WO2025142539A1 - Biological particle analysis system, information processing device, and information processing method - Google Patents

Abstract

A biological particle analysis system according to the present disclosure comprises: an acquisition unit that acquires measurement data measured from biological particles included in a sample; a compression unit that performs a data compression process on the measurement data acquired by the acquisition unit; a gate unit that gates the measurement data compressed by the compression unit into measurement data for learning and measurement data for verification, and adds a label to the measurement data for learning; a learning unit that constructs a learning model using the measurement data for learning and the label; an estimation unit that inputs the measurement data for verification to the learning model and outputs the degree of certainty of the measurement data for verification; and a threshold value setting unit that, on the basis of the degree of certainty, sets a threshold value for separating the sample.

Description

Biological particle analysis system, information processing device, and information processing method

This disclosure relates to a bioparticle analysis system, an information processing device, and an information processing method.

In fields such as medicine and biochemistry, it is common to use flow cytometers to rapidly measure the characteristics of large numbers of particles. A flow cytometer is a device that measures the characteristics of each particle by irradiating flowing particles such as cells or beads with a beam of light and detecting the fluorescence emitted by the particles.

In addition, devices have been developed that separate particles that emit specific fluorescence from a measurement sample by controlling the movement of the particles based on the fluorescence information detected by a flow cytometer. Such separation devices are also called cell sorters.

In recent years, studies have been conducted on flow cytometers to enable more detailed analysis of particles by increasing the number of fluorescent substances that can be measured at one time. However, increasing the number of fluorescent substances increases the number of dimensions of the measurement data, making analysis by the flow cytometer more complicated.

Therefore, various methods for analyzing measurement data in flow cytometers have been investigated. For example, the following Patent Document 1 discloses a technique for estimating shape information of a biological object based on the peak position of a pulse waveform detected from the biological object irradiated with a light beam.

JP 2017-58361 A

On the other hand, in a cell sorter or other sorting device, it is necessary to measure and analyze the flowing particles, and to determine whether or not to sort the particles based on the measurement and analysis results within the limited time that the particles flow through the device.

Therefore, there is a need for cell sorters and other sorting devices to be able to determine more quickly and in real time whether or not a particle is a target for sorting.

The bioparticle analysis system of the first disclosure includes an acquisition unit that acquires measurement data measured from bioparticles contained in a sample, a compression unit that performs a data compression process on the measurement data acquired by the acquisition unit, a gating unit that gates the measurement data compressed by the compression unit into training measurement data and verification measurement data and adds a label to the training measurement data, a learning unit that constructs a learning model using the training measurement data and the label, an estimation unit that inputs the verification measurement data to the learning model and outputs a confidence level of the verification measurement data, and a threshold setting unit that sets a threshold for separating the sample based on the confidence level.

1 is a block diagram showing an example of the configuration of a biological particle analysis system according to an embodiment. FIG. 4 is an explanatory diagram illustrating a filter-type detection mechanism of the measurement unit. FIG. 2 is an explanatory diagram illustrating a spectral detection mechanism of the measurement unit. 2 is a block diagram showing a configuration example of an information processing device according to the embodiment. FIG. FIG. 4 is a table showing an example of information regarding the fluorescence of biogenic particles obtained from the sorting device. FIG. 11 is an explanatory diagram showing a result of a clustering process. FIG. 11 is an explanatory diagram showing a result of a clustering process. FIG. 13 is an explanatory diagram showing the results of dimensionality reduction processing of information on the expression level of each fluorescent substance in a biological particle to two dimensions using the t-SNE algorithm. FIG. 4 is a diagram showing verification data according to the first embodiment; FIG. 13 is a diagram showing a screen showing the purity and efficiency of dimensionally reduced measurement data according to the first embodiment. FIG. 11 is a diagram showing classes and confidence levels of dimension-reduced measurement data according to the first embodiment; FIG. 13 is a diagram showing a screen showing the relationship between the mode and purity and efficiency in the first embodiment. 10A and 10B are diagrams for explaining a case where a threshold is set for each measurement data according to the first embodiment; 10 is a diagram for explaining a case where a threshold is set using an ROC curve of verification measurement data in the first embodiment. FIG. FIG. 11 is a diagram illustrating a display example of dimensionally compressed measurement data according to the first embodiment. 11A and 11B are diagrams showing an example of display in which measurement data before fluorescence correction of cells to be measured is displayed in different colors according to similarity in the first embodiment; 11A and 11B are diagrams showing an example of displaying measurement data after fluorescence correction of cells as measurement objects in a manner that changes color according to similarity in the first embodiment; 1 is a functional block diagram illustrating a configuration of an information processing device according to a first embodiment for dividing measurement data in deep learning. FIG. 11 is a flowchart for explaining the sorting of measurement data in deep learning of the information processing device according to the first embodiment. FIG. 11 is a block diagram showing a configuration example of a biological particle analysis system according to a modified example of the first embodiment. FIG. 11 is a functional block diagram of an information processing system according to a modified example of the first embodiment. FIG. 11 is a functional block diagram illustrating a modified example of the information processing system according to the first embodiment. FIG. 13 is a diagram for explaining the concept of thresholds for clustering sorting according to the second embodiment. FIG. 13 is a diagram for explaining the concept of the range when the threshold value is set to 50% in the clustering sorting according to the second embodiment. FIG. 11 is a functional block diagram of an information processing device according to a second embodiment for performing clustering sorting; FIG. 13 is a diagram illustrating a first example of a circuit of FlowSOM according to a second embodiment. FIG. 13 is a diagram illustrating a second example of a circuit of FlowSOM according to the second embodiment. FIG. 13 is a diagram illustrating a third example of a circuit of FlowSOM according to the second embodiment. 13 is a flowchart for explaining clustering sorting of an information processing device according to a second embodiment. FIG. 13 is a functional block diagram of an information processing system according to a modified example of the second embodiment. 11 is a flowchart for explaining the operation of a first example of a circuit of FlowSOM according to a second embodiment; 13 is a flowchart for explaining the operation of a third example of the circuit of FlowSOM according to the second embodiment. FIG. 13 is a functional block diagram showing IFCM fractionation of an information processing device according to a third embodiment. 13 is a flowchart for explaining IFCM sorting in an information processing device according to a third embodiment. FIG. 13 is a functional block diagram of an information processing system according to a modified example of the third embodiment. FIG. 2 is a hardware configuration diagram illustrating an example of a computer that realizes a calculation unit of the information processing device according to the embodiment.

Below, a preferred embodiment of the present disclosure will be described in detail with reference to the attached drawings. In this specification and drawings, components having substantially the same functional configuration will be denoted by the same reference numerals to avoid duplicated explanations. The explanation will be given in the following order.

0. Basic Concept 0.1. Configuration of Bioparticle Analysis System 0.2. Configuration of Information Processing Device 1. First Embodiment 1.1. Fractionation Based on Certainty Factor 1.2. How to Use Certainty Factor 1.3. Setting of Threshold Value 1.3.1. Threshold Value Setting Method 1
1.3.2. Threshold setting method 2
1.3.3. Threshold setting method 3
1.4. Visualization using similarity and certainty 1.5. Functional block diagram of information processing device 300 1.6. Operational description 1.7. Modification 2. Second embodiment 2.1. Sorting (clustering) based on certainty
2.2. Threshold for clustering sorting 2.2.1. When threshold judgment is performed for each parameter 2.2.2. When threshold judgment is performed by averaging all parameters 2.3. Functional block diagram of information processing device 400 2.4. FlowSOM circuit 2.5. Operation description 2.6. Modification 2.7. Flowchart for FlowSOM sorting 3. Third embodiment 3.1. Sorting based on confidence factor (imaging flow cytometer)
3.2. Functional block diagram of information processing device 600 3.3. Operational description 3.4. Modifications 4. Hardware configuration

<0. Basic concept>
In recent years, a method called machine learning sorting has been developed that uses machine learning to separate target cells from a sample containing cells, etc., based on measurement data (including, for example, the intensity of fluorescence emitted from labeled cells, etc.) The basic concept of machine learning sorting is disclosed in Patent Document 2, and the contents of Patent Document 2 may be referred to as appropriate in the present disclosure.

JP 2020-193877 A

<0.1. Configuration of bioparticle analysis system>
A bioparticle analysis system is described.

First, the configuration of a bioparticle analysis system 1 according to an embodiment will be described with reference to FIG. 1. FIG. 1 is a block diagram showing an example of the configuration of a bioparticle analysis system 1 according to an embodiment.

As shown in FIG. 1, the bioparticle analysis system 1 according to this embodiment includes a fractionation device 10 that acquires measurement data from a sample S and fractionates particles to be sorted based on the judgment of an information processing device 20, and an information processing device 20 that analyzes the measurement data acquired by the fractionation device 10 and judges whether the particles are to be sorted. The bioparticle analysis system 1 can be used, for example, as a so-called cell sorter.

The sample S is, for example, a particle of biological origin, such as a cell, a microorganism, or a biologically-related particle, and includes a plurality of groups of biologically-derived particles. The sorting device 10 can classify the biologically-derived particles into a plurality of groups that are internally bound and externally separated, and sort a specific group by analyzing the measurement data of the sample S. The sample S may be, for example, a cell such as an animal cell (e.g., a blood cell), or a plant cell, a microorganism such as bacteria such as E. coli, viruses such as tobacco mosaic virus, or fungi such as yeast, a biologically-related particle that constitutes a cell, such as a chromosome, a liposome, a mitochondria, or various organelles (cell organelles), or a biologically-related macromolecule, such as a nucleic acid, a protein, a lipid, a sugar chain, or a complex of these, which is derived from a biological organism.

The sample S includes, for example, synthetic particles such as latex particles, gel particles, and industrial particles. The industrial particles may be, for example, organic or inorganic polymeric materials, metals, etc. Organic polymeric materials include polystyrene, styrene-divinylbenzene, polymethyl methacrylate, etc. Inorganic polymeric materials include glass, silica, magnetic materials, etc. Metals include gold colloids, aluminum, etc. The shape of these microparticles may be spherical, but may also be non-spherical. The microparticles may have a cavity and may be configured to capture biological particles in the cavity. The size and mass of these microparticles may be appropriately selected by those skilled in the art and are not particularly limited.

Here, the sample S is labeled (stained) with one or more fluorescent dyes. The labeling of the sample S with the fluorescent dyes can be performed by known methods. For example, if the sample S is a cell, the cells to be measured can be labeled with the fluorescent dye by mixing a fluorescently labeled antibody that selectively binds to an antigen present on the cell surface with the cells to be measured and binding the fluorescently labeled antibody to the antigen on the cell surface.

