Compound docking calculation processing device, compound docking calculation processing method, and program

Abstract

The ability to perform combinatorial optimization in compound docking calculations based on fragments. [Solution] The compound docking calculation processing device includes: a fragment decomposition unit that acquires fragment data indicating fragments of a compound to be used for docking calculations with a target biomolecule; a fragment docking calculation unit that performs a fragment docking calculation for each fragment; a fragment interaction evaluation unit that calculates evaluation values ​​for fragment placement candidates and their relationships based on the results of the fragment docking calculation for each fragment placement candidate and the positional relationships between the fragment placement candidates; an evaluation function generation unit that generates an evaluation function that gives an evaluation value for the fragment pose set based on the evaluation value; and a combinatorial optimization calculation processing unit that calculates the fragment pose set by determining the state for optimizing the evaluation function based on combinatorial optimization.

Description

[Technical Field]

[0001] The present invention relates to a compound docking calculation processing apparatus, a compound docking calculation processing method, and a program. [Background technology]

[0002] In drug discovery, the use of computers to improve efficiency is being widely pursued. Among these methods, Structure-Based Drug Design (SBDD) is a technique that uses the three-dimensional structural information of target proteins and other biomolecules to search for drug candidate compounds. Compared to Ligand-Based Drug Design (LBDD), which searches for drug candidate compounds by referring to the characteristics of known active compounds of target proteins and other biomolecules, SBDD is said to be able to select drugs with novel skeletons.

[0003] As a method for realizing SBDD (Structural-Based Drug Determination), there is a growing emphasis on methods that perform docking calculations between the three-dimensional structural information of target proteins and other biomolecules and compounds that could potentially be drug candidates, a number that has been rapidly increasing in recent years. Accordingly, there is a demand for high-speed docking calculations with compounds. In docking calculations, it is necessary to consider the rotation and translation of the entire compound within the binding site of the target protein or other biomolecule, and compounds have many internal degrees of freedom due to the rotation of some covalent bonds. Even with conventional techniques, processing one case can be completed in about tens of seconds, but if there are hundreds of millions of compounds to be evaluated, the computation time becomes impractically enormous.

[0004] In recent years, the concept of ultra-large compound libraries has emerged, and with the rise of generative deep learning, even virtually generated compounds can become drug candidates. This has led to demands for docking calculations of hundreds of millions to billions of compounds. Currently, docking calculations for such a vast number of compounds are impossible. Therefore, the number of candidates is reduced through simpler selection methods, but expanding the number of compounds that can be subjected to docking calculations is considered important to avoid missing out on potentially useful drug candidates.

[0005] On the other hand, "fragment-based compound docking" is known as a method for achieving high-speed docking calculations with a vast number of compounds to be evaluated. In fragment-based compound docking, a compound is broken down into substructures called fragments, and then fragment docking calculations are performed between each fragment and target biomolecules such as proteins. The results of the fragment docking calculations for each fragment are then combined to obtain the docking calculation results for the compound before it was broken down.

[0006] While there is no single unique method for decomposing a compound into fragments, it is common practice to decompose them so that each fragment has no internal degrees of freedom, and standard decomposition methods exist. Docking calculations for each fragment after decomposition are straightforward, but the resulting fragment docking calculations often result in positions where each fragment collides with others or where covalent bonding with other fragments is impossible. Therefore, it is necessary to select a consistent combination from the various fragment docking calculation results ("combinatorial optimization"), which arguably complicates the problem. However, fragment-based compound docking calculations remain highly attractive because compounds in drug discovery are created from known compounds through partial substitution, resulting in extremely similar structures, and the number of unique fragments constituting a compound is relatively small. However, conventional fragment-based compound docking calculation methods discarded intermediate results and recalculated every compound, even if the constituent fragments were similar, resulting in computation time proportional to the number of compounds. In contrast, a fragment-based compound docking calculation, which first performs docking calculations for a relatively small number of fragments and then combines them as appropriate, may be able to process data faster than conventional methods that require computation time proportional to the number of compounds.

[0007] Fragment-based compound docking calculations require combinatorial optimization. Conventional techniques reduced combinatorial optimization to a graph theory problem called the maximum clique problem and solved it using classical algorithms (Non-Patent Literature 1). Another conventional technique completely avoided the process of searching for consistent combinations and instead generated multiple "possible conformations" externally for each compound, utilizing the information of the generated "possible conformations" (Non-Patent Literature 2). [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] "Journal of Molecular Graphics and Modelling," July 2007, Vol. 26, No. 1, pp. 198-212. [Non-Patent Document 2] "ACS Omega," August 19, 2022, Vol. 7, No. 34, pp. 30265-30274 [Non-Patent Document 3] "Bioinformatics", March 30, 2017, Volume 33, Issue 23, p.3836-3843 [Non-Patent Document 4] "Journal of Chemical Information and Modeling," July 19, 2021, Vol. 61, No. 8, pp. 3891-3898. [Non-Patent Document 5] "Entropy", February 28, 2022, Volume 24, Issue 3, p.354 [Overview of the project] [Problems that the invention aims to solve]

[0009] As mentioned above, the potential demand for fragment-based compound docking calculations is rapidly increasing in line with the growing scale of compound libraries, and there is a strong need to perform combinatorial optimization calculations inherent in fragment-based compound docking calculations. The goal of this combinatorial optimization is to select an appropriate set of fragments (fragment pose set) from a group of possible fragment placement candidates on the surface of a target biomolecule that are consistent with each other in constructing the compound.

[0010] The present invention has been made in view of the above points, and provides a compound docking calculation processing device, a compound docking calculation processing method, and a program that can perform combinatorial optimization in compound docking calculations based on fragments. [Means for solving the problem]

[0011] The compound docking calculation processing device of the embodiment includes a fragment decomposition unit, a fragment docking calculation unit, a fragment interaction evaluation unit, an evaluation function generation unit, and a combinatorial optimization calculation processing unit. The fragment decomposition unit acquires fragment data that shows fragments, which are partial chemical structures of a compound to be docked with a target biomolecule. The fragment docking calculation unit performs a fragment docking calculation between each fragment shown in the fragment data and the target biomolecule. The fragment interaction evaluation unit calculates an evaluation value between each fragment arrangement candidate and each other based on the results of the fragment docking calculation for each fragment arrangement candidate, which is a candidate arrangement when the fragment binds to the target biomolecule, and the positional relationship between the fragment arrangement candidates. The evaluation function generation unit generates an evaluation function based on the evaluation value that gives an evaluation value of the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule. The combinatorial optimization calculation processing unit calculates the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule by determining the state for optimizing the evaluation function based on combinatorial optimization. [Brief explanation of the drawing]

