (e.g., 10 μM). -8 M or lower, such as 10 -8 M to 10 -13 M, such as 10 -9 M to 10 -13 The dissociation constant (KD) of M).

[0486] In one mode, for antibody binding activity, a ligand capture method is used, such as the BIACORE® T200 or BIACORE® 4000 (GE Healthcare, Uppsala, Sweden), which uses surface plasmon resonance analysis as the measurement principle. For machine operation, for example, BIACORE® control software is used. In another mode, molecules for ligand capture, such as anti-tag antibodies, anti-IgG antibodies, and protein A on a sensor chip coated with carboxymethyl dextran (GE Healthcare, Uppsala, Sweden), are immobilized using an amine conjugation kit (GE Healthcare, Uppsala, Sweden) according to the supplier's instructions. The ligand capture molecules are diluted with 10 mM sodium acetate solution at the appropriate pH and injected at an appropriate flow rate and injection time. In measurements of binding activity, a buffer containing 0.05% polysorbate 20 (another name: Tween(R)-20) is used as the measurement buffer, and the measurement is performed at a flow rate of 10 to 30 μL / min at a measurement temperature of, for example, 25°C or 37°C. When the measurement is performed by capturing the antibody as a ligand using a molecule that allows ligand capture, the antibody is injected to allow capture of the expected amount of antibody, and then a serially diluted product of the antigen and / or Fc receptor (analyte) prepared with the measurement buffer is injected. When the measurement is performed by capturing the antigen and / or Fc receptor as a ligand using a molecule that allows ligand capture, the antigen and / or Fc receptor are injected to allow capture of the expected amount of antibody and / or Fc receptor, and then a serially diluted product of the antibody prepared with the measurement buffer (analyte) is injected.

[0487] In one mode, the measurement results are analyzed using the BIACORE® evaluation software. When calculating kinetic parameters, a 1:1 binding model is used to simultaneously fit the association and dissociation sensor maps, and the association rate (kon or ka), dissociation rate (koff or kd), and equilibrium dissociation constant (KD) can be calculated. If the binding activity is weak, especially if the kinetic parameters are difficult to calculate due to rapid dissociation, a steady-state model can be used to calculate the equilibrium dissociation constant (KD). As another parameter of binding activity, "the amount of analyte bound per unit amount of ligand" can also be calculated by dividing the amount of analyte bound at a specific concentration (RU) by the amount of ligand captured (RU).

[0488] If the antigen is a soluble molecule, KD (dissociation rate constant) can be used as a value for antigen binding activity; if the antigen is a membrane-associated molecule, apparent kd (apparent dissociation rate constant) can be used. Both kd (dissociation rate constant) and apparent KD (apparent dissociation rate constant) can be measured using methods known to those skilled in the art, and for example, using Biacore (GE Healthcare), flow cytometry, etc.

[0489] Another characterization modality involves, for example, using selection techniques for antigen-binding molecules using display libraries. One example of this modality is panning using phage display. For instance, in affinity assessment, phages presenting antigen-binding molecules that interact with the target antigen can be concentrated through a process in which phage libraries presenting different antigen-binding molecules are prepared, the target antigen is contacted with the prepared phages, and then unbound phages are washed out. Sequences with affinity for the target antigen can be identified by analyzing the nucleic acid sequences encoding the antigen-binding molecules contained in the concentrated phages. Another example of this modality is panning using mammalian cell display. In pharmacological activity assessments using display systems, for example, cells containing genes for antigen-binding molecules with the desired pharmacological activity can be isolated using a process such as flow cytometry, in which libraries containing multiple different antigen-binding molecules are expressed in the targeted mammalian cells, and the reporter gene activity is altered based on its effects on the cells. In physical property assessments using display systems, for example, cells possessing genes for antigen-binding molecules capable of stable high-level expression can be isolated using a process such as flow cytometry, in which a library containing multiple different antigen-binding molecules is expressed in targeted mammalian cells, and the expression level is examined by staining with antibodies specific to that antigen-binding molecule. Antigen-binding molecule characterization via panning is not limited to techniques using bacteriophages or mammalian cells, but can utilize various techniques that allow for the presentation of antigen-binding molecules, including but not limited to techniques allowing ribosome presentation, techniques allowing mRNA presentation, techniques allowing viral presentation other than bacteriophages, and techniques allowing bacterial presentation such as *E. coli*.

[0490] Another characterization modality is, for example, a method for obtaining antibody gene sequences from immune cells derived from an individual or a method for obtaining antibody protein sequences from serum. In affinity assessments involving the extraction of antibody gene sequences from immune cells, sequences with affinity for the target antigen can be identified, for example, by a process involving administering the target antigen protein to an individual to induce immune sensitization, and extracting the antibody gene from immune cells possessing the gene to obtain an antibody that binds to the target antigen.

[0491] The antigens that cause immune sensitization are not limited to proteins in this technology, and may include, but are not limited to, genes encoding the protein or cells expressing the protein. Examples of individuals who may be subjects include, but are not limited to, humans, mice, rats, hamsters, rabbits, monkeys, chickens, camels, llamas, and alpacas.

[0492] Examples of techniques for analyzing nucleic acid sequence or appearance frequencies include, but are not limited to, techniques for cloning recombinant organisms with nucleic acid sequences having different antigen-binding molecules and analyzing them using capillary electrophoresis via the Sanger method and techniques for analysis using next-generation sequencers.

[0493] In nucleic acid sequence analysis, the intensity of a feature can be further determined based on its frequency of occurrence. For example, the feature of an antigen-binding molecule encoded by a sequence with a high frequency of occurrence in the concentrated nucleic acid sequence can be estimated as high, while the feature of an antigen-binding molecule encoded by a sequence with a low frequency of occurrence in the concentrated sequence can be estimated as lower than the feature of an antigen-binding molecule encoded by a sequence with a high frequency of occurrence.

[0494] Techniques for obtaining information about antigen-binding molecules derived from display libraries or individuals are applicable to various types of characterization, and are not limited to those described above.

[0495] - Pharmacological activity assessment

[0496] The techniques for assessing the pharmacological activity of a molecule are not limited in any way and can be evaluated, for example, by measuring the neutralizing activity, agonist activity, or cytotoxic activity exhibited by the molecule. Examples of cytotoxic activity assessment as examples of pharmacological activity include antibody-dependent cell-mediated cytotoxicity (ADCC), complement-dependent cytotoxicity (CDC), T-cell-dependent cytotoxicity (TDCC), and antibody-dependent phagocytosis (ADCP). CDC activity is cytotoxic activity due to the complement system. ADCC activity is activity in which immune cells, etc., bind to the Fc region of an antigen-binding molecule containing an antigen-binding domain, which binds via a γFc receptor expressed on the immune cell to a membrane-associated molecule expressed on the cell membrane of a target cell, and the immune cell damages the target cell. TDCC activity is an activity characterized by target cell disorder resulting from the close proximity of T cells to target cells via a bispecific antibody containing an antigen-binding domain that binds to membrane-associated molecules expressed on the target cell membrane and an antigen-binding domain targeting any of the constituent subunits of the T cell receptor (TCR) complex on the T cell, particularly the CD3ε chain. Whether the antigen-binding molecule of interest possesses ADCC, CDC, TDCC, or ADCP activity can be determined by known methods.

[0497] Neutralizing activity is the activity of inhibiting ligands that possess biological activity against cells (such as viruses and toxins). In other words, a substance with neutralizing activity binds to such a ligand or the receptor to which the ligand binds, thereby inhibiting the binding between the ligand and the receptor. Receptors that have been prevented from binding to a ligand by neutralizing activity are thus prohibited from exercising the biological activity mediated by that receptor. When the antigen-binding molecule is an antibody, antibodies with this neutralizing activity are generally called neutralizing antibodies, and neutralizing activity can be determined by measuring the inhibitory activity of ligand-receptor binding. Ligands with biological activity against cells are not limited to viruses, toxins, etc., and the inhibitory activity of physiological effects induced by the binding of endogenous ligands (such as cytokines and chemokines) to receptors is also understood as neutralizing activity. Neutralizing activity is not limited to the inhibition of ligand-receptor binding, and the activity of inhibiting the function of biologically active proteins is also understood as neutralizing activity; examples of protein function include enzyme activity.

[0498] - Physical property assessment

[0499] The techniques for assessing the physical properties of molecules are not limited in any way, and examples of physical properties include thermal stability, chemical stability, solubility, viscosity, light stability, long-term storage stability, nonspecific adsorption, lipophilicity, and membrane permeability. In the various physical property assessments illustrated, they can be measured using methods known to those skilled in the art. The assessment methods are not limited in any way, and in the assessment of stability (such as thermal stability, chemical stability, light stability, stability to mechanical stimulation, and long-term storage stability), for example, an assessment can be made by measuring the decomposition, chemical modification, and association of the molecule of interest before and after treatments intended for stability assessment (such as heat treatment, exposure to low pH environments, exposure to light, mechanical stirring, and long-term storage). Examples of a non-limiting mode of measurement involving such stability assessments include, but are not limited to, techniques using chromatography (such as ion exchange chromatography and size exclusion chromatography, mass spectrometry, and electrophoresis), and measurements can be performed using a variety of techniques known to those skilled in the art.

[0500] Examples of physical property assessments other than those described above include, but are not limited to, assessing protein solubility using polyethylene glycol precipitation, assessing viscosity using small-angle X-ray scattering, and assessing nonspecific binding based on the assessment of binding to the extracellular matrix (ECM).