A fluorescently labeled antibody is an antibody to which a fluorescent dye is bound as a label. Specifically, a fluorescently labeled antibody may be an antibody to which avidin-bound fluorescent dye is bound by an avidin-biodin reaction, which is then bound to a biotin-labeled antibody. Alternatively, a fluorescently labeled antibody may be an antibody to which a fluorescent dye is directly bound. The antibody may be either a polyclonal antibody or a monoclonal antibody. Furthermore, the fluorescent dye for labeling cells is not particularly limited, and at least one or more well-known dyes used for staining cells, etc. may be used.

The fractionation device 10 includes a measurement unit and a fractionation unit. The fractionation device 10 may be a so-called flow cell type fractionation device 10, or may be a microchannel chip type fractionation device.

The measurement unit measures the fluorescence emitted from the sample S by irradiating the sample S with a beam of light such as a laser beam. Specifically, the measurement unit aligns the sample S in one direction by forming a laminar flow in the sheath liquid in which the sample S is dispersed. At this time, the measurement unit irradiates the aligned sample S with a laser beam having a wavelength capable of exciting the fluorescent dye that labels the sample S, and photoelectrically converts the fluorescence generated from the sample S irradiated with the laser beam using a known photoelectric conversion element such as a CCD (Charge Coupled Device), CMOS (Complementary Metal Oxide Semiconductor), photodiode, or PMT (Photo Multiplier Tube). This allows the measurement unit to acquire the fluorescence from the sample S.

The detection mechanism for the fluorescence from the sample S in the measurement unit may be either a filter type or a spectral type. Here, the detection mechanism for the fluorescence from the sample S will be described with reference to Figures 2 and 3. Figure 2 is an explanatory diagram for explaining the detection mechanism of the filter type, and Figure 3 is an explanatory diagram for explaining the detection mechanism of the spectral type.

As shown in Figure 2, in the filter-type detection mechanism, the sample S flowing through the flow path 13 is irradiated with light from the light source 11, and the resulting fluorescence is separated by dichroic mirrors 15A, 15B, and 15C. As a result, the filter-type detection mechanism can obtain the intensity of the fluorescence for each predetermined wavelength band using photodetectors 17A, 17B, and 17C.

Specifically, dichroic mirrors 15A, 15B, and 15C are mirrors that reflect light in specific wavelength bands and transmit light in other wavelength bands. As a result, the measurement unit can separate the fluorescence into wavelength bands by providing dichroic mirrors 15A, 15B, and 15C that reflect light in different wavelength bands on the optical path of the fluorescence from sample S. For example, the measurement unit can separate the fluorescence from sample S into wavelength bands by providing, in order from the side where the fluorescence from sample S is incident, dichroic mirror 15A that reflects light in the red wavelength band, dichroic mirror 15B that reflects light in the green wavelength band, and dichroic mirror 15C that reflects light in the blue wavelength band.

As shown in FIG. 3, in the spectral detection mechanism, the sample S passing through the flow path 13 is irradiated with light from the light source 11, and the resulting fluorescence is dispersed by the prism 16. This allows the spectral detection mechanism to obtain a continuous fluorescence spectrum at the photodetector array 18.

Specifically, the prism 16 is an optical element that disperses the incident light. As a result, the measurement unit disperses the fluorescence from the sample S using the prism 16, and is able to detect a continuous spectrum of the fluorescence using the photodetector array 18, which has multiple photoelectric conversion elements arranged in an array.

The fractionation unit fractionates a portion of the sample S to be fractionated. Specifically, the fractionation unit first generates droplets of the sample S and charges the droplets of the sample S to be fractionated. Next, the fractionation unit moves the generated droplets into the electric field generated by the polarizing plate. At this time, the charged droplets are attracted to the charged polarizing plate, so the direction of movement of the droplets is changed. This allows the fractionation unit to separate droplets of the sample S to be fractionated from droplets of the sample S that are not to be fractionated, making it possible to fractionate the biological particles to be fractionated. The fractionation method of the fractionation unit may be either a jet-in-air method or a cuvette flow cell method. The sample S may be fractionated by being ejected outside the flow cell or the microchannel chip, or may be fractionated inside the microchannel chip. The decision as to whether or not to collect sample S may be made by a logic circuit (e.g., a field-programmable gate array (FPGA) circuit) provided in the collection device 10, or may be made based on an instruction from the information processing device 20.

The information processing device 20 analyzes the measurement data of the sample S acquired by the measurement unit and presents the analyzed data to the user. By checking the data analyzed by the information processing device 20, the user can identify the group of biological particles to be separated.

<0.2. Configuration of information processing device>
Next, a more specific configuration of the information processing device 20 included in the biological particle analysis system 1 according to the present embodiment will be described with reference to Fig. 4. Fig. 4 is a block diagram showing an example of the configuration of the information processing device 20 according to the present embodiment.

As shown in FIG. 4, the information processing device 20 includes an acquisition unit 201, an analysis unit 203, a reference spectrum storage unit 205, a data compression processing unit 207, an interface unit 209, a learning unit 211, a learning model storage unit 213, and a discrimination unit 215.

The acquisition unit 201 acquires information about the fluorescence of the biological particles from the sorting device 10. Specifically, the sorting device 10 detects the light of the biological particles using a spectral detection mechanism, and the acquisition unit 201 acquires information about the spectrum of the light of the biological particles. The light of the biological particles may be either scattered light or fluorescence from the biological particles irradiated with laser light, or it may be both. The acquisition unit 201 may acquire information about the light of the biological particles from the sorting device 10 via a network, for example, or may acquire information about the light of the biological particles from the sorting device 10 via a wired or wireless LAN (Local Area Network) or a wired cable.

For example, the information about the light of the biological particles acquired by the acquisition unit 201 may be information as shown in FIG. 5. FIG. 5 is a table showing an example of information about the light of the biological particles acquired from the sorting device 10.

As shown in FIG. 5, the information about the light of the biogenic particles may be represented by the gains detected by N photomultiplier tubes (PMTs) arranged in a photodetector array for each identification number of a cell (i.e., a biogenic particle) as "PMT1" to "PMTN". These N photomultiplier tubes are arranged in a line in an array in the direction of light dispersion by a prism. Therefore, by continuously arranging the gains of these N photomultiplier tubes as a histogram, the spectrum of the light of the cell can be obtained. FIG. 5 shows the measurement results of the gains of N photomultiplier tubes for each of N cells.

The analysis unit 203 derives information about the characteristics of the biological particles by analyzing information about the light of the biological particles measured by the sorting device 10. Specifically, the analysis unit 203 separates each of the fluorescent light contained in the fluorescent spectrum measured by the sorting device 10, and derives the expression amount in the biological particles of the fluorescent substance corresponding to each of the fluorescent light.

The biological particles to be measured are labeled with multiple fluorescent substances that emit fluorescence with overlapping wavelength distributions. Therefore, the analysis unit 203 can derive the expression level of each fluorescent substance by weighting and fitting the wavelength distribution of the fluorescence emitted from each fluorescent substance to the fluorescence spectrum measured by the fractionation device 10.

More specifically, first, the analysis unit 203 acquires from the reference spectrum storage unit 205 reference spectra indicating the wavelength distribution of the fluorescence emitted by the fluorescent substances labeling the biological particles. Next, the analysis unit 203 superimposes the reference spectra of each fluorescent substance and fits them to the fluorescence spectrum measured by the fractionation device 10 using the weighted least squares method, thereby estimating the expression level of each fluorescent substance.

The reference spectrum storage unit 205 stores reference spectra that indicate the wavelength distribution of fluorescence emitted by fluorescent substances capable of labeling biological particles. The reference spectrum storage unit 205 may be provided in either the information processing device 20 or the fractionation device 10, or may be provided in another information processing device or information processing server that can communicate via a network.

The data compression processing unit 207 performs data compression processing on the optical information of the biological particles analyzed by the analysis unit 203.

Data compression processing includes both nonlinear processing and linear processing. For example, nonlinear processing may include dimensionality reduction processing, clustering processing, or grouping processing. For example, linear processing may include processing for generating fluorescence information for each fluorescent dye from the optical spectrum information of biological particles by performing fluorescence separation.

Note that any algorithm, whether supervised or unsupervised machine learning or weakly supervised machine learning, may be used for the nonlinear processing. However, it is preferable that the machine learning algorithm used for the nonlinear processing is different from the machine learning algorithm used by the learning unit 211 described below.

Specifically, the data compression processing unit 207 may perform clustering processing on information related to the expression level of each fluorescent substance in the bioparticles. In this way, the data compression processing unit 207 can classify the bioparticles into a plurality of groups that are externally separated and internally combined.

The algorithm for the clustering process is not particularly limited, and any known clustering algorithm can be used. For example, the data compression processing unit 207 may perform the clustering process using an algorithm that can specify the number of clusters, such as k-means, or may perform the clustering process using an algorithm that automatically determines the number of clusters, such as flowsom.

The results of the clustering process performed by the data compression processing unit 207 may be presented to the user in a format as shown in Figures 6 and 7. Figures 6 and 7 are explanatory diagrams showing the results of the clustering process.

For example, as shown in FIG. 6, the clustering results by the data compression processing unit 207 may be presented to the user in a table format.

In FIG. 6, a group of 1000 cells (i.e., biological particles) is divided into N clusters, and the identification numbers assigned to the clusters and cells indicate which cells belong to which cluster. Specifically, in FIG. 6, the cells with identification numbers "1", "2", "3", and "10" belong to the cluster with identification number "1", the cells with identification numbers "11", "12", "22", and "31" belong to the cluster with identification number "2", the cells with identification numbers "4" to "6", "14", and "15" belong to the cluster with identification number "3", and the cell with identification number "1000" belongs to the cluster with identification number "N". By presenting the data to the user in such a tabular format, the cell's belonging to each cluster can be simply shown.

For example, as shown in FIG. 7, the clustering results by the data compression processing unit 207 may be presented to the user in a minimum spanning tree format.

In Figure 7, radar charts painted in multiple colors (in Figure 7, colors are distinguished by the type of hatching) are arranged in a tree shape that is connected to each other. Each radar chart represents a cell (i.e., a biological particle). Specifically, the distribution and size of each radar chart represents a vector corresponding to the expression level of each fluorescent substance in the cell. Here, the areas painted in each color represent the cluster to which each cell belongs. For example, cells shown in radar charts painted in the same color (i.e., the same type of hatching) belong to the same cluster.

Furthermore, in Figure 7, the distance between radar charts corresponds to the similarity between the cells represented by the radar charts. In other words, Figure 7 shows that cells represented by radar charts that are close to each other are similar to each other, and cells represented by radar charts that are far from each other are dissimilar to each other. By presenting the data to the user in this minimum spanning tree format, it is possible to show the similarity between the cells in addition to the cluster affiliation of the cells.

