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

Abstract

This information processing system for identifying molecules that are suitable for a drug candidate comprises: a characteristic prediction unit that, for molecules constituting elements of a sequence information set that is a set of sequence information for a plurality of different molecules, calculates a characteristic prediction value for each molecule using a prediction model for predicting the characteristics of the molecules from the sequence information for the molecules, and estimates the uncertainty of the prediction; and a candidate molecule identification unit that searches for candidates for molecules having desired characteristics on the basis of the characteristic prediction value and an estimated value of the uncertainty of the prediction.

Description

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

[0001] One aspect of the present disclosure relates to an information processing system, a molecular design device, an information processing method, a program, and a method for producing a molecular compound.

[0002] In recent years, machine learning information processing technology has been utilized in the pharmaceutical field to reduce the burden of drug discovery (Patent Document 1). Attempts are being made to efficiently discover molecules with properties required as drugs using predictions made by machine learning.

[0003] International Publication No. 2018 / 132752

[0004] Since predictions based on machine learning involve uncertainty, using machine learning predictions without taking this into consideration can lead to unexpected results. In addition, there are many targets that need to be analyzed in drug discovery, and it is difficult to apply machine learning predictions to all targets that are feasible.

[0005] In view of the above circumstances, one aspect of the present disclosure aims to provide a technology that further reduces the burden required for drug discovery.

[0006] One aspect of the present disclosure provides the following: [A1] An information processing system for identifying molecules suitable as drug candidates, comprising: a property prediction unit that calculates a predicted property value of a molecule that is an element of a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, using a prediction model for predicting the property of the molecule from the building block combination information of the molecule and estimates the uncertainty of the prediction; and a candidate molecule identification unit that searches for candidate molecules having desired properties based on the predicted property value and the estimated uncertainty of the prediction. [A2] The information processing system described in A1, wherein the building block combination information is the molecular sequence information. [A3] The information processing system described in A1 or A2, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound. [A4] The information processing system described in any of A1 to A3, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is amino acid sequence information. [A5] The information processing system of any one of A1 to A4, wherein the characteristic is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety. [A6] The information processing system of any one of A1 to A5, wherein the molecule is a molecule that binds to a target molecule, and the characteristic is binding ability to the target molecule. [A7] The information processing system of any one of A1 to A6, further comprising a prediction information processing unit that calculates a characteristic quality value based on the characteristic prediction value and an estimated value of uncertainty of the prediction, wherein the candidate molecule identifying unit identifies at least one candidate molecule from elements of the building block combination information set based on the characteristic quality value. [A8] The information processing system of any one of A1 to A7, wherein the characteristic quality value increases as the characteristic prediction value increases and decreases as the estimated value of uncertainty of the prediction decreases.[A9] The information processing system of A8, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the characteristic prediction value and an estimate of the prediction uncertainty. [A10] The information processing system of any of A7 to A9, wherein the prediction information processing unit calculates a mean variance as an objective function that gives the characteristic quality value. [A11] The information processing system of any of A1 to A10, wherein the estimate of the prediction uncertainty is the standard deviation of the characteristic prediction value. [A12] The information processing system of any of A1 to A11, wherein the characteristic quality value increases as the characteristic prediction value increases, and the building block combination information set is obtained using a combinatorial optimization algorithm as the estimate of the prediction uncertainty decreases, and extraction parameters of the combinatorial optimization algorithm are updated based on the characteristic prediction value of the molecule or the characteristic quality value. [A13] The information processing system of A12, wherein the combinatorial optimization algorithm is a tree-structured Parzen estimator. [A14] 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 building block combination information of a plurality of molecules for training and results of property evaluation of the molecules.

[0007] [B1] A molecular design device comprising a control unit that infers properties of a molecule from building block combination information of the molecule using a predetermined prediction model, wherein the control unit: acquires a building block combination information set that is a collection of building block combination information of a plurality of different molecules; calculates a predicted property value of a molecule that is an element of the building block combination information set from the building block combination information of the molecule using the prediction model and estimates the uncertainty of the prediction; and searches for candidate molecules having desired properties based on the predicted property value and the estimated uncertainty of the prediction. [B2] The molecular design device according to B1, wherein the building block combination information is sequence information of the molecule. [B3] The molecular design device according to B1 or B2, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a small molecule compound. [B4] The molecular design device according to any one of B1 to B3, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is information on an amino acid sequence. [B5] The molecular design device according to any one of B1 to B4, wherein the characteristic is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety. [B6] The molecular design device according to any one of B1 to B5, wherein the molecule is a molecule that binds to a target molecule, and the characteristic is binding ability to the target molecule. [B7] The molecular design device according to any one of B1 to B6, wherein the control unit further calculates a characteristic quality value based on the characteristic prediction value and an estimated value of uncertainty of the prediction, and identifies at least one candidate molecule from elements of the building block combination information set based on the characteristic quality value. [B8] The molecular design device according to any one of B7, wherein the characteristic quality value increases as the characteristic prediction value increases and decreases as the estimated value of uncertainty of the prediction decreases. [B9] The molecular design device according to B8, wherein the characteristic quality value is an output value of a predetermined function that takes as input two variables, the predicted characteristic value and an estimated value of the uncertainty of the prediction.[B10] The molecular design device according to any one of B7 to B9, wherein the control unit calculates a mean variance as an objective function that provides the characteristic quality value. [B11] The molecular design device according to any one of B1 to B10, wherein the estimated value of prediction uncertainty is the standard deviation of the characteristic prediction value. [B12] The molecular design device according to any one of B1 to B11, wherein the control unit obtains the building block combination information set using a combinatorial optimization algorithm, and updates extraction parameters of the combinatorial optimization algorithm based on the characteristic prediction value or the characteristic quality value. [B13] The molecular design device according to any one of B12, wherein the combinatorial optimization algorithm is a tree-structured Parzen estimator. [B14] The molecular design device according to any one of B1 to B13, further comprising an output unit that outputs building block combination information of the candidate molecule. [B15] The information processing system according to any one of B1 to B14, wherein the prediction model is a prediction model generated by learning based on building block combination information of a plurality of molecules for training and results of property evaluation of the molecules.

[0008] [C1] An information processing system for identifying molecules suitable as drug candidates, comprising: a building block combination information processing unit that acquires a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, in accordance with a combinatorial optimization algorithm; a property prediction unit that calculates a predicted property value of a molecule that is an element of the building block combination information set, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule; and a candidate molecule identification unit that searches for candidate molecules having desired properties based on the predicted property value, wherein the building block combination information processing unit updates extraction parameters of the combinatorial optimization algorithm based on the predicted property value of the molecule so that molecules having more desired properties are included. [C2] The information processing system of C1, wherein the candidate molecule identifying unit searches for candidate molecules having desired properties based on a first predicted property value calculated for molecules that are elements of a first building block combination information set, which is the building block combination information set obtained before the extraction parameters were updated, and a second predicted property value calculated for molecules that are elements of a second building block combination information set, which is the building block combination information set obtained after the extraction parameters were updated. [C3] The information processing system of C2, wherein the first building block combination information set and the second building block combination information set are sets including building block combination information of different molecules. [C4] The information processing system of any of C1 to C3, wherein the building block combination information is sequence information of the molecule. [C5] The information processing system of any of C1 to C4, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound.[C6] 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 building block combination information is information on an amino acid sequence. [C7] The information processing system according to any one of C1 to C6, wherein the property is at least one property selected from the group consisting of binding ability, pharmacological activity, physical properties, kinetics, and safety. [C8] 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 property is binding ability to the target molecule. [C9] The information processing system according to any one of C1 to C8, wherein the property prediction unit calculates the predicted property value and an estimated value of uncertainty of the prediction. [C10] The information processing system according to any one of C1 to C9, wherein the building block combination information processing unit updates extraction parameters of a combinatorial optimization algorithm so as to include molecules having more desirable properties, based on the predicted property value of the molecule and the estimated value of uncertainty of the prediction. [C11] The information processing system according to C9 or C10, further comprising a prediction information processing unit that processes prediction information output from the property prediction unit, the prediction information processing unit including the property prediction value and / or an estimate of the prediction uncertainty, wherein the prediction information processing unit calculates a property quality value based on the property prediction value and the estimate of the prediction uncertainty, and the candidate molecule identification unit identifies at least one molecule from the building block combination information set based on the property quality value. [C12] The information processing system according to C11, wherein the property quality value increases as the property prediction value increases and decreases as the estimate of the prediction uncertainty decreases. [C13] The information processing system according to C12, wherein the property quality value is the output value of a predetermined function that takes as input two variables, the property prediction value and the estimate of the prediction uncertainty. [C14] The information processing system according to any of C11 to C13, wherein the prediction information processing unit calculates a mean variance as an objective function that gives the property quality value. [C15] The information processing system according to any one of C9 to C14, wherein the estimated value of the uncertainty of the prediction is a standard deviation of the predicted characteristic value.[C16] The information processing system according to any one of C1 to C15, wherein a tree-structured Parzen estimator is used as the combinatorial optimization algorithm.

[0009] [D1] A molecular design device comprising a control unit that infers the properties of a molecule from building block combination information of the molecule using a predetermined prediction model, wherein the control unit: obtains a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; calculates the property prediction values ​​for molecules that are elements of the building block combination information set using the prediction model; updates extraction parameters of the combinatorial optimization algorithm based on the property prediction values ​​for each molecule obtained by the prediction model so that molecules having more desired properties are included; and searches for candidate molecules having desired properties based on the property prediction values. [D2] The molecular design device according to D1, wherein the control unit: acquires a first building block combination information set which is a set of building block combination information of a plurality of different molecules before updating the extraction parameters, calculates a first predicted property value for molecules which are elements of the first building block combination information set using the prediction model, further acquires a second building block combination information set which is a set of building block combination information of a plurality of different molecules after updating the extraction parameters, calculates a second predicted property value for molecules which are elements of the second building block combination information set using the prediction model, and searches for candidate molecules having desired properties based on the first predicted property value and the second predicted property value. [D3] The molecular design device according to D2, wherein the first building block combination information set and the second building block combination information set are sets including sequence information of molecules which are different from each other. [D4] The molecular design device according to any of D1 to D3, wherein the building block combination information is sequence information of the molecules. [D5] The molecular design apparatus according to any one of D1 to D5, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound.[D6] The molecular design device according to any one of D1 to D5, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is information on an amino acid sequence. [D7] The molecular design device according to any one of D1 to D6, wherein the property is at least one property selected from the group consisting of binding ability, pharmacological activity, physical properties, kinetics, and safety. [D8] The molecular design device according to any one of D1 to D7, wherein the molecule is a molecule that binds to a target molecule, and the property is binding ability to the target molecule. [D9] The molecular design device according to any one of D1 to D8, wherein the control unit calculates the predicted property value and further calculates an estimated value of the uncertainty of the prediction. [D10] The information processing system according to any one of D1 to D9, wherein the control unit acquires the building block combination information set so as to include molecules having more desirable properties based on the predicted property value of the molecule and the estimated value of the uncertainty of the prediction. [D11] The molecular design device according to D10, wherein the control unit further calculates a characteristic quality value based on the predicted characteristic value and an estimated value of the prediction uncertainty, and the candidate molecule identification unit identifies at least one molecule from the building block combination information set based on the characteristic quality value. [D12] The molecular design device according to D11, wherein the characteristic quality value increases as the predicted characteristic value increases and decreases as the estimated value of the prediction uncertainty decreases. [D13] The molecular design device according to D12, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the predicted characteristic value and the estimated value of the prediction uncertainty. [D14] The molecular design device according to any of D11 to D13, wherein the control unit calculates a mean variance as an objective function that gives the characteristic quality value. [D15] The molecular design device according to any of D9 to D14, wherein the estimated value of the prediction uncertainty is the standard deviation of the predicted characteristic value. [D16] The molecular design device according to any one of D1 to D15, wherein a tree-structured Parzen estimator is used as the combinatorial optimization algorithm.[D17] The molecular design device according to any one of D1 to D16, wherein the prediction model is a model trained based on building block combination information that differs for each molecule and training data that indicates the properties of the molecule.

[0010] [E1] An information processing method in an information processing system that identifies molecules suitable as drug candidates, the information processing method comprising the steps of: calculating a predicted property value of a molecule that is an element of a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule, and estimating the uncertainty of the prediction; and searching for candidate molecules having more desirable properties based on the predicted property value and the uncertainty of the prediction. [E2] The information processing method described in E1, wherein the building block combination information is information on the sequence of the molecule. [E3] The information processing method described in E1 or E2, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound. [E4] The information processing method described in any of E1 to E3, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is information on an amino acid sequence. [E5] The information processing method according to any one of E1 to E4, wherein the characteristic is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety. [E6] 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 characteristic is binding ability to the target molecule. [E7] The information processing method according to any one of E1 to E6, further comprising the step of calculating a characteristic quality value based on the predicted characteristic value and an estimated value of uncertainty of the prediction, wherein the searching step includes the step of identifying at least one molecule from the building block combination information set based on the characteristic quality value. [E8] The information processing method according to E7, wherein the characteristic quality value increases as the predicted characteristic value increases and decreases as the estimated value of uncertainty of the prediction decreases. [E9] The information processing method according to E8, wherein the characteristic quality value is the output value of an arbitrary function that takes as input two variables, the predicted characteristic value and the estimated value of uncertainty of the prediction.[E10] The information processing method according to any one of E7 to E9, wherein the step of calculating the characteristic quality value calculates a mean variance as an objective function that gives the characteristic quality value. [E11] The information processing method according to any one of E1 to E10, wherein the estimated value of the uncertainty of the prediction is the standard deviation of the predicted value. [E12] The information processing method according to any one of E1 to E11, comprising a step of outputting building block combination information of the identified candidate molecule. [E13] The information processing method according to E12, further outputting information on the predicted property value of the identified molecule. [E14] The information processing method according to any one of E7 to E13, wherein the searching step comprises a step of identifying molecules whose characteristic quality value is equal to or greater than a predetermined value from the building block combination information set. [E15] The information processing method according to any one of E7 to E13, wherein the searching step comprises a step of identifying a ranking of the characteristic quality value for each molecule and identifying molecules within the predetermined ranking. [E16] The information processing method according to any one of E1 to E15, wherein the searching step includes a step of selecting at least one molecule from the building block combination information set, the molecule having the property prediction value and the prediction uncertainty each satisfying predetermined conditions. [E17] The information processing method according to any one of E1 to E16, wherein the building block combination information set is obtained using a combinatorial optimization algorithm, and extraction parameters of the combinatorial optimization algorithm are updated based on the property prediction value or the property quality value. [E18] The information processing method according to any one of E1 to E18, wherein the combinatorial optimization algorithm is a tree-structured Parzen estimator. [E19] The information processing method according to any one of E1 to E18, wherein the prediction model is a model trained based on building block combination information that differs for each molecule and training data indicating the properties of the molecules.