[0501] Even for assessing protein expression levels, evaluating binding to resins or ligands used for purification, and evaluating surface charge, assessments can be made as physical property assessments, provided that they can be measured using techniques known to those skilled in the art.

[0502] - Dynamics assessment

[0503] Techniques for assessing molecular kinetics are not limited in any way and can be made by administering the molecule to animals (such as mice, rats, monkeys, and dogs) and measuring the amount of the molecule in the blood over time after administration, and can also be made using techniques well known to those skilled in the art as pharmacokinetic (PK) assessments. In addition to techniques for directly assessing PK, the kinetic behavior of a molecule can be predicted from its amino acid sequence by using software to calculate the molecule's surface charge, isoelectric point, etc.

[0504] - Security Assessment

[0505] Techniques for assessing the safety of molecules are not limited in any way, and examples include immunogenicity prediction tools such as immunogenicity screening based on the ISPRI network (EpiVax, Inc.), HLA binding of fragment peptides of antigen-binding molecules, detection of T cell epitopes or assessment of T cell growth using MAPP (MHC-related peptide proteomics), and assessment of immunogenicity. Assessments can be made as long as they can be measured using techniques known to those skilled in the art, such as assessing immune responses and platelet aggregation using binding to rheumatoid factor (RF), PBMCs, or whole blood.

[0506] (Definitions of terms and techniques used in machine learning)

[0507] - MBO (Model-Based Optimization)

[0508] Model-based optimization (MBO) refers to a process where the feature values ​​themselves are not the direct object to be optimized, but rather the predicted values ​​of the feature values ​​estimated by a model are the object to be optimized.

[0509] - TPE (Tree-structured Parzen estimator)

[0510] Tree-structured Parzen estimators (TPEs) are a type of Bayesian optimization. A TPE involves a process in which, for the function to be optimized, the expected improvement in the output value given a particular input value is calculated based on the conditional probability of the output value given the input value and the probability of the output value. In other words, a TPE is a technique for optimizing a function using input values ​​that maximizes the expected improvement calculated as described above.

[0511] (Object production method)

[0512] In one aspect of this disclosure, the method for producing an object suitable for a drug candidate can be implemented by means well known to those skilled in the art.

[0513] When the object is an antibody, the antibody can be produced by, for example, the recombinant method or configuration described in U.S. Patent No. 4,816,567. One mode is a method for preparing an antibody, comprising culturing a host cell containing a nucleic acid encoding the antibody under conditions suitable for expressing an antibody as a candidate molecular compound described herein, and optionally collecting the antibody from the host cell (or host cell culture medium). The isolated nucleic acid encoding the antibody may encode an amino acid sequence of a VL containing the antibody and / or an amino acid sequence of a VH containing the antibody (e.g., the light chain and / or heavy chain of the antibody). The host cell containing such nucleic acid contains (1) a vector containing a nucleic acid encoding an amino acid sequence of a VL containing the antibody and an amino acid sequence of a VH containing the antibody, or (2) a first vector containing a nucleic acid encoding an amino acid sequence of a VL containing the antibody and a second vector containing a nucleic acid encoding an amino acid sequence of a VH containing the antibody (e.g., the host cell is transformed). In one mode, the host cell is a eukaryotic cell (e.g., Chinese hamster ovary (CHO) cell) or a lymphocyte (e.g., Y0, NSO, Sp2 / O cell). Suitable host cells for cloning or expressing vectors encoding antibodies include prokaryotic or eukaryotic cells. For example, bacteria can be used to produce antibodies, particularly when glycosylation and Fc effector function are not required. For expression of antibody fragments and peptides in bacteria, see, for example, U.S. Patent Nos. 5,648,237, 5,789,199, and 5,840,523 (see also Charlton, Methods in Molecular Biology, Vol. 248 (edited by BKCLo, Humana Press, Totowa, NJ, 2003), pp. 245-254). The antibody can be separated from the bacterial cell paste into a soluble fraction after expression and can be further purified.

[0514] When the target is a peptide or cyclic peptide compound, it can be prepared by liquid-phase synthesis, solid-phase synthesis involving Fmoc synthesis, Boc synthesis, etc., or combinations thereof. Liquid-phase and solid-phase synthesis can be performed using methods known to those skilled in the art. Solid-phase synthesis is a method in which a compound is bound to a solid and the compound is chemically reacted with a reagent on a solid resin to synthesize the compound of interest. Solid-phase synthesis of peptides is a method in which a desired amino acid or peptide is bound to a solid resin and conjugates the desired amino acid or peptide to the solid resin, thereby sequentially linking the desired amino acid or peptide to extend the peptide chain, thus synthesizing the peptide. The desired peptide can be obtained by separating the peptide bound to the solid resin from the solid resin.

[0515] (First embodiment)

[0516] Figure 1 shows a block diagram illustrating an example of a drug discovery system 100 including a molecular design device 1 according to a first embodiment.

[0517] Drug discovery system 100 is a system for creating new objects suitable for drug candidates. This system provides a method for generating new objects having predetermined properties, such as specific physiological activities (e.g., binding to a specific protein). Examples of drugs include, but are not limited to, potential activators such as low-molecular-weight drugs, medium-molecular-weight drugs, biological agents, cells, nucleic acid drugs, biopharmaceuticals, and other activators. Objects include molecular structures having desired or defined biological activities (e.g., binding to a specific protein preferentially over other proteins). Molecules that are drug candidates include biomolecules and compounds, including a variety of molecules such as nucleic acids, peptides, cyclic peptides, proteins, antibodies, target-binding molecules, high-molecular-weight compounds, medium-molecular-weight compounds, and low-molecular-weight compounds.

[0518] Drug discovery system 100 may include, for example, means for selecting molecules that interact with a target of a drug and means for generating lead molecules, which are not described herein. Drug discovery system 100 may be, for example, an information processing system that includes the disclosure of WO2020 / 246617.

[0519] The drug discovery system 100 includes a molecular design device 1. The molecular design device 1 searches for candidate molecules possessing desired characteristics and outputs information about the identified candidates. The output information is information about the structural unit combinations of the candidate molecules. In other words, the molecular design device 1 identifies candidate molecules and outputs information about the structural unit combinations of the identified candidate molecules. Here, a candidate molecule is a molecule expected to possess the desired characteristics. The candidate molecule structural unit combination information is information about the candidate molecule; it is information about some or all of the structural unit combinations of the candidate molecule. The output candidate molecule structural unit combination information may include information representing one candidate molecule, or it may include information representing multiple candidate molecules.

[0520] Drug discovery system 100 uses candidate molecule structural unit combination information output from molecular design device 1 to select new targets suitable for drug candidates. Drug discovery system 100 generates candidate molecules based on the candidate molecule structural unit combination information output from molecular design device 1, experimentally evaluates the characteristics of the candidate molecules, and selects molecules with desired characteristics as new targets suitable for drug candidates based on the characteristic evaluation results. In other words, using the candidate molecule structural unit combination information output from molecular design device 1, drug discovery system 100 can create new targets suitable for drug candidates based on the results of characteristic evaluation when actually generating candidate molecules. In this case, the candidate molecules can be considered as candidate molecules that can be validated for narrowing down the main components of a drug.

[0521] Information about the combination of structural units of a molecule can be information about the combination of some or all of the structural units of that molecule. When the information about the combination of structural units is about the combination of a portion of the structural units of the molecule, the range of its sequence can be arbitrarily set.

[0522] Structural units are the units that make up a molecule. Information about the combination of structural units in a molecule refers to information about the combinations of structural units that make up the molecule. In this application, a sequence containing a single constituent element may be referred to as a "combination". The term "sequence" can be used as an example of a combination of structural units.

[0523] For example, the designed molecule is a protein. When the molecule is a protein, the structural units are amino acids, and the information about the combination of structural units of the molecule is, for example, the sequence information of the amino acids of the protein.

[0524] The designed molecule is, for example, a nucleic acid. When the molecule is a nucleic acid, the structural unit is a nucleotide. The sequence of the molecule is, for example, the nucleotide sequence of a nucleic acid. Information about the combination of the structural units of the molecule is information about the nucleotide sequence.

[0525] More specifically, when the designed molecule is an antibody, the molecular sequence is an amino acid sequence, and the structural units are amino acids. Information about the combination of structural units of the molecule is, for example, the full-length sequence of the antibody, such as the amino acid sequence of VH or VL, or a portion of the antibody sequence, such as CDR or FR.

[0526] For example, the designed molecule is a cyclic peptide, with a molecular sequence that includes non-natural amino acids, and structural units that are both natural and non-natural amino acids. Information about the combination of structural units in the molecule refers to information about the amino acid sequence (including non-natural amino acids).

[0527] When the designed molecule is a small molecule, the structural units of the molecule are a combination of fragments, and the structural units are fragments (molecules that make up fragments of a small molecule).

[0528] For example, when the designed molecule is a nucleic acid, the molecular sequence is a nucleotide sequence and the structural unit is a base.

[0529] The desired characteristic is one required for a novel drug candidate and can be arbitrarily set. Examples of non-limiting patterns of this characteristic include, but are not limited to, the ability to bind to a predetermined in vivo target, binding affinity, pharmacological activity, physical properties, kinetics, and safety. For example, when the designed molecule is an antibody, the characteristic might be, for example, the ability of the reagent to bind to a predetermined antigen. For example, when the designed molecule is mRNA (messenger RNA), the characteristic might be, for example, the ability to translate proteins.

