WO2023106342A1

WO2023106342A1 - Method and apparatus for detection, identification, and quantification of fine particles

Info

Publication number: WO2023106342A1
Application number: PCT/JP2022/045152
Authority: WO
Inventors: 典彦直野; 弘泰武居
Original assignee: アイポア株式会社
Priority date: 2021-12-08
Filing date: 2022-12-07
Publication date: 2023-06-15
Also published as: JP2023085018A

Abstract

This apparatus is characterized by comprising a Coulter measuring device configured to detect as a pulse signal a transient change in an ionic current between two electrodes that are in contact with an electrolytic solution on both sides of a fine hole, said transient change occurring when detection target particles dispersed in the electrolytic solution pass through the fine hole, a feature quantity extracting unit for calculating a feature quantity of a pulse waveform that accompanies the passage of the particle through the fine hole, an incidental measurement condition storage unit for storing an incidental measurement condition of the fine hole, and an AI program which learns the pulse waveform, the apparatus being configured such that: a first incidental measurement condition including at least one selected from the group comprising a hole diameter, shape, and fine hole thickness of a first fine hole of a first Coulter measuring device used to measure a known particle of a known type is stored in the incidental measurement condition storage unit; the feature quantity extracting unit calculates a first feature quantity from a first pulse waveform obtained from the first Coulter measuring device in conjunction with the passage of the known particle through the first fine hole; a trained AI program is created by training the AI program using the first feature quantity and the first incidental measurement condition as teacher data, and using the type of the known particle as a teacher label; a second incidental measurement condition including at least one selected from the group comprising a hole diameter, shape, and fine hole thickness of a second fine hole of a second Coulter measuring device used to measure an unknown particle is stored in the incidental measurement condition storage unit; and detection, identification, or quantification of the detection target particle is performed by inputting into the trained AI program the second incidental measurement condition and an unknown feature quantity calculated by the feature quantity extracting unit from an unknown pulse waveform obtained from an output of the second Coulter measuring device in conjunction with the passage of the unknown particle through the second fine hole.

Description

Method, device for detection, identification and quantification of microparticles

The present invention relates to devices and methods for detecting, identifying and quantifying microparticles in samples.

In the diagnosis, treatment, and prevention of diseases, techniques for detecting, identifying, and quantifying biological microparticles such as cells, microorganisms, viruses, or exosomes dispersed in solutions have become essential in recent years. One of the most widely used techniques in clinical practice is flow cytometry.
In the flow cytometry method, microparticles flowing in a line in a capillary are irradiated with excitation light, fluorescence and scattered light emitted from the microparticles are detected by multiple detectors, and based on the output data of the detectors Detect and classify microparticles. In recent years, in order to analyze microparticles in more detail, a method of acquiring spatial distribution information of signals emitted by microparticles (Patent Document 1), labeling an object with a plurality of fluorescent dyes, and simultaneously detecting fluorescence of different wavelengths A method of detecting and analyzing (Patent Document 2) is used.

The scattered light intensity of microparticles varies depending on the size and internal state of the particles, and the fluorescence intensity depends on the amount of the antigen to be labeled held by the particles. The peak value of scattered light or fluorescence obtained from one particle is plotted on a two-dimensional scatter diagram, and based on that information, the types of microparticles are classified one by one. Therefore, for accurate sample analysis, it is necessary to define appropriate discrimination boundaries for scattered light and fluorescence intensity (Patent Document 3).

For example, in the analysis of blood cell populations, discrimination boundaries are first set based on the scattering properties of the cells. Forward and side scatter can provide approximate cell size and granularity information. Since red blood cells are smaller than white blood cells, both forward and side scattering intensities are low. Among white blood cells, monocytes and lymphocytes clearly differ in forward scattering light intensity, while granulocytes have high side scattering intensity. . Furthermore, by extracting a population of lymphocytes and analyzing their fluorescence intensity, it is possible to further classify them into subpopulations based on surface markers. As described above, the flow cytometry method is a technique that enables various analyzes by combining excitation light, scattered light emitted from microparticles, and fluorescence intensity, wavelength, and angle.

On the other hand, as another technique for detecting, identifying, and quantifying biological microparticles such as cells, microorganisms, viruses, or exosomes, target microparticles are suspended in an electrolytic solution, and particles are generated by electrophoresis. A pore electrical resistance method, the so-called Coulter method, which detects a transient change in ion current accompanying passage through a pore is sometimes used (Non-Patent Document 1). In recent years, high-precision thin-film pores using semiconductor photolithography have significantly improved the accuracy of particle identification, and it has become possible to identify clinical specimens containing bioparticles as small as 100 nm with high accuracy. (Non-Patent Document 2). In the analysis of a blood cell population, as in the flow cytometry method, the particle size coefficient and size are estimated from the signal intensity associated with the passage of cells through pores (Patent Document 4). Compared to the flow cytometry method, it is possible to analyze objects smaller than the wavelength of light, and the simple configuration of the device facilitates miniaturization and cost reduction of the device.

As described above, the flow cytometry method and the Coulter method are useful methods for detecting and classifying a large amount of microparticles and analyzing the amount of particles present, regardless of the purity of the object. A common feature of these measuring means is that each time one particle passes through the measuring section, one pulse-like signal is observed, and the distribution of the pulses is the object of analysis. These methods are widely used for blood cell counting and particle size measurement in blood.

WO2017/073737 Japanese Patent Application Laid-Open No. 2021-143988 JP-A-2009-236798 Japanese translation of PCT publication No. 2000-503772

As a means of counting fine particles present in a biological sample, the Coulter method, which measures transient changes in electrical resistance due to passage of fine particles through pores, and the scattered light of fine particles that pass through a narrow tube in a single line, are irradiated with a laser beam. Flow cytometric methods are used to measure transient changes in Each of these methods is a technique of acquiring transient changes in electrical resistance or scattered light intensity as a pulse signal, and detecting, identifying, and quantifying fine particles by analyzing the pulse signal. In analysis by conventional methods, the height and width of each pulse generated by measuring each particle are expressed in a scatter diagram. Identify and quantify. In the following, detection refers to confirming the presence or absence of particles to be detected (cells, microorganisms, and proteins) in the sample, identification refers to distinguishing between a plurality of particles to be detected in the sample, and Estimating the concentration of is called quantification, respectively.

There are five problems related to pulse analysis in the conventional Coulter method and flow cytometry. The first is the overlap of widths and heights between different grain types, the second is the cluster shift caused by variations in each measurement value, the third is the presence of contaminant particles other than the particles to be measured, and the fourth is the presence of contaminant particles in the biological sample. There are difficulties when the number of particles to be measured is small, and the fifth point is that small particles such as proteins cannot be measured.

The first problem is that even if the measurement results are plotted on a scatter diagram, if the overlap is large, the identification boundaries cannot be determined, and identification and quantification cannot be performed. FIG. 1 shows a conceptual diagram of measurement results of a sample in which three types of particles to be measured, ie, first, second, and third particle types are mixed. FIG. 1 represents the first grain type 110, the second grain type 120 and the third grain type 130, each with a differently shaped plot. One plot represents one pulse. A first axis 101 and a second axis 102 are quantities representing characteristics of each pulse obtained as a measurement result. For example, in the Coulter method or flow cytometry, the first axis may be pulse width (duration) and the second axis may be pulse height (peak power value). In the case of flow cytometry using two sensors, the first axis may be the pulse height of the first sensor, and the second axis may be the pulse height of each second sensor. For example, in the case of five types of leukocytes (neutrophils, lymphocytes, monocytes, eosinophils, and basophils) by flow cytometry, the scatter plots of the scattered light or fluorescence peak intensities are well separated. It is easy to distinguish between However, in both the Coulter method and flow cytometry according to the prior art, different types of particles having similar particle sizes generally have unclear clusters, as shown in FIG. Therefore, it is difficult to distinguish between grain types, and it is also difficult to detect, identify, and quantify the grains.

The second problem is that even if the clusters can be classified, if the variation between measurements is large, the pulses are erroneously identified, resulting in large errors in the case of quantification. FIG. 2 shows a conceptual diagram of measurement results of a sample in which three types of particles to be measured, ie, first, second, and third particle types are mixed. A first axis 201 and a second axis 202 are each a quantity that characterizes each pulse obtained as a result of the measurement. FIG. 2(a) shows the result of preliminary measurement for determination of the identification boundary, and the first grain type 210, the second grain type 220 and the third grain type 230 are plotted in different forms. On this scatterplot, the measurements for each particle form clusters and are relatively well separated. FIGS. 2(b) and 2(c) schematically show measurement results of measurement target samples of unknown grain types. After determining the identification boundary 200 of the three grain types in FIG. They can each be classified as a second particle and, if within the third region 239, a third particle. However, in reality, the plot for each grain type shifts greatly on the scatter plot. For example, in the measurement result of an unknown sample in FIG. 2B, there are many misidentifications 221 in which particles belonging to the cluster of the first grain type are classified as the second grain type. Further, in the measurement result of the unknown sample in FIG. 2(c), misidentification 231 often occurs in which the particles belonging to the cluster of the third grain type are classified as the second grain type. As described above, in the conventional technology, the accuracy of detection and identification deteriorates due to inter-measurement variations in measurement results. Moreover, when the above three grain types are quantified based on this data, the accuracy is greatly impaired.

The third problem is that it is difficult to separate the pulses derived from the particles to be measured and the pulses derived from particles other than the particles to be measured, that is, contaminants. The measurement in FIG. 1 assumes the measurement result of a sample consisting of only the first to third grain types, but in the measurement of an actual biological sample, it is common that some particles to be measured are present in contaminants. . In general, contaminants are composed of many kinds of particles, and their origins are unknown. FIG. 3 is a conceptual diagram of such measurement results. A first axis 301 and a second axis 302 are quantities that characterize each pulse obtained as a measurement result. FIG. 3(a) shows measurement results of a sample containing particles to be measured, and FIG. 3(b) shows measurement results of a sample not containing particles to be measured. Contaminant particles are distributed in a region 311 in FIG. 3A, and particles to be measured are distributed in a region 312, respectively. In this case, the contaminant particles are unknown particles and consist of a mixture of particles of various sizes and properties. Therefore, with conventional techniques, it is difficult to extract and quantify only particles to be measured from the measurement results of such samples, or to detect the presence or absence of particles to be measured for each sample. For example, the former corresponds to the case of quantifying a trace amount of viruses, bacteria, or proteins from a specimen sample, and the latter corresponds to the case of detecting whether or not a specific virus, bacteria, or protein is contained in a specimen sample. Both of these technologies are highly demanded for practical use, but conventional technologies have not been able to exhibit sufficient performance.

The fourth problem is that in the case of a sample containing only a few particles to be measured in contaminants, it is difficult to detect, identify, and quantify the particles. For example, in the schematic diagram of FIG. Have difficulty. Not only in the Coulter method and flow cytometry, but in general, the most important performance index in practical examination of biological samples is the sensitivity of particles to be measured, and the solution of the fourth problem is strongly desired.