[0012] [Figure 1] This figure shows an example of the functional configuration of the computing processing device 1 according to the first embodiment of the present invention. [Figure 2] This figure shows an example of the processing content of each functional unit of the computing processing device 1 according to the first embodiment of the present invention. [Figure 3] This figure shows an example of a comparison between a polyomino puzzle and a compound docking calculation based on fragments according to the first embodiment of the present invention. [Figure 4] This figure shows an example of a comparison between a polyomino puzzle and a fragment-based compound docking calculation for the representation of binary variables, element fitness, collisions between elements, coordination between elements, and the number of element selections according to the first embodiment of the present invention. [Figure 5]It is a diagram showing an example of comparison between a polyomino puzzle regarding collision between elements according to the first embodiment of the present invention and a compound docking calculation based on fragments. [Figure 6] It is a diagram showing an example of comparison between a polyomino puzzle regarding cooperation between elements according to the first embodiment of the present invention and a compound docking calculation based on fragments. [Figure 7] It is a diagram showing an example of the flow of processing of a compound docking calculation based on fragments according to the first embodiment of the present invention. [Figure 8] It is a diagram showing an example of a state in which fragment placement candidates are obtained by fragment docking calculation from fragment data according to the first embodiment of the present invention. [Figure 9] It is a diagram showing an example of a state in which a Hamiltonian according to the first embodiment of the present invention is generated. [Figure 10] It is a diagram showing an example of a state in which a state minimizing the Hamiltonian is obtained by combinatorial optimization using a pseudo quantum annealer according to the first embodiment of the present invention. [Figure 11] It is a diagram showing an example of a state in which a covalent bond is formed between fragments according to the first embodiment of the present invention. [Figure 12] It is a diagram showing the structures of a target protein and a compound according to an example of the first embodiment of the present invention. [Figure 13] It is a diagram showing docking calculation parameters according to an example of the first embodiment of the present invention. [[ID=@25]] [Figure 14] It is a diagram showing the results of a docking calculation with a fragment alone according to an example of the first embodiment of the present invention. [Figure 15] It is a diagram showing the number of fragment placement candidates obtained as a result of a docking calculation with a fragment alone according to an example of the first embodiment of the present invention, and the maximum and minimum values of their binding free energy scores. [Figure 16] It is a diagram showing the distribution of binding free energy scores for each fragment according to an example of the first embodiment of the present invention. [Figure 17]This figure shows the distribution of binding free energy scores for each fragment according to an embodiment of the first embodiment of the present invention. [Figure 18] This figure shows the distribution of binding free energy scores for each fragment according to an embodiment of the first embodiment of the present invention. [Figure 19] This figure shows the distribution of binding free energy scores for each fragment according to an embodiment of the first embodiment of the present invention. [Figure 20] This is a scatter plot of the Hamiltonian value after performing a combinatorial optimization calculation according to an embodiment of the first embodiment of the present invention, and the RMSD for the correct structure of the docking calculation. [Figure 21] This figure shows the fragment pose set of the best local optimum among the combinatorial optimization calculations according to an embodiment of the first embodiment of the present invention. [Figure 22] This figure shows the docking calculation results of a compound after local structure optimization according to an example of the first embodiment of the present invention. [Figure 23] This figure shows an example of the functional configuration of the computing processing device 1a according to a second embodiment of the present invention. [Figure 24] This figure shows an example of the processing content of each functional unit of the computing processing device 1a according to the second embodiment of the present invention. [Figure 25] This figure shows an example of the processing flow for compound docking calculations based on fragments according to a second embodiment of the present invention. [Modes for carrying out the invention]

[0013] (First embodiment) Embodiments of the present invention will be described in detail below with reference to the drawings. The compound docking calculation processing device according to this embodiment is referred to as the calculation processing device 1. Figure 1 is a diagram showing an example of the functional configuration of the calculation processing device 1 according to this embodiment. The calculation processing device 1 is a device that performs compound docking calculations based on fragments based on combinatorial optimization. The objective of this combinatorial optimization is to select an appropriate set of fragments (fragment pose set) that are mutually consistent in constructing a compound from a group of candidate fragment arrangements that can be considered on the surface of a target biomolecule, which is the target biomolecule.

[0014] In this embodiment, the bond may be either a covalent bond or a non-covalent bond. Non-covalent bonds include, for example, electrostatic interactions, van der Waals forces, and hydrophobic interactions.

[0015] The computational processing unit 1 comprises a fragment decomposition unit 10, a fragment docking calculation unit 11, a fragment interaction evaluation unit 12, an evaluation function generation unit 13, a combinatorial optimization calculation processing unit 14, a post-processing unit 15, a local structure optimization unit 16, an output unit 17, and a storage unit 2. The fragment decomposition unit 10, the fragment docking calculation unit 11, the fragment interaction evaluation unit 12, the evaluation function generation unit 13, the combinatorial optimization calculation processing unit 14, the post-processing unit 15, the local structure optimization unit 16, and the output unit 17 are each implemented, for example, by the CPU reading a program from ROM (Read Only Memory), expanding it into RAM (Random Access Memory), and executing processing according to that program. The ROM and RAM are included in the storage unit 2.

[0016] The storage unit 2 stores various types of information. For example, the storage unit 2 stores compound data 20. Compound data 20 is data that identifies compounds. The storage unit 2 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device. The compound data 20 may also be stored in a database provided outside the computing device 1.

[0017] The computing processing unit 1 may be implemented as a single computer or as a virtual server. Each functional unit of the computing processing unit 1 may be distributed across multiple servers. The computing processing unit 1 may also be implemented as a cloud server.

[0018] Here, with reference to Figure 2, we will describe each functional unit of the computing device 1. Figure 2 is a diagram showing an example of the processing content of each functional unit of the computing device 1 according to this embodiment. In this embodiment, in the compound docking calculation performed by the computing device 1, the number of input compounds is one, and the output is the docking calculation result for the compound to the target biomolecule. The target biomolecule is, for example, a protein (referred to as the target protein). The target biomolecule may also be a biomacromolecule such as nucleic acid, nucleic acid-protein complex, or a molecule in a living organism. Below, we will describe an example where the target biomolecule is a target protein. The docking calculation result is a list of candidate poses for the compound to bind to the target protein, and an evaluation score for each candidate pose. In the explanation of Figure 2, the compound is assumed to consist of four fragments, for example.

[0019] The fragment decomposition unit 10 decomposes the compound into fragments. The decomposition by the fragment decomposition unit 10 generates fragment data, which is data representing the fragments. A fragment is a partial chemical structure of a compound. In this embodiment, the fragments are assumed to have no internal degrees of freedom. Internal degrees of freedom are degrees of freedom due to the rotation of a portion of the covalent bonds. As an example, the fragment decomposition unit 10 decomposes the compound into fragments based on the decomposition method described in Non-Patent Literature 3 and obtains the fragment data generated by the decomposition. The fragment decomposition unit 10 may use other decomposition methods in this decomposition. The fragment decomposition unit 10 may also obtain the fragment data generated by an information processing device separate from the computing processing device 1 from that information processing device.

[0020] Furthermore, when the original compound is broken down into fragments, it is possible to attach capping atoms to each atom that was at either end of the broken covalent bond. These caps indicate the presence of an atom or group of atoms at the cap's position. These caps may also be used as potential covalent bonding points (attachment points) when evaluating the coordination between fragments, as described later.

[0021] The fragment docking calculation unit 11 performs a docking calculation between each fragment and the target protein for each fragment indicated by the fragment data. This allows the fragment docking calculation unit 11 to obtain fragment docking calculation results for each fragment. The fragment docking calculation results consist of a group of candidate fragment placements on the target protein for that fragment, and a fragment docking score for each candidate fragment placement. Candidate fragment placements are candidate positions for when the fragment binds to the target protein. These placements are specified by the orientation of the fragment and its position in three-dimensional space. Note that candidate fragment placements are also referred to as fragment poses.

[0022] The fragment docking calculation unit 11 performs a fragment docking calculation with the target protein for each of the fragment groups (for example, four fragments) indicated by the fragment data. Known docking algorithms and known docking score calculation methods may be used for the fragment docking calculation. For example, the fragment docking calculation unit 11 uses a predetermined algorithm that comprehensively performs fragment rotation and translation as the docking calculation algorithm. Also, for example, the fragment docking calculation unit 11 uses parameters described in Non-Patent Document 4 as the evaluation value of the interaction energy as the docking score calculation method. Here, the interaction energy is the energy of the interaction between a single fragment and the target protein.

[0023] In the processing described later, in order to optimize the combinations between fragment placement candidates, if the parameters specifying rotation and translation are changed too finely, the number of fragment placement candidates may become too large. For this reason, the aforementioned algorithm may be designed to allow adjustment so that the amount of change in the parameters specifying rotation and translation exceeds a predetermined amount.

[0024] Furthermore, in order to introduce diversity into the candidates selected by the fragment interaction evaluation unit 12 described later, the fragment docking calculation unit 11 may perform the fragment docking calculation not once for the entire binding site, but multiple times independently for each sub-region within the binding site. For example, the binding site of the target protein may be divided into cubic regions with sides of 2 Å, and the fragment docking calculation may be performed independently for each region. In this case, the fragment interaction evaluation unit 12 can handle the fragment docking calculation results for each region using independent criteria. As another example, the fragment interaction evaluation unit 12 may also retain fragment placement candidates from the fragment docking calculation results for the entire target protein, taking into account diversity in position.

[0025] The fragment interaction evaluation unit 12 calculates evaluation values ​​for each fragment placement candidate and for the relationships between the fragment placement candidates based on the fragment docking score for each candidate and the positional relationships between the candidates. The positional relationships between the candidates include collisions and cooperation between them.

[0026] In this embodiment, the fragment interaction evaluation unit 12 performs the above-described processing as a primary evaluation process and a secondary evaluation process on the fragment docking calculation results from the fragment docking calculation unit 11. In the primary evaluation process, the contribution of each individual fragment placement candidate to the interaction with the target protein is evaluated for the vast group of fragment placement candidates. In the secondary evaluation process, the collision and cooperation between two fragment placement candidates when they are paired together is evaluated for the vast group of fragment placement candidates.

