JP2007052507A - Biological information processor, biological information processing method, and biological information processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technology excellent in validity of prediction and classification reference in classifying biological information into three or more classes. <P>SOLUTION: A multi-margin support vector machine (MM-SVM) partitions input space into three or more regions by producing two or more mutually parallel separation faces in the input space. Therefore, risk of "over learning" or "excess adaptation" having been apt to occur in an existing support vector machine can be reduced. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a biological information processing apparatus, a biological information processing method, and a biological information processing program.

  The most basic problem in pattern recognition is to develop a classifier that determines which class an object belongs to from an input vector obtained by measuring an unknown recognition object. For this purpose, it is necessary to learn as a knowledge a stochastic correspondence between an input vector and a class from a training sample whose class membership is known. In order to identify an unknown recognition target, a method for estimating (determining) which class it belongs to using the learned probabilistic knowledge must be specified. At that time, it is desirable to minimize the probability of erroneous identification. In the ideal case where the stochastic correspondence between the input vector and the class is completely known, there is a theoretically optimal identification method (Bayes identification method). However, in an actual pattern recognition problem, it is rare that a stochastic correspondence between a feature vector and a class is completely known, and it is necessary to learn such a stochastic relation from training data.

  As such a pattern recognition technique, a pattern recognition technique called a support vector machine (Support Vector Machine, SVM) has recently attracted attention. The support vector machine is a technique for constructing two classes of pattern classifiers using linear threshold elements.

  The support vector machine, which has been extended so that a nonlinear discriminant function can be constructed using a method called kernel trick, is considered to be one of the learning models with the best recognition performance among the currently known methods. Yes. In general, in order for classifiers trained using the kernel learning method to exhibit high discrimination performance even for unlearned data that is not included in training samples, it is necessary to devise measures to improve generalization ability It is.

  The reason why the support vector machine can exhibit excellent recognition performance is because there is a device for obtaining high identification performance (generalization performance) for unlearned data. That is, the support vector machine learns the parameters of the linear threshold element from the training sample on the basis of “margin maximization”. However, the support vector machine is basically a learning method for constructing a discriminator for discriminating two classes. In order to construct a multi-class discriminator such as biological information, a plurality of support vectors are used. It is necessary to devise such as combining vector machines.

  On the other hand, in recent years, research and development for using a support vector machine for biological information processing has been actively conducted in a technical field called bioinformatics in response to completion of decoding of the human genome and development of a DNA microarray.

  A technique related to diagnosis, prognosis, and identification of treatment target candidates of multiple myeloma based on conventional gene expression profiling is described in Patent Document 1. The technique described in that document obtains gene expression data by means of a DNA microarray; and further on said data by a method selected from the group consisting of logistic circuits, decision trees, voting groups, naive Bayes, Bayesian networks and support vector machines. Perform statistical analysis; including each step.

  Further, as a conventional technique for detecting the occupant state, for example, there is one described in Patent Document 2. The technique described in this document measures a pressure pattern by arranging a plurality of sensor portions on a seat portion where pressure is generated by the seating of an occupant, and inputs the measured pressure pattern data to a support vector machine (SVM). The SVM performs feature extraction and discrimination by the SVM separation curved surface, and classifies adults / children for the seated occupant.

  Moreover, as a technique regarding the grasping of the state of an animal using acquisition and analysis of a conventional biological signal, for example, there is one described in Patent Document 3. The technique described in this document is based on a support vector learned from a reference vector of a predetermined database that reflects the behavior, intention, and emotion in a specific state for each animal type with respect to a feature vector acquired by a predetermined method. Applying a machine shunt to grasp the animal's feelings and intentions, including the animal's hunger, tension and fear, and stool desire.

  Further, as a technique related to the conventional three-dimensional lesion detection of a computer, there is one described in Patent Document 4, for example. The technique described in that document includes distinguishing the set of lesion candidates using at least one of a linear discriminator, a secondary discriminator, a neural network, and a support vector machine.

  Moreover, as a technique regarding the molecular signature of a highly lethal cancer, there is one described in Patent Document 5, for example. The technique described in the document includes: a) obtaining a value for one or more features for each of a plurality of members of each class; b) determining a Wilcoxon rank score for each of the features, and non-predicting Excluding features, and c) ranking the remaining features by predicted accuracy using a support vector machine.

  Moreover, as a technique for identifying a pattern of a conventional biological system, for example, there is one described in Patent Document 6. The technique described in that document involves the use of Support Vector Machine (SVM) and RFE (Repeated Feature Elimination) for pattern identification that is useful for medical diagnosis, prognosis and treatment. SVM-RFE can be used with variable data sets.

JP 2005-512557 A JP 2003-344196 A Japanese Patent Laid-Open No. 2004-138 JP 2005-506140 A JP-T-2005-503779 JP 2005-502097 gazette

  However, the prior art described in the above literature has room for improvement in the following points when it is intended to be applied to biological information processing.

  In general, as a common problem when a learning machine such as a support vector machine is applied to biological information processing, it is a problem to overcome the risk of “over-learning” or “over-fitting”.

  However, the support vector machine is basically a learning method for constructing a discriminator for discriminating two classes. In order to construct a multi-class discriminator such as a classification of biological information, a plurality of discriminators are used. It is necessary to devise such as combining support vector machines. Therefore, when the input vector or feature vector is classified into three or more classes, the support vector machine may “overlearn” or “overfit” the training pattern for each of the plurality of support vector machines. Therefore, the risk of “overlearning” or “overfit” as a whole tends to increase.

  Further, as described in paragraph [0088] of Patent Document 6, the risk of such “overlearning” or “overfit” is studied by the number of dimensions of the input vector or feature vector, for example, in a microarray. It is likely to occur when the number of training patterns is large, such as thousands of genes, and relatively small, such as a few dozen patients.

  In this regard, in the techniques described in Patent Documents 1 to 5, no contrivance for overcoming the risk of “over-learning” or “over-fitting” is provided. For this reason, in the techniques described in Patent Documents 1 to 5, when processing biological information in which the number of feature vectors (input vectors) is relatively large and the number of training patterns is relatively small, The risk of “over-learning” or “over-fitting” tends to increase.

  For this reason, in the techniques described in Patent Documents 1 to 5, the support vector machine is “over-learned” or “over-fitted” to the training pattern, and as a result, the validity of prediction in solving a general problem related to biological information. In some cases, this was not sufficient. Therefore, the techniques described in Patent Documents 1 to 5 have room for further improvement in terms of validity of prediction or classification criteria when classifying biological information into three or more classes.

  On the other hand, in the technique described in Patent Document 6, as described in paragraph [0088] of Patent Document 6, the dimension of the feature space is reduced in order to overcome the risk of “over-learning” or “over-fitting”. The idea is given as an issue.

  In other words, the technique described in Patent Document 6 uses a combination of a support vector machine and RFE (repetitive feature exclusion), thereby reducing the dimension of the feature space and reducing the risk of “over-learning” or “over-fitting”. It is reduced to some extent. However, the reduction of the risk of “over-learning” or “over-fitting” by such RFE (repetitive feature exclusion) has not been sufficiently satisfactory.

  In the technique described in Patent Document 6, some useful information included in the excluded dimension cannot be used for classification by reducing the dimension of the feature space using RFE (repetitive feature exclusion). there is a possibility.

  For this reason, in the technique described in Patent Document 6, the support vector machine is “over-learned” or “over-fitted” to the training pattern to some extent, and some useful information included in the excluded dimensions is used for classification. In some cases, the validity of predictions when solving general problems related to biological information is not sufficient. Therefore, the technique described in Patent Document 6 has room for further improvement in terms of the validity of prediction or classification criteria when classifying biological information into three or more classes.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a technique excellent in the validity of prediction or classification criteria when classifying biological information into three or more classes.

  According to the present invention, there is provided a biological information processing apparatus for predicting a classification of biological information, and acquiring a plurality of known biological information classified into three or more classes. An input vector generation unit that generates a plurality of input vectors including a plurality of biological information acquired by the unit and the known information acquisition unit in the input space, and corresponds to the plurality of input vectors and the plurality of input vectors, respectively. A separation unit that generates two or more parallel separation planes in the input space based on a plurality of biological information classifications, and separates the input space into three or more regions respectively corresponding to three or more classes; An unknown information acquisition unit that acquires biological information whose classification is unknown, and an unknown vector that includes biological information acquired by the unknown information acquisition unit is separated into three or more regions by two or more separation planes. Enter Biology corresponding to an unknown vector based on a region where an unknown vector is arranged in an input space that is separated into three or more regions by two or more separation planes and an unknown vector generation unit that generates in the space A biology comprising: a prediction determination unit that predicts and determines a classification of the target information; and a prediction classification output unit that outputs a classification of biological information corresponding to the unknown vector predicted and determined by the prediction determination unit. An information processing apparatus is provided.

  According to this configuration, when the biological information is classified into three or more classes, the input space is separated into three or more regions respectively corresponding to the three or more classes by two or more parallel separation surfaces. Therefore, it is possible to achieve high discrimination while avoiding “over-learning” or “over-fitting”, so that the validity of prediction when classifying biological information into three or more classes can be improved.

  According to the present invention, there is provided a biological information processing apparatus for predicting a classification of biological information, and acquiring a plurality of known biological information classified into three or more classes. , An input vector generation unit for generating a plurality of input vectors including a plurality of biological information acquired by the known information acquisition unit in the input space, and converting the plurality of input vectors in the input space by nonlinear mapping A plurality of feature vectors respectively corresponding to the plurality of input vectors in a higher-order feature space than the input space, and a plurality of feature vectors and a plurality of feature vectors respectively corresponding to the plurality of feature vectors. Based on the classification of biological information, two or more parallel separation planes are generated in the feature space, and the feature space is separated into three or more regions respectively corresponding to three or more classes. A separation unit, an unknown information acquisition unit that acquires biological information whose classification is unknown, and an unknown vector that includes biological information acquired by the unknown information acquisition unit by non-linear mapping, An unknown vector generation unit that generates a corresponding transformed unknown vector in a feature space that is separated into three or more regions by two or more separation surfaces, and a feature space that is separated into three or more regions by two or more separation surfaces Of these, the prediction determination unit that predicts and determines the classification of biological information corresponding to the conversion unknown vector based on the area where the conversion unknown vector is arranged, and the conversion unknown vector that is predicted and determined by the prediction determination unit There is provided a biological information processing apparatus comprising: a predicted classification output unit that outputs a classification of biological information.

