JP7218274B2 - Compound Property Prediction Apparatus, Compound Property Prediction Program, and Compound Property Prediction Method for Predicting Properties of Compound - Google Patents

Description

The present invention relates to a compound property prediction device, a compound property prediction program, and a compound property prediction method for predicting properties of a compound.

In drug discovery research, a new drug candidate compound (hereinafter referred to as a lead compound) that has been discovered as a new drug candidate is used as a standard, and by repeating design, synthesis, and evaluation so as to change the structure of the lead compound, the compound as a new drug is developed. An operation (hereinafter referred to as an optimization program) is performed to gradually optimize the structure. At this time, while maintaining the main activity (drug effect) as a drug in the lead compound, it is possible to improve the compound's absorption, distribution, metabolism, excretion, and toxicity in humans and animals. Compounds are searched to improve the ADMET attribute that indicates the property.

In the optimization program from the lead compound in drug discovery research, the compound to be predicted is not randomly selected from a huge set of compounds, but the properties such as ADMET attributes proposed in the optimization program are known. It is often preferable to select a compound that is structurally similar to the compound of . Therefore, in order to shorten the period of the optimization program and reduce the cost, there is a demand for a technique for predicting the ADMET attributes of compounds that have not yet been examined in the optimization program.

One aspect of the present invention is a compound property prediction device for predicting properties of a compound, comprising a compound database in which the property measured for each of a plurality of compounds is associated with each of the compounds. A teacher that is accessible to a stored storage means and that includes at least a combination of two compounds selected from the compound database as selected compounds, the common structure and differential structure of the selected compounds, and the properties of the selected compounds. a property learning means for constructing a compound property prediction model that is subjected to machine learning for predicting the property of a compound to be predicted using training data, and a compound selected from the compound to be predicted and the compound database; a property prediction means for obtaining a prediction result of the property of the compound to be predicted as an output of the compound property prediction model by inputting the common structure and the difference structure of the compound to the compound property prediction model. It is a compound property prediction device characterized by

Here, it is preferable that the property learning means uses a graph neural network (GNN) to use the common structure as a common graph structure and the difference structure as a difference graph structure as the supervised training data.

It is also preferred to determine the consensus structure of said selected compounds by maximum common substructure analysis (MCS).

It is also preferred that the property is at least one ADMET attribute for the compound.

In addition, it is preferable that the compound database includes compounds obtained in an optimization program for lead compounds in drug discovery research and properties actually measured for the compounds.

Further, the data contained in the compound database is arranged in chronological order for each optimization program and divided, the data of the front part is used as the supervised training data, and the data of the back part is used as verification data or evaluation data. It is preferable to perform the machine learning.

In addition, the compound database includes compounds obtained in a plurality of the optimization programs for the lead compound and properties actually measured for the compounds, and the verification data is sequentially selected for each optimization program, and the Iterative machine learning is preferred.

Another aspect of the present invention is a compound property prediction program for predicting the properties of a compound, comprising a compound database in which the properties actually measured for each of a plurality of compounds are associated with each of the compounds. A computer having access to a stored storage means is configured to, using two compounds selected from the compound database as selected compounds, at least combine common structures and differential structures of the selected compounds and the properties of the selected compounds. property learning means for constructing a compound property prediction model that is subjected to machine learning for predicting the property of a compound that is a prediction target using supervised training data including the compound, and from the compound that is the prediction target and the compound database Functioning as a property prediction means for obtaining a prediction result of the property of the compound to be predicted as an output of the compound property prediction model by inputting the common structure and differential structure of the selected compound into the compound property prediction model It is a compound property prediction program characterized by allowing

Another aspect of the present invention is a compound property prediction method for predicting the properties of a compound, comprising: a compound database in which the properties actually measured for each of a plurality of compounds are associated with each of the compounds; Using a computer having access to the stored storage means, two compounds selected from the compound database are used as selected compounds, and the common structure and differential structure of the selected compounds are combined with the properties of the selected compounds. using as supervised training data containing at least a property prediction step of obtaining prediction results of the properties of the compound to be predicted as an output of the compound property prediction model by inputting common structures and differential structures of compounds selected from a database into the compound property prediction model; A compound property prediction method characterized by comprising:

An object of the present invention is to provide a compound property prediction device, a compound property prediction program, and a compound property prediction method that enable prediction of the properties of a target compound in an optimization program for a lead compound. Let it be one. Other objects of embodiments of the present invention will become apparent by reference to the specification as a whole.