A fifth problem is that the Coulter method and flow cytometry cannot detect, identify, and quantify small particles of 10 nm or less, such as proteins. Since flow cytometry uses light scattering, it is difficult to measure scattered light pulses from such small particles. In the case of the Coulter principle, it is theoretically possible to measure by using a device with pores of about 10 nm. ELISA (Enzyme-Linked Immuno Sorbent Assay), which is used for protein quantification, requires a long measurement time and an expensive apparatus. There is a demand for a technology that enables fast, low-cost detection, identification, and quantification of proteins.

The present invention has been made in view of such circumstances, and can provide the following aspects.

configured to detect, as a pulse signal, a transient change in ionic current between two electrodes in contact with the electrolytic solution on both sides of the pore when particles to be detected dispersed in the electrolytic solution pass through the pore. a coulter measuring device that
a feature quantity extraction unit that calculates a feature quantity of a pulse waveform associated with passage of the particles through the pore;
an incidental measurement condition storage unit that stores incidental measurement conditions of the pores;
An AI program that learns the pulse waveform,
A first attachment containing one or more selected from the group consisting of the hole diameter, shape, and thickness of the first pore of the first pore of the first Coulter measurement device used for measuring known particles of known type storing measurement conditions in the incidental measurement condition storage unit;
The feature amount extraction unit calculates a first feature amount from a first pulse waveform obtained from the first Coulter measurement device as the known particles pass through the first pore,
creating a learned AI program by learning the AI program using the first feature quantity and the first incidental measurement condition as teacher data and the type of the known particle as a teacher label;
A second incidental measurement condition including one or more selected from the group consisting of the hole diameter, shape, and pore thickness of the second pore of the second Coulter measurement device used for measuring unknown particles, stored in the incidental measurement condition storage unit,
The unknown feature value calculated by the feature value extractor from the unknown pulse waveform obtained from the output of the second Coulter measuring device as the unknown particle passes through the second pore and the second incidental measurement. A device characterized by being configured to input conditions into the learned AI program to detect, identify, or quantify the particles to be detected.

A laser beam is irradiated to a measurement part including a transparent flow path through which particles in a sample pass in a row, and a pulse signal from scattered light or fluorescence from particles passing through the measurement part each time one particle passes through the measurement part. a flow cytometry instrument configured to obtain a
a feature amount extraction unit that calculates a feature amount of a pulse waveform obtained from a pulse signal obtained as the particles pass through the measurement unit;
an incidental measurement condition storage unit for storing incidental measurement conditions of flow cytometry measurement means;
An AI program that learns the pulse waveform,
First light source characteristic information, first light receiver angle information, first flow velocity information, and first fluorescent label characteristics representing the characteristics of the first flow cytometry measurement means used for measuring known particles of known types storing a first incidental measurement condition including one or more selected from the group consisting of information and first sheath liquid physical property information in the incidental measurement condition storage unit;
calculating a first feature quantity from a first pulse waveform obtained from the first flow cytometry measurement means as the known particles pass through the first measurement unit of the first flow cytometry measurement means; death,
creating a learned AI program by learning the AI program using the first feature amount and the first incidental measurement condition as teacher data and the type of the known particle as a teacher label;
Second light source characteristic information, second light receiver angle information, second flow velocity information, and second fluorescent label representing the characteristics of the second flow cytometry measuring means used in the second measurement used to measure unknown particles storing second incidental measurement conditions including one or more selected from the group consisting of characteristic information and second sheath liquid physical property information in the incidental measurement condition storage unit;
a second feature amount calculated by the feature amount extraction unit based on a second pulse waveform obtained by the unknown particle passing through the second measurement unit of the second flow cytometry measurement means; An apparatus configured to input the second contingent measurement condition into the learned AI program to detect, identify, or quantify the particles.

An AI program for identifying first type particles and second type particles mixed in a sample,
The instructions in the AI program, when executed by a processor, include:
A transient change in ionic current between two electrodes in contact with the electrolyte on both sides of the pore when one particle dispersed in the electrolyte passes through the pore is detected as one pulse signal. a pore device comprising:
A feature quantity extraction unit that calculates N feature quantity groups for each pulse waveform;
Using a Coulter measuring device equipped with an AI program that learns pulse waveforms,
The first pulse waveform obtained by measuring the first sample with the Coulter measuring device is clustered in an N-dimensional feature amount space based on the feature amount vector having the feature amount group calculated for each of the first pulse waveforms. classifying into a first estimated pulse waveform group estimated to be particles of the first type and a second estimated pulse waveform group estimated to be particles of the second type;
The second pulse waveform obtained by measuring the second sample with the Coulter measurement device is clustered in an N-dimensional feature amount space based on the feature amount vector having the feature amount group calculated for each of them as an element. classifying into the first estimated pulse waveform group and the second estimated pulse waveform group,
Each of the feature quantity groups calculated from the first estimated pulse waveform group is used as teacher data, the first type is used as a teacher label, and the feature quantity group calculated from the second estimated pulse waveform group is An AI program characterized by being configured to learn the AI program using each as teacher data and the second type as a teacher label, and to create a learned AI discriminator.

An AI program for estimating the presence or absence of a protein to be detected in a sample,
The AI program is
By mixing anti-protein beads modified with an antibody that specifically binds to the detection target protein into a sample known to contain the detection target protein and contaminants, the detection target protein is bound therein. a first mixed sample prepared containing bound beads;
using a second mixed sample prepared by mixing the anti-protein beads with a sample known to contain the contaminants but not the protein to be detected;
configured to detect transient changes in ionic current between two electrodes in contact with the electrolyte on both sides of the pore as a pulse waveform when particles dispersed in the electrolyte pass through the pore. used in measurements by Coulter measuring devices,
In the measurement,
The first mixed sample is measured by the Coulter measuring device to obtain a first pulse measurement result, and the second mixed sample is measured by the Coulter measuring device to obtain a second pulse measurement result. can be,
The instructions in the AI program, when executed by a processor, include:
calculating a contamination pulse probability that the first pulse measurement result is not the bound bead by calculating the first pulse measurement result and the second pulse measurement result;
Learning the AI program using the pulse measurement result of the pulse whose probability of contamination is below the threshold as teacher data and positive as a correct label,
An AI program that is configured to learn the AI program using the second pulse measurement result as teacher data and negative as a correct label to create a trained AI detector.

A method using an AI program for detecting whether or not a microorganism to be detected is contained in a sample,
A bead-mixed negative sample is prepared by mixing antimicrobial beads modified with an antibody that specifically binds to the microorganism to be detected to a known negative sample that does not contain the microorganism to be detected,
Mixing the antimicrobial beads with a known positive sample containing the microorganism to be detected to prepare a bead-mixed positive sample containing microorganism-bound beads to which the microorganism to be detected and the antimicrobial beads are bound,
A Coulter device configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of a pore when fine particles of the bead-mixed negative sample dispersed in an electrolytic solution pass through the pore. By measuring with, a negative pulse measurement result is obtained,
A positive pulse measurement result is obtained by measuring the bead-mixed positive sample with the Coulter device,
Using an AI program, AI learning is performed using the negative pulse measurement result as teacher data and the negative label as the correct answer,
A method comprising using the AI program to perform AI learning using the positive pulse measurement result as teacher data and the positive label as a correct answer to create a trained AI detector.

A method using an AI program for detecting whether the sample contains exosomes to be detected,
Create a bead-mixed negative sample by mixing anti-exosome beads modified with an antibody that specifically binds to the exosomes to be detected in a known negative sample that does not contain exosomes to be detected,
The anti-exosome beads are mixed with the known positive sample containing the exosomes to be detected, and the exosomes to be detected and the anti-exosome beads are bound to create a bead mixed positive sample containing exosome-bound beads,
A Coulter device configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of a pore when fine particles of the bead-mixed negative sample dispersed in an electrolytic solution pass through the pore. By measuring with, a negative pulse measurement result is obtained,
A positive pulse measurement result is obtained by measuring the bead-mixed positive sample with the Coulter device,
Using an AI program, AI learning is performed using the negative pulse measurement result as teacher data and the negative label as the correct answer,
A method comprising using the AI program to perform AI learning using the positive pulse measurement result as teacher data and the positive label as a correct answer to create a trained AI detector.

A method for detecting whether a target protein is contained in a sample, comprising:
Configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal only when detectable particles having a size equal to or larger than the detection limit size dispersed in the electrolytic solution pass through the pore. with a Coulter device that
preparing a first sample by mixing anti-protein beads that specifically bind to the protein to be detected with a negative sample that does not contain a protein having a size less than the detection limit size;
Mixing the anti-protein beads with the positive sample containing the protein to create a second sample containing bound beads in which the anti-protein beads are bound to each other;
A first pulse number of the first pulse measurement result obtained by measuring the first sample with the Coulter device, and a second pulse measurement result of the second pulse measurement result obtained by measuring the second sample with the Coulter device Determine the positive pulse number threshold from the number of pulses,
For the unknown sample in which the third pulse number in the third pulse measurement result obtained by measuring the unknown sample in which it is unknown whether or not the protein to be detected is contained is above the positive pulse number threshold, A method comprising determining that the protein to be detected is included.

A method using an AI program for detecting whether or not a microorganism to be detected is contained in a sample,
A bead-mixed negative sample is prepared by mixing labeled beads modified with an antibody that specifically binds to the microorganism to be detected to a known negative sample that does not contain the microorganism to be detected,
Centrifuging the negative sample container containing the bead-mixed negative sample,
The sediment in the lower part of the negative sample container is configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of the pore when fine particles dispersed in the electrolytic solution pass through the pore. By measuring with the Coulter device that is used, a concentrated negative pulse measurement result is obtained,
Mixing the labeled beads with a known positive sample containing the microorganism to be detected to prepare a bead-mixed positive sample containing microorganism-bound beads to which the microorganism to be detected and the labeled beads are bound,
Centrifuging the positive sample container containing the bead-mixed positive sample,
By measuring the sediment at the bottom of the positive sample container with the Coulter device, a concentrated positive pulse measurement result is obtained,
Using an AI program, AI learning is performed using the concentrated negative pulse measurement result as teacher data and negative as a correct label,
using the AI program to create a trained AI detector by performing AI learning using the enriched positive pulse measurement results as teacher data and using positive as a correct label.

A method using an AI program for identifying which of two types of microorganisms are present in a sample, comprising:
A first sample known to contain a first microorganism, a first antimicrobial bead modified with an antibody that specifically binds to said first microorganism and an antibody that specifically binds to a second microorganism A first mixed bead sample containing the first microorganism-bound beads bound to the first microorganism and the first anti-microbial bead by mixing two kinds of mixed beads consisting of the second anti-microbial beads modified with and create
A second sample known to contain a second microorganism is mixed with the two-species mixed beads to form a second microorganism-bound bead in which the second microorganism and the second antimicrobial bead are bound. creating a second bead mixture sample comprising
The first bead mixture sample is configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal when the fine particles dispersed in the electrolyte pass through the pore. A first pulse measurement result is obtained by measuring with a Coulter device,
obtaining a second pulse measurement result by measuring the second bead-mixed sample with the Coulter device;
Using an AI program, AI learning is performed using the first pulse measurement result as teacher data and the first microorganism as a teacher label,
A method comprising using the AI program to perform AI learning using the second pulse measurement result as teacher data and the second microorganism as a teacher label to create a learned AI discriminator.