  According to this configuration, when biological information is classified into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a feature space higher in order than the input space. By separating the feature space into three or more regions respectively corresponding to three or more classes by using two or more parallel separation surfaces, high discriminability can be achieved while avoiding “over-learning” or “over-fitting”. Since it is realizable, the validity of prediction in classifying biological information into three or more classes can be improved.

  According to the present invention, there is provided a biological information processing apparatus for predicting a classification of biological information, and acquiring a plurality of known biological information classified into three or more classes. , An input vector generation unit for generating a plurality of input vectors including a plurality of biological information acquired by the known information acquisition unit in the input space, and converting the plurality of input vectors in the input space by nonlinear mapping A plurality of feature vectors respectively corresponding to the plurality of input vectors in a higher-order feature space than the input space, and a plurality of feature vectors and a plurality of feature vectors respectively corresponding to the plurality of feature vectors. Based on the classification of biological information, two or more parallel separation planes are generated in the feature space, and the feature space is separated into three or more regions respectively corresponding to three or more classes. Inverse transformation that generates two or more separation surfaces in the input space by separating the input space into three or more regions by inversely transforming the separation unit and two or more separation surfaces generated by the separation unit by a non-linear mapping An unknown vector including biological information acquired by the unknown information acquisition unit and the unknown information acquisition unit into three or more regions by two or more separation planes. Based on the area where the unknown vector is arranged in the input space generated by the unknown vector generation unit generated in the input space to be separated and the input space separated into three or more areas by two or more separation planes. A prediction determination unit that predicts and determines a corresponding biological information classification; and a prediction classification output unit that outputs a biological information classification corresponding to the unknown vector predicted and determined by the prediction determination unit. The biological information processing apparatus is provided.

  According to this configuration, when biological information is classified into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a feature space higher in order than the input space. The feature space is separated into three or more regions respectively corresponding to three or more classes by two or more parallel separation surfaces, and the generated two or more separation surfaces are inversely transformed by a non-linear mapping to obtain an input space. By generating in the input space two or more separation planes that separate three into three or more regions, it is possible to achieve high discrimination while avoiding “overlearning” or “overfitting”. The validity of prediction when classifying scientific information into three or more classes can be improved.

  According to the present invention, a biological information processing apparatus that generates a classification standard for classifying biological information, and acquires a plurality of known biological information classified into three or more classes. A known information acquisition unit; an input vector generation unit that generates a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space; a plurality of input vectors and a plurality of input vectors; Two or more parallel separation planes are generated in the input space based on a plurality of biological information classifications corresponding respectively to three or more regions corresponding to the three or more classes, respectively. There is provided a biological information processing apparatus comprising: a separation unit that separates the information into two parts; and a classification reference output unit that outputs a classification reference including information defining two or more separation planes generated by the separation unit.

  According to this configuration, when the biological information is classified into three or more classes, the input space is separated into three or more regions respectively corresponding to the three or more classes by two or more parallel separation surfaces. Since high discrimination can be realized while avoiding “overlearning” or “overfitting”, the validity of the classification criteria when classifying biological information into three or more classes can be improved.

  According to the present invention, a biological information processing apparatus that generates a classification standard for classifying biological information, and acquires a plurality of known biological information classified into three or more classes. A known information acquisition unit, an input vector generation unit that generates a plurality of input vectors including a plurality of biological information acquired by the known information acquisition unit in the input space, and a plurality of input vectors in the input space are nonlinear By converting by mapping, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a higher-order feature space than the input space, and a plurality of feature vectors and a plurality of feature vectors, respectively Based on the classification of the corresponding plurality of biological information, two or more parallel separation planes are generated in the feature space, and the feature space corresponds to three or more classes respectively. A separation unit for separating the frequency, a classification reference output unit for outputting the classification criteria including information defining two or more separate surface produced by separation unit, the biological information processing apparatus including a is provided.

  According to this configuration, when biological information is classified into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a feature space higher in order than the input space. By separating the feature space into three or more regions respectively corresponding to three or more classes by using two or more parallel separation surfaces, high discriminability can be achieved while avoiding “over-learning” or “over-fitting”. Since it can be realized, the validity of the classification criteria when classifying biological information into three or more classes can be improved.

  According to the present invention, a biological information processing apparatus that generates a classification standard for classifying biological information, and acquires a plurality of known biological information classified into three or more classes. A known information acquisition unit, an input vector generation unit that generates a plurality of input vectors including a plurality of biological information acquired by the known information acquisition unit in the input space, and a plurality of input vectors in the input space are nonlinear By converting by mapping, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a higher-order feature space than the input space, and a plurality of feature vectors and a plurality of feature vectors, respectively Based on the classification of the corresponding plurality of biological information, two or more parallel separation planes are generated in the feature space, and the feature space corresponds to three or more classes respectively. Generates two or more separation surfaces in the input space by separating the input space into three or more regions by inversely transforming the separation unit that divides into regions and two or more separation surfaces generated by the separation unit by nonlinear mapping There is provided a biological information processing apparatus comprising: an inverse transforming unit that outputs a classification criterion that includes information defining two or more separation planes generated by the inverse transforming unit.

  According to this configuration, when biological information is classified into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors are generated in a feature space higher in order than the input space. The feature space is separated into three or more regions respectively corresponding to three or more classes by two or more parallel separation surfaces, and the generated two or more separation surfaces are inversely transformed by a non-linear mapping to obtain an input space. By generating in the input space two or more separation planes that separate three into three or more regions, it is possible to achieve high discrimination while avoiding “overlearning” or “overfitting”. The validity of classification criteria when classifying scientific information into three or more classes can be improved.

  Note that the above-described device is one embodiment of the present invention, and the device of the present invention may be any combination of the above components. The method, system, computer program, recording medium, etc. of the present invention have the same configuration.

  According to the present invention, biological information is classified into three or more classes in order to separate the input space or feature space into three or more regions respectively corresponding to three or more classes by two or more parallel separation surfaces. Improve the validity of prediction or classification criteria.

  In the present invention, the above-described conversion unit can be configured to perform calculation when an input vector is converted to a feature vector using a kernel function that satisfies semi-fixed property.

  According to this configuration, it is possible to linearly separate an input vector that is difficult to be linearly separated by performing a calculation when the input vector is converted into a feature vector using a kernel function that satisfies the semi-fixed property. Since it can be converted into possible feature vectors, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

  In the present invention, the kernel function may be one or more kernel functions selected from the group consisting of a linear kernel function, a polynomial kernel function, and an RBF kernel function.

  According to this configuration, since one or more types of kernel functions selected from the group consisting of a linear kernel function, a polynomial kernel function, and an RBF kernel function are all kernel functions that satisfy semi-positive stationary properties, The generalization of prediction or classification criteria when classifying into three or more classes can be improved.

  According to the present invention, the linear separation unit can be configured to generate two or more parallel separation planes using a support vector machine.

  According to this configuration, since the support vector machine has a high discrimination performance, in combination with the above-described various devices, the discrimination of the prediction or the classification criterion when classifying the biological information into three or more classes is performed. Can be improved.

  According to the present invention, the support vector machine can be configured to maximize the margin width of the separation surface having the smallest margin width among the two or more parallel separation surfaces.

  According to this configuration, since the support vector machine separates the input vector or the feature vector with a margin as much as possible, the identification of the prediction or classification criterion when classifying the biological information into three or more classes. And generalization can be improved.

  According to the present invention, the separation unit can be configured to arrange the separation surface so as to be positioned at the center of the margin width.

  According to this configuration, since the support vector machine can uniquely determine the separation plane, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

  According to the present invention, the support vector machine may be a support vector machine extended by a soft margin method.

  According to this configuration, the support vector machine can linearly separate the input vector or the feature vector by the soft margin method even when it is difficult to linearly separate the input vector or the feature vector. The generalization of prediction or classification criteria when classifying into classes can be improved.

In the present invention, two or more parallel separation planes are expressed as follows: b = w ^T x (where x is an input vector or feature vector, w is a parameter vector, b is a bias, T Is an arithmetic symbol indicating transposition), w is a parameter vector that is the same on the two or more separation planes, and b is a value different from each other on the two or more separation planes. A bias that takes

According to this configuration, two or more separation planes represented by b = w ^T x are parallel to each other, and high discrimination can be realized while avoiding “overlearning” or “overfitting”. The validity of prediction or classification criteria when classifying biological information into three or more classes can be improved.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, the same reference numerals are given to the same components, and the description will be omitted as appropriate.

<Embodiment 1>
FIG. 1 is a functional block diagram showing an outline of the configuration of the biological information processing apparatus 100. The biological information processing apparatus 100 is an apparatus for predicting the classification of biological information whose classification is unknown. Alternatively, the biological information processing apparatus 100 is also an apparatus that generates a classification standard for classifying biological information whose classification is unknown.

  The biological information processing apparatus 100 is provided with a known information acquisition unit 102. As shown in FIG. 2, the known information acquisition unit 102 acquires a plurality of known biological information classified in advance into three or more classes from the outside.

  In FIG. 2, an example of an input file (biological information whose classification is known) acquired from the outside by the known information acquisition unit 102 is described. In this example, the known information acquisition unit 102 obtains predetermined medical observation data (such as drug resistance, disease level, survival rate, gene name, white blood cell volume, blood pressure, etc.) for each sample acquired from each patient. get.

  These medical observation data are classified in advance into a plurality of three or more classes by doctors, dentists, nurses, clinical technologists, engineers of clinical contractors, researchers in the field of biology, and the like. In the example of FIG. 2, the first stage, the second stage,... The Mth stage, and M classes are classified in advance. For example, in the example of FIG. 2, a researcher who studies an anticancer drug can be classified into low, medium, and high for anticancer drug resistance.

  In addition, the biological information processing apparatus 100 is provided with a feature amount extraction unit 118. The feature quantity extraction unit 118 is based on predetermined extraction criteria (drug resistance, morbidity, survival rate, gene, etc.) from biological information whose classification for each sample acquired from each patient included in the input file is known. Extract feature quantities (such as name, white blood cell volume, blood pressure). For example, the top d feature quantity 1, feature quantity 2,... Feature quantity D can be extracted based on a predetermined extraction criterion. In this case, the types of extracted feature quantities are the same for the biological information of the samples obtained from each patient. On the other hand, the value of each feature amount differs for each biological information of the sample acquired from each patient.

