Patent Application
20210140938
The present invention relates to an identification method for identifying nonconforming data included in measurement data obtained from a measurement system, a classification analysis method performing the classification analysis using the data from which the nonconforming data are removed, an identification device, a classification analysis device and a storage medium.
For example, as described in Non-Patent Document 1, devices for measuring minute and trace objects have been developed one after another in the field of advanced sensing device development such as nano-sensing, minute measurement, and quantum measurement.
However, many of the above-mentioned advanced sensing devices output only partial information of the target because the measurement system and measurement target are minute, and the output is often affected by thermal noise, quantum noise, etc. For this reason, when the noise signal level is larger than the target signal, that is, the SN ratio is often very bad, the measurement accuracy is too low at the primary measurement stage, which is not suitable for practical use. Further, under such a measurement situation, there is no room for adopting a general noise removal method of removing a noise component by signal strength on the assumption that a small signal is noise and a large signal represents a target.
In addition, it is difficult to apply various kinds of noise filtering using the knowledge about the object and the specific properties of the problem, when the knowledge and properties are not clear. In particular, in the next generation DNA sequencer using single molecule measurement technology, the influence of noise becomes a serious problem because there are many unknown parts in the properties of target molecules and measurement systems and the noise signal is large.
An object of the present invention is to provide an identification method which can identify appropriately nonconforming data from a measurement data set, for example, can contribute to improve the reliability of measurement results by an advanced sensing device, a classification analysis method which can perform the classification analysis with high accuracy for measurement data, an identification device, a classification analysis device, a storage medium for identification and a storage medium for classification analysis.
In view of the above problems, the present invention focuses on a machine learning method for learning a classifier from a positive example set and an unknown set. For example, there is used a classifier constituted by a PU classification method suitable for binary classification of positive example/negative example (Classification of Positive and Unlabeled Examples). Therefore, this is an invention made based on the knowledge that non-conforming data can be identified with high accuracy from a measurement pattern. a classifier constituted by. Details of the PU classification method are described in Non-Patent Document 3.
The first form of the present invention is the identification method comprising the steps of
introducing a sample containing an analyte into a measurement space, obtaining pulse signal data detected due to said introduction, and
identifying nonconforming data detected by elements other than said analyte from said pulse signal data by execution of a computer control program, wherein
said computer control program includes an identification analysis program using the machine learning to learn a classifier that classifies positive and negative examples from positive example data of a positive example set and unknown data of an unknown set in which either positive or negative example is unknown,
when type 1 data of a pulse signal are obtained under first measurement condition measured by introducing a sample not containing an analyte in said measurement space and type 2 data of a pulse signal are obtained under second measurement condition measured by introducing a sample containing an analyte in said measurement space, a storage means is included for storing said type 1 data and said type 2 data, and
said nonconforming data included in said type 2 data is identified by executing said identification analysis program, through using said type 1 data as said positive example data and said type 2 data as said unknown data.
The second form of the present invention is the classification analysis method comprising the steps of
introducing a sample containing an analyte into a measurement space, obtaining pulse signal data detected due to said introduction,
obtaining analyzed data through removing said nonconforming data detected by elements other than said analyte from said pulse signal data, and
performing a classification analysis of said analyzed data by execution of a computer control program,
wherein
a nonconforming data storage means is included for storing said nonconforming data identified by the identification method according to claim 1,
said computer control program includes a classification analysis program that performs said classification analysis using the machine learning,
a feature value is obtained in advance which indicates a feature of waveform form of said pulse signal,
said feature value obtained in advance is set as the learning data for said machine learning,
said feature value obtained from said pulse signal of said analyzed data removed said nonconforming data is set as a variable, and
said classification analysis on said analyte is performed by executing said classification analysis program.
The third form of the present invention is the classification analysis method according to the second form, wherein said feature value is one or more selected from a group of
a wave height value of the waveform in a predetermined time width,
a pulse wavelength ta,
a peak position ratio represented by ratio tb/ta of time ta and tb leading from the pulse start to the pulse peak,
a kurtosis which represents the sharpness of the waveform,
a depression representing the slope leading from the pulse start to the pulse peak,
an area representing total sum of the time division area dividing the waveform with the predetermined times,
an area ratio of sum of the time division area leading from the pulse start to the pulse peak to the total waveform area,
a time inertia moment determined by mass and rotational radius when the mass is constructive to said time division area centered at the pulse start time and the rotational radius is constructive to time leading from said center to said time division area,
a normalized time inertia moment determined when said time inertia moment is normalized so as that the wave height becomes a reference value,
a mean value vector whose vector component is the mean value of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak,
a normalized mean value vector which is normalized so as that the wavelength becomes a standard value for said mean value vector,
a wave width mean value inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to mean value difference vector whose vector component is mean value difference of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak and the rotational center is constructive to time axis of waveform foot,
a normalized wave width mean value inertia moment determined when said wave width mean value inertia moment is normalized so as that the wavelength becomes a standard value,
a wave width dispersion inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to dispersion vector whose vector component is dispersion in which the wave form is equally divided in the wave height direction and the dispersion is calculated from time value for each division unit and the rotational center is constructive to time axis of waveform foot, and
a normalized wave width dispersion inertia moment determined when said wave width dispersion inertia moment is normalized so as that the wavelength becomes a standard value.
The fourth form of the present invention is the identification device comprising
a means for introducing a sample containing an analyte into a measurement space,
a means for obtaining pulse signal data detected due to said introduction, and
a means for identifying nonconforming data detected by elements other than said analyte from said pulse signal data by execution of a computer control program, wherein
said computer control program includes an identification analysis program using a machine learning to learn a classifier that classifies positive and negative examples from positive example data of a positive example set and unknown data of an unknown set in which either positive or negative example is unknown, when type 1 data of a pulse signal are obtained under first measurement condition
measured by introducing a sample not containing an analyte in said measurement space and type 2 data of a pulse signal are obtained under second measurement condition measured by introducing a sample containing an analyte in said measurement space, an storage means is included for storing said type 1 data and said type 2 data, and
said nonconforming data included in said type 2 data are identified by executing said identification analysis program, by using said type 1 data as said positive example data and said type 2 data as said unknown data.
The fifth form of the present invention is the classification analysis device comprising
a means for introducing a sample containing an analyte into a measurement space,
a means for obtaining pulse signal data detected due to said introduction,
a means for obtaining analyzed data through removing said nonconforming data detected by elements other than said analyte from said pulse signal data, and
a means for performing a classification analysis of said analyzed data by execution of a computer control program,
wherein
a nonconforming data storage means is included for storing said nonconforming data identified by the identification method according to claim 4,
said computer control program includes a classification analysis program that performs said classification analysis using the machine learning,
a feature value is obtained in advance which indicates a feature of waveform form of said pulse signal,
said feature value obtained in advance is set as the learning data for said machine learning,
said feature value obtained from said pulse signal of said analyzed data removed said nonconforming data is set as a variable, and
said classification analysis on said analyte is performed by executing said classification analysis program.
The sixth form of the present invention is the classification analysis device according to the fifth form, wherein said feature value is one or more selected from a group of
a wave height value of the waveform in a predetermined time width,
a pulse wavelength ta,
a peak position ratio represented by ratio tb/ta of time ta and tb leading from the pulse start to the pulse peak,
a kurtosis which represents the sharpness of the waveform,
a depression representing the slope leading from the pulse start to the pulse peak,
an area representing total sum of the time division area dividing the waveform with the predetermined times,
an area ratio of sum of the time division area leading from the pulse start to the pulse peak to the total waveform area,
a time inertia moment determined by mass and rotational radius when the mass is constructive to said time division area centered at the pulse start time and the rotational radius is constructive to time leading from said center to said time division area,
a normalized time inertia moment determined when said time inertia moment is normalized so as that the wave height becomes a reference value,
a mean value vector whose vector component is the mean value of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak,
a normalized mean value vector which is normalized so as that the wavelength becomes a standard value for said mean value vector,
a wave width mean value inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to mean value difference vector whose vector component is mean value difference of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak and the rotational center is constructive to time axis of waveform foot,
a normalized wave width mean value inertia moment determined when said wave width mean value inertia moment is normalized so as that the wavelength becomes a standard value,
a wave width dispersion inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to dispersion vector whose vector component is dispersion in which the wave form is equally divided in the wave height direction and the dispersion is calculated from time value for each division unit and the rotational center is constructive to time axis of waveform foot, and
a normalized wave width dispersion inertia moment determined when said wave width dispersion inertia moment is normalized so as that the wavelength becomes a standard value.
The seventh form of the present invention is the storage medium for identification comprising a storage medium in which said computer control program is stored.
The eighth form of the present invention is the storage medium for classification analysis comprising a storage medium in which said computer control program is stored.
According to the first form, the identification method comprises the steps of introducing a sample containing an analyte into a measurement space, obtaining pulse signal data detected due to said introduction, and identifying nonconforming data detected by elements other than said analyte from said pulse signal data by execution of a computer control program.
In addition, said computer control program includes an identification analysis program using the machine learning to learn a classifier that classifies positive and negative examples from positive example data of a positive example set and unknown data of an unknown set in which either positive or negative example is unknown. When type 1 data of a pulse signal are obtained under first measurement condition measured by introducing a sample not containing an analyte in said measurement space and type 2 data of a pulse signal are obtained under second measurement condition measured by introducing a sample containing an analyte in said measurement space, a storage means is included for storing said type 1 data and said type 2 data.
Furthermore, said nonconforming data included in said type 2 data can be identified by executing said identification analysis program, through using said type 1 data as said positive example data and said type 2 data as said unknown data.
Therefore, in the present embodiment, a classifier based on the PU classification method can be configured to identify nonconforming data contained in the pulsed signal obtained as a result of measurement with high accuracy, and for example, can contribute to improve the reliability of the measurement result by the advanced sensing device.
In particular, since the classifier in this embodiment can be configured with each data of a nonconforming data set collected in the past and an actually measured data set in which positivity or negativity is unknown, it has the excellent removal performance of nonconforming data that cannot be achieved by the conventional method of identifying by simple signal strength, and has a wide applicability to analysis of various measurement data.
According to the second form, the classification analysis method comprises the steps of introducing a sample containing an analyte into a measurement space, obtaining pulse signal data detected due to said introduction, obtaining analyzed data through removing said nonconforming data detected by elements other than said analyte from said pulse signal data, and performing a classification analysis of said analyzed data by execution of a computer control program.
In addition, a nonconforming data storage means is included for storing nonconforming data identified by the identification method according to the first form, said computer control program includes a classification analysis program that performs said classification analysis using the machine learning, a feature value is obtained in advance which indicates a feature of waveform form of said pulse signal, said feature value obtained in advance is set as the learning data for said machine learning, said feature value obtained from said pulse signal of said analyzed data removed said nonconforming data is set as a variable, and said classification analysis on said analyte is performed by executing said classification analysis program.
Therefore, in this embodiment, it is possible to perform the classification analysis on the analyte with high accuracy by the analyzed data from which the nonconforming data identified with high accuracy by the classifier based on the PU classification method according to the first embodiment is removed.