A method using an AI program for quantifying the concentration of protein contained in a sample, comprising:
A first sample known to contain the protein at a first concentration is mixed with labeled beads modified with an antibody that specifically binds to the protein to form a first sample containing a conjugate of the protein and the labeled beads. Create a bead mixture sample,
mixing the labeled beads with a second sample known to contain a second concentration of the protein to form a second bead mixture sample containing a conjugate of the protein and the labeled beads;
The first bead mixture sample is configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal when the fine particles dispersed in the electrolyte pass through the pore. Obtaining a measurement first teacher pulse measurement result extracted by the pulse extraction means from the measurement result by the Coulter device,
Obtaining a second teacher pulse measurement result extracted by the pulse extraction means from the measurement result of the second bead mixed sample by the Coulter device,
Using an AI program, AI learning is performed using the first teacher pulse measurement result as teacher data and the first concentration as a correct label,
and creating a learned AI quantifier by performing AI learning using the AI program with the second teacher pulse measurement result as teacher data and the second concentration as a correct label.

According to the present invention, cells such as blood cells, microorganisms such as viruses and bacteria, or microparticles such as proteins in a biological sample are detected and detected using measurement means that generate pulses as a result, such as the Coulter method and flow cytometry. Identification and quantification can be achieved with higher precision than in the prior art.

It is a figure explaining the measurement result in the case of the sample in which three types of measurement object particle|grains of the 1st, 2nd, and 3rd particle types are mixed, and the problem in the prior art. It is a figure explaining the measurement result in the case of the sample in which three types of measurement object particle|grains of the 1st, 2nd, and 3rd particle types are mixed, and the problem in the prior art. FIG. 2 is a diagram for explaining measurement results of a sample containing contaminant particles and problems in the prior art. 1 shows an example of a measuring device that constitutes a fine particle measuring apparatus using the Coulter method used in an embodiment of the present invention. FIG. 7 shows an example of a pulse waveform obtained by the apparatus of FIG. 4 or FIG. 6, etc. FIG. 1 shows an example of a measurement device that constitutes a flow cytometry particle measurement apparatus used in an embodiment of the present invention. 1 shows an example configuration of a detection, identification or quantification device according to an embodiment of the invention. 4 is a flow chart showing an example of AI program learning according to an embodiment of the present invention, and an unknown sample detection/discrimination/quantitation process using the same. It is a figure explaining the process of extracting the feature-value of a pulse from a pulse waveform. An example of padding after the end of a pulse as AI data preprocessing for a deep neural network is shown. An example of a feature value obtained from a pulse and an example of expressing it in p-dimensional space are shown. An example of teacher label assignment to measurement results using a Coulter device for learning feature values is shown. An example of supervised label assignment to measurement results using flow cytometry for learning feature values is shown. FIG. 10 is a conceptual diagram of the result of calculating each discriminant boundary for each measurement; Another example of supervised labeling for identifying microorganisms is shown. 1 is a diagram schematically showing a deep neural network; FIG. 4 shows an example of result output by the AI program according to the embodiment of the present invention. Fig. 3 shows an example configuration of a detection, identification or quantification device according to another embodiment of the invention; Fig. 3 shows an example configuration of a detection, identification or quantification device according to yet another embodiment of the present invention; FIG. 4 is a diagram schematically showing the pulse number ratio in the measurement results of a biological sample containing particles to be measured and a biological sample not containing particles to be measured. An example of a measurement result in which contaminants and particles to be measured are mixed in a sample to be measured and a discrimination boundary cannot be determined, and a measurement result of a sample containing only contaminants without particles to be measured will be shown. When the mixing ratio a2000 of the 220 nm beads, which are the particles to be measured shown in FIG. The results of performance evaluation are shown. The concept of specimen pretreatment in detecting whether or not a microorganism to be detected exists in a sample is shown. The concept of sample pretreatment for identifying the types of viruses present in the sample is shown. A sample pretreatment concept for quantifying the target protein concentration in a sample is shown. Fig. 3 shows the pulse height (transient change in ion current) when measuring polystyrene beads of various diameters with a Coulter apparatus having pores with a hole diameter of 300 nm and a thickness of 50 nm. 1 shows an example of a sample that is a target of a detection method according to an embodiment of the present invention. FIG. 10 is a diagram expressing 6 measurements performed for each pulse by selecting two of the feature amounts extracted by the feature amount extraction means from the pulses extracted by the pulse extraction means and performing them in a scatter diagram. A confusion matrix that compares a teacher label and an AI detector output on a pulse-by-pulse basis and its AI detection result are shown.

The fine particles to be detected, identified, or quantified in the present invention are not limited, and can be arbitrarily selected according to the type and performance of the measuring device. Suitable examples of microparticles include protein molecules, microorganisms (fungi, bacteria, viruses, etc.), exosomes, and the like.

The program according to the present invention has a function of performing the method according to the present invention by storing it in an arbitrary storage medium (RAM, storage, etc.) as containing software instructions and executing the instructions on the processor. The type of computer device having such a processor is not particularly limited, and may be any computing device such as a personal computer, smart phone, or server. In addition, hardware for measuring fine particles, hardware for learning an AI program (executing software), and analysis (detection, identification, or quantification) of fine particles using the learned AI program ( The hardware (which executes the software) may each be separate physical devices or they may be the same device. If they are separate devices, they may be connected to each other via a network, or some of them may operate standalone (eg for security reasons).

FIG. 4 shows an example of a measuring device that constitutes a fine particle measuring apparatus based on the Coulter method used in an embodiment of the present invention. Device 400 has a cross-sectional structure in which two

chambers

410 and 420 are connected via pore 440 .

Electrodes

412 and 422 are installed in the two chambers, respectively. A sample containing the particles to be measured suspended in the electrolytic solution is introduced into the chamber 410 through the introduction port 411 , and the electrolytic solution is introduced into the chamber 420 through the introduction port 421 , and a voltage is applied to the two electrodes by the voltage source 452 . For example, when a voltage is applied so that the electrode 412 becomes a negative electrode and the electrode 422 becomes a positive electrode, an ion current flows through the pores. Then, the negatively charged particles in chamber 410 move to chamber 420 through pores 440 by electrophoresis.

In the example of FIG. 4, the negatively charged particles in the sample move downward from the top of the drawing due to the Coulomb force. Therefore, every time one particle in the sample passes through the pore 440, the ion current between the

electrodes

412 and 422 decreases transiently, and the ammeter 451 produces a signal as schematically shown in FIG. is output. In one example of FIG. 6, the horizontal axis 601 is time and the vertical axis 602 is the output having a positive correlation with the resistance value between the

electrodes

412 and 422. This is the time change of the resistance value when passing through FIG. 6 is an example in which 10 particle passing pulses were observed. For example, width 621 is the duration of example pulse 615 . The output of the ammeter 451 in FIG. 4 is a downward pulse that decreases when a particle passes through. , the pulse can be either upward (increase as the particle passes) or downward (decrease as the particle passes).

FIG. 5 shows an example of a measuring section of a flow cytometry device used in an embodiment according to the present invention. A sample containing bioparticles such as cells is introduced from an injection port 512, and a sheath liquid is introduced from a sheath liquid introduction port 511 into a sample introduction tube 515 and a sheath liquid introduction tube 516, respectively. The sample and the sheath liquid pass through the mixing section 513 and are introduced into the measuring tubule 510 . Particles in the sample pass through the measurement unit 500 . Incident light 506 generated by the light source 502 passes through the measurement tubule 510 and reaches the measurement section 500 . When the particles contained in the sample pass through the measurement unit 500 , they scatter and radiate around the measurement unit 500 . As processing before injecting the sample into the injection port 512, the particles to be measured may be subjected to fluorescence dyeing processing in which the particles are excited by light of a specific wavelength and emit light. When the light source 502 emits light containing the specific wavelength, when the particles pass through the measurement unit 500, the dyed particles are excited by the incident light and emit fluorescence, and the fluorescence is emitted around the measurement capillary. . In flow cytometry, a light receiver (detector) for detecting a transient change in intensity of emitted scattered light or fluorescence is installed around the measurement unit 500 . In one example of FIG. 5, a photodetector A503, a photodetector B504, and a photodetector C505 are installed.

In the example of FIG. 5, the particles in the sample move downward from the top of the drawing. Therefore, every time one particle in the sample passes through the measurement unit 500, scattered light or fluorescence increases transiently, and a signal as schematically shown in FIG. 6 is output from the light receiver. In an example of FIG. 6, the horizontal axis 601 is time, and the vertical axis 602 is the current output having a positive correlation with the received light intensity of the detector. It is the time variation of scattered light or fluorescence intensity received by one receiver. FIG. 6 is an example in which 10 particle passing pulses were observed. For example, width 621 is the duration of example pulse 615 . The output of the photodetector may be current or voltage.

FIG. 7 is an example configuration of a detection, identification or quantification device according to an embodiment of the invention. The sensors X703 to Z705 may be those capable of outputting a signal in response to an external stimulus, and may be, for example, the ammeter 451 in FIG. 4 according to the Coulter method, or the photodetectors A503 to C505 in FIG. 5 for flow cytometry. The number of sensors may be any number as long as it is one or more. The output from the sensor is amplified and A/D converted by the input signal processing means 710 and then sent to the information processing means 730 via the input/output means 720 . The digitizing sampling interval is preferably 1/5 or less of the pulse duration, but in the present invention, the shorter the sampling interval, the higher the identification performance.

The information processing method of the present invention consists of two stages. The first is learning and the second is detection, identification or quantification. The first learning is a process of optimizing the AI parameters of the AI program using samples with known types and concentrations of particles contained in the samples. Also, an optimized AI program is called a learned AI program. The second detection, identification, or quantification is a process of inputting pulses of unknown particles to the learned AI program and estimating the presence, type, or concentration of the particles estimated by the learned AI program as its output.

AI parameters differ depending on the AI program used. For example, in the case of linear discriminative transformation, the coefficient vector and bias term for each discriminant class, in the case of deep learning, the structure of the neural network, the coefficients of each node, the activation function, and the coupling coefficient , a convolutional layer filter of a convolutional neural network, etc., which are parameters to be optimized in the AI program used in the implementation of the present invention, and may be anything. In the present invention, any type of algorithm may be implemented in the AI program. For example, the k-nearest neighbor method, linear discriminant method, various deep learning methods, support vector machines, decision trees, Boltzmann machines, or an ensemble method combining various algorithms may be used.

(Learning AI programs)
FIG. 8(a) shows an example of AI program learning according to an embodiment of the present invention. In the following description, an example of learning using the apparatus of FIG. 7 is shown. For example, the output from the photodetector in a Coulter method ammeter or flow cytometry results in a noisy peak as shown in FIG. For this reason, in this embodiment, first, as illustrated in FIG. 9, after estimating the baseline 901 of the pulse (step S811), the start and end times of the pulse edge 902 are estimated (step S812), and the extracted pulse Intra-pulse data 903 is extracted (step S813). For example, the example pulse width 621 in FIG. 6 is the time difference of the pulse edge 902 . Intra-pulse data refers to all data sets of times and current values between the extracted pulse ends as starting and ending points. The time interval of intra-pulse data is determined by the A/D conversion sampling interval in the input/output processing means (A/D converter) 450 or 710 . In this example, the intra-pulse data is extracted using the baseline estimation, but the present invention may extract the intra-pulse data by any method. Intra-pulse data may be extracted by using a portion exceeding .