[0011] [F1] A program for causing a computer to execute the following steps: for a molecule that is an element of a building block combination information set, which is a set of building block combination information of a plurality of different molecules, calculating a predicted property value of the molecule using a prediction model for predicting the property of the molecule from the building block combination information of the molecule and estimating the uncertainty of the prediction; and searching for candidate molecules having more desirable properties based on the predicted property value and the uncertainty of the prediction. [F2] The program described in F1, wherein the building block combination information is sequence information of the molecule. [F3] The program described in F1 or F2, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound. [F4] The program described in any of F1 to F3, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is amino acid sequence information. [F5] The program according to any one of F1 to F4, wherein the characteristic is at least one of binding affinity, pharmacological activity, physical properties, kinetics, and safety. [F6] The program according to any one of F1 to F5, wherein the molecule is a molecule that binds to a target molecule, and the characteristic is binding affinity to the target molecule. [F7] The program according to any one of F1 to F6, further comprising a step of calculating a characteristic quality value based on the predicted characteristic value and an estimated value of uncertainty of the prediction, wherein the searching step includes a step of identifying at least one molecule from the building block combination information set based on the characteristic quality value. [F8] The program according to F7, wherein the characteristic quality value increases as the predicted characteristic value increases and decreases as the estimated value of uncertainty of the prediction decreases. [F9] The program according to F8, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the predicted characteristic value and the estimated value of uncertainty of the prediction.[F10] The program according to any one of F7 to F9, wherein the step of calculating the characteristic quality value calculates a mean variance as an objective function that gives the characteristic quality value. [F11] The program according to any one of F1 to F10, wherein the estimated value of uncertainty is the standard deviation of the predicted value. [F12] The program according to any one of F1 to F11, including a step of searching for and outputting building block combination information of identified candidate molecules having more desirable properties. [F13] The program according to F12, further outputting information on predicted property values ​​of identified candidate molecules having more desirable properties. [F14] The program according to any one of F7 to F13, wherein the step of searching includes a step of identifying molecules having the characteristic quality value equal to or greater than a predetermined value from the building block combination information set. [F15] The program according to any one of F7 to F13, wherein the step of searching includes a step of identifying a ranking of the characteristic quality value for each molecule and identifying molecules within the predetermined ranking. [F16] The program according to any one of F1 to F15, wherein the searching step includes a step of selecting at least one molecule from the building block combination information set, the molecule having the property prediction value and the estimated value of the prediction uncertainty each satisfying predetermined conditions. [F17] The program according to any one of F1 to F16, further including a step of updating extraction parameters for obtaining the building block combination information set based on the property prediction value or the property quality value of the molecule using a combinatorial optimization algorithm. [F18] The program according to F17, wherein the combinatorial optimization algorithm is a tree-structured Parzen estimator. [F19] The program according to any one of F1 to F18, wherein the prediction model is a model trained based on building block combination information that differs for each molecule and training data indicating the properties of the molecules.

[0012] [G1] A method for producing a molecular compound, comprising: an input step of accessing a building block combination information set, which is a collection of building block combination information for a plurality of different molecules, and inputting the building block combination information set into a prediction model; an inference step of searching for a molecule having more desired properties from the building block combination information set based on a predicted property value and an estimated value of prediction uncertainty for each molecule included in the building block combination information set, which are output from the prediction model; an output step of outputting building block combination information for the candidate molecule; and a generation step of generating the molecular compound having a molecular sequence indicated in the building block combination information. [G2] The method described in G1, wherein the candidate molecule has a biological sequence, and the building block combination information is sequence information. [G3] The method described in G1 or G2, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a small molecule compound. [G4] The method according to any one of G1 to G3, wherein the molecule is a protein, antibody, peptide, or cyclic peptide, and the building block combination information is amino acid sequence information. [G5] The method according to any one of G1 to G4, wherein the property is at least one property selected from the group consisting of binding ability, pharmacological activity, physical properties, kinetics, and safety. [G6] The method according to any one of G1 to G5, wherein the molecule is a molecule that binds to a target molecule, and the property is binding ability to the target molecule. [G7] The method according to any one of G1 to G6, further comprising the steps of calculating a property quality value based on the property prediction value and an estimated value of the uncertainty of the prediction, and identifying at least one candidate molecule from the building block combination information set based on the property quality value. [G8] The method according to G7, wherein the property quality value increases as the property prediction value increases and decreases as the estimated value of the uncertainty of the prediction decreases.[G9] The method described in G8, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the predicted characteristic value and an estimate of the uncertainty of the prediction. [G10] The method described in any of G7 to G9, wherein the step of calculating the characteristic quality value calculates a mean variance as an objective function that gives the characteristic quality value. [G11] The method described in any of G1 to G10, wherein the estimate of uncertainty is the standard deviation of the predicted values. [G12] The method described in any of G1 to G11, wherein the information on the candidate molecule includes building block combination information of the candidate molecule. [G13] The method described in any of G1 to G12, wherein the information on the candidate molecule includes a predicted characteristic value of the candidate molecule. [G14] The method described in any of G7 to G13, wherein the step of searching for candidate molecules includes a step of identifying, from the building block combination information set, molecules whose characteristic quality value is equal to or greater than a predetermined value as the candidate molecule. [G15] The method of any one of G7 to G13, wherein the step of searching for candidate molecules includes a step of identifying a ranking of the property quality value for each molecule and identifying molecules within the predetermined ranking as the candidate molecule. [G16] The method of any one of G1 to G15, wherein the step of searching for candidate molecules includes a step of selecting at least one molecule from the building block combination information set, the property predicted value and the estimated value of the prediction uncertainty each satisfy predetermined conditions. [G17] The method of any one of G1 to G16, wherein the building block combination information set is obtained using a combinatorial optimization algorithm, and further comprising a step of updating extraction parameters of the combinatorial optimization algorithm based on the property predicted value or the property quality value of the molecule. [G18] The method of G17, wherein the combinatorial optimization algorithm is a tree-structured Parzen estimator. [G19] The method of any one of G1 to G18, comprising a selection step of selecting a candidate molecule having the best experimental value from experimental values ​​of the properties of each of a plurality of candidate molecules.

[0013] [H1] An information processing method in an information processing system for identifying molecules suitable as drug candidates, comprising the steps of: acquiring a building block combination information set, which is a collection of building block combination information for a plurality of different molecules, according to a combinatorial optimization algorithm; calculating a predicted property value for a molecule that is an element of the building block combination information set, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule; searching for a molecule having more desirable properties based on the predicted property value; and updating extraction parameters of the combinatorial optimization algorithm so that molecules having more desirable properties are included, based on the predicted property value of the molecule. [H2] The information processing method according to H1, comprising a step of searching for candidate molecules having desired properties based on a first predicted property value calculated for molecules that are elements of a first building block combination information set, which is the building block combination information set obtained before the extraction parameters are updated, and a second predicted property value calculated for molecules that are elements of a second building block combination information set, which is the building block combination information set obtained after the extraction parameters are updated. [H3] The information processing method according to H2, wherein the first building block combination information set and the second building block combination information set are sets including building block combination information of different molecules. [H4] The information processing method according to any of H1 to H3, wherein the building block combination information is information on the sequence of the molecule. [H5] The information processing system according to any of H1 to H4, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound. [H6] The information processing method according to any one of H1 to H5, wherein the molecule is a protein, an antibody, a peptide, or a cyclic peptide, and the building block combination information is information on an amino acid sequence.[H7] The information processing method according to any one of H6, wherein the property is at least one property selected from the group consisting of binding ability, pharmacological activity, physical properties, kinetics, and safety. [H8] 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 property is binding ability to the target molecule. [H9] The information processing method according to any one of H1 to H8, comprising a step of calculating the property prediction value and an estimate of the uncertainty of the prediction. [H10] The information processing method according to H9, comprising a step of calculating a property quality value based on the property prediction value and the estimate of the uncertainty of the prediction, wherein the searching step includes a step of identifying at least one molecule from the building block combination information set based on the property quality value. [H11] The information processing method according to H10, wherein the property quality value increases as the property prediction value increases and decreases as the estimate of the uncertainty of the prediction decreases. [H12] The information processing method according to H11, wherein the characteristic quality value is an output value of a predetermined function having two input variables, the characteristic predicted value and the prediction uncertainty. [H13] The information processing method according to any of H10 to H12, wherein the step of calculating the characteristic quality value includes a step of calculating a mean variance as an objective function that gives the characteristic quality value. [H14] The information processing method according to any of H9 to H13, wherein the estimated value of the prediction uncertainty is the standard deviation of the characteristic predicted value. [H15] The information processing method according to any of H1 to H14, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm for updating the extraction parameters.

[0014] [I1] A program for causing a computer to execute the following steps: an acquisition procedure for acquiring a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; a prediction procedure for calculating predicted property values ​​of molecules that are elements of the building block combination information set, using a prediction model for predicting the properties of a molecule from the building block combination information of the molecule; an update procedure for updating extraction parameters of the combinatorial optimization algorithm, based on the predicted property values ​​of each molecule, so that molecules with more desired properties are included; and a search procedure for searching for molecules with desired properties based on the predicted property values. [I2] The program described in I1, wherein the acquisition step includes: a step of acquiring a first building block combination information set which is a set of building block combination information of a plurality of different molecules before updating the extraction parameters; and a step of further acquiring a second building block combination information set which is a set of building block combination information of a plurality of different molecules after updating the extraction parameters; the prediction step includes: a step of calculating a first predicted property value for molecules which are elements of the first building block combination information set using the prediction model; and a step of calculating a second predicted property value for molecules which are elements of the second building block combination information set using the prediction model; and the search step includes: a search step of searching for candidate molecules having desired properties from the first building block combination information set and the second building block combination information set based on the first predicted property value and the second predicted property value. [I3] The program according to I2, wherein the first building block combination information set and the second building block combination information set are sets containing building block combination information of different molecules. [I4] The program according to any one of I1 to I3, wherein the building block combination information is information on the sequence of the molecule.[I5] The program described in any one of I1 to I4, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound. [I6] The program described in any one of I1 to I5, wherein the molecule is a protein, an antibody, a peptide, or a cyclic peptide, and the building block combination information is amino acid sequence information. [I7] The program described in any one of I1 to I6, wherein the property is at least one of binding ability, pharmacological activity, physical properties, kinetics, and safety. [I8] The program described in any one of I1 to I7, wherein the molecule is a molecule that binds to a target molecule, and the property is binding ability to the target molecule. [I9] The program described in any one of I1 to I8, comprising the steps of calculating the property prediction value and calculating an estimate of the uncertainty of the prediction. [I10] The program described in I9, further comprising a step of calculating a characteristic quality value based on the predicted characteristic value and an estimated value of the prediction uncertainty, wherein the search step comprises a step of identifying at least one molecule from the building block combination information set based on the characteristic quality value. [I11] The program described in I10, wherein the characteristic quality value increases as the predicted characteristic value increases and decreases as the estimated value of the prediction uncertainty decreases. [I12] The program described in I11, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the predicted characteristic value and the estimated value of the prediction uncertainty. [I13] The program described in any one of I10 to I12, wherein the step of calculating the characteristic quality value calculates a mean variance as an objective function that gives the characteristic quality value. [I14] The program described in any one of I9 to I13, wherein the estimated value of the prediction uncertainty is the standard deviation of the predicted characteristic value. [I15] The program according to any one of I1 to I14, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm for updating the building block combination information set.[I16] The program according to any one of I1 to I15, wherein the prediction model is a model trained based on building block combination information that differs for each molecule and training data that indicates the properties of the molecule.

[0015] [J1] A method for producing a molecular compound, comprising: an acquisition step of acquiring a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; an input step of inputting the building block combination information set into a prediction model; an update step of updating extraction parameters of the combinatorial optimization algorithm based on a property prediction value for each molecule included in the building block combination information set, which is output from the prediction model, so that molecules having more desired properties are included; a step of searching for a molecule having more desired properties from the building block combination information set based on the property prediction value, and identifying it as a candidate molecule; an output step of outputting building block combination information related to the candidate molecule; and a generation step of generating the molecular compound having a molecular sequence indicated in the building block combination information. [J2] The method of J1, comprising a step of searching for the candidate molecule based on a first predicted property value calculated for a molecule that is an element of a first building block combination information set, which is the building block combination information set obtained before the extraction parameters are updated, and a second predicted property value calculated for a molecule that is an element of a second building block combination information set, which is the building block combination information set obtained after the extraction parameters are updated. [J3] The method of J2, wherein the first building block combination information set and the second building block combination information set are sets containing building block combination information of different molecules. [J4] The method of any of J1 to J3, wherein the candidate molecule has a biological sequence. [J5] The method of any of J1 to J4, wherein the building block combination information is sequence information of the molecule.[J6] The method according to any one of J1 to J5, wherein the molecule is at least one of a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a small molecule compound. [J7] The method according to any one of J1 to J6, wherein the molecule is a protein, an antibody, a peptide, or a cyclic peptide, and the building block combination information is amino acid sequence information. [J8] The method according to any one of J1 to J7, wherein the property is at least one property selected from the group consisting of binding ability, pharmacological activity, physical properties, kinetics, and safety. [J9] The method according to any one of J1 to J8, wherein the molecule is a molecule that binds to a target molecule, and the property is binding ability to the target molecule. [J10] The method according to any one of J1 to J9, comprising the steps of calculating the property prediction value and an estimate of the uncertainty of the prediction. [J11] The method of J10, comprising a step of calculating a characteristic quality value based on the predicted characteristic value and an estimated value of the prediction uncertainty, wherein the inference step of searching for the candidate molecule includes a step of identifying at least one molecule from the building block combination information set as the candidate molecule based on the characteristic quality value. [J12] The method of J11, wherein the characteristic quality value increases as the predicted characteristic value increases and decreases as the estimated value of the prediction uncertainty decreases. [J13] The method of J12, wherein the characteristic quality value is the output value of a predetermined function that takes as input two variables, the predicted characteristic value and the estimated value of the prediction uncertainty. [J14] The method of any of J11 to J13, wherein the step of calculating the characteristic quality value includes a step of calculating a mean variance as an objective function that gives the characteristic quality value. [J15] The method of any of J10 to J14, wherein the estimated value of the prediction uncertainty is the standard deviation of the predicted characteristic value. [J16] The method according to any one of J1 to J15, wherein a tree-structured Parzen estimator is used as a combinatorial optimization algorithm for updating the building block combination information set.[J17] The method according to any one of J1 to J16, comprising a step of selecting a candidate molecule having the best experimental value from among experimental values ​​of the properties of each of a plurality of candidate molecules.