  The feature quantity extraction method used in the feature quantity extraction unit 118 includes, for example, a one-way analysis of variance F-test or t-test, two-group comparison t-test (equal variance, unequal variance), and values at each stage. In some cases, a correlation coefficient or the like can be used.

  The biological information with known classification from which the feature quantity is extracted by the feature quantity extraction unit 118 is stored in the known information storage unit 104 and sent to the classification reference generation unit 106. As shown in FIG. 3 to be described later, the classification reference generation unit 106 uses, for example, a learning machine such as a multi-margin support vector machine developed by the present inventor based on the acquired biological information with known classification. Generating a classification criterion for predicting the classification of target information.

  FIG. 3 is a conceptual diagram for explaining a difference in function between an existing support vector machine and the multi-margin support vector machine of the present embodiment used in the separation unit 214. In an existing support vector machine (SVM), vectors of a predetermined dimension (d-dimensional vectors, that is, vectors including d feature value values) are classified into two classes. On the other hand, the multi-margin support vector machine (MM-SVM) developed by the present inventors classifies vectors of a predetermined dimension into three or more classes (M stages). M is an integer of 3 or more. The multi-margin support vector machine will be described later with reference to FIGS.

  Returning to FIG. 1, the classification standard generated by the classification standard generation unit 106 is stored in the classification standard storage unit 108 and sent to the classification prediction determination unit 112 or sent to the output unit 116, and directly to the outside. Is output. The classification standard may be a separation plane itself generated by a multi-margin support vector machine, which will be described later, or a combination of parameters for defining the separation plane.

  On the other hand, the biological information processing apparatus 100 is provided with an unknown information acquisition unit 110 in addition to the known information acquisition unit 102. The unknown information acquisition unit 110 acquires biological information whose classification is unknown from the outside.

  The input file (biological information whose classification is unknown) acquired from the outside by the unknown information acquisition unit 110 is, for example, a predetermined (drug resistance, disease level, survival rate, gene, for each sample acquired from each patient) Ordered medical observation data (name, white blood cell volume, blood pressure, etc.). However, unlike the example of FIG. 2, these medical observation data are obtained by doctors, dentists, nurses, clinical technologists, clinical contractor engineers, biology researchers, etc. They are not pre-classified into classes.

  The feature quantity extraction unit 118 is based on predetermined extraction criteria (drug resistance, morbidity, survival rate, gene, etc.) from biological information whose classification for each sample acquired from each patient included in the input file is unknown. Extract feature quantities (such as name, white blood cell volume, blood pressure). For example, the top d feature quantity 1, feature quantity 2,... Feature quantity D can be extracted based on a predetermined extraction criterion. In this case, the types of extracted feature quantities are the same as in the case of biological information whose classification is known. On the other hand, the value of each feature amount differs for each biological information of the sample acquired from each patient. The biological information with unknown classification from which the feature amount is extracted by the feature amount extraction unit 118 is sent to the classification prediction determination unit 112.

  The classification prediction determination unit 112 predicts and determines the classification of biological information whose classification is unknown. The classification prediction determination unit 112 applies the biological information with unknown classification to the classification reference acquired from the classification reference storage unit 108 by applying the biological information with unknown classification extracted by the feature amount extraction unit 110. Predictive classification of The classification prediction result of the biological information with unknown classification determined by the classification prediction determination unit 112 is stored in the classification prediction storage unit 114 and sent to the output unit 116.

  The output unit 116 outputs the classification prediction result of the biological information with unknown classification determined by the classification prediction determination unit 112 to the outside, or outputs two or more separation planes generated by the classification reference generation unit 106. The classification criteria including the information to be defined is output to the outside.

  In addition, the biological information processing apparatus 100 is provided with a cross validation unit 120. As will be described later with reference to FIG. 29, the cross-validation unit 120 performs classification prediction determination results obtained from the classification criteria generated from the above-described known information for the known information used as the learning data of the above-described multi-margin support vector machine. Perform cross-validation. Note that details of the cross-validation method will be described later with reference to FIG. 29, and therefore description thereof will not be repeated here.

  Further, the cross validation unit 120 calculates the estimated prediction rate of the above-described classification standard based on the result of the cross validation. The estimated prediction rate obtained in this way is stored in the estimated prediction rate storage unit 122, output by the output unit 116, and presented to the user of the biological information processing apparatus 100.

  In the biological information processing apparatus 100, the cross-validation unit 120 performs cross-validation of the learning data, thereby calculating the estimated prediction rate of the above-described classification standard and presenting it to the user via the output unit 116. For this reason, the user of the biological information processing apparatus 100 grasps the estimated prediction rate of the classification standard generated from the learning data, and then applies unknown data to the classification standard to determine the degree of reliability. A classification prediction determination result can be obtained. For this reason, the risk that the user overconfidents the classification prediction determination result based on the classification criterion with low reliability is reduced, and the risk in research and development or medical care is reduced.

  FIG. 4 is a functional block diagram illustrating details of the configuration of the classification reference generation unit 106. The biological information acquisition unit 202 acquires biological information whose classification is known from the known information storage unit 104 and stores the biological information in the biological information storage unit 204. On the other hand, the classification acquisition unit 210 acquires a known classification corresponding to the biological information and stores it in the classification storage unit 212.

  The input vector generation unit 206 generates, in the input space, a plurality of input vectors including a plurality of biological information from biological information whose classification is known acquired from the known information storage unit 104. Specifically, as shown in FIG. 2, the sample data derived from each patient including the top d feature quantities extracted by the feature quantity extraction unit is converted into a d-dimensional vector to obtain the input vector. Generate. As the input space, for example, a space based on d-dimensional Euclidean geometry is used. The input vector generated by the input vector generation unit 206 is stored in the input vector storage unit 208 and sent to the separation unit 214.

Here, in order to realize pattern recognition using a support vector machine or the like, it is necessary to measure (extract) some feature amount from the recognition target. In general, not only one type of feature amount but also a plurality of feature amounts are measured and used in many cases. Such feature amounts are generally represented collectively as an input vector x ^T = (x ₁ ,..., X _d ). Here, ^{x T} represents the transpose of vector x. This is merely an example of an input vector.

  As shown in FIG. 3, the separation unit 214 uses the multi-margin support vector machine based on a plurality of input vectors and a plurality of known classifications so that the input space corresponds to three or more classifications (classes). To separate areas. The multi-margin support vector machine generates two or more mutually parallel separation planes in the input space, thereby separating the input space into three or more regions respectively corresponding to three or more classifications (classes). The separation surface generated by the separation unit 214 is stored in the separation surface storage unit 216 and sent to the classification reference output unit 218.

  FIG. 5 is a functional block diagram illustrating details of the configuration of the separation unit 214. The input vector acquisition unit 602 acquires an input vector from the input vector storage unit 208 and sends it to the support vector machine 606. The classification acquisition unit 604 acquires the classification corresponding to the input vector from the classification storage unit 212 and sends it to the support vector machine 606. The support vector machine 606 separates the input space in which the input vectors are arranged into three or more regions by generating two or more parallel separation surfaces in the input space. The generated separation surface is output from the separation unit 214 by the separation surface output unit 608 and stored in the separation surface storage unit 216.

  Again, as shown in FIG. 3, in the existing support vector machine (SVM), a vector of a predetermined dimension (d-dimensional vector, that is, a vector including d feature value values) is classified into two classes. . On the other hand, the multi-margin support vector machine (MM-SVM) developed by the present inventors classifies vectors of a predetermined dimension into three or more classes (M stages). M is an integer of 3 or more.

  The multi-margin support vector machine (MM-SVM) developed by the inventor generates two or more parallel separation planes in the input space, thereby separating the input space into three or more regions. It is possible to reduce the risk of “over-learning” or “over-fitting” that is likely to occur in the support vector machine. It has been experimentally verified that the risk of such “overlearning” or “overfit” can be reduced, as will be described later.

  In general, biological information has a lot of ordered data (drug resistance, morbidity, survival rate, etc.), so these various types of biological information can be integrated into doctors, dentists, nurses, clinical Classifications that are specifically determined by laboratory technicians, clinical contractor engineers, biology researchers, etc. according to the situation are also ordered data, such as low, medium, and high. There are many cases. Therefore, if two or more parallel separation planes are generated as in the multi-margin support vector machine developed by the present inventor, ordered data can be appropriately classified.

  In FIG. 4, the classification reference generation unit 218 generates a classification reference including information defining two or more separation planes generated by the separation unit 214. Then, the generated classification standard is stored in the classification standard storage unit 218. The classification standard may be a separation plane itself generated by a multi-margin support vector machine, which will be described later, or a combination of parameters for defining the separation plane.

  FIG. 6 is a conceptual diagram for explaining in detail the difference in learning method between the existing support vector machine and the multi-margin support vector machine used in the first embodiment. An existing support vector machine (SVM) is a method of constructing a two-class pattern classifier using the simplest linear threshold element as a neuron model. The existing support vector machine learns the parameters of the linear threshold element from the training sample set (input vector set) on the basis of “margin maximization”.

Here, the input vector is expressed as x ^T = (x ₁ ,..., X _d ). Incidentally, ^{x T} represents the transpose of vector x. D is the number of feature quantities. The total number of classes to be recognized is M (an integer of 3 or more), and each class is C ₁ , C ₂ ,..., C _M.

At this time, the linear threshold element used in the existing support vector machine is a linear discriminant function (score function) for the input vector, and the following expression: g (x) = g (x | w) = w ^T x = w ₁ x ₁ +... + w _d x _d In this linear threshold element, the identification plane when g (x) = b corresponds to a separation plane that separates the input space. Here, w is a parameter corresponding to the synaptic load, and b is a threshold value. This is geometrically equivalent to dividing the input space into two by the separation plane (identification plane).

  6 to 11, it is assumed that this set of input vectors can be linearly separated. That is, it is assumed that the set of input vectors can be divided without error by adjusting the parameters of the linear threshold elements well.

  Even if a set of input vectors is linearly separable, generally, a parameter that divides an input vector set without error is not uniquely determined. In the existing support vector machine, a separation plane that separates input vectors as much as possible is required instead of passing the input vectors. Specifically, the margin with the nearest input vector is measured by an amount called a margin, and a separation plane that maximizes the margin is obtained.