According to the third form, each feature value described above is a feature value derived from the waveform form of the pulse signal, and by using one or more feature values in these feature value groups, the classification analysis by the machine learning can be performed with higher accuracy.
In the present embodiment, the classification analysis is not limited to the case where the classification analysis is performed using at least one or more selected from the feature value group, but the combination analysis of two or more selected from the feature value group can be performed.
According to the fourth form, the identification device comprises a means introducing a sample containing an analyte into a measurement space, a means obtaining pulse signal data detected due to said introduction, and a means identifying nonconforming data detected by elements other than said analyte from said pulse signal data by execution of a computer control program.
In addition, said computer control program includes an identification analysis program using a machine learning to learn a classifier that classifies positive and negative examples from positive example data of a positive example set and unknown data of an unknown set in which either positive or negative example is unknown, when type 1 data of a pulse signal are obtained under first measurement condition measured by introducing a sample not containing an analyte in said measurement space and type 2 data of a pulse signal are obtained under second measurement condition measured by introducing a sample containing an analyte in said measurement space, an storage means is included for storing said type 1 data and said type 2 data, and said nonconforming data included in said type 2 data can be identified by executing said identification analysis program, by using said type 1 data as said positive example data and said type 2 data as said unknown data.
Therefore, in the present embodiment, a classifier based on the PU classification method can be configured to identify the nonconforming data included in the pulse signal obtained as a result of measurement with high accuracy. For example, it is possible to provide an identification device contributing to that the reliability of the measurement result by the advanced sensing device can be improved in performance.
In particular, since the classifier in the present embodiment can be configured with a nonconforming data set collected in the past and an actually measured data set in which positivity or negativity is unknown without using the specific properties to knowledge and problem on the object, it has the excellent removal performance of nonconforming data that cannot be achieved by the conventional method of identifying by simple signal strength, and has a wide applicability to analysis of various measurement data.
According to the fifth form, the classification analysis device comprises a means introducing a sample containing an analyte into a measurement space, a means obtaining pulse signal data detected due to said introduction, a means obtaining analyzed data through removing said nonconforming data detected by elements other than said analyte from said pulse signal data, and a means performing a classification analysis of said analyzed data by execution of a computer control program.
In addition, a nonconforming data storage means is included for storing nonconforming data identified by the identification method according to the fourth form, said computer control program includes a classification analysis program that performs said classification analysis using the machine learning, a feature value is obtained in advance which indicates a feature of waveform form of said pulse signal, said feature value obtained in advance is set as the learning data for said machine learning, said feature value obtained from said pulse signal of said analyzed data removed said nonconforming data is set as a variable, and said classification analysis on said analyte can be performed by executing said classification analysis program.
Therefore, in the present embodiment, the classification analysis apparatus can be provided which can perform a classification analysis on the analyte with high accuracy from the analyzed data from which the nonconforming data identified with high accuracy by the classifier based on the PU classification method according to the fourth embodiment is removed.
According to the sixth form, each feature value described above is a feature value derived from the waveform form of the pulse signal, and by using one or more feature values in these feature value groups, the classification analysis by the machine learning can be performed with higher accuracy.
In the present embodiment, the classification analysis is not limited to the case where the classification analysis is performed using at least one or more selected from the feature value group, but the combination analysis of two or more selected from the feature value group can be performed.
According to the seventh form, it is possible to provide the storage medium for identification that stores the computer control program according to the first form. Therefore, since the storage medium according to the present form has the effect of the computer control program described in the first form, it is possible to install the computer control program stored in the storage medium for identification in the computer and execute the identification operation by the computer, so that it is possible to perform the identification analysis with high accuracy.
According to the eighth form, it is possible to provide the storage medium for classification analysis that stores the computer control program according to the second form. Therefore, since the storage medium according to the present form has the effect of the computer control program described in the second form, it is possible to install the computer control program stored in the storage medium for classification analysis in the computer and execute the classification analysis operation by the computer, so that it is possible to perform the classification analysis with high accuracy.
As the storage medium according to the seventh and eighth forms, any one of storage media readable by a computer such as a flexible disk, a magnetic disk, an optical disk, a CD, an MO, a DVD, a hard disk, a mobile terminal and the like can be selected.
According to the present invention, by using a computer terminal, it is possible to be utilized to an information compression technology of DNA storage media and a drug discovery using an artificial base pairing, or to measure a fine dust mixed in a measurement sample or an analysis substance contained in. a body fluid, and it is possible to perform the data analysis with high accuracy in the field of technology in identification and removal of nonconforming data caused by fine substances such as red blood cells, white blood cells, and platelets.
The classification analysis device according to one embodiment of the present invention will be described below with reference to the drawings. In the present embodiment, there will be explained the base species analysis form for classification analysis of DNA constituent molecule as an example of the analyte.
(1A) of
The measurement system includes a measurement space MS configured by a storage container that stores a solution containing base molecules, and a pair of fine-shaped electrodes D1 and D2 that are arranged to face each other in the measurement space MS. The electrodes D1 and D2 are nanogap electrodes formed of gold (Au) element which are arranged at a minute distance from each other. The fine distance is formed to be about 1 nm. The measurement sample in the measurement space MS is a solution sample containing a solvent (pure water) and DNA constituent molecules mixed in the solvent.
As described in Non-Patent Document 4, the nanogap electrode is a device expected as a next-generation DNA sequencer. This electrode is an electrode gap having a very fine gap created by using a technique called mechanical fracture joining. When a constant voltage is applied to the electrode gap, a current due to the quantum mechanical tunnel effect (tunnel current: see the broken line in
By measuring the tunnel current pulse due to this nanogap electrode, it is possible to identify the type of DNA base molecule in single molecule units, and to identify the amino acid sequence of a peptide and a modified amino acid molecule that becomes a disease marker, which was difficult with existing technology. In the measurement system of
As the molecules to be measured, two kinds of dithiophene uracil derivatives (hereinafter abbreviated as BithioU) and TTF uracil derivatives (hereinafter abbreviated as TTF), which are artificial nucleobases, were used. These molecules are obtained by chemically modifying an epigenetic site (an acquired modification site where DNA methylation or the like occurs) for easy identification. As indicated by the arrow F, as the driving force source for allowing DNA molecules to pass through the vicinity of the gap, in addition to the Brownian motion of the molecules themselves, those by electrophoresis, electroosmotic flow, or dielectrophoresis can be used.
As shown in (1B), the pulse determination part of the pulse signal is a ⅓ part of the center of the measurement waveform, and this pulse waveform data is used as the analyzed data.
2A1 and 2A2 in (2A) show examples of waveform when the base molecule BithioU is detected. 2B1 and 2B2 in (2B) show examples of waveform when the base molecule TTF is detected. 2A3, 2A4 of (2A) and 2B3, 2B4 of (2B) show examples of noise waveform mixed when a base molecule is detected.
In the measurement system of
The PC 1 has the identification processing function of nonconforming data and the classification analysis processing function. However, in the present invention, the PC 1 can be configured by two dedicated terminals having separately the identification processing function and the classification analysis processing function.
The identification processing of the PC 1 is performed by the processing procedure based on the following identification method. The identification/classification analysis program includes the identification processing program based on the PU method using a machine learning to learn a classifier that classifies positive and negative examples from positive example data of a positive example set and unknown data of an unknown set in which either positive or negative example is unknown,
(Process 1-1) Type 1 data of a pulse signal obtained under first measurement condition measuring by introducing a sample (solvent only) not containing an analyte (DNA constituent molecule) in the measurement space MS is taken into and stored in the RAM 4 of the storage means.
(Process 1-2) Type 2 data of a pulse signal obtained under second measurement condition measuring by introducing a sample (solvent and DNA constituent molecule) containing an analyte (DNA constituent molecule) in the measurement space MS is taken into and stored in the RAM 4 of the storage means.
(Process 1-3) In order to match with the input format of the identification processing program, the attribute vectors of the type 1 data and the type 2 data are created.
(Process 1-4) The identification processing program is executed with the type 1 data as the positive example data and the type 2 data as the unknown data.
(Process 1-5) The probability p (s=1|x) is extracted and obtained by executing the identification processing program. The probability data is stored and saved in a predetermined area of the RAM 4. Note that the attribute vector used in the analysis by the following PU method is based on multidimensional data and is represented by a vector, but in the following description, the vector notation is particularly omitted.
(Process 1-6) Based on the probability, nonconforming data detected due to elements other than the analyte (fluctuations and impurities of metal atoms on the electrode surface, not from the base molecule mentioned above) included in the type 2 data is detected and identified. The detected nonconforming data is stored and saved in a predetermined area of the RAM 4.
The classifier software of machine learning platform freeware Weka disclosed in Non-Patent Document 2 can be used for the identification processing program.
The measured pulse waveform data varies in both wavelength and wave height, so it is necessary to use the attribute vectors with uniform dimensions as input in order to identify the base type by the machine learning classifier. When the machine learning processing program is executed, as the preprocessing matching with the input format, the preprocessing that performs a kind of coarse-graining and creates the attribute vector reflecting the pulse waveform is performed in (processing 1-1) and (processing 1-2).
The measured pulse waveform data varies in both wavelength and wave height, so it is necessary to use the attribute vectors with uniform dimensions as input in order to identify the base type by the machine learning classifier. When the identification processing program is executed, as the preprocessing matching with the input format, the preprocessing that performs a kind of coarse-graining and creates the attribute vector reflecting the pulse waveform is performed in (processing 1-3).
As shown in
As shown in
In the present embodiment, in order to configure a binary classifier, two feature values of wave height and wavelength are used. In the verification experiment for verifying the identification accuracy of nonconforming data according to the present embodiment, the verification was performed using the following eight attribute vectors V1 to V8 created from one pulse waveform data.
(V1) Pulse height vector (hvNrmd) with the pulse peak value normalized to “1”
(V2) Unnormalized wave height vector (hvRaw)
(V3) Wavelength direction time vector (wvNrmd) in which the pulse wavelength time is normalized to “1”
(V4) Unnormalized wavelength direction time vector (wvRaw)
(V5) (dh+2dw) dimensional vector connecting V1 and V2
(V6) (dh+2dw) dimensional vector connecting V1 and V4
(V7) (dh+2dw) dimensional vector connecting V2 and V3
(V8) (dh+2dw) dimensional vector connecting V2 and V4
In the verification experiment, the above eight attribute vectors are created, and their identification accuracies are compared. The number of division at the time of creating the attribute vector is set to be dh=10 and dw=5 uniformly after the preliminary analysis.
In the case of a normal binary classifier, the classifier is generated by learning from data in which the positive examples and the negative examples are given. On the other hand, in the present embodiment, since the nonconforming data are mixed in the measurement data, the classifier based on the PU method is used. As described in detail in Non-Patent Document 3, the PU method used in the present embodiment is a kind of learning algorithm semi-supervised for learning from positive examples and unlabeled data and performing binary classification of positive examples/negative examples. The outline of the processing procedure of the learning algorithm of the PU method stored in the ROM 3 is as follows.