[0016] [K] A computer system comprising a processor and a memory, wherein the memory is configured to store one or more instructions that cause the processor to: calculate a predicted property value of a molecule that is an element of a building block combination information set, which is a set of building block combination information of a plurality of different molecules, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule, estimate the uncertainty of the prediction; and search for candidate molecules having more desirable properties based on the predicted property value and the uncertainty of the prediction.

[0017] [L] A non-transitory computer-readable storage medium storing one or more instructions that cause a computer to: calculate a predicted property value of a molecule that is an element of a building block combination information set, which is a set of building block combination information of a plurality of different molecules, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule, estimate the uncertainty of the prediction; and search for candidate molecules that have more desirable properties based on the predicted property value and the uncertainty of the prediction.

[0018] [M] A computer system comprising a processor and a memory, the memory being configured to store one or more instructions, which cause the processor to: obtain a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; calculate predicted property values ​​of molecules that are elements of the building block combination information set using a prediction model for predicting the properties of a molecule from the building block combination information of the molecule; update extraction parameters of the combinatorial optimization algorithm based on the predicted property values ​​for each molecule so that molecules with more desired properties are included; and search for molecules with desired properties based on the predicted property values.

[0019] [N] A non-transitory computer-readable storage medium storing one or more instructions, which cause a computer to: obtain a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; calculate predicted property values ​​of molecules that are elements of the building block combination information set using a prediction model for predicting the properties of a molecule from the building block combination information of the molecule; update extraction parameters of the combinatorial optimization algorithm based on the predicted property values ​​for each molecule so that molecules with more desired properties are included; and search for molecules with desired properties based on the predicted property values.

[0020] According to one aspect of the present disclosure, the burden required for drug discovery can be further reduced.

[0021] 1 is a diagram showing an example of a drug discovery system including a molecular design device according to a first embodiment. 2 is a diagram showing an example of a drug discovery system including a molecular design device according to a second embodiment. 3 is an explanatory diagram illustrating an example of a plurality of molecular sequence information according to an embodiment of the present application. 4 is a diagram showing an example of the hardware configuration of a molecular design device according to an embodiment of the present application. 5 is a diagram showing an example of the configuration of a control unit according to an embodiment of the present application. 6 is a flowchart showing an example of a processing flow executed by the molecular design device according to the first embodiment. 7 is a flowchart showing an example of a processing flow executed by the molecular design device according to the second embodiment. 8 is a flowchart showing an example of a combinatorial optimization process according to the second embodiment. 9 is a flowchart showing another example of the combinatorial optimization process according to the second embodiment. 10 is a schematic diagram of optimization problem setting according to an application example of an embodiment of the present application. 11 is a diagram showing an example of calculation of objective function values ​​by optimization with TPE according to a first verification example. 12 is a diagram showing the relationship between the predicted mean value and the predicted standard deviation value of sequences obtained by sampling with TPE according to the first verification example. 13 is a diagram showing the density distribution of edit distances of sequences obtained by sampling with TPE according to the first verification example. 14 is a diagram showing the distribution of pseudo-correct model scores for proposed sequences according to the first verification example. 15 is a diagram showing the edit distance of proposed sequences according to the first verification example. 16 is a diagram showing the predicted standard deviation value of proposed sequences according to the first verification example. 1 is a table showing examples of amino acid candidates for each mutation candidate site in the search space according to the first verification example; a table showing example parameter settings for TPE according to the first verification example; a diagram illustrating the distribution of predicted means and predicted standard deviations of sequences obtained by sampling with TPE according to the second verification example; a t-SNE visualization diagram illustrating sequences obtained by sampling with TPE according to the second verification example; a diagram illustrating the predicted mean value of proposed sequences according to the second verification example; a diagram illustrating the predicted variance value of proposed sequences according to the second verification example; a diagram illustrating the distribution of expression levels of proposed sequences according to the second verification example; a diagram illustrating the distribution of octet values ​​for each sequence according to the second verification example.

[0022] I. Definitions The term "and / or" is used herein to refer to each of the objects listed before and after "and / or" or any combination thereof. For example, "A, B and / or C" includes each of the objects "A," "B," and "C," as well as the combinations "A and B," "A and C," "B and C," and "A and B and C."

[0023] Amino acids: As used herein, the term "amino acids" may include natural amino acids and unnatural amino acids. Natural amino acids are represented by one-letter codes, three-letter codes, 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, and Val / V.

[0024] Amino acid modification Amino acids in the amino acid sequence of an antigen-binding molecule can be modified by known methods, such as site-directed mutagenesis (Kunkel et al. (Proc. Natl. Acad. Sci. USA (1985) 82, 488-492)) and overlap extension PCR, as appropriate. Furthermore, several known methods can be used to modify amino acids by substituting amino acids other than natural amino acids (Annu. Rev. Biophys. Biomol. Struct. (2006) 35, 225-249, Proc. Natl. Acad. Sci. USA (2003) 100 (11), 6353-6357). For example, a cell-free translation system (Clover Direct (Protein Express)) containing a tRNA in which a non-natural amino acid is bound to an amber suppressor tRNA complementary to the UAG codon (amber codon), a type of stop codon, may also be used.

[0025] Antigen: As used herein, the term "antigen" is not limited to a specific structure, as long as it contains an epitope to which an antigen-binding domain binds. In one embodiment, the antigen is a peptide, polypeptide, or protein consisting of four or more amino acids. Examples of the antigen include membrane-type molecules expressed on the cell membrane and soluble molecules secreted extracellularly from cells.

[0026] Antigen-binding domain: As used herein, the term "antigen-binding domain" refers to any domain with any structure that binds to a target antigen. Examples of such domains include the variable regions of the heavy and light chains of an antibody, a module called an A domain of approximately 35 amino acids contained in Avimer, a cell membrane protein present in living organisms (International Publication Nos. WO2004 / 044011 and WO2005 / 040229), Adnectin, which contains the 10Fn3 domain that binds to proteins in fibronectin, a glycoprotein expressed on the cell membrane (International Publication No. WO2002 / 032925), Affibody, which uses an IgG-binding domain consisting of a 58-amino acid, three-helix bundle of Protein A as a scaffold (International Publication No. WO1995 / 001937), and DARPins (Designed Ankyrin Repeats), which are regions exposed on the molecular surface of ankyrin repeats (AR), which have a structure in which a 33-amino acid turn, two antiparallel helices, and a loop subunit are repeatedly stacked. Examples of antigen-binding domains of the present disclosure include those comprising the variable regions of antibody heavy and light chains (WO 2002 / 020565), anticalin, which is a four-loop region supporting one side of a barrel structure in which eight highly conserved antiparallel strands twist toward the center in lipocalin molecules such as neutrophil gelatinase-associated lipocalin (NGAL) (WO 2003 / 029462), and a concave region of a parallel sheet structure within a horseshoe-shaped structure in which leucine-rich repeat (LRR) modules are repeatedly stacked in the variable lymphocyte receptor (VLR) that does not have an immunoglobulin structure and is part of the adaptive immune system of jawless fish such as lampreys and hagfish (WO 2008 / 016854). Examples of antigen-binding domains of the present disclosure include antigen-binding domains comprising the variable regions of antibody heavy and light chains.Examples of such antigen-binding domains include "scFv (single chain Fv)," "single chain antibody," "Fv," "scFv2 (single chain Fv 2)," "Fab," and "F(ab')2."

[0027] Antigen-binding molecules. In the present disclosure, the term "antigen-binding molecule containing an antigen-binding domain" is used in the broadest sense and specifically includes various molecular types as long as it contains an antigen-binding domain. An antigen-binding molecule may be a molecule consisting of only an antigen-binding domain, or may be a molecule containing an antigen-binding domain and other domains. For example, when an antigen-binding molecule is a molecule in which an antigen-binding domain and an Fc region are linked, examples include complete antibodies and antibody fragments. Antibodies may include single monoclonal antibodies (including agonist and antagonist antibodies), human antibodies, humanized antibodies, chimeric antibodies, etc. Scaffold molecules, in which existing stable α / β barrel protein structures or other three-dimensional structures are used as scaffolds, and only partial structures of these structures are compiled into libraries for constructing antigen-binding domains, are also included in the antigen-binding molecules of the present disclosure.

[0028] The term "antibody" is used herein in the broadest sense and encompasses various antibody structures, including, but not limited to, monoclonal antibodies, polyclonal antibodies, multispecific antibodies (e.g., bispecific antibodies), and antibody fragments, as long as they exhibit the desired antigen-binding activity. Antibodies can be isolated from natural sources, such as plasma or serum, or from the culture supernatant of antibody-producing hybridoma cells, or can be partially or completely synthesized using techniques such as genetic recombination. Examples of antibodies include immunoglobulin isotypes and their isotypic subclasses. Nine known classes (isotypes) of human immunoglobulins are IgG1, IgG2, IgG3, IgG4, IgA1, IgA2, IgD, IgE, and IgM. Antibodies of the present disclosure may include IgG1, IgG2, IgG3, and IgG4 of these isotypes. For the human IgG1, human IgG2, human IgG3, and human IgG4 constant regions, multiple allotype sequences due to genetic polymorphisms are described in "Sequences of proteins of immunological interest," NIH Publication No. 91-3242, and any of these sequences may be used in the present disclosure. In particular, for the human IgG1 sequence, the amino acid sequence at positions 356-358 (EU numbering) may be either DEL or EEM. Furthermore, for the human Igκ (Kappa) constant region and human Igλ (Lambda) constant region, multiple allotype sequences due to genetic polymorphisms are described in "Sequences of proteins of immunological interest," NIH Publication No. 91-3242, and any of these sequences may be used in the present disclosure.

[0029] "Antibody fragment" refers to a molecule other than an intact antibody that contains a portion of the intact antibody that binds to the antigen to which the intact antibody binds. Examples of antibody fragments include, but are not limited to, Fv, Fab, Fab', Fab'-SH, F(ab')2; diabodies; linear antibodies; single-chain antibody molecules (e.g., scFv); and multispecific antibodies formed from antibody fragments.

[0030] The terms "full length antibody," "complete antibody," and "whole antibody" are used interchangeably herein and refer to an antibody having a structure substantially similar to a native antibody structure or having a heavy chain that includes an Fc region as defined herein.

[0031] The term "variable region" or "variable domain" refers to the domain of an antibody heavy or light chain that is involved in binding the antibody to an antigen. The heavy and light chain variable domains (VH and VL, respectively) of natural antibodies typically have similar structures, with each domain containing four conserved framework regions (FR) and three hypervariable regions (HVR). (See, for example, Kindt et al., Kuby Immunology, 6th ed., W.H. Freeman and Co., page 91 (2007)). A single VH or VL domain may be sufficient to confer antigen-binding specificity. Furthermore, antibodies that bind to a specific antigen may be isolated by screening a complementary library of VL or VH domains, respectively, using a VH or VL domain from an antibody that binds to that antigen. See, e.g., Portolano et al., J. Immunol. 150:880-887 (1993); Clarkson et al., Nature 352:624-628 (1991).

[0032] As used herein, "molecular weight" refers to the sum of the atomic weights of the atoms constituting a compound molecule (unit: "g / mol"), and is obtained by calculating the sum of the atomic weights of the atoms included in the molecular structural formula. In this specification, the unit of molecular weight may be omitted. Note that the molecular weight can be measured, for example, by liquid chromatography mass spectrometry (LC / MS).

[0033] As used herein, a "medium-sized compound" or "medium molecule" refers to a compound having a molecular weight of 500 g / mol or more and less than 30,000 g / mol. Examples of medium-sized compounds include compounds having a molecular weight of 500 g / mol or more and less than 6,000 g / mol, compounds having a molecular weight of 500 g / mol or more and less than 4,000 g / mol, compounds having a molecular weight of 600 g / mol or more and less than 4,000 g / mol, and compounds having a molecular weight of 700 g / mol or more and less than 3,000 g / mol. Examples of medium-sized compounds include peptide compounds containing peptide chains, nucleic acids, or sugar chains. These compounds may be peptide compounds, 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 more and less than 6000 g / mol, or may be a peptide compound having a molecular weight of 500 g / mol or more and less than 4000 g / mol, or may be a peptide compound having a molecular weight of 600 g / mol or more and less than 4000 g / mol, or may be a peptide compound having a molecular weight of 700 g / mol or more and less than 3000 g / mol.

[0034] In this specification, a low molecular weight compound is a compound having a molecular weight of less than 500 g / mol, and a high molecular weight compound is a compound having a molecular weight of 30,000 g / mol or more.

[0035] (Definition of Experimental Method) [Characteristics Evaluation] In one aspect of the present disclosure, a prediction model is generated by machine learning based on molecular sequence information and evaluation result information of the molecular characteristics. Non-limiting examples of molecular characteristics evaluation include, but are not limited to, molecular binding ability evaluation, pharmacological activity evaluation, physical property evaluation, kinetic evaluation, and safety evaluation.