  FIG. 7 and FIG. 8 are conceptual diagrams for explaining the difference in the mathematical expression of the learning method between the existing support vector machine and the multi-margin support vector machine used in the first embodiment. As shown in FIG. 7, in the existing support vector machine, if the set of input vectors can be linearly separated, the set of input vectors is formed by two hyperplanes indicated by dotted lines on both sides of the separation plane indicated by the solid line. It is completely separated and no input vector exists between the two hyperplanes. At this time, the distance (margin size) between the separation plane and these hyperplanes is 1 / || w ||. Further, the sum of the margins on both sides is 2 / || w || as shown by the double-headed arrow in FIG.

Then, (defining the separation surface to maximize the gap) of the maximum and the margin existing support vector machine is a problem of finding the parameters w and b, the parameters that minimizes the (1/2) || w || ² This is equivalent to the problem of obtaining w and b. Further, as a constraint condition, an input vector with a classification (−) is constrained to be below the separation plane, and an input vector with a classification (+) is constrained to be above the separation plane. This optimization problem is known as a quadratic programming problem in the field of mathematical programming, and can be solved by various known numerical calculation methods. However, an existing support vector machine can generate only one separation plane per support vector machine.

On the other hand, as shown in FIG. 8, the multi-margin support vector machine used in the first embodiment is a method of defining the separation surface by the same method as described above. It differs from existing support vector machines in that they are parallel to each other. In FIG. 8, two separation planes are generated as an example. In any of the two or more separation surfaces, the parameter w is the same, but the threshold value takes different values of b ₁ and b ₂ . In Figure 8, the separation surface of the lower, the threshold is b _1, the upper side of the separation surface, the threshold is b _2.

In this case, how to determine the value of w becomes a problem, but in the multi-margin support vector machine developed by the present inventor, w is determined by finding the separation plane that maximizes the narrowest gap. Yes. That is, the value of w is the same in any two or more separation planes, and the condition that no input vector exists in the margin area on both sides of any two or more separation planes is satisfied. | w || ² is minimized.

That is, the parameters w and b ₁ , b _2, which maximize the margin (specify two or more parallel separation surfaces that maximize the gap of the narrowest gap) in the multi-margin support vector machine developed by the present inventors _. b _M-1 the problem of finding the, (1/2) || w || parameters w and _b 1 ^{to 2} to _minimize, b _{2, b} be a problem equivalent to obtaining the _M-1. In addition, as a constraint condition, an input vector whose classification is (1) is below the separation plane 1, and an input vector whose classification is (2) is above the separation plane 1 and below the separation plane 2, and so on. Thus, the input vector of the separation surface (M) is constrained to be above the separation surface (M-1). This optimization problem is known as a quadratic programming problem in the field of mathematical programming, and can be solved by various known numerical calculation methods.

  Thus, since the input vector is converted into a d-dimensional vector and two or more parallel separation planes are generated in the separation unit, the classification prediction determination unit 112 shown in FIG. A creature corresponding to an unknown vector is generated based on a region where the unknown vector is arranged in an input space that is generated as a vector of dimensions and is separated into three or more regions by two or more separation planes. The classification of scientific information is predicted.

FIG. 9 is a conceptual diagram for explaining the difference in the mathematical expression of the learning method between the existing support vector machine and the multi-margin support vector machine used in the first embodiment. In an existing support vector machine, parameters are set so that (1/2) || w || ² is minimized. The constraint condition at this time is that an input vector with a classification (−) is g (x _i | w) −b ≦ −1, and an input vector with a classification (+) is g (x _i | w) −b. Restrained to be ≧ + 1. Note that g (x | w) = w ^T x.

On the other hand, in the multi-margin support vector machine developed by the present inventors, it sets parameters so as to minimize the (1/2) || w || ^2. The constraint condition at this time is that the input vector with the classification (1) is g (x _i | w) −b ₁ ≦ −1, and the input vector with the classification (2) is g (x _i | w) −. b ₁ ≧ + 1 and g (x _i | w) −b ₂ ≦ −1. Similarly, the input vector of the separation plane (M) is g (xi | w) −b _M−1 ≧ + 1. Restrained to be. Note that g (x | w) = w ^T x.

FIG. 10 is a conceptual diagram for explaining the problem that the optimum bias becomes indefinite in the multi-margin support vector machine. In the multi-margin support vector machine developed by the present inventor, optimal values may not be uniquely determined for the biases such as b ₁ and b ₂ described above. This is because the bias is uniquely determined on the separation plane with the smallest margin, but the bias can take a plurality of values on the separation plane with the smallest margin. For this reason, it is considered preferable to set a criterion for uniquely determining the bias based on some criterion.

FIG. 11 is a conceptual diagram for explaining that the bias is positioned at the center in the multi-margin support vector machine in order to solve the problem described in FIG. As a criterion for uniquely determining the bias, in the multi-margin support vector machine developed by the present inventor, the bias is set to be located at the center of the two hyperplanes on both sides in the separation plane where the margin is not minimum. That is, in the case of b ₁ in FIG. 11, the score function is a value when projected in a form perpendicular to the score function: g (x) for all input vectors, and the maximum score of classification (1) and classification 2 Set a value that is the center of the minimum score.

  The reason why the bias is positioned in the center will be described. The original purpose of the learning machine, including the multi-margin support vector machine developed by the present inventor, is to minimize the rate (misidentification rate) of erroneously identifying test samples (input vectors whose classification is unknown). is there. The reason why the misidentification rate decreases when the bias (threshold value) that defines the separation plane is placed in the center is omitted in detail, but from the viewpoint of the known VC theory, the expected loss is minimized. This is because, from an input vector with a known classification, only experience loss can be known, and instead there is no choice but to minimize experience loss. This method is a known method called experience loss minimization (ERM). That is, in order to minimize the experience loss in the above case, it is most reasonable to position the bias in the center. Therefore, in the multi-margin support vector machine developed by the present inventor, the bias is positioned in the center. I am going to do that.

  FIG. 12 is a functional block diagram showing details of the configuration of the support vector machine. The support vector machine 606 is provided with a parameter vector setting unit 702 for setting the parameter vector w described above. Since the calculation when the parameter vector setting unit 702 sets the parameter vector w has been described above, it will not be repeated.

In addition, the support vector machine 606 is provided with a bias setting unit 704 that sets the above-described biases b ₁ , b _{2, and} b _M−1 . Since the calculation when the bias setting unit 704 sets the bias b has been described above, it is not repeated.

  Further, the support vector machine 606 is provided with a soft margin unit 706 that extends the support vector machine by a soft margin method. When the support vector machine is expanded by the soft margin method, the configuration and operation of the soft margin unit 706 that sets the value of the slack function will be described in detail later with reference to FIGS.

  FIG. 13 is a conceptual diagram for explaining a case where linear separation is impossible by a support vector machine. The above-described existing support vector machine and the multi-margin support vector machine developed by the present inventor are discussions about a case where a set of input vectors can be linearly separated. In the actual problem of pattern recognition, as shown in FIG. In addition, neither a set of learning input vectors classified into two classes or a set of learning input vectors classified into M classes (M is an integer of 3 or more) may not be linearly separable.

  For this reason, when linear separation is impossible in this way, a parameter that satisfies the condition in the above-described equation of FIG. 9 has no solution. Therefore, further ingenuity is necessary to use the support vector machine for practical problems. Other than the method using the kernel trick described later, the first conceivable method is to relax the constraint so as to allow some identification errors. This is called “soft margin”.

FIG. 14 is a conceptual diagram for explaining a case where the support vector machine is relaxed by soft margin. In the soft margin method, the margin 1 / || w || (the sum of both sides is 2 / || w ||) is maximized, and as shown in FIG. 14, some input vectors are hyperplane H1 or hyperplane. Allow to enter the other side beyond H2. If the distance of how far into the opposite side is expressed as ξ _i / || w || using the parameter ξ _i (≧ 0), the sum is preferably as small as possible.

FIG. 15 is a conceptual diagram for explaining a learning method when the support vector machine is relaxed by soft margin. The soft margin method existing support vector machine, the problem of finding an optimum separation surface from the above conditions, to minimize the ^{(1/2) || w || 2 +} CΣ (i = 1 → N) ξ i Parameters To the problem of seeking. Note that Σ (i = 1 → N) ξ _i means obtaining the sum from ξ ₁ to ξ _N. The newly introduced parameter C is a constant that determines the balance between the size of the margin of the first term and the degree of protrusion of the second term.

The constraint condition at this time is that an input vector with a classification (−) is g (x _i | w) −b ≦ −1 + ξ _i , and an input vector with a classification (+) is g (x _i | w) −. It restrains so that it may become b> = + 1- _xi . Note that g (x | w) = w ^T x.

On the other hand, in the multi-margin support vector machine developed by the present inventors, the problem of finding an optimum separation surface from the above ^{conditions, (1/2) || w || 2} + C {Σ (m = 1 → M-1 ) Σ (i∈I _m ) ξ _i ⁻ + Σ (m = 2 → M) Σ (i∈I _m ) ξ _i ⁺ } is reduced to the problem of obtaining a parameter that minimizes. Note that Σ (iεI _m ) X _i means obtaining the sum of ξ _i for _i belonging to the subscript set of observation values belonging to class m. The newly introduced parameter C is a constant that determines the balance between the size of the margin of the first term and the degree of protrusion of the second term.

The constraint condition at this time is that an input vector with classification (1) is g (x _i | w) −b ₁ ≦ −1 + ξ _i ⁻ , and an input vector with classification (2) is g (x _i | w ) −b ₁ ≧ + 1−ξ _i ⁺ , and g (x _i | w) −b ₂ ≦ −1 + ξ _i ⁻ . Similarly, the input vector of the separation plane (M) is g (xi | w) -B _M-1 ≧ + 1−ξ _i ⁺ is constrained. Note that g (x | w) = w ^T x.

  FIG. 16 is a conceptual diagram for explaining a solution of the learning method when the support vector machine is relaxed by soft margin. The solution when the above-described existing support vector machine is relaxed and when the multi-margin support vector machine developed by the present inventor is relaxed is only one kind of optimization problem.

  Although the method for solving these optimization problems will not be described in detail, it can be basically solved by a known calculation method using the mathematical formula shown in FIG. 16 as in the case where linear separation is possible.