[0530] Molecular design apparatus 1 is an example of an inference apparatus for inferring molecules with desired characteristics from predictive information, which will be described later. Examples of desired characteristics include the ability to bind to target molecules, efficacy, and pharmaceutical properties, including membrane permeability. Hereinafter, the inference of molecules with desired characteristics from predictive information can be restated as the identification of candidate molecules.

[0531] In this embodiment, molecules expected to have the desired characteristics are shown as molecules with good prediction values ​​and low prediction uncertainty when using a prediction model, with respect to the desired characteristics.

[0532] The following section will describe the case where sequence information is processed into structural unit combination information.

[0533] The molecular design device 1 includes an inferrer 111, which includes, for example, a sequence information processor 111a, a feature predictor 111b, a prediction information processor 111c, and a candidate molecule discriminator 111d.

[0534] Sequence information processor 111a provides a set of sequence information and outputs the provided set of sequence information to feature predictor 111b. The set of sequence information is a collection of sequence information about multiple molecules. Sequence information processor 111a can autonomously generate sequence information about each molecule, or it can input sequence information about each molecule from another device. The set of sequence information can be reformulated as a set of structural unit combination information, and sequence information processor 111a can be reformulated as a structural unit combination information processor.

[0535] Feature predictor 111b predicts the features of the molecule based on the sequence information of each element of the sequence information set that is input from sequence information processor 111a, and outputs the prediction information related to the predicted features to prediction information processor 111c.

[0536] The prediction information processor 111c obtains prediction information about each molecule from the feature predictor 111b. The prediction information processor 111c outputs the obtained prediction information to the candidate molecule discriminator 111d.

[0537] The candidate molecule discriminator 111d identifies sequence information about at least one candidate molecule based on prediction information. The candidate molecule discriminator 111d may output data representing the identified sequence information to another device, or the output data may be stored in memory 14 (described later).

[0538] In this embodiment, the sequence information set may be a set of sequence information representing virtual molecular sequences (sometimes referred to as "virtual sequences") generated through machine learning, a set of sequence information about real molecular sequences (sometimes referred to as "actual sequences"), or a set including both sequence information representing virtual sequences and sequence information representing actual sequences. For example, the sequence information set may include sequence information about virtual sequences generated by a virtual sequence generation model. The sequence information set may include sequence information representing actual sequences obtained as, for example, existing databases and experimental results. The sequence information set may include sequence information extracted from all candidates for combinations of structural units through combinatorial optimization, which will be described later.

[0539] Sequence information processor 111a processes information to provide a sequence information set. Sequence information processor 111a may further include a sequence information set acquirer (not shown) configured to perform an acquisition process for using a machine learning model to acquire sequence information as output about certain input information. Specifically, in this case, sequence information processor 111a acquires the sequence information set and outputs the acquired sequence information set to feature predictor 111b.

[0540] In this embodiment, sequence information about multiple molecules is input from sequence information processor 111a to feature predictor 111b.

[0541] The feature predictor 111b calculates feature prediction values ​​and estimates of prediction uncertainty for the input sequence information of various molecules.

[0542] In this embodiment, the feature predictor 111b includes a prediction value calculator 111x and a prediction uncertainty estimator 111y. The prediction value calculator 111x takes sequence information about the molecule as input and uses a prediction model to calculate feature prediction values. The feature prediction values ​​are the feature values ​​predicted using the prediction model. The prediction uncertainty estimator 111y estimates the uncertainty of the feature prediction values ​​calculated by the prediction value calculator 111x. The value representing the estimated uncertainty is called the estimate of the prediction uncertainty.

[0543] The predictive model is generated, for example, through learning based on training data, which includes multiple sets of sequence information about each molecule and the molecule's evaluation properties.

[0544] The prediction calculator 111x can calculate the predicted values ​​of multiple features. When predicting the feature values ​​of multiple features, the prediction uncertainty estimator 111y can estimate the prediction uncertainty for each feature's predicted value, or it can estimate the prediction uncertainty for one of the feature's predicted values.

[0545] Feature predictor 111b outputs the predicted feature value and an estimate of the uncertainty of the prediction for that feature value as prediction information about each molecule to prediction information processor 111c.

[0546] The candidate molecule discriminator 111d processes the predicted information to infer molecules with the desired characteristics and to identify candidate molecules.

[0547] In this embodiment, the candidate molecule discriminator 111d infers molecules with desired features based on feature prediction values ​​and estimates of prediction uncertainty for sequence information about each molecule, output from the feature predictor 111b. The candidate molecule discriminator 111d identifies at least one candidate molecule from a plurality of molecules, the sequence information of which is contained in a sequence information set from the sequence information processor 111a, based on prediction information obtained from the prediction information processor 111c. The candidate molecule discriminator 111d may use feature prediction values ​​of at least one feature or multiple features in the identification of candidate molecules. The candidate molecule discriminator 111d may use multiple feature prediction values ​​and estimates of prediction uncertainty for at least one feature prediction value as prediction information in the identification of candidate molecules. That is, in the identification of candidate molecules, it is permissible to constrain a second feature value (the prediction uncertainty for the feature prediction value) to optimize the first feature value (the feature prediction value). The candidate molecule discriminator 111d may identify molecules with desired features as candidate molecules based on feature quality values ​​described later.

[0548] For example, the candidate molecule discriminator 111d can select at least one molecule as a candidate molecule, wherein the predicted feature value and the prediction uncertainty respectively satisfy predetermined conditions. The candidate molecule discriminator 111d can also select at least one molecule as a candidate molecule, wherein the estimated value of the prediction uncertainty for at least one predicted feature value respectively satisfies predetermined conditions.

[0549] The candidate molecule discriminator 111d can identify at least one candidate molecule based on a feature quality value calculated based on a feature prediction value and an estimate of the uncertainty of the prediction. In this case, for example, the feature quality value can be calculated in the prediction information processor 111c, or it can be calculated in the feature predictor 111b.

[0550] Feature predictions are values ​​based on the predicted feature values ​​and estimates of the prediction uncertainty. Feature quality values ​​can also be considered as objective variables calculated using a predetermined function that uses the predicted feature values ​​and estimates of the prediction uncertainty as explanatory variables.

[0551] For example, the feature quality value can be an index value that gives a smaller value as the estimated uncertainty of the feature prediction increases, and a larger value as the predicted feature value increases. In other words, the feature quality value only needs to increase with the predicted feature value and decrease with the estimated uncertainty of the prediction. As an example, the feature quality value is the difference between the predicted feature value and a predetermined coefficient multiple of the estimated uncertainty of the prediction in a linear region, or the difference between a predetermined coefficient multiple of the predicted feature value and the estimated uncertainty of the prediction. The predetermined coefficient is a positive constant and represents the degree to which the feature value or the estimated uncertainty of the prediction contributes to the feature quality value.

[0552] Therefore, when a feature prediction value is defined that increases with the degree to which the desired feature is predicted, or increases with the degree of the feature itself, the feature quality value is an indicator that increases with the increase of the feature prediction value and the decrease of the estimate of the prediction uncertainty. In this case, the larger the feature quality value, the higher the probability that the actual generated molecule will exhibit the expected feature. That is to say, in molecular design, estimation using feature quality values ​​can be considered a risk-averse estimation process.

[0553] The feature quality value does not need to be the difference between the predicted value of the feature in the linear region and a predetermined coefficient multiple of the estimated uncertainty of the prediction. The feature quality value can be obtained by: exponentializing one of the predicted value of the feature and the estimated uncertainty of the prediction a predetermined number of times, and calculating the difference between the obtained value and other values, or by calculating the quotient of the obtained value and other values. Therefore, when defining a feature prediction value that gives a smaller value as the probability of having the desired feature increases or as the degree of that feature increases, the feature prediction value can be determined by dividing the feature prediction value by the estimated uncertainty of the prediction or by other types of calculations (such as calculating these values ​​in the logarithmic domain). The calculation of the feature quality value only requires derivation using the prediction value and its uncertainty, and there are no restrictions on the related functions, processes, etc.

[0554] The candidate molecule discriminator 111d identifies molecules with the desired feature based on the calculated feature quality value. As an example, in the candidate molecule discriminator 111d, among molecules whose feature prediction values ​​are obtained in the feature predictor 111b, molecules that give a better feature quality value are more likely to be identified as molecules with the desired feature. Here, the term "good feature quality value" means obtaining a feature quality value that indicates a high probability of being predicted to have the desired feature or a high degree of that feature.

[0555] As another example, candidate molecule discriminator 111d can identify molecules whose feature quality values ​​satisfy predetermined conditions as molecules with desired features from molecules whose feature prediction values ​​are calculated in feature predictor 111b. As yet another example, candidate molecule discriminator 111d can rearrange multiple molecules whose feature prediction values ​​are calculated in feature predictor 111b, ranking the molecules based on feature quality values, and identifying molecules within a predetermined range of rearranged molecules as molecules with desired features. Here, candidate molecule discriminator 111d can identify a certain number of molecules as molecules with desired features, such that molecules with better feature quality values ​​are given higher priority. Candidate molecule discriminator 111d can identify molecules whose feature prediction values ​​and estimates of prediction uncertainty satisfy predetermined conditions as molecules with desired features.