Alternatively, the data compression processing unit 207 may perform a dimensionality compression process on information relating to the expression levels of each fluorescent substance in the biological particles. In this way, the data compression processing unit 207 can compress the dimensions of high-dimensional data including the expression levels of multiple fluorescent substances, thereby making it possible to easily visualize the relationships between each piece of high-dimensional data on a low-dimensional map. Therefore, by checking the low-dimensional information after the dimensionality compression process, the user can classify the biological particles into multiple groups more easily than with the high-dimensional information before the dimensionality compression process. The data compression processing unit 207 only needs to be able to perform a dimensionality compression process that reduces the number of dimensions by at least one or more, but for example, by compressing the dimensions of the information relating to the expression levels of each fluorescent substance in the biological particles to three dimensions or less, it is possible to more clearly visualize the relationships between each piece of high-dimensional data.

The algorithm for the dimensionality compression process is not particularly limited, and any known dimensionality compression algorithm can be used. For example, the data compression processing unit 207 may perform the dimensionality compression process using an algorithm such as PCA, t-SNE, or Umap.

The results of the dimensionality compression process by the data compression processing unit 207 may be presented to the user in a format as shown in FIG. 8. FIG. 8 is an explanatory diagram showing the results of dimensionality compression process of information on the expression levels of each fluorescent substance in biological particles to two dimensions using the t-SNE algorithm.

For example, in Figure 8, the Euclidean distance of high-dimensional data, such as the expression levels of each fluorescent substance in a cell, is converted into a probability using the probability distribution of Student's t-distribution and mapped onto two-dimensional coordinates. This allows the user to more simply compare the similarity of the expression levels of each fluorescent substance in cells without having to compare the expression levels of each fluorescent substance individually. For example, in Figure 8, cells belonging to the same population are shown in different colors. Referring to Figure 8, it can be seen that the dimensionality reduction process allows cells belonging to the same population to be grouped with appropriate internal connections and external separation.

The interface unit 209 includes an output device and an input device, and performs input and output of information with the user. Specifically, the interface unit 209 may present the information after nonlinear processing by the data compression processing unit 207 to the user using a CRT (Cathode Ray Tube) display device, a liquid crystal display device, an OLED (Organic Light Emitting Diode) display device, or the like. The interface unit 209 may also accept user input specifying the biological particles to be separated, using an input device such as a touch panel, a keyboard, a mouse, a button, a microphone, a switch, or a lever.

By checking the information after data compression processing output from the interface unit 209, the user can more easily specify the population of biological particles to be separated. For example, by checking the information after clustering processing, the user can identify the cluster of biological particles to be separated. Alternatively, by checking the information after dimensional compression processing, the user can specify the range of the population of biological particles to be separated.

The construction of the learning model performed by the learning unit 211 will be described later.

The constructed learning model may be stored, for example, in a learning model storage unit 213 provided in the information processing device 20. In this way, the sorting device 10 can sort the biological particles to be sorted by sorting control from the information processing device 20. Alternatively, the constructed learning model may be implemented in a logic circuit such as an FPGA circuit provided in the sorting device 10. For example, the sorting device 10 is provided with a discrimination unit 215, and the FPGA circuit provided in the sorting device 10 may be implemented with logic designed based on the type of discrimination unit 215 and for executing the constructed learning model. The logic for executing the constructed learning model may be designed by the learning unit 211.

The machine learning algorithm performed by the learning unit 211 is supervised learning that uses information on the fluorescence spectrum of the biological particles identified as the target for sorting as a teacher. For example, the learning unit 211 may construct a learning model using a machine learning algorithm such as a random forest, a support vector machine, or deep learning.

The bioparticle analysis system 1 according to this embodiment uses various non-standardized information as training data, and therefore can suitably use a random forest machine learning algorithm that does not require standardization. In addition, the random forest machine learning algorithm is easy to implement as a learning model in hardware, and therefore can be suitably used in the bioparticle analysis system 1 according to this embodiment, in which it is important to quickly determine whether or not a particle of biological origin is a target for separation.

The learning unit 211 may determine whether a learning model capable of sufficiently identifying the separation target has been constructed, and notify the user. For example, when the number of pieces of learned information on biological particles, or the proportion of the total, exceeds a threshold, the learning unit 211 may notify the user that a learning model capable of sufficiently identifying the separation target has been constructed.

In addition, when the accuracy rate of the learning model exceeds a threshold, the learning unit 211 may notify the user that a learning model capable of sufficiently discriminating between the separation targets has been constructed. The accuracy rate of the learning model can be determined, for example, by N-fold-cross validation. Specifically, the entire information to be used as a teacher is divided into N parts, learning is performed using the information contained in the N-1 divided parts to construct a learning model, and then the accuracy rate of the constructed learning model can be determined by discriminating between the information contained in the remaining divided part.

The learning model storage unit 213 stores the learning model constructed by the learning unit 211. The learning model storage unit 213 may store the learning model in the form of hardware using a field-programmable gate array (FPGA) circuit or the like. This makes it possible to more quickly determine whether or not a biological particle is a target for separation.

The discrimination unit 215 discriminates whether or not the fluorescent biogenic particles measured by the fractionation device 10 are to be separated, based on the learning model stored in the learning model storage unit 213. If it is determined that the biogenic particles are to be separated, the discrimination unit 215 instructs the fractionation device 10 to separate the biogenic particles.

The learning model storage unit 213 and the discrimination unit 215 may be provided in the fraction collection device 10.

Furthermore, when the sorting device 10 is capable of separately sorting a plurality of groups of biogenic particles, the discrimination unit 215 may instruct the sorting device 10 not only whether or not the biogenic particles are to be sorted, but also which collection unit the biogenic particles should be collected in. In such a case, the learning unit 211 performs machine learning using as training data information on the fluorescence spectrum of the biogenic particles that further specifies which collection unit the biogenic particles should be collected in after sorting. In this way, the discrimination unit 215 can output an instruction to the sorting device 10 to separately sort a plurality of groups of biogenic particles.

As described above, in machine learning sorting, the biological particles are separated according to the discrimination made by the discrimination unit 215. In machine learning sorting, the discrimination result with the highest degree of certainty is output, so even if the degree of certainty is low, if it is higher than the other discrimination results, there is a possibility that the target particles will be separated. For this reason, it is not preferable when greater purity of the measurement data (certainty in the correct answer) is required.

The bioparticle analysis system according to the embodiment inputs cell information into a machine learning model, determines whether or not to separate the particles, and then determines whether or not to separate the particles based on a threshold value. The following describes an embodiment of the bioparticle analysis system.

1. First embodiment
1.1. Fractional collection based on confidence
When using machine learning to separate samples, the decision as to whether or not to separate samples is based on past trends (learning data), which necessarily contains ambiguity.

In addition, when using the Softmax function in the output layer of deep learning, the confidence levels for each class are calculated to sum to 100%.

If the class determined with the highest probability is sorted without setting a threshold, even if there is an event where the probability is determined to be class 0 = 20%, class 1 = 40%, class 2 = 30%, and class 3 = 10%, the data will be sorted into class 1, which has the highest probability. However, for users who want to increase purity (certainty), such cases should not be sorted because the probability of class 1 is low at 40%. In this case, the efficiency of sorting will decrease. Here, "class" refers to the category or group of data.

Therefore, in the embodiment, a threshold is set to separate out only events with a high degree of certainty. Note that this threshold can be set variably and can be adjusted according to the user's intentions.

<1.2. How to use confidence level>
The confidence level in deep learning in the first embodiment is used in the following operations. Here, the "confidence level" is the probability that the estimation result in deep learning is correct.

Step 1: First, run some samples to reduce the dimensions.

Step 2: Specify (gate) the range of the population of bioparticles you want to separate from the dimensionality reduction results.

Step 3: Divide some of the dimensionally reduced samples into training data and validation data.

Step 4: After training, vary the threshold using validation data to check for variations in purity and efficiency.

In the above operation, the training and validation data are dimensionally compressed together and then gated, but if the dimensionality compression algorithm maintains reproducibility for newly added data, it is possible to perform dimensionality compression and gating on just the training data, and then add new validation data to the dimensionality compression result to give it a correct label. A "label" indicates which class each piece of data belongs to.

FIG. 9 is a diagram showing the verification data according to the first embodiment. As shown in FIG. 9, the verification data includes events of cell 1, cell 2, cell 3, cell 4, ..., and each event is associated with a "correct answer," "estimate," and "confidence." The "correct answer" data is added in the gating process described below, and the "estimate" and "confidence" data are added in the inference process using the verification data. Here, the "correct answer" indicates the class that the cell should actually be included in. The "estimate" indicates the class estimated in machine learning.

The events of "Cell 1" are associated with a "correct answer" class of "1", a "estimate" class of "2", and a "confidence" of 55%. The events of "Cell 2" are associated with a "correct answer" class of "3", a "estimate" class of "3", and a "confidence" of 80%. The events of "Cell 3" are associated with a "correct answer" class of "5", a "estimate" class of "5", and a "confidence" of 98%. The classes of "Cell 4" are associated with a "correct answer" class of "2", a "estimate" class of "4", and a "confidence" of 40%.

In Figure 9, for example, by setting the threshold to 60%, events of "cell 1" and "cell 4" are not collected, and therefore purity can be increased. However, if the threshold is set too high, such as 90%, there is a high probability that even events that are correctly estimated will not be collected, and the efficiency of event acquisition will decrease.

FIG. 10 is a diagram showing a screen showing the purity and efficiency (yield) of dimensionally compressed measurement data according to the first embodiment. As shown in FIG. 10, the screen displays purity and efficiency based on a threshold value, and also shows which measurement data has been separated. In FIG. 10, the X-axis shows the value of the dimensionally compressed first-dimensional measurement data, and the Y-axis shows the value of the dimensionally compressed second-dimensional measurement data. Although FIG. 10 shows an example of two-dimensional measurement data, the measurement data may be displayed in three dimensions.

Here, "purity" is the percentage of measurement data that is correctly labeled, and "efficiency" is the percentage of correct measurement data contained in the labeled measurement data.

In Figure 10, black stars, crosses, black squares, black triangles, and black circles indicate dimensionally compressed measurement data, and the solid line-enclosed parts of the squares indicate the areas to which labels are assigned. It is assumed that the correct labels are black stars for label 101, black squares for label 102, black triangles for label 103, and black circles for label 104.

For example, when the threshold is 0% (left diagram in Figure 10), when the range of label 101 is separated, the purity is 100% and the efficiency is 100%, when the range of label 102 is separated, the purity is 70% and the efficiency is 70%, when the range of label 103 is separated, the purity is 80% and the efficiency is 100%, and when the range of label 104 is separated, the purity is 100% and the efficiency is 70%.

When the threshold is 70% (center diagram in Figure 10), when the range of label 101 is separated, the purity is 100% and the efficiency is 100%, when the range of label 102 is separated, the purity is 75% and the efficiency is 60%, when the range of label 103 is separated, the purity is 88.9% and the efficiency is 100%, and when the range of label 104 is separated, the purity is 100% and the efficiency is 60%.