[0027] In the primary evaluation process, the fragment interaction evaluation unit 12 calculates an evaluation value for the binding of each fragment placement candidate to the target protein based on the fragment docking score for that fragment placement candidate alone. The evaluation value calculated by the primary evaluation process is referred to as the primary evaluation value.

[0028] In one example of this embodiment, the fragment interaction evaluation unit 12 calculates a primary evaluation value for the binding of a single fragment to the target protein based on the size of the fragment corresponding to the fragment placement candidate, the number of fragment placement candidates for each subregion within the binding site of the target protein, and the fragment docking score for the fragment placement candidate. The value indicating the size of the fragment is calculated based on one or more of the following, for example, the number of heavy atoms and the spatial size.

[0029] As an example, the fragment interaction evaluation unit 12 calculates a primary evaluation value by dividing the fragment docking score obtained by the fragment docking calculation unit 11 by a value indicating the size of the fragment. This primary evaluation value corresponds to the contribution of each fragment size to the fragment docking score. As another example, the fragment interaction evaluation unit 12 may calculate the primary evaluation value by adding a value corresponding to the contribution of each fragment size to the fragment docking score to the original fragment docking score. As yet another example, the fragment interaction evaluation unit 12 may calculate the primary evaluation value by adding a value corresponding to the rarity of locations where other fragments cannot be placed, according to the number of candidate fragment placements for each partial region within the binding site of the target protein.

[0030] Calculating a primary evaluation value based on a value corresponding to the contribution of each fragment size to the fragment docking score is similar to the concept of "ligand efficiency" in drug discovery, which is considered for the compound as a whole. In the concept of "ligand efficiency," it is assumed that if a small fragment makes a large contribution, the remaining volume where that fragment is not located can be used to further increase the binding affinity.

[0031] On the other hand, larger fragments may be rewarded to favor them. For example, the value obtained by adding points to the fragment docking score of a large fragment may be used as the primary evaluation value. This is based on the idea that a compound structure that uses large complex rings as fragments is more characteristic of a compound than one filled with small, fragmentary fragments. In addition, the value of the fragment docking score obtained by the fragment docking calculation unit 11 may be used directly as the primary evaluation value in the primary evaluation process.

[0032] Thus, the primary evaluation process can be adjusted to reflect various drug discovery knowledge, allowing for adjustments to the evaluation of the superiority or inferiority of each fragment placement candidate. To reduce the computational burden of the secondary evaluation process and beyond, fragment placement candidates whose primary evaluation value is below a predetermined level may be excluded in the primary evaluation process. In that case, a predetermined number of fragment placement candidates may be retained for each region within the binding site, as described above, to ensure spatial diversity.

[0033] In the secondary evaluation process, the fragment interaction evaluation unit 12 calculates evaluation values ​​for collisions between fragment placement candidates and evaluation values ​​for cooperation between fragment placement candidates. In the secondary evaluation process, for all fragment placement candidates that were not excluded in the primary evaluation process, the fragment interaction evaluation unit 12 calculates evaluation values ​​for collisions and cooperation between two fragment placement candidates when those two candidates are paired together. In other words, the secondary evaluation process evaluates both deductions for collisions between fragment placement candidates and additions for cooperation. The evaluation value for collisions calculated in the secondary evaluation process is referred to as the collision evaluation value. The evaluation value for cooperation calculated in the secondary evaluation process is referred to as the cooperation evaluation value.

[0034] For collision evaluation, known methods for determining collisions between partial chemical structures can be used. However, in that case, as mentioned above, the rotation and translation of candidate arrangements are represented more coarsely than in known collision determination methods (that is, the change in the parameters specifying rotation and translation is adjusted to be greater than or equal to a predetermined amount, so as not to result in too many candidate fragment arrangements). Therefore, in the collision evaluation in this embodiment, it is preferable to relax the conditions for determination compared to known collision determination methods. For example, a certain percentage of collisions may be allowed. As another example, if a collision is possible if the orientation or position of candidate fragment arrangements is slightly shifted even if they do not appear to be colliding, a penalty of less than or equal to a predetermined amount may be imposed on the collision evaluation.

[0035] Regarding the evaluation of cooperation, the cooperation evaluation value is calculated as follows: For two paired fragment placement candidates, if the aforementioned covalent bonding points (attachment points) are located in close proximity to each other and at an angle that allows for covalent bonding, then the two paired fragment placement candidates have the potential to form a covalent bond and connect. Accordingly, points are added to the cooperation evaluation value for that pair. In this case, to avoid overreacting to placement errors, the criteria for determining whether a covalent bond is formed may be relaxed.

[0036] In the secondary evaluation process, if the distance between two paired fragment placement candidates is sufficiently large, the relationship between the two paired fragment placement candidates may be evaluated as 0 without performing the detailed evaluation of collisions and coordination described above. For example, the collision evaluation value and the coordination evaluation value may both be set to 0. In the case of quantum annealing, which will be described later, the evaluation values ​​obtained by the secondary evaluation process (collision evaluation value and coordination evaluation value) are realized as relationships between binary variables, so evaluating weak relationships that do not have effect for a pair as 0 has the advantage of reducing the computational cost.

[0037] The evaluation function generation unit 13 generates an evaluation function based on the evaluation values ​​calculated by the fragment interaction evaluation unit 12. The evaluation function is a function that gives an evaluation value of the fragment pose set of the fragments constituting the compound on the surface of the target protein. In this embodiment, the evaluation function generated by the evaluation function generation unit 13 is a Hamiltonian, which is a function that mimics the energy of the system, and therefore the lower its value, the better the solution. This Hamiltonian can be designed to correspond to, for example, a quadratic unconstrained binary optimization (QUBO) problem, which is a type of mathematical optimization problem, and in that case it can be uniquely represented by a QUBO matrix. The evaluation function generation unit 13 generates a QUBO matrix that represents the Hamiltonian, and the generated QUBO matrix is ​​input to the combinatorial optimization calculation processing unit 14.

[0038] In this embodiment, for example, compound docking calculations based on fragments that consider the internal degrees of freedom of the compound are formulated as a QUBO problem and performed. Any method can be used to model the evaluation values ​​(primary evaluation value, collision evaluation value, and cooperative evaluation value) obtained by the fragment interaction evaluation unit 12 as a QUBO problem. As an example, good fragment placement candidates (fragment placement candidates that remain without being excluded in the primary and secondary evaluation processes) as the result of the fragment docking calculation for each fragment can be used as candidates, and a binary variable indicating whether or not to adopt these fragment placement candidates can be used to use the primary evaluation value as the primary term of the QUBO problem, and the secondary evaluation values ​​(collision evaluation value and cooperative evaluation value) as the secondary term of the QUBO problem. The primary and secondary terms of the QUBO problem are collectively represented as a QUBO matrix.

[0039] The following shows an example of a specific configuration in this embodiment. When formulating the compound docking calculation based on fragments as a QUBO problem, the following elements should be considered. (i) Represent a compound using multiple fragments; (b) The pose resulting from independently docking the fragments is considered a candidate; (h) Each fragment will only have one pose selected from the possible pose set; (ii) Consider collisions between fragments; (e) For covalently bonded fragments, consider the covalent bond distance between pauses.

[0040] These elements have similarities to the components of the Hamiltonian when protein-compound docking calculations are approximated as a generalized polyomino puzzle (see Non-Patent Document 5). Although the polyomino puzzle itself is merely a geometric puzzle, it has properties very similar to protein-compound docking calculations and will be referenced later for illustrative purposes. Figure 3 shows a comparison between the polyomino puzzle and a fragment-based compound docking calculation.

[0041] In combinatorial optimization, binary variables that take values ​​of 0 or 1 serve as explanatory variables for the problem (sometimes spins that take values ​​of -1 or 1 are used). The selection of binary variables for the problem is an important point, as it greatly changes the difficulty of the combinatorial optimization problem. As described above, in this embodiment, the problem is formulated by treating each candidate fragment placement in 3D space as a single binary variable.

[0042] Since there are countless potential fragment placements in three-dimensional space, it is necessary to limit their number based on some condition. The simplest condition is to select fragment placement candidates that have good elemental fit (fragment docking score between the target protein and each candidate fragment placement). However, it has been pointed out that among the candidate fragment placements that constitute the compound structure in the three-dimensional structure of a protein-compound complex, there are fragments that do not have good binding free energy with the target protein themselves, but that act as linkers connecting fragment placements with good binding free energy. Therefore, when selecting candidate fragment placements as binary variables, it is necessary to consider not only good fragment docking scores but also ensuring diversity in positions.