[0556] The accuracy of the feature predictions calculated by feature predictor 111b affects molecule identification. The accuracy of sequence-based feature predictions varies depending on the training data of the feature prediction model. Here, prediction accuracy can be restated as prediction reliability or prediction certainty. Low prediction accuracy can be restated as low prediction reliability, low prediction certainty, or high prediction uncertainty. High prediction accuracy can be restated as high prediction reliability, high prediction certainty, or low prediction uncertainty.

[0557] In the molecular design apparatus 1 of this embodiment, not only feature prediction values ​​but also estimates of the prediction uncertainty are used as indicators for processing to infer molecules with the desired features. That is, in the molecular design apparatus 1 of this embodiment, instead of a mathematical model for predicting feature values, a mathematical model for predicting feature values ​​and estimating the prediction uncertainty is used as the object to be used by the candidate molecule discriminator when identifying candidate molecules. Accordingly, the inference process can prevent the identification of candidate molecules with high feature prediction values ​​but low prediction certainty as candidate molecules. In other words, it prevents the identification of molecules that may not have the expected high feature quality when actually generated. Accordingly, this embodiment can further reduce the burden associated with drug discovery.

[0558] (Second Embodiment)

[0559] Figure 2 shows a block diagram illustrating an example of a drug discovery system 100 including a molecular design device 1 according to a second embodiment. The following description mainly includes the differences between this embodiment and the first embodiment. Unless otherwise stated, descriptions of the same points as in the first embodiment will be incorporated.

[0560] The molecular design apparatus 1 according to the invention is an example of an inference apparatus for inferring molecules expected to have desired characteristics from predictive information, which will be described later.

[0561] The molecular design device 1 includes an inferrer 111, which includes, for example, a sequence information processor 111a, a feature predictor 111b, a prediction information processor 111c, and a candidate molecule discriminator 111d.

[0562] According to the inferrer 111 of this embodiment, a combinatorial optimization algorithm (sometimes referred to as "combinatorial optimization" in this application) is executed to infer molecules that are expected to have the desired characteristics.

[0563] Sequence information processor 111a performs combinatorial optimization to extract sequence information about an arbitrary number of molecules from a preset search space. That is, sequence information processor 111a repeatedly processes, through combinatorial optimization, extracting sequence information about some molecules expected to have desired features from candidates of structural unit sequences in the search space using updated extraction parameters. The search space can be arbitrarily set. The search space can be set based, for example, on training data used in training a predictive model for use in feature prediction. The number of times sequence information is acquired, i.e., the number of repetitions of acquiring sequence information, can be any natural number greater than or equal to 1. In inferrer 111, the number of times sequence information is acquired can be preset. Inferrer 111 can count the number of times sequence information processor 111a actually extracts sequence information as the repetition count. In this embodiment, sequence information about molecules contained in the search space can be elements of a sequence information set.

[0564] The sequence information processor 111a outputs the extracted sequence information about multiple molecules to the feature predictor 111b.

[0565] According to this embodiment, the feature predictor 111b uses a prediction model to generate prediction information representing the features of each molecule based on the sequence information of each molecule input from the sequence information processor 111a, and then outputs the prediction information generated for each molecule to the prediction information processor 111c. The prediction model is generated, for example, through learning based on training data, which includes multiple sets of sequence information about each molecule and results of evaluating the properties of the molecule.

[0566] The prediction information for each molecule includes, for example, the predicted feature values ​​for each molecule. Therefore, the feature predictor 111b may include a prediction value calculator 111x, which is configured to calculate the predicted feature values ​​from the sequence information for each molecule.

[0567] The prediction information for each molecule may include, for example, the predicted feature value for each molecule and an estimate of the uncertainty of that predicted feature value. The feature predictor 111b may include a prediction uncertainty estimator 111y, which is configured to calculate an estimate of the uncertainty of the predicted feature value. Accordingly, the prediction information for each molecule output from the feature predictor 111b may include both the predicted feature value for each molecule and an estimate of the uncertainty of that predicted feature value.

[0568] The prediction information processor 111c obtains prediction information about each molecule from the feature predictor 111b. When the prediction information includes a feature prediction value and an evaluation value of the uncertainty of that feature prediction value, the prediction information processor 111c can calculate a feature quality value based on the feature prediction value of each molecule and the uncertainty of the feature prediction value about the sequence information of each molecule contained in the sequence information set. In this embodiment, the sequence information processor 111a uses the feature quality value calculated by the prediction information processor 111c as prediction information to update the extraction parameters. Specific examples of extraction parameters will be described later.

[0569] If the counted number of repetitions is less than the preset number of times sequence information is acquired, the sequence information processor 111a updates the extraction parameters based on the acquired prediction information. Therefore, in combinatorial optimization, a series of processes are repeated, including extracting sequence information based on the updated extraction parameters in the sequence information processor 111a, acquiring prediction information based on the extracted sequence information by the feature predictor 111b, and updating the extracted parameters by the sequence information processor 111a.

[0570] When the counted repetitions reach the preset number of times sequence information is obtained, the sequence information processor 111a stops processing with the combinatorial optimization algorithm, and the prediction information processor 111c outputs the prediction information about each molecule input from the feature predictor 111b to the candidate molecule discriminator 111d. Based on the prediction information about each molecule input from the prediction information processor 111c, the candidate molecule discriminator 111d identifies at least one candidate molecule from the sequence information set at the corresponding time. When the number of times sequence information is obtained is N, the prediction information processor 111c identifies the sequence information about at least one candidate molecule from the N extracted sequence information about the molecule. The candidate molecule discriminator 111d can identify the sequence information about at least one candidate molecule from the set of N extracted sequence information about the molecule. When the prediction information includes feature quality values, the candidate molecule discriminator 111d can identify candidate molecules based on the feature quality values ​​of each molecule.

[0571] The candidate molecule discriminator 111d outputs sequence information about the identified candidate molecules.

[0572] In this embodiment, Bayesian optimization can be applied to combinatorial optimization, but this embodiment is not limited to this. Examples of Bayesian optimization techniques include TPE (Tree-based Parzen Estimator), and as an evolutionary algorithm technique, any algorithm can be used, such as NSGA II (Non-dominated sorting genetic algorithm with elitist strategy).

[0573] In this embodiment, by using combinatorial optimization, molecules with desired characteristics can be inferred without using their sequence information to predict the feature values ​​of all molecules contained in the search space. If a large number of molecules were to be covered as objects to be processed, significant computational time would be required. However, by performing combinatorial optimization on the search space as in this embodiment, molecules with desired characteristics can be inferred without covering all molecules. This further reduces the burden associated with drug discovery.

[0574] In this embodiment, the molecular design device 1 can use either a mathematical model for estimating feature values ​​from sequence information or a mathematical model for estimating feature quality values ​​as the objective function for combinatorial optimization. The output of the mathematical model is optimized through combinatorial optimization. When using a mathematical model for estimating feature quality values, molecules in which the features indicated by the feature prediction values ​​are preferred but the prediction certainty is low can be prevented from being identified as candidate molecules. That is, it is possible to reduce the likelihood of identifying molecules that may not have the expected high feature quality when actually generated as candidate molecules. Accordingly, the burden accompanying drug discovery can be further reduced.

[0575] Figure 3 illustrates an example of the sequence information set used to describe the first and second embodiments. The “Sequence ID” in Figure 3 is a discriminator that identifies the sequence information about each molecule contained in the sequence information set. Figure 3 shows the sequence information about each molecule contained in the sequence information set. The sequence information about each molecule represents multiple structural units and arrangements of the corresponding structural unit. An arrangement of a single structural unit is shown between information D101 and information D102. For example, information D101 is information about the structural unit in the molecular sequence associated with position H1. VHL0001, VHL0002, VHL0003, VHL0004, etc., each represent a molecule in which the structural unit at position H1 is M.

[0576] Figure 4 shows a schematic diagram illustrating an example of the hardware configuration of the molecular design device 1 according to the first and second embodiments. The molecular design device 1 includes a controller 11, which includes a processor 91 (such as a CPU) and memory 92 interconnected via a bus, and includes a computer system executing a predetermined program. The computer system can also be considered to function as the molecular design device 1, including the controller 11, input terminal 12, communicator 13, storage device 14, and output terminal 15, by executing the program. In this application, the phrase "execution of a program" or "execution of a program" includes the meaning of performing processing as instructed by each of one or more commands written in the program.

[0577] More specifically, processor 91 reads a program stored in memory 14 and stores the read program in memory 92. As processor 91 executes the program stored in memory 92, it performs the functions of molecular design device 1, which includes controller 11, input terminal 12, communicator 13, memory 14, and output 15. The program is stored in memory 14 in advance, for example.

[0578] The controller 11 controls the operation of various functional units of the molecular design device 1. The controller 11 has, for example, the function of an inferrer 111, which performs inference processing, wherein a molecule with desired characteristics is inferred.

[0579] Input terminal 12 includes input devices such as a mouse, keyboard, and touch panel. Input terminal 12 can be configured to connect the input devices to the interface of molecular design device 1. Input terminal 12 receives various types of information about molecular design device 1. For example, training data of a prediction model can be input to input terminal 12.

[0580] The communicator 13 includes a communication interface for connecting the molecular design device 1 to an external device. The communicator 13 may or may not use a wire to communicate with the external device. The external device is a means of sending training data for a prediction model. The external device may be a means of sending sequence information about candidate molecules to it. The molecular design device 1 is connectable to the Internet and is implemented as a server device capable of communicating with each of one or more terminal devices.