When the threshold is 90% (the diagram on the right side of Figure 10), when the range of label 101 is separated, the purity is 98% and the efficiency is 84%, when the range of label 102 is separated, the purity is 85.7% and the efficiency is 60%, when the range of label 103 is separated, the purity is 100% and the efficiency is 87.5%, and when the range of label 104 is separated, the purity is 100% and the efficiency is 60%.

By displaying a screen like that shown in Figure 10, the user can set the threshold while checking the quantitative changes in purity and efficiency according to the threshold, and the qualitative changes in which plots of measurement data are judged to be fractions.

FIG. 11 shows the classes and confidence levels of dimensionally compressed measurement data in the first embodiment.

For example, in FIG. 11, the user clicks on a measurement data item, or a gate or the like selects multiple events. If only one measurement data item is selected, the confidence level of each class of the selected measurement data item is displayed. The user can check the confidence level of each class of the selected measurement data item.

When multiple measurement data are selected, the confidence level for each class is displayed using the average or median of the selected measurement data. The user can check the confidence level for each class of the selected measurement data.

On the left side of FIG. 11, a table 105 is shown showing the classes and confidence levels of multiple selected measurement data. On the right side of FIG. 11, a table 106 is shown showing the classes and confidence levels of one selected measurement data.

1.3. Setting the Threshold
<1.3.1. Threshold setting method 1>
The thresholds may be set in advance for each mode (threshold setting method 1). For example, the thresholds are set as follows: Purity mode = 95%, Normal mode = 75%, Yield mode = 0%, etc. In method 1, the user selects a mode, and the thresholds are set in response to the user's selection.

FIG. 12 is a diagram showing a screen showing the relationship between the mode and purity and efficiency according to the first embodiment. FIG. 12 shows the purity and efficiency of the Yield mode, Normal mode, and Purity mode, as well as which measurement data is being collected. The user may refer to the screen shown in FIG. 12 to determine which mode to select.

In this type of threshold setting method, an algorithm is provided in which a mode is selected and the threshold is set according to the selected mode. This makes it easier for users who find it difficult to set thresholds to use the threshold setting.

<1.3.2. Threshold setting method 2>
The threshold value may be set by the user inputting an arbitrary threshold value on a GUI (Graphical User Interface) (threshold setting method 2). The threshold value may be input by direct input of a value, input using a slide bar, or the like.

<1.3.3. Threshold setting method 3>
In threshold setting method 1, a threshold is determined in advance for each mode based on past data. In threshold setting method 3, an appropriate threshold is automatically calculated for each mode for each measurement data.

FIG. 13 is a diagram for explaining a case where a threshold is set for each measurement data in the first embodiment. In FIG. 13, the thick line indicates the purity of the measurement data for verification, the thick dotted line indicates the three-section average moving line of the purity of the measurement data for verification, the thin line indicates the efficiency of the measurement data for verification, and the thin dotted line indicates the three-section average moving line of the efficiency of the measurement data for verification.

Focusing on the purity in FIG. 13, the Normal mode may be set at the certainty level where the slope becomes gentle, with the threshold set to 62-63%, and the Purity mode may be set at the purity level where the slope changes from gentle to steep, with the threshold set to 87-88%. Whether the slope is gentle or steep may be determined, for example, by determining that the slope is gentle when the difference in slope between the threshold levels before and after the certainty level threshold is less than a predetermined difference, and determining that the slope is steep when the difference in slope between the threshold levels before and after the certainty level threshold is greater than a predetermined difference. Note that the Purity mode may be set at the point where the slope becomes gentle again, near 99% of the certainty level threshold. In other words, the threshold can be set at a point that has some characteristic feature for the slope of the purity, etc.

In addition, the mode and threshold may be set based on the slope of efficiency, the slope of a combination of purity and efficiency, the slope of the moving average line of purity, the slope of the moving average line of efficiency, etc., instead of "purity". Also, the threshold may be calculated using a method that does not use the slope.

As a method for automatically determining the threshold, a receiver operating characteristic (ROC) curve may be used. Figure 14 is a diagram for explaining a case where a threshold is set using an ROC curve of the measurement data for verification according to the first embodiment.

In Figure 14, the true positive rate (TPR) is the percentage of all positives that were correctly determined to be positive when they were actually positive. The false positive rate (FPR) is the percentage of all negatives that were actually negative but were mistakenly determined to be positive.

The threshold that balances purity and efficiency is the threshold that is located closest to the upper left (0, 1) when drawing the ROC curve, so this value can be used as the threshold.

To calculate the threshold closest to (0, 1), you can search using the Euclidean distance, or you can use other methods.

1.4. Visualization using similarity and certainty
Dimensionality reduction compresses multidimensional information into lower dimensional information, so relationships in a multidimensional space cannot be fully represented in a lower dimensional space.

As a result, even similar cell types, such as CD4+ T cells and CD8+ T cells, can be distributed at large distances. This reduces the efficiency of analyzing such similar cell types.

The analysis method disclosed herein selects a cell group using gating or other methods in dimensionality reduction, calculates the similarity and confidence between the selected cell group and the cells being measured for each measurement data, and displays the calculated similarity and confidence in different colors.

As an example of a visualization method, the measurement data may be visualized by changing the shade of the measurement data based on the similarity, or by changing the color.

The similarity may be calculated using a distance-based calculation such as Euclidean distance, Manhattan distance, or Chebyshev distance, or a similarity-based calculation such as cosine similarity, Jaccard coefficient, or Dice coefficient, or it may be calculated using other methods.

This visualization can be done for analytical purposes, or it can be done on measurement data after fractionation.

FIG. 15 is a diagram showing an example of the display of dimensionally compressed measurement data according to the first embodiment. In FIG. 15, the dimensionally compressed measurement data is displayed according to the similarity to the measurement data 111 of the selected cell group. In FIG. 15, the measurement data is displayed in a darker color the higher the similarity to the measurement data 111 of the selected cell group.

In addition, visualization using similarity and confidence can be performed not only on plots on dimensionality reduction, but also on data before fluorescence correction as shown in Figure 16, or on data after fluorescence correction as shown in Figure 17.

FIG. 16 shows an example of a display in which the measurement data before fluorescence correction of the cells to be measured in the first embodiment is displayed in a different color according to the similarity. As shown in FIG. 16, the measurement data is displayed in a different color according to the similarity to the measurement data 111 of the selected cell group.

When displaying the measurement data of the cells to be measured before fluorescence correction, the channel values of each light receiving system may be displayed for each channel. The horizontal axis may represent the fluorescence intensity of each fluorescent dye, and the vertical axis may represent the channel values of each light receiving system. Figure 17 shows an example of a display in which measurement data after fluorescence correction of the cells to be measured according to the first embodiment is displayed in different colors according to similarity. Here, the X and Y axes in Figure 17 represent the fluorescence intensity after fluorescence correction of each fluorescent dye (Color) included in the measurement data.

<1.5. Functional block diagram of information processing device 300>
FIG. 18 is a functional block diagram illustrating the sorting of measurement data in deep learning of the information processing device 300 according to the first embodiment.

As shown in FIG. 18, a measuring device 311 is connected to the information processing device 300. The measuring device 311 measures a sample (e.g., a cell, etc.), adds necessary data (e.g., the color of the cell's fluorescence, the intensity of the fluorescence, etc.) to the measured measurement data, and outputs the data to the information processing device 300. In the measurement, at least an event of the measurement data (e.g., cell 1, etc.) is measured.

The information processing device 300 has an acquisition unit 312, a preprocessing unit 313, a dimensional compression unit 314, a gate unit 315, a division unit 316, a learning unit 317, an estimation unit 318, a threshold setting unit 319, a display unit 320, and a fractionation unit 321.

The acquisition unit 312 acquires multiple pieces of measurement data from a measurement device 311 external to the information processing device 300. The preprocessing unit 313 performs downsampling and narrowing down the target population on the measurement data measured by the acquisition unit 312.

The dimensionality reduction unit 314 performs dimensionality reduction on the measurement data that has been preprocessed by the preprocessing unit 313. "Dimensionality reduction" refers to finding common features in multidimensional data and expressing it in low dimensions while preserving as much as possible the relationships of data distribution in multidimensional space.

The dimensionality compression unit 314 determines the range to be separated after compressing the dimensions of the measurement data. The measurement data compressed by the dimensionality compression unit 314 includes measurement data for verification and measurement data for learning.

The explanatory variables for the measurement data may be raw values before fluorescence correction, such as spectra, or may be data after fluorescence correction. In addition, when performing fluorescence correction, an inverse matrix calculation is performed, and the Gauss-Jordan method may be used to solve the problem. Furthermore, algorithms such as normalization may be used as preprocessing for clustering in order to suppress batch effects.

The gate unit 315 gates the measurement data (including the measurement data for verification and the measurement data for learning) that has been dimensionally compressed by the dimensional compression unit 314. The gate unit 315 also adds a label to the measurement data for learning that has been dimensionally compressed by the dimensional compression unit 314. The division unit 316 divides the multiple pieces of dimensionally compressed measurement data gated by the gate unit 315 into multiple pieces of measurement data for learning and multiple pieces of measurement data for verification.

The learning unit 317 performs machine learning using the learning measurement data (measurement data before or after fluorescence correction) split by the splitting unit 316 and the labels added to the learning measurement data by the gate unit 315 to construct a learning model. The learning model estimates the measurement data and estimates the confidence level for determining whether or not the biological particles are to be separated.

The estimation unit 318 inputs at least a portion of the multiple measurement data (measurement data for verification) into the learning model created by the learning unit 317, and infers whether or not the data is a target for separation.

The estimation unit 318 estimates the accuracy of the multiple measurement data for verification among the multiple measurement data acquired by the acquisition unit 312, and estimates the confidence level of the estimate. Specifically, the estimation unit 318 estimates the accuracy of the multiple measurement data for verification using the learning model generated by the learning unit 317.

The estimation unit 318 has a confidence calculation unit that calculates the confidence of the estimation result based on the multiple measurement data used in the inference by the estimation unit 318 and the information obtained by the data compression process.

The threshold setting unit 319 sets a threshold for dividing the multiple measurement data acquired by the acquisition unit 312 into measurement data for the confidence level estimated by the estimation unit 318.

The display unit 320 displays the measurement data for verification, thresholds, classification (class), thresholds, mode, purity of the measurement data for verification, efficiency, etc. on the screen. The display unit 320 can display the results of the estimation by the estimation unit 318.

The sorting unit 321 sorts out the measurement data to be sorted out from the multiple measurement data acquired by the acquisition unit 312 based on the threshold value set by the threshold setting unit 319. Specifically, the sorting unit 321 classifies the remaining measurement data and the verification measurement data whose estimation and confidence level have been estimated by the estimation unit 318 into classes, and sorts out the measurement data included in the classified classes using the set threshold value.