[0036] Binding Ability Assessment The method for assessing the binding ability of a target molecule-binding molecule to a target molecule is not particularly limited, but can be performed by quantitatively assessing 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, if the target molecule is an antigen, it can be assessed 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 the molecule's binding partner (e.g., an antigen). Here, "binding activity" is not strictly limited to a 1:1 interaction between members of a binding pair (e.g., an antibody and an antigen). For example, when members of a binding pair reflect a monovalent 1:1 interaction, binding activity refers to the intrinsic binding affinity (sometimes simply referred to as "affinity"). When members of a binding pair are capable of both monovalent and multivalent binding, binding activity is the sum of their binding strengths. The binding activity of molecule X to its partner Y can generally be expressed by the dissociation constant (KD) or "amount of analyte bound per unit amount of ligand." Binding activity can be measured by conventional methods known in the art, including those described herein. Conditions other than the concentration of the target tissue-specific compound can be appropriately determined by those skilled in the art. In certain embodiments, the antigen-binding molecule provided herein is an antibody, and the binding activity of the antibody is ≦1 μM, ≦100 nM, ≦10 nM, ≦1 nM, ≦0.1 nM, ≦0.01 nM, or ≦0.001 nM (e.g., 10 -8 M or less, e.g. 10 -8 M to 10 -13 M, e.g. 10 -9 M to 10 -13 is the dissociation constant (KD) of the ATPase (M).

[0037] In one embodiment, antibody binding activity is measured using surface plasmon resonance analysis, e.g., a ligand capture method using a BIACORE® T200 or BIACORE® 4000 (GE Healthcare, Uppsala, Sweden). The instrument is operated using, for example, BIACORE® Control Software. In one embodiment, an amine coupling kit (GE Healthcare, Uppsala, Sweden) is used according to the supplier's instructions to immobilize a ligand capture molecule, e.g., an anti-tag antibody, an anti-IgG antibody, or protein A, on a carboxymethyldextran-coated sensor chip (GE Healthcare, Uppsala, Sweden). The ligand capture molecule is diluted with 10 mM sodium acetate solution at an appropriate pH and injected at an appropriate flow rate and injection time. Binding activity measurements are performed using a buffer containing 0.05% polysorbate 20 (also known as Tween® 20) as the measurement buffer, at a flow rate of 10-30 μL / min, and at a measurement temperature of, for example, 25°C or 37°C. When measurements are performed by capturing an antibody as a ligand on a ligand capture molecule, the antibody is injected and captured in a desired amount, and then a serial dilution (analyte) of an antigen and / or Fc receptor prepared using the measurement buffer is injected. When measurements are performed by capturing an antigen and / or Fc receptor as a ligand on a ligand capture molecule, the antigen and / or Fc receptor is injected and captured in a desired amount, and then a serial dilution (analyte) of an antibody prepared using the measurement buffer is injected.

[0038] In one embodiment, the measurement results are analyzed using BIACORE® Evaluation Software. Kinetic parameter calculations are performed by simultaneously fitting the binding and dissociation sensorgrams using a 1:1 binding model, and the binding rate (k or ka), dissociation rate (k or k), and equilibrium dissociation constant (KD) can be calculated. When binding activity is weak, particularly when dissociation is rapid and calculation of kinetic parameters is difficult, the equilibrium dissociation constant (KD) may be calculated using a steady state model. Another parameter of binding activity is the "amount of analyte bound per unit amount of ligand," which can be calculated by dividing the amount of analyte bound (RU) at a specific concentration by the amount of ligand captured (RU).

[0039] As the value of antigen-binding activity, KD (dissociation rate constant) can be used when the antigen is a soluble molecule, and apparent kD (apparent dissociation rate constant) can be used when the antigen is a membrane-type molecule. kD (dissociation rate constant) and apparent KD (apparent dissociation rate constant) can be measured by methods known to those skilled in the art, for example, using Biacore (GE Healthcare), a flow cytometer, etc.

[0040] Another aspect of property evaluation is a method for selecting antigen-binding molecules using a display library. One aspect includes panning using phage display. For example, affinity evaluation can be performed by preparing a phage library displaying multiple different antigen-binding molecules, contacting the prepared phages with a target antigen, and then washing away unbound phages to concentrate phages displaying antigen-binding molecules that interact with the target antigen. 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 aspect includes panning using mammalian cell display. For example, pharmacological activity evaluation using this display system can be performed by expressing a library containing multiple different antigen-binding molecules in target mammalian cells and varying reporter activity or the like depending on the effect the molecules have on the same cells, allowing cells carrying antigen-binding molecule genes with the desired pharmacological activity to be isolated using a flow cytometer or the like. Furthermore, as an example of physical property evaluation using this display system, a library containing multiple different antigen-binding molecules is expressed in target mammalian cells, and the expression level is stained with an antibody specific to the antigen-binding molecule, allowing cells harboring antigen-binding molecule genes that can be stably and highly expressed to be isolated using a flow cytometer, etc. Evaluation of the properties of antigen-binding molecules by panning is not limited to techniques using phages or mammalian cells, and various techniques can be used as long as they can display antigen-binding molecules, including, but not limited to, techniques in which the antigen-binding molecules are displayed on ribosomes, mRNA, viruses other than phages, and bacteria such as Escherichia coli.

[0041] Another aspect of property evaluation is a method of obtaining an antibody gene sequence from immune cells derived from an individual, or a method of obtaining an antibody protein sequence from serum. Taking affinity evaluation, in which antibody gene sequences are extracted from immune cells, as an example, a target antigen protein is administered to an individual to induce immune sensitization, and then genes are extracted from immune cells carrying antibody genes that bind to the target antigen, thereby identifying sequences with affinity for the target antigen. The antigen that induces immune sensitization is not limited to the above-mentioned protein; genes encoding the protein or cells expressing the protein can also be used. Target individuals include, but are not limited to, humans, mice, rats, hamsters, rabbits, monkeys, chickens, camels, llamas, and alpacas. Methods for analyzing the nucleic acid sequence or occurrence frequency include, but are not limited to, cloning genetically modified organisms containing the nucleic acid sequence of each antigen-binding molecule and analyzing them using the Sanger method with capillary electrophoresis, and using a next-generation sequencer. When analyzing the nucleic acid sequence, it is also possible to determine the strength of the property based on the occurrence frequency. For example, it is possible to estimate that antigen-binding molecules encoded by sequences that occur frequently after enrichment based on nucleic acid sequence analysis have high characteristics, whereas antigen-binding molecules encoded by sequences that occur less frequently after enrichment have lower characteristics than antigen-binding molecules encoded by sequences that occur more frequently. Furthermore, the techniques for obtaining information on antigen-binding molecules derived from the display library or an individual can be applied to various property evaluations, and are not limited to those described above.

[0042] Pharmacological Activity Evaluation The method for evaluating the pharmacological activity of a molecule is not particularly limited, and can be, for example, by measuring the neutralizing activity, agonist activity, or cytotoxic activity of the molecule. Examples of cytotoxic activity evaluation include antibody-dependent cell-mediated cytotoxicity (ADCC) activity, complement-dependent cytotoxicity (CDC) activity, T-cell-dependent cytotoxicity (TDCC) activity, and antibody-dependent cellular phagocytosis (ADCP) activity. CDC activity refers to cytotoxic activity mediated by the complement system. ADCC activity refers to the activity of immune cells, etc., binding to the Fc region of an antigen-binding molecule containing an antigen-binding domain that binds to a membrane-type molecule expressed on the cell membrane of the target cell via the Fcγ receptor expressed on the immune cell, causing damage to the target cell. TDCC activity refers to the activity of T cells damaging target cells by bringing them into close proximity using a bispecific antibody containing an antigen-binding domain that binds to a membrane-type molecule expressed on the cell membrane of the target cell and an antigen-binding domain that binds to one of the constituent subunits of the T cell receptor (TCR) complex on T cells, particularly an antigen-binding domain that binds to the CD3 epsilon chain. Whether or not a target antigen-binding molecule has ADCC activity, CDC activity, TDCC activity, or ADCP activity can be measured using known methods. Neutralizing activity refers to the activity of inhibiting the biological activity of a ligand that has biological activity against cells, such as a virus or toxin. In other words, a substance with neutralizing activity refers to a substance that binds to the ligand or the receptor to which the ligand binds, and inhibits the binding of the ligand to the receptor. A receptor whose binding to a ligand is blocked by neutralizing activity is no longer able to exert biological activity via the receptor.When the antigen-binding molecule is an antibody, an antibody having such neutralizing activity is generally called a neutralizing antibody, and neutralizing activity can be measured by measuring the inhibitory activity of the binding between the ligand and receptor. Ligands having biological activity against cells are not limited to viruses and toxins, and the inhibitory activity of physiological actions induced by the binding of endogenous ligands such as cytokines and chemokines to receptors is also understood as neutralizing activity. Furthermore, neutralizing activity is not limited to the inhibition of ligand-receptor binding, but also includes the activity of inhibiting the function of a protein having biological activity, and an example of the function of the protein is enzymatic activity.

[0043] Physical Property Evaluation Methods for evaluating the physical properties of molecules are not particularly limited, and examples of physical properties include thermal stability, chemical stability, solubility, viscosity, light stability, long-term storage stability, nonspecific adsorption, lipophilicity, and membrane permeability. The various physical properties exemplified above can be measured by methods known to those skilled in the art. The evaluation method is not particularly limited, and stability evaluations such as thermal stability, chemical stability, light stability, stability against mechanical stimuli, and long-term storage stability can be evaluated by measuring the decomposition, chemical modification, or aggregation of the molecule before and after the treatments, such as heat treatment, exposure to a low pH environment, light exposure, mechanical stirring, and long-term storage, that are the objective of the stability evaluation. Non-limiting examples of measurement methods for performing such stability evaluations include, but are not limited to, ion exchange chromatography, size exclusion chromatography, and other chromatographic methods, mass spectrometry, and electrophoresis, and can be measured by various methods known to those skilled in the art. Examples of physical property evaluation other than those described above include, but are not limited to, evaluation of protein solubility using polyethylene glycol precipitation, evaluation of viscosity using small-angle X-ray scattering, evaluation of nonspecific binding based on binding to extracellular matrix (ECM), etc. Physical property evaluations such as evaluation of protein expression level, evaluation of binding to purification resins or purification ligands, and evaluation of surface charge can also be performed as long as they are measurable by methods known to those skilled in the art.

[0044] Kinetic evaluation The method for evaluating the kinetics of a molecule is not particularly limited, but can be performed 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, a method widely known to those skilled in the art as pharmacokinetic (PK) evaluation. In addition to methods for directly evaluating PK, it is also possible to predict kinetic behavior from the amino acid sequence of the molecule by calculating the surface charge, isoelectric point, etc. of the molecule using software.

[0045] Safety assessment: Methods for assessing the safety of molecules are not particularly limited, and examples include immunogenicity prediction tools such as ISPRI Web-Based Immunogenicity Screening (EpiVax), HLA binding assessment of fragment peptides of antigen-binding molecules, detection of T cell epitopes and assessment of immunogenicity using MAPPs (MHC-Associated Peptide Proteomics) or T cell proliferation assessment, etc. Safety assessment can also be performed as long as it is measurable by methods known to those skilled in the art, such as binding to rheumatoid factor (RF), assessment of immune responses using PBMCs or whole blood, and platelet aggregation assessment.

[0046] (Definitions of terms and methods used in machine learning) - MBO (Model Based Optimization) MBO (Model Based Optimization) is an optimization method that does not directly target the characteristic value of a certain characteristic, but rather targets the predicted value of the characteristic value of a certain characteristic, estimated by some kind of model.

[0047] TPE (Tree-structured Parzen Estimator) TPE (Tree-structured Parzen Estimator) is a type of Bayesian optimization. TPE involves calculating the expected improvement of an output value for a given input value based on the conditional probability of the output value for the input value and the probability for the output value for the function to be optimized. In other words, TPE is a method of optimizing a function using input values ​​that maximize the expected improvement calculated in this way.

[0048] (Method for Producing Targets) In one aspect of the present disclosure, methods for producing targets suitable for obtaining drug candidates can be carried out by methods known to those skilled in the art. When the target is an antibody, the antibody can be produced using recombinant methods or constructs, for example, as described in U.S. Patent No. 4,816,567. One embodiment of the present disclosure includes a method for producing an antibody, comprising culturing a host cell containing nucleic acid encoding the antibody under conditions suitable for expression of the antibody, a candidate molecular compound described herein, and, optionally, recovering the antibody from the host cell (or host cell culture medium). The isolated nucleic acid encoding the antibody may encode an amino acid sequence comprising the antibody VL and / or an amino acid sequence comprising the antibody VH (e.g., the antibody light chain and / or heavy chain). The host cell containing such nucleic acid contains (e.g., is transformed with) (1) a vector containing nucleic acid encoding an amino acid sequence comprising the antibody VL and an amino acid sequence comprising the antibody VH, or (2) a first vector containing nucleic acid encoding an amino acid sequence comprising the antibody VL and a second vector containing nucleic acid encoding an amino acid sequence comprising the antibody VH. In one embodiment, the host cell is eukaryotic (e.g., Chinese hamster ovary (CHO) cells) or lymphoid cells (e.g., Y0, NS0, Sp2 / 0 cells). Suitable host cells for cloning or expressing antibody-encoding vectors include prokaryotic or eukaryotic cells. For example, antibodies may be produced in bacteria, particularly if glycosylation and Fc effector functions are not required. See, e.g., U.S. Patent Nos. 5,648,237, 5,789,199, and 5,840,523 for expression of antibody fragments and polypeptides in bacteria. (See also Charlton, Methods in Molecular Biology, Vol. 248 (BKC Lo, ed., Humana Press, Totowa, NJ, 2003), pp. 245-254, which describes the expression of antibody fragments in E. coli.) After expression, the antibody may be isolated in a soluble fraction from the bacterial cell paste and can be further purified.

[0049] When the target compound is a peptide compound or a cyclic peptide compound, it can be produced by liquid-phase synthesis, solid-phase synthesis using Fmoc synthesis, Boc synthesis, or a combination thereof. Liquid-phase synthesis and solid-phase synthesis can be performed by methods well known to those skilled in the art. Solid-phase synthesis is a method in which a compound is bound to a solid and then the compound is chemically reacted with a reagent on the solid resin to synthesize the target compound. Solid-phase peptide synthesis is a method in which a desired amino acid or peptide is bound to a solid resin, and further desired amino acids or peptides are sequentially linked to the amino acids or peptides bound to the solid resin to extend the peptide chain, thereby synthesizing the peptide. The peptide bound to the solid resin can be cleaved from the solid resin to obtain the target peptide.

