JP2019095957A - Attribute value prediction device and attribute value prediction program - Google Patents

Abstract

To accurately predict a physical property value of a chemical compound.SOLUTION: An attribute value prediction device 10 comprises a graph generation unit 12 for generating a graph in which a chemical compound is shown by a graph structure, a feature quantity extraction unit 14 for extracting feature quantity of the chemical compound by subjecting a node signal applied to a node of the graph to graph wavelet transform, and an attribute value prediction unit 16 for predicting an attribute value of the chemical compound.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an attribute value prediction device and an attribute value prediction program.

  Non-Patent Document 1 discloses a method of predicting the effect of mutation of a protein using a feature amount called mCSM signature.

  The mCSM signature represents the frequency of atoms present in the radius r, centering on the amino acid at the mutation site, and by setting a plurality of different r, the atoms present in the periphery of the mutation site are graded Can be converted to feature quantities.

PiCs, D. E., Ascher, D. B., & Blundell, T. L. (2013). MCSM: predicting the effects of mutations in proteins using graph-based signatures. Bioinfomatics, 30 (3), 335-342.

  In the above-mentioned prior art, the radius r is used as the threshold of distance to convert the environment of a part (around the mutation) of the compound (protein) into the feature value. In the above-mentioned prior art, in order to discretize the signal depending on whether or not it is included within the radius r, for example, when an atom is present in the vicinity of the radius r, the feature quantity becomes unstable due to the fluctuation in the solvent of the compound. And there is a problem that attribute values such as physical property values can not be predicted accurately.

  Further, in the above-mentioned prior art, although the radius r as the threshold must be set appropriately, there is a problem that the radius r is difficult to adjust because it varies depending on the size of the compound and the problem to be solved.

  The present invention has been made in view of the above problems, and an object thereof is to provide an attribute value prediction device and an attribute value prediction program capable of accurately predicting an attribute value of a compound.

  In order to achieve the above object, the attribute value prediction device of the invention according to claim 1 comprises a graph generation unit for generating a graph representing a compound in a graph structure, and a node given to a node of the graph The feature amount extraction unit extracts the feature amount of the compound by graph wavelet transform of the signal, and the attribute value prediction unit predicts the attribute value of the compound based on the feature amount.

  In the invention according to claim 2, the graph generation unit generates a proximity graph, which generates a proximity graph from the distance matrix, and a distance matrix calculation unit which calculates a distance matrix representing a distance between a plurality of atoms constituting the compound. The feature amount extraction unit extracts a feature amount of the compound by performing a graph wavelet transform on a node signal of the neighborhood graph.

  An attribute value prediction program according to a third aspect of the present invention is an attribute value prediction program for causing a computer to function as each part of the attribute value prediction device according to the first or second aspect.

  According to the present invention, it is possible to predict physical property values of a compound with high accuracy.

It is a functional block diagram of an attribute value prediction device. It is a block diagram of the hardware constitutions of an attribute value prediction device. It is a flowchart of attribute value prediction processing.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

First Embodiment

  FIG. 1 is a diagram showing a configuration example of an attribute value prediction device 10 according to the present embodiment. The attribute value prediction device 10 includes a graph generation unit 12, a feature quantity extraction unit 14, and an attribute value prediction unit 16.

  The graph generation unit 12 generates a graph representing the compound in a graph structure. Specifically, the graph generation unit 12 includes a distance matrix calculation unit 18 and a proximity graph generation unit 20.

  The distance matrix calculating unit 18 calculates a distance matrix representing the distance between a plurality of atoms constituting the compound.

  The neighborhood graph generation unit 20 generates a neighborhood graph from the distance matrix calculated by the distance matrix calculation unit 18.

  The feature quantity extraction unit 14 extracts the feature quantity of the compound by subjecting the node signal given to the node of the neighborhood graph generated by the neighborhood graph generation unit 20 to graph wavelet transformation.

  The attribute value prediction unit 16 predicts the attribute value of the compound based on the feature amount extracted by the feature amount extraction unit 14. In addition, although this embodiment demonstrates the case where an attribute value is a physical property value of a compound, it is not restricted to this.

  FIG. 2 is a block diagram showing the hardware configuration of the attribute value prediction device 10. As shown in FIG.

  The attribute value prediction device 10 has a function as a computer, and includes a central processing unit (CPU) 30A, a read only memory (ROM) 30B, a random access memory (RAM) 30C, a non-volatile memory 30D, a hard disk 30E, and a communication interface. (I / F: Interface) part 30F is provided. The CPU 30A, the ROM 30B, the RAM 30C, the non-volatile memory 30D, the hard disk 30E, and the communication interface unit 30F are connected to one another via a bus 30G.

  The CPU 30A is a central processing unit and executes various programs and controls each part. At this time, the program is executed with the RAM 30C as a work area.

  The non-volatile memory 30D is a memory that retains data contents even when the power to the attribute value prediction device 10 is turned off.

  For example, an attribute value prediction program described later is stored in the hard disk 30E.

  The communication interface unit 30F communicates with an external device (not shown).

  Next, an attribute value prediction process executed by the CPU 30A will be described with reference to the flowchart shown in FIG. 3 as the operation of the present embodiment. The CPU 30A reads and executes the attribute value prediction program stored in the hard disk 30E.

In step S100, a distance matrix representing the distances between a plurality of atoms constituting the compound is calculated. For example, in a certain compound C, the physical property values of N atoms contained in the compound are represented as f = {f ₁ , f ₂ ,..., F _N }. Further, the three-dimensional position of N atoms contained in the compound C is represented as x = {x ₁ , x ₂ ,... X _N }.