The remaining measurement data is measurement data for measuring samples other than the learning measurement data samples and the verification measurement data samples. This measurement data sample for measurement is sent to the measurement device 311 after an instruction is given from the information processing device 300 to the measurement device 311. The measurement device 311 then collects an aliquot of the sample and outputs the measurement data of the collected sample to the acquisition unit 312 of the information processing device 300. The instruction from the information processing device 300 to the measurement device 311 is given, for example, after the threshold setting unit 319 has set a threshold.

<1.6. Operation Description>
FIG. 19 is a flowchart for explaining the sorting of measurement data in deep learning of the information processing device 300 according to the first embodiment.

First, a portion of the multiple samples is passed through the measuring device 311 and the portion of the multiple samples is measured (step S1). Next, pre-processing such as downsampling of the measurement data of the portion of the multiple samples and narrowing down of the target group is performed (step S2).

Next, the preprocessed portion of the multiple measurement data is dimensionally compressed (step S3), and the dimensionally compressed portion of the multiple measurement data is gated (step S4). Here, the data to be dimensionally compressed and the explanatory variables during learning may be raw values before fluorescence correction, such as spectra, or may be data after fluorescence correction. In addition, an inverse matrix calculation is performed when performing fluorescence correction, and the Gauss-Jordan method may be used to solve this. Furthermore, an algorithm such as normalization may be used as a preprocessing step for dimensionality compression in order to suppress batch effects.

Next, the multiple measurement data that have been gated by the gate unit 315 and have been dimension-compressed are divided into multiple measurement data for learning and multiple measurement data for validation (step S5).

Next, the divided multiple measurement data for learning are used to perform learning and generate a learning model (step S6). Then, the generated learning model is used to estimate the correct answer for the multiple measurement data for validation and the confidence level for the estimation for the multiple measurement data for validation (step S7).

Then, a threshold value for the estimated confidence level is set (step S8). The threshold value may be set by user instruction or automatically. Next, the user checks the purity and efficiency values and the plot of the measurement data displayed on the display unit 320 (step S9), and if the threshold value setting is not appropriate (NG in step S9), the process returns to step S8 and the threshold value is set again.

On the other hand, if the threshold setting is appropriate (OK in step S9), the remaining samples are flushed (step S10), the remaining measurement data measured on the remaining samples is fractionated (step S11), and a decision is made on fractionation of the measurement data to be fractionated based on the confidence level of the measurement data for the fractionated measurement data class (step S12).

1.7. Modifications
In the first embodiment, the case where the information processing device 300 collects the remaining measurement data has been described. However, since the collection of the remaining measurement data takes time, the collection of the remaining measurement data may be performed on the measuring device 311 side.

FIG. 20 is a block diagram showing a configuration example of a bioparticle analysis system 1 according to a modified example of the first embodiment. The same parts as those in FIG. 4 are described with the same reference numerals.

The biological particle analysis system 1 according to the modified example is an example in which the functions of the information processing device 20 shown in FIG. 4 are divided and provided in a fractionation device 10 connected via a network.

Specifically, as shown in FIG. 20, in the modified biological particle analysis system, the fractionation device 10 has an analysis unit 203, a reference spectrum storage unit 205, a data compression processing unit 207, and a learning unit 211. The fractionation device 10 acquires measurement data from the sample S, and separates particles to be separated based on the discrimination of the information processing device 20. The information processing device 20 has an acquisition unit 201, a learning model storage unit 213, and a discrimination unit 215.

The information processing device 20 and the fractionation device 10 may be connected to each other so as to be able to communicate with each other via a network such as the Internet, a public line network such as a telephone network or a satellite communication network, various LANs (Local Area Networks) including Ethernet (registered trademark), or a WAN (Wide Area Network).

In the bioparticle analysis system according to the modified example, functions with a large computational load (e.g., the analysis unit 203, the data compression processing unit 207, and the learning unit 211) can be handled by the fraction collection device 10. On the other hand, since delays due to networks, etc., must be avoided for rapid discrimination, and the computational load is not large, the functions of the discrimination unit 215 and the learning model storage unit 213 may be handled by the information processing device 20 that is directly connected to the fraction collection device 10.

FIG. 21 is a functional block diagram of an information processing system according to a modified example of the first embodiment. Note that the same parts as those in FIG. 18 are denoted by the same reference numerals. As shown in FIG. 21, the fractionation unit 321 provided in the information processing device 300 may be provided in the measurement device 311.

The preprocessing unit 313 and the threshold setting unit 319 may be provided in the measurement device 311.

As shown in FIG. 21, the threshold set by the threshold setting unit 319 and the remaining measurement data not used for verification and learning by the estimation unit 318 are output from the information processing device 300 to the measurement device 311.

The fractionation unit 321 of the measurement device 311 receives the threshold and the estimated remaining measurement data output from the information processing device 300, and fractionates the remaining measurement data using the received threshold.

The sorting section (determination section) of the measuring device 311, which is a biological particle sorting device, inputs optical information measured from the biological particles to be sorted into a learning model created by the learning section 317, infers whether the biological particles to be sorted are to be sorted, and if it is inferred that they are to be sorted, makes a sorting determination based on the threshold set by the threshold setting section 119. The biological particle sorting device sorts the particles to be sorted based on the sorting determination by the determination section. The biological particles to be sorted are included in the sample. Next, a modified example of the information processing system according to the first embodiment will be described. FIG. 22 is a functional block diagram showing a modified example of the information processing system according to the first embodiment.

In a modified example of the information processing system according to the first embodiment, as shown in FIG. 22, functions requiring a large amount of calculation (e.g., preprocessing unit 313, dimensional compression unit 314, gate unit 315, division unit 316, learning unit 317, estimation unit 318, threshold setting unit 319) can be assigned to a device with higher computing power (information processing server 301 in the example of FIG. 22).

The information processing device 300 may be a cloud computer connected to the measurement device 311 via a network. In this case, the cloud computer may execute some of the functions of the information processing device 300, such as the dimensional compression unit 314, the machine learning learning unit 317, and the threshold setting unit 319.

On the other hand, in order to make rapid judgments and to avoid delays due to networks, etc., and functions that do not impose a large computational load, the information processing device 300 directly connected to the measurement device 311 may be responsible.

According to the information processing system according to the modified example of the first embodiment, measurement data can be appropriately classified, similar to the information processing device 300 according to the first embodiment.

2. Second embodiment
2.1. Confidence-Based Clustering
The second embodiment is to set a threshold value for clustering sorting. When sorting is performed using a clustering algorithm, the cluster is always classified into the cluster with the highest relative similarity among all clusters. However, it is unclear whether the classification results are close in absolute value.

If the absolute distance between the classified clusters is far, a user who prioritizes purity may want to exclude the classified cluster from separation. In the second embodiment, a threshold is set so that measurement data is only separated when the distance is closer than a certain value.

2.2. Threshold for clustering fractionation
FIG. 23 is a diagram for explaining the concept of thresholds for clustering sorting according to the second embodiment.

In Figure 23, the parameters on the horizontal axis indicate, for example, the type of fluorescent dye antibody, antigen marker, or CD classification, and the vertical axis indicates the fluorescence intensity of an event (e.g., a cell). The solid line indicates the representative value of the cluster, and the dotted line indicates the target event (the measured value of the remaining measurement data).

As shown in FIG. 23, for example, if a threshold of 50% is set for the representative value of the cluster corresponding to the leftmost parameter, then 25% to 75% of the cluster's representative value is set as the threshold range, as shown in FIG. 24. FIG. 24 is a diagram for explaining the concept of the range when the threshold is set to 50% in clustering sorting according to the second embodiment. If the measured value (fluorescence intensity) of the parameter of the leftmost target event shown in FIG. 23 falls within 25% to 75% of the representative value of this cluster, then it is set as the measured value to be sorted. In the case of FIG. 23, the measured value of the parameter of the leftmost event does not fall within the threshold range, so it is not set as the target for sorting.

For example, if a threshold of 50% is set for the representative value of the cluster corresponding to the second parameter from the left, then 25% to 75% of the cluster's representative value will be set as the threshold range, as shown in Figure 24. If the measured value of the parameter of the second target event from the left falls within 25% to 75% of this cluster's representative value, then it will be set as the measured value to be sampled. In the case of Figure 23, the measured value second from the left does not fall within the threshold range, so it is not set as the target for sampled.

For example, if a threshold of 50% is set for the representative value of the cluster corresponding to the third parameter from the left, then 25% to 75% of the cluster's representative value will be set as the threshold range, as shown in Figure 24. If the third measurement value from the left falls within the range of 25% to 75% of this cluster's representative value, it will be set as the measurement value to be sampled. In the case of Figure 24, the third measurement value from the left falls within the threshold range, so it will be set as the measurement value to be sampled.

In the second embodiment, the threshold for clustering may be determined as follows:

<2.2.1. When threshold determination is performed for each parameter>
- Enter an absolute threshold value, and if the measured values for all parameters are within the cluster representative value ± the threshold value, the sample is collected.
- Enter a percentage threshold, and if the measured values for all parameters are within the cluster representative value ± representative value × threshold, the sample will be sorted.
- For each cluster, if the frequency distribution of each parameter is within the threshold value entered by the user, the parameters are separated.

<2.2.2. When threshold is determined by averaging all parameters>
Input an absolute threshold value, and if the mean (|measured value-representative value|) falls within the threshold value, the sample is collected.
Enter a percentage threshold, and if the mean (|measured value - representative value|) is within the average of the representative values x the threshold value, the sample is collected.
Here, "mean" means average. When using a random forest as an algorithm, the number or ratio of the number of decision trees may be set as a threshold when performing majority voting of the decision trees. The threshold may be determined automatically based on the measurement data, or may be determined by the user.

The threshold value may be determined not only by the average value but also by the median value of multiple measurement data included in a cluster. The threshold value may also be determined by using a representative value determined by the learning unit 317.

<2.3. Functional block diagram of information processing device 400>
FIG. 25 is a functional block diagram of the information processing device 400 according to the second embodiment for performing clustering sorting.

As shown in FIG. 25, a measuring device 411 is connected to the information processing device 400. The measuring device 411 measures a sample (e.g., a cell, etc.), adds necessary data (e.g., the color of the cell's fluorescence, the intensity of the fluorescence, etc.) to the measured measurement data, and outputs the data to the information processing device 400. In the measurement, at least an event of the measurement data (e.g., cell 1, etc.) is measured.

The information processing device 400 has an acquisition unit 412, a preprocessing unit 413, a classification and clustering unit 414, a cluster selection unit 415, a display unit 416, a threshold setting unit 417, and a fractionation unit 418.

The acquisition unit 412 acquires multiple pieces of measurement data from a measurement device 411 external to the information processing device 400. The preprocessing unit 413 performs downsampling and narrowing down the target population on the measurement data measured by the acquisition unit 412.