Next, AI data preprocessing of intra-pulse data is performed for AI program learning (step S814). Examples of AI data preprocessing according to the present invention include the generation of smoothed pulse data by applying a low-pass filter to the intra-pulse data to reduce high-frequency fluctuations, and the standardization of the maximum voltage of each intra-pulse data as 1. generation of normalized pulse data, moving average of measured values before and after each measurement time, etc. Also, the AI data preprocessing in the present invention is not limited to smoothing or normalization, and any calculation may be performed on part or all of the intra-pulse data or incidental measurement conditions. These preprocessings are used to optimize the AI program by allowing the AI model to better represent the intra-pulse data and finding the best solution among the many extreme values of the evaluation function using limited computational resources. Any pre-learning data processing for guidance may be used.

Note that AI performance in learning does not necessarily improve as the sampling frequency of the A/

D converter

450 or 710 increases. The optimal frequency at which the AI program can express the correlation with the pulse shape and measurement conditions well depends on the AI algorithm used. Optimal sampling can be found experimentally. The lower the sampling frequency, the lower the cost of amplifiers and digitizers.

When using an AI program with a large number of input nodes such as a deep neural network, all intra-pulse data can be directly input to the AI input layer, or data after AI data preprocessing (S814) can be directly input to the AI input layer. However, the pulse waveforms output from the sensors X703 to Z705 exemplified as an embodiment according to the present invention are not constant in pulse width in both the Coulter method and flow cytometry. Therefore, even when a pulse waveform is directly input to the input layer of the deep neural network, AI data preprocessing such as normalization and padding in the time direction can be performed before input to the input layer. FIG. 10 is an example in which padding 1000 is applied after the end of pulse 1 as AI data preprocessing of the deep neural network. In the example of FIG. 10, the first pulse 1001 has a shorter duration than the second pulse 1002 . Without AI data preprocessing, the value after the end of pulse 1001 cannot be input to the input layer, and some of the input nodes of deep learning become indefinite. In order to prevent this in the example of FIG. 10, an output voltage of 0 is added after the pulse 1001 ends. In another embodiment, the preprocessing for aligning short duration pulses to long pulses is not limited to this, and can be anything.

Next, the feature quantity extraction means 732 extracts a feature quantity representing the feature of each pulse (step S815). In this embodiment, by extracting various feature amounts as shown in FIG. 11A for each pulse, more information about each particle can be extracted from the pulse waveform. The example in FIG. 11(a) is that the feature quantity a is the width, b is the height, c is the asymmetry of the pulse time (0≦c≦1), d is the asymmetry of the pulse area (0≦d≦1), e is the angle between the two ends of the pulse and the straight line connecting the pulse peak, f is the pulse area, g is the moment of the pulse with respect to the time axis, and h is the moment of the pulse with respect to the voltage axis. One pulse is represented by a vector having such feature quantities as elements. As shown in FIG. 11(b), if the number of feature quantities used for AI learning is p (p≧1), one pulse is represented by one point on the p-dimensional space. For example, the width of the pulse and the peak value within the pulse have been used to analyze the measurement results of the Coulter method. In addition, for analysis of measurement results of conventional flow cytometry, for example, intra-pulse peak values of forward scattering (light receiver 503 in FIG. 5) and side scattering (light receiver 505 in FIG. 5) are used, and two-dimensional analysis is performed. was often done. On the other hand, in the present invention, when the number of sensors is q, AI learning searches for a discrimination boundary in a p×q-dimensional space. As a result, even in the situation shown in FIG. 1, where the conventional method cannot determine the identification boundary, it is possible to obtain the identification boundary with high accuracy, thereby mainly solving the first problem.

The pulse feature amount of the present invention is not limited to those illustrated in FIG. 11(a), and any amount that expresses the pulse waveform feature may be used. Moreover, the number of feature amounts may be any number as long as it is one or more. In another embodiment, pulse data extraction may not be performed for feature amount calculation. For example, a method of estimating a peak value directly from comparison with data around a certain time without extracting intra-pulse data may be used.

Furthermore, in an embodiment of the present invention, additional measurement conditions that differ for each measurement are used as feature amounts for AI learning, thereby enabling more accurate identification and quantification.

In the case of the Coulter method, the incidental measurement conditions are, for example, the hole diameter and shape of the pore 440 in FIG. may be used as In reality, manufacturing variations for each sensor are unavoidable, and this causes the cluster shift shown in FIG. Therefore, the analysis device 740 acquires the characteristics of the sensor device such as the hole diameter and the partition wall thickness for each sensor device (Coulter device) 400 used at the time of measurement from the input means 722, and stores them in the incidental measurement condition storage means 442. Then, it is learned as a feature in AI learning (step S818). Supplementary measurement conditions to be learned are not limited to hole diameter and partition wall thickness, but may be anything that affects the pulse waveform, such as channel structure, surface treatment of channels and pores, and electrical characteristics. . In addition, the conditions may be different for each measurement, such as the result of measuring the hole diameter of the device 400 one by one, may vary for each manufacturing lot, or may be uniformly determined by the design specifications. .

In the case of flow cytometry, additional measurement conditions include conditions that affect the pulse waveform, such as the turbidity of the specimen, based on the light source characteristic information, the receiver angle information, the flow velocity information, the fluorescent label characteristic information, and the sheath fluid physical property information. can be used. These incidental measurement conditions include unavoidable variations between measurements or production lots, which causes the cluster shift shown in FIG. Therefore, the analyzer 740 acquires these incidental measurement conditions from the input means 722 for each measurement, stores them in the incidental measurement condition storage means 442, and makes them learned as features in AI learning (step S818). Any incidental measurement condition to be learned can be used as long as it affects the pulse waveform. In addition, even under conditions that differ for each measurement, such as sample turbidity and the number of measurements after cleaning the measurement capillaries, or conditions that vary from production lot to production lot, or even under conditions such as light source characteristics, receiver angle, flow velocity, fluorescent labels, and other design specifications, uniformity can be achieved. It may be a condition that is determined by

In both the Coulter method and flow cytometry, the additional measurement conditions may be measurable physical properties, or discontinuous numbers or symbols such as manufacturing lot numbers and serial numbers of devices and reagents. In this way, by performing AI learning with incidental measurement conditions included in teacher data, it is possible to create an AI program that includes variations for each measurement, and solve or alleviate the second problem. This method is effective in all cases of detection, identification and quantification.

For example, convolutional and pooling layers in neural networks can also be used as preprocessing to extract pulse features. For example, as shown in FIG. 5, in the case of a measuring device having multiple sensors, each output from multiple sensors may be used as a channel in the convolution. Equation (1) below represents the convolution process from the L-1th layer to the Lth layer. k is the channel number, K is the number of channels, H is the filter size, p is the coordinate in the filter (0<p<H−1), h _pkm is the convolution coefficient in the filter, and S is the stride.

In the example of FIG. 5, receivers A, B and C correspond to k=0, 1 and 2, respectively. The convolution coefficient h _pkm may be uniquely determined, and the convolution processing corresponds to AI data preprocessing (step S814). Alternatively, the convolution coefficient h _pkm may be included in the optimization parameters for learning, in which case the convolution process can be regarded as AI data preprocessing (step S814) or AI learning (step S818). Further, as shown in FIG. 8A, the present invention may or may not perform feature extraction.

A teacher label is assigned to data on which AI data preprocessing (step S814) or pulse feature quantity extraction (step S815) has been performed. Teacher label assignment for AI learning will be described using the schematic diagrams of FIGS. 12 to 14. FIG.

FIG. 12 shows an example of teacher label assignment to measurement results using the Coulter device for learning the feature values illustrated in FIG. 11(a). FIG. 12(a) is an example of teacher labeling for AI learning to discriminate using the measurement result of a sample in which the presence or absence of microorganisms, proteins, exosomes, etc. to be detected in the sample is known. In this example, the teacher label 1205 is a binary value of "Positive" indicating that the microorganism to be detected is included and "Negative" indicating that it is not included. FIG. 12(b) is an example of teacher labeling for quantifying AI learning using measurement results of samples with known concentrations of microorganisms, proteins, exosomes, etc. in the sample. The teacher label 1225 may be the concentration or amount of detection target microorganisms, proteins, exosomes, or the like. In an example of FIG. 12(a), one line represents one measurement pulse, and for each measurement pulse, the pore hole diameter and pore thickness are set as incidental measurement conditions 1203, and the feature amounts a to A teacher label is assigned in association with p teacher data. This is just an example, and the incidental measurement conditions and feature amounts may be of any kind, and the incidental measurement conditions may be absent.

In the example of FIG. 12(a), the sample with sample number 1 is measured by two different devices. A pulse feature amount result 1211 is measured by the first device, and a pulse feature amount result 1212 is measured by the second device. Also, the measurement of the sample number 2 is performed by the third device, and the result is 1213. One row represents one pulse, and results 1211 and 1212 are the results of measuring a plurality of pulses from each one device. The pore diameters and pore thicknesses of these first to third devices vary due to manufacturing variations. Such variations in incidental measurement conditions affect the pulse waveform, causing cluster shift as shown in FIG. 2, for example, which causes the second problem. According to the present invention, as shown in FIG. 12, incidental measurement conditions 1203 that cause variations are targeted for AI learning together with pulse waveform feature amounts 1204 . Since the AI program also learns the effect of incidental measurement conditions on the pulse waveform, it is possible to perform detection, identification, or quantification in consideration of the causes of variations due to incidental measurement conditions.

A sample to be learned is one whose presence (in case of detection), type (in case of identification), or concentration (in case of quantification) of microorganisms, proteins, etc. to be measured is known. In the example of FIG. 12(a), the correct answer for sample number 1 is known as "Positive". Therefore, the accompanying measurement conditions and the pulse feature quantities 1211 and 1212 are all given a teacher label of “Positive”. Since the correct answer for sample number 2 is known to be "Negative", the incidental measurement conditions and the pulse feature quantity 1213 are all labeled "Negative". Similarly, in FIG. 12(b), since the concentration of the microorganism or protein to be measured for sample number 1 is known to be "2.32", all the measurement results are given the teacher label "2.32". ing. Similarly, all the measurement results of sample number 2 are given the teacher label "0.856". In FIG. 12, the Coulter apparatus was used for the description, but the measurement results of flow cytometry can also be assigned teacher labels in the same manner.

The results of detection, identification or quantification of samples used for AI learning must be known. The results of measuring and analyzing the sample to be learned by another detection, identification, or quantification method may be used as teacher labels for AI learning. For example, the results of existing detection, identification or quantification means such as PCR, ELISA, immunochromatography and turbidimetric methods may be used.

In addition, in the measurement for learning, the biological reaction, chemical reaction, and physical characteristics of the microorganisms and cell types to be measured, such as density gradient centrifugation, immunoprecipitation, chromatography, salting out, SDS-PAGE, and cell sorter, are used to determine the learning target. It may be possible to prepare a sample of known grain type that is generally free of all but one type of grain. In such cases, known grain types may be learned as teacher labels for all pulses obtained from such samples. FIG. 13(a) shows an example in which a sample containing generally only one type of leukocyte was measured by flow cytometry and assigned a teacher label by the method of the present invention.