And the distance between each atom

Calculate Here, A is an N × N distance matrix representing distances between N atoms. The interatomic distance A _{i, j} between atom i and atom j is expressed by the following equation.

  In step S102, a proximity graph is generated from the distance matrix calculated in step S100. In the present embodiment, the compound is expressed using a graph in a so-called graph theory. A graph in graph theory is composed of a set of nodes and a set of edges connecting nodes and nodes. Here, a node represents an atom of a compound, and a node signal given to the node represents a physical property value of the atom. Moreover, the side connecting each node represents the distance between atoms.

  As a method of generating a proximity graph from the distance A between atoms, for example, as a first method, there is a method of connecting a predetermined number of nodes in order from nodes closer in distance based on a distance matrix.

  Further, as a second method, there is a method of connecting nodes whose distance is within a predetermined threshold value or less based on a distance matrix.

  As a third method, there is a method of connecting nodes using the inverse of the distance as the weight of the graph.

  The method of generating the neighborhood graph is not limited to the first to third methods described above.

Here, let Wt be the weight of the generated neighborhood graph. However, for the atom i, j, Wt _{i, i} = 0 and Wt _{i, j} = Wt _{j, i} . As in the third method, when the reciprocal of the distance is used as the weight of the graph, Wt _{i, j} = 1 / A _{i, j} . That is, in the neighborhood graph, the weight of the graph decreases as the distance increases.

  In step S104, the node signal given to the nodes of the graph is graph wavelet transformed using the neighborhood graph generated in step S102. The graph wavelet transform will be described below.

  The graph Laplacian of the neighborhood graph generated in step S102 is L. In the present embodiment, L is a random walk normalized graph Laplacian as an example, and is expressed by the following equation.

L = D- ¹ (D-W)

here,

It is. Also, let it be assumed that the l-th eigenvalue of the graph Laplacian is λ _l and the eigenvector is u _l .

  Then, the physical property value f of the atom which is a node signal is graph-wavelet transformed by the following equation.

... (1)

  Here, W represents the feature value obtained by graph wavelet transformation. Also, s is a scale parameter, and a is a position parameter. Also, g is a kernel function, which corresponds to the wavelet function in the normal wavelet transform. The kernel function g may be, for example, a Mexican hat function, but is not limited thereto.

  The graph wavelet transform is a process of projecting the function f in the eigenvector direction (frequency direction) of the graph, filtering it in the frequency space, and returning it to the original space. By this graph wavelet transformation, it is possible to obtain the feature value W in which the physical property values of the surrounding atoms are convoluted.

In step S106, the physical property value y is predicted by a predetermined prediction process g _θ based on the feature amount W and the other feature amounts X obtained by the graph wavelet transformation in step S104.

A well-known learning method can be used for the prediction process g _θ of the physical property value y. As prediction process g _theta can be used, for example SVM (Support Vector Machine), xgboost (eXtreme Gradient Boosting) and the like.

Here, with the feature amount W and the other feature amount X as input, the prediction process g _θ predicts the physical property value y using a predetermined parameter θ. Although the parameter θ will be described later, for example, a value estimated in advance from learning data is used.

As another feature amount X, for example, when the compound is a protein, PSSM (Position Specific Scoring Matrix, genetic feature amount) or the like is used. PSSM is, for example, a feature that indicates which amino acid is present more at a certain position when a certain protein is compared with a similar protein. The physical property value y of the unknown compound is obtained as an output by executing the prediction process g _θ with the feature amount W and the other feature amounts X as inputs.

Next, optimization of the parameter θ of the prediction process g _θ will be described.

For optimization of the parameter θ, a plurality of compounds whose physical property values have already been measured are used. Assuming that the characteristic amount of this compound is W _train and X _train , and the measured physical property value is y _train , using these, the parameter θ is set so that y _train = g _θ (W _train , X _train ). Optimize. Thus, the physical property value y can be predicted from the feature amounts W and X for a new compound whose physical property value is not measured.

  Thus, in the present embodiment, by capturing the compound as a graph, it is possible to obtain a data structure that is not easily affected by the size of the compound and the like. Then, by extracting feature quantities using wavelet functions such as Mexican hat function used in wavelet transform, it does not depend on parameters such as threshold as in the prior art, and influences fluctuation in the solvent of the compound. It is difficult to do so, and it is possible to generate feature quantities that include the entire compound. Thereby, the physical property value of the compound can be accurately predicted.

DESCRIPTION OF SYMBOLS 10 Attribute value prediction apparatus 12 Graph generation part 14 Feature quantity extraction part 16 Attribute value prediction part 18 Distance matrix calculation part 20 Neighborhood graph generation part

Claims (3)

A graph generation unit that generates a graph representing the compound in a graph structure;
A feature amount extraction unit that extracts feature amounts of the compound by performing graph wavelet transform on node signals given to nodes of the graph;
An attribute value prediction unit that predicts the attribute value of the compound based on the feature amount;
Attribute value prediction device equipped with The graph generation unit includes a distance matrix calculation unit that calculates a distance matrix that indicates distances between a plurality of atoms that constitute the compound, and a neighborhood graph generation unit that generates a neighborhood graph from the distance matrix,
The attribute value prediction device according to claim 1, wherein the feature amount extraction unit extracts feature amounts of the compound by performing graph wavelet transform on node signals of the neighborhood graph.   The attribute value prediction program for functioning a computer as each part of the attribute value prediction apparatus of Claim 1 or Claim 2.