[0581] Memory 14 is formed using a computer-readable storage medium (non-transitory computer-readable storage medium) such as a hard disk or semiconductor storage device. Memory 14 stores various information about the molecular design device 1. Memory 14 stores information input via, for example, input terminal 12 or communicator 13. Memory 14 stores various information generated by processes performed by, for example, controller 11. Memory 14 stores, for example, prediction models, sequence information, and various parameters (which may include the extracted parameters described above) for use in combinatorial optimization. Memory 14 stores, for example, the programs described above.

[0582] Output terminal 15 outputs various information. Output terminal 15 includes a display device, such as a CRT (cathode ray tube) display, a liquid crystal display, or an organic EL (electroluminescent) display. Output terminal 15 can be configured as an interface to connect the output device to the molecular design device 1. Output terminal 15 outputs information input to, for example, input terminal 12 or communicator 13. Output terminal 15 can output various information generated by processing performed by, for example, controller 11.

[0583] Figure 5 shows a schematic diagram illustrating an example configuration of the controller 11 according to the first and second embodiments. The controller 11 includes an inferrer 111, an input controller 112, a communication controller 113, a storage controller 114, and an output controller 115. The inferrer 111 performs inference processing. The input controller 112 controls the operation of the input terminal 12. The communication controller 113 controls the operation of the communicator 13. The storage controller 114 controls the operation of the memory 14. The output controller 115 controls the operation of the output terminal 15.

[0584] Next, examples of inference processing according to various embodiments will be described. Figure 6 shows a flowchart illustrating an example of the processing flow performed by the molecular design device 1 in the first embodiment.

[0585] Sequence information processor 111a acquires a set of sequence information including sequence information about multiple molecules (step S101).

[0586] Feature predictor 111b calculates feature prediction values ​​for each molecule that forms the sequence information set, and then calculates an estimate of the uncertainty of the feature values ​​(step S102).

[0587] Candidate molecule discriminator 111d identifies candidate molecules based on the calculated feature prediction values ​​and the estimated uncertainty of the prediction for each molecule (step S103). The technique used by feature predictor 111b to estimate the uncertainty of the prediction is not limited, as long as it can quantify the uncertainty. For example, the standard deviation of the feature prediction values ​​can be calculated as an estimate of the uncertainty of the prediction. Conformal prediction can be used in the uncertainty quantification. Feature predictor 111b can divide the learning data into multiple parts, calculate the prediction error (prediction error) for each part, and quantify the uncertainty based on the distribution of the error.

[0588] The candidate molecule discriminator 111d uses output terminal 15 to output the sequence information of the discriminated candidate molecule as the sequence information of the candidate molecule (step S104). Thereafter, the controller 11 completes the processing shown in FIG6.

[0589] Prior to step S103, the prediction information processor 111c may calculate a feature quality value based on the feature prediction value and the estimate of the prediction uncertainty obtained in step S102. The prediction information processor 111c may calculate the mean-variance (MV) as an example of the feature quality value. MV is obtained by subtracting the penalty function g(x) from the product of the risk resistance parameter p and the feature prediction value f'(x) using a predetermined prediction model (MV = pf'(x) - g(x)). The risk resistance parameter is a positive real value representing the resistance of the uncertainty of the feature prediction value f'(x) to g(x), i.e., the reliability of the prediction value f'(x). The prediction information processor 111c calculates the standard deviation as an example of the estimate of uncertainty g(x). MV can be applied to portfolio optimization in financial engineering. In financial engineering, MV can be used as an indicator in which the expected return is integrated as the mean and the variance of the return is integrated as the risk. Regardless of the optimization algorithm, MV is a metric that can be used in a variety of fields and for a variety of purposes (see, for example, Q. Zhu and VYF Tan: Thompson Sampling Algorithms for Mean-Variance Bandits (2020 ICML), S. Takemori: Distributionally-Aware Kernelized Bandit Problems for Risk Aversion (2022 ICML)).

[0590] The prediction information processor 111c can perform calculations using the uncertainty g(x) as an instance of the feature quality value as a constraint. For example, the prediction information processor 111c can provide a threshold τ for the uncertainty g(x), and set the feature quality value to the feature prediction value f'(x) if the uncertainty g(x) is equal to or less than τ, and set the feature quality value to a sufficiently small value (e.g., the minimum value that can be given by f'(x)) if the uncertainty g(x) is greater than τ. This configuration is represented by expression (1).

[0591] [Expression 1]

[0592]

[0593] The prediction information processor 111c can calculate a feature quality value based on multiple physical property values. For example, the prediction information processor 111c provides a threshold τ1 for a first physical property value g1(x) and a threshold τ2 for a second physical property value g2(x). When the first physical property value g1(x) is less than τ1 and the second physical property value g2(x) is less than τ2, the prediction information processor 111c can set the feature quality value to the feature prediction value f'(x). On the other hand, when the first physical property value g1(x) is equal to or greater than τ1 or the second physical property value g2(x) is equal to or greater than τ2, the prediction information processor 111c can set the feature quality value to a sufficiently small value (e.g., the minimum possible value of f'(x)).

[0594] Alternatively, the predictive information processor 111c can calculate the feature quality value based on the predicted value Pr (x ∈ F|x) of the probability of satisfying the constraints. That is, the predictive information processor 111c can multiply the predicted value f'(x) of the feature obtained using the predictive model by the predicted value of the probability of satisfying the constraints, and sample only the sequence information where the probability of satisfying the constraints is predicted to be high. In this example, the feature quality value is close to 0 when the probability of satisfying the constraints is low, and the feature quality value approaches the target index itself as the probability of satisfying the constraints becomes higher (closer to 1). This feature quality value is represented by expression (2).

[0595] [Expression 2]

[0596]

[0597] The number of constraints can be one, two, or more. For example, when calculating a characteristic quality value based on multiple physical property values, the predictive information processor 111c can set constraints for each physical property value and multiply the predicted characteristic value f'(x) obtained using the predictive model by the probability of satisfying all conditions. In this case, the probability of satisfying all conditions can be the product of the probabilities of satisfying the constraints on the physical property values.

[0598] Figure 7 shows a flowchart illustrating an example of the processing flow performed by the molecular design device 1 in the second embodiment.

[0599] Sequence information processor 111a acquires a dataset D comprising multiple data pairs as training data, each data pair being a set of input values ​​x and measurement values ​​y (step S201). For example, the dataset D input from input terminal 12 or communicator 13 is pre-stored in memory 14. Sequence information processor 111a reads the dataset D from memory 14.

[0600] The sequence information processor 111a uses the acquired training data to learn, for example, a predictive model for predicting input values ​​and measurements by performing Gaussian process regression (step S202).

[0601] Inferrer 111 performs combinatorial optimization to identify candidate molecules (step S203). Candidate molecule discriminator 111d uses output terminal 15 to output sequence information about the identified candidate molecules (step S204). After this, the processing shown in Figure 7 is completed.

[0602] <Examples of applications in MBO>

[0603] Next, an example of an application to MBO will be described.

[0604] In this application example, offline MBO (a computer simulation-based drug design method) is combined and applied with molecular design based on structural units.

[0605] Offline MBO is a method for searching for the best molecule from a "proxy" predictive model generated from acquired data. This method involves black-box optimization of the surrogate model as a black box, also known as inverse analysis. Here, MBO is a technique in which a predictive model is learned in advance using accumulated experimental data as training data, drug candidate molecules are evaluated based on the feature values ​​obtained using the learned predictive model, and the predictive model is updated based on the evaluation results. In the following description, the predictive model may be referred to as the surrogate model.

[0606] In this application example, discrete input values ​​can be the objects to be optimized using MBO. In MBO, it is assumed that discrete input values ​​x ∈ X and measured values ​​y are given as an unknown function f(x), with a dataset D (where n is an integer of 2 or greater) containing n sets of discrete input values ​​x and measured values ​​y, and no additional input values ​​x and measured values ​​y are available. The discrete input values ​​x and measured values ​​y give the correct solution (also called an oracle) to the function f(x). General MBO is designed to determine the input values ​​x that maximize the function f(x) in the discrete value space X. The above technique can be viewed as a problem of determining the input values ​​x that maximize the feature quality value f(x) - λg(x) rather than the function f(x). Here, a sequence is applied as the input x, and the function f(x) corresponds to a function used to predict the real values ​​corresponding to the feature values. The penalty function g(x) is a function that gives a real value representing the uncertainty of the value of the function f(x). λ corresponds to the coefficient α. The input value x is expressed using a vector representation.

[0607] In offline MBO, the sequence information processor 111a is pre-trained using a dataset D comprising multiple sets of discrete input values ​​x and measured values ​​y as training data. During the learning of the prediction model, the sequence information processor 111a determines the parameters of the prediction model such that the objective function of the discrete input value x yields the predicted value y', and the exponential value representing the magnitude of the difference between the predicted value y' and the measured value y of the discrete input value x is minimized. In this application example, as described below, the inferrer 111 samples the discrete input values ​​x and updates the combined maximized parameters based on the predictions obtained using the prediction model for each sample.

[0608] The inferrer 111 can use, for example, TPE in combinatorial optimization. TPE is a black-box optimization algorithm in which Parzen window density estimation is combined with Bayesian optimization. TPE can be adapted to class parameters and therefore can be applied to molecular design based on structural units.