  FIG. 17 is a flowchart for explaining the operation of biological information processing apparatus 100 according to the present embodiment. First, when the operation of the biological information processing apparatus 100 starts, the known information acquisition unit 102 acquires known biological information classified into three or more classes from the outside (S102). Next, the acquired biological information whose classification is known is adjusted by extracting the number of dimensions of the feature amount based on a predetermined extraction criterion in the feature amount extraction unit 118 (S104). Then, an input vector is generated in the input space by the input vector generation unit 206 from the biological information from which the feature amount is extracted (S106).

  Thereafter, based on the set of generated input vectors and the classification corresponding to each of the input vectors, the separation unit 214 generates two or more parallel separation planes in the input space (S108). Based on the generated separation plane, the classification standard generation unit 106 sets a classification standard for classifying and predicting an input vector (unknown vector) whose classification is unknown (S110). The set classification standard is output by the output unit 116 (S112).

  Next, biological information whose classification is unknown is acquired from the outside by the unknown information acquisition unit 110 (S114). Next, the acquired biological information whose classification is unknown is adjusted by extracting the number of dimensions of the feature amount based on a predetermined extraction criterion in the feature amount extraction unit 118 (S116). Then, based on the extracted feature quantity, an unknown vector whose classification is unknown is generated (S118).

  Then, the unknown vector classification prediction is determined by applying the unknown vector to the set classification standard (S120). Thereafter, the determined classification prediction is output by the output unit 116 (S122), and the series of operations ends.

Hereinafter, the operational effects of the biological information processing apparatus 100 will be described.
According to the biological information processing apparatus 100, when the biological information is classified into three or more classes, the multi-margin support vector machine developed by the present inventor is used to separate two or more parallel separation surfaces. By separating the input space into three or more regions corresponding to three or more classes, high discrimination can be realized while avoiding “over-learning” or “over-fitting”.

  That is, the multi-margin support vector machine (MM-SVM) developed by the present inventors generates two or more parallel separation planes in the input space, thereby separating the input space into three or more regions. It is possible to reduce the risk of “over-learning” or “over-fit” that is likely to occur in existing support vector machines. It has been experimentally verified that the risk of such “overlearning” or “overfit” can be reduced.

  And in general, biological information has a lot of ordered data (drug resistance, morbidity, survival rate, etc.), so these various types of biological information can be integrated into doctors, dentists, nurses. Classifications that are specifically determined by clinical technologists, clinical contractor engineers, biology researchers, etc. according to the situation are also ordered data, such as low, medium, high, etc. Often. Therefore, if two or more parallel separation planes are generated as in the multi-margin support vector machine developed by the present inventor, the ordered data can be appropriately classified.

  In addition, since the multi-margin support vector machine developed by the present inventor has high discrimination performance by various devices, it improves the discriminability of prediction or classification criteria when classifying biological information into three or more classes. it can. Therefore, according to the biological information processing apparatus 100, it is possible to improve the validity of the classification criteria when classifying biological information into three or more classes. Therefore, it is possible to improve the validity of prediction when classifying biological information into three or more classes.

  In the multi-margin support vector machine developed by the present inventors, two or more separation planes are defined by two or more biases for a single score function, as shown in FIG. Therefore, the score of this single score function can play a role as a kind of evaluation index in the ordered medical observation data. In other words, even if the vectors belong to the same class, a kind of ordering in the ordered medical observation data is performed according to the score function value of the mapping when the projection is conceptually orthogonal to the score function. Can do.

  Furthermore, according to the biological information processing apparatus 100, the multi-margin support vector machine developed by the present inventor maximizes the margin width of the separation surface having the smallest margin width among two or more parallel separation surfaces. Therefore, the input vector or the feature vector is separated with a margin as much as possible, so that the identifiability and generalization of the prediction or classification criteria when the biological information is classified into three or more classes can be improved. .

  Further, according to the biological information processing apparatus 100, the multi-margin support vector machine developed by the present inventor arranges the separation plane so as to be located at the center of the margin width. Therefore, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

More specifically, according to the biological information processing apparatus 100, two or more parallel separation surfaces are expressed by the following equation: b = w ^T x (where x is an input vector or feature vector, and w Is a parameter vector, b is a bias, T is an arithmetic symbol indicating transposition), and w is a parameter vector that is the same in the two or more separation planes, Since b may be biases having different values from each other on the two or more separation surfaces, the two or more separation surfaces represented by b = w ^T x are parallel to each other and are “over-learning” or “excessive”. Since high discrimination can be achieved while avoiding “adaptation”, it is possible to improve the validity of prediction or classification criteria when classifying biological information into three or more classes.

  According to the biological information processing apparatus 100, since the support vector machine may be a support vector machine extended by the soft margin method, it is difficult to linearly separate an input vector or a feature vector. However, since it is possible to perform linear separation by the soft margin method, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

<Embodiment 2>
The present embodiment is basically a modification obtained by expanding the first embodiment by the kernel trick method, and has the same configuration as that of the first embodiment, unless otherwise specified. First, in order to understand the significance of this embodiment, an outline of the kernel trick method will be described below.

  When training samples (a set of input vectors) are learned using a learning machine such as a multi-margin support vector machine developed by the present inventor, linear separation is possible by using the soft margin method as described above. However, the parameter of the linear threshold element (separation surface) can be used even in the case of not.

  However, even if the soft margin method is used, a multi-margin support vector machine with good performance cannot always be constructed for an essentially nonlinear and complex set of input vectors. As a method for dealing with an essentially nonlinear set of input vectors, as shown in FIG. 20 to be described later, the input vectors are nonlinearly transformed to generate corresponding feature vectors, and linear identification is performed in the space. A method called “kernel trick” to do can be used. By using this method, the performance of the multi-margin support vector machine can be dramatically improved.

  In general, linear separability becomes more difficult as the number of samples increases, and conversely, if the dimension of the feature vector is larger than the number of samples of the input vector, linear separation is possible for any classification pattern. However, in order to make it possible to linearly separate a set of nonlinear and complex input vectors, it must be mapped to a dimension as large as the number of samples in the input vector, resulting in a huge amount of computation. End up.

  At this time, if the above calculation is performed using a function called a semi-fixed kernel function, while calculating in a high dimension, the calculation of the feature vector in the mapped space is actually avoided, and only the calculation of the kernel is performed. An optimal discriminant function can be constructed. This technique is called “kernel trick”.

  FIG. 19 is a functional block diagram illustrating an outline of the configuration of the classification reference generation unit 106 included in the biological information processing apparatus. The overall configuration of the biological information processing apparatus of the present embodiment is the same as that of the first embodiment, and thus the description thereof is omitted. In FIG. 19 and subsequent drawings, the same components as those of the first embodiment are denoted by the same reference numerals, and the description thereof is omitted as appropriate.

  The classification reference generation unit 106 includes a conversion unit 302 that converts an input vector to generate a feature vector. The conversion unit 302 acquires an input vector from the input vector storage unit 208 and performs nonlinear conversion to generate a feature vector. The generated feature vector is stored in the feature vector storage unit 304 and sent to the separation unit 214. In addition, a known classification corresponding to each feature vector is sent from the classification storage unit 212 to the separation unit 214.

  In other words, the conversion unit 302 converts a plurality of input vectors in the input space by nonlinear mapping, thereby converting a plurality of feature vectors respectively corresponding to the plurality of input vectors into a feature space having a higher order than the input space. Generate. Details of this nonlinear transformation will be described later with reference to FIGS.

  As shown in FIG. 20 to be described later, the separation unit 214 classifies the feature vectors using a learning machine such as a multi-margin support vector machine developed by the present inventor based on the acquired known classification feature vectors. Two or more parallel separation planes for predicting are generated in the feature space.

  The generated separation surface is stored in the separation surface storage unit 216, sent as it is to the classification reference output unit 218, output to the classification reference storage unit 108 outside the classification reference generation unit 106, and stored therein.

FIG. 20 is a conceptual diagram for explaining the operation of the kernel trick in the classification reference generation unit. FIG. 20 shows a kernel trick for a classification problem in a two-dimensional space. If a set of input vectors in the left two-dimensional input space is nonlinearly mapped using (x ₁ ² , 2 ^1/2 x ₁ x ₂ , x ₂ ² ) as features, the right three-dimensional features In a set of feature vectors in space, two classes can be identified on a linear plane. This corresponds to configuring an elliptical identification surface in the input space.

  FIG. 21 is a functional block diagram illustrating details of the configuration of the conversion unit and the separation unit. In the conversion unit 302, the input vector acquisition unit 502 acquires an input vector from the input vector storage unit 208 and sends it to the nonlinear conversion unit 504. The non-linear conversion unit 504 converts an input vector in the input space by a non-linear mapping, and converts it into a feature vector in a higher-dimensional feature space. The generated feature vector is output by the feature vector output unit 506 to the feature vector storage unit 208 outside the conversion unit 302 and stored.

  In the separation unit 214, the feature vector acquisition unit 702 acquires a feature vector from the feature vector storage unit 208 and sends it to the support vector machine 606. Also, the classification acquisition unit 604 acquires a known classification corresponding to the feature vector from the classification storage unit 212 and sends it to the support vector machine 606. The support vector machine 606 linearly separates a set of feature vectors in the feature space into three or more classes using the multi-margin support vector machine developed by the present inventors in the same manner as in the first embodiment. Generate two or more parallel separation planes. The generated separation surface is output and stored by the separation surface output unit 608 in the separation surface storage unit 216 outside the separation unit 214.

  FIG. 22 is a conceptual diagram showing an example of a kernel function used for extending the multi-margin support vector machine developed by the present inventors. The multi-margin support vector machine developed by the inventor, which has been extended nonlinearly using kernel tricks, automatically supports a small number of feature vectors near the separation plane based on the criterion of "margin maximization" on the separation plane with the smallest margin. Only the kernel (kernel feature) to be selected is selected, and the optimum discriminant function (separation surface) is constructed.

  At this time, as a kernel function suitable for expansion of the multi-margin support vector machine developed by the present inventors, a kernel function satisfying semi-stationary property can be used. Using a kernel function that satisfies the semi-fixed property, the calculation when the input vector is converted to the feature vector (performing the kernel trick) makes the calculation easier and the calculation result in the kernel This is because it is possible to correspond to the calculation result when the kernel is uniquely removed.

  As a kernel function satisfying the semi-fixed property, for example, as shown in FIG. 22, a linear kernel function, a polynomial kernel function, an RBF kernel function, and the like are preferably exemplified. The definition of semi-stationarity is shown in FIG. In principle, any kernel function may be used as long as the kernel function satisfies this definition.