[0050] First Embodiment FIG. 1 is a block diagram illustrating an example of a drug discovery system 100 including a molecular design apparatus 1 according to a first embodiment. The drug discovery system 100 is a system for creating new objects suitable as drug candidates. This system provides a method for generating new objects with predetermined properties, such as specific physiological activity (e.g., binding to a specific protein). Drugs include, but are not limited to, small molecule drugs, medium molecule drugs, biological drugs, cells, nucleic acid drugs, biopharmaceuticals, or potential active agents such as other active agents. Objects include molecular structures with desired or defined biological activity (e.g., preferentially binding to a specific protein over other proteins). Molecules that serve as drug candidates include biomolecules and compounds, including various molecules such as nucleic acids, peptides, cyclic peptides, proteins, antibodies, target molecule-binding molecules, polymeric compounds, medium molecule compounds, and small molecule compounds. Although not described herein, the drug discovery system 100 may also include a device for selecting molecules that interact with drug targets, a device for creating lead molecules, and the like. The drug discovery system 100 may be, for example, an information processing system configured including the disclosure of WO2020 / 246617.

[0051] The drug discovery system 100 is configured to include a molecular design device 1. The molecular design device 1 searches for candidate molecules having desired properties and outputs information on the identified candidates. The output information is building block combination information of the candidate molecules. In other words, the molecular design device 1 identifies candidate molecules and outputs building block combination information of the identified candidate molecules. Here, a candidate molecule is a molecule expected to have desired properties. The candidate molecule building block combination information is information about the candidate molecule, and is building block combination information for some or all of the candidate molecules. The output candidate molecule building block combination information may include information indicating one candidate molecule, or may include information indicating multiple candidate molecules.

[0052] The drug discovery system 100 selects new objects suitable for drug candidates using the candidate molecular building block combination information output from the molecular design device 1. For example, the drug discovery system 100 generates candidate molecules based on the candidate molecular building block combination information output from the molecular design device 1, experimentally evaluates the properties of the candidate molecules, and selects molecules with desired properties as new objects suitable for drug candidates based on the results of the property evaluation. In other words, the drug discovery system 100 can create new objects suitable for drug candidates based on the results of property evaluation performed by actually generating candidate molecules using the candidate molecular building block combination information output from the molecular design device 1. In such cases, candidate molecules can be considered to be molecules that can be used as verification targets for narrowing down candidates for the main components of pharmaceuticals.

[0053] The building block combination information of a molecule is information about the combination of some or all of the building blocks of a molecule. When the building block combination information is information about the combination of some of the building blocks of a molecule, the range of the sequence may be arbitrarily settable. A building block is a unit that constitutes a molecule. The building block combination information of a molecule relates to the combination of the building blocks that constitute the molecule. In this application, a sequence containing individual components may be referred to as a "combination." The term "sequence" may also be used as an example of a building block combination.

[0054] The molecule to be designed is, for example, a protein. When the molecule is a protein, the building blocks are amino acids, and the building block combination information of the molecule is, for example, information on the amino acid sequence of the protein. The molecule to be designed is, for example, a nucleic acid. When the molecule is a nucleic acid, the building blocks are nucleotides. The sequence of the molecule is, for example, the nucleotide sequence of a nucleic acid. The molecular building block combination information is information on the nucleotide sequence. More specifically, when the molecule to be designed is an antibody, the molecular sequence is an amino acid sequence, and the building blocks are amino acids. The building block combination information of the molecule is, for example, the full-length antibody sequence, such as the amino acid sequence of VH or VL, or the sequence of a portion of an antibody, such as CDR or FR. Furthermore, when the molecule to be designed is, for example, a cyclic peptide, the molecular sequence is an amino acid sequence containing unnatural amino acids, and the building blocks are natural amino acids and unnatural amino acids. The building block combination information of the molecule is information on the amino acid sequence containing unnatural amino acids. When the molecule to be designed is a small molecule, the building block combination of the molecule is a combination of fragments, and the building blocks are fragments (fragment molecules that make up the small molecule). For example, when the molecule to be designed is a nucleic acid, the molecular sequence is a base sequence, and the building blocks are bases.

[0055] The desired properties are properties required for a new target suitable for a drug candidate, and can be set arbitrarily. Non-limiting examples of properties include, but are not limited to, the ability to bind to a predetermined in vivo target, binding ability, pharmacological activity, physical properties, kinetics, and safety. For example, if the molecule to be designed is an antibody, the property is, for example, the drug's ability to bind to a predetermined antigen. For example, if the molecule to be designed is mRNA (messenger-RNA), the property is, for example, the protein translation ability.

[0056] The molecular design device 1 is an example of an inference device that infers a molecule having desired properties from prediction information, which will be described later. Examples of desired properties include the ability to bind to a target molecule, efficacy, and drug-like properties, including membrane permeability. Hereinafter, inferring a molecule having desired properties from prediction information may be rephrased as identifying a candidate molecule. In this embodiment, a molecule expected to have desired properties is a molecule that exhibits good prediction values ​​and low prediction uncertainty for the desired properties by a prediction model.

[0057] The following describes the case where sequence information is processed as building block combination information. The molecular design device 1 has an inference unit 111, which includes, for example, a sequence information processing unit 111a, a property prediction unit 111b, a prediction information processing unit 111c, and a candidate molecule identification unit 111d. The sequence information processing unit 111a prepares a sequence information set and outputs the prepared sequence information set to the property prediction unit 111b. The sequence information set is a set of sequence information for multiple molecules. The sequence information processing unit 111a may autonomously generate sequence information for each molecule or may input sequence information for each molecule from another device. Note that the sequence information set may also be referred to as a building block combination information set, and the sequence information processing unit 111a may also be referred to as a building block combination information processing unit. The property prediction unit 111b predicts the properties of molecules based on the sequence information of each element of the sequence information set input from the sequence information processing unit 111a, and outputs prediction information related to the predicted properties to the prediction information processing unit 111c. The prediction information processing unit 111c acquires prediction information for each molecule from the property prediction unit 111b. The prediction information processing unit 111c outputs the acquired prediction information to the candidate molecule identification unit 111d. The candidate molecule identification unit 111d identifies sequence information for at least one candidate molecule based on the prediction information. The candidate molecule identification unit 111d may output output data indicating the identified sequence information to another device or may store it in the storage unit 14 (described below).

[0058] In this embodiment, the sequence information set may be a set of sequence information indicating a virtual molecular sequence (sometimes referred to as a "virtual sequence") generated by machine learning, a set of sequence information of a real molecular sequence (sometimes referred to as a "real sequence"), or a set including both sequence information indicating a virtual sequence and sequence information indicating a real sequence. For example, the sequence information set may include sequence information of a virtual sequence generated by a virtual sequence generation model. The sequence information set may also include sequence information indicating a real sequence obtained from an existing database or experimental results. The sequence information set may also be sequence information extracted from all possible combinations of building blocks by combinatorial optimization, which will be described later. The sequence information processing unit 111a performs a process of preparing the sequence information set. The sequence information processing unit 111a may further include a sequence information set acquisition unit (not shown) that uses a machine learning model to execute an acquisition process for acquiring sequence information that will be output for certain input information. In this case, specifically, the sequence information processing unit 111a acquires the sequence information set and outputs the acquired sequence information set to the property prediction unit 111b.

[0059] In this embodiment, the property prediction unit 111b receives sequence information for multiple molecules from the sequence information processing unit 111a. The property prediction unit 111b calculates a predicted property value and an estimated value of prediction uncertainty for each of the input sequence information for the multiple molecules. In this embodiment, the property prediction unit 111b includes a predicted value calculation unit 111x and a prediction uncertainty estimation unit 111y. The predicted value calculation unit 111x receives the sequence information for the molecules and calculates a predicted property value using a prediction model. The predicted property value is a property value predicted using the prediction model. The prediction uncertainty estimation unit 111y estimates the uncertainty of the predicted property value calculated by the predicted value calculation unit 111x. A value indicating the estimated uncertainty is referred to as an estimated value of prediction uncertainty. The prediction model is generated, for example, by learning based on training data including multiple sets of sequence information for individual molecules and property evaluations for the molecules. The predicted value calculation unit 111x may calculate predicted property values ​​for multiple properties. When characteristic values ​​are predicted for multiple characteristics, the prediction uncertainty estimator 111y may estimate the prediction uncertainty for the predicted characteristic values ​​of each of the characteristics, or may estimate the prediction uncertainty for any one of the predicted characteristic values. The characteristic predictor 111b outputs the predicted characteristic value and an estimated value of the prediction uncertainty for the predicted characteristic value to the prediction information processor 111c as prediction information for each molecule.

[0060] The candidate molecule specifying unit 111d performs a process of inferring molecules having desired properties based on the prediction information, thereby specifying candidate molecules. In this embodiment, the candidate molecule specifying unit 111d infers molecules having desired properties based on the property prediction value and the estimated value of the prediction uncertainty for the sequence information of each molecule output from the property prediction unit 111b. The candidate molecule specifying unit 111d specifies at least one candidate molecule from a plurality of molecules whose sequence information is included in the sequence information set from the sequence information processing unit 111a, based on the prediction information obtained from the prediction information processing unit 111c. In specifying the candidate molecule, the candidate molecule specifying unit 111d may use a property prediction value for at least one property, or may use property prediction values ​​for multiple properties. In specifying the candidate molecule, the candidate molecule specifying unit 111d may use, as prediction information, multiple property prediction values ​​and an estimated value of the prediction uncertainty for at least one property prediction value. That is, in identifying a candidate molecule, when optimizing the first characteristic value (predicted characteristic value), constraints on the second characteristic value (predicted characteristic value) may be taken into consideration. Furthermore, the candidate molecule identifying unit 111d may identify a molecule having a desired characteristic as a candidate molecule based on a characteristic quality value, which will be described later. For example, the candidate molecule identifying unit 111d may select at least one molecule whose predicted characteristic value and prediction uncertainty each satisfy predetermined conditions as a candidate molecule. The candidate molecule identifying unit 111d may select at least one molecule whose multiple predicted characteristic values ​​and an estimated value of the prediction uncertainty for at least one predicted characteristic value each satisfy predetermined conditions as a candidate molecule.

[0061] The candidate molecule identifying unit 111d may identify at least one candidate molecule based on a characteristic quality value calculated based on a predicted characteristic value and an estimated value of the prediction uncertainty. In this case, the characteristic quality value may be calculated, for example, by the prediction information processing unit 111c or the characteristic prediction unit 111b. The characteristic quality value is a value based on the predicted characteristic value and the estimated value of the prediction uncertainty. The characteristic quality value may also be considered as a response variable calculated using a predetermined function with the predicted characteristic value and the estimated value of the prediction uncertainty as explanatory variables. For example, the characteristic quality value may be an index value that decreases as the estimated value of the prediction uncertainty increases, and increases as the predicted characteristic value increases. In other words, the characteristic quality value may increase as the predicted characteristic value increases and decrease as the estimated value of the prediction uncertainty increases. As an example, the characteristic quality value is the difference between the predicted characteristic value in a linear region and a predetermined coefficient multiplied by the estimated value of the prediction uncertainty, or the difference between the predicted characteristic value multiplied by a predetermined coefficient and the estimated value of the prediction uncertainty. The predetermined coefficient is a positive constant that indicates the degree of contribution of the characteristic value or the estimated value of the prediction uncertainty to the characteristic quality value. Therefore, if a characteristic prediction value is defined that gives a larger value the more a desired characteristic is predicted to be possessed or the degree of that characteristic is higher, the characteristic quality value is an index that indicates a larger value as the characteristic prediction value increases and the estimated value of the prediction uncertainty decreases. In this case, a larger characteristic quality value means that the molecule actually generated is more likely to exhibit the expected characteristics. In other words, estimation processing using characteristic quality values ​​in molecular design can be said to realize risk-averse estimation processing.

[0062] The characteristic quality value does not necessarily have to be the difference between the characteristic prediction value in the linear domain and the prediction uncertainty estimate multiplied by a predetermined coefficient. The characteristic quality value may be the difference or quotient between either the characteristic prediction value or the prediction uncertainty estimate raised to a predetermined power. Furthermore, when a prediction value that decreases with increasing probability of having a desired characteristic or the degree of the characteristic is defined, the characteristic prediction value may be determined by dividing the characteristic prediction value by the prediction uncertainty estimate or by other types of calculations, such as a calculation in the logarithmic domain. The characteristic quality value may be calculated using both the prediction value and its uncertainty, and the function, procedure, etc., used are not limited.

[0063] The candidate molecule specifying unit 111d identifies molecules having the desired property based on the calculated property quality value. As an example, the candidate molecule specifying unit 111d, among the molecules for which the property prediction unit 111b has obtained a property prediction value, has a higher property quality value, which means that the molecule is more likely to be identified as a molecule having the desired property. Here, "good property quality value" means that a property quality value indicating the possibility of being predicted to have the desired property or a high degree of the property is obtained. As another example, the candidate molecule specifying unit 111d may identify, among the molecules for which the property prediction unit 111b has calculated a property prediction value, a molecule whose property quality value satisfies a predetermined condition as a molecule having the desired property. As yet another example, the candidate molecule specifying unit 111d may rank and sort a plurality of molecules for which the property prediction unit 111b has calculated a property prediction value using the property quality value, and identify molecules within a predetermined range from the sorted molecules as molecules having the desired property. Here, the candidate molecule specifying unit 111d may identify a certain number of molecules as molecules having the desired property, so that molecules with better property quality values ​​are given priority. Alternatively, the candidate molecule specifying unit 111d may specify, as a molecule having a desired property, a molecule whose predicted property value and estimated value of the uncertainty of the prediction each satisfy predetermined conditions.

[0064] The accuracy of the predicted property values ​​calculated by the property prediction unit 111b affects the identification of molecules. Depending on the training data for the property prediction model, differences in the accuracy of property predictions based on sequence information can occur. Here, prediction accuracy may be rephrased as prediction reliability or prediction certainty. Low prediction accuracy can be rephrased as low prediction reliability, low prediction certainty, or high prediction uncertainty. High prediction accuracy can be rephrased as high prediction reliability, high prediction certainty, or low prediction uncertainty. In the molecular design device 1 of this embodiment, the inference process for molecules having desired properties uses not only predicted property values ​​but also estimated values ​​of prediction uncertainty as indicators. That is, in the molecular design device 1 of this embodiment, when identifying candidate molecules, the candidate molecule identification unit uses a mathematical model that predicts property values ​​and estimates prediction uncertainty, rather than a mathematical model that predicts property values. This prevents molecules with high predicted property values ​​but low prediction certainty from being identified as candidate molecules through the inference process. That is, when a molecule is actually generated, it is possible to prevent the identification of a molecule that is likely to have properties that are not as good as expected. As a result, this embodiment can further reduce the burden required for drug discovery.