[0609] In this application example, the inferrer 111 implements the functionality of TPE and is applied in molecular design as follows. TPE is a technique designed to maximize the expected improvement (EI) of an objective function. First, the sequence information processor 111a applies combinations of structural units that form drug candidate molecules to a search space and assigns the output from the prediction model to a target score set Y corresponding to the set of input molecules X sampled from the search space. For example, the sequence information processor 111a sets the feature prediction values ​​obtained using the prediction model as the target score of the TPE. The sequence information processor 111a samples the set of input molecules X from the search space pairs, that is, it assigns candidates of structural units to each change position in the combination of structural units. For example, when the drug candidate molecule is a protein, candidate amino acids are assigned to each change position in the sequence of that protein. As an example, the Optuna implementation can be used to perform the TPE sampler.

[0610] Next, feature predictor 111b uses a prediction model for the input values ​​determined by sampling to compute an estimate (evaluation) of the objective function.

[0611] The prediction information processor 111c outputs the selected input values ​​and the calculated estimates to the sequence information processor 111a, and the sequence information processor 111a uses the input values ​​and estimates from the prediction information processor 111c to update the parameters of the optimization algorithm. Here, the sequence information processor 111a assigns the selected input values ​​and the calculated estimates to the set of input molecules X and the set of target scores Y sampled from the search space, respectively. The inferrer 111 repeats the process of data partitioning, sampling, evaluation, and updating, thereby incorporating the evaluation results of the input values ​​(sampled values) obtained through sampling into the prediction model, and performing a probabilistic search for combinations of structural units with large objective function estimates.

[0612] Figure 8 illustrates an example of a combinatorial optimization process performed by the molecular design device 1 according to this application example.

[0613] Inferrer 111 sets the initial value of 0 for the number of times sequence information is acquired. Loop R20 includes the processing in steps S203a to S203c. Controller 11 defines the condition for executing loop R20 as the number of repetitions being equal to or less than a predetermined number of samples N. The number of samples corresponds to the number of times sequence information is acquired.

[0614] Sequence information processor 111a acquires sequence information about multiple molecules according to a combinatorial optimization algorithm (step S203a). For example, when TPE is used as the combinatorial optimization algorithm, sequence information processor 111a divides the set of input values ​​into two sets based on the estimated values ​​(corresponding to feature prediction values) obtained using a prediction model for each input value and a predetermined threshold γ. One set (called "set 1") includes input values ​​when the estimated value is equal to or greater than the threshold γ. The other set (called "set 2") includes input values ​​when the estimated value is less than the threshold γ. Sequence information processor 111a samples the input values ​​that maximize the expected improvement of the objective function. The expected improvement corresponds to the amount of increase in the expected value of the objective function before and after the update, and is known to be proportional to the input value x with p(x|y1) / p(x|y2) belonging to set 1. p(x|y1) represents the density distribution of input value x in set 1. p(x|y2) represents the density distribution of input value x in set 2. That is, EI is an instance of extracted parameters. The sequence information processor 111a samples the input value that maximizes the calculated EI.

[0615] Feature predictor 111b uses a prediction model to perform sequence evaluation. Here, feature predictor 111b uses the prediction model to compute an estimate of the objective function of the sampled input values ​​(step S203b).

[0616] The sequence information processor 111a assigns the selected input values ​​and the calculated estimates to the set of input molecules X sampled from the search space and the set of target scores Y. Therefore, the new input values ​​are added to the set of input molecules X that matches the objective function and is sampled from the search space, and thus the extraction parameters of the combinatorial optimization algorithm involved in the sequence information set extraction are updated (step S203c).

[0617] The inferrer 111 updates the repetition count by incrementing by 1 each time the processing in steps S203a to S203c is completed. When the repetition count is equal to or less than the number of samples N, the inferrer 111 repeats the processing in steps S203a to S203c.

[0618] When the number of repetitions exceeds the number of samplings N, the inferrer 111 completes the processing of loop R20. The candidate molecule discriminator 111d can use output terminal 15 to output a predetermined number of structural unit combinations in descending order of the estimated value of the objective function as sequence information about the candidate molecules. After this, the processing shown in Figure 8 is completed.

[0619] Figure 9 illustrates another example of the combinatorial optimization process performed by the molecular design device 1 according to this application example. The process in Figure 9 is identical to the process in Figure 8 in terms of repeating steps S203a to S203c, but differs from the process in Figure 8 in that steps S203d and S203e are provided instead of looping R20. In step S203d, the inferrer 111 sets the initial value of the sampling number N and the number of repetitions to 0. After the processing in step S203b, the process proceeds to step S203e. In step S203e, the inferrer 111 increments the number of repetitions here by 1 and determines whether the number of repetitions has reached N. When it is determined that the number of repetitions has reached N (yes in step S203e), the process in Figure 9 is completed. When it is determined that the number of repetitions has not yet reached N (no in step S203e), the process proceeds to step S203c. After the processing in step S203c, the process proceeds to step S203a. Therefore, the processing in Figure 9 is the same as... Figure 8 The difference in the processing is that when the number of repetitions exceeds the number of samplings N, the inferr 111 does not update the extraction parameters.

[0620] In the examples in Figures 8 and 9, it is assumed that the completion condition is set to the time required for the number of repetitions to exceed the number of times sequence information is acquired, but the completion condition is not necessarily limited to this. For example, the completion condition could be that the target value of the estimated objective function is preset, and the estimated objective function reaches the target value.

[0621] In the verification examples of this application, the TPE calculated from the predicted mean and predicted variance of Gaussian process regression as the objective function can be called MV-TPE. The TPE using the predicted mean of Gaussian process regression as the objective function can be called mean-TPE. In this application, the predicted mean of Gaussian process regression as the objective function refers to the predicted value itself.

[0622] <First Verification Example>

[0623] Next, a validation example will be described in this embodiment. In the first validation example, the GFP dataset is used as training data for the prediction model to validate the security of this application example. The GFP dataset includes information on 56,086 GFP (green fluorescent protein) sequences and their brightness as characteristics. GFP is a widely used sample in medical and biological research. The generation of optical GFP can be used as a benchmark for model-based optimization performance as a research objective. A 756-dimensional vector representation is used to express the input values ​​representing each GFP sequence. In this validation, the task of evaluating the protein embedding method (TAPE) is used as the protein embedding model in the acquisition of the vector representation.

[0624] As training data, a Gaussian process (GP) was used to learn a predictive model based on GFP sequences, where the number of mutations relative to the parental sequence avGFP (aequorea victoria GFP) is 2 or less. In the description below, the predictive model obtained through learning can be referred to as the surrogate model. The parental sequence avGFP can simply be referred to as the parental sequence or template GFP. This process validates sequences in which a small number of residues from the parental sequence are substituted, thus enabling a practical molecular optimization process. In this validation, a GFP sequence with 2 or fewer residue substitutions (edit distance) was used, as shown in Figure 10. LightGBM (Lightweight Gradient Boosting Machine) was learned using all data from the GFP dataset as the pseudo-correct solution model. LightGBM can be used for ranking, classification, and other tasks based on decision tree algorithms. In this configuration, the GFP sequences surrounding the parental sequence are covered by the surrogate model, and a wider GFP space is covered by the pseudo-correct solution model as the search space.

[0625] The search space is defined as the mutations among the top 100 sequences in the training data. The search space includes, for example,... Figure 17 The 37 mutation candidate positions shown are represented by 2 to 5 amino acid candidates at each position. The TPE parameters are set to "Number of samples: 3,000" and "Multivariate: None" (i.e., the objective function is about one variable), as shown in Figure 18. The top 10 sequences from the training data were used for warm-start initialization in MV-TPE and mean-TPE, respectively.

[0626] Typically, in drug discovery, 10 to 100 sequences are evaluated in a single batch. In this validation, as... Figure 10 As shown in the figure, the 96 highest-scoring sequences (i.e., the estimates of the objective function when using the pseudo-corrected solution model) are selected as the proposed sequences and used as the objects to be evaluated.

[0627] In this validation, the processes shown in Figures 7 and 8 were performed using MV and the mean as objective functions, respectively. The mean was used as a comparison example. MV and the mean each include features. Among these features, the brightness of GFP was used. Figure 11 shows the trajectory of TPE optimization, with the mean, as an estimate of the objective function, shown in the upper half and MV in the lower half. The results of 10 optimization processes are shown in each half. Dashed lines represent the estimated values ​​of the samples in each process.

[0628] Figure 12 shows the relationship between the mean (GP mean) and standard deviation (GP Std) of the sample values ​​during the process. Compared to using the mean, using MV as the objective function results in a smaller standard deviation. This means that using MV can reduce the uncertainty of the estimates.

[0629] Figure 13 shows the density distribution of edit distances for the samples, with the template GFP as a reference. Using MV as the objective function results in a smaller overall edit distance compared to using the mean. This means that by utilizing MV, GFP with fewer mutations relative to the template GFP is selected as the sample.

[0630] Figure 14 shows the distribution of scores for the proposed sequence obtained using the pseudo-correct solution model. This score represents an estimate of the fluorescence intensity emitted by GFP. Using MV as the objective function results in higher intensity compared to the case using the mean.