The classifying and clustering unit 414 classifies the multiple pieces of measurement data acquired by the acquiring unit 412 into classes. The classifying and clustering unit 414 also classifies the multiple pieces of measurement data acquired by the acquiring unit 412 into clusters.

The cluster selection unit 415 selects a cluster to be collected from the classes classified by the classification and clustering unit 414. The display unit 416 displays a screen showing the efficiency of the classified measurement data (e.g., measurement data, class, threshold, mode, purity, efficiency-classified measurement data, cluster of classified measurement data). The threshold setting unit 417 sets a threshold for a representative value of the cluster, which is the average of multiple measurement data included in the cluster selected by the cluster selection unit 415.

The threshold value may be the median value of the multiple measurement data included in the cluster. The threshold value may also be determined using a representative value determined by the learning unit 317.

The fractionation unit 418 fractionates the measurement data to be fractionated from among the measurement data contained in the clusters classified by the classification and clustering unit 414 based on the threshold set by the threshold setting unit 417.

Specifically, if all the measurement values of the multiple measurement data included in the cluster classified by the classifying and clustering unit 414 are within a representative value ±threshold, the sorting unit 418 sorts the sampling data included in the cluster classified by the classifying and clustering unit 414 as the target for sorting.

The fractionation unit 418 may fractionate the sampling data included in the cluster classified by the clustering and clustering unit 414 as a fractionation target, if all the measurement values of the multiple measurement data included in the cluster classified by the classifying and clustering unit 414 are within the range of a representative value ± representative value × threshold value.

<2.4. FlowSOM circuit>
26 is a diagram showing a first example of a circuit of FlowSOM according to the second embodiment. FlowSOM is a known clustering algorithm. As shown in FIG. 26, event data a (d dimension) and data b of node (cluster) 1 containing a representative value of the d dimension are input to a difference calculator 551, and a difference (a-b) is calculated.

The squarer 552 calculates the square (a−b) ² of the difference (a−b) calculated by the differentiator 551, and outputs it to a summation calculator 553. The summation calculator 553 calculates the sum Σ(a−b ⁾² of the square (a−b) ² of the difference (a−b) calculated by the squarer 552, and outputs it to a comparator 554.

The comparator 554 compares the minimum distance held in the minimum distance holder 555 with the sum Σ(a−b) ² output from the summation unit 553 , and holds the smaller distance in the minimum distance holder 555 as the minimum distance.

Specifically, the comparator 554 replaces the event data a and data b held in the minimum distance holder 555 with the sum Σ(a−b) ² , which has the closest Euclidean distance between them. That is, the comparator 554 performs comparison to search for the node with the smallest error. As a result, the data is classified into the node (cluster) with the smallest distance held in the minimum distance holder 555.

The data b from node 1, node 2, ..., node N are input serially to the difference calculator 551, but the data b from node 1, node 2, ..., node N may be processed in parallel.

FIG. 27 is a diagram showing a second example of a circuit of FlowSOM according to the second embodiment. As shown in FIG. 27, data of 100 nodes 1 to 100 is input with a parallel number of 10.

Specifically, data b containing d-dimensional representative values of node 1, node 2, node 3, ..., node 10 is input in parallel to subtractors 551_1 to 551_10. Event data a (d-dimension) is also input to subtractors 551_1 to 551_10.

The data b of node 1, node 11, node 21, ..., node 91 is input in order. The data b of node 2, node 12, node 22, ..., node 92 is input in order, the data b of node 3, node 13, node 23, ..., node 93 is input in order, ..., the data b of node 10, node 20, node 30, ..., node 100 is input in order.

Differentiators 551_1 to 551_10 receive event data a (d dimension) and data b from nodes (clusters) 1, 2, ..., and N that contain the representative value of the d dimension, and calculate the difference (a - b).

The squarers 552_1 to 552_10 calculate the squares (a-b) ² of the differences (a-b) calculated by the differencers 551_1 to 551_10, respectively, and output the squares to the summations 553_1 to 553_10. The summations 553_1 to 553_10 calculate the sums Σ(a-b) ² of the squares (a-b) ² of the differences (a-b) calculated by the squarers 552_1 to 552_10, respectively, and output the sums to the comparators 554_1 to 554_10, respectively.

Comparators 554_1 to 554_10 compare the minimum distances held in minimum distance holders 555_1 to 555_10 with the sums Σ(a−b) ² output from summaries 553_1 to 553_10, respectively, and hold the smaller distance as the minimum distance in minimum distance holders 555_1 to 555_10, respectively.

As a result, the nodes (clusters) that are the shortest distance among node 1, node 11, node 21, ..., node 91 are classified into the nodes (clusters) that are the shortest distance among node 2, node 12, node 22, ..., node 92, ..., node 10, node 20, node 30, ..., node 100 are classified into the nodes (clusters) that are the shortest distance among node 1, node 11, node 21, ..., node 91, ..., node 100.

Comparator 556 compares the minimum distances stored in minimum distance holder 555_1 to minimum distance holder 555_10, and stores the smaller distance as the minimum distance in minimum distance holder 257. As a result, minimum distance holder 257 classifies nodes (clusters) with the smallest distance from node 1 to node 100.

In FIG. 27, the number of nodes is 100 and the number of parallel connections is 10, but these values may be flexible depending on the circuit resources. Also, in FIG. 27, the case where one comparator 556 is used is shown, but multiple comparators 556 may be used to perform parallel processing.

FIG. 28 is a diagram showing a third example of a circuit of FlowSOM according to the second embodiment. In FIG. 28, ◇ indicates a metacluster, and □ indicates a node associated with the metacluster for which the minimum value is selected.

In FIG. 28, the number of metaclusters is 8, and the number of nodes linked to the metacluster for which the minimum value was selected is 10, but the number of metaclusters and nodes is not limited to this. In FIG. 28, the case where nodes 1 to 10 linked to the metacluster are calculated in series is shown, but the calculations may be performed in parallel.

In the third example of the FlowSOM circuit shown in Figure 28, the metacluster with the smallest distance is found among metaclusters 1-8 by the processing of the difference calculator 571 to the minimum distance holder 575. After that, the 10 nodes linked to the metacluster with the smallest distance are classified into the node with the final distance.

In FIG. 28, event data a (d dimension) and data b of a node (cluster) linked to a selected meta cluster with the smallest error containing the representative value of the d dimension are input to a subtractor 571, and the difference (a-b) is calculated.

The squarer 572 calculates the square (a−b) ² of the difference (a−b) calculated by the differentiator 571, and outputs it to a summation calculator 573. The summation calculator 573 calculates the sum Σ(a−b ⁾² of the square (a−b) ² of the difference (a−b) calculated by the squarer 572, and outputs it to a comparator 574.

The comparator 574 compares the minimum distance held in the minimum distance holder 555 with the sum Σ(a−b) ² output from the summation unit 573 , and holds the smaller distance in the minimum distance holder 575 as the minimum distance.

Specifically, the comparator 574 replaces the event data a and data b held in the minimum distance holder 575 with the sum Σ(a−b) ² , which has the closest Euclidean distance between them. That is, the comparator 574 performs comparison to search for the node with the smallest error. As a result, the node (cluster) with the smallest distance held in the minimum distance holder 575 is clustered.

The data b from node 1, node 2, ..., node 10 linked to the meta cluster with the smallest error distance is input in series to the difference calculator 571, but the data b from node 1, node 2, ..., node 10 may be processed in parallel.

<2.5. Operation Description>
FIG. 29 is a flowchart for explaining the clustering sorting of the information processing device 400 according to the second embodiment.

First, a portion of the multiple samples is passed through the measuring device 411 and the portion of the multiple samples is measured (step S21). Next, pre-processing such as downsampling of the measurement data of the portion of the multiple samples and narrowing down of the target group is performed (step S22).

Next, classification is performed to classify the preprocessed part of the measurement data into classes (step S23). A cluster to be collected is selected from the classified clusters (step S24).

Next, a threshold is set for a representative value that is the average of multiple measurement data included in the selected cluster (step S25). Note that the threshold may be set for the median value of multiple measurement data included in the selected cluster. Next, the user checks the efficiency value displayed on the display unit 416 (step S26), and if the efficiency is not 100% (NG in step S26), the process returns to step S25, where the threshold is set again. Note that the efficiency value may not be 100% but may be any value determined by the user.

On the other hand, if the efficiency is 100% (OK in step S26), the remaining sample is passed through (step S27), and clustering is performed on the remaining measurement data (step S28). Next, from the remaining measurement data contained in the clusters classified based on the set threshold, the measurement data to be collected is classified using the set threshold (step S29).

Here, the explanatory variables for the data to be clustered may be raw values before fluorescence correction, such as spectra, or may be data after fluorescence correction. In addition, an inverse matrix calculation is performed when performing fluorescence correction, and the Gauss-Jordan method may be used to solve this. Furthermore, algorithms such as normalization may be used as preprocessing for clustering in order to suppress batch effects.

2.6. Modifications
In the second embodiment, the case where the information processing device 400 classifies the remaining measurement data has been described. However, since classification of the remaining measurement data takes time, it may be performed on the measuring device 411 side.

FIG. 30 is a functional block diagram of an information processing system according to a modified example of the second embodiment. Note that the same parts as those in FIG. 25 are denoted by the same reference numerals. As shown in FIG. 30, the fractionation unit 418 provided in the information processing device 400 may be provided in the measurement device 411.

As shown in FIG. 30, the threshold set by the threshold setting unit 417 and the clusters clustered by the classification and clustering unit 414 are output from the information processing device 400 to the measurement device 411.

The fractionation unit 418 of the measurement device 411 receives the threshold and the clustered clusters output from the information processing device 400, and fractionates the measurement data contained in the clusters using the received threshold.

The information processing system according to the modified example of the second embodiment can appropriately classify measurement data, similar to the information processing device 400 according to the second embodiment.

<2.7. Flowchart for FlowSOM fractionation>
Next, a description will be given of the operation of the first example of the circuit of FlowSOM shown in Fig. 26. Fig. 31 is a flowchart for explaining the operation of the first example of the circuit of FlowSOM according to the second embodiment.

As shown in FIG. 31, i = 0 is set (step S31), and it is determined whether i < d (d: number of dimensions) (step S32). If i < d in step S32 (Yes in step S32), the difference between the i-th dimension value of the representative vector of each node and the i-th dimension value of the event to be sorted is calculated (step S33).

Next, the difference between the i-dimension value of the representative vector of each node calculated in step S33 and the i-dimension value of the event to be sorted is squared (step S34), and the squared difference value is integrated (step S35). Next, i = i + 1 is set (step S36), and the process returns to step S32.

If i<d is not satisfied in step S32 (No in step S32), the node with the smallest integrated value of the squared differences is calculated (step S37), and the process ends. As a result, the node (cluster) with the smallest error distance is clustered.

Next, the operation of the third example of the FlowSOM circuit shown in FIG. 28 will be described. FIG. 32 is a flowchart for explaining the operation of the third example of the FlowSOM circuit according to the second embodiment.