[0065] Second Embodiment Fig. 2 is a block diagram showing an example of a drug discovery system 100 including a molecular design apparatus 1 according to a second embodiment. The following description will focus on differences from the first embodiment. Unless otherwise specified, the description of the first embodiment will be used for points in common with the first embodiment.

[0066] The molecular design device 1 according to this embodiment is an example of an inference device that infers molecules expected to have desired properties from prediction information, which will be described later. The molecular design device 1 includes an inference unit 111 that includes, for example, a sequence information processing unit 111 a, a property prediction unit 111 b, a prediction information processing unit 111 c, and a candidate molecule identification unit 111 d.

[0067] The inference unit 111 according to this embodiment executes a combinatorial optimization algorithm (sometimes simply referred to as "combinatorial optimization" in this application) to infer molecules expected to have desired properties. The sequence information processing unit 111a executes combinatorial optimization to extract sequence information for an arbitrary number of molecules from a predetermined search space. That is, the sequence information processing unit 111a repeatedly extracts sequence information for a portion of molecules expected to have desired properties from candidate building block sequences in the search space using updated extraction parameters through combinatorial optimization. The search space can be set arbitrarily. For example, the search space may be set based on training data used to train a prediction model used for property prediction. The number of sequence information acquisitions, i.e., the number of repetitions of sequence information acquisition, may be any natural number equal to or greater than 1. The number of sequence information acquisitions may be preset in the inference unit 111. The inference unit 111 may count the number of times the sequence information processing unit 111a actually extracts sequence information as the number of repetitions. In this embodiment, the sequence information of molecules included in the search space can be an element of a sequence information set.

[0068] The sequence information processing unit 111a outputs the extracted sequence information of multiple molecules to the property prediction unit 111b. The property prediction unit 111b according to this embodiment uses a prediction model to generate prediction information indicating the properties of each molecule based on the sequence information of the molecule input from the sequence information processing unit 111a, and outputs the prediction information generated for each molecule to the prediction information processing unit 111c. The prediction model is generated, for example, by learning based on training data including multiple sets of sequence information for each molecule and property evaluation results for the molecule. The prediction information for each molecule includes, for example, a predicted property value for each molecule. Therefore, the property prediction unit 111b may include a predicted value calculation unit 111x that calculates a predicted property value from the sequence information for each molecule. The prediction information for each molecule may include, for example, a predicted property value for each molecule and an estimated value for the uncertainty of the predicted property value. The property prediction unit 111b may also include a prediction uncertainty estimation unit 111y that calculates an estimated value for the uncertainty of the predicted property value. Therefore, the prediction information for each molecule output from the characteristic prediction unit 111b can include not only the predicted characteristic value for each molecule but also an estimated value of the uncertainty of the predicted characteristic value.

[0069] The prediction information processing unit 111c acquires prediction information for each molecule from the property prediction unit 111b. If the prediction information includes a property prediction value and an evaluation value for the uncertainty of the property prediction value, the prediction information processing unit 111c may calculate a property quality value for each molecule of sequence information included in the sequence information set based on the property prediction value and the uncertainty of the property prediction value. In this embodiment, the sequence information processing unit 111a uses the property quality value calculated by the prediction information processing unit 111c as prediction information to update the extraction parameters. Specific examples of extraction parameters will be described later. If the counted number of repetitions does not reach a predetermined number of sequence information acquisitions, the sequence information processing unit 111a updates the extraction parameters based on the acquired prediction information. Therefore, in the combinatorial optimization, a series of processes is repeated, including the extraction of sequence information based on the updated extraction parameters in the sequence information processing unit 111a, the acquisition of prediction information based on the extracted sequence information by the property prediction unit 111b, and the updating of the extraction parameters by the sequence information processing unit 111a.

[0070] When the counted number of repetitions reaches a preset number of sequence information acquisitions, the sequence information processing unit 111a stops processing the combinatorial optimization algorithm, and the prediction information processing unit 111c outputs the prediction information for each molecule input from the property prediction unit 111b to the candidate molecule identification unit 111d. The candidate molecule identification unit 111d identifies at least one candidate molecule from the set of sequence information for each iteration based on the prediction information for each molecule input from the prediction information processing unit 111c. When the number of sequence information acquisitions is N, the prediction information processing unit 111c identifies sequence information for at least one candidate molecule from the sequence information for molecules extracted N times. The candidate molecule identification unit 111d may also identify sequence information for at least one candidate molecule from the set of sequence information for molecules extracted N times. Note that when the prediction information includes a property quality value, the candidate molecule identification unit 111d can identify a candidate molecule based on the property quality value for each molecule. The candidate molecule identification unit 111d outputs sequence information for the identified candidate molecule.

[0071] In this embodiment, Bayesian optimization can be applied to the combinatorial optimization, but is not limited to this. As a Bayesian optimization method, for example, a Tree-Structured Parzen Estimator (TPE) may be used, and as an evolutionary algorithm method, for example, an Elitist Non-Dominated Sorting Genetic Algorithm (NSGA II) may be used.

[0072] In this embodiment, by using combinatorial optimization, it is possible to infer molecules with desired properties without predicting property values ​​for each piece of sequence information for all molecules included in the search space. If a huge number of molecules were to be processed, it would take an enormous amount of computation time. However, by performing combinatorial optimization on the search space as in this embodiment, it is possible to infer molecules with desired properties without dealing with all molecules. This further reduces the burden required for drug discovery.

[0073] In this embodiment, the molecular design device 1 may use, as the objective function of combinatorial optimization, a mathematical model for estimating a characteristic value from sequence information, or a mathematical model for estimating a characteristic quality value. The output of the mathematical model is then optimized by combinatorial optimization. When using a mathematical model for estimating a characteristic quality value, it is possible to prevent a molecule with a low prediction probability from being identified as a candidate molecule, even if the characteristic indicated by the characteristic prediction value is favorable. In other words, it is possible to reduce the possibility of identifying, as a candidate molecule, a molecule that is unlikely to have the characteristics expected when actually generated. This further reduces the burden required for drug discovery.

[0074] FIG. 3 is an explanatory diagram illustrating an example of a sequence information set in the first and second embodiments. The "sequence ID" in FIG. 3 is an identifier that identifies the sequence information of each molecule included in the sequence information set. FIG. 3 shows sequence information for each molecule included in the sequence information set. The sequence information for each molecule indicates multiple building blocks and the arrangement of each building block. The arrangement of each building block is shown between information D101 and information D102. For example, information D101 is information on the building block at position H1 in the sequence of each molecule. VHL0001, VHL0002, VHL0003, VHL0004, etc. all indicate molecules in which the building block at position H1 is M.

[0075] 4 is a diagram showing an example of the hardware configuration of a molecular design device 1 according to the first and second embodiments. The molecular design device 1 includes a computer system that includes a control unit 11 having a processor 91 such as a CPU and a memory 92 interconnected by a bus, and that executes a predetermined program. The computer system can also be considered to function as the molecular design device 1 that includes the control unit 11, input unit 12, communication unit 13, storage unit 14, and output unit 15 by executing the program. In this application, "executing a program" or "running a program" includes the meaning of executing a process instructed by each of one or more instructions written in the program.

[0076] More specifically, the processor 91 reads out a program stored in the storage unit 14 and stores the read out program in the memory 92. The processor 91 executes the program stored in the memory 92, thereby functioning as a molecular design device 1 including a control unit 11, an input unit 12, a communication unit 13, a storage unit 14, and an output unit 15. The program is stored in advance in the storage unit 14, for example.

[0077] The control unit 11 controls the operations of various functional units included in the molecular design apparatus 1. The control unit 11 includes, for example, the function of an inference unit 111 that executes an inference process for inferring a molecule having desired properties.

[0078] The input unit 12 includes input devices such as a mouse, a keyboard, and a touch panel. The input unit 12 may be configured as an interface that connects these input devices to the molecular design device 1. The input unit 12 accepts input of various information to the molecular design device 1. For example, training data for a prediction model is input to the input unit 12.

[0079] The communication unit 13 includes a communication interface for connecting the molecular design apparatus 1 to an external device. The communication unit 13 communicates with the external device via wired or wireless communication. The external device is, for example, a device that transmits training data for a prediction model. The external device may also be a device to which candidate molecule sequence information is transmitted. The molecular design apparatus 1 may be realized as a server device connected to the Internet and capable of communicating with one or more terminal devices.

[0080] The storage unit 14 is configured using a computer-readable storage medium device (non-transitory computer-readable recording medium) such as a magnetic hard disk drive or a semiconductor storage device. The storage unit 14 stores various information related to the molecular design apparatus 1. The storage unit 14 stores information input via the input unit 12 or the communication unit 13, for example. The storage unit 14 stores various information generated by processing executed by the control unit 11, for example. The storage unit 14 stores, for example, a prediction model, sequence information, various parameters used in combinatorial optimization (which may include the above-mentioned extracted parameters), and the like. The storage unit 14 stores, for example, the above-mentioned program.

[0081] The output unit 15 outputs various types of information. The output unit 15 includes a display device such as a CRT (Cathode Ray Tube) display, a liquid crystal display, or an organic EL (Electro-Luminescence) display. The output unit 15 may be configured as an interface that connects these display devices to the molecular design apparatus 1. The output unit 15 outputs information input to the input unit 12 or the communication unit 13, for example. The output unit 15 may also output various types of information generated by processing executed by the control unit 11, for example.

[0082] 5 is a diagram showing an example of the configuration of the control unit 11 according to each of the first and second embodiments. The control unit 11 includes an inference unit 111, an input control unit 112, a communication control unit 113, a storage control unit 114, and an output control unit 115. The inference unit 111 executes the above-described inference process. The input control unit 112 controls the operation of the input unit 12. The communication control unit 113 controls the operation of the communication unit 13. The storage control unit 114 controls the operation of the storage unit 14. The output control unit 115 controls the operation of the output unit 15.

[0083] Next, an example of an inference process according to each embodiment will be described. FIG. 6 is a flowchart showing an example of the flow of the process executed by the molecular design apparatus 1 according to the first embodiment. The sequence information processing unit 111a acquires a sequence information set including sequence information of multiple molecules (step S101). The property prediction unit 111b calculates a predicted property value for each molecule in the sequence information set and calculates an estimated value of the uncertainty of the property value (step S102). The candidate molecule identification unit 111d identifies a candidate molecule based on the predicted property value and the estimated value of the prediction uncertainty calculated for each molecule (step S103). The method used by the property prediction unit 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 predicted property value may be calculated as the estimated value of the uncertainty of the prediction. Furthermore, conformal prediction may be used to quantify the uncertainty. The property prediction unit 111b may divide the training data into multiple parts, calculate the error (prediction error) for each part, and quantify the uncertainty based on the distribution of the errors. The candidate molecule identification unit 111d outputs the sequence information of the identified candidate molecule as candidate molecule sequence information using the output unit 15 (step S104). Thereafter, the control unit 11 ends the process of FIG. 6.

[0084] Prior to step S103, the prediction information processing unit 111c may calculate a characteristic quality value based on the characteristic prediction value obtained in step S102 and an estimated value of the prediction uncertainty. The prediction information processing unit 111c may calculate mean variance (MV) as an example of the characteristic quality value. MV is obtained by subtracting a penalty function g(x) from the product of a risk tolerance parameter ρ and a characteristic prediction value f'(x) based on a predetermined prediction model (MV = ρf'(x) - g(x)). The risk tolerance parameter is a positive real value indicating the tolerance of the characteristic prediction value f'(x) to the uncertainty g(x), i.e., the reliability of the predicted value f'(x). The prediction information processing unit 111c calculates the standard deviation as an example of the uncertainty estimate g(x). MV is sometimes applied to portfolio optimization in financial engineering. In financial engineering, MV is sometimes used as an index that combines the expected reward (average) and the variance of the reward (risk). MV is an index that can be used in various fields and for various purposes, regardless of the optimization algorithm (see, for example, Q. Zhu and V. Y. F. Tan: Thompson Sampling Algorithms for Mean-Variance Bandits (2020 ICML), S. Takemori: Distributionally-Aware Kernelized Bandit Problems for Risk Aversion (2022 ICML)).

[0085] The prediction information processing unit 111c may perform calculations using uncertainty g(x) as a constraint, which is an example of a characteristic quality value. For example, the prediction information processing unit 111c may set a threshold value τ for a certain uncertainty g(x), and when the uncertainty g(x) is equal to or smaller than τ, set the characteristic quality value to the characteristic predicted value f'(x), and when the uncertainty g(x) is greater than τ, set the characteristic quality value to a sufficiently small value (for example, the minimum value that f'(x) can take). Such settings are expressed by equation (1).

[0086] The prediction information processing unit 111c may calculate the characteristic quality value based on multiple physical property values. For example, the prediction information processing unit 111c may set a threshold value τ1 for the first physical property value g1(x) and a threshold value τ2 for the second physical property value g2(x). When the first physical property value g1(x) is smaller than τ1 and the second physical property value g2(x) is smaller than τ2, the prediction information processing unit 111c may set the characteristic quality value to the characteristic predicted 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 processing unit 111c may set the characteristic quality value to a sufficiently small value (e.g., the minimum value that f'(x) can take).

[0087] Alternatively, the prediction information processing unit 111c may calculate the characteristic quality value based on the predicted value Pr(x∈F|x) of the probability of satisfying the constraint. That is, the prediction information processing unit 111c may multiply the characteristic predicted value f'(x) by the prediction model by the predicted value of the probability of satisfying the constraint, and sample only the sequence information with a high predicted probability of satisfying the constraint. In this example, when the probability of satisfying the constraint is low, the characteristic quality value approaches 0, and the higher the probability of satisfying the constraint (the closer it is to 1), the characteristic quality value approaches the value of the objective index itself. Such a characteristic quality value is expressed by Equation (2).