[0631] Figure 15 shows the edit distance of the proposed sequence with template GFP as a reference. Using MV as the objective function results in a smaller edit distance compared to using the mean. This means that by using MV, sequences with few mutations can be sampled, i.e., safe optimization can be achieved.

[0632] Figure 16 shows the standard deviation (GP Std) of the proposed sequence. Using MV as the objective function results in a smaller standard deviation compared to using the mean. This indicates that safe optimization can be achieved by sampling sequences with minimal brightness uncertainty as feature values ​​using MV.

[0633] <Second Verification Example>

[0634] Next, a second validation example will be described. In this second validation example, data on the bispecific antibodies is used as training data, and a Gaussian process is performed to learn the predictive model. The training data used in this validation example consists of feature values ​​representing the sequences and binding abilities of the bispecific antibodies that recognize MarvelD3 and CD3 as antigens. Binary octet values ​​are used as the feature values ​​for binding ability. Vector representations obtained using tape are used as input values ​​for the predictive model. The sequences of the bispecific antibody proteins used as samples are represented by vector representations.

[0635] MarvelD3 is a tight junction protein with a 4-transmembrane structure. In this validation example, MarvelD3 was designated as a target candidate for an anticancer drug. The development of bispecific antibodies bridging cancer antigens and T-cell antigens holds promise for cancer therapy. In this validation, a learned predictive model was used to identify anti-MarvelD3 candidates with better characteristics from lead antibodies.

[0636] In this validation, the binary octet values ​​of the antibody sequences were measured multiple times, and the antibody sequences obtained through measurement and their binary octet values ​​were used as input to the prediction model. Batch measurements were performed to obtain the binary octet values ​​of typically 100 or fewer antibody sequences in a single measurement.

[0637] Here, the antibody used in the measurement was obtained through the following procedure. First, a plasmid encoding a pre-designed heavy or light chain was provided, and recombinant antibodies were transiently expressed using Expi293F cells. The antibody was removed from the culture supernatant using protein A and dissolved in a buffer solution. The dissolved buffer solution was mixed under reducing conditions to prepare the MarvelD3 / CD3 bispecific antibody. Here, charge repulsion was applied between identical heavy chains to achieve selective heavy chain heterodimerization. The concentration of the antibody in the buffer solution was determined by absorbance at 280 nm. Subsequently, the buffer solution containing the bispecific antibody was subjected to ion-exchange chromatography, confirming the provision of the expected MarvelD3 / CD3 antibody.

[0638] The Octet HTX system is used for the measurement of binary eight-bit values. Extracellular vesicles with CD81 protein and human Marvel D3 on their surface are captured on a sensor chip using an anti-CD81 antibody. After a 600-second baseline step in D-PBS(-) solution containing 0.1% BSA, the association and dissociation responses are measured, respectively, for 900 seconds and 1,500 seconds in the same buffer containing 20 nM antibody. The antibody binding capacity is expressed as a wavelength change between the baseline step and the end of the association phase. Measurements are performed at 30°C and a vibration rate of 1,000 Hz during the baseline step, association, and dissociation phases.

[0639] In the optimization of the prediction model, MV-TPE and mean-TPE are performed as described in the first validation example. In this validation, the 48 sequences with the highest ranking in terms of the estimated objective function among the candidate sequences obtained through sampling are selected as proposed sequences for evaluation.

[0640] Figure 19 shows the distribution and standard deviation of the mean of the estimated values ​​obtained by sampling in each process for MV-TPE and mean-TPE, respectively. The standard deviation represents uncertainty, and a higher mean of the estimated value indicates a better binary octet value. According to Figure 19, the pattern value of the standard deviation is 1.0 when using the mean as the objective function, and 0.3 when using MV. The pattern value of the mean is 2.3 when using the mean as the objective function, and 1.5 when using MV. These results show that the uncertainty of the binary octet value as a feature prediction is higher when using the mean than when using MV as the objective function.

[0641] Figure 20 illustrates a visualization of t-SNE (t-distributed random neighborhood embedding) of sequences obtained through sampling. T-SNE is a technique in which high-dimensional data is compressed to low dimensions for visualization. In Figure 20, the distributions representing the vector representations of individual sequences based on TAPE are shown on two-dimensional planes for the training data (Train), the mean (Mean), and MV, respectively. According to Figure 20, the sampled sequences using the mean as the objective function differ significantly from the training data, while the sampled sequences using MV are closer to the training data. This shows that, compared to mean-TPE, MV-TPE can suppress uncertainty to a greater extent while searching for safer regions, and MV-TPE avoids ill-conditioned samples.

[0642] Figure 21 shows the estimated mean (GP Mean) of the binary octet values ​​of the 48 sequences to be evaluated when the objective function of TPE is the mean and MV, respectively. According to Figure 21, the mean of the binary octet values ​​obtained using MV is lower than the mean of the binary octet values ​​obtained using the mean as the objective function.

[0643] Figure 22 shows the standard deviation (GP Std) of the estimates of the mean and MV, respectively. According to... Figure 22 The standard deviation of the binary octet obtained using MV is lower than that obtained using the mean as the objective function. Based on these results, it can be inferred that for sequences sampled using MV, stable expression and binding capacity are unlikely to deteriorate due to low uncertainty.

[0644] Figure 23 shows the expression level distribution of the 48 sequences to be evaluated for the mean and MV, respectively. According to Figure 23, when the mean is used as the objective function, the expression levels of the top 48 sequences are extremely low, and there is not enough sample size to measure the binary octet values. On the other hand, when MV is used as the objective function, almost all of the top 48 sequences are significantly expressed.

[0645] Figure 24 shows the distribution of binary octet values ​​for MV and the training data sequences, respectively. The distribution trends of MV and the training data in terms of binary octet values ​​are similar. This demonstrates that using MV as the objective function yields sequences comparable to the training data in terms of binding ability, and thus, parent sequences can be obtained. Furthermore, several sequences with binary octet values ​​higher than the maximum value in the training data were obtained using MV.

[0646] <Conclusion>

[0647] The above validation examples demonstrate that the offline model-based optimization (MBO) according to this application example can be applied to unit-based molecular design, a common practice in drug design, thereby enabling computer simulation (i.e., via a calculator). Here, the Tree-based Parzen Estimator (TPE), a technique for Bayesian optimization, can be applied to unit-based molecular design related to the optimization of combinations of categorical variables. It has also been shown that when applying offline MBO, the MV used in financial engineering helps to search for safer sequences by avoiding pathological behavior. That is, by incorporating risk aversion predictions into the MBO used for molecular design, overestimation of extrapolation regions when predicting the characteristics of candidate molecules can be prevented, thus preventing the proposal of molecules with low predictive reliability. Through the above validation examples, this embodiment has been shown to be usable for searching for safe sequences in the design of therapeutic antibodies.

[0648] The molecular design apparatus 1 of this embodiment, configured as described above, performs Bayesian optimization using a model configured to obtain characteristic quality values ​​based on the molecular sequence information of drug candidates as the objective function. Therefore, the burden accompanying drug discovery can be further reduced.

[0649] (Variation)

[0650] The molecular design device 1 can be assembled using multiple information processing devices communicatively connected via a network. In this case, the functional units of the molecular design device 1 can be distributed across multiple information processing devices. The drug discovery system 100 may include a factory (not shown) that generates the optimal molecule estimated by the molecular design device 1. The molecular design device 1 outputs information representing the biological sequence of the optimal molecule estimated by the above-described technique to this factory. The factory performs the step of generating a molecular compound having the biological sequence indicated by the output information input from the molecular design device 1.

[0651] All or part of the functionality of the molecular design device 1 can be implemented using hardware such as an ASIC (Application-Specific Integrated Circuit), a PLD (Programmable Logic Device), or a FPGA (Field-Programmable Gate Array). The program can be recorded on a computer-readable storage medium. For example, a "computer-readable storage medium" is a storage device such as a portable medium, such as a floppy disk, magneto-optical disk, ROM, and CD-ROM, as well as a hard disk built into a computer system. The program can be transmitted via a telecommunications line.

[0652] Molecular design device 1 is an example of an estimation device. This function can be achieved by another type of device, such as a device whose primary function is not molecular design, or an information device that includes a public computer system.

[0653] For example, this embodiment can be implemented as a system including a processor and memory. In this system, the memory is configured to store one or more commands, and the commands may be commands for causing the processor to perform the following processes: calculating a predicted feature value using a prediction model for a molecule that forms elements of a set of sequence information as a set of sequence information about a variety of different molecules, the prediction model being used to predict the features of the molecule from the sequence information about the molecule, and then estimating the uncertainty of the prediction; and a process of searching for molecules with more desired features based on the predicted feature value and the uncertainty of the prediction.

[0654] The command can be used to cause the processor to perform the following processes: obtaining a set of sequence information based on a combinatorial optimization algorithm, the set of sequence information being a collection of sequence information for multiple different molecules; using a prediction model to calculate the feature prediction values ​​of the molecules that form the elements of the set of sequence information, the prediction model being used to predict the features of the molecules from the sequence information about the molecules; updating the extraction parameters of the combinatorial optimization algorithm based on the feature prediction values ​​of each molecule to include molecules with more desired features; and searching for molecules with more desired features based on the feature prediction values.