[0043] The element's fit is the local gain obtained by selecting the corresponding fragment placement candidate (setting the value of the binary variable to 1). In fragment-based compound docking calculations, the compound's docking score is expressed as the sum of the fragment docking scores of each selected fragment placement candidate. Since the sum of the fragment docking scores of the fragment placement candidates corresponding to each binary variable can be described by a first-order term, each fragment placement candidate is represented by i, and its fragment docking score to the target protein is represented by ΔG. i The binary variable corresponding to i is x i In this case, the bond free energy score of the compound can be written as shown in equation (1).

[0044]

number

[0045] Figure 4 shows a comparison between polyomino puzzles and fragment-based compound docking calculations in terms of binary variable representation, element fit, element collisions, element coordination, and the number of element selections.

[0046] Regarding collisions between elements, for example, in a polyomino puzzle, each polyomino must not overlap with another. Similarly, a state in which colliding fragment arrangement candidates are selected simultaneously is inappropriate as a compound structure, and a penalty should be imposed for such collisions. Therefore, the penalty for collisions is expressed as shown in equation (2) using the function clash(i,j). The range of possible values ​​for the function clash(i,j) is a real number greater than or equal to 0.

[0047]

number

[0048] Figure 5 shows a comparison of element-to-element collisions between polyomino puzzles and fragment-based compound docking calculations.

[0049] On the other hand, the cooperation between elements, in contrast to the collision between elements described above, is expressed using the function conn(i,j) as shown in equation (3). The range of values ​​that the function conn(i,j) can take is a real number less than or equal to 0.

[0050]

number

[0051] Figure 6 shows a comparison of the coordination between elements between polyomino puzzles and fragment-based compound docking calculations.

[0052] Let's explain the number of elements to select. Compound docking calculation estimates the binding mode between a target protein and a compound included in compound data 20. For example, suppose a compound is composed of three fragments (fragment a, fragment b, and fragment c). In this case, even if there are two or more fragments with the same structural formula, these two or more fragments are treated as separate fragments. In this case, to reconstruct the compound structure, only one pose of fragment a must be selected. Similarly, only one pose each must be selected for fragment b and fragment c. Note that this is exactly the same as the condition in a polyomino puzzle that "each polyomino is used once," and the constraint on the number of elements to select can be written as shown in equation (4). Here, F is the set of fragments that make up the compound, f i This is the fragment corresponding to fragment placement candidate i.

[0053]

number

[0054] Using the four terms shown by equations (1) to (4) above, the Hamiltonian H, which is the objective function to be minimized, is designed as follows.

[0055]

number

[0056] Here, coefficients A, B, C, and D are the weight parameters for each term. Each of coefficients A, B, C, and D is a real number greater than or equal to 0. In one example of this embodiment, for each term's coefficient (weight parameter), the value of coefficient A was first set to 1, and then a preliminary experiment was conducted. For the remaining coefficients, the values ​​of coefficient B, C, and D were set to 5, 5, and 25, respectively.

[0057] The evaluation function generation unit 13 generates a QUBO matrix representing the Hamiltonian H shown by equation (5). The evaluation function generation unit 13 outputs the generated QUBO matrix in a predetermined format. The predetermined format is, for example, the MM (Matrix Market) format, which is standardly used to represent matrices, or the HDF5 (Hierarchical Data Format 5) format, which is a binary representation.

[0058] The combinatorial optimization calculation processing unit 14 calculates the fragment pose set of the compound fragments on the surface of the target protein by determining the state that optimizes the evaluation function generated by the evaluation function generation unit 13 based on combinatorial optimization. In one example of this embodiment, the combinatorial optimization calculation processing unit 14 calculates the fragment pose set by determining the state that minimizes the Hamiltonian H based on combinatorial optimization using a QUBO matrix. The combinatorial optimization calculation processing unit 14 inputs the QUBO matrix generated by the evaluation function generation unit 13 into, for example, a pseudo-quantum annealer to obtain a solution to the QUBO problem (local optimum) that minimizes the Hamiltonian H. The local optimum is expected to represent a solution among the various fragment arrangement candidates expressed as binary variables in which the fragment docking scores with each target protein are good, there are no collisions with each other, and covalent bonds can be appropriately formed between them. Note that the combinatorial optimization calculation processing unit 14 may use other methods that can obtain a solution to the QUBO problem, such as a quantum annealer or simulated annealing, instead of a pseudo-quantum annealer.

[0059] As a pseudo-quantum annealer, Toshiba Digital Solutions' quantum-inspired optimization solution, SQBM+, can be used. SQBM+ is a quantum-inspired optimization solution centered on the Simulated Bifurcation Machine (SBM), a combinatorial optimization solver using the Simulated Bifurcation (SB) algorithm. With SQBM+, by using a unique SB algorithm, a large number of diverse local optima can be obtained in a single calculation. Note that other solutions may also be used as pseudo-quantum annealers.

[0060] The post-processing unit 15 performs post-processing on the fragment pose set calculated by the combinatorial optimization calculation unit 14. As a result of the post-processing, the post-processing unit 15 outputs a small number of effective solutions from among the various local optima.

[0061] For example, the post-processing unit 15 verifies the local optima obtained by the pseudo-quantum annealer. In the Hamiltonian H shown by equation (5) above, the fourth term H4 represents the so-called one-hot constraint, which is necessary to realize the condition that each fragment is always placed in only one location. In other words, the Hamiltonian H includes a constraint that each fragment is selected only once as a candidate for fragment placement. This constraint must always be satisfied, but among the various local optima, there may be unacceptable solutions, such as one fragment being placed in two locations. The post-processing unit 15 removes these unacceptable solutions.

[0062] Furthermore, the post-processing unit 15 may analyze not only the value of the Hamiltonian H but also the contribution of each term constituting the Hamiltonian H to the local optimum obtained by the pseudo-quantum annealer, and present this analysis to the user.

[0063] The local structure optimization unit 16 forms covalent bonds between fragments for each effective solution shown by the post-processing results, constructing a structure that can be interpreted as a single compound molecule. Furthermore, the local structure optimization unit 16 performs local structure optimization on the molecular structure. The local structure optimization unit 16 removes distortions from the molecular structure, for example, based on an energy minimization method. Note that the energy minimization method is a known method and is a standard technique used to remove structural distortions in compound modeling and other applications.

[0064] The output unit 17 outputs the docking calculation results after local structure optimization processing has been performed by the local structure optimization unit 16. As an example, the output unit 17 outputs only the single combined pose with the best evaluation score as the docking calculation result. Alternatively, the output unit 17 may output a predetermined number of combined pose candidates in descending order of evaluation score, taking diversity into consideration, as the docking calculation result.

[0065] The output unit 17 outputs the docking calculation results. For example, the output unit 17 outputs the docking calculation results to an external device (server, display device, etc.) separate from the calculation processing unit 1. The output unit 17 may also store the docking calculation results in the storage unit 2 as a database.

[0066] Through the above procedure, the computational processing device 1 according to this embodiment can obtain the docking calculation result (one or more bond poses and an evaluation score) of a compound by combinatorial optimization using a pseudo-quantum annealer, based on the fragment docking calculation results performed for each fragment of a specific input compound.

[0067] Furthermore, the calculation may be repeated based on the contribution of each term constituting the Hamiltonian H analyzed in the post-processing unit 15, or the final docking calculation result. Whether or not to repeat the calculation may be decided by the user, or it may be decided automatically by the calculation processing unit 1. For example, the calculation can be repeated while slightly changing the weight parameters of each term constituting the Hamiltonian H. If only the weight parameters are changed, the process from the evaluation function generation unit 13 to the output unit 17 of the above process only needs to be repeated.

[0068] Furthermore, if the post-processing unit 15 detects a tendency for frequent collisions between candidate fragment placements, or if the local structure optimization unit 16 detects many candidate solutions with unnatural covalent bond angles, the process can return to the fragment interaction evaluation unit 12 to adjust the thresholds for judgments regarding penalties for collisions or rewards for cooperation. These threshold adjustments can largely be incorporated as automatic functions by acquiring data while repeatedly performing compound docking calculations.

[0069] Figure 7 shows an example of the processing flow for compound docking calculations based on fragments according to this embodiment.

