Analysis device, analysis system, analysis method, and program

Abstract

Provided are an analysis device and the like with improved prediction accuracy and robustness.　The analysis device comprises: an acquisition unit that acquires data about optical characteristics that interact with a subject and change behavior; a learning unit that generates a trained model on the basis of the data acquired for a known subject by the acquisition unit; and an analysis unit that inputs data acquired for an unknown subject by the acquisition unit to the trained model generated by the learning unit, and outputs an analysis result for the unknown subject, wherein the learning unit includes a weak learner generation unit that generates a plurality of weak learners from learning data based on the data by using a neural network, a strong learner generation unit that generates a strong learner by combining the plurality of weak learners via boosting, and a trained model generation unit that generates a trained model from the strong learner.

Description

Analysis Device, Analysis System, Analysis Method, and Program

[0001] The present disclosure relates to an analysis device, an analysis system, an analysis method, and a program.

[0002] Methods for discriminating the type of a subject, detecting abnormalities in the subject, estimating the performance of the subject, etc. using various analysis methods have been widely developed. In such analysis methods, data acquired in a large number of samples in advance is machine-learned to construct a learned model. Based on the constructed learned model, the actually acquired data is analyzed to derive a determination result or an estimation result.

[0003] Generally, the larger the number of sample data used for analysis, the more accurately the determination or estimation can be made. However, actual data often has variations and may include values (outliers) that deviate significantly from the normal range. To address this problem, methods for improving the prediction accuracy of analysis by excluding outliers have been proposed (Patent Documents 1 to 3).

[0004] In Patent Documents 1 and 2, a prediction formula is generated from the acquired data by discriminant analysis, and outliers are detected as abnormalities in the specimen. In Patent Document 3, by updating the accumulated learning data group, even if the sensor detects an outlier, it can be immediately deleted.

[0005] JP-A-2021-89484 JP-A-2023-71993 JP-A-2015-35118

[0006] However, only these methods for excluding outliers are insufficient in terms of prediction accuracy and robustness, and further improvement is required. In addition, these methods for excluding outliers significantly narrow the analysis range as the number of outliers increases, and the prediction accuracy and robustness tend to decrease.

[0007] The problem to be solved by the present disclosure is to provide an analysis device, etc. with improved prediction accuracy and robustness.

[0008] To solve the above problems, the analysis device of this disclosure comprises: an acquisition unit that acquires data on optical properties whose behavior changes when interacting with a subject; a learning unit that generates a trained model based on the data acquired by the acquisition unit for a known subject; and an analysis unit that inputs the data acquired by the acquisition unit for an unknown subject into the trained model generated by the learning unit and outputs an analysis result for the unknown subject. The learning unit includes: a weak learner generation unit that generates a plurality of weak learners from training data based on the data using a neural network; a strong learner generation unit that generates a strong learner by combining the plurality of weak learners by boosting; and a trained model generation unit that generates a trained model from the strong learner.

[0009] The analysis system of this disclosure comprises the analysis device and a measuring device for measuring optical properties whose behavior changes in interaction with a subject.

[0010] The analysis method of this disclosure is an analysis method performed by an analysis device, comprising: an acquisition step of acquiring data on optical properties whose behavior changes in interaction with a subject; a learning step of generating a trained model for a known subject based on the data acquired in the acquisition step; and an analysis step of inputting the data acquired in the acquisition step for an unknown subject into the trained model generated by the learning step and outputting an analysis result for the unknown subject, wherein the learning step comprises: a weak learner generation step of generating a plurality of weak learners from training data based on the data using a neural network; a strong learner generation step of generating a strong learner by combining the plurality of weak learners by boosting; and a trained model generation step of generating a trained model from the strong learner.

[0011] The program of this disclosure is a program that causes the computer of an analysis device to function as an acquisition unit that acquires data on optical properties whose behavior changes when interacting with a subject; a learning unit that generates a trained model based on the data acquired by the acquisition unit for a known subject; and an analysis unit that inputs the data acquired by the acquisition unit for an unknown subject into the trained model generated by the learning unit and outputs an analysis result for the unknown subject, wherein the learning unit generates a plurality of weak learners from training data based on the data using a neural network, generates a strong learner by combining the plurality of weak learners by boosting, and generates a trained model from the strong learner.

[0012] According to this disclosure, the predictive accuracy and robustness of the analysis can be improved.

[0013] This is a block diagram illustrating the schematic configuration of the analysis system. This is a schematic cross-sectional diagram of a laminate. This is a schematic cross-sectional diagram of a laminate coated with a sample. This is a schematic cross-sectional diagram of a laminate coated with a light-emitting probe. This is a schematic cross-sectional diagram of an organic EL element. This is a diagram explaining the mechanism by which light is emitted when the target substance and light-emitting dye molecules interact. This is a diagram explaining the synthesis method of light-emitting dye molecules. This shows a flowchart of the measurement method. This is a flowchart showing an example of the operation of the analysis system. This is a flowchart showing an example of the operation of the learning process in the analysis system. This is an explanatory diagram showing the structure of a neural network. This is an explanatory diagram of a neuron.

[0014] Hereinafter, one or more embodiments of this disclosure will be described with reference to the drawings. However, the scope of this disclosure is not limited to the disclosed embodiments.