In an example of FIG. 13(a), the light source wavelength, the light receiver angle, the flow velocity and the fluorescence wavelength when blood cells are measured by flow cytometry are used as incidental measurement conditions 1303, and the pulse waveform feature quantity 1304 in FIG. 11 is used as teacher data. , a teacher label 1305 is given. Measurement results 1311 and 1312 are the measurement results of the light receiver 503 and the light receiver 505, respectively. In FIG. 13A, the AI is learned with the light receiver angles of 179 degrees and 90 degrees, respectively. In the example of FIG. 13A, for example, the teacher data is the result of measuring a plurality of samples that do not generally contain particles other than "LYMPH", which is the learning target particle, and the teacher labels 1305 are all "LYMPH". is given.

In addition, it may not be possible to purify a sample containing only particles to be learned in this way. If the grain type can be estimated for each particle, that is, for each pulse, instead of for each sample, by a conventional measurement method that determines teacher labels, a teacher label may be assigned for each grain type as shown in FIG. 13(b). . However, when measuring a sample mixed with different target particles as shown in FIG. Since optimization is performed by assuming that the teacher label of the conventional method is correct, it is not possible to achieve identification accuracy exceeding that of the apparatus of the conventional method. However, with respect to cluster shift for each measurement as shown in FIG. Thus, teacher labeling (step S816) can also be performed.

For example, it is assumed that the results of measuring a sample in which three types of particles A to C are mixed using a Coulter apparatus or flow cytometry are as shown in FIG. In the method according to the present invention, the teacher label used for AI learning is not uniquely determined by clustering data obtained by mixing multiple measurement results, but is determined based on the discrimination boundary calculated by clustering for each measurement. FIG. 14 is a conceptual diagram of the result of calculating the identification boundary 1410 for measurement 1 and the identification boundary 1420 for measurement 2, respectively. In assigning teacher labels in step S816, for example, pulses in

regions

1418 and 1419 are labeled with grain type "A", pulses in

regions

1428 and 1429 with grain type "B", and pulses in

regions

1438 and 1439 with grain type "B". "C" is given as a teacher label to each.

For clustering at this time, what kind of algorithm is used, such as simple linkage method, complete linkage method, group average method, hierarchical method such as Ward's method, non-hierarchical method such as k-means method, clustering by random variable model, etc. may be used. In general, the computational complexity of clustering a large amount of data is large. For this reason, in the prior art, performing clustering calculations on a large amount of measurement data in real time and detecting, discriminating, or quantifying beyond measurement-to-measurement variability was impossible or required high-cost hardware. On the other hand, in the method according to the present invention, clustering can be limited to preprocessing in learning, and at the time of detection, identification or quantification, identification is possible by inputting measurement data to one function called a trained AI program. is. Therefore, identification can be performed with practical hardware and speed. As a result, in the present invention, even if a conventional measuring device is used to acquire training data, the AI absorbs the variations in each measurement that caused erroneous identification in the conventional method, enabling high-precision identification. It can be carried out.

FIG. 15 is another example of supervised labeling for identifying microorganisms. In FIG. 15, each of the voltage output groups 1514 for each time 1513 of the intra-pulse data 1501 of the pulse number and the

incidental measurement conditions

1511 and 1512 are combined to give one teacher label "N1H1". While FIGS. 12 and 13 assign teacher labels to a set of incidental measurement conditions and a set of pulse feature values for each pulse, in the example of FIG. , one teacher label is assigned to all sensor outputs (intra-pulse data) at each time.

Such attachment of teacher labels to data to be learned is repeated for all pulses given to the AI program (step S817). When this is completed, it is input to the AI program 733, and the parameter optimization means 734 executes learning (step S818). Here, learning refers to the process of optimizing the parameters to be optimized included in the AI program, for example, the process of finding the extreme value of the evaluation function in the AI program. The optimized AI program parameters are stored in the AI parameter storage means 747 (step S819). The optimized parameters optimized in step S818 are stored in the AI parameter storage means 747 and used for subsequent identification.

For example, as an example of learning when teacher labels are assigned as shown in FIG. 15, use of an AI model having a plurality of nodes in an input layer such as a deep neural network is shown. FIG. 16 schematically shows a deep neural network, and some nodes and connections between nodes are omitted. Taking pulse number 1 1501 in FIG. Each value of the data 1514 of the voltage output group of the sensor output is input to the input node 1614 . A teacher label 1515 is given to such a neural network as a correct answer, and the coefficients of each node and connection are optimized by, for example, the back propagation method. The example in FIG. 16 has two output nodes and performs binary classification of the types of microorganisms contained in the sample. The output layer may output multiple values or some continuous quantity. The structure of the AI program shown in FIG. 16 is merely an example. For example, there may be an arbitrary number of output amounts, or a continuous amount may be output. In the case of detection to check whether or not particles to be measured are included in a sample, the output is binary classification, and in the case of identification in which a plurality of types of particles to be measured are included, it is multi-value classification. Also, in the case of performing quantification in which a continuous quantity is given as a teacher label as shown in FIG. 12(b), an AI program that outputs a continuous quantity can be used.

In the description of the embodiment of the present invention with reference to FIGS. 12, 13, and 15, the incidental measurement conditions have been described using an example in which the AI program is learned for each pulse as a feature amount. The present invention is not limited to this, and it is sufficient that the incidental measurement conditions are input in some form for learning of the AI program. In another embodiment, for example, an AI program may be independently trained using pulse feature amounts or intra-pulse data for each incidental measurement condition, and the results may be ensembled after the training.

(Detection, identification or quantification by AI)
Next, the detection, identification or quantification process shown in FIG. 8(b) will be described. In one embodiment, a learned AI program is created by the learning process shown in FIG. 8(a), and optimized AI parameters are stored in the AI parameter storage means 747 (step S819). Unknown analytes to be detected, identified or quantified are then measured with a Coulter device or flow cytometry device. In the measurement for this identification, it is desirable that the configuration of the measuring device and the incidental measurement conditions are the same as those at the time of learning. In another embodiment of the present invention, one AI program is trained with measurement results under various measurement conditions to create an AI program that can identify even if the configuration of the measurement device and the incidental measurement conditions are not constant. can also

In the identification process shown in FIG. 8B, baseline estimation (step S821) to pulse feature quantity extraction (step S825) are the same as steps S811 to S815 at the time of learning, so detailed description thereof will be omitted. FIG. 17 shows an example of result output by an AI program according to an embodiment of the invention. FIG. 17(a) shows a case where the output 1713 of the AI program is binary "Positive" and "Negative", and FIG. 17(b) shows a case where the output 1723 of the AI program is a continuous quantity. FIG. 17(b) can be used for detection, identification, and quantification of fine particles. Using an AI program that outputs a continuous quantity as shown in FIG. 17(b), when the desired result is detection or identification, classification is performed by setting a boundary value for the output value. For example, in the binary classification of "Positive" and "Negative", if the probability that the result output 1723 is positive is 0.5 or more is classified as "Positive", and less than 0.5 is classified as "Negative". The threshold at this time may also be a learning target in the AI program.

Also, the configuration for realizing the present invention is not limited to FIG. 7, and in another embodiment, for example, as shown in FIG. you can In still another embodiment, AI server 1930 connected via network 1900 may perform part of the processing according to the present invention, as shown in FIG. 19, for example. Any partitioning of each element constituting the present invention may be used.

In this way, in the present invention, by having the AI program learn various characteristics of waveforms and measurement conditions, for example, when the scatter diagram of feature values is greatly superimposed as in FIG. The AI program was able to detect, discriminate and quantify with high accuracy even when the clusters for each species shift. This makes it possible to solve or reduce the first and second problems.

(Contaminant sorting AI)
A third problem to be solved by the present invention is that biological samples contain particles (contaminants) other than those to be measured, and their identities are unknown. FIG. 20(a) 2010 schematically represents the pulse number ratio in the measurement results of the biological sample containing the measurement target particles to be detected, identified or quantified, and FIG. 20(b) 2020 the biological sample not containing the measurement target particles. It is what I did. For example, in a measurement intended to detect microorganisms, FIG. 20(a) is positive and FIG. 20(b) is negative. FIGS. 21(a) and 21(b) schematically represent the measurement results of the samples shown in FIGS. 20(a) and 20(b), respectively. The measurement result of FIG. 21(a) is a mixture of pulses 2112 originating from particles to be measured and pulses 2111 originating from contaminants. Therefore, when performing AI learning, it is not possible to assign teacher labels as described with reference to FIGS. In the present invention, the measurement result in which contaminants and particles to be measured are mixed in the sample to be measured and the discrimination boundary cannot be determined as shown in FIG. Pulses derived from contaminants are separated by using the measurement result of a sample containing only contaminants. In this way, the present invention provides a method for identifying and quantifying particles to be measured with higher accuracy than in the prior art.

Let x be a feature vector obtained from one pulse for the measurement results of the particles to be measured as shown in FIGS. R is the feature space.

Then, y=1 when one pulse is truly derived from impurities, and y=-1 when it is truly derived from the particles to be measured.

Originally, AI should be learned by assigning a teacher label only to the pulse of y=-1, but this cannot be done with the measurement result mixed with contaminants as shown in FIG. 21(a). Therefore, in the embodiment of the present invention, the measurement result of only impurities not containing the target particles as shown in FIG. 21(b) is used for learning. This is a technique that utilizes the fact that, in many biological particle measurements, it is not possible to make a sample of only target particles, but it is possible to make a sample of contaminants only.

Here, let s=1 be a pulse that can determine whether one pulse is a contaminant or a particle to be measured as a result of observation, and s=0 for a pulse that cannot be determined to be a contaminant or a particle to be measured. All pulses in FIG. 21(b) have s=1 and all pulses in FIG. 21(a) have s=0.

At this time, the probability that the observed value x of one pulse of the mixed sample result of the particles to be measured and contaminants is really contaminants is expressed as follows.

This equation (5) estimates the probability that y=1 for each pulse in FIG. 21(a) for each pulse. Then, the left side of the equation (5) removes pulses exceeding the threshold value for removing impurities, and AI learning is performed only with the remaining pulses. As a result, in the method according to the present invention, for measurement data in which particles to be measured and contaminants are mixed as shown in FIG. separation became possible. In the embodiment of the present invention, AI learning can be performed by calculating the probability that a certain pulse is a contaminant and excluding the pulse exceeding the threshold, without being limited to the formula (5). Learning of the contaminant sorting AI of the present invention may be learning of pulse feature values as shown in FIGS. 12 and 13 in some embodiments. In another embodiment, the sensor output value for each time may be learned as shown in FIG.

In order to verify the effect of the present invention, two types of polystyrene beads having similar particle sizes of 200 nm and 220 nm modified with COOH groups were used. 2012 are simulated as 220 nm beads, and the particle passing pulse is measured with a Coulter apparatus having the device shown in FIG. We verified the effect. FIG. 22 shows a 200 nm single sample 2020 of FIG. This is the result of evaluating the binary discrimination performance between and. The vertical axis 2202 is the accuracy rate when the learning result by deep learning as shown in FIG. 16 is evaluated on a sample-by-sample basis, where 1.0 is all questions correct and 0 is all questions incorrect. Therefore, 0.5 indicates that no discrimination is possible (score when AI randomly outputs the discrimination results of FIGS. 20(a) and (b)). According to FIG. 22, it can be seen that the result 2220 when the technique of formula (5) is used achieves significantly higher discrimination performance than the result 2210 when the technique of formula (5) is not used, especially when the particle mixture ratio is low.