[0070] Step S10: The fragment decomposition unit 10 decomposes the compound into fragments. As an example, the fragment decomposition unit 10 reads data of a compound specified by the user from the compound data 20 stored in the memory unit 2. The calculation processing unit 1 is equipped with an operation unit (not shown in Figure 1) that accepts operations from the user. Based on the read compound data, the fragment decomposition unit 10 decomposes the compound into fragments.

[0071] Step S20: Furthermore, the fragment decomposition unit 10 generates fragment data based on the results of the decomposition process.

[0072] Step S30: The fragment docking calculation unit 11 performs a fragment docking calculation between each fragment shown in the fragment data and the target protein. The fragment docking calculation unit 11 calculates the fragment docking calculation results (candidate fragment placements and their fragment docking scores). Figure 8 shows how the fragment docking calculation results are obtained from the fragment data through docking calculation.

[0073] As described above, the group of candidate fragment placements to be optimized must consist of placements with good binding free energy while maintaining diversity, and it is necessary to efficiently retain such candidates. Therefore, after decomposing the compound into fragments, the binding site of the target protein is divided into numerous 2Å×2Å×2Å subregions, and docking calculations between the target protein and fragments are performed independently for each subregion, outputting up to 20 poses as candidate fragment placements for each fragment in each subregion.

[0074] In the docking calculation performed by the fragment docking calculation unit 11, a known docking calculation algorithm is used, and local optimization is not performed to ensure structural diversity. Only fragment configuration candidates with negative binding free energy scores (considered to be good solution candidates) are output. Furthermore, to ensure that each fragment configuration candidate is non-redundant, the root-mean-square deviation of atomic positions (RMSD) is set to be at least 1.0 Å apart. Note that the RMSD of fragment configuration candidates output from each subregion may be less than 1.0 Å. Therefore, if there are pairs of fragment configuration candidates whose RMSD is again less than 1.0 Å after combining the fragment docking calculation results from each subregion, only fragment configuration candidates with good fragment docking scores are retained.

[0075] Step S40: The fragment interaction evaluation unit 12 performs primary evaluation processing and secondary evaluation processing.

[0076] In the primary evaluation process, the fragment interaction evaluation unit 12 calculates a primary evaluation value for the binding of each fragment placement candidate to the target protein based on the fragment docking score of that candidate. The fragment interaction evaluation unit 12 excludes fragment placement candidates whose primary evaluation value is below a predetermined threshold.

[0077] The fragment interaction evaluation unit 12 performs a secondary evaluation on the fragment placement candidates that were not excluded in the primary evaluation process. In the secondary evaluation process, the fragment interaction evaluation unit 12 calculates a collision evaluation value for collisions between fragment placement candidates and a cooperation evaluation value for cooperation between fragment placement candidates. The collision evaluation value is used as a function clash(i,j) that represents collisions included in the second term H2 of the Hamiltonian H shown by equation (5). The cooperation evaluation value is used as a function conn(i,j) that represents covalent bonds between fragments included in the third term H3 of the Hamiltonian H.

[0078] The functions clash(i,j) and conn(i,j) are expressed in terms of physicochemical parameters, respectively. These physicochemical parameters include, for example, the bond length, bond angle, and dihedral angle of covalent bonds, as well as the interatomic distance between each atom. The molecular mechanical force field calculates the interaction energy from the values ​​of these physicochemical parameters, and it is desirable to use the interaction energy calculated by the molecular mechanical force field. However, there is a limit to the number of fragment arrangement candidates that can be considered in combinatorial optimization, and in particular, a set of fragment arrangement candidates that shows appropriate bond length and bond angle for covalent bonds cannot always be obtained. Therefore, after obtaining the energy score using the molecular mechanical force field, arrangements are accepted if the strain energy or intermolecular collision energy is below a predetermined value.

[0079] The fragment interaction evaluation unit 12 calculates the interaction energy ENB(i, j) when the fragment placement candidate i and the fragment placement candidate j do not form a covalent bond.

[0080] When the fragment placement candidate i and the fragment placement candidate j form a covalent bond, the fragment interaction evaluation unit 12 calculates the interaction energy EB(i, j) when forming a covalent bond.

[0081] The fragment interaction evaluation unit 12 corresponds to the fragment f corresponding to the fragment placement candidate i and the fragment placement candidate j respectively i , and the fragment f j are covalently connected in the original compound, and EB(i, j) ≤ th E If so, conn(i, j) = -1 is set. The fragment interaction evaluation unit 12 uses the fragment f i , and the fragment f j are covalently connected in the original compound, and EB(i, j) > th E If so, conn(i, j) = 0 is set. Here, the energy score threshold th E is a predetermined threshold.

[0082] When conn(i, j) = 0 and ENB(i, j) > th E If so, clash(i, j) = 1 is set. When conn(i, j) ≠ 0 or ENB(i, j) ≤ th E If so, clash(i, j) = 0 is set.

[0083] The energy score was calculated based on the Universal Force Field (UFF) force field using RDKit (version 2023.09.5). Also, in order to allow sufficient structural distortion, the energy score threshold th E was set to 500 kcal / mol.

[0084] Step S50: The evaluation function generation unit 13 generates the Hamiltonian H based on the evaluation values ​​calculated by the fragment interaction evaluation unit 12. As described above, the evaluation function generation unit 13 generates the Hamiltonian H by generating the QUBO matrix. Figure 9 shows how the Hamiltonian H is generated. The first evaluation value is ΔG, which is the fragment docking score included in the first term H1. i These are used as follows: The collision evaluation value is used as the function clash(i,j) which represents collisions included in the second term H2. The cooperative evaluation value is used as the function conn(i,j) which represents covalent bonds between fragments included in the third term H3.

[0085] Step S60: The combinatorial optimization calculation processing unit 14 calculates the fragment pose set on the surface of the target protein of the fragments constituting the compound by determining the state to optimize the evaluation function generated by the evaluation function generation unit 13 based on combinatorial optimization. The combinatorial optimization calculation processing unit 14 inputs the QUBO matrix generated by the evaluation function generation unit 13 into a pseudo-quantum annealer to obtain a solution to the QUBO problem that minimizes the Hamiltonian H. In this embodiment, by using SQBM+ as the pseudo-quantum annealer, a number of local optima of the QUBO problem are obtained. More specifically, using SQBM+ for AWS (version 2.0.1), a calculation was performed for 300 seconds (timeout=300) when the QUBO matrix was input with single-precision floating-point accuracy, and a number of outputs were obtained. Figure 10 shows how a solution to the QUBO problem is obtained by combinatorial optimization using a pseudo-quantum annealer. Figure 10 shows one of the many local optima obtained by SQBM+ from a large group of candidate fragment placements.

[0086] Step S70: The post-processing unit 15 performs post-processing on the fragment pose set calculated by the combinatorial optimization calculation unit 14. Among the local optima output by SQBM+, there may be solutions in which more than one fragment placement candidate is selected for a single fragment. In order to output a structure that can be interpreted as a single compound molecule, only one fragment placement candidate must be selected for each fragment. Therefore, in post-processing, from among the multiple local optima obtained, only the solutions in which all fragments constituting the compound are selected one by one are selected.

[0087] Step S80: The local structure optimization unit 16 forms covalent bonds between fragments for the solution selected by the post-processing unit 15, and outputs each of the obtained solution candidates as a structure that can be interpreted as a single compound molecule. Figure 11 shows how covalent bonds are formed between fragments.

[0088] Furthermore, the local structure optimization unit 16 performs local structure optimization on molecular structures that can be interpreted as a single compound molecule obtained through the formation of covalent bonds.

[0089] Step S90: The output unit 17 outputs the docking calculation results for the compound. The output unit 17 outputs to an external device (server, display device, etc.) separate from the calculation processing device 1.

[0090] Step S100: The calculation processing unit 1 determines whether or not to repeat the calculation. The calculation processing unit 1 determines to repeat the calculation, for example, when it receives an operation from the user indicating that the calculation should be repeated. The user checks the final docking calculation result displayed on the display device, or the contribution analysis result by the post-processing unit 15, and decides whether or not to repeat the calculation. The contribution analysis result by the post-processing unit 15 is the analysis result of the contribution of each term that constitutes the Hamiltonian H. As another example, the calculation processing unit 1 may automatically determine whether or not to repeat the calculation based on the final docking calculation result, or the contribution analysis result by the post-processing unit 15.