Case x (input data) is an attribute vector related to a pulse waveform, y is its class label, and s is a flag indicating whether or not a class label is attached to the case. In the set of input cases, only a part of positive examples (y=1) is labeled (s=1), and other positive examples and all negative examples (y=0) are not labeled (s=0). That is, if the sample is a negative example, the probability of being labeled is zero and p (s=1|x, y=0)=0. Using such a case set as an input to the learning algorithm of the binary classifier, the probability g (x)=p (s=1|x) that the sample is labeled can be obtained. Furthermore, since what is originally desired to get is not g(x) but p (y=1|x), p (y=1|x) is extracted by applying the following correction.
In all cases set, G (x)=p (s=1|x) showing the probability that a sample is labeled, is derived to the relation of g(x)=p (y=1|x) p (=1|y=1) by the derivation process shown in
Here, the probability of labeling in the positive example set is uniformly random, that is, it is assumed that p (s=1|y=1, x)=p (s=1|y=1)=c is constant regardless of x. Assumes value. This is because the measurement data handled is not arbitrary data that is intentionally biased.
Here, the constant c can be estimated as follows. If it is labeled to be uniformly random in the positive example set, g(x) matches the proportion of labeled example set contained in the positive case when x is positive, and g(x)=p (s=1|y=1)=c holds.
Therefore, by using g(x) obtained by the normal binary classifier that does not depend on the PU method, c can be estimated as the mean in the labeled example set L that is a positive example (c given in the following equation 1).
In process P1-1, the process of learning g(x) from the learning data set is performed. Next, in process P1-2, the process of estimating c from the verification data set based on Equation 1 is performed. In process P1-3, the process of identifying positive/negative examples for test data is performed by g (y=1|x) determined from the relationship g (y=1|x) g(x)/c. Is called. The judgement criterion in this case can be set as g (y=1|x)>0.5.
The present invention can extract the probability that the unlabeled example is the positive example in the configuration of the classifier based on the PU method shown in
All labeled examples are positive, but unlabeled examples can be either positive or negative. If the probability that the unlabeled example is the positive example is w(x), the probability that it is the negative example is 1−w(x). Therefore, all unlabeled examples are duplicated, one is treated as the positive example, and the other is treated as the negative example. A weight w(x) is given to the unlabeled example x treated as the positive example, and the weight “1−w(x)” is given to the unlabeled example x treated as the negative example. Since all labeled examples cases are positive examples, they are treated as the positive examples with the weight “1”. The classifier is created by using these weighted examples set as learning data.
Here, assuming that c and g(x)=p (s=1|x) are obtained by the method shown in
The verification experiment of the identification accuracy according to the present embodiment will be described below.
From a set of measurement pulses measured using the nanogap electrode with the DNA constituent molecule as an analyte, 1) First, the classifier based on the PU method is constructed as the pre-processing, and the noise-derived pulses (nonconforming data) are identified (see
For noise removal, the tunnel current pulses measured with the nanogap electrode are obtained in advance only for a solvent that does not contain the base (BithioU, TTF). This pulse set is the pulse derived from noise not relating to the base and is referred to as “noise pulse set”. Next, the current pulse measured for the solvent mixed with the base BithioU is obtained. The TTF is acquired in the same manner. This pulse set includes both the “base pulse” derived from the base and the noise pulse. Therefore, this is called “base+noise pulse set”.
Since the pulse in the noise pulse set is always the noise pulse, it is regarded as the positive example set (data set of the first data) and on the other hand, the pulse in the base+noise pulse set is unknown which pulse it is, so it is considered as the unlabeled example set. By the PU classifier process shown in
The PU classifier processing is used only once to classify the positive and negative examples of the noise pulse and base pulse, and no problem due to overlearning will occur, so the PU classifier is created using the entire pulse set as the learning data, and therefore, the classification analysis was performed to separate all pulse sets into the noise pulses and the base pulses. In this way, the BithioU base pulse set was obtained from the BithioU base+noise pulse set by the PU classifier, and the TTF base pulse set was obtained from the TTF base+noise pulse set by the PU classifier.
In order to evaluate the identification accuracy for the base pulse and the base+noise pulse, the identification experiment of the base type using the normal binary classifier was performed for the base pulse set of BithioU and TTF from which noise data was removed and separated. In the identification experiment, when either of the number of base pulses of the two kinds of bases was less than 10, it was excluded from the experiment object because there were too few learning examples.
F-measure (F-scale shown in
In addition, in order to see the effect of noise removal by the PU classifier, the identification experiment of BithioU and TTF was also performed on the base+noise pulse set before performing the noise data removal. The accuracy evaluation was performed by 10 CV on the pulse set obtained by random extraction of N components from the base+noise pulse set for each of BithioU and TTF as in the case of the base pulse.
The experimental conditions for the verification experiment are described below.
With the machine learning software shown in
As a pulse extraction parameter, the two parameters of α and k (the adjustment factor shown in
The ensemble learning “Rotation Forest” is adopted as the classifier for the classification analysis, and as base classifier used internally, there are used the 22 types of classifiers among those implemented in Weka, which can perform the binary classification of the input example continuous value vector. As the PU method, two methods of g(x) and w (x) shown in
The identification experiment of nonconforming data conducted under the above experimental conditions were performed for 3272 cases where the number of base pulses was 10 or more for both bases, while selecting from all combinations given by the pulse extraction parameter 16 types x the attribute vector 8 types x 22 classifiers implemented in Weka x PU method 2 types. In this identification experiment, for the sake of simplicity, the conditions used for these three parties of 1) BithioU noise removal, 2) TTF noise removal, and 3) 2-base identification after noise removal are all common, where the above conditions consist of the pulse extraction parameter, the attribute vector, the classifier, and the PU classifier method. Under the same conditions, the identification experiment was also performed on the base+noise pulse set without using the noise removal by the PU classifier.
As can be seen from
As can be seen from
The classification analysis device by the PC 1 has the classification analysis function with high accuracy for the analyzed data from which the nonconforming data identified by above identification processing is removed. This classification analysis function includes the following analysis procedures.
(C1) The nonconforming data detected by the above identification processing which is caused by elements other than the analyte are stored in a predetermined area of the RAM 4 of the nonconforming data storage means. Not only the nonconforming data is detected and stored by the PC 1, but also a nonconforming data file stored in advance in an external terminal may be introduced and stored in the PC 1.
(C2) The data group obtained by removing the stored nonconforming data from the pulse signal data detected by introducing the sample containing the analyte into the measurement space is stored in a predetermined area of the RAM 4 as the analyzed data. The data group from which the nonconforming data have been removed in advance may be introduced into the PC 1 and stored as the analyzed data.
(C3) The computer control program installed in the PC 1 includes the classification analysis program for performing the classification analysis using machine learning, and is stored in the ROM 3.
(C4) The feature value is obtained in advance which indicates the feature of waveform form of the pulse signal, the feature value obtained in advance is set as the learning data for the machine learning, the feature value obtained from the pulse signal of the analyzed data removed the nonconforming data is set as the variable, and the classification analysis on the analyte can be performed by executing the classification analysis program.
By the classification analysis device through the PC 1, the classification analysis on the analyte can be performed with high accuracy based on the analyzed data, because the feature value is obtained in advance which indicates the feature of waveform of pulse signal, the feature value obtained in advance is set as the learning data for machine learning, the nonconforming data identified with high accuracy by the classifier based on the PU classification method are removed from the analyzed data, the feature quantity obtained from the analyzed data removed the nonconforming data is set as the variable, and the classification analysis on the analyte is performed by executing the classification analysis program.
In the present embodiment, since it is possible to perform the identification of the nonconforming data with high accuracy and the classification analysis, for example, a way to enable the identification of artificial bases is opened, and it is possible to develop applications in the information compression technology for DNA storage media and the pharmaceutical creation using artificial base pairs.
The present invention is not limited to the electric current output waveform in the embodiment described above and can be applied to the identification of nonconforming data and the classification analysis for a wide range of output waveforms such as a voltage waveform and an impedance waveform.
The present invention is not limited to the detection waveform obtained by the measurement system using the nanogap electrode, and can be applied to the detection waveform due to the measurement system of a microstructure equivalent to the sample object through which the sample object passes, such as a through hole, a well (concave), a pillar (convex), and a flow path. The applicable range of measurement data in the present invention is all of time series measurement data, and is not limited to electrical measurement, and can include detection data of physical phenomena such as optical measurement and sound.
The removal target caused by the elements other than the analyte in the present invention is not limited to the quantum level elements in the embodiment described above, and for example, can be applied to signals based on elements other than the analyte existing in a solution, a measuring instrument and a measuring device. That is, the present invention can be applied to the identification/removal technology of the nonconforming data caused by for example, fine dust mixed in a measurement sample and minute substances such as red blood cells, white blood cells and platelets in the case where measurement object is blood.
The feature value used in the identification processing and the classification analysis according to the present invention is not limited to the above-described wave height and wavelength, and various feature values derived from the waveform form can be used. The present inventors have come up to grasp the effective feature values from analysis of waveform data that appears in particle detection technology using a micro/nanopore device. The particle detection technique using a micro/nanopore device is disclosed in Patent Document 1 and the like. Hereinafter, the feature values effective for the present invention and the classification analysis processing when using them will be described in detail.
The particle detection device is constituted by the micro-nanopore device 8 and an ionic current detection portion. The micro-nanopore device 8 has a chamber 9, a partition wall 11 partitioning the chamber 9 into upper and lower accommodation spaces, and a pair of electrodes 13, 14 arranged on the front and back sides of the partition wall 11. The partition wall 11 is formed on a substrate 10. A small through-hole 12 is formed in the vicinity of the center of the partition wall 11. Below the through-hole 12, a recess portion 18 is formed by removing a part of the substrate 10 downward in a concave shape.
The micro-nanopore device 8 is fabricated using a manufacturing technique (for example, an electron beam drawing method or photolithography) of a semiconductor device or the like. That is, the substrate 10 is made of Si material, and a partition wall 11 made of Si3N4 film is formed as a thin fin on the surface. The recess portion 18 is formed by removing a part of the substrate 10 by etching.
The partition wall 11 is formed by laminating SiN film with 50 nm thickness on Si substrate having a size of 10 mm square and a thickness of 0.6 mm. A resist is applied to the Si3N4 film, and a circular opening pattern having a diameter of 3 μm is formed on it by an electron beam writing method, and the through-hole 12 is bored. On the back side of the through-hole 12, wet etching with KOH is performed to form a 50 μm square opening to provide the recess portion 18. The formation of the recess portion 18 is not limited to wet etching, but it can be performed by isotropic etching etc. using the dry etching with CF4 gas or the like, for example.
In addition to the SiN film, the insulating film such as Si2 film, Al2O3 film, glass, sapphire, ceramic, resin, rubber, elastomer, or the like can be used for the film of the partition wall 11. The substrate material of the substrate 10 is not limited to Si, and glass, sapphire, ceramic, resin, rubber, elastomer, SiO2, SiN, Al2O3, or the like can be used.