  FIG. 23 is a conceptual diagram showing an example of a mathematical expression of a learning method using a kernel function used for expansion of a multi-margin support vector machine developed by the present inventors. Although detailed explanation of the calculation method is omitted, the learning algorithm using the kernel function, the score function using the kernel function, the matrix, the notation regarding the vector, the experience mapping, the subscript set, the kernel matrix, and the class label matrix shown in FIG. The multi-margin support vector machine developed by the present inventor can be extended by a kernel trick using a known calculation method using a sub-matrix formula of a class label matrix.

  FIG. 24 is a conceptual diagram showing an example of a mathematical expression used for a bias calculation method in the case of performing a soft margin in a learning method using a kernel function used for expansion of a multi-margin support vector machine developed by the present inventors. . In this way, the multi-margin support vector machine developed by the present inventor can be relaxed by further soft margin after being expanded by the kernel trick.

  The description about the soft margin is not repeated because it has been described in the first embodiment. At this time, a method for restoring slack variables ξ + and ξ− and calculating a bias can be executed by a person skilled in the art using a known calculation method by using the mathematical formula shown in FIG.

  FIG. 25 is a flowchart for explaining the operation of the biological information processing apparatus when the multi-margin support vector machine developed by the present inventor is extended by kernel tricks. First, when the operation of the biological information processing apparatus 100 starts, the known information acquisition unit 102 acquires known biological information classified into three or more classes from the outside (S202).

  Next, the acquired biological information whose classification is known is adjusted by extracting the number of dimensions of the feature amount based on a predetermined extraction criterion in the feature amount extraction unit 118 (S204). Then, an input vector is generated in the input space by the input vector generation unit 206 from biological information from which the feature quantity is extracted and whose classification is known (S206).

  Subsequently, the input vector in the generated input space is nonlinearly transformed by the conversion unit 302 to generate a feature vector in a higher-order feature space (S208). Thereafter, based on the generated set of feature vectors and the classification corresponding to each of the feature vectors, the separation unit 214 generates two or more parallel separation planes in the feature space (S210).

  Next, based on the formed separation plane, the classification standard generation unit 218 sets a classification standard for classifying and predicting an unknown vector, which will be described later (S212). The set classification standard is output by the classification standard output unit 218 (S214).

  Thereafter, the unknown information acquisition unit 110 acquires biological information whose classification is unknown from the outside (S216). Next, for the acquired biological information whose classification is unknown, the feature quantity extraction unit 118 extracts the dimension number of the feature quantity based on a predetermined extraction criterion (S218). Then, based on the extracted feature quantity, an unknown vector whose classification is unknown is generated (S220).

  Subsequently, the classification prediction determining unit 112 converts the input vector whose classification is unknown by nonlinear mapping, and generates a conversion unknown vector (S222). Then, the classification prediction determination unit 112 determines the classification prediction of the conversion unknown vector by applying the conversion unknown vector to the set classification criterion (S224). Thereafter, the determined classification prediction is output by the output unit 116 (S226), and the series of operations ends.

  In the following, a description will be given of the specific operational effects of the biological information processing apparatus 100 when the multi-margin support vector machine developed by the present inventor is extended by kernel tricks. The effects that are not particularly mentioned are the same as in the first embodiment.

  When the multi-margin support vector machine developed by the present inventor is extended by kernel tricks, when classifying biological information into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors, By generating in a feature space of higher order than the input space and separating the feature space into three or more regions respectively corresponding to three or more classes by two or more parallel separation planes, "Or" over-fitting "can be avoided while achieving high discrimination.

  At this time, the unknown vector is transformed by a non-linear mapping to generate a transformed unknown vector in the feature space, thereby avoiding “over-learning” or “over-fitting”, and by using a separation surface having high discrimination characteristics. The classification prediction of the transformed unknown vector can be determined in the space, and the validity of the prediction when the biological information is classified into three or more classes can be improved.

  At this time, when the multi-margin support vector machine developed by the present inventor is extended by kernel tricks, even if the input vector is an essentially nonlinear input vector set, the input vector is nonlinearly transformed, A method called “kernel trick” in which corresponding feature vectors are generated and linear identification is performed in the space can be used. By using this method, if the dimension of the feature vector is larger than the number of samples of the input vector, linear separation can be performed for any classification pattern. Therefore, the performance of the multi-margin support vector machine can be dramatically improved by using this method.

  In addition, when the multi-margin support vector machine developed by the present inventor is expanded by kernel tricks, the conversion unit calculates when the input vector is converted into a feature vector using a kernel function that satisfies semi-fixed property. Therefore, it is possible to construct an optimum discriminant function only by calculating a kernel while actually calculating feature vectors in the mapped space while mapping in a high dimension. Therefore, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

  Further, when the multi-margin support vector machine developed by the present inventor is extended by kernel tricks, any one or more kernel functions selected from the group consisting of a linear kernel function, a polynomial kernel function, and an RBF kernel function are all Since it is a kernel function that satisfies the semi-fixed property, it is possible to improve generalization of prediction or classification criteria when classifying biological information into three or more classes.

<Modification of Embodiment 2>
FIG. 26 is a functional block diagram illustrating details of the configuration of the classification reference generation unit used in the modification of the second embodiment. This modified example is basically the same as that of the second embodiment, and has the same configuration except for a case where particularly mentioned.

  In this modification, unlike the case of the second embodiment, the separation surface in the feature space generated by the multi-margin support vector machine developed by the present inventor is stored in the separation surface storage unit 216 by the separation unit 214. , And sent to the inverse transform unit 402.

  Then, in the inverse transform unit 402, two or more parallel linear separation surfaces in the feature space are inversely transformed into two or more nonlinear separation surfaces in the input space by the inverse transformation of the transformation by the above-described nonlinear mapping. The That is, two or more separation surfaces that separate the input space into three or more regions are generated in the input space by inversely transforming the two or more separation surfaces generated by the separation unit 214 using the above-described nonlinear mapping.

  The reversely converted separation plane is stored in the reverse converted separation plane storage unit 408, sent to the classification reference output unit 218, and output to the classification reference storage unit 108 outside the classification reference generation unit 106 for storage. Is done.

  FIG. 27 is a flowchart for explaining the operation of the biological information processing apparatus according to the modification of the second embodiment. In this modified example, first, when the operation of the biological information processing apparatus 100 is started, the known information that has been classified into three or more classes is acquired from the outside by the known information acquisition unit 102 (S302).

  Next, the acquired biological information whose classification is known is adjusted by extracting the dimensionality of the feature amount based on a predetermined extraction criterion in the feature amount extraction unit 118 (S304). Then, an input vector is generated in the input space by the input vector generation unit 206 from biological information from which the feature quantity is extracted and whose classification is known (S306).

  Subsequently, the input vector in the generated input space is nonlinearly transformed by the conversion unit 302 to generate a feature vector in a higher-order feature space (S308). Thereafter, based on the generated set of feature vectors and the classification corresponding to each of the feature vectors, the separation unit 214 generates two or more parallel separation planes in the feature space (S310).

  Thereafter, the generated separation plane in the feature space is inversely transformed using the above-described nonlinear mapping by the inverse transformation unit 402 to generate an inversely transformed separation plane in the input space (S312). Then, based on the generated reverse-transformed separation plane, the classification standard generation unit 106 sets a classification standard for classifying and predicting an unknown vector to be described later (S314). Next, the generated classification standard is output by the output unit 116 (S316).

  Furthermore, biological information whose classification is unknown is acquired from the outside by the unknown information acquisition unit 110 (S318). Next, the acquired biological information whose classification is unknown is extracted by the feature amount extraction unit 118 based on a predetermined extraction criterion (S320). Then, based on the extracted feature quantity, an unknown vector whose classification is unknown is generated (S322).

  Subsequently, the classification prediction determination unit 112 determines the classification prediction of the unknown vector by applying the unknown vector to the classification criterion set (S324). Thereafter, the determined classification prediction is output by the output unit 116 (S326), and the series of operations ends.

  Hereinafter, the specific operation and effect of the biological information processing apparatus 100 in this modification will be described. The effects that are not particularly mentioned are the same as those in the second embodiment.

  When the multi-margin support vector machine developed by the present inventor is extended by kernel tricks, when classifying biological information into three or more classes, a plurality of feature vectors respectively corresponding to a plurality of input vectors, By generating in a feature space of higher order than the input space and separating the feature space into three or more regions respectively corresponding to three or more classes by two or more parallel separation planes, "Or" over-fitting "can be avoided while achieving high discrimination.

  Then, the separation plane in the feature space thus obtained is inversely transformed by the above-described nonlinear mapping, whereby a nonlinear separation plane can be generated also in the input space. For this reason, it is possible to perform classification prediction determination with high distinctiveness while avoiding “overlearning” or “overfitting” of unknown vectors in the input space, and when biological information is classified into three or more classes. Can improve the validity of predictions.

<Summary>
FIG. 28 is a conceptual diagram for explaining the outline of the classification prediction system for ordered medical observation data using the biological information processing apparatus according to the embodiment described so far. As shown in FIG. 28, the biological information processing apparatus according to the first and second embodiments can be suitably used as an ordered medical observation data classification prediction system.

  For example, for input data (known), for each of patients A, B, etc., doctors, nurses, clinical technologists, etc. provide data such as gene expression data, blood test data, interview data, X-ray data, etc. Create by measuring and putting together. The doctor examines the input data and classifies the order of the ordered medical observation data (such as drug resistance, morbidity, and survival rate) into one of three stages of low, medium, and high.

  The classification known input data thus obtained is computer-learned by a multi-margin support vector machine, which is a new technique developed by the present inventor, to generate a separation criterion. The obtained classification standard is estimated in advance by a cross-validation.

  When predicting the disease of a new patient X, doctors, nurses, clinical technologists, etc. of the new patient X obtain data such as gene expression data, blood test data, interview data, X-ray data, etc. Similarly, input data with unknown classification is created by collecting and measuring. Then, by applying the input data to the above classification criteria, the order of the ordered medical observation data (drug resistance, disease level, survival rate, etc.) of the new patient X can be any of three levels, low, medium, and high. Step by step.

  According to this system, when the biological information is classified into three or more classes, the multi-margin support vector machine developed by the present inventor is used. Therefore, the input space is divided by two or more parallel separation surfaces. By separating into three or more regions corresponding to three or more classes, high discrimination can be realized while avoiding “overlearning” or “overfitting”.