It is a figure which shows the structure of the compound property prediction apparatus in embodiment of this invention. 4 is a flow chart showing compound property prediction processing in the embodiment of the present invention. It is a figure which shows the example of the structure of the compound in embodiment of this invention. It is a figure which shows the example of the compound database in embodiment of this invention. It is a figure which shows the optimization program in embodiment of this invention. It is a figure explaining the data division process in embodiment of this invention. FIG. 4 is a diagram for explaining the process of obtaining common structures and differential structures of compounds according to the embodiment of the present invention; It is a figure for demonstrating the machine learning in embodiment of this invention. FIG. 4 is a diagram for explaining prediction processing of properties of a compound in the embodiment of the present invention;

A compound property prediction apparatus 100 according to an embodiment of the present invention includes a processing unit 10, a storage unit 12, an input unit 14, an output unit 16 and a communication unit 18, as shown in FIG.

The compound property prediction device 100 can be configured by a general computer. The processing unit 10 includes a CPU and the like, and integrally performs processing in the compound property prediction device 100 . The processing unit 10 executes the compound property prediction program stored in the storage unit 12 to perform compound property prediction processing in the present embodiment. The storage unit 12 stores information necessary for the compound property prediction process, such as a compound property prediction model (compound property predictor) used in the compound property prediction process, and a compound database that associates compounds obtained in drug discovery research with their properties. Remember. The storage unit 12 can be composed of, for example, a semiconductor memory, a hard disk, or the like. The storage unit 12 may be provided inside the compound property prediction apparatus 100, or may be provided outside so as to be accessible from the processing unit 10 using an information network such as wireless or wired. Input unit 14 includes means for inputting information to compound property prediction apparatus 100 . The output unit 16 includes means for displaying information processed by the compound property prediction device 100 . The communication unit 18 includes an interface for exchanging information with an external device (server or the like). The communication unit 18 enables communication with external devices by being connected to an information communication network such as the Internet.

[Compound Property Prediction Processing]
The compound property prediction process according to the present embodiment will be described below with reference to the flowchart of FIG. The compound property prediction apparatus 100 executes a compound property prediction program to perform machine learning for predicting compound properties using known learning data including compounds and their properties, thereby generating a compound property prediction model. Then, the compound property prediction model is used to predict the property of the compound to be predicted.

In the present embodiment, a compound database is created by associating compounds and ADMET attributes that have been evaluated in an optimization program that repeats design, synthesis, and evaluation so as to change the structure of a lead compound found as a candidate for a new drug in drug discovery research. is stored in the storage unit 12 as. Machine learning uses the compounds and ADMET attributes contained in the compound database as supervised training data.

However, compounds used for learning and compounds to be predicted are not limited to compounds related to drug discovery. Further, the property of the compound is not limited to the ADMET attribute, and may be any information related to the compound. Also, all items of the ADMET attribute may be used, or only some of the items may be used.

FIG. 3 shows examples of compounds. FIG. 3(a) shows the structural formula of the lead compound that serves as the starting point for the optimization program. FIGS. 3(b) to 3(c) show structural formulas of analogous compound A and analogous compound B in which a part of the structure of the lead compound is changed to another structure. The structure of the lead compound and the structure of the analogous compound are examples, and the present invention is not limited to these. The analogous compound may have, for example, a structure obtained by removing a part of the structure from the lead compound, a structure obtained by replacing a part of the structure of the lead compound with another structure, or a structure obtained by replacing the structure of the lead compound with another structure. may be added.

FIG. 4 shows an example of a compound database. The compound database includes an optimization program ID (PID) uniquely assigned to each series of optimization programs performed in drug discovery research, an optimization program name, a compound ID uniquely assigned to each compound, a compound name, It is a database in which the structure of a compound, the property of the compound evaluated in the optimization program, and the date and time of evaluation are stored in association with each other. As the structure of the compound, as illustrated in FIG. 3, the atoms constituting the compound and their bonding states are stored. The structure of a compound may be registered in a compound database using, for example, the SMILES notation.

Note that FIG. 4 shows an example in which the compound ID, compound name, compound structure and properties are associated with only three compounds evaluated in the optimization program a. The properties of a large number of compounds derived from compounds are evaluated and stored as a compound database. Further, in the present embodiment, only the ADMET attribute CYP3A4 inhibition rate and JP1 solubility value were exemplified as the properties of the compound, but the present invention is not limited to these, and the values of other items of the ADMET attribute and the compound Other properties may be used.