The through-hole 12 is not limited to the case of forming the thin film on the above substrate, and for example, by attaching a thin film sheet having the through-hole 12 onto the substrate, the partition wall having the through-hole may be formed.
The ionic current detection portion is constituted by an electrode pair of the electrodes 13 and 14, a power supply 15, an amplifier 16, and a voltmeter 20. The electrodes 13, 14 are arranged to face each other through the through-hole 12. The amplifier 16 is constituted by an operational amplifier 17 and a feedback resistor 19. The (−) input terminal of the operational amplifier 17 and the electrode 13 are connected. The (+) input terminal of the operational amplifier 17 is grounded. The voltmeter 20 is connected between the output side of the operational amplifier 17 and the power supply 15. The applied voltage of 0.05 to 1V can be used between the electrodes 13 and 14 by the power supply 15, but in this embodiment, 0.05V is applied. The amplifier 16 amplifies the current flowing between the electrodes and outputs it to the voltmeter 20 side. The electrode material of the electrodes 13 and 14 are, for example, Ag/AgCl electrode, Pt electrode, Au electrode or the like, preferably Ag/AgCl electrode can be used.
The chamber 9 is a flowable substance accommodation container which hermetically surrounds the micro-nanopore device 8, and can be made of electrically and chemically inert materials such as glass, sapphire, ceramic, resin, rubber, elastomer, SiO2, SiN, Al2O3, or the like.
An electrolytic solution 24 containing the subject 21 is filled in the chamber 9 from an injection port (not shown). The subject 21 is, for example, an analyte such as a bacterium, a microparticulate substance, a molecular substance or the like. The subject 21 is mixed in the electrolytic solution 24 which is a flowable substance, and the measurement is performed by the micro-nanopore device 8. At the end of the measurement by the ion current detection portion, the filling solution can be discharged from the discharge port (not shown). As the electrolytic solution, for example, in addition to phosphate buffered saline (PBS), Tris-EDTA (TE) buffer and dilution solutions thereof, all electrolytic solution agents similar thereto can be used. The measurement is not limited to the case in which it is always performed when the subject-containing electrolytic solution is introduced into the chamber 9 and filled. The subject-containing electrolytic solution (flowable substance) is pumped out from the solution reservoir by a simple pumping device and injected from the injection port into the chamber 9 and discharged from the discharge port after the measurement. Furthermore, new solution is stored in the solution reservoir or another solution reservoir, and newly pumped out to perform next measurement, so that the continuous measurement system can be constituted.
When a voltage from the power source 15 is applied between the upper and lower electrodes 13, 14 of the through-hole 12 in a state where the electrolytic solution 24 is filled in the chamber 9, a constant ion current proportional to the through-hole 12 flows between the electrodes. When a subject such as bacteria etc. existing in the electrolytic solution 24 passes through the through-hole 12, a part of the ion current is inhibited by the subject, so that the pulsed ion current reduction can be measured by the voltmeter 20. Therefore, according to the particle detecting device using the micro-nanopore device 8, by detecting the change in the waveform of the measured current, it is possible to detect each individual presence of the particles contained in the flowable substance by passing through the through-hole 12 for each subject (for example, particle) with high accuracy. The measurement mode is not limited to the case where the measurement is performed while forcibly flowing the flowable substance but can include a case of measurement while flowing the flowable substance non-forcibly.
The measurement output of the ion current by the voltmeter 20 can be externally outputted. The external output is converted into digital signal data (measured current data) by a conversion circuit device (not shown), temporarily stored in a storage device (not shown), and then stored in the data file storage portion 5. Measurement current data acquired in advance by the particle detection device using the micro-nanopore device 8 can be externally input to the data file storage portion 5.
The classification analysis processing of
(a) As a result of measurement by the nanopore device 8a on the flowable material containing the predetermined analyte (for example, E. coli Ec or Bacillus subtilis Bs), the pulse signals De and Db corresponding to the passage of analytes through the through-hole 8b are obtained as the detection signals for each type, and the feature values indicating the features of their waveform forms are determined in advance. The pulse signals De and Db are the signals obtained by passing through the through holes 8b of E. coli Ec and B. subtilis Bs, respectively.
(b) The computer analysis unit 1a incorporates the classification analysis program for performing the classification analysis by the machine learning. The feature values obtained in advance in (a) are the feature values obtained from the known data of E. coli Ec and Bacillus subtilis Bs, and are used in the computer analysis unit 1a as the learning data for the machine learning.
(c) For example, when the mixture mixed in the flowable material with the unknown state on the content ratio or the content number of E. coli Ec and B. subtilis Bs is used as the classified analyte Mb, in the same manner as the case of obtaining known data of (a), the measurement is performed by the nanopore device 8c. By this measurement, a pulse signal Dm is obtained as the analyte data by passage of the classified analyte Mb through the through hole 8d.
(d) Using the feature value based on known data as the learning data and the feature value obtained from the pulse signal Dm of the analyzed data as the variable, by executing the classification analysis program, it is possible to perform the classification analysis relating to the predetermined analyte in the analyzed data.
According to the classification analysis described above, the classification analysis by the machine learning is performed based on the feature value, and the analyzed data of unknown type can be classified into one 1b derived from the passage of E. coli Ec or B. subtilis Bs and one not derived from them. In addition, the feature value according to the present invention may be createdin the computer analysis unit 1a, or may be given to the computer analysis unit 1a after being created using another feature value creation program.
Main control processes include input process (step S100), feature value acquisition process (step S101) for acquiring feature values from input data, classification analysis process (step S104), number analysis process (step S105) and output process (step S106). In the input process (step S100), various inputs necessary for PC operation, start input of built-in program, execution instruction input of various analyses, input of measurement current data and/or feature value data, setting input of output mode, input of the designated feature value when the feature value is designated at the analysis time, and so on. This input process also includes the removal process of the nonconforming data. The classification analysis process (step S104) or the number analysis process (step S105) can be performed (steps S102 and S103) by performing the operation of specifying the analysis type by the input means 6. The classification analysis process can be classified and analyzed using the vector value data of the feature value acquired from the input data in the feature value acquisition process (step S101). The number analysis process can perform the number analysis using scalar data of the feature value acquired from the input data in the feature value acquisition process. The present embodiment is an embodiment having the number analysis process function in addition to the classification analysis process function, but the present invention can be performed by an embodiment having only the classification analysis process function.
The computer control program according to the present embodiment includes a number analysis program for analyzing the number or the number distribution of particle types. In the number analysis process (step S105), the number analysis program can be executed. In the output process (step S106), the analysis result data can be output in the classification analysis process (step S104) and the number analysis process (step S105). For example, various analysis result data are displayed and output on the display means 7. When a printer (not shown) is connected to the PC 1 as the output means, the print output of various analysis result data is possible.
The classification analysis device according to the present embodiment is configured such that a flowable material (electrolytic solution 24) including one or more types of particles as an analysis target (an example of analyte) is supplied to the upper side surface of the partition wall 11 by execution of the number analysis program. The change of current flow of between the electrodes 13 and 14 is caused by the particles passing through the through-hole 12, and it has an analysis function such that the number or the number distribution of the particle types is analyzed based on the data of detection signal of the change. That is to say, the PC 1 can perform the number analysis process for the measured current data stored in the data file storage portion 5 by executing the number analysis program stored in the ROM 3 under the control of the CPU 2. The number analysis process is configured from the following processes in which the probability density estimation is performed for the data group based on the feature value indicating the feature of the waveform of the pulse signal corresponding to particle passage contained in the detection signals and the automatic analysis of the particle number of each type is performed based on the number analysis method deriving the particle number for each particle type.
The number analysis program includes the probability density function module program for obtaining the probability density function from the data group based on the feature value indicating the feature of the waveform of the pulse signal corresponding to the particle passing through the through-hole 12 obtained as the detection signal, and the particle type distribution estimation program for deriving the number of each particle type from the result of the probability density estimation. Furthermore, the number analysis program includes the feature value extraction program for extracting the feature value indicating the feature of the waveform of the pulse signal with reference to the baseline extracted from the data group, and the data file creation program for creating the data file due to the pulse feature value data for each particle obtained based on the extracted feature value.
The classification analysis process and the number analysis process are performed for the data created by the data file creation program. The feature value extraction program includes the baseline estimation process program for extracting the baseline from the original measurement current data. In the feature value acquisition process (step S101), the feature value extraction program and the data file creation program are executed to create the feature values from the data input in the input process (step S100), and to store it in the data file for feature value storage of the RAM 4. The input data for classification analysis are known data required to create the feature quantities used as the learning data, and data for analysis (analysis data). The feature value data created from the known data is stored in the feature value storage data file DA due to the known data, and the feature value data created from the analysis data is stored in the feature value storage data file DB due to the analysis data.
When the classification analysis is performed, the analysis process can be performed by acquiring the vector value data of the feature values from the data files DA and DB. The input data for number analysis is only the data for analysis (analysis data). The feature value data created from the input data for number analysis is stored in the data file for number analysis DC, and when performing the number analysis, the scalar data of the feature value is acquired from the data file DC and the analysis process can be performed.
Since the form of the true probability density function is unknown as the premise of particle type distribution estimation, the execution of the probability density function module program performs the nonparametric (not specifying the functional form) probability density estimation called Kernel method. The original data of the estimation target is the pulse appearance distribution data including, for example, a pulse height h, a time width Δt, an appearance number, etc. obtained from the pulse signals. Each data of the original measurement data distribution is expressed by a Gaussian distribution introducing measurement error uncertainty, and a probability density function is obtained by superimposing each Gaussian distribution. The probability density estimation process is performed by executing the probability density function module program and the original data can be represented by an unknown complex probability density function based on the original data (for example, pulse height, pulse width, appearance probability of the feature value).
As described above, the probability density estimation process is performed based on the Kernel method which is one of estimation methods of the nonparametric density function. The Kernel method is an estimation method in which a function (Kernel function) at one data point is applied, this is applied to all data points, and the arranged functions are superimposed, which is suitable for obtaining a smooth estimation value.
By executing the probability density function module program and by considering the multivariable multidimensional probability density from data such as pulse wave height, pulse width, etc. of the measured current waveform, the weighted optimum estimation is performed extending to two or more dimensions and the estimation process of the particle type number distribution is performed. EM algorithm software executed based on Hasselblad iteration method is used for the weighted optimum estimation. The EM algorithm is preinstalled in the PC 1. The particle type number distribution result obtained by the estimation process of the particle type number distribution can be output and displayed as the histogram of appearance frequency (number of particles) for each particle type on the display means 7.
The feature value according to the present invention is, as the parameters derived from the pulse signal, either of first type showing local feature of waveforms of said pulse signals and second type showing global feature of waveforms of said pulse signals. By carrying out the number analysis using one or two or more of these feature values, it is possible to analyze the number or number distribution corresponding to the type of analyte such as particle type etc. with high accuracy.