  And in general, biological information has a lot of ordered data (drug resistance, morbidity, survival rate, etc.), so these various types of biological information can be integrated into doctors, dentists, nurses. Classifications that are specifically determined by clinical technologists, clinical contractor engineers, biology researchers, etc. according to the situation are also ordered data, such as low, medium, high, etc. Often. Therefore, if two or more parallel separation planes are generated as in the multi-margin support vector machine developed by the present inventor, ordered data can be appropriately classified.

  Therefore, according to this system, the order of the ordered medical observation data (drug resistance, morbidity, survival rate, etc.) of the new patient X is determined in one of three stages of low, medium, and high. In this case, it is possible to perform a highly reliable stage determination by suppressing “over-learning” or “over-fitting”.

  As mentioned above, although embodiment of this invention was described with reference to drawings, these are the illustrations of this invention, Various structures other than the above are also employable.

  FIG. 29 is a diagram illustrating an example of a prediction rate by cross-validation for confirming the reliability of the biological information processing apparatus according to the embodiment. In this example, antibiotic resistance prediction was performed on cultured cells derived from 44 human cancers using the biological information processing apparatus 100 according to the second embodiment. As the input data, as shown in the table of FIG. 25, input data that has already been classified into three classes of low, medium, and high is evaluated by the researcher on the anticancer drug resistance experimental results for each cell. Using. The features included in the input data were gene expression data on thousands of different genomes.

  Of these 44 pieces of input data, 43 pieces of input data are extracted in accordance with predetermined criteria (feature parameters by t-test and feature parameters by correlation coefficient), and the top few feature amounts are extracted. An input vector was generated from the input data. Then, the input vector is converted into a feature vector by nonlinear mapping, the multi-margin support vector machine developed by the present inventor is extended by the RBF kernel function, and further relaxed by the soft margin method. The calculation was performed. As a result, two parallel separation surfaces were obtained that separated the feature space into three. A separation standard was set from the obtained separation surface. Then, the remaining one input data among these 44 input data was applied to the separation criterion, and the obtained classification prediction determination result was examined.

  Thus, in order to measure the prediction rate of the separation criterion, 43 cells out of 44 cells were learned and 1 cell was predicted 44 times. Then, the classification prediction determination results using the feature parameter by the t test and the feature parameter by the correlation coefficient as input vectors are summarized as a table in FIG. As shown in FIG. 29, the prediction rate with the feature amount by the t-test method was 93.2% (41/44). Moreover, the prediction rate with the feature-value by the Pearson correlation coefficient method using an actual tolerance amount was 93.22% (41/44).

  As a result of the cross-validation described above, anticancer drug resistance prediction was performed on 44 human cancer-derived cultured cells using the biological information processing apparatus 100 according to the second embodiment. It was experimentally revealed that the classification criteria are highly valid and the classification prediction results using this classification criteria are also highly valid.

  In the above, this invention was demonstrated based on the Example. It is to be understood by those skilled in the art that this embodiment is merely an example, and that various modifications are possible and that such modifications are within the scope of the present invention.

  As described above, the biological information processing apparatus according to the present invention separates an input space or a feature space into three or more regions respectively corresponding to three or more classes by two or more parallel separation surfaces. The biological information processing apparatus, the biological information processing method, the biological information processing program, and the like have the effect of improving the validity of prediction or classification criteria when classifying scientific information into three or more classes. Useful as.

1 is a functional block diagram illustrating an outline of a configuration of a biological information processing apparatus according to a first embodiment. FIG. 3 is a conceptual diagram for explaining the operation of the biological information processing apparatus according to the first embodiment. It is a conceptual diagram for demonstrating the difference in the function of the existing support vector machine and the multi-margin support vector machine used for Embodiment 1. 3 is a functional block diagram illustrating details of a configuration of a classification reference generation unit used in Embodiment 1. FIG. FIG. 3 is a functional block diagram illustrating details of a configuration of a separation unit used in the first embodiment. It is a conceptual diagram for demonstrating the difference in the learning method of the existing support vector machine and the multimargin support vector machine used for Embodiment 1. FIG. It is a conceptual diagram for demonstrating the difference of the numerical formula of the learning method of the existing support vector machine and the multimargin support vector machine used for Embodiment 1. FIG. It is a conceptual diagram for demonstrating the difference of the numerical formula of the learning method of the existing support vector machine and the multimargin support vector machine used for Embodiment 1. FIG. It is a conceptual diagram for demonstrating the difference of the numerical formula of the learning method of the existing support vector machine and the multimargin support vector machine used for Embodiment 1. FIG. FIG. 9 is a conceptual diagram for explaining a problem that an optimum bias becomes indefinite in the multi-margin support vector machine used in the first embodiment. FIG. 6 is a conceptual diagram for explaining that a bias is positioned at the center in the multi-margin support vector machine used in the first embodiment. 3 is a functional block diagram showing details of the configuration of a support vector machine used in Embodiment 1. FIG. It is a conceptual diagram for demonstrating the case where linear separation is impossible with the support vector machine used for Embodiment 1. FIG. It is a conceptual diagram for demonstrating the case where the support vector machine used for Embodiment 1 relaxes by soft margin-ization. It is a conceptual diagram for demonstrating the learning method in the case of relaxing the support vector machine used for Embodiment 1 by soft margin-ization. It is a conceptual diagram for demonstrating the solution of the learning method in the case of relaxing the support vector machine used for Embodiment 1 by soft margining. 3 is a flowchart for explaining an operation of the biological information processing apparatus according to the first embodiment. FIG. 3 is a conceptual diagram for explaining the operational effect of the biological information processing apparatus according to the first embodiment. 10 is a functional block diagram showing details of a configuration of a classification reference generation unit used in Embodiment 2. FIG. FIG. 10 is a conceptual diagram for explaining a kernel trick operation in a classification reference generation unit used in the second embodiment. 10 is a functional block diagram illustrating details of configurations of a conversion unit and a separation unit used in Embodiment 2. FIG. FIG. 10 is a conceptual diagram illustrating an example of a kernel function used in the second embodiment. 6 is a conceptual diagram illustrating an example of a mathematical expression of a learning method using a kernel function used in Embodiment 2. FIG. FIG. 10 is a conceptual diagram illustrating an example of a mathematical expression used in a bias calculation method when soft margining is performed in a learning method using a kernel function used in the second embodiment. 10 is a flowchart for explaining an operation of the biological information processing apparatus according to the second embodiment. FIG. 10 is a functional block diagram illustrating details of a configuration of a classification reference generation unit used in a modification of the second embodiment. 10 is a flowchart for explaining an operation of the biological information processing apparatus according to the modification of the second embodiment. It is a conceptual diagram for demonstrating the outline | summary of the classification | category prediction system of the ordered medical observation data using the biological information processing apparatus which concerns on Embodiment 1 and 2. FIG. It is a figure which shows the example of the prediction rate by the cross-validation for confirming the reliability of the biological information processing apparatus which concerns on an Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Biological information processing apparatus 102 Known information acquisition part 104 Known information storage part 106 Classification reference generation part 108 Classification reference storage part 110 Unknown information acquisition part 112 Classification prediction determination part 114 Classification prediction storage part 116 Output part 118 Feature-value extraction part 120 Cross validation unit 122 Estimated prediction rate storage unit 202 Biological information acquisition unit 204 Biological information storage unit 206 Input vector generation unit 208 Input vector storage unit 210 Classification acquisition unit 212 Classification storage unit 214 Separation unit 216 Separation plane storage unit
218 Classification reference output unit 302 Conversion unit 304 Feature vector storage unit 402 Inverse conversion unit 408 Inversely converted separation plane storage unit 502 Input vector acquisition unit 504 Nonlinear conversion unit 602 Input vector acquisition unit 604 Classification acquisition unit 606 Support vector machine 608 Separation plane Output unit 702 Parameter vector setting unit 704 Bias setting unit 706 Soft margin unit

Claims (32)