[0091] If the calculation processing unit 1 determines that the calculation should be repeated (step S100; YES), it will execute the processes from step S50 to step S90 again. On the other hand, if the calculation processing unit 1 determines that the calculation should not be repeated (step S100; NO), it will terminate the process. This concludes the explanation of the processing flow for compound docking calculations based on fragments according to this embodiment.

[0092] Next, an example of compound docking based on fragments by the computational processing unit 1 according to this embodiment will be described. First, a redocking experiment was performed to evaluate the performance of the computing device 1. A redocking experiment involves extracting only the ligand structure for a protein and ligand whose binding structure is experimentally known, and then re-estimating the binding structure of that protein and ligand using docking calculations to evaluate the performance of the computing device 1. In the performance evaluation, it was considered that the three-dimensional structure was correctly estimated if the distance between the experimentally known binding structure and the binding structure obtained by docking calculations was less than or equal to a predetermined distance (for example, 2.5 Å in RMSD).

[0093] Aldose reductase (ALDR) was used in the redocking experiment. As shown in Figure 12(A), ALDR consists of just over 300 amino acid residues and is known as an enzyme that carries out reactions such as the reduction of glucose to sorbitol using nicotinamide adenine dinucleotide phosphate (NADP). ALDR is an enzyme that is also thought to be involved in neuropathy in diabetes, and it is one of the target proteins for which drug design is underway, with the existence of an approved drug called epalrestat that inhibits this protein.

[0094] First, the cocrystal structure (PDB ID: 2HV5) of ALDR and a known inhibitor was obtained from the Protein Data Bank (PDB). The molecular structure of the inhibitor is shown in Figure 12(B). Based on the compound degradation algorithm described in Non-Patent Literature 3, the inhibitor molecule was decomposed into fragments. As a result of fragmentation, the inhibitor molecule was decomposed into four fragments as shown in Figure 12(C). The ionization state and three-dimensional structure of each fragment were estimated using Schrodinger's LigPrep.

[0095] Next, the binding sites for fragment docking calculations performed on individual fragments were defined as shown in the docking calculation parameters in Figure 13. The box center is the center of the cocrystal inhibitor's three-dimensional structure, and the docking region was determined visually as the size of the cube covering the entire inhibitor's three-dimensional structure.

[0096] Next, Figure 14 shows the results of fragment docking calculations for individual fragments. The number of fragment placement candidates obtained from fragment docking calculations for individual fragments is 3005. Figure 14 shows all 3005 of these fragment placement candidates. By dividing the binding site into numerous sub-regions, the fragment placement candidates are distributed so as to spread across the entire binding site of the target protein.

[0097] The number of fragment placement candidates obtained for each fragment, as well as the maximum and minimum values ​​of their fragment docking scores, are shown in the table in Figure 15. Note that in Figure 15, the structural formula of each fragment is represented by SMILES (Simplified Molecular Input Line Entry System).

[0098] Furthermore, the distribution of fragment docking scores for each fragment is shown in Figures 16 to 19. It is noteworthy that, in addition to poses with good fragment docking scores, poses with less favorable fragment docking scores are also retained as potential fragment placement candidates. By obtaining fragment placement candidates with good scores at the subregion level, it is possible to obtain fragment placement candidates that can function as linkers even in regions with little direct interaction with the protein. Therefore, it is suggested that the obtained group of fragment placement candidates ensures diversity in terms of position.

[0099] Combinatorial optimization calculations using SQBM+ yielded 7298 local optima. First, 5414 solutions (74.2%) were extracted from these local optima, in which all fragments were used exactly once. Figure 20 shows scatter plots of the Hamiltonian value after combinatorial optimization calculations and the RMSD relative to the correct structure of the docking calculation results for the top 1000 solutions with good Hamiltonian H values ​​(combinatorial energy). The distribution of the higher-performing solutions among these 1000 solutions shows a funnel-shaped pattern where the RMSD decreases as the Hamiltonian value improves, indicating that the Hamiltonian H is appropriately designed.

[0100] Next, Figure 21 shows the best fragment pose set among the obtained local optima. According to the fragment pose set shown in Figure 21(A), this best solution shows that each fragment is located in close proximity to the others, resulting in a structure that does not contain any major inconsistencies as a single compound structure. Figure 21(B) shows the result of superimposing the best fragment pose set with the co-crystal structure and the protein three-dimensional structure. The RMSD of this fragment pose set relative to the correct docking structure is 1.26 Å, confirming that the binding pose can be adequately estimated.

[0101] The fragment pose set shown in Figure 21(A) retains structural distortions, making it inappropriate to directly use it as the result of docking calculations. Therefore, the distortions in the molecular structure are removed by local structure optimization. In local structure optimization, an energy minimization method was used to optimize the distorted compound structure composed of the fragment pose set within the protein three-dimensional structure, which is treated as a rigid body. Schrodinger's Maestro software (version 2020-2) was used for local structure optimization.

[0102] The three-dimensional structure of the target protein ALDR (PDB ID: 2HV5) and the complex predicted structure obtained by superimposing the structure that can be interpreted as a compound molecule obtained through combinatorial optimization and post-processing were processed using the Protein Preparation Wizard's Preprocess and H-bond assignment. The protein and coenzyme (NADP) were treated as rigid bodies, and the energy minimization of the compound structure was performed using the OPLS3e force field.

[0103] Figure 22 shows the docking calculation results of the compound after local structure optimization. Figure 22(A) shows the fragment pose set before local structure optimization and the docking calculation results after local structure optimization superimposed. Figure 22(B) shows the docking calculation results after local structure optimization and the cocrystal structure (experimentally known bonding structure) superimposed. According to the results shown in Figure 22(B), the RMSD between the docking calculation results and the cocrystal structure of the obtained compound was 0.27 Å, resulting in a structure that closely matches the cocrystal structure.

[0104] In this embodiment, an example of using Hamiltonian H shown by equation (5) has been described, but it is not limited to this. Hamiltonians other than Hamiltonian H shown by equation (5) may be used. For example, in Hamiltonian H, the second term H2 or the third term H3 shown in equation (5) may be omitted, and the first term H1, the second term H2, or the third term H3 may be modified from equations (1), (2), or (3) described above. In addition, values ​​other than those described above may be set as the weight parameters for each term. However, in order to improve the estimation accuracy of the calculation based on combinatorial optimization of the compound docking calculation based on fragments, it is preferable to use Hamiltonian H shown by equation (5).

[0105] Furthermore, one or more of the post-processing unit 15, local structure optimization unit 16, and output unit 17 may be omitted from the configuration of the computing processing unit 1. In that case, each of the processes performed by the post-processing unit 15, local structure optimization unit 16, and output unit 17 may be performed by an information processing unit separate from the computing processing unit 1, and the computing processing unit 1 may acquire the processing results from the information processing unit, or may have the processing results from the computing processing unit 1 output to the information processing unit.

[0106] As described above, the compound docking calculation processing device (calculation processing device 1 in this embodiment) according to this embodiment comprises a fragment decomposition unit 10, a fragment docking calculation unit 11, a fragment interaction evaluation unit 12, an evaluation function generation unit 13, and a combinatorial optimization calculation processing unit 14. The fragment degradation unit 10 acquires fragment data that shows fragments, which are partial chemical structures of the compound to be used for docking calculations with the target biomolecule (in this embodiment, the target protein). The fragment docking calculation unit 11 performs a fragment docking calculation between each fragment indicated by the fragment data and the target biomolecule (in this embodiment, the target protein) for each fragment. The fragment interaction evaluation unit 12 calculates evaluation values ​​(primary evaluation value, collision evaluation value, and cooperative evaluation value in this embodiment) for each fragment placement candidate and for the relationships between the fragment placement candidates, based on the fragment docking score for each fragment placement candidate, which is a candidate for the placement of the fragment when it binds to a target biomolecule (target protein in this embodiment), and the positional relationship between the fragment placement candidates. The evaluation function generation unit 13 generates an evaluation function (Hamiltonian H in this embodiment) that gives an evaluation value of the fragment pose set on the surface of the target biomolecule (target protein in this embodiment) of the fragment constituting the compound, based on the evaluation value. The combinatorial optimization calculation unit 14 calculates the fragment pose set on the surface of the target biomolecule (in this embodiment, the target protein) of the fragments constituting the compound by determining the state for optimizing the evaluation function based on combinatorial optimization.