Based upon the difference of the attribute of the pulse waveform form of particle passage described above, the feature values used as a base for creating the probability distribution can be extracted for each particle type (E. coli and B. subtilis) from the pulse waveform data.
The feature value of the first type is one selected from a group of
the wave height value of the waveform in a predetermined time width,
the pulse wavelength ta,
the peak position ratio represented by ratio tb/ta of time ta and tb leading from the pulse start to the pulse peak,
the kurtosis which represents the sharpness of the waveform,
the depression representing the slope leading from the pulse start to the pulse peak,
the area representing total sum of the time division area dividing the waveform with the predetermined times, and
the area ratio of sum of the time division area leading from the pulse start to the pulse peak to the total waveform area.
The 5a to 5d in
(1) Wavelength (pulse width) Δt: Δt=te−ts (ts is the start time of the pulse waveform, te is the end time of the pulse waveform, Δt=ta).
(2) Wave height |h|:h=xp−xo (the height of pulse waveform up to xp of the pulse peak PP with reference to xo of BL).
(3) Peak position ratio r:r (tp−ts)/(te−ts) (the ratio of the time t=(tp−ts) from the pulse start to the pulse peak pp to the pulse wavelength (=Δt)).
(4) Peak kurtosis κ: It is normalized so as that the wave height |h|=1, ts=0 and te=1 hold, and there are collected the time set [T]=[[ti]| i=1, . . . , m] which is the time crossing the horizontal line of 30% in wave height from the pulse peak PP, and then the K is obtained so that the dispersion of the data of the time set [T] is calculated as the pulse waveform spread as shown in the following equation 2.
(5) As shown in (34 A), the depression θ is the slope leading from the pulse start to the pulse peak and is defined by the following equation 3.
(6) The area m is defined as the area [m] by the inner product of the unit vector [u] and the wave height vector [p] as shown in the following equation 4. In the following description, the vector notation of variable A is indicated by [A]. For example, as shown in the 10-division example of (34 B), the area m is the area representing the total sum of the time division area hi (hi=hx×hy, i=1 to 10 when width hi and height hy).
m=(u,p)=Σi1·hi [Equation 4]
Here, it is necessary to calculate and obtain the d-dimensional wave height vector [p](=(h1, h2, . . . , hd)) in advance as preparation for the feature value calculation.
As shown in (35 A), by equally dividing the wavelength into d number for one waveform data, the d number of data groups are differentiated. Next, as shown in (35B), the values of wave height are averaged for each group (each divided interval), for example, when dividing equally into 10, the average values A1 to A10 are obtained. This averaging can include a case where the wave height value is not normalized and a case where the wave height value is normalized. The area [m] described by equation 4 indicates a case where the normalization is not performed. The d-dimensional vector having the average values thus determined as components is defined as a “wave height vector”.
As shown in (36 A), when the sampling rate related to acquisition of pulse data is large, since the number of steps (number of data) T in the pulse part exceeds the dimension number d of the vector, by the above acquisition steps, it is possible to obtain the wave height vector of which component is the average of each section. On the other hand, as the sampling rate is lowered, the number of steps T in the pulse part falls below the dimension number d (>T) of the vector. In the case of T<d, since the average value of each section cannot be acquired by the acquisition procedure described above, it is possible to acquire the d-dimensional wave height vector by cubic spline interpolation.
The feature value extraction program includes the wave height vector acquisition program for acquiring wave height vector data. In the case where the pulse step number T exceeds the dimension number d of the vector (T>d)(T=d) by executing the wave height vector acquisition program, the average value of each division equally divided in the time direction is obtained, In the case where the pulse step number T is smaller than the dimension number d of the vector (T<d), a cubic spline interpolation is executed to obtain the d-dimensional wave height vector. That is, by performing the interpolation process using the cubic spline interpolation method, even when the number of pulse steps is small, the number of dimensions of the vector can be made constant.
(7) The area ratio rm is defined as the area ratio of the sum of the time interval area hi shown in (34B) in the section leading from the pulse start to the pulse peak to the total waveform area. The following Equation 5 shows the area ratio rm.
r
m=Σt<tpht/Σht [Equation 4]
The feature value of the first type is uniquely derived from the waveform of the pulsed signal such as the pulse wave height, the pulse wavelength, the pulse area and the like, so that it is the feature value showing the local feature. The second type of the feature value is the feature value indicating the global feature with respect to the first type of the local feature.
The second type of the feature value is one selected from a group of
the time inertia moment determined by mass and rotational radius when the mass is constructive to said time division area centered at the pulse start time and the rotational radius is constructive to time leading from said center to said time division area,
the normalized time inertia moment when said time inertia moment is normalized so as that the wave height becomes a standard value,
the mean value vector whose vector component is the mean value of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak,
the normalized mean value vector which is normalized so as that the wavelength becomes a standard value for said mean value vector,
the wave width mean value inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to mean value difference vector whose vector component is mean value difference of the same wave height position in which the wave form is equally divided in the wave height direction and the mean value of time values is calculated for each division unit in before and after each pulse peak and the rotational center is constructive to time axis of waveform foot, the normalized wave width mean value inertia moment when said wave width mean value inertia moment is normalized so as that the wavelength becomes a standard value, the wave width dispersion inertia moment determined by mass distribution and rotational center when the mass distribution is constructive to dispersion vector whose vector component is dispersion in which the wave form is equally divided in the wave height direction and the dispersion is calculated from time value for each division unit and the rotational center is constructive to time axis of waveform foot, and, the normalized wave width dispersion inertia moment when said wave width dispersion inertia moment is normalized so as that the wavelength becomes a standard value.
(8) The time inertia moment, as in (34B), is the feature value determined by mass and rotational radius when the mass is constructive to the time division area hi formed by equally dividing one waveform in i-dimension with a predetermined time interval and the rotational radius is constructive to the time leading from the center to the time division area hi. That is, the feature value of the time inertia moment is defined by [I] which is the inner product of the vector [v] and the wave height vector [p] as shown in the following equation 6. Here, when the dimension of the vector is n, [v]=(12, 22, 32, . . . , n2) and [p]=(h1, h2, . . . , hd). For example, the time inertia moment, as shown in the example of 10 divisions of (37 A), is the feature value determined when the time division area hi (hi=hx×hy, i=1 to 10 with width hx and height hy) is regarded as the mass and the time leading from the center to the time division area hi is regarded as the turning radius, where one waveform is divided into 10 with a predetermined time-interval as in (34 B), so that in the same manner of the area m of (6), it can be obtained from the wave height vector.
I=(v,p)Eii2·hi [Equation 6]
(9) The normalized time inertia moment is calculated by using the waveform normalized in the wave height direction in which the wave height becomes the reference value “1” for the waveform for which the time division area is created as shown in (8), and is the feature value defined by equation 6 using the wave height vector hi in the same manner shown in (8).
(10) The average value vector divides one waveform equally with i-dimension in the wave height direction as shown in the example of 10 division of (37 B), and the average value of the time values is calculated for each division unit (division region wi) in before and after each pulse peak, and the average value of the same wave height position of the division region wi is set as the component of the vector.
(11) The normalized average value vector is the feature value in the case where the wavelength is normalized to be the reference value with respect to the average value vector of (10).
(12) The wave width mean value inertia moment is, as shown in 10 division example of (37B), obtained by equally dividing one waveform into i-dimensions in the wave height direction in the same way as in (8) and by calculating the mean value of time values for each division unit (divided region wi) in before and after the pulse peak, so that the difference vector of the mean value having the difference of the mean value of the same wave height positions in the divided region wi as the components of the vector is regarded as the mass distribution hi (the dimension number of the vector is n, i=1 to n) and it is the feature value defined as the moment of inertia when the time axis At of the waveform foot is to be the rotation center. The definition formula is the same as equation 6, and the feature value of (12) can be obtained by the inner product of the vector [v] and the mass distribution hi.
(13) The normalized wave width mean value inertia moment is the feature value which uses the waveform normalized in the wavelength direction in which the wavelength becomes the reference value “1” for the waveform creating the division region wi shown in (12) and is the feature value defined by equation 6 using the mass distribution hi created through the same manner as in (12).
(14) The wave width dispersion inertia moment, similar to the wave width mean value inertia moment, is obtained by equally dividing one waveform into i-dimensions in the wave height direction and by calculating the dispersion from the time values for each division unit (divided region wi) in before and after the pulse peak, so that the dispersion vector having the dispersion as the components of the vector is regarded as the mass distribution hi (the dimension number of the vector is n, i=1 to n) and it is the feature value defined as the moment of inertia when the time axis At of the waveform foot is to be the rotation center, and this feature value is defined by equation 6, in the same manner as the wave width mean value inertia moment.
(15) The normalized wave width dispersion inertia moment is the feature value which uses the waveform normalized in the wavelength direction in which the wavelength becomes the reference value “1” for the waveform creating the divided region wi shown in (14) and is the feature value defined by equation 6 using the mass distribution hi created through the same manner as in (14).
The wave width mean value inertia moment and the wave width dispersion inertia moment are the feature values defined by equation 6 as described above, and the vector [p] in the definition is the difference vector of mean values of the time values in the case of the wave width mean value inertia moment, and the dispersion vector of the time values in the case of the wave width dispersion inertia moment. In the following description, the vector [p] in the moment of inertia with respect to the wave widths of (12) to (15) is expressed as [pw].
For the creation of moment of inertia with respect to the wave widths of (12) to (15), there is performed using the wave width vector [pw] (=[p1, p2, . . . , pdw]) in which the vertical and horizontal axes of the wave height vector shown in
The wave width vector is the difference vector or the dispersion vector of the mean value shown in the definition of the feature value of (12) to (15). By grasping the wave width vector as the density distribution, the wave width mean value inertia moment of (12) and (13) and the wave width dispersion inertia moment of (14) and (15) can be obtained. The wave width vector is the dw dimensional vector, for which the pulse waveform data is equally divided with dw dimensions in the wave height direction and whose components are the difference or dispersion of the mean values of the peak values obtained for each division. In the case of (37B), the dimension of the wave width vector is 10 dimensions. The time axis At shown in (37B) is different from the base line BL, and is the rotation axis line around the pulse foot obtained from the wave width vector.
The feature value extraction program includes the wave width vector acquisition program for acquiring the wave width vector by the creation calculation processing of the wave width vector of the following dw dimension.