As described above, the contaminant sorting AI according to the present invention can solve or reduce the third problem by stochastically eliminating pulses derived from indefinite contaminants.

(sensitization)
A fourth problem to be solved by the present invention is the problem that the performance of detection, identification or quantification deteriorates when the number of particles to be measured is extremely small, for example a<0.1 in FIG. In real biological samples there are many demands for the detection, identification and quantification of low concentrations of cells, microorganisms, proteins and exosomes. In the present invention, in addition to the method of the present invention, sample pretreatment using an antibody enables detection, identification, and quantification of even low-concentration particles to be measured.

According to an embodiment of the present invention, in the sample pretreatment for sensitization, in the first step, the sample is mixed with beads modified with antibodies that specifically bind to the protein to be detected, microorganisms, exosomes, etc. , concentration is performed in the second step, and then the measurement, AI learning, and AI detection/discrimination/quantification described above are performed.

The pretreatment for detection of microorganisms, proteins, exosomes, etc. and AI learning according to the present invention are described. FIG. 23 is a conceptual diagram of specimen pretreatment in detecting whether or not a microorganism to be detected exists in a sample. In the embodiment shown in FIGS. 20 and 21, in the AI learning 2310 of the positive sample 2311 consisting of the virus 2301 and contaminants 2302 shown in FIG. AI learning to remove objects was explained. On the other hand, in this embodiment, beads 2303 modified with an antibody that specifically binds to the virus 2301 to be detected are mixed in both the positive sample and the negative sample, resulting in the state shown in FIG. 23(b). Virus-bound bound beads 2304 are produced in the bead-mixed positive sample 2321 but not in the bead-mixed negative sample 2322 .

In some embodiments of the present invention, if there are significantly fewer contaminants in the sample than the bound beads, if the pulse of contaminants is too small to be measured by a Coulter or flow cytometry device, or if in feature space If the pulses from the contaminants and the bound beads can be clearly separated (hereafter, these three are collectively referred to as "AI unnecessary requirements"), the presence or absence of pulses from the bound beads can be measured without AI learning. Detection of microorganisms is possible. In the former two cases, the detected pulses are mostly bound beads, so the presence or absence of the detected pulses can be used to determine the presence or absence of the microorganisms to be detected. In addition, even if there are contaminant pulses or pulses of unbound antibody-modified beads, if they can be separated from the bound beads in the feature value space, by specifying the bound beads by clustering, etc., the presence or absence of the microorganisms to be detected can be detected. can be determined. This is the same even if the detection target is protein or exosome. However, many samples do not meet the AI-free requirement, in which case the following AI learning and detection process according to another embodiment of the invention is performed.

For samples that do not satisfy the AI unnecessary requirement, AI learning 2320 is performed by assigning teacher labels of "positive" and "negative" to the bead-mixed positive sample and the bead-mixed negative sample, respectively (step S818 in FIG. 8). Create an AI program. Since the size and shape of these bound beads are different from those of contaminants, when a bead-mixed sample is measured by a Coulter apparatus or flow cytometry, the pulses obtained from bound beads are different from those of viruses and contaminants. Therefore, higher performance than AI learning 2310 is realized. Furthermore, in AI learning 2320, a method using foreign matter sorting AI according to formula (5) may be used. In this case, setting s=1 in the formula (4) may be the measurement result of the negative sample 2312 or the measurement result of the bead-mixed negative sample 2322 . This allows AI programs to learn the characteristics of negative samples that contain unbound beads, so that sorting of unbound beads 2323 contained in bead-mixed positive samples can be expected to perform well even for dilute samples. . The preparation of the sample mixed with such beads may be performed by the same device as the device that operates the AI program, or may be performed by another device. The same applies to other samples exemplified in this specification.

Bound beads 2304 in FIG. 23(b) are larger in size and mass than contaminants and unbound beads. Bound beads can be concentrated as follows. In addition, for example, if a magnetic substrate is used for the antibody-modified beads 2303, the bound beads can be concentrated by applying a magnetic field to the bead-mixed sample.

In the present invention, if the sample concentrated in this way satisfies the requirement that AI is not required, then FIG. 23(c) is more sensitive than the case where FIG. can confirm the presence or absence of microorganisms and proteins to be detected.

For samples that do not satisfy the AI unnecessary requirement, AI learning 2330 is performed by assigning "positive" and "negative" labels to the bead-mixed positive enriched sample 2331 and the bead-mixed negative enriched sample 2332, respectively (step S818 in FIG. 8). , create a trained AI program. Since the number of bound beads in the bead mixture positive enrichment sample 2331 is larger than that in the bead mixture positive sample 2321, even if the concentration of cells, microorganisms, proteins, exosomes, etc. to be measured is low, AI learning The learned AI program created in 2330 has higher performance than those in

AI learning

2310 and 2320. Furthermore, in AI learning 2330, a method using foreign matter sorting AI according to formula (5) may be used. In this case, the measurement result of the negative sample 2312 or the bead-mixed negative sample 2322 may be used to set s=1 in the formula (4), but it is preferable to set the measurement result of the bead-mixed negative concentrated sample 2332 . This allows the AI program to learn the characteristics of negative samples that have undergone the enrichment process, so that selection of contaminants, unbound beads or unbound virus that remain in the bead-mixed positive enrichment sample can lead to more dilute samples. Good performance can also be expected for

In the above-described pretreatment and the detection of FIG. 8(b), the labeling 2329 shown in FIG. Detection may be performed by inputting into a trained AI program.

Although the detection of microorganisms was described as an example in FIG. 23, detection is possible by using beads modified with antibodies that specifically bind to the protein to be detected or exosomes to be detected instead of the virus 2301. becomes. Also note that this method is applicable to both the Coulter apparatus and flow cytometry.

Next, the pretreatment for identification of microorganisms, proteins, exosomes, etc. and AI learning according to the present invention will be described. FIG. 24 is a conceptual diagram of sample pretreatment for identifying virus types present in a sample. FIG. 24( a ) shows a corona-positive sample 2411 consisting of coronavirus 2401 and contaminants 2403 and an influenza-positive sample 2412 consisting of influenza virus 2402 and contaminants 2403 . In this embodiment, corona positive sample 2411, influenza positive sample 2412 and negative sample 2413 are combined with anti-corona beads 2404 modified with anti-corona antibodies that specifically bind to coronavirus and anti-influenza beads that specifically bind to influenza virus. Both of the antibody-modified anti-influenza beads 2405 are mixed to obtain the state shown in FIG. 24(b). Corona-positive bead mixture sample 2421 produces corona-bound beads 2406 but not flu-bound beads 2407 . Influenza-positive bead-mix sample 2422 produces influenza-bound beads 2407 but not corona-bound beads 2406 . No bound beads are produced in the negative bead mixture sample 2423 . In this embodiment, anti-corona beads 2404 and anti-influenza beads 2405 are preferably different in shape, size or material.

In the present invention, for a sample that satisfies the requirement that AI is not required, if corona-bound beads and influenza-bound beads can be separated in the pulse feature space in an example of FIG. can do However, since many samples do not satisfy the AI-free requirement, the following AI learning and discrimination processing according to the present invention is performed in that case.

Next, the bead-mixed corona sample, the bead-mixed influenza sample, and the bead-mixed negative sample are labeled as “corona positive,” “influenza positive,” and “negative,” respectively, and AI learning 2420 is performed (step S818). Create an AI program. Since corona-bound beads, influenza-bound beads, and contaminants have different shapes, sizes, and materials, the pulse waveform obtained from the bound beads is Since it is different, generation of a trained AI program with higher performance than AI learning 2410 is expected. Furthermore, in AI learning 2420, a method using foreign matter sorting AI according to formula (5) may be used. In this case, s=1 in formula (4) may be the measurement result of the bead-mixed negative sample 2423 . For example, the bead-mixed corona sample 2421 is s = 0, the bead-mixed negative sample 2423 is s = 1, and the first contaminant selection is performed using equation (5). After performing the second contaminant selection with the negative sample 2423 as s = 1, ternary classification between the first contaminant-selected data, the second contaminant-selected data and the negative sample By performing AI learning, corona-positive samples, influenza-positive samples, and negative samples can be distinguished with higher performance.

Corona-bound beads 2306 and influenza-bound beads 2307 in an example of FIG. 24(b) are larger in size and mass than contaminants and unbound beads. Bound beads can be concentrated as in 24(c). Further, for example, if a magnetic substrate is used for the anti-corona beads 2404 and anti-influenza beads 2405, which are antibody-modified beads, the bound beads can be concentrated by applying a magnetic field to the bead-mixed sample.

In the present invention, if the sample concentrated in this way satisfies the requirement that AI is not required, the presence or absence of the pulse of the bound beads will result in a higher figure than the case of analyzing FIG. 24(c) can identify microorganisms and proteins to be identified with higher sensitivity.

For samples that do not meet the no-AI requirement, bead-mixed enriched corona sample 2431 was labeled "corona positive", bead-mixed enriched influenza sample 2432 was labeled "influenza positive", and enriched sample was labeled "negative" to generate a ternary AI. Learning 2430 is performed (step S818) to create a trained AI program. Since the number of bound beads in the positive sample in FIG. The AI program created in is expected to have higher performance than those of

AI learning

2410 and 2420. The sample in FIG. 24(c) may also be processed by the contaminant sorting AI in the same manner as described in FIG. 24(b).

In the above preprocessing and the identification of FIG. 8(b), after performing the labeling 2429 shown in FIG. It may be input into a trained AI program for identification.

In FIG. 24, virus detection was explained as an example, but multiple types of detection target proteins, exosomes, etc. were modified with antibodies that specifically bind to each of these proteins instead of coronavirus 2401 and influenza virus 2402. With protein beads, similarly sensitized AI learning and discrimination are possible. This method is applicable to both the Coulter apparatus and flow cytometry.

Subsequently, pre-quantification of microorganisms, proteins, exosomes, etc. and AI learning according to the present invention will be described. FIG. 25 is a conceptual diagram of sample pretreatment for quantifying the target protein concentration in the sample.

Samples

2511, 2512, and 2513, in which the concentration of the protein to be quantified is known, are mixed with anti-protein beads 2503 modified with an antibody that specifically binds to the protein 2501 to be quantified, and the state shown in FIG. 25(b) is obtained. do. For example, the concentrations of proteins to be quantified contained in

samples

2512 and 2513 are set to a% and b%, respectively (b>a). Bead aggregates 2504 and 2505 having a size corresponding to the concentration of the protein to be quantified are generated in the sample mixed with the beads. No bead aggregates are generated in the bead-mixed sample 2511 containing no protein to be detected.

In the present invention, for a sample that satisfies the AI-unnecessary requirement, it is possible to quantify target proteins and microorganisms using feature amounts that are influenced by the number of pulses of bead aggregates and the size of bead aggregates. This is because there is a positive correlation between the amount of the target protein or microorganism contained in the sample and the number of bead aggregates and the size of the bead aggregates.