[0107] With this configuration, the computational processing device 1 according to this embodiment calculates evaluation values ​​for each fragment placement candidate and the positional relationships between the fragment placement candidates based on the fragment docking score for each candidate and the positional relationships between the candidates. Using these evaluation values, it generates an evaluation function and determines the state for optimizing this evaluation function based on combinatorial optimization. This enables the calculation of the fragment pose set on the surface of the target biomolecule (target protein in this embodiment) of the fragments constituting the compound, thus enabling combinatorial optimization in compound docking calculations based on fragments. As described above, in one example of this embodiment, the QUBO problem and the corresponding evaluation function (Hamiltonian H in this embodiment) are designed, and the combinatorial optimization can be efficiently performed by, for example, inputting the QUBO matrix into a pseudo-quantum annealer.

[0108] (Second embodiment) A second embodiment of the present invention will be described in detail below with reference to the drawings. The compound docking calculation processing device according to this embodiment will be referred to as the calculation processing device 1a. Note that components identical to those in the first embodiment described above will be denoted by the same reference numerals, and descriptions of identical components and operations may be omitted.

[0109] Figure 23 shows an example of the functional configuration of the computing device 1a according to this embodiment. The computing device 1a comprises a fragment decomposition unit 10, a fragment docking calculation unit 11, a fragment interaction evaluation unit 12, an evaluation function generation unit 13, a combinatorial optimization calculation processing unit 14a, a post-processing unit 15, a local structure optimization unit 16, an output unit 17, and a storage unit 2. Comparing the computing device 1a according to this embodiment (Figure 23) with the computing device 1 according to the first embodiment (Figure 1), the differences are that the combinatorial optimization calculation processing unit 14a and the compound dataset 20a are stored in the storage unit 2. Here, the functions of the other components (fragment decomposition unit 10, fragment docking calculation unit 11, fragment interaction evaluation unit 12, evaluation function generation unit 13, post-processing unit 15, local structure optimization unit 16, and output unit 17) are the same as in the first embodiment. However, in the second embodiment, there are some adjustments to the detailed operation of each unit that differ from those in the first embodiment.

[0110] Figure 24 is a diagram showing an example of the processing content of each functional unit of the computing processing device 1a according to this embodiment. In this embodiment, the input fragment is not a fragment obtained by decomposing a single compound, but rather a fragment set, which is a collection of fragments corresponding to multiple compounds, is used as input. Preferably, the fragment set is a collection of diverse fragments corresponding to a broad compound space (a collection of compounds). The fragment set can be obtained, for example, by decomposing multiple compounds. The compound dataset 20a is data representing the group of compounds. The fragment decomposition unit 10 generates the fragment dataset by sequentially decomposing these multiple compounds into fragments. In this case, as the number of compounds increases, it can be expected that many fragments common to multiple compounds will be generated. In this embodiment, it is important that the commonality of these fragments contributes to improving the overall efficiency of the processing.

[0111] The subsequent processing steps are generally the same as in the first embodiment, but there are three main differences from the first embodiment. The first difference lies in the design of the Hamiltonian generated by the evaluation function generation unit 13. In the first embodiment, it was obvious that each fragment obtained by decomposition from the input compound would be used exactly once, so a one-hot constraint was introduced in the fourth term H4. On the other hand, in this embodiment, fragments are often not used, and all but a few fragments selected from the fragment dataset are not used. Therefore, the weighting of the one-hot constraint is weakened, and instead, a term indicating the constraint of how many fragments are allowed in total is introduced into the Hamiltonian. A similar idea is shown in Non-Patent Literature 5 in the field of polyomino puzzle solving, and a specific Hamiltonian design method is also shown, so a similar method can be used.

[0112] The second difference is that in this embodiment, there may be a greater number of solutions compared to the first embodiment. The combinatorial optimization calculation processing unit 14a calculates a fragment pose set from fragment data obtained from multiple compounds by combinatorial optimization based on an evaluation function. At this time, there may be multiple compounds among the input compounds that can be considered to be the correct docking calculation result, and furthermore, there may be a completely new compound formed by combining fragments obtained from multiple compounds that can be considered to be the correct docking calculation result. For this reason, if an information processing device that can only obtain one solution is used in the combinatorial optimization calculation processing unit 14a, there is a possibility that a variety of solutions will be overlooked.

[0113] Therefore, in order to cover a variety of solutions, it is preferable to use a pseudo-quantum annealer that has the function of obtaining a large number of local optima. As a pseudo-quantum annealer that is excellent in terms of its ability to cover diversity, for example, Toshiba Digital Solutions' SQBM+ described in the first embodiment is suitable. Furthermore, as described in the first embodiment, the diversity of solutions can be expanded in this embodiment as well by changing the weights of each term of the Hamiltonian, or by changing the collision and cooperation criteria, respectively.

[0114] Furthermore, in this embodiment, a bias is introduced in the primary evaluation process by the fragment interaction evaluation unit 12 to set a higher primary evaluation value for a specific fragment, that is, to preferentially adopt a specific fragment. For example, a positive primary bias is applied to a large number of fragments based on the user's expectations, such as pharmaceutical knowledge. In another example, by performing combinatorial optimization calculations while mechanically changing the target fragment to which the bias is applied, compounds containing the target fragment are more likely to be found preferentially, and a more diverse solution is obtained overall. By repeatedly performing such information processing based on user input or automatically, or by performing it in parallel using many pseudo-quantum annealers, evaluation of a variety of compounds can be realized.

[0115] The third difference is that, in this embodiment, instead of obtaining a fragment dataset by decomposing multiple compounds with the fragment decomposition unit 10, it is possible to realize the function of the fragment decomposition unit 10 by preparing a fragment dataset that is convenient for the user in advance and reading the said fragment dataset from a file or the like. As the fragment dataset to be prepared in advance, for example, a fragment set that is frequently used within a pharmaceutical company may be selected, or a fragment set that has been designed and distributed for a wider range of users may be selected.

[0116] In the first embodiment, it is conceivable to pre-create a database of the fragment decomposition of compounds, but the only effect of this is to reduce the computation time for the relatively light calculation of fragment decomposition for a single target compound. In contrast, this embodiment differs significantly in that it attempts to obtain docking calculation results for multiple compounds at once by inputting a fragment set containing many representative fragments. In this case, the selection of the fragment set affects the result.

[0117] Figure 25 shows an example of the processing flow for compound docking calculation based on fragments according to this embodiment. Steps S130, S140, S150, S170, S180, and S110 are the same as steps S30, S40, S50, S70, S80, and S100 in Figure 7, so their explanation is omitted.

[0118] Step S110: The fragment decomposition unit 10 decomposes multiple compounds into fragments. For example, the fragment decomposition unit 10 reads data for multiple compounds specified by the user from the compound dataset 20a stored in the memory unit 2. Based on the read compound data, the fragment decomposition unit 10 sequentially decomposes the multiple compounds into fragments.

[0119] Step S120: Furthermore, the fragment decomposition unit 10 generates a fragment dataset based on the results of the decomposition process. The fragment dataset is a set of fragment data showing each fragment, which is a substructure of multiple compounds.

[0120] Furthermore, in steps S110 and S120, if a file of fragment sets is stored in advance instead of the compound dataset 20a in the memory unit 2, it is possible to omit the decomposition process and directly use the read fragment set. The compound dataset 20a can be computationally efficient if it is pre-collected a group of compounds that are expected to have high binding affinity to the target protein (excellent docking score). Even when directly loading a fragment set instead of compounds, it is efficient if it is pre-collected a group of fragments that are expected to be likely to be adopted. It is desirable to efficiently select these datasets by utilizing general knowledge in drug discovery science and prior information on the target protein.

[0121] Step S160: The combinatorial optimization calculation unit 14a calculates the fragment pose set on the surface of the target protein of the fragment dataset based on the evaluation function generated by the evaluation function generation unit 13. As described above, the fragment pose set may include the fragment docking calculation results between multiple fragments, which are substructures of multiple different compounds, and the target protein. Therefore, the combinatorial optimization calculation unit 14a calculates the fragment pose set on the surface of the target protein of multiple fragments, which are substructures of multiple different compounds, based on the evaluation function generated by the evaluation function generation unit 13.

[0122] Step S190: The output unit 17 outputs the docking calculation results for the compound group. The output unit 17 outputs to an external device (server, display device, etc.) separate from the calculation processing device 1a. This concludes the explanation of the processing flow for compound docking calculations based on fragments according to this embodiment.