Since the pulse waveform data distribute at various intervals in the wave height direction, there may be cases that it is contained one or more non-existent regions Bdin which data points do not exist in the section divided in the wave height direction. In (38A), an example of the non-existent region Bd is indicated by an arrow. In the non-existent region Bd, the data points do not exist because the data interval becomes coarse, so that it is impossible to obtain the component of the inertia moment with respect to the wave width defined by the above equation 6. Therefore, as in the case of the above-described pulse waveform spread, when the wave height to the pulse peak is equally divided by dw, the time set [Tk]=[[ti] i=1, . . . , m] of each wave height is collected and the components of the wave width vector are created. At this time, in the non-existent region Bd in which no data point exists, the component data is acquired by the linear interpolation. The linear interpolation is performed on two consecutive data that extends over the values of (10k+5)% (k=0, 1, 2, 3, . . . ) of the pulse peak. The (38B) shows an example of the linear interpolation point tk with respect to the height k of the non-existent region Bd occurring between the data points ti and ti+1. In the creation of the wave width vector, as shown in (38 C), when a discrepancy of the wave height data occurs in the foot region UR of the pulse waveform data, the wave height data Du on the side far from the pulse peak is truncated to align to the side close to the pulse peak. The execution process of the wave width vector acquisition program includes the linear interpolation process for the non-existent area Bd and the truncation process of the pulse height data Du for the discrepancy of the pulse height data.
The (39A) is the example in which the waveform is equally divided into ten intervals in the wave height direction, and there are shown the division region 39b of the wave width vector and the rotation axis line 39c which are obtained by performing the above linear interpolation process and the truncation process for one waveform 39.
As shown in (39 B), for each division unit, the mean value of the time values is calculated before and after the pulse peak, and there can be obtained the wave width vector which is the difference vector of the mean value having the difference of the mean value at the same wave height position in the divided region as the component of the vector. By regarding this difference vector of the mean value as the mass distribution, the wave width mean value inertia moment of (12) with the rotation axis line 39c (time axis) as the rotation center can be created. Further, by calculating the dispersion from the time value of each division unit, it is possible to obtain the dispersion vector having the dispersion as the component of the vector. By regarding this dispersion vector as the mass distribution, it is possible to create the wave width dispersion inertia moment of (14) using the rotation axis line 39c (time axis) as the rotation center. In addition, the average value vectors of (10) and (11) are the vectors in which the average value of time values is calculated and whose component is the average value at the same wave height position of the division region, and it is represented by the time vector with 2 dw dimensions in the case of the equal division of dw.
The vector dimension number of the wave height vector and the wave width vector used for creating the feature value need not be limited to the division number but can be arbitrarily set. The wave height vector and the wave width vector are obtained by dividing in one direction of the wavelength or the wave height, but the vectors which are divided into a plurality of directions can be used for creating the feature value.
The (40A) shows the data map 40a obtained by dividing one waveform data into the mesh shape. The data map 40a shows the distribution state of the number of data points in the matrix form by dividing the waveform data with the dn division in the time axis direction of the horizontal axis and the dw division in the wave height direction of the vertical axis. The (40B) shows the distribution state in which a part of the matrix-like section (lattice) is enlarged. In the distribution state of (40 B), 0 to 6 data points are distributed in 11×13 grids. By this matrix division, the waveform vector in which the number of data points in each lattice/the total number of data points is a component of the dn×dw dimensional vector is converted into a vector in which data groups of the matrix array are rearranged in a scanning manner, so that it is possible to create the feature value instead of the wave height vector and the wave width vector.
Generally, bacteria or the like are minute objects having finely the different forms. For example, in the case of average Escherichia coli, the body length is 2 to 4 μm and the outer diameter is 0.4 to 0.7 μm. In the case of the average Bacillus subtilis, the body length is 2 to 3 μm and the outer diameter is 0.7 to 0.8 μm. In addition, flagella of 20 to 30 nm are attached to Escherichia coli and the like.
When using bacteria or the like as subject particles, if slight differences are missed from the pulse waveform data, the number judgment accuracy will be lowered. For this reason, in order to accurately calculate the feature value and use it as the estimation basis of the probability distribution, it is necessary to accurately grasp the particle passage pulse wave height. For this purpose, it is necessary to estimate the base line of the measurement signal. However, because the base line of the original data of the measurement signal contains fluctuations due to noise data and weak measured current, the pulse wave height and the like need to be detected after the base line excluding the fluctuation component etc. is determined. It is preferable that the base line estimation (hereinafter referred to as BL estimation) is practiced online (instantly) by the computer in practice.
As a method for estimating BL on a computer, by using the Kalman filter suitable for estimating the amounts that change from moment to moment from observations with discrete errors, the disturbances (system noise and observation noise) are removed, the base line BL can be estimated.
The Kalman filter is a method for estimating the value at the time [t] of updatable state vector [x], in which the discrete control process is defined by the linear difference equation shown in (6A) of
It is assumed that the state vector [x] is indirectly estimated by the observation model shown in (6B) of
The measured current data [X] is not a vector but a scalar, further various matrices are also scalars, and it can be regarded as [F]=[G]=[H]=[1]. Therefore, when letting [xt], [yt] and [vt] be the base line level of the actual current value at the time t, the current measured at the time t, and the observation noise at the time t, respectively, [xt] and [yt] are expressed as shown in (6 C) of
The estimation due to the Kalman filter is performed by repetition of prediction and updating. The estimation of the base line is also executed by repetition of prediction and updating due to the Kalman filter.
When executing the BL estimation process, the values of the start time m, the constants k, a of the adjustment factors necessary for the prediction and update process in the Kalman filter are adjusted (tuned) and decided to appropriate values according to the data attribute of the estimation target in advance. The value of α is the value for adjusting the dispersion of the estimated value of the base line. The value of k is the value related to the number of executions of update Ain the Kalman filter shown in
When the step portion (10 C) of the base line shown in (10B) is detected from the waveform of the bead model, the immediately preceding period becomes the initial value calculation period. For example, when m=100000, the 11 to 12 pulses having significance can be visually confirmed in the period excluding the initial value calculation period.
The (12 A) of
Comparing the three kinds of simulation results in
The BL estimation process of
|et|>α·σv,t et=yt−{circumflex over (x)}t|t-1 [Equation 7]
When the condition of the equation 7 is not satisfied, the update B of the Kalman filter at the time t is executed, and the processes of the steps S53 to S55 are repeated for each data whose unit time has elapsed. When the above condition of Equation 7 is satisfied, the number value is cumulatively stored in the count area of the RAM 23 every time (steps S54 and S56). Next, on the basis of the count value, it is judged whether or not the condition of the equation 7 has been satisfied by k times consecutively starting from the time s (step S57). If it is not consecutive by k times, the process proceeds to step S55 and the update Bis performed.
When the k times consecutive, the process proceeds to step S58, and it is judged that the hold necessary period for determining the BL has started. At this time, the hold start time of the hold necessary period is stored as s in the RAM 23, and the operation result of the Kalman filter between the time (s+1) and the time (s+k−1) is not stored but discarded.
By the start of the hold necessary period, the drop maximum value of the pulse at the time t is stored in the RAM 23in a updatable manner (step S59). Next, similarly to step S54, it is judged whether or not the condition of the following equation 8 is satisfied during the hold necessary period (step S60).
|yt−{circumflex over (x)}s|s-1|≤α·σv,s [Equation 8]
When the condition of the above formula 8 is not satisfied, the pulse drop maximum value is updated (steps S59, S60). When the condition of Equation 8 is satisfied, the number value is cumulatively stored in the count area of the RAM 23 at each time (steps S60, S61). Next, on the basis of the count value, it is judged whether or not the condition of the equation 8 has been satisfied by k times consecutively starting from the time s2 (step S62). If it is not k times consecutive, the process returns to step S59.
If k times consecutive, the process proceeds to step S63, and the maximum drop value of the pulse which is updated and stored at this time is stored in the RAM 23 as the estimation value of the pulse wave height value. The estimation value of the pulse wave height value is stored together with the data of the pulse start time and the pulse end time. When the estimation of the pulse wave height value is completed, it is determined that the hold necessary period is ended. By this termination, the hold end time of the hold necessary period is stored in the RAM 23 as s2 (step S64). Next, in step S65, the value of the time s is retroactively calculated for the period from the time s2 to the time (s+k−1) as the initial value at the restart of the operation process of the Kalman filter and the operation of the Kalman filter is executed After step S65, it is judged whether or not the BL estimation process of all pulse waveform data has been performed (step S66), and the process is terminated at the completion of estimation of all pulse waveform data, and when there is remaining data, the process goes to step S53.
The feature value extraction process becomes executable on condition that the extraction data of the pulse wave height value (wave height |h|) is present by execution of the BL estimation process of
The first type of feature values are calculated in steps S71 to S76. The wavelength (pulse width) Δt is sequentially calculated and stored with respect to the extracted data group of the pulse wave height value (step S71). The calculated feature value is stored in the memory area for storing the feature value of the RAM 4. The pulse width is obtained by calculating Δt (=te−ts; ts is the start time of the pulse waveform and te is the end time of the pulse waveform). The peak position ratio r is sequentially calculated and stored with respect to the extraction data group of the pulse wave height value (step S72). The peak position ratio r is calculated by r=(tp−ts)/(te−ts) (the ratio of the pulse width Δt and the time (tp−ts) leading from the pulse start to the pulse peak pp.
The peak kurtosis κ is sequentially calculated and stored with respect to the extraction data group of pulse wave height values (step S73). It is normalized so as that the wave height |h|=1, ts=0 and te=1 hold, and there are collected the time set [T]=[[ti]| i=1, . . . , m] which is the time crossing the horizontal line of 30% in wave height from the pulse peak PP, and then the K is obtained so that the dispersion of the data of the time set [T] is calculated as the pulse waveform spread.
The depression θ is obtained based on the time from pulse start to pulse peak and wave height data and the calculation of Equation 3 above (step S74). The area m is obtained from the wave height vector data, and is calculated and stored by obtaining the time division area hi according to the number of division (number of divisions set in advance: 10) and calculating the total sum thereof. The area ratio rm is calculated and stored by calculating the total waveform area and the partial sum of time division area hi in each division leading from the pulse start to the pulse peak and by calculating the area ratio of the partial sum to the total waveform area (Step S76).
The second type feature values are calculated in steps S77 to S82. The time inertia moment is obtained from the data of the wave height vector, and are calculated and stored based on the time division area hi obtained according to the number of divisions and the calculation of Equation 6 above (step S77). The normalized time inertia moment of (9) is stored as the normalization data (step S78) by the process normalized in the wave height direction (the inner product of the wave height vector and the normalized vector) so that the wave height becomes “1” of the reference value with respect to the time inertia moment obtained in step S77. The wave width mean value inertia moment is calculated from the data of the wave width vector (the difference vector of the mean value) obtained in steps S42 to S44 by using the time value calculated for each division unit (the number of divisions set in advance: 10) before and after the pulse peak and the calculation of Equation 7 and is stored (step S79). The wave width mean value inertia moment (11) is stored as the normalization data (step S80) normalized in the wavelength direction (the inner product of the difference vector of the mean value and the normalized vector) so that the wavelength becomes the reference value “1” for the wave width mean value inertia moment obtained in step S79. The wave width dispersion inertia moment is calculated from the data of the wave width vector (dispersion vector) based on the dispersion of the time value calculated for each division unit and the calculation of the above equation 7 and stored (step S81). The normalized wave width dispersion inertia moment of (13) is stored as the normalization data (step S82) normalized in the wavelength direction (the inner product of the dispersion vector and the normalized vector) so that the wavelength becomes the reference value “1” with respect to the wave width dispersion inertia moment obtained in step S81.