For samples that do not satisfy the AI unnecessary requirement, AI learning 2520 is performed by assigning teacher labels of “0”, “a” and “b” to the bead mixed samples 2521 to 2523 (step S818). create a program For quantitation, AI algorithms may be employed where the inputs and outputs are continuous quantities. Since bead aggregates differ in size and shape from contaminants, their pulse waveforms are clearly different. Therefore, higher quantitative performance than AI learning 2510 is realized. Furthermore, in AI learning 2520, a method using foreign matter sorting AI according to formula (5) may be used. In this case, s=1 in formula (4) may be the bead-mixed sample 2521 containing no quantitative protein. For example, the measurement result of the sample 2511 may be set to s=1, and the

mixed bead samples

2521, 2522 and 2523 may all be set to s=0 and the contaminant sorting AI may be used. That is, AI quantification with better performance can be achieved by performing AI learning 2520 after removing the contaminants 2502 from each of the

samples

2521, 2522, and 2523 as much as possible.

For example, in an example of FIG. 25(b), since the size and mass of the aggregated bead clusters are larger than the contaminants, for example, by centrifuging and removing the supernatant, the beads aggregate as shown in FIG. 25(c). The mass can be concentrated. In addition, if a magnetic base material is used for the anti-protein beads 2503, bead aggregates can be concentrated by applying a magnetic field to the bead-mixed sample. The measurement results of each sample in FIG. AI learning 2530 for learning quantities is executed (step S818) to create a learned AI program. For those containing the protein to be quantified in FIG. 25(c), the number of aggregated bead clusters is greater than that of the positive sample in FIG. 25(b). Also, the AI program created by AI learning 2430 is expected to have higher performance than those of

AI learning

2410 and 2420 . The sample in FIG. 25(c) may also be processed by the contaminant sorting AI as in the case of FIGS. 22 and 23 .

In an embodiment of the present invention, if the concentrated sample satisfies the AI-free requirement as shown in FIG. Microorganisms to be identified, proteins and exosomes can be identified with higher sensitivity than FIG. 25(b) without concentration.

In the above preprocessing and the identification of FIG. 8(b), after performing the labeling 2529 shown in FIG. Detection may be performed by inputting into a trained AI quantifier.

In FIG. 25, protein quantification was explained as an example, but instead of the quantification target microorganisms, exosomes, etc. as the quantification target protein 2501, beads modified with antibodies that specifically bind to those microorganisms can be used. It is possible to learn and quantify AI by feeling. This method is also applicable to both the Coulter apparatus and flow cytometry.

In general, both the Coulter apparatus and flow cytometry have a lower limit to the particle size that can be measured as a pulse waveform. FIG. 26 shows pulse heights (transient changes in ion current) when measuring polystyrene beads with various diameters using a Coulter apparatus having pores with a hole diameter of 300 nm and a thickness of 50 nm. As shown in FIG. 26, the smaller the bead diameter 2601, the smaller the ion current drop (pulse height) 2602 when the beads pass through the pores. When a pore with a hole diameter of 300 nm is used, the pulses of particles of 100 nm or less are buried in baseline noise, and thus pulse data extraction in steps S813 and S823 in FIG. 8 cannot be performed. In order to measure particles below 100 nm with this Coulter device, a pore Coulter device with a smaller hole diameter of 100 nm is utilized.

In the labeling and concentration according to the present invention described in FIGS. 23 to 25, the object of measurement is bound beads, and contaminants and pulses from antibody-modified beads are noise that should be removed. In one embodiment of the present invention, the size of contaminants and antibody-modified beads is adjusted to the region 2610 in FIG. can be selected appropriately. For example, in FIG. 26, if the size of the antibody-modified beads is about 80 nm and the pore diameter is 300 nm, the pulses from the antibody-modified beads that are not bound to the particles to be measured (for example, 2524 and 2534 in FIG. 25) are AI learning. Because they are excluded from target or AI detection, identification or quantification of target pulses, a trained AI program with higher performance can be created.

23 and 24 and proteins in FIG. 25, the methods according to the present invention can be applied to antibodies that specifically bind to any microorganism, any protein, exosome, or the like. Anything that can produce modified beads can be used. Thereby, the fourth and fifth problems can be solved or reduced.

In order to demonstrate the effect of the present invention, the two types of samples shown in FIGS. gone. FIG. 27(a) shows Au particles 2700 with a diameter of 50 nm modified with an anti-IgG antibody using steptavidin and biotin dispersed in 1×PBS (phosphate buffered saline). FIG. 27(b) shows the same Au particles dispersed in a sample containing 0.1 mg/mL anti-IgG antibody 2701 in 1×PBS. Since the Au particles in FIG. 27(a) have a negative charge, they repel each other and are well dispersed. Some of the Au particles in the sample in FIG. 27(b) bind to streptavidin 2712, biotin 2714 and anti-IgG antibody 2701 coated on the Au particles, forming aggregates 2710, as shown in the enlarged view of FIG.

FIGS. 27(a) and (b) were filled into the channel (chamber) 410 of the Coulter device as shown in FIG. A voltage of 0.1 V was applied and the pulse signal accompanying the passage of the particles was measured by the ammeter 451 . For each sample in FIGS. 27(a) and 27(b), 3 measurements each were performed using different Coulter devices for a total of 6 measurements. The pore 440 was circular with a diameter of 300 nm. FIG. 28 expresses six measurements performed in a scatter diagram for each pulse by selecting two of the feature quantities extracted by the feature quantity extraction means from the pulses extracted by the pulse extraction means 731 . Other features are omitted. The horizontal axis 2801 is the feature amount a in FIG. 11, and the vertical axis 2802 is the feature amount b. The detection limit particle size for the Coulter device of FIG. 4 is 90 nm. Therefore, a pulse through unagglomerated 50 nm Au particles is not detected by the Coulter device 400 used here. As can be seen from FIG. 28, few pulses are observed from the sample in FIG. 27(a), but many pulses are observed from the sample in FIG. 27(b). This result indicated that the anti-IgG antibody was detected by pulse measurement using the Coulter device 400 .

Furthermore, regarding the data shown in FIG. 28, the feature amount of the pulse measured with the sample in FIG. ” as the correct answers, the AI detector was trained, and the results were cross-validated, and the results are shown in FIG. FIG. 29(a) is a confusion matrix for comparing the teacher label 2910 and the AI detector output 2920 on a pulse-by-pulse basis, and the pulse F value was about 0.71. Since no particles of 90 nm or less were detected in this experiment, the pulses observed from the sample in FIG. Even for Au particle agglomerates of similar size, the AI detector according to the present invention was shown to distinguish between agglomerated agglomerates and non-agglomerated agglomerates by the enlarged view of FIG. 27(b). Furthermore, FIG. 29(b) shows the results of AI detection for each sample by aggregating the results of FIG. 29(a). As a result, the F-value for correctness of detection was 1.0, demonstrating that the use of the trained AI detector according to the present invention exhibits high performance in anti-IgG antibody detection on a sample-by-sample basis.

27 to 29, the effectiveness of the present invention was verified for anti-IgG antibodies, but the method of the present invention described in FIGS. and each measurement means can be applied regardless of the Coulter method or flow cytometry.