A biological information processing apparatus for predicting the classification of biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit that generates a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
An unknown information acquisition unit for acquiring biological information whose classification is unknown;
An unknown vector generation unit that generates an unknown vector including the biological information acquired by the unknown information acquisition unit in the input space separated into the three or more regions by the two or more separation surfaces;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A prediction determination unit for determining;
A prediction classification output unit that outputs a classification of the biological information corresponding to the unknown vector predicted and determined by the prediction determination unit;
A biological information processing apparatus comprising: A biological information processing apparatus for predicting the classification of biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit for generating a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Conversion that generates a plurality of feature vectors corresponding to the plurality of input vectors in a feature space of a higher order than the input space by converting the plurality of input vectors in the input space by nonlinear mapping. And
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
An unknown information acquisition unit for acquiring biological information whose classification is unknown;
By converting the unknown vector including the biological information acquired by the unknown information acquisition unit by the nonlinear mapping, the converted unknown vector corresponding to the unknown vector is converted into the three or more separation surfaces by the two or more separation planes. An unknown vector generation unit that generates the feature space separated into regions;
Classification of the biological information corresponding to the converted unknown vector based on a region where the converted unknown vector is arranged in the feature space separated into the three or more regions by the two or more separation planes. A prediction determination unit for predicting and determining
A predicted classification output unit that outputs a classification of the biological information corresponding to the transformed unknown vector predicted and determined by the prediction determination unit;
A biological information processing apparatus comprising: A biological information processing apparatus for predicting the classification of biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit for generating a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Conversion that generates a plurality of feature vectors corresponding to the plurality of input vectors in a feature space of a higher order than the input space by converting the plurality of input vectors in the input space by nonlinear mapping. And
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
Inverse transformation that generates two or more separation surfaces in the input space by separating the input space into three or more regions by inversely transforming the two or more separation surfaces generated by the separation unit by the nonlinear mapping. And
An unknown information acquisition unit for acquiring biological information whose classification is unknown;
An unknown vector generation unit that generates an unknown vector including the biological information acquired by the unknown information acquisition unit in the input space separated into the three or more regions by the two or more separation surfaces;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A prediction determination unit for determining;
A prediction classification output unit that outputs a classification of the biological information corresponding to the unknown vector predicted and determined by the prediction determination unit;
A biological information processing apparatus comprising: The biological information processing apparatus according to claim 2 or 3,
The biological information processing apparatus, wherein the conversion unit is configured to perform a calculation when the input vector is converted into the feature vector using a kernel function satisfying semi-fixed property. The biological information processing apparatus according to claim 4,
The biological information processing apparatus, wherein the kernel function is one or more kernel functions selected from the group consisting of a linear kernel function, a polynomial kernel function, and an RBF kernel function. The biological information processing apparatus according to any one of claims 1 to 5,
The biological information processing apparatus, wherein the linear separation unit is configured to generate the two or more parallel separation planes using a support vector machine. The biological information processing apparatus according to claim 6,
The biological information processing characterized in that the support vector machine is configured to maximize the margin width of a separation surface having a minimum margin width among the two or more parallel separation surfaces. apparatus. The biological information processing apparatus according to claim 6 or 7,
The biological information processing apparatus, wherein the separation unit is configured to arrange the separation surface so as to be positioned at the center of the margin width. The biological information processing apparatus according to any one of claims 6 to 8,
The biological information processing apparatus, wherein the support vector machine is a support vector machine extended by a soft margin method. The biological information processing apparatus according to any one of claims 1 to 9,
The two or more parallel separation surfaces have the following formula: b = w ^T x
(Where x is an input vector or feature vector, w is a parameter vector, b is a bias, and T is an arithmetic symbol indicating transposition)
Is defined by the score function represented by
The w is a parameter vector that is the same in the two or more separation planes;
The biological information processing apparatus, wherein b is a bias having different values on the two or more separation planes. A biological information processing device for generating classification criteria for classifying biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit that generates a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
A classification criterion output unit that outputs the classification criterion including information defining the two or more separation planes generated by the separation unit;
A biological information processing apparatus comprising: A biological information processing device for generating classification criteria for classifying biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit that generates a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Conversion that generates a plurality of feature vectors corresponding to the plurality of input vectors in a feature space of a higher order than the input space by converting the plurality of input vectors in the input space by nonlinear mapping. And
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
A classification criterion output unit that outputs the classification criterion including information defining the two or more separation planes generated by the separation unit;
A biological information processing apparatus comprising: A biological information processing device for generating classification criteria for classifying biological information,
A known information acquisition unit for acquiring a plurality of known biological information classified into three or more classes;
An input vector generation unit that generates a plurality of input vectors including the plurality of biological information acquired by the known information acquisition unit in an input space;
Conversion that generates a plurality of feature vectors corresponding to the plurality of input vectors in a feature space of a higher order than the input space by converting the plurality of input vectors in the input space by nonlinear mapping. And
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is A separation unit that separates into three or more regions respectively corresponding to the three or more classes;
Inverse transformation that generates two or more separation surfaces in the input space by separating the input space into three or more regions by inversely transforming the two or more separation surfaces generated by the separation unit by the nonlinear mapping. And
A classification criterion output unit that outputs the classification criterion including information defining the two or more separation planes generated by the inverse transform unit;
A biological information processing apparatus comprising: The biological information processing apparatus according to claim 12 or 13,
The biological information processing apparatus, wherein the conversion unit is configured to perform a calculation when the input vector is converted into the feature vector using a kernel function satisfying semi-fixed property. The biological information processing apparatus according to claim 14, wherein
The biological information processing apparatus, wherein the kernel function is one or more kernel functions selected from the group consisting of a linear kernel function, a polynomial kernel function, and an RBF kernel function. The biological information processing apparatus according to any one of claims 11 to 15,
The biological information processing apparatus, wherein the linear separation unit is configured to generate the two or more parallel separation planes using a support vector machine. The biological information processing apparatus according to claim 16, wherein
The biological information processing characterized in that the support vector machine is configured to maximize the margin width of a separation surface having a minimum margin width among the two or more parallel separation surfaces. apparatus. The biological information processing apparatus according to claim 16 or 17,
The biological information processing apparatus, wherein the separation unit arranges the separation surface so as to be positioned at the center of the margin width. The biological information processing apparatus according to any one of claims 16 to 18,
The biological information processing apparatus, wherein the support vector machine is a support vector machine extended by a soft margin method. The biological information processing apparatus according to any one of claims 11 to 19,
The two or more parallel separation surfaces have the following formula: b = w ^T x
(Where x is an input vector or feature vector, w is a parameter vector, b is a bias, and T is an arithmetic symbol indicating transposition)
Is defined by the score function represented by
The w is a parameter vector that is the same in the two or more separation planes;
The biological information processing apparatus, wherein b is a bias having different values on the two or more separation planes. A biological information processing method for predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating in the input space a plurality of input vectors including the plurality of biological information obtained by obtaining the biological information;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is Separating into three or more regions respectively corresponding to the three or more classes;
Obtaining biological information with unknown classification;
An unknown vector including the biological information obtained by obtaining biological information whose classification is unknown is input into the input space separated into the three or more regions by the two or more separation surfaces. Generating step;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A determining step;
Outputting the biological information classification corresponding to the unknown vector predicted and determined by predicting and determining the biological information classification;
A biological information processing method comprising: A biological information processing method for predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Obtaining biological information with unknown classification;
By converting the unknown vector including the biological information acquired by the step of acquiring the biological information whose classification is unknown, the transformed unknown vector corresponding to the unknown vector is converted into the 2 Generating in the feature space separated into the three or more regions by the separation surface;
Classification of the biological information corresponding to the converted unknown vector based on a region where the converted unknown vector is arranged in the feature space separated into the three or more regions by the two or more separation planes. Predicting and determining
Outputting the biological information classification corresponding to the transformed unknown vector predicted and determined by predicting and determining the biological information classification;
A biological information processing method comprising: A biological information processing method for predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Step of generating two or more separation surfaces in the input space for separating the input space into three or more regions by inversely transforming the two or more separation surfaces generated by the separating step by the nonlinear mapping. When,
Obtaining biological information with unknown classification;
An unknown vector including the biological information obtained by obtaining biological information whose classification is unknown is input into the input space separated into the three or more regions by the two or more separation surfaces. Generating step;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A determining step;
Outputting the biological information classification corresponding to the unknown vector predicted and determined by predicting and determining the biological information classification;
A biological information processing method comprising: A biological information processing method for generating a classification criterion for classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is Separating into three or more regions respectively corresponding to the three or more classes;
Outputting the classification criteria including information defining the two or more separation surfaces generated by the separation unit;
A biological information processing method including: A biological information processing method for generating a classification criterion for classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Outputting the classification criteria including information defining the two or more separation surfaces generated by the separating step;
A biological information processing method including: A biological information processing method for generating a classification criterion for classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more separation planes are generated in the feature space, and the feature space is the three or more feature spaces. Separating into three or more regions each corresponding to a class of
Two or more parallel separation surfaces that separate the input space into three or more regions are transformed into the input space by inversely transforming the two or more separation surfaces generated by the separating step by the nonlinear mapping. Generating step;
Outputting the classification criteria including information defining the two or more separation surfaces generated by the step of generating separation surfaces in an input space by the inverse transformation;
A biological information processing method including: A biological information processing program for causing a computer to execute a process of predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is Separating into three or more regions respectively corresponding to the three or more classes;
Obtaining biological information with unknown classification;
An unknown vector including the biological information obtained by obtaining biological information whose classification is unknown is input into the input space separated into the three or more regions by the two or more separation surfaces. Generating step;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A determining step;
Outputting a classification of the biological information corresponding to the unknown vector predicted and determined by the predicting and determining step;
A biological information processing program characterized by causing a computer to execute. A biological information processing program for causing a computer to execute a process of predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Obtaining biological information with unknown classification;
By converting the unknown vector including the biological information acquired by the step of acquiring the biological information whose classification is unknown, the transformed unknown vector corresponding to the unknown vector is converted into the 2 Generating in the feature space separated into the three or more regions by the separation surface;
Classification of the biological information corresponding to the converted unknown vector based on a region where the converted unknown vector is arranged in the feature space separated into the three or more regions by the two or more separation planes. Predicting and determining
Outputting a classification of the biological information corresponding to the transformed unknown vector predicted and determined by the predicting and determining step;
A biological information processing program characterized by causing a computer to execute. A biological information processing program for causing a computer to execute a process of predicting a classification of biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Step of generating two or more separation surfaces in the input space for separating the input space into three or more regions by inversely transforming the two or more separation surfaces generated by the separating step by the nonlinear mapping. When,
Obtaining biological information with unknown classification;
An unknown vector including the biological information obtained by obtaining biological information whose classification is unknown is input into the input space separated into the three or more regions by the two or more separation surfaces. Generating step;
A classification of the biological information corresponding to the unknown vector is predicted based on a region where the unknown vector is arranged in an input space separated into the three or more regions by the two or more separation planes. A determining step;
Outputting a classification of the biological information corresponding to the unknown vector predicted and determined by the predicting and determining step;
A biological information processing program characterized by causing a computer to execute. A biological information processing program for generating a classification criterion for causing a computer to execute a process of classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Based on the plurality of input vectors and the classification of the plurality of biological information respectively corresponding to the plurality of input vectors, two or more parallel separation planes are generated in the input space, and the input space is Separating into three or more regions respectively corresponding to the three or more classes;
Outputting the classification criteria including information defining the two or more separation surfaces generated by the separating step;
Biological information processing program that causes a computer to execute. A biological information processing program for generating a classification criterion for causing a computer to execute a process of classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Outputting the classification criteria including information defining the two or more separation surfaces generated by the separation unit;
Biological information processing program that causes a computer to execute. A biological information processing program for generating a classification criterion for causing a computer to execute a process of classifying biological information,
Obtaining a plurality of known biological information classified into three or more classes;
Generating, in an input space, a plurality of input vectors including the plurality of biological information obtained by obtaining the known biological information;
Generating a plurality of feature vectors respectively corresponding to the plurality of input vectors in a feature space of higher order than the input space by transforming the plurality of input vectors in the input space by nonlinear mapping; When,
Based on the plurality of feature vectors and the classification of the plurality of biological information respectively corresponding to the plurality of feature vectors, two or more parallel separation planes are generated in the feature space, and the feature space is Separating into three or more regions respectively corresponding to the three or more classes;
Step of generating two or more separation surfaces in the input space for separating the input space into three or more regions by inversely transforming the two or more separation surfaces generated by the separating step by the nonlinear mapping. When,
Outputting the classification criteria including information defining the two or more separation surfaces generated by the step of generating separation surfaces in the input space by the inverse transformation;
Biological information processing program that causes a computer to execute.

Images