Upon completing the extraction of the feature value from all the data, the file of each data is stored and it is judged whether or not there is another data group (steps S83, S84). If there is the data group of another file, the above process (steps S71 to S82) can be repeatedly executed. When there is no more data to be processed, the extraction process of the feature value ends (step S85). In the above extraction process, all of the first type and the second type of feature values are obtained, but it is also possible to designate a desired feature value by designation input of the input means 6 and it is possible to extract only the feature value due to this designation.
Since the pulse waveform to be measured is not necessarily constant even for the same type of particles, the probability density function of the pulse waveform of particle type is preliminarily estimated from the test data as preparation for the particle type distribution estimation. The appearance probability of each pulse can be expressed by the probability density function derived through the estimation of the probability density function.
The (15 B) of
Since the true density function of the pulse width Δt and the pulse wave height h is unknown, it is necessary to estimate the nonparametric probability density function. In the present embodiment, the Kernel density estimation using a Gaussian function as the Kernel function is used.
Kernel density estimation is a method of assuming the probability density distribution given by Kernel function to the measurement data and regarding the distribution obtained by superimposing these distributions as the probability density function. When using a Gaussian function as the Kernel function, it is possible to assume a normal distribution for each data and to regard the distribution obtained by superimposing them as the probability density function.
The probability density function [p(x)] for the input data [x] is expressed by the following equation 9 using the number of teacher data [N], the teacher data [μi], and the variance covariance matrix [Σ].
Furthermore, the probability density function [p(x)] can be expressed by the product of the Gaussian functions of each dimension as shown in the following Equation 10.
As can be seen from Equation 10, it is equivalent to assuming that each pulse attribute is an independent random probability variable that follows the normal distribution, which can be expanded to three or more dimensions as well. Therefore, in the present embodiment, it is possible to analyze the number of two or more types of particle types.
The probability density function module program has a function of computing and obtaining the probability density function for two types of feature values. That is, in the case of using the estimation target data due to two feature values [(β, γ)], the probability density function [p(β, γ)] in the Kernel density estimation using the Gaussian function as the Kernel function is expressed by the following equation 11.
The probability density function estimation process executed by the probability density function module program based on the equation 11 performs the estimation process of the probability density function in two feature values, as described in detail later with reference to
In the present embodiment, the probability density function data (see Expression 10) obtained by estimating the probability density function of the particle type obtained in advance is stored in the RAM 23 as the analysis reference data. The particle type number analysis is performed by determining the appearance frequency of the entire data to be analyzed based on the equation 12 through identifying the number of particle types to be matched from each analysis data. The number analysis is performed by estimating histograms of different particle types (appearance frequency (number of particles) for particle type).
In the particle type estimation process shown in
<About Maximum Likelihood Estimation (the Method of Point Estimation of Population of Probability Distribution that it Follows from Data Given in Statistics)>
It is assumed that a data set [D]=[x1, x2, x3, . . . xN] has been obtained as an actual pulse estimation result. The likelihood at which the estimated j-th pulse height data appears is expressed by the following equation 13.
Then, the likelihood of appearance of the data set D is expressed by the following equation 14.
The particle type distribution that maximizes the likelihood in Equation 14 is the particle type distribution with the most likelihood value set [n]=[n1, . . . , nk]T.
<About Lagrangian Undetermined Multiplier Method (it is an Analytical Method to Optimize Under Constraint Conditions, Prepare the Undetermined Multipliers for Each Constraint Condition and Make New Functions of Linear Combinations Using these Multipliers as Coefficients (Unknown Multipliers are New Variables), and it is a Method to Solve the Constraint Problem as the Normal Extreme Problem>
Maximizing the likelihood of occurrence of data set D is equivalent to maximizing the logarithmic likelihood that data set [D] appears. The following equation 15 shows the process of deriving the logarithmic likelihood to check the suitability of the Lagrange undetermined multiplier method.
In Equation 15, the coefficient 1/NN in the middle is omitted in the final expression.
Here, the value set n=[n1, . . . , nk]T of the particle diameter number distribution has the constraint “the total is N” (see the following equation 16).
Therefore, since the proposition of obtaining the most likelihood particle type distribution becomes a problem of constrained logarithmic likelihood maximization, it is possible to perform optimization by the Lagrangian undetermined multiplier method. The constrained logarithmic likelihood maximization equation that optimizes by the Lagrange undetermined multiplier method can be expressed by the following equation 17.
From the constrained logarithmic likelihood maximization equation shown in Equation 17, [k] simultaneous equations shown in the following Equation 18 can be derived through the mathematical derivation process shown in
To solve numerically the simultaneous equations shown in Equation 18, it can be performed using the iterative method proposed by Hasselblad. According to the Hasselblad iteration method, the iterative calculation of the following Equation 19 may be performed. The details of this iterative method are described in the proposal paper (Hasselblad V., 1966, Estimation of parameters for a mixture of normal distributions. Technomerics, 8, pp. 431-444).
For the iterative calculation of Equation 19, it is performed using the software of EM algorithm available on the market. As clear from the origin of naming, the EM algorithm is the method of calculating the probability distribution parameter by maximizing the likelihood function, that is, the algorithm which can maximize the expectation of the probability distribution being the likelihood function. According to the EM algorithm, the initial value of the desired parameter is set, the likelihood (expected value) is calculated from the initial value, and in many cases, using the condition that the partial differential of the likelihood function becomes zero, the maximum likelihood parameters can be calculated by iterative calculation. In the Hasselblad iterative arithmetic operation performed by using the EM algorithm, the initial value of the parameter to be obtained is set, the likelihood (expected value) is calculated from the initial value, and furthermore, under the condition that the partial differential of the likelihood function becomes zero, the maximum likelihood parameter is calculated by carrying out the iterative calculation.
In order to execute the particle type estimation process shown in
By using the input means 6 of the PC 1, it is possible to perform the designation operations of k (2 in the embodiment) feature values for creating the data file. The combination input of the designated feature values is set in the RAM 23 (step S30). The data of the feature value data file for each feature value setting is read into the work area of the RAM 23 (step S31). The feature value data file is estimated and extracted by the BL estimation process in
The matrix data of N rows and k columns is created by designating k feature values used for the number estimation (step S32). The created matrix data is outputted to the particle type distribution estimation data file and stored for each designated feature value (step S33). Upon completion of creation of all the data files for the designated feature value, the process ends (step S34).
Data of the data file of the probability density function estimation target created in the data file creation process (step S1) is read to form the matrix [D] of N rows and 2 columns (steps S20 and S21). The dispersion shown in the following equation 20 for each row of the matrix [D] is calculated (step S22).
σD
Next, the dispersion parameter shown in the following Equation 21 is set using the standard deviation coefficient c as shown in the following Equation 22 (Step S23).
σβ2,σγ2 [Equation 21]
σβ2=c2σD
The probability density function is obtained by substituting each line of the dispersion parameter and matrix [D] as teacher data shown in the following Equation 23 and stored in the predetermined area of the RAM 23 (Steps S24 and S25). The process in steps S20 to S25 is performed until the probability density function is derived from all the process target data (step S26).
μβi,μγi [Equation 23]
First, similarly to the above-described steps S20 and S21, the data of the data file of the particle number estimation target created in the data file creation process is read, and the matrix [D] of N rows and 2 columns is created (step S10, S 11). For the matrix [D] data, the estimation process by the Hasselblad iteration method is executed (step S12).
First, after setting the initial value (process 23A), the number calculation based on the probability density function is sequentially executed (process 23B) (steps S12a and S12b). The iteration of the number calculation is executed until the convergence condition shown in (23C) is satisfied (step S12c). The execution result (the estimated number data for each particle type) of the EM algorithm is stored in the predetermined area of the RAM 23 (step S12d).
In step 4, the estimated number data for each particle type obtained by the particle number estimation process is edited into the particle type number distribution data, and according to display designation, the histogram display output to the display means 7 becomes possible. Although not shown in
The present inventors performed the verification 1 of the analysis performance of the particle type number under the following evaluation conditions using the measured current data of Escherichia coli and Bacillus subtilis in the above example.
The evaluation conditions of verification 1 are as follows.
(1) It is evaluated with the 1000 kHz experimental measurement data of Escherichia coli and Bacillus subtilis.
(2) As the feature values, four first type feature values of wavelength Δt, wave height h, peak position ratio r, and peak kurtosis k are calculated and used.
(3) The number estimation process on combinations of each feature value is performed.
(4) The actual measurement data of Escherichia coli and Bacillus subtilis are estimated and evaluated at random dividing for learning and testing. This estimation and evaluation is repeated 10 times, and the mean accuracy and standard deviation of them are calculated. In this case, the cross validation is used to evaluate the accuracy close to actual.
(5) A part of measured data of verification particles (Escherichia coli and Bacillus subtilis) are individually number-analyzed, the rest are randomly mixed with the predetermined mixture ratio δ for verification, and the number analysis results are compared. The data mixing program for the mixing random data is stored in the ROM 3, the random mixing of the data is executed using the PC1, and the number estimation for the randomly mixed data is performed. That is, in the matrix data of step S32 of
The (25A) and (25B) of
The number of all pulses obtained by this verification was 146 in E. coli and 405 in B. subtilis.
The (26A) and (26B) of
Evaluation of the particle type number can be performed by “weighted mean relative error” represented by the equation shown in (27B) of
The (27A) in
The (28A) and (28B) of
The (29A) to (29D) of
The (30A) to (30D) of
The (31A) and (31B) of
The (32A), (32B) and (32C) of
From the above performance evaluation experiments, the following evaluation results were obtained.
(1) In the data scatter diagrams of
(2) From the estimation result of the type number distribution shown in (27 A) of
Using the measured current data of E. coli and B. subtilis in the above example, the verification 2 of the analysis performance of particle type number different from verification 1 is performed. In the verification 2, unlike the verification 1, the feature values of the first type and the second type (13 types of (1) to (13)) are calculated and used, and there were verified the relationship between the feature value and the number of sampling data, and the analysis performance according to each combination.
The (42A) and (42B) in
The estimation evaluation results for each combination in these tables are obtained by the cross validation method in the same as (4) of verification 1. The mean accuracy is described in the upper side and the standard deviation indicated in parenthesis is in the lower side. The inertia I, inertia I (normalization), inertia I_w, inertia I_wv, inertia I_w (normalization), and inertia I_wv (normalization) in the table, respectively, show the feature values as the time inertia moment of (8), the normalized time inertia moment of (9), the wave width mean value inertia moment of (10), the wave width dispersion inertia moment of (12), the normalized wave width mean value inertia moment of (11), and the normalized wave width dispersion inertia moment of (13).
The values on the vertical axis in
The results obtained from Verification 2 are as follows.