[0123] In the embodiments described above, an example was given in which the docking calculation result between a compound and a target protein is obtained by determining the state for optimizing the Hamiltonian based on combinatorial optimization based on the QUBO matrix, but this is not limited to this example. The Hamiltonian may also be optimized based on optimization algorithms other than combinatorial optimization based on the QUBO matrix.

[0124] For example, the fragment interaction evaluation unit 12 may perform one or more of the following: tertiary evaluation and quaternary evaluation, in addition to the primary and secondary evaluation processes described above. A tertiary evaluation is a process that evaluates collisions and cooperation between three candidate fragment placements. A quaternary evaluation is a process that evaluates collisions and cooperation between four candidate fragment placements. Increasing the order of the evaluation process leads to an increase in the number of triplets or quadruplets of candidate fragment placements to be evaluated, and consequently, an increase in computational complexity. However, by considering particularly important features for solving the problem, it may be possible to improve the quality of the solution. When performing evaluation processes of order 3 or higher, the Polynomial Unconstrained Binary Optimization (PUBO) method is used instead of Quadratic Optimization of Binary Variables (QUBO), but the essential configuration remains the same as in the embodiments described above.

[0125] Furthermore, some parts of the calculation processing units 1 and 1a in the above-described embodiments, such as the fragment decomposition unit 10, the fragment docking calculation unit 11, the fragment interaction evaluation unit 12, the evaluation function generation unit 13, the combinatorial optimization calculation processing units 14 and 14a, the post-processing unit 15, the local structure optimization unit 16, and the output unit 17, may be implemented using a computer. In that case, the program for realizing this control function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be loaded into a computer system and executed. Here, "computer system" refers to a computer system built into the calculation processing units 1 and 1a, and includes hardware such as an operating system (OS) and peripheral devices. Furthermore, "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs (Read Only Memory), CD-ROMs (Compact Disc-Read Only Memory), and storage devices such as hard disks built into a computer system. Furthermore, "computer-readable recording media" may include those that dynamically hold programs for a short period of time, such as communication lines used when transmitting programs via networks such as the Internet or communication lines such as telephone lines, as well as those that hold programs for a certain period of time, such as volatile memory within a computer system that acts as a server or client in such cases. In addition, the above-mentioned program may be for the purpose of realizing some of the functions described above, and may also be a program that can realize the above-mentioned functions in combination with a program already recorded in the computer system. Furthermore, some or all of the computing devices 1 and 1a in the above-described embodiments may be implemented as integrated circuits such as LSIs (Large Scale Integrations). Each functional block of the computing devices 1 and 1a may be individually implemented as a processor, or some or all of them may be integrated into a single processor. In addition, the method of implementing integrated circuits is not limited to LSIs; dedicated circuits or general-purpose processors may also be used. Furthermore, if advances in semiconductor technology lead to the emergence of integrated circuit implementation technologies that can replace LSIs, integrated circuits using such technologies may be used.

[0126] Although one embodiment of this invention has been described in detail above with reference to the drawings, the specific configuration is not limited to that described above, and various design changes can be made without departing from the spirit of this invention. [Explanation of symbols]

[0127] 1, 1a...Computational processing unit (compound docking calculation processing unit), 10...Fragment decomposition unit, 11...Fragment docking calculation unit, 12...Fragment interaction evaluation unit, 13...Evaluation function generation unit, 14, 14a...Combinatorial optimization calculation processing unit

Claims

1. A fragment degradation unit that acquires fragment data showing fragments, which are partial chemical structures of compounds used in docking calculations with target biomolecules, A fragment docking calculation unit performs a fragment docking calculation between each fragment shown in the aforementioned fragment data and the target biomolecule. A fragment interaction evaluation unit calculates an evaluation value between the fragment configuration candidates and the fragment configuration candidates based on the results of fragment docking calculations for each fragment configuration candidate, which are candidate configurations when the fragment binds to the target biomolecule, and the positional relationship between the fragment configuration candidates. An evaluation function generation unit generates an evaluation function based on the evaluation value that gives an evaluation value of the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule, A combinatorial optimization calculation processing unit that calculates the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule by determining the state for optimizing the evaluation function based on combinatorial optimization, A compound docking calculation processing device equipped with the following features.

2. The positional relationships between the aforementioned fragment placement candidates include collisions between the aforementioned fragment placement candidates. The fragment interaction evaluation unit calculates the evaluation value for collisions between the candidate fragment placements. The compound docking calculation apparatus according to claim 1.

3. The positional relationship between the aforementioned fragment placement candidates includes coordination between the aforementioned fragment placement candidates. The fragment interaction evaluation unit calculates the evaluation value for the coordination between the candidate fragment placements. The compound docking calculation apparatus according to claim 1.

4. The fragment interaction evaluation unit calculates the evaluation value for the binding of each fragment arrangement candidate to the target biomolecule based on the fragment docking score for each fragment arrangement candidate. The compound docking calculation apparatus according to claim 1.

5. The fragment interaction evaluation unit calculates the evaluation value for the binding of the fragment placement candidate to the target biomolecule on its own, based on the size of the fragment corresponding to the fragment placement candidate and the fragment docking score for the fragment placement candidate. The compound docking calculation apparatus according to claim 4.

6. The aforementioned evaluation function is a Hamiltonian. The compound docking calculation apparatus according to claim 1.

7. The combinatorial optimization calculation unit calculates the fragment pose set by finding the state that minimizes the Hamiltonian based on unconstrained binary variable optimization. The compound docking calculation apparatus according to claim 6.

8. The Hamiltonian includes a constraint that each fragment is selected exactly once as a candidate for fragment placement. The compound docking calculation apparatus according to claim 6.

9. The fragment decomposition unit acquires fragment data that shows a collection of fragments corresponding to multiple compounds. The compound docking calculation apparatus according to claim 1.

10. The fragment docking calculation unit outputs a fragment docking calculation result between each fragment and the target biomolecule, for each fragment indicated by the fragment data representing the set of fragments corresponding to the plurality of compounds. The compound docking calculation apparatus according to claim 9.

11. The combinatorial optimization calculation unit calculates the fragment pose set on the surface of the target biomolecule for a plurality of fragments constituting the compound that are all different from each other, based on the evaluation function. The compound docking calculation processing apparatus according to claim 10.

12. A post-processing unit that performs post-processing on the aforementioned fragment pose set, A local structure optimization unit performs local structure optimization on the molecular structure shown by the results of the post-processing described above, An output unit that outputs the docking calculation results between the compound and the target biomolecule based on the results of the local structure optimization, The compound docking calculation apparatus according to claim 1, further comprising

13. A fragment degradation step to obtain fragment data showing fragments, which are partial chemical structures of a compound used for docking calculations with a target biomolecule, A fragment docking calculation step is performed for each fragment indicated by the aforementioned fragment data, which involves performing a fragment docking calculation between the fragment alone and the target biomolecule. A fragment interaction evaluation step that calculates an evaluation value between the fragment placement candidates and the fragment placement candidates based on the results of a fragment docking calculation for each fragment placement candidate, which is a candidate arrangement when the fragment binds to the target biomolecule, and the positional relationship between the fragment placement candidates. An evaluation function generation step of generating an evaluation function that gives an evaluation value of the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule, based on the evaluation value, A combinatorial optimization calculation step to calculate the fragment pose set of the fragments constituting the compound on the surface of the target biomolecule by determining the state for optimizing the evaluation function based on combinatorial optimization, A compound docking calculation method having the following characteristics.

14. On the computer, A fragment degradation step to obtain fragment data showing fragments, which are partial chemical structures of a compound used for docking calculations with a target biomolecule, A fragment docking calculation step is performed for each fragment indicated by the aforementioned fragment data, which involves performing a fragment docking calculation between the fragment alone and the target biomolecule. A fragment interaction evaluation step that calculates an evaluation value between the fragment placement candidates and the fragment placement candidates based on the results of a fragment docking calculation for each fragment placement candidate, which is a candidate arrangement when the fragment binds to the target biomolecule, and the positional relationship between the fragment placement candidates. An evaluation function generation step of generating an evaluation function that gives an evaluation value of the fragment pose set of the fragment constituting the compound on the surface of the target biomolecule, based on the evaluation value, A combinatorial optimization calculation step to calculate the fragment pose set of the fragments constituting the compound on the surface of the target biomolecule by determining the state for optimizing the evaluation function based on combinatorial optimization, A program to execute.

Images