[0655] This embodiment can be implemented as a non-transitory computer-readable medium for storing one or more commands. The command can be a command to cause a processor to perform the following processes: calculating a predicted feature value using a prediction model for molecules forming elements of a set of sequence information that is a set of sequence information about a variety of different molecules, the prediction model being used to predict the features of the molecule from the sequence information about the molecule, and then estimating the uncertainty of the prediction; and a process of searching for candidates of molecules with more desired features based on the predicted feature value and the uncertainty of the prediction. The command can also be a command to cause a computer to perform the following processes: obtaining a set of sequence information according to a combinatorial optimization algorithm, the set of sequence information about a variety of different molecules; a prediction process using a prediction model to calculate predicted feature values ​​for molecules forming elements of the set of sequence information, the prediction model being used to predict the features of the molecule from the sequence information about the molecule; an update process based on the predicted feature values ​​of each molecule to update the extraction parameters of the combinatorial optimization algorithm to include molecules with more desired features; and a search process based on the predicted feature values ​​to search for molecules with more desired features.

[0656] This embodiment can be implemented as a computer-based method using an artificial intelligence engine. The computer-based method involves processing a candidate set, which is a set of elements having sequence information about drug candidate molecules as elements. This sequence information pertains to the sequence of units constituting the drug candidate molecule. The method includes: a control step of estimating an optimal molecule by performing combinatorial optimization, the optimal molecule being the drug candidate molecule that gives the optimal value of the desired feature among the drug candidate molecules represented by the elements of the candidate set; and a step of outputting information about the optimal molecule. The objective function of the combinatorial optimization is used for: executing a mathematical model for estimating the eigenvalues ​​of the desired feature based on the sequence information about the drug candidate molecules and obtaining the uncertainty of the estimation result; and obtaining a feature quality value calculated based on the determination of the eigenvalues ​​and the uncertainty.

[0657] While some embodiments have been described above, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments may be implemented in various other forms and may be subject to various omissions, substitutions, and changes without departing from the spirit of the invention. These embodiments and variations thereof are within the scope of the invention as set forth in the claims and their equivalents, because they are within the scope and spirit of the invention. References cited herein are incorporated by reference.

[0658] [List of Reference Symbols]

[0659] 1... Molecular design device, 11... Controller, 12... Input terminal, 13... Communicator, 14... Memory, 15... Output terminal, 100... Drug discovery system, 111... Inferrer, 112... Input terminal controller, 113... Communication controller, 114... Memory controller, 115... Output terminal controller, 91... Processor, 92... Memory.

Claims

1. An information processing system for identifying molecules suitable for drug candidates, The information processing system includes: A feature predictor is configured to use a prediction model to calculate a feature prediction value for a molecule that forms an element of a set of structural unit combination information that is a set of structural unit combination information for a variety of different molecules, and then estimate the uncertainty of the prediction. The prediction model is used to predict the features of the molecule from the structural unit combination information of the molecule. as well as A candidate molecule discriminator is configured to search for candidates of molecules having the desired features based on the predicted value of the features and an estimate of the uncertainty of the prediction.

2. The information processing system according to claim 1, further comprising a predictive information processor, The prediction information processor is configured to calculate a feature quality value based on the predicted feature value and an estimate of the uncertainty of the prediction. The candidate molecule discriminator identifies at least one candidate molecule from the elements of the set of structural unit combination information based on the characteristic quality value.

3. The information processing system according to claim 2, wherein, The characteristic quality value is the output value of a predetermined function. In the predetermined function, two variables are used as inputs, one of which is the predicted value of the feature, and the other of which is an estimate of the uncertainty of the prediction.

4. The information processing system according to claim 2 or 3, wherein, The predictive information processor calculates the mean and variance as the objective function to give the feature quality value.

5. The information processing system according to any one of claims 2 to 4, in, The feature quality value increases as the feature prediction value increases, and decreases as the estimated uncertainty of the prediction decreases.

6. The information processing system according to any one of claims 2 to 5, comprising a structural unit combination information processor, the structural unit combination information processor being configured to use a combinatorial optimization algorithm to obtain the structural unit combination information set based on the feature prediction value or the feature quality value of the molecule.

7. The information processing system according to any one of claims 2 to 6, in, The feature quality value increases as the feature prediction value increases, and decreases as the estimated uncertainty of the prediction decreases. The set of structural unit combination information is obtained using a combinatorial optimization algorithm, and the extraction parameters of the combinatorial optimization algorithm are updated based on the feature prediction value or the feature quality value of the molecule.

8. The information processing system according to claim 7, wherein, The combined optimization algorithm is a tree-structured Parzen estimator.

9. The information processing system according to any one of claims 1 to 8, wherein, The prediction model is generated by learning based on the combination information of structural units of various molecules used for training and the results of evaluating the features of the molecules.

10. The information processing system according to any one of claims 1 to 9, wherein, The structural unit combination information is the sequence information of the molecule.

11. The information processing system according to any one of claims 1 to 10, wherein, The molecule is at least one of nucleic acids, peptides, cyclic peptides, proteins, antibodies, and low molecular weight compounds.

12. The information processing system according to any one of claims 1 to 11, wherein, The feature is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety.

13. The information processing system according to any one of claims 1 to 12, wherein, The molecule is a molecule that binds to the target molecule, and the feature is the ability to bind to the target molecule.

14. An information processing method, performed in an information processing system for identifying molecules suitable for drug candidates, the method comprising the following steps: Regarding a molecule that forms an element of a set of structural unit combination information that is a set of information about combinations of structural units of various different molecules, a prediction model is used to calculate the feature prediction value of the molecule, and then the uncertainty of the prediction is estimated. The prediction model is used to predict the features of the molecule from the structural unit combination information of the molecule. as well as Candidate molecules with more desirable features are searched based on the predicted feature values ​​and the uncertainty of the predictions.

15. A program configured to cause a computer to perform the following procedures: Regarding molecules that form elements of a set of structural unit combination information that is a set of information about combinations of structural units of various different molecules, a prediction model is used to calculate the predicted feature values ​​of the molecule, and then the uncertainty of the prediction is estimated. The prediction model is used to predict the features of the molecule from the structural unit combination information of the molecule; and Candidate molecules with more desirable features are searched based on the predicted feature values ​​and the uncertainty of the predictions.

16. A method for manufacturing a molecular compound, the method comprising: The input steps involve storing and retrieving a set of structural unit combination information, and inputting the set of structural unit combination information into the prediction model. The set of structural unit combination information is a collection of structural unit combination information for various molecules. The inference step involves searching for molecules with more desirable features from the set of structural unit combination information based on the feature prediction values ​​of each molecule included in the set of structural unit combination information output from the prediction model and the estimated value of the prediction uncertainty, and then identifying them as candidate molecules. The output step outputs information about the structural unit combination of the candidate molecule; as well as The generation step generates the molecular compound having a molecular sequence represented by the combination information of the structural units.

17. An information processing system for identifying molecules suitable for drug candidates, said information processing system comprising: A structural unit combination information processor, wherein the structural unit combination information processor obtains a set of structural unit combination information according to a combination optimization algorithm, wherein the set of structural unit combination information is a collection of structural unit combination information of various molecules; A feature predictor, wherein the feature predictor uses a prediction model to calculate feature prediction values ​​for a molecule with respect to the elements forming the set of structural unit combination information, the prediction model being used to predict features of the molecule from the structural unit combination information of the molecule; and A candidate molecule discriminator searches for candidates of molecules possessing desired features based on the predicted feature values. The structural unit combined information processor The extraction parameters of the combinatorial optimization algorithm are updated based on the predicted feature values ​​of the molecules to include molecules with more desirable features.

18. An information processing method, performed in an information processing system for identifying molecules suitable for drug candidates, the method comprising the following steps: The structural unit combination information set is obtained according to the combinatorial optimization algorithm. The structural unit combination information set is a collection of structural unit combination information of various molecules. For molecules that form the set of structural unit combination information, a prediction model is used to calculate the feature prediction value of the molecule, the prediction model being used to predict the features of the molecule from the structural unit combination information of the molecule; Based on the predicted feature values, search for molecules with more desirable features; as well as The extraction parameters of the combinatorial optimization algorithm are updated based on the predicted feature values ​​of the molecules to include molecules with more desirable features.

19. A program configured to cause a computer to execute: The acquisition process involves obtaining a set of structural unit combination information based on a combinatorial optimization algorithm. This set of structural unit combination information is a collection of structural unit combination information for various molecules. The prediction process uses a prediction model to calculate the feature prediction values ​​of molecules that form the set of structural unit combination information. The prediction model is used to predict the features of the molecules from the structural unit combination information of the molecules. The update process updates the extraction parameters of the combined optimization algorithm based on the predicted feature value of each molecule, so as to include molecules with more desirable features. as well as The search process searches for molecules with the desired features based on the predicted feature values.

20. A method for manufacturing a molecular compound, the method comprising: The acquisition step involves obtaining a set of structural unit combination information based on a combinatorial optimization algorithm. This set of structural unit combination information is a collection of structural unit combination information for various molecules. The input step involves inputting the set of structural unit combination information into the prediction model; The update step involves updating the extraction parameters of the combinatorial optimization algorithm based on the feature prediction values ​​of each molecule included in the set of structural unit combination information output from the prediction model, so as to include molecules with more desirable features. The step of searching for molecules with more desired features from the set of structural unit combination information based on the predicted feature values, and then identifying them as candidate molecules; The output step outputs information about the structural unit combination of the candidate molecule; as well as The generation step generates the molecular compound having a molecular sequence represented by the combination information of the structural units.

Images