(R1) As shown in
(R2) As shown in
(R3) As shown in
(R4) As can be seen from (R1) to (R3), the highly accurate number estimation can be carried out even by using the combination of feature value of the first type and the second type. Furthermore, according to the number analyzing method of the present invention, even if the sampling number is not sufficiently large, when the predetermined sampling number can be obtained, the number analysis can be performed with the same accuracy as when it is sufficient. For example, in the combination of the kurtosis k and the peak position ratio r examined in Verification 1, a maximum error of 12% was generated, but for example, in the case of using the feature value of the wavelength Δt—inertia I, it is possible to perform the number estimation process with high accuracy of about 9% using the high density sampling data at 1 MHz to 125 kHz even if all data is not used, that is, it is partial data. Therefore, the number analysis function according to the present embodiment can be applied not only to the stationary number analysis, but also to the quarantine inspection and the medical site requiring urgency, so that it can be used as a suitable inspection tool that can be implemented quickly for judgement of the presence or absence of particles or the number of bacteria or the like.
Since in the number estimation, the calculation time is required for the iterative calculation due to the Hasselblad method, the comparison and examination of the feature values are verified in verification 3 with respect to the relation between the required calculation time and the sampling frequency. In the comparative examination example of Verification 3, there are used four kinds of the feature value combinations such as the wavelength Δt—area m, the wavelength Δt—inertia I, the kurtosis k—wave height |h| and the kurtosis k—peak position ratio r shown in (51B) of
As shown in (52 A), the feature value combination G1 of the wavelength Δt—area m and the wavelength Δt—inertia I is almost the same total calculation time, and the feature value combination G2 of the kurtosis k—wave height |h| and the kurtosis k—the peak position ratio r has approximately the same total calculation time. As shown in (52 B), the calculation time required for generating each feature value of the feature value combination G1 is the same, and the calculation time required for generating each feature value of the feature value combination G2 is the same. As shown in
Obviously from the comparison result of the feature value combinations G1 and G2 of verification 3, even if it is the same type combination in the first type and the second type, even if it is the different mixing combination, the required calculation time using the feature value can be shortened. Therefore, according to the number analyzing device of the present embodiment, in addition to the stationary number analysis, for example, it is possible to quickly perform the process of discriminating the presence or absence of particles and the number of bacteria or the like in the quarantine inspection or the medical field requiring urgency.
As can be understood from the above performance evaluation, based on the data group of the detection signal detected by the nanopore device 8, there is executed the particle type distribution estimation program which is the number deriving means in the computer control program (number analysis program), and it is possible to perform the probability density estimation from the data group based on the feature value showing the feature of the waveform of the pulse signal corresponding to the particle passage obtained as the detection signal and it is possible to derive the number of the particle type. Therefore, by using the number analyzing device, it is possible to analyze the number or the number distribution corresponding to the type of analyte such as, for example, bacteria, microparticulate material, etc. with high accuracy, so simplification and cost reduction in the number analyzing inspection can be realized. By incorporating the detection signal from the nanopore device 8 directly into the number analyzing device so that data can be stored, the particle type integration analyzing system integrating inspection and analysis may be constructed.
The probability density estimation is performed from the data group based on the feature value, and the result of deriving the particle type number is displayed on the display means 7 as the output means or printed out on the printer. Therefore, according to the present embodiment, highly accurate derivation results (particle number, particle number distribution, estimation accuracy, etc.) can be notified promptly in the output form of, for example, the histogram or the dispersive diagram, so that for example, it is possible to use the number analysis function according to the present embodiment as the useful inspection tool in the quarantine portion or the medical field requiring urgency.
The present invention can be applied not only to a computer terminal such as a specific PC or the like mounted with a identification process program but also to a storage medium for identification analysis which stores a part or all of the identification process program. That is, since the identification analysis program stored in the identification analysis storage medium can be installed in a predetermined computer terminal and the desired computer can be operated to perform the identification analysis operation, it is possible to carry out the identification analysis simply and inexpensively. As the storage medium applicable to the present invention, there are a flexible disk, a magnetic disk, an optical disk, a CD, an MO, a DVD, a hard disk, a mobile terminal, or the like, and any storage medium readable by a computer can be selected and used.
The computer analysis unit 1a in
For example, when E. coli Ec and B. subtilis Bs are used as analytes of which particle types are specified (specific analytes), each of the specific analytes is measured by the nanopore device 8a, and each pulse signal data is input to the PC 1 as known data, and the input data is stored in a memory area for storing known data in the RAM 4. When the content state of the specific analytes is unknown in the analysis target, data of the pulse signal obtained by performing the measurement by the nanopore device 8a is input as analysis data to the PC 1, and the input data is a memory area for storing analysis data in the RAM 4
When the classification analysis process is activated by the activation operation, it is determined whether or not the known data is input (step S110). When the known data has not been input, the display means 7 displays a guidance prompting the user so as to input the known data. In
When the known data is input, it is determined whether or not the feature value is specified (steps S110 and S111). When the feature value is designated, the vector value data of the feature value designated from the feature value storage data file DA based on the known data of the RAM 4 is taken into the learning data storage area of the RAM 4 (step S113). When the feature value is not specified, the vector value data of all the feature values are taken into the learning data storage area of the RAM 4 from the feature value storage data file DA based on the known data of the RAM 4 (step S112).
Next, it is determined whether or not the analysis data is input (step S114). When there is no analysis data input, the display means 7 displays a guidance prompting input of the analysis data. When the analysis data is input, the acquired analysis data is stored in the analysis data storage memory area in the RAM 4 (step S100). When the analysis data is input, as described, the feature value related to the analysis data is created and stored in the RAM 4 (step S101). When the analysis data is input, the vector value data of the feature value is fetched from the feature value storage data file DB based on the analysis data of the RAM 4 into the variable data storage area of the RAM 4 (step S115).
In the state of acquisition of feature values under the input of the known data and the analysis data, the guidance display for prompting execution of the classification analysis is performed. By performing the predetermined instruction operation in accordance with the guidance display, the classification analysis program is activated and the execution process of the classification analysis due to the machine learning is performed (step S116). In the present embodiment, for example, the machine learning classification analysis program configured by the algorithm based on Random Forest Method is stored in advance in the ROM 3.
The feature value due to the known data is set as the learning data, the feature value obtained from the analyzed data is set as a variable and by executing the classification analysis program, the classification analysis relating to the specified analyte in the analyzed data can be performed. At the time of execution of the classification analysis program, the pulse waveforms are converted into the numerical vectors of the same dimension, and the classification analysis is performed by identifying the individual pulses and determining how each vector is different.
The classification analysis method due to the machine learning according to the present invention is not limited to Random Forest Method, and, for example, it is possible to use the group learning such as K-nearest Neighbor Algorithm, Naïve Bayes Classifier, Decision Tree, Neural Network, Support Vector Machine, Bagging Method, and Boosting Method.
When the execution process of classification analysis by machine learning is executed for all of the feature values by analysis data, the classification analysis process is finished, and the output process of the classification analysis result is performed (step S117). In the output process, for each of analysis data of unknown type, the display means 7 can display a classification result, and as an example of the specific analyte, it is displayed a ratio of those derived from the passage of E. coli Ec or B. subtilis Bs. The display mode that can be output is not limited to the classification result for each analysis data, and the display mode such as the corresponding total number of the analyte (for example, E. coli Ec or B. subtilis Bs) and the corresponding ratio of both can be used.
About the process accuracy of the above-mentioned classification analysis process, the classification analysis is tried by applying the analysis method based upon various machine learnings and the accuracy of the classification analysis process due to the present embodiment is verified.
In (57A) of
The analysis sample is two kinds of bacterial species (E. coli, Bacillus subtilis) as shown in (57B). For each bacterial species, the pulse shapes are obtained by measuring the passing waveform using the micro-nanopore device 8 with the inner diameter of through-hole 12 of 4.5Φ and the penetration distance (pore depth) of through-hole 12 of 1500 mm and the 42 signal data (all of the measurement pulses in the case of E. coli and 42 pulses out of 265 measurement pulses in the case of B. subtilis) are used. During execution of the classifier, about 90% of the pulse signal data are used as the learning data and the remaining data are assigned to the variables.
As shown in (57A), the evaluation items are represented by F-measure (F-Measure), which consists of true positive rate (TPRate), false positive rate (PRate), precision rate (Precision), recall rate (Recall), F value (FMeasure) and receiver operating characteristic curve area (ROC (Receiver Operating Characteristic) Curve Area).
As shown in (58A), for the real numbers of the two bacterial species (E. coli real number: P, B. subtilis real number: N), when the predicted value of each bacterial species is assigned, the F-measure is expressed by 2TP/(2TP+FP+FN) as shown in (58B), assuming that the sum of true positive (TP), false positive (FP), true negative (FN) and false negative (TN) in each combination is 1.
In this verification, the classification analysis is attempted on about 4000 patterns by using the 67 kinds of classifiers with different algorithms and using various feature values or combinations of feature values. As a result, the significant analysis results are obtained for the combinations of 60 kinds of feature values. The (57A) of
As shown in (57A), the feature values in the top ten places include the 13-dimensional feature value vector (abbreviated as ┌hv&F┘ in the table) arranging 13 kinds of feature values of (1) to (11), (14) and (15), a combination (abbreviated as ┌h&wV┘ in the table) of the wave height vector (abbreviated as ┌h┘ in the table) and the mean value vector of (10) (abbreviated as ┌wV┘ in the table), and a combination (abbreviated as ┌h&wNrmdV┘ in the table) of the wave height vector and the normalized mean value vector of (11) (abbreviated as ┌wNrmdV┘ in the table). The most excellent classification accuracy in (57A) is the case of the Random Forest Method classifier (┌4 meta. Random Committee┘) using a combination of h&wV as the feature value, and the classification accuracy shows the high accuracy of approximately 98.9%.
The present invention is not limited to a computer terminal such as the specific PC equipped with the classification analysis program, and can be applied to the classification analysis storage medium storing a part or all of the classification analysis program. That is, since the classification analysis program stored in the storage medium for classification analysis can be installed in the predetermined computer terminal and the classification analysis operation can be performed on the desired computer, the classification analysis can be performed simply and inexpensively. As the storage medium to which the present invention can be applied, any of computer-readable storage medium such as flexible disk, magnetic disk, optical disk, CD, MO, DVD, hard disk, and mobile terminal can be selected and used.
It is to be understood that the present invention is not limited to the above-described embodiments, but includes various modifications, design changes and the like within the technical scope without departing from the technical idea of the present invention.
According to the present invention, since the identification of the nonconforming data and the classification analysis are performed with high accuracy, it is possible to be utilized to an information compression technology of DNA storage media and a drug discovery using an artificial base pairing, or in the fields of a fine dust mixed in a measurement sample and an analysis substance contained in a body fluid, it can be applied and developed in the fields of the identification/removal technologies of the nonconforming data derived from the fine substances such as red blood cells, white blood cells and platelets. In particular, the present invention can be applied to data analysis in a detection technique in which a sample containing DNA or RNA and contaminants is analyzed, for example, for performing DNA content analysis in sewage to detect the occurrence of a virus.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2017-092075 | May 2017 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2018/014926 | 4/9/2018 | WO | 00 |