Claims

configured to detect, as a pulse signal, a transient change in ionic current between two electrodes in contact with the electrolytic solution on both sides of the pore when particles to be detected dispersed in the electrolytic solution pass through the pore. a coulter measuring device that
a feature quantity extraction unit that calculates a feature quantity of a pulse waveform associated with passage of the particles through the pore;
an incidental measurement condition storage unit that stores incidental measurement conditions of the pores;
An AI program that learns the pulse waveform,
A first attachment containing one or more selected from the group consisting of the hole diameter, shape, and thickness of the first pore of the first pore of the first Coulter measurement device used for measuring known particles of known type storing measurement conditions in the incidental measurement condition storage unit;
The feature amount extraction unit calculates a first feature amount from a first pulse waveform obtained from the first Coulter measurement device as the known particles pass through the first pore,
creating a learned AI program by learning the AI program using the first feature quantity and the first incidental measurement condition as teacher data and the type of the known particle as a teacher label;
A second incidental measurement condition including one or more selected from the group consisting of the hole diameter, shape, and pore thickness of the second pore of the second Coulter measurement device used for measuring unknown particles, stored in the incidental measurement condition storage unit,
The unknown feature value calculated by the feature value extractor from the unknown pulse waveform obtained from the output of the second Coulter measuring device as the unknown particle passes through the second pore and the second incidental measurement. A device characterized by being configured to input conditions into the learned AI program to detect, identify, or quantify the particles to be detected.
A laser beam is irradiated to a measurement part including a transparent flow path through which particles in a sample pass in a row, and a pulse signal from scattered light or fluorescence from particles passing through the measurement part each time one particle passes through the measurement part. a flow cytometry instrument configured to obtain a
a feature amount extraction unit that calculates a feature amount of a pulse waveform obtained from a pulse signal obtained as the particles pass through the measurement unit;
an incidental measurement condition storage unit for storing incidental measurement conditions of flow cytometry measurement means;
An AI program that learns the pulse waveform,
First light source characteristic information, first light receiver angle information, first flow velocity information, and first fluorescent label characteristics representing the characteristics of the first flow cytometry measurement means used for measuring known particles of known types storing a first incidental measurement condition including one or more selected from the group consisting of information and first sheath liquid physical property information in the incidental measurement condition storage unit;
calculating a first feature quantity from a first pulse waveform obtained from the first flow cytometry measurement means as the known particles pass through the first measurement unit of the first flow cytometry measurement means; death,
creating a learned AI program by learning the AI program using the first feature amount and the first incidental measurement condition as teacher data and the type of the known particle as a teacher label;
Second light source characteristic information, second light receiver angle information, second flow velocity information, and second fluorescent label representing the characteristics of the second flow cytometry measuring means used in the second measurement used to measure unknown particles storing second incidental measurement conditions including one or more selected from the group consisting of characteristic information and second sheath liquid physical property information in the incidental measurement condition storage unit;
a second feature amount calculated by the feature amount extraction unit based on a second pulse waveform obtained by the unknown particle passing through the second measurement unit of the second flow cytometry measurement means; An apparatus configured to input the second contingent measurement condition into the learned AI program to detect, identify, or quantify the particles.
An AI program for identifying first type particles and second type particles mixed in a sample,
The instructions in the AI program, when executed by a processor, include:
A transient change in ionic current between two electrodes in contact with the electrolyte on both sides of the pore when one particle dispersed in the electrolyte passes through the pore is detected as one pulse signal. a pore device comprising:
A feature quantity extraction unit that calculates N feature quantity groups for each pulse waveform;
Using a Coulter measuring device equipped with an AI program that learns pulse waveforms,
The first pulse waveform obtained by measuring the first sample with the Coulter measuring device is clustered in an N-dimensional feature amount space based on the feature amount vector having the feature amount group calculated for each of the first pulse waveforms. classifying into a first estimated pulse waveform group estimated to be particles of the first type and a second estimated pulse waveform group estimated to be particles of the second type;
The second pulse waveform obtained by measuring the second sample with the Coulter measurement device is clustered in an N-dimensional feature amount space based on the feature amount vector having the feature amount group calculated for each of them as an element. classifying into the first estimated pulse waveform group and the second estimated pulse waveform group,
Each of the feature quantity groups calculated from the first estimated pulse waveform group is used as teacher data, the first type is used as a teacher label, and the feature quantity group calculated from the second estimated pulse waveform group is An AI program characterized by being configured to learn the AI program using each as teacher data and the second type as a teacher label, and to create a learned AI discriminator.
When executed by a processor, the instructions contained in the AI program further:
The feature quantity group calculated from each of the third pulse waveform group obtained by measuring the third sample with the third Coulter measuring device is input to the learned AI discriminator, and the third pulse waveform group is obtained. 4. The AI program according to claim 3, wherein each of the AI programs is configured to identify whether it belongs to the first type or the second type.
5. The AI program according to claim 3, wherein N>3.
An AI program for estimating the presence or absence of a protein to be detected in a sample,
The AI program is
By mixing anti-protein beads modified with an antibody that specifically binds to the detection target protein into a sample known to contain the detection target protein and contaminants, the detection target protein is bound therein. a first mixed sample prepared containing bound beads;
using a second mixed sample prepared by mixing the anti-protein beads with a sample known to contain the contaminants but not the protein to be detected;
configured to detect, as a pulse waveform, transient changes in ionic current between two electrodes in contact with the electrolyte on both sides of the pore when particles dispersed in the electrolyte pass through the pore. used in measurements by Coulter measuring devices,
In the measurement,
The first mixed sample is measured by the Coulter measuring device to obtain a first pulse measurement result, and the second mixed sample is measured by the Coulter measuring device to obtain a second pulse measurement result. can be,
The instructions in the AI program, when executed by a processor, include:
calculating a contamination pulse probability that the first pulse measurement result is not the bound bead by calculating the first pulse measurement result and the second pulse measurement result;
Learning the AI program using the pulse measurement result of the pulse whose probability of the mixed pulse is lower than the threshold as teacher data and positive as a correct label,
An AI program, wherein the AI program is learned using the second pulse measurement result as teacher data and negative as a correct label to create a learned AI detector.
Said measurement further comprises:
A third mixed sample prepared by mixing the anti-protein beads with the unknown sample in which the presence or absence of the protein to be detected is unknown is also used,
In the measurement,
A third pulse measurement result is obtained by measuring the third mixed sample with the Coulter measurement device,
The instructions contained in the AI program, when executed by a processor, further:
The third pulse measurement result is input to the learned AI detector, and whether or not the unknown sample contains the protein to be detected is output. 7. The AI program according to item 6.
A method using an AI program for detecting whether or not a microorganism to be detected is contained in a sample,
A bead-mixed negative sample is prepared by mixing antimicrobial beads modified with an antibody that specifically binds to the microorganism to be detected to a known negative sample that does not contain the microorganism to be detected,
Mixing the antimicrobial beads with a known positive sample containing the microorganism to be detected to prepare a bead-mixed positive sample containing microorganism-bound beads to which the microorganism to be detected and the antimicrobial beads are bound,
A Coulter device configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of a pore when fine particles of the bead-mixed negative sample dispersed in an electrolytic solution pass through the pore. By measuring with, a negative pulse measurement result is obtained,
A positive pulse measurement result is obtained by measuring the bead-mixed positive sample with the Coulter device,
Using an AI program, AI learning is performed using the negative pulse measurement result as teacher data and the negative label as the correct answer,
A method comprising using the AI program to perform AI learning using the positive pulse measurement result as teacher data and the positive label as a correct answer to create a trained AI detector.
creating a bead-mixed unknown sample by mixing the antimicrobial beads with an unknown sample in which it is unknown whether or not the microorganism to be detected is included;
Using the AI program, the unknown pulse measurement result obtained by measuring the bead-mixed unknown sample with the Coulter device is input to the learned AI detector, and the detection target microorganism is present in the unknown sample. 9. The method of claim 8, further comprising detecting whether to do so.
A method using an AI program for detecting whether the sample contains exosomes to be detected,
Create a bead-mixed negative sample by mixing anti-exosome beads modified with an antibody that specifically binds to the exosomes to be detected in a known negative sample that does not contain exosomes to be detected,
The anti-exosome beads are mixed with the known positive sample containing the exosomes to be detected, and the exosomes to be detected and the anti-exosome beads are bound to create a bead mixed positive sample containing exosome-bound beads,
A Coulter device configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of a pore when fine particles of the bead-mixed negative sample dispersed in an electrolytic solution pass through the pore. By measuring with, a negative pulse measurement result is obtained,
A positive pulse measurement result is obtained by measuring the bead-mixed positive sample with the Coulter device,
Using an AI program, AI learning is performed using the negative pulse measurement result as teacher data and the negative label as the correct answer,
A method comprising using the AI program to perform AI learning using the positive pulse measurement result as teacher data and the positive label as a correct answer to create a trained AI detector.
Create a bead-mixed unknown sample by mixing the anti-exosome beads with an unknown sample that is unknown whether or not it contains the exosomes to be detected,
Using the AI program, the unknown pulse measurement results obtained by measuring the bead-mixed unknown sample with the Coulter device are input to the learned AI detector, and the detection target exosomes are present in the unknown sample. 11. The method of claim 10, further comprising detecting whether to do so.
A method for detecting whether a target protein is contained in a sample, comprising:
Configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal only when detectable particles having a size equal to or larger than the detection limit size dispersed in the electrolytic solution pass through the pore. with a Coulter device that
preparing a first sample by mixing anti-protein beads that specifically bind to the protein to be detected with a negative sample that does not contain a protein having a size less than the detection limit size;
Mixing the anti-protein beads with the positive sample containing the protein to create a second sample containing bound beads in which the anti-protein beads are bound to each other;
A first pulse number of the first pulse measurement result obtained by measuring the first sample with the Coulter device, and a second pulse measurement result of the second pulse measurement result obtained by measuring the second sample with the Coulter device Determine the positive pulse number threshold from the number of pulses,
For the unknown sample in which the third pulse number in the third pulse measurement result obtained by measuring the unknown sample in which it is unknown whether or not the protein to be detected is contained is above the positive pulse number threshold, A method comprising determining that the protein to be detected is included.
13. The method of claim 12, wherein said anti-protein bead size is less than said detection limit size and said bound bead size is greater than said detection limit size.
14. The method of claim 12 or 13, wherein the positive pulse number threshold is less than or equal to 1/5 of the first pulse number.
A method using an AI program for detecting whether or not a microorganism to be detected is contained in a sample,
A bead-mixed negative sample is prepared by mixing labeled beads modified with an antibody that specifically binds to the microorganism to be detected to a known negative sample that does not contain the microorganism to be detected,
Centrifuging the negative sample container containing the bead-mixed negative sample,
The sediment in the lower part of the negative sample container is configured to detect, as a pulse signal, a transient change in electrical resistance between both sides of the pore when fine particles dispersed in the electrolytic solution pass through the pore. By measuring with the Coulter device that is used, a concentrated negative pulse measurement result is obtained,
Mixing the labeled beads with a known positive sample containing the microorganism to be detected to prepare a bead-mixed positive sample containing microorganism-bound beads to which the microorganism to be detected and the labeled beads are bound,
Centrifuging the positive sample container containing the bead-mixed positive sample,
By measuring the sediment at the bottom of the positive sample container with the Coulter device, a concentrated positive pulse measurement result is obtained,
Using an AI program, AI learning is performed using the concentrated negative pulse measurement result as teacher data and negative as a correct label,
using the AI program to create a trained AI detector by performing AI learning using the enriched positive pulse measurement results as teacher data and using positive as a correct label.
mixing the labeled beads with an unknown sample in which it is unknown whether or not it contains the microorganism to be detected to prepare a bead-mixed unknown sample;
Using the AI program, the unknown pulse measurement result obtained by measuring the bead-mixed unknown sample with the Coulter device is input to the learned AI detector, and the detection target microorganism is present in the unknown sample. 16. The method of claim 15, further comprising detecting whether to do so.
A method using an AI program for identifying which of two types of microorganisms are present in a sample, comprising:
A first sample known to contain a first microorganism, a first antimicrobial bead modified with an antibody that specifically binds to said first microorganism and an antibody that specifically binds to a second microorganism A first mixed bead sample containing the first microorganism-bound beads bound to the first microorganism and the first anti-microbial bead by mixing two kinds of mixed beads consisting of the second anti-microbial beads modified with and create
A second sample known to contain a second microorganism is mixed with the two-species mixed beads to form a second microorganism-bound bead in which the second microorganism and the second antimicrobial bead are bound. creating a second bead mixture sample comprising
The first bead mixture sample is configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal when the fine particles dispersed in the electrolyte pass through the pore. A first pulse measurement result is obtained by measuring with a Coulter device,
obtaining a second pulse measurement result by measuring the second bead-mixed sample with the Coulter device;
Using an AI program, AI learning is performed using the first pulse measurement result as teacher data and the first microorganism as a teacher label,
A method comprising using the AI program to perform AI learning using the second pulse measurement result as teacher data and the second microorganism as a teacher label to create a learned AI discriminator.
creating a bead-mixed unknown sample by mixing the two kinds of mixed beads into an unknown sample in which it is unknown which of the first microorganism and the second microorganism is included;
Using the AI program, the unknown pulse measurement result obtained by measuring the bead-mixed unknown sample with the Coulter device is input to the learned AI classifier, and the first microorganism and the unknown sample are included in the unknown sample. 18. The method of claim 17, further comprising identifying which of said second microorganisms are present.
19. The method of claims 17-18, wherein the first antimicrobial beads and the second antimicrobial beads differ in at least one of shape, material and/or size.
A method using an AI program for quantifying the concentration of protein contained in a sample, comprising:
A first sample known to contain the protein at a first concentration is mixed with labeled beads modified with an antibody that specifically binds to the protein to form a first sample containing a conjugate of the protein and the labeled beads. Create a bead mixture sample,
mixing the labeled beads with a second sample known to contain a second concentration of the protein to form a second bead mixture sample containing a conjugate of the protein and the labeled beads;
The first bead mixture sample is configured to detect a transient change in electrical resistance between both sides of the pore as a pulse signal when the fine particles dispersed in the electrolyte pass through the pore. Obtaining a first teacher pulse measurement result extracted by the pulse extraction means from the measurement result by the Coulter device,
Obtaining a second teacher pulse measurement result extracted by the pulse extraction means from the measurement result of the second bead mixed sample by the Coulter device,
Using an AI program, AI learning is performed using the first teacher pulse measurement result as teacher data and the first concentration as a correct label,
and creating a learned AI quantifier by performing AI learning using the AI program with the second teacher pulse measurement result as teacher data and the second concentration as a correct label.
mixing the labeled beads with an unknown sample with an unknown protein concentration to prepare a bead-mixed unknown sample;
Using the AI program, the unknown pulse measurement result extracted by the pulse extraction means from the measurement result of measuring the bead-mixed unknown sample with the Coulter device is input to the learned AI quantifier, and the unknown sample is 21. The method of claim 20, further comprising estimating the concentration of said protein contained therein.
the size of the labeled bead is below the lower limit size of extraction that can be extracted as a pulse signal by the pulse extraction means;
22. The method of any one of claims 20-21, wherein the size of the conjugate exceeds the lower extraction limit size.