As shown in FIG. 32, i = 0 is set (step S41), and it is determined whether i < d (d: number of dimensions) (step S42). If i < d in step S42 (Yes in step S42), the difference between the i-th dimension value of the representative vector of each metacluster and the i-th dimension value of the event to be sorted is calculated (step S43).

Next, the difference between the i-dimension value of the representative vector of each metacluster calculated in step S43 and the i-dimension value of the event to be sorted is squared (step S44), and the squared difference value is integrated (step S45). Next, i = i + 1 is set (step S46), and the process returns to step S42.

In step S42, if i<d is not satisfied (No in step S42), the metacluster with the smallest squared difference is calculated (step S47), and j is set to 0 (step S48).

Next, it is determined whether j<d (d: number of dimensions) (step S49). If j<d is true in step S49 (Yes in step S49), the difference between the j-dimensional value of the representative vector of each node belonging to the metacluster with the smallest squared difference value and the j-dimensional value of the event to be sorted is calculated (step S50).

Next, the difference between the j-dimension value of the representative vector of each node belonging to the meta-cluster calculated in step S50 and the j-dimension value of the event to be sorted is squared (step S51), and the squared difference value is integrated (step S52). Next, j = j + 1 is set (step S53), and the process returns to step S49.

If j<d is not satisfied in step S49 (No in step S49), the node with the smallest integrated value of the squared differences is calculated (step S54), and the process ends. As a result, the node (cluster) with the smallest error distance is clustered.

In the third example, a metacluster with the shortest Euclidean distance is first selected, and then the distance to each node belonging to that metacluster is calculated, which is expected to reduce computing resources and increase processing speed.
3. Third embodiment
3.1. Fractionation based on confidence
In the second embodiment, the explanation is given from the viewpoint of a general FCM (Flow Cytometer) that does not use images and focuses mainly on fluorescence intensity. In the third embodiment, the explanation is given for a case where certainty-based sorting is applied to an IFCM (Imaging Flow Cytometer).

In IFCM, in addition to being able to measure fluorescence intensity like regular FCM, it is also possible to take images of individual cells. In the third embodiment, the fluorescence intensity or image is used as input to identify the group to be separated using dimensionality reduction, clustering, etc. (objective variable), and then the fluorescence intensity or image is used as the explanatory variable for learning. After that, an appropriate threshold is set and separation is performed.

Here, the fluorescence intensity data may be either data before or after fluorescence correction, and the image may be either the raw image data or data that has undergone preprocessing such as convolution. In addition, the method described in <1.3. Setting the threshold> may be used to set the threshold.

<3.2. Functional block diagram of information processing device 600>
FIG. 33 is a functional block diagram showing IFCM fractionation of an information processing device 600 according to the third embodiment.

As shown in FIG. 33, a measuring device 611 is connected to the information processing device 600. The measuring device 611 measures the sample, adds necessary data to the measured measurement data, and outputs the data to the information processing device 600. In the measurement, at least an event of the measurement data (e.g., cell 1, etc.) is measured.

The information processing device 600 has an acquisition unit 612, a preprocessing unit 613, a determination unit 614, a dimensionality reduction/clustering unit 615, a population identification unit 616, a division unit 617, a learning unit 618, an estimation unit 619, a display unit 620, a threshold setting unit 621, and a fractionation unit 622.

The acquisition unit 612 acquires multiple pieces of measurement data from a measurement device 611 external to the information processing device 600. The pre-processing unit 613 performs downsampling and narrowing down the target population on the measurement data measured by the acquisition unit 612.

The determination unit 614 determines whether to input the fluorescence data or image data contained in the measurement data acquired by the acquisition unit 612 from among the multiple measurement data. The dimensionality reduction/clustering unit 615 performs dimensionality reduction or classifies the fluorescence data or image data determined by the determination unit into clusters.

The population identification unit 616 identifies a population to be separated from the dimensionally compressed fluorescent data or image data classified by the dimensionality compression/clustering unit 615, or from the classified clusters. The division unit 617 divides the fluorescent data or image data identified by the population identification unit 616 into fluorescent data or image data for learning and the fluorescent data or image data for verification.

The learning unit 618 performs learning using the multiple pieces of measurement data for learning split by the splitting unit 617, and generates a learning model. The estimation unit 619 estimates an estimate for multiple pieces of measurement data for verification among the measurement data included in the population identified by the population identification unit 616, and estimates the confidence level of the estimate. Specifically, the estimation unit 619 estimates the confidence level of the fluorescence data or image data for verification using the learning model generated by the learning unit 618.

The display unit 620 displays the purity and efficiency of the measurement data for verification, as well as the measurement data for verification, thresholds, classification (class), thresholds, mode, etc., as necessary.

The threshold setting unit 621 sets a threshold for classifying the multiple measurement data acquired by the acquisition unit 612 based on the confidence level estimated by the estimation unit 619. The sorting unit 622 sorts the dimensionally compressed fluorescent data or image data classified by the dimensionality compression/clustering unit 615, or the measurement data included in the classified cluster, as the measurement data to be sorted, based on the threshold set by the threshold setting unit 621.

The sorting unit 622 sorts out the remaining fluorescence data or image data other than the fluorescence data or image data for verification and the fluorescence data or image data for learning from the multiple measurement data acquired by the acquisition unit 612.

Specifically, if all the measurement values of the multiple measurement data included in the cluster classified by are within the representative value ±threshold, the sorting unit 622 sorts the sampling data included in the cluster classified by as the target for sorting.

The fractionation unit 622 may fractionate the sampling data included in the cluster classified by the clustering unit as a fractionation target if all the measurement values of the multiple measurement data included in the cluster classified by fall within the range of the representative value ± the representative value × the threshold value.

<3.3. Operation Description>
FIG. 34 is a flowchart for explaining IFCM sorting in information processing device 600 according to the third embodiment.

First, a portion of the sample is passed through the measuring device 611 and a portion of the multiple samples is measured (step S131). Next, pre-processing such as downsampling of the measurement data of the portion of the multiple samples that have been measured and narrowing down of the target group is performed (step S132).

Next, it is determined whether to input the fluorescence or the image of the portion of the multiple measurement data that has been preprocessed (step S133). Then, dimensionality reduction and clustering are performed on the fluorescence or image determined in step S33 (step S134). Next, a population to be sorted is identified from the clusters that have been clustered in step S34 (step S135). Here, a "population" is an island that has been dimensionally reduced for the fluorescence or image, and the dimensionality reduced fluorescence or image that constitutes this island is gated.

Here, the input data for dimensionality reduction and clustering, and the explanatory variables during learning, may be raw values before fluorescence correction, such as spectra, or may be data after fluorescence correction. When using images, the raw data may be used, or it may be used after preprocessing such as convolution. In addition, an inverse matrix calculation is performed when performing fluorescence correction, and the Gauss-Jordan method may be used to solve this. Furthermore, an algorithm such as normalization may be used as preprocessing to suppress batch effects.

Next, the multiple measurement data included in the population identified in step S135 are divided into multiple measurement data for learning and multiple measurement data for validation (step S136).

Next, the divided multiple pieces of measurement data for learning are used to perform learning using the fluorescence or images as explanatory variables to generate a learning model (step S137). Then, the generated learning model is used to estimate the correct answer for the multiple pieces of measurement data for validation and the confidence level for the estimate for the multiple pieces of measurement data for validation (step S138).

Then, a threshold value for the estimated confidence level is set (step S139). Next, the user checks the purity and efficiency values and the plot of the measurement data displayed on the display unit 320 (step S140), and if the threshold setting is not appropriate (NG in step S140), the process returns to step S139, and the threshold setting is performed again.

On the other hand, if the threshold setting is appropriate (OK in step S140), the remaining measurement data is sent (step S141), and the remaining samples measured are sorted into clusters (step S142). Next, from the remaining measurement data contained in the classified clusters based on the set threshold, the measurement data to be sorted is sorted based on the confidence level (step S143).

3.4. Modifications
In the third embodiment, the case where the information processing device 600 classifies the remaining measurement data has been described. However, since classification of the remaining measurement data takes time, it may be performed on the measuring device 611 side.

FIG. 35 is a functional block diagram of an information processing system according to a modified example of the third embodiment. Note that the same parts as those in FIG. 30 are denoted by the same reference numerals. As shown in FIG. 35, the fractionation unit 622 provided in the information processing device 600 may be provided in the measurement device 611.

As shown in FIG. 35, the threshold set by the threshold setting unit 621 and the clusters clustered by the dimensionality reduction/clustering unit 615 are output from the information processing device 600 to the measurement device 611.

The fractionation unit 622 of the measurement device 611 receives the threshold and the clustered clusters output from the information processing device 600, and fractionates the measurement data contained in the clusters using the received threshold.

The information processing system according to the modified example of the third embodiment allows for proper IFCM fractionation.

4. Hardware Configuration
FIG. 36 is a hardware configuration diagram showing an example of a computer that realizes the arithmetic unit of the information processing devices 20, 300, 400, 600 and the measurement devices 311, 411, 611 according to the embodiment.

Computer 1000 has a CPU 1100, RAM 1200, ROM (READ ONLY MEMORY) 1300, HDD (HARD DISK DRIVE) 1400, communication interface 1500, and input/output interface 1600. Each part of computer 1000 is connected by bus 1050.

The CPU 1100 operates based on the programs stored in the ROM 1300 or the HDD 1400 and controls each component. For example, the CPU 1100 loads the programs stored in the ROM 1300 or the HDD 1400 into the RAM 1200 and executes processes corresponding to the various programs.

The ROM 1300 stores boot programs such as the BIOS (BASIC INPUT OUTPUT SYSTEM) that is executed by the CPU 1100 when the computer 1000 starts up, as well as programs that depend on the hardware of the computer 1000.

HDD 1400 is a computer-readable recording medium that non-temporarily records programs executed by CPU 1100 and data used by such programs. Specifically, HDD 1400 is a recording medium that records application programs related to the present disclosure, which are an example of program data 1450.

The communication interface 1500 is an interface for connecting the computer 1000 to an external network 1550 (e.g., the Internet). For example, the CPU 1100 receives data from other devices and transmits data generated by the CPU 1100 to other devices via the communication interface 1500.

The input/output interface 1600 is an interface for connecting the input/output device 1650 and the computer 1000. For example, the CPU 1100 receives data from an input device such as a keyboard or a mouse via the input/output interface 1600. The CPU 1100 also transmits data to an output device such as a display, a speaker or a printer via the input/output interface 1600. The input/output interface 1600 may also function as a media interface that reads programs and the like recorded on a specific recording medium. Media include, for example, optical recording media such as DVD (DIGITAL VERSATILE DISC) and PD (PHASE CHANGE REWRITABLE DISK), magneto-optical recording media such as MO (MAGNETO-OPTICAL DISK), tape media, magnetic recording media, or semiconductor memory.