Also, as shown in FIG. 5, in drug discovery research, multiple optimization programs are run from a single lead compound. Therefore, when multiple optimization programs are executed, numerical values indicating chronological order relationships such as compound ID, compound name, compound structure/property, and evaluation date and time are associated with each optimization program as a compound database. remembered.

In step S10, division processing of the compound database is performed. Through the processing of this step, the compound property prediction device 100 functions as data dividing means. In the machine learning of the compound property prediction apparatus 100, the data stored in the compound database is divided into supervised training data, verification data and evaluation data for use.

The processing unit 10 reads out the compound database from the storage unit 12 and performs the following processing. In the present embodiment, as shown in FIG. 6, the data stored in the compound database is sorted chronologically along the evaluation date and time for each optimization program and divided into training data, verification data, and evaluation data. do.

Here, the training data is data for constructing a compound property prediction model of the compound property prediction device 100 by machine learning. Validation data is data used to determine hyperparameters in machine learning and to select a model. Evaluation data is data used to evaluate whether a compound property prediction model constructed by machine learning is appropriate.

The processing unit 10 arranges the data stored in the compound database in chronological order for each optimization program and divides the data into two. Then, among the multiple optimization programs (programs a to d), the data corresponding to the latter group out of the groups divided chronologically in the optimization program (program a) targeted for prediction is evaluated. data. In addition, the data corresponding to the latter group among the groups before and after the groups divided chronologically in the optimization programs (programs b to d) other than the optimization program (program a) from which the evaluation data was extracted is the verification data. and

Also, in order to perform machine learning more appropriately, cross-validation may be applied to change verification data and repeat machine learning. For example, as shown in learning processes 1 to 3 in FIG. 6, optimization programs for extracting verification data from optimization programs (programs b to d) other than the optimization program (program a) from which the evaluation data was extracted are sequentially selected. It may be modified so that machine learning is performed repeatedly.

In the present embodiment (FIG. 6), the data included in the compound database is divided into the first half 30% and the second half 70% for each optimization program, but it is not limited to this, and the data is divided into other ratios. You may In other words, it may be divided into ratios such that the compound property prediction model is machine-learned so that the property of the compound to be predicted is appropriately output.

In step S12, compound pair selection processing is performed. Through the processing of this step, the compound property prediction device 100 functions as compound selection means. The processing unit 10 selects two compounds as selected compounds from the training data (train data) associated with the same optimization program from the compound database stored in the storage unit 12 . Data associated with the selected compounds are used as supervised training data for machine learning compound property prediction models.

Here, when selecting two compound pairs as selected compounds from the training data associated with the same optimization program, simply learning the combination of all compound pairs as a data set yields a compound property prediction model. Overfitting may occur. One of the paired compounds is then uniformly sampled from the training data associated with the same optimization program. As a result, one compound of the pair of compounds to be the selected compound is sampled without bias from the training data. Then, compounds that can be paired with the sampled compounds are uniformly sampled from the remaining data included in the training data, and two compounds are combined and selected as a selection compound. By performing such processing, at least one of the two compounds forming a pair can be selected from the training data without bias. In implementation, a set of data of compounds forming a pair may be set for each compound contained in the compound database, and compounds in the pair may be selected in order.

Similarly, the processing unit 10 selects two compounds from the verification data from the compound database stored in the storage unit 12 as selected compounds. Similarly, the processing unit 10 selects two compounds from the evaluation data from the compound database stored in the storage unit 12 as selected compounds.

In step S14, a process of extracting common structures and differential structures of compounds is performed. Through the processing of this step, the compound property prediction device 100 functions as structural analysis means. The processing unit 10 extracts and vectorizes a common structure and an uncommon chemical structure for each pair of two compounds selected as the selected compounds in step S12 as a common structure and a difference structure, respectively. For example, rdFMCS. By using FindMCS( ), it is possible to extract the common structure of the two selected compounds. Further, structures other than the common structure are extracted as differential structures for each of the two compounds. A common structure and a difference structure can be represented by, for example, the SMILES notation.

For example, as shown in FIG. 7, a common structure and a difference structure are extracted for each pair of selected compounds (selected compound 1 and selected compound 2). Here, a structure that exists in the selected compound 1 but does not exist in the selected compound 2 is defined as a differential structure 1, and a structure that exists in the selected compound 2 but does not exist in the selected compound 1 is defined as a differential structure 2. are extracting.