[0088] The number of constraints may be one or more. For example, when calculating a characteristic quality value based on multiple physical property values, the prediction information processing unit 111c may set constraint conditions for each physical property value and multiply the characteristic predicted value f'(x) by the prediction model by the probability that all conditions are satisfied. In this case, the probability that all conditions are satisfied may be the product of the probabilities that the constraint conditions for each physical property value are satisfied.

[0089] FIG. 7 is a flowchart showing an example of the flow of processing executed by the molecular design apparatus 1 according to the second embodiment. The sequence information processing unit 111a acquires, as training data, a dataset D including a plurality of data pairs, each of which is a pair of an input value x and a measurement value y (step S201). For example, the storage unit 14 pre-stores the dataset D input from the input unit 12 or the communication unit 13. The sequence information processing unit 111a reads the dataset D from the storage unit 14. Using the acquired training data, the sequence information processing unit 111a performs, for example, Gaussian process regression to learn a prediction model for predicting a measurement value for an input value (step S202). The inference unit 111 performs combinatorial optimization to identify candidate molecules (step S203). The candidate molecule identification unit 111d outputs sequence information of the identified candidate molecules using the output unit 15 (step S204). The processing of FIG. 7 then terminates.

[0090] <Application Example to MBO> Next, an application example to MBO will be described. In this application example, offline MBO, a type of in-silico drug design method, is combined with building block-based molecular design. Offline MBO is a method for searching for optimal molecules within a "proxy" predictive model generated from acquired data. This method involves black-box optimization, also known as inverse analysis, which treats a representative model as a black box. Here, MBO is a technique that trains a predictive model using previously accumulated experimental data as training data, evaluates drug candidate molecules based on characteristic values ​​obtained using the trained predictive model, and updates the predictive model based on the evaluation results. In the following description, the predictive model may be referred to as a representative model (proxy model).

[0091] In this application example, discrete input values ​​can be the optimization targets of MBO. In MBO, measurement values ​​y for discrete input values ​​x∈X are given as an unknown function f(x), and a dataset D is given containing n sets of measurement values ​​y for discrete input values ​​x (n is an integer greater than or equal to 2), and it is assumed that measurement values ​​y for additional input values ​​x cannot be obtained. The measurement values ​​y for discrete input values ​​x provide a correct solution (also called an oracle) for the function f(x). A typical MBO aims to find an input value x in a discrete value space X that maximizes the function f(x). The above method can be viewed as a problem of finding an input value x that maximizes the characteristic quality value f(x)-λg(x) instead of the function f(x). Here, an array is applied as the input value x, and the function f(x) corresponds to a function for predicting a real value corresponding to the characteristic value. The penalty function g(x) is a function that provides a real value corresponding to the uncertainty of the value of the function f(x). λ corresponds to the coefficient α mentioned above. The input value x is represented using a vector representation.

[0092] In offline MBO, the array information processing unit 111a trains a prediction model in advance using a data set D having multiple pairs of measurement values ​​y for discrete input values ​​x as training data. In training the prediction model, the array information processing unit 111a sets the function value of the objective function for the discrete input value x as a predicted value y', and determines parameters of the prediction model so as to minimize an index value representing the magnitude of the difference between the predicted value y' and the measurement value y for the discrete input value x. In this application example, as described below, the inference unit 111 samples the discrete input value x and updates the parameters of the combinatorial optimization based on the prediction result obtained for each sample using the prediction model.

[0093] The inference unit 111 can use, for example, the above-mentioned TPE in combinatorial optimization. TPE is a black-box optimization algorithm that combines Bayesian optimization with Parzen window density estimation. Because TPE can handle categorical parameters, it can be applied to building block-based molecular design.

[0094] In this application example, the inference unit 111 implements the functionality of TPE and is applied to molecular design as follows. TPE is a method aimed at maximizing the expected improvement (EI) of an objective function. First, the sequence information processing unit 111a applies building block combinations constituting drug candidate molecules to a search space, and assigns the output from the prediction model to an objective score set Y corresponding to an input molecule set X sampled from the search space. For example, the sequence information processing unit 111a sets the predicted property value obtained by the prediction model as the objective score for TPE. The sequence information processing unit 111a samples the input molecule set X from the search space, i.e., assigns candidate building blocks to each modification position in the building block combination. For example, if the drug candidate molecule is a protein, candidate amino acids are assigned to each modification position in the protein sequence. As an example, an Optuna implementation can be used to execute the TPE sampler.

[0095] Next, the property prediction unit 111b calculates an estimate of the objective function using the prediction model for the input values ​​determined by sampling (evaluation). The prediction information processing unit 111c outputs the selected input values ​​and the calculated estimates to the sequence information processing unit 111a, and the sequence information processing unit 111a updates the parameters of the optimization algorithm using the input values ​​and estimates input from the prediction information processing unit 111c (update). Here, the sequence information processing unit 111a assigns the selected input values ​​and the calculated estimates to an input molecule set X sampled from the search space and a set Y of objective scores, respectively. The inference unit 111 repeats the processes of data division, sampling, evaluation, and update, so that the evaluation results of the input values ​​(sample values) obtained by sampling are reflected in the prediction model, and a combination of building blocks with a larger estimated value of the objective function is probabilistically searched for.

[0096] An example of the flow of the combinatorial optimization process executed by the molecular design apparatus 1 according to this application example is shown in FIG. 8. The inference unit 111 sets an initial value of 0 for the number of sequence information acquisitions. Loop R20 includes the processes of steps S203a to S203c. The control unit 11 sets the execution condition for loop R20 when the number of repetitions is equal to or less than a predetermined number of samplings N. The number of samplings corresponds to the number of sequence information acquisitions described above.

[0097] The sequence information processing unit 111a acquires sequence information for multiple molecules according to a combinatorial optimization algorithm (step S203a). For example, when using TPE as the combinatorial optimization algorithm, the sequence information processing unit 111a divides a set of input values ​​into two sets based on estimated values ​​(corresponding to the above-mentioned predicted characteristic values) obtained for each input value using a prediction model and a predetermined threshold γ. One set (referred to as the "first set") includes input values ​​whose estimated values ​​are equal to or greater than the threshold γ. The other set (referred to as the "second set") includes input values ​​whose estimated values ​​are less than the threshold γ. The sequence information processing unit 111a samples input values ​​that maximize the expected improvement of the objective function. The expected improvement corresponds to the increase in the expected value of the objective function before and after the update and is known to be proportional to p(x|y1) / p(x|y2) belonging to the first set for the input value x. p(x|y1) indicates the density distribution of the input value x for the first set. p(x|y2) denotes the density distribution of the input value x for the second set. That is, EI is an example of an extraction parameter. The sequence information processing unit 111a samples an input value that maximizes the calculated EI. The property prediction unit 111b performs sequence evaluation using a prediction model. Here, the property prediction unit 111b calculates an estimate of the objective function for the sampled input value using the prediction model (step S203b). The sequence information processing unit 111a assigns the selected input value and the calculated estimate to an input molecule set X sampled from the search space and a set Y of objective scores. Therefore, by adding a new input value to the input molecule set X sampled from the search space in association with the objective function, the extraction parameters of the combinatorial optimization algorithm related to the extraction of a sequence information set are updated (step S203c).

[0098] The inference unit 111 updates (increments) the number of repetitions by adding 1 each time the processing of steps S203a to S203c is completed. When the number of repetitions is equal to or less than the sampling number N, the inference unit 111 repeats the processing of steps S203a to S203c. When the number of repetitions exceeds the sampling number N, the inference unit 111 terminates the processing of loop R20. The candidate molecule identification unit 111d may output a predetermined number of building block combinations in descending order from the one with the highest estimated value of the objective function as candidate molecule sequence information using the output unit 15. Thereafter, the processing of FIG. 8 terminates.

[0099] FIG. 9 shows another example of the flow of the combinatorial optimization process executed by the molecular design apparatus 1 according to this application example. The process of FIG. 9 is the same as the process of FIG. 8 in that steps S203a to S203c are repeated, but differs from the process of FIG. 8 in that steps S203d and S203e are included instead of loop R20. In step S203d, the inference unit 111 sets the sampling count N and the initial value of the repetition count to 0. After the process of step S203b, the process proceeds to step S203e. In step S203e, the inference unit 111 adds 1 to the repetition count at that time and determines whether the repetition count has reached N. If it is determined that it has reached N (YES in step S203e), the process of FIG. 9 ends. If it is determined that it has not reached N (NO in step S203e), the process proceeds to step S203c. After the process of step S203c, the process proceeds to step S203a. 9 differs from the processing in FIG. 8 in that the inference unit 111 does not update the extracted parameters when the number of repetitions exceeds the number of samplings N.

[0100] 8 and 9, it is assumed that the termination condition is set when the number of repetitions exceeds the number of times sequence information is acquired, but this is not necessarily limited to this. For example, the termination condition may be set such that a target value for the estimated value of the objective function is set in advance and the estimated value of the objective function reaches that target value.

[0101] In the verification examples of the present application, a TPE that uses MV calculated from the predicted mean and predicted variance of Gaussian process regression as the objective function may be referred to as MV-TPE. A TPE that uses the predicted mean (Mean) of Gaussian process regression as the objective function may be referred to as Mean-TPE. In the present application, the predicted mean of Gaussian process regression as the objective function refers to the characteristic value itself.

[0102] <First Verification Example> Next, a verification example of this embodiment will be described. In this first verification example, the safety of this application example was verified using a GFP dataset as training data for the prediction model. The GFP dataset contains 56,086 GFP (Green Fluorescent Protein) sequences and their characteristic brightness information. GFP is a sample widely used in medical and biological research. Generating bright GFP is a research goal, and it is sometimes used as a benchmark for model-based optimization performance. Input values ​​representing individual GFP sequences are expressed using a 756-dimensional vector representation. In this verification, the Task Assessing Protein Embeddings method (TAPE) was used as a protein embedding model to obtain the vector representation.

[0103] A predictive model was trained using a Gaussian process (GP) based on GFP sequences with two or fewer mutations from the parent sequence avGFP (Aequoria victoria GFP) as training data. In the following description, the predictive model obtained through this training may be referred to as a proxy model. The parent sequence avGFP may also be simply referred to as the parent sequence or template GFP. This procedure validates sequences with fewer residue substitutions from the parent sequence, thereby realizing a practical molecular optimization process. For this validation, GFP sequences with two or fewer residue substitutions (edit distance), as illustrated in Figure 10, were used. A Light Gradient Boosting Machine (LightGBM) was trained as a pseudo-ground-truth model using all data from the GFP dataset. LightGBM is sometimes used for ranking, classification, and other tasks based on decision tree algorithms. With this setting, the representative model covers all GFP sequences around the parent sequence, and the pseudo-ground truth model covers a wider GFP space as a search space.

[0104] The search space was defined as the mutations in the top 100 sequences of the training data. The search space included 2 to 5 amino acid candidates for each of the 37 candidate mutation sites, as shown in Figure 17. The TPE parameters were set as follows: sampling count 3000, no multivariate (i.e., the objective function had one variable), as shown in Figure 18. The top 10 sequences of the training data were used for warm-start initialization for both MV-TPE and Mean-TPE.

[0105] Typically, in drug discovery, 10 to 100 sequences are evaluated in one batch. In this study, as shown in Figure 10, the pseudo-ground truth model was used to select the top 96 sequences with the highest scores, i.e., the highest estimated values ​​of the objective function, and these sequences were used as proposed sequences for evaluation.

[0106] In the verification, the processes shown in Figures 7 and 8 were performed using MV and mean as the objective functions. The mean was used as a comparative example. MV and mean each contain characteristic values, and GFP brightness was used as the characteristic value. Figure 11 shows the optimization trajectory by TPE, with the mean as the estimated value of the objective function shown in the upper row and MV shown in the lower row. Each shows the results of 10 optimization processes. The dashed lines indicate the estimated values ​​for each sample in each run. Figure 12 shows the relationship between the mean (GP Mean) and standard deviation (GP Std) of the sample values ​​in each run. Using MV as the objective function reduces the standard deviation compared to using the mean. This means that using MV reduces the uncertainty of the estimated values. Figure 13 shows the density distribution of the edit distance of each sample relative to the template GFP. Using MV as the objective function reduces the overall edit distance compared to using the mean. This means that using MV allows GFPs with fewer mutations from the template GFP to be selected as samples.

[0107] Figure 14 shows the distribution of scores for proposed sequences obtained using the pseudo-ground truth model. The score represents an estimate of the brightness of the fluorescence emitted by GFP. Using MV as the objective function results in higher brightness than when the average is used. Figure 15 shows the edit distance of proposed sequences based on the template GFP. Using MV as the objective function results in a smaller edit distance than when the average is used. This indicates that using MV allows sequences with fewer mutations to be sampled, i.e., safe optimization can be achieved. Figure 16 shows the standard deviation (GP Std) of proposed sequences. Using MV as the objective function results in a smaller standard deviation than when the average is used. This confirms that using MV allows sequences with less uncertainty in the brightness, which is a characteristic value, to be sampled, thereby enabling safe optimization.

[0108] <Second Verification Example> Next, a second verification example will be described. In the second verification example, data related to bispecific antibodies was used as training data, and a Gaussian process was executed to train a prediction model. The training data used in this verification example was data indicating the sequences and characteristic values ​​of binding ability of bispecific antibodies whose antigens are MarvelD3 and CD3. Octet values ​​were used as the characteristic values ​​of binding ability. A vector representation obtained using TAPE was used as the input value for the prediction model. This vector representation represents the protein sequence of the bispecific antibody sample.

[0109] MarvelD3 is a tight junction protein with four transmembrane domains. In this study, MarvelD3 was selected as a candidate target for anticancer drugs. The development of bispecific antibodies that cross-link cancer antigens with antigens on T cells is expected to be applied to cancer treatment. In this study, the trained prediction model was used to identify candidate anti-MarvelD3 sequences with superior properties from the lead antibody.