Note that the CPU 1100 reads and executes the program data 1450 from the HDD 1400, but as another example, the CPU 1100 may obtain these programs from other devices via the external network 1550.

The above describes in detail preferred embodiments of the present disclosure with reference to the attached drawings, but the technical scope of the present disclosure is not limited to such examples. It is clear that a person with ordinary knowledge in the technical field of the present disclosure can conceive of various modified or revised examples within the scope of the technical ideas described in the claims, and it is understood that these also naturally fall within the technical scope of the present disclosure.

Furthermore, the effects described in this specification are merely descriptive or exemplary and are not limiting. In other words, the technology disclosed herein may achieve other effects that are apparent to a person skilled in the art from the description in this specification, in addition to or in place of the above effects.

The present technology can also be configured as follows.
[1]
an acquisition unit that acquires measurement data measured from biogenic particles contained in a sample;
a compression unit that performs a data compression process on the measurement data acquired by the acquisition unit;
a gate unit that gates the measurement data compressed by the compression unit into training measurement data and verification measurement data and adds a label to the training measurement data;
A learning unit that constructs a learning model using the learning measurement data and the labels;
an estimation unit that inputs the verification measurement data to the learning model and outputs a confidence level of the verification measurement data;
and a threshold setting unit that sets a threshold for separating the sample based on the degree of certainty.
[2]
a display unit that displays the efficiency and yield of the biological particles based on the output confidence level and the threshold value;
The bioparticle analysis system according to [1].
[3]
The bioparticle analysis system according to claim 1 or 2, wherein the learning measurement data and the verification measurement data included in the measurement data are different from each other.
[4]
a determination unit that inputs measurement data measured from the biological particles for sorting into the learning model, infers whether the biological particles for sorting are targets for sorting, and, when it is inferred that the biological particles for sorting are targets for sorting, determines whether the biological particles for sorting are targets for sorting based on a threshold value set by the threshold setting unit;
The bioparticle analysis system according to any one of [1] to [3].
[5]
the biogenic particle sorting device includes a sorting unit that sorts particles to be sorted based on the sorting determination of the determination unit,
The bioparticle analysis system according to [4].
[6]
The biological particle analysis system according to [4], wherein the biological particles for separation are contained in the sample.
[7]
The biological particle analysis system according to any one of [1] to [6], wherein the threshold value is set to a predetermined threshold value.
[8]
The bioparticle analysis system according to [7], wherein the predetermined threshold is determined according to one or more modes.
[9]
The bioparticle analysis system according to any one of [1] to [6], wherein the threshold value is set by a user.
[10]
The data compression process is a dimensional compression process,
The bioparticle analysis system according to any one of [1] to [9], wherein a range to be sorted is determined after the dimensionality reduction.
[11]
a compression unit that performs a data compression process on measurement data obtained by measuring the biogenic particles contained in a sample;
a gate unit that gates the measurement data compressed by the compression unit into training measurement data and verification measurement data and adds a label to the training measurement data;
a learning unit that uses the learning measurement data and the label to construct a learning model for determining whether the biogenic particles are to be sorted;
an inference unit that inputs the verification measurement data into a learning model constructed by the learning unit and infers whether the data is a target for collection;
a certainty factor calculation unit that calculates a certainty factor of the verification measurement data used in the inference;
a threshold setting unit that sets a threshold for dividing the sample based on the certainty calculated by the certainty calculation unit;
An information processing device having the above configuration.
[12]
The constructed learning model is output to a microparticle sorting device.
The information processing device according to [11].
[13]
a compression step of performing a data compression process on the measurement data measured from the biogenic particles contained in the sample;
a gating step of gating the data-compressed measurement data into training measurement data and verification measurement data, and adding a label to the training measurement data;
a learning step of constructing a learning model for determining whether the biological particle is a separation target by using the learning measurement data and the label; and an inference step of inputting the verification measurement data into the learning model constructed by the learning step and inferring whether the biological particle is a separation target.
a confidence level calculation step of calculating a confidence level of the verification measurement data used in the inference;
a threshold setting step of setting a threshold for dividing the sample based on the certainty calculated by the certainty calculation step;
An information processing method comprising the steps of:
[14]
The information processing method according to the above [13], further comprising a measuring step in which the measurement data is measured using a microparticle analysis device.
[15]
The information processing method according to item [14], further comprising the steps of: inputting optical information measured from the biological particles for separation in the microparticle analysis device into a learning model constructed by the learning step; inferring whether the biological particles for separation are targets for separation; and, when it is inferred that the biological particles for separation are targets for separation, making a separation determination based on the threshold value set by the threshold setting step.
[16]
The information processing method according to [15] above, further comprising the step of separating particles to be separated based on the separation determination.
[17]
an acquisition unit that acquires a plurality of pieces of measurement data including optical information measured from biogenic particles contained in a sample;
a clustering unit that classifies the plurality of pieces of measurement data acquired by the acquisition unit into a plurality of clusters;
a cluster selection unit that selects a cluster to be collected from the clusters classified by the clustering unit;
a threshold setting unit that sets a threshold based on the plurality of pieces of measurement data included in the cluster selected by the cluster selection unit.
[18]
The clustering unit includes:
Classifying the plurality of pieces of measurement data acquired by the acquisition unit into clusters;
The information processing device according to [17], further comprising a fractionation unit that fractionates the measurement data to be fractionated from the measurement data included in the clusters classified by the clustering unit based on the threshold value set by the threshold setting unit.
[19]
The threshold setting unit is
The information processing device according to any one of claims 17 to 18, further comprising: setting a threshold for a representative value of the cluster or a threshold for a median value of the plurality of pieces of measurement data included in the cluster selected by the cluster selection unit.

300, 400, 600 Information processing device 311, 411, 611 Measuring device 312, 412, 612 Acquisition unit 313, 413, 613 Preprocessing unit 314 Dimensional compression unit 315 Gating unit 316 Division unit 317, 618 Learning unit 318, 619 Estimation unit 319, 417, 621 Threshold setting unit 320, 416, 620 Display unit 321, 418, 622 Sorting unit 414 Classifying and clustering unit 415 Cluster selection unit 614 Determination unit 615 Dimensional compression/clustering unit 616 Population identification unit

Claims (19)

an acquisition unit that acquires measurement data measured from biogenic particles contained in a sample;
a compression unit that performs a data compression process on the measurement data acquired by the acquisition unit;
a gate unit that gates the measurement data compressed by the compression unit into training measurement data and verification measurement data and adds a label to the training measurement data;
A learning unit that constructs a learning model using the learning measurement data and the labels;
an estimation unit that inputs the verification measurement data to the learning model and outputs a confidence level of the verification measurement data;
and a threshold setting unit that sets a threshold for separating the sample based on the degree of certainty. a display unit that displays the efficiency and yield of the biological particles based on the output confidence level and the threshold value;
The bioparticle analysis system according to claim 1 . The bioparticle analysis system according to claim 1 , wherein the learning measurement data and the verification measurement data included in the measurement data are different from each other. a determination unit that inputs measurement data measured from the biological particles for sorting into the learning model, infers whether the biological particles for sorting are targets for sorting, and, when it is inferred that the biological particles for sorting are targets for sorting, determines whether the biological particles for sorting are targets for sorting based on a threshold value set by the threshold setting unit;
The bioparticle analysis system according to claim 1 . the biogenic particle sorting device includes a sorting unit that sorts particles to be sorted based on the sorting determination of the determination unit,
The bioparticle analysis system according to claim 4 . The biological particle analysis system according to claim 4 , wherein the biological particles for separation are contained in the sample. The bioparticle analysis system according to claim 1 , wherein the threshold value is set to a predetermined threshold value. The bioparticle analysis system according to claim 7 , wherein the predetermined threshold is determined according to one or more modes. The bioparticle analysis system according to claim 1 , wherein the threshold value is set by a user. The data compression process is a dimensional compression process,
The bioparticle analysis system according to claim 1 , wherein a fractionation target range is determined after the dimensionality reduction. a compression unit that performs a data compression process on measurement data obtained by measuring the biogenic particles contained in a sample;
a gate unit that gates the measurement data compressed by the compression unit into training measurement data and verification measurement data and adds a label to the training measurement data;
a learning unit that uses the learning measurement data and the label to construct a learning model for determining whether the biogenic particles are to be sorted;
an inference unit that inputs the verification measurement data into a learning model constructed by the learning unit and infers whether the data is a target for collection;
a certainty factor calculation unit that calculates a certainty factor of the verification measurement data used in the inference;
a threshold setting unit that sets a threshold for dividing the sample based on the certainty calculated by the certainty calculation unit;
An information processing device having the above configuration. The constructed learning model is output to a microparticle sorting device.
The information processing device according to claim 11. a compression step of performing a data compression process on the measurement data measured from the biogenic particles contained in the sample;
a gating step of gating the data-compressed measurement data into training measurement data and verification measurement data, and adding a label to the training measurement data;
a learning step of constructing a learning model for determining whether the biological particle is a separation target by using the learning measurement data and the label; and an inference step of inputting the verification measurement data into the learning model constructed by the learning step and inferring whether the biological particle is a separation target.
a confidence level calculation step of calculating a confidence level of the verification measurement data used in the inference;
a threshold setting step of setting a threshold for dividing the sample based on the certainty calculated by the certainty calculation step;
An information processing method comprising the steps of: The information processing method according to claim 13 , further comprising a measuring step in which the measurement data is measured using a microparticle analysis device. 15. The information processing method according to claim 14, further comprising the steps of: inputting optical information measured from the biological particles for separation in the microparticle analysis device into a learning model constructed by the learning step; inferring whether the biological particles for separation are targets for separation; and, when it is inferred that the biological particles for separation are targets for separation, making a separation determination based on the threshold value set by the threshold setting step. The information processing method according to claim 15 , further comprising the step of separating particles to be separated based on the separation determination. an acquisition unit that acquires a plurality of pieces of measurement data including optical information measured from biogenic particles contained in a sample;
a clustering unit that classifies the plurality of pieces of measurement data acquired by the acquisition unit into a plurality of clusters;
a cluster selection unit that selects a cluster to be collected from the clusters classified by the clustering unit;
a threshold setting unit that sets a threshold based on the plurality of pieces of measurement data included in the cluster selected by the cluster selection unit. The clustering unit includes:
Classifying the plurality of pieces of measurement data acquired by the acquisition unit into clusters;
18. The information processing device according to claim 17, further comprising a sorting unit that sorts the measurement data to be sorted out from the measurement data included in the clusters classified by the clustering unit based on the threshold value set by the threshold setting unit. The threshold setting unit is
The information processing apparatus according to claim 17 , wherein the threshold is set to a representative value of the cluster or a median value of the plurality of pieces of measurement data included in the cluster selected by the cluster selection unit.

Images