In step S16, machine learning is performed to construct a compound property prediction model. Through the processing of this step, the compound property prediction device 100 functions as property learning means. As shown in FIG. 8, the processing unit 10 converts the common structure and difference structure (difference structure 1 and difference structure 2) of the selected compound of the training data extracted in step S14 into a common graph structure and a difference graph structure, respectively. By combining the properties of the two compounds that make up the selected compound stored as a compound database for the input containing the common structure and the difference structure of the two compounds that are the selected compounds, the input of the compound A compound property prediction model is machine-learned so that properties are output.

It is preferable to apply a graph neural network (GNN) to the compound property prediction model. GNN is a neural network that handles graph structures, and many models have been proposed. In order to build a compound property prediction model for configuring the compound property prediction device 100, it is preferable to apply GIN (Graph Isomorphism Network) [Xu+, ICLR2019], although it is not particularly limited to this. be. In the machine learning model, it is preferable to appropriately select the number of layers of the neural network, the activation function, the loss function, and the like.

Specifically, for example, the following processing may be performed. The common structure of the selected compounds extracted in step S14 is used as an input for the GNN, and graph convolution is performed using the common structure as a subgraph, and readout is performed for the entire graph structure of the common structure. By doing so, we obtain the feature vector of the entire graph of the common structure. Here, the readout is a process of calculating the sum (Sum) of the vectors assigned to all nodes (atoms) in the graph structure or finding the maximum value (Max). . Further, the differential structure (differential structure 1 and differential structure 2) of the selected compound extracted in step S14 is used as an input for the GNN, and the differential structure is used as a subgraph to perform graph convolution processing (Convolution). A feature vector of the differential structure is obtained by reading out the structure. At this time, instead of the differential structure, graph convolution processing (Convolution) is performed on the structures of the two compounds themselves of the selected compounds, and then readout is performed limited to the differential structure. You may read out (Readout) with respect to a structure (all the nodes of the graph structure of a compound).

In addition, not only the common structure and differential structure of the selected compounds of the training data, but also the types of atoms constituting two compounds of the selected compounds and the bonding state between atoms may be input as supervised training data for machine learning. . Also, the properties of two compounds constituting the selected compound may be used as teacher data, or the difference between the properties of two compounds may be used as teacher data.

In this way, machine learning is performed for a compound property prediction model that outputs the properties of the selected compound with input of training data that includes at least the common structure and differential structure of the selected compound included in the training data. Furthermore, using the verification data selected in step S12, the hyperparameters in the compound property prediction model obtained using the common structure and differential structure of the selected compound included in the verification data and the property data of the selected compound are determined. to select the optimal compound property prediction model. Further, using the evaluation data selected in step S12, the property of the selected compound obtained by actual evaluation of the common structure and differential structure of the selected compound included in the evaluation data is is output.

Also, when applying cross-validation, machine learning is repeated by changing the validation data. For example, as shown in learning processes 1 to 3 in FIG. 6, the optimization program for extracting verification data is sequentially changed to repeat machine learning.

In step S18, a process of predicting properties of the compound to be predicted is performed. Through the processing of this step, the compound property prediction device 100 functions as property prediction means. First, an input of structural data of a compound whose properties are to be predicted in any optimization program is accepted. The structure of the compound to be predicted may be received using the input unit 14 or may be stored in the storage unit 12 in advance. Next, the processing unit 10 selects one compound belonging to the same optimization program as the compound to be predicted in the compound database, and calculates the common structure and difference between the structure of the compound and the structure of the compound to be predicted. Extract and vectorize the structure. For example, rdFMCS. A common structure of two compounds can be extracted by using FindMCS(). Further, structures other than the common structure are extracted as differential structures for each of the two compounds. Then, by inputting the vectorized common structure and differential structure of the compound into the compound property prediction model obtained in step S16, the property prediction result output of the compound is obtained.

In step S16, as training data, not only the common structure and differential structure of the selected compound, but also the types of atoms constituting two compounds of the selected compound and the bonding state between atoms are input as supervised training data, and machine learning is performed. In such a case, the types of atoms constituting the compound and the state of bonding between atoms may be input into the compound property prediction model.