[0110] In this validation, the octet values ​​of antibody sequences were measured multiple times, and the antibody sequences and the octet values ​​obtained from the measurements were used as inputs to the prediction model. Batch measurements typically yield octet values ​​of up to 100 antibody sequences per measurement. The antibodies used in the experiments were obtained using the following procedure. First, plasmids encoding predesigned heavy or light chains were prepared, and recombinant antibodies were transiently expressed in Expi293F cells. Antibodies were captured from the culture supernatant using protein A and eluted into buffer. The eluted buffers were mixed under reducing conditions to prepare MarvelD3 / CD3 bispecific antibodies. Selective heavy chain heterodimerization was achieved by applying charge repulsion between identical heavy chains. The antibody concentration in the buffer was determined by absorbance at 280 nm. The buffer containing the bispecific antibody was then subjected to ion exchange chromatography to confirm the preparation of the desired MarvelD3 / CD3 antibody.

[0111] Octet values ​​were measured using the Octet HTX system. Extracellular vesicles bearing CD81 and human MarvelD3 proteins on their surfaces were captured on a sensor chip using an anti-CD81 antibody. After a 600-second baseline step in D-PBS(-) containing 0.1% BSA, the association and dissociation responses were measured for 900 and 1500 seconds, respectively, in the same buffer containing 20 nM antibody. The antibody binding activity was measured as the wavelength shift between the baseline step and the end of the association phase. Measurements were performed at 30°C and an oscillation rate of 1000 cycles per minute during the baseline step, association, and dissociation phases.

[0112] In optimizing the prediction model, MV-TPE and Mean-TPE were performed as in the first verification example. In this verification, of the candidate sequences obtained by sampling, the top 48 sequences with the estimated value of the objective function were evaluated as proposed sequences.

[0113] Figure 19 shows the distribution of the mean and standard deviation of the estimated values ​​obtained by sampling each time for both MV-TPE and Mean-TPE. The standard deviation indicates uncertainty, and a higher mean estimate indicates a better octet value. According to Figure 19, when the mean (Mean) is used as the objective function, the most frequent standard deviation is 1.0, whereas when MV is used as the objective function, it is 0.3. Furthermore, when the mean is used as the objective function, the most frequent mean value is 2.3, whereas when MV is used, it is 1.5. These results indicate that the uncertainty of the octet value as a characteristic prediction value is higher when the mean is used as the objective function than when MV is used.

[0114] Figure 20 shows a visualization of the sequences obtained by sampling using t-SNE (t-distributed Stochastic Neighbor Embedding). t-SNE is a method for compressing high-dimensional data into low-dimensional data for visualization. Figure 20 shows the distribution of vector representations by TAPE, representing individual sequences, on a two-dimensional plane for the training data (Train), mean (Mean), and MV. Figure 20 shows that when the mean is used as the objective function, the sampled sequences are far from the training data, whereas when MV is used, they are closer to the training data. This indicates that MV-TPE can explore safer regions while suppressing uncertainty compared to Mean-TPE, and can avoid pathological samples.

[0115] Figure 21 shows the mean (GP Mean) of the estimated octet values ​​of the 48 sequences evaluated when the TPE objective function was set to mean (Mean) and when it was set to MV. As shown in Figure 21, the mean octet values ​​obtained using MV were lower than the mean octet values ​​obtained using mean as the objective function. Figure 22 shows the standard deviation (GP Std) of the estimated values ​​for both mean (Mean) and MV. As shown in Figure 22, the standard deviation of the octet values ​​obtained using MV was lower than the standard deviation of the octet values ​​obtained using mean as the objective function. This result suggests that sequences sampled using MV have low uncertainty and are therefore less likely to lose stable expression and binding ability.

[0116] Figure 23 shows the distribution of expression levels of the 48 sequences evaluated for both mean and MV. According to Figure 23, when mean was used as the objective function, the expression levels of the top 48 sequences were extremely low, and insufficient samples were obtained for measuring octet values. In contrast, when MV was used as the objective function, almost all of the top 48 sequences were significantly expressed. Figure 24 shows the distribution of octet values ​​for each sequence for both MV and training data. The distribution trends of octet values ​​are similar between MV and training data. This indicates that using MV as the objective function makes it possible to obtain training data, and therefore sequences with binding ability similar to that of the parent sequence. Furthermore, using MV yielded several sequences with octet values ​​higher than the maximum value in the training set.

[0117] <Summary> The above-described demonstration example demonstrates that offline model-based optimization (MBO) according to this application example can be applied to building-block-based molecular design, a common method in drug design, and can be implemented in silico (i.e., computer-based). Here, the tree-structured Parzen estimator (TPE), a Bayesian optimization technique, can be applied to building-block-based molecular design for optimizing combinations of categorical variables. Furthermore, it has been shown that applying MV, a technique used in financial engineering, to offline MBO can avoid pathological behavior and contribute to the search for safe sequences. Specifically, incorporating risk averse prediction into MBO for molecular design prevents overestimation of the extrapolation range in predicting the properties of candidate molecules and prevents the proposal of molecules with low prediction reliability. Furthermore, the above-described demonstration example demonstrates that this embodiment is useful for searching for safe sequences in therapeutic antibody design.

[0118] The molecular design device 1 of the embodiment configured as described above performs Bayesian optimization using a model that obtains characteristic quality values ​​based on drug candidate molecular sequence information as an objective function, thereby further reducing the burden required for drug discovery.

[0119] (Modification) The molecular design device 1 may be implemented using a plurality of information processing devices communicably connected via a network. In this case, each functional unit of the molecular design device 1 may be distributed and implemented in the plurality of information processing devices. The drug discovery system 100 may also include a plant (not shown) that generates an optimal molecule estimated by the molecular design device 1. The molecular design device 1 outputs output information indicating the biological sequence of the optimal molecule estimated by the above-described method to the plant. The plant executes a process of generating a molecular compound having a biological sequence indicated by the output information input from the molecular design device 1.

[0120] All or part of the functions of the molecular design apparatus 1 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The program may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems. The program may be transmitted via a telecommunications line.

[0121] The molecular design device 1 is an example of an estimation device. The above functions may be realized by other types of devices, such as devices whose main function is not molecular design, or information devices equipped with a general-purpose computer system.

[0122] For example, this embodiment may be realized as a system including a processor and a memory. In this system, the memory may be configured to store one or more instructions, which may cause the processor to execute the following steps: for a molecule that is an element of a sequence information set, which is a collection of sequence information of a plurality of different molecules, calculating a predicted property value of the molecule using a prediction model for predicting the property of the molecule from the sequence information of the molecule and estimating the uncertainty of the prediction; and searching for candidate molecules having more desirable properties based on the predicted property value and the prediction uncertainty. Furthermore, the instructions may cause the processor to execute the following steps: acquiring a sequence information set, which is a collection of sequence information of a plurality of different molecules, using a combinatorial optimization algorithm; predicting a predicted property value of a molecule that is an element of the sequence information set using a prediction model for predicting the property of the molecule from the sequence information of the molecule; updating extraction parameters of the combinatorial optimization algorithm based on the predicted property value of each molecule so that molecules having more desirable properties are included; and searching for molecules having desirable properties based on the predicted property value.

[0123] This embodiment may be realized as a non-transitory computer-readable medium storing one or more instructions. The instructions may cause a computer to execute the following steps: for a molecule that is an element of a sequence information set, which is a collection of sequence information of a plurality of different molecules, calculating a predicted property value of the molecule using a prediction model for predicting the property of the molecule from the sequence information of the molecule and estimating the uncertainty of the prediction; and searching for candidate molecules having more desirable properties based on the predicted property value and the prediction uncertainty. The instructions may also cause the computer to execute the following steps: an acquisition step of acquiring a sequence information set, which is a collection of sequence information of a plurality of different molecules, using a combinatorial optimization algorithm; a prediction step of calculating predicted property values ​​of molecules that are elements of the sequence information set using a prediction model for predicting the property of the molecule from the sequence information of the molecule; an update step of updating extraction parameters of the combinatorial optimization algorithm based on the predicted property value of each molecule so that molecules having more desirable properties are included; and a search step of searching for molecules having desirable properties based on the predicted property values.

[0124] This embodiment may be realized as a computer-implemented method using an artificial intelligence engine. The computer-implemented method includes a control step of performing combinatorial optimization to estimate an optimal molecule, which is a drug candidate molecule that provides an optimal value of a desired property, among each of the drug candidate molecules represented by each element of the candidate set, by performing combinatorial optimization, and a step of outputting information on the optimal molecule, wherein the objective function of the combinatorial optimization is a model that executes a mathematical model that estimates a characteristic value of the desired property based on the drug candidate molecule sequence information and obtains an uncertainty of the estimation result, and obtains a characteristic quality value calculated based on the characteristic value and the uncertainty.

[0125] Although several embodiments have been described above, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are intended to be included within the scope and spirit of the invention, as well as within the scope of the invention described in the claims and their equivalents. All documents cited in this specification are incorporated herein by reference.

[0126] 1... molecular design device, 11... control unit, 12... input unit, 13... communication unit, 14... memory unit, 15... output unit, 100... drug discovery system, 111... inference unit, 112... input control unit, 113... communication control unit, 114... memory control unit, 115... output control unit, 91... processor, 92... memory

Claims

1. An information processing system for identifying molecules suitable as drug candidates, comprising: a property prediction unit that calculates a property predicted value of a molecule that is an element of a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule, and estimates the uncertainty of the prediction; and a candidate molecule identification unit that searches for candidate molecules having desired properties based on the property predicted value and the estimated value of the uncertainty of the prediction.

2. The information processing system of claim 1, further comprising a prediction information processing unit that calculates a characteristic quality value based on the characteristic prediction value and an estimated value of the uncertainty of the prediction, and the candidate molecule identification unit identifies at least one candidate molecule from the elements of the building block combination information set based on the characteristic quality value.

3. The information processing system according to claim 2, wherein the characteristic quality value is an output value of a predetermined function having as input two variables, the characteristic prediction value and an estimate of the uncertainty of the prediction.

4. The information processing system according to claim 2 or 3, wherein the prediction information processing unit calculates a mean variance as an objective function that gives the characteristic quality value.

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

6. An information processing system according to any one of claims 2 to 5, comprising a building block combination information processing unit that uses a combinatorial optimization algorithm to obtain the building block combination information set based on the predicted property value or the property quality value of the molecule.

7. An information processing system according to any one of claims 2 to 6, wherein the characteristic quality value increases as the characteristic prediction value increases, and the building block combination information set is obtained using a combinatorial optimization algorithm as the prediction uncertainty estimate decreases, and extraction parameters of the combinatorial optimization algorithm are updated based on the molecular characteristic prediction value or the characteristic quality value.

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

9. The information processing system according to any one of claims 1 to 8, wherein the predictive model is a predictive model generated by learning based on building block combination information of a plurality of molecules for training and the results of property evaluation of the molecules.

10. The information processing system according to any one of claims 1 to 9, wherein the building block combination information is 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 a nucleic acid, a peptide, a cyclic peptide, a protein, an antibody, and a low molecular weight compound.

12. The information processing system according to any one of claims 1 to 11, wherein the characteristic 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 a target molecule, and the property is a binding ability to the target molecule.

14. An information processing method in an information processing system for identifying molecules suitable as drug candidates, comprising: a step of calculating a predicted property value of a molecule that is an element of a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, using a predictive model for predicting the properties of the molecule from the building block combination information of the molecule, and estimating the uncertainty of the prediction; and a step of searching for candidate molecules having more desirable properties based on the predicted property value and the uncertainty of the prediction.

15. A program for causing a computer to execute the steps of: calculating a predicted characteristic value of a molecule that is an element of a building block combination information set, which is a set of building block combination information of a plurality of different molecules, using a prediction model for predicting the characteristics of the molecule from the building block combination information of the molecule, and estimating the uncertainty of the prediction; and searching for candidate molecules having more desired characteristics based on the predicted characteristic value and the uncertainty of the prediction.

16. A method for producing a molecular compound, comprising: an input step of accessing a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, and inputting the building block combination information set into a prediction model; an inference step of searching for a molecule having more desired properties from the building block combination information set based on a property prediction value and an estimated value of prediction uncertainty for each molecule included in the building block combination information set, which are output from the prediction model, and identifying the molecule as a candidate molecule; an output step of outputting building block combination information related to the candidate molecule; and a generation step of generating the molecular compound having a molecular sequence indicated in the building block combination information.

17. An information processing system for identifying molecules suitable as drug candidates, comprising: a building block combination information processing unit that acquires a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, in accordance with a combinatorial optimization algorithm; a property prediction unit that calculates a predicted property value of a molecule that is an element of the building block combination information set, using a prediction model for predicting the properties of the molecule from the building block combination information of the molecule; and a candidate molecule identification unit that searches for candidate molecules having desired properties based on the predicted property value, wherein the building block combination information processing unit updates extraction parameters of the combinatorial optimization algorithm so as to include molecules having more desired properties based on the predicted property value of the molecule.

18. An information processing method in an information processing system for identifying molecules suitable as drug candidates, comprising the steps of: obtaining a building block combination information set, which is a collection of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; calculating a predicted property value of a molecule that is an element of the building block combination information set using a predictive model for predicting the properties of the molecule from its building block combination information; searching for a molecule having more desirable properties based on the predicted property value; and updating extraction parameters of the combinatorial optimization algorithm so that molecules having more desirable properties are included based on the predicted property value of the molecule.

19. A program for causing a computer to execute the following steps: an acquisition step of acquiring a building block combination information set, which is a set of building block combination information of a plurality of different molecules, in accordance with a combinatorial optimization algorithm; a prediction step of calculating a predicted property value of a molecule that is an element of the building block combination information set, using a prediction model for predicting the properties of the molecule from its building block combination information; an update step of updating extraction parameters of the combinatorial optimization algorithm, based on the predicted property value for each molecule, so that molecules having more desired properties are included; and a search step of searching for molecules having desired properties based on the predicted property values.

20. A method for producing a molecular compound, comprising: an acquisition step of acquiring a building block combination information set, which is a set of building block combination information of a plurality of different molecules, according to a combinatorial optimization algorithm; an input step of inputting the building block combination information set into a prediction model; an update step of updating extraction parameters of the combinatorial optimization algorithm based on a property prediction value for each molecule included in the building block combination information set output from the prediction model so that molecules having more desired properties are included; a step of searching for a molecule having more desired properties from the building block combination information set based on the property prediction value and identifying it as a candidate molecule; an output step of outputting building block combination information related to the candidate molecule; and a generation step of generating the molecular compound having a molecular sequence indicated in the building block combination information.

Images