As described above, the compound property prediction apparatus 100 according to the present embodiment performs machine learning using training data including common structures and differential structures of compounds, thereby more appropriately predicting properties of compounds whose properties are unknown. It is possible to construct a compound property prediction model that can be used. In addition, by arranging the data used for machine learning in chronological order and dividing it, the data corresponding to the latter group can be used as evaluation data or verification data, so that the properties of compounds whose properties are unknown can be more appropriately identified. It is possible to construct a compound property prediction model that can predict

In the compound property prediction device 100 of the present embodiment, the data division means, compound selection means, structure analysis means, property learning means, and property prediction means are realized in one device. It may be implemented by a different device or a different execution entity. For example, some of these means may be implemented by the server computer, and the remaining means may be implemented by the client computer.

10 processing unit, 12 storage unit, 14 input unit, 16 output unit, 18 communication unit, 20 processing unit, 22 storage unit, 24 input unit, 26 output unit, 28 communication unit, 100 compound property prediction device.

Claims (7)

A compound property prediction device for predicting properties of a compound,
For a plurality of compounds, a storage means storing a compound database in which compounds obtained in a lead compound optimization program in drug discovery research and properties actually measured for the compounds are associated with each of the compounds,
Using two compounds selected from the compound database as selected compounds as supervised training data including at least a combination of the common structure and the differential structure of the selected compounds and the properties of the selected compounds, a prediction target In order to predict the property of a certain compound, the data contained in the compound database is arranged in time series and divided for each of the optimization programs, the data of the front part is used as the supervised training data, and the data of the back part A property learning means for constructing a compound property prediction model that is machine-learned using as verification data or evaluation data ,
By inputting the common structure and differential structure of the compound to be predicted and the compound selected from the compound database to the compound property prediction model, the property of the compound to be predicted as the output of the compound property prediction model A property prediction means for obtaining a prediction result of
A compound property prediction device comprising: The compound property prediction device according to claim 1,
The compound property prediction model uses a graph neural network (GNN) as the common structure as a common graph structure and the difference structure as a difference graph structure as the supervised training data. . The compound property prediction device according to claim 1 or 2,
A compound property prediction device, wherein the common structure of the selected compound is determined by maximum common substructure analysis (MCS). The compound property prediction device according to any one of claims 1 to 3,
A compound property prediction device, wherein the property is at least one ADMET attribute for the compound. The compound property prediction device according to claim 1 ,
The compound database includes compounds obtained in a plurality of the optimization programs for the lead compound and properties actually measured for the compounds, and the verification data is sequentially selected for each optimization program to perform the machine learning. A compound property prediction device characterized by repeatedly performing. A compound property prediction program for predicting properties of a compound,
For a plurality of compounds, a computer capable of accessing a storage means storing a compound database in which compounds obtained in a lead compound optimization program in drug discovery research and properties actually measured for the compounds are associated with each of the compounds. ,
Using two compounds selected from the compound database as selected compounds as supervised training data including at least a combination of the common structure and the differential structure of the selected compounds and the properties of the selected compounds, a prediction target In order to predict the property of a certain compound, the data contained in the compound database is arranged in time series and divided for each of the optimization programs, the data of the front part is used as the supervised training data, and the data of the back part as verification data or evaluation data, and a property learning means for constructing a compound property prediction model that is machine-learned;
By inputting the common structure and differential structure of the compound to be predicted and the compound selected from the compound database to the compound property prediction model, the property of the compound to be predicted as the output of the compound property prediction model A compound property prediction program characterized by functioning as a property prediction means for obtaining a prediction result of. A compound property prediction method for predicting the properties of a compound,
For a plurality of compounds, a computer capable of accessing a storage means storing a compound database in which compounds obtained in a lead compound optimization program in drug discovery research and properties actually measured for the compounds are associated with each of the compounds. make use of,
Using two compounds selected from the compound database as selected compounds as supervised training data including at least a combination of the common structure and the differential structure of the selected compounds and the properties of the selected compounds, a prediction target In order to predict the property of a certain compound, the data contained in the compound database is arranged in time series and divided for each of the optimization programs, the data of the front part is used as the supervised training data, and the data of the back part A property learning step of constructing a compound property prediction model subjected to machine learning using as verification data or evaluation data ;
By inputting the common structure and differential structure of the compound to be predicted and the compound selected from the compound database to the compound property prediction model, the property of the compound to be predicted as the output of the compound property prediction model A property prediction step of obtaining a prediction result of
A compound property prediction method comprising:

Images