[0015] [Configuration of the Analysis System] Figure 1 is a block diagram showing the schematic configuration of the analysis system 100 of this embodiment. The analysis system 100 comprises a measuring device 10 and an analysis device 20. The measuring device 10 and the analysis device 20 may communicate with each other. Alternatively, the measuring device 10 and the analysis device 20 may be integrated. The measuring device 10 comprises a measuring unit 11, a communication unit 12, and a control unit 13. The analysis device 20 comprises a display unit 21, an operation unit 22, a communication unit 23, and a control unit 24.

[0016] As mentioned above, the data actually obtained from subjects often contains variability. There are two factors that cause this variability: first, variability in the subjects themselves, and second, variability in the measuring device itself. The variability caused by the first factor is important for detailed analysis of the subjects. The variability caused by the second factor significantly reduces the predictive accuracy and robustness of the analysis of the subjects, so it is preferable to reduce it.

[0017] The measuring device 10 measures the optical properties that change in behavior in response to interaction with the subject. Generally, devices that measure the behavior of light have good sensitivity and can acquire complex data, but on the other hand, the data is easily altered by even slight factors. Therefore, variations in the acquired data are likely to occur due to individual differences that occur during the manufacturing of the device, minute changes in the luminescent probe contained within the device, etc.

[0018] In other words, even if the same sample is measured multiple times, differences are likely to occur in the data obtained each time. If the trained model determines that this difference is due to differences in the sample itself, the analysis results will change each time, even when analyzing the same sample. However, for the analysis device 20, it is preferable that the analysis results be consistent even when measuring the same sample multiple times. In this specification, the degree to which the analysis results are consistent when the same sample is measured multiple times is referred to as "robustness (reproducibility)." A higher degree of consistency in the analysis results indicates higher robustness.

[0019] In this embodiment, the subject is analyzed using machine learning. In machine learning using neural networks, features can be automatically learned from training data, and the nonlinear and complex relationship between input and output can also be learned to generate a trained model (predictive model). On the other hand, in machine learning using neural networks, if data that is significantly different from the training data is input to the trained model, the accuracy and robustness of the output prediction tend to decrease.

[0020] Therefore, multiple weak learners are generated using a neural network, and a strong learner is generated by combining these multiple weak learners through boosting. A trained model is generated from the strong learner, and the subject is analyzed using the generated trained model. As a result, even if data that is significantly different from the training data is input to the trained model, that is, even if the data is variable due to variability in the measurement itself, the analysis can be performed with high predictive accuracy and robustness.

[0021] The details of each device are described below.

[0022] [Measurement Device] (Measurement Unit) The number of data points measured by the measurement unit 11 and acquired by the control unit 24 of the analysis device 20 is not particularly limited, but the more data points there are, the better the prediction accuracy of the analysis. The number of data points is appropriately selected depending on the measurement method, etc. It is preferable that the number of data points be equal to or greater than the number of explanatory variables. If a large number of data points can be acquired, it is preferable that the number of data points be 10 times or more the number of explanatory variables, and more preferably 100 times or more.

[0023] The measurement unit 11 is not particularly limited as long as it is a device that measures optical properties whose behavior changes in interaction with the sample. One such measurement method is to irradiate the sample with light from a light source and acquire spectral data of the light obtained from the sample. Examples of light irradiated from the light source include X-rays, ultraviolet rays, visible light, infrared rays, etc.

[0024] In this embodiment, the measurement unit 11 preferably uses a light-emitting probe to acquire spectral data of light obtained from the sample. In this embodiment, a "light-emitting probe" refers to a substance that has the property of binding to or associating with the target substance and emits light when returning from an excited state to a ground state. By using a light-emitting probe, accurate data about the target substance can be obtained, and the prediction accuracy of the analysis can be improved.

[0025] If the sample contains multiple substances, the luminescent probe may bind to or associate only with specific target substances within the sample. The luminescent probe may be a molecule, a molecular aggregate, or a particle containing molecules.

[0026] The method for acquiring spectral data of light obtained from a sample using a light-emitting probe is not particularly limited. For example, one method is to use an organic electroluminescent (EL) element. Another method is to use a light-emitting dye molecule having a main chain of a specific structure as the light-emitting probe.

[0027] (Organic EL Elements) The following configurations are possible for organic electroluminescent (EL) elements: (i) Transparent substrate / anode / light-emitting layer (detection region) / cathode (ii) Transparent substrate / anode / light-emitting layer (detection region / electron transport layer) / cathode (iii) Transparent substrate / anode / light-emitting layer (hole transport layer / detection region) / cathode (iv) Transparent substrate / anode / light-emitting layer (hole transport layer / detection region / electron transport layer) / cathode (v) Transparent substrate / anode / light-emitting layer (hole transport layer / detection region / electron transport layer / electron injection layer) / cathode (vi) Transparent substrate / anode / light-emitting layer (hole injection layer / hole transport layer / electron blocking layer / detection region / hole blocking layer / electron transport layer) / cathode

[0028] The above-described organic EL element has a light-emitting layer between its electrodes, and the light-emitting layer is the detection region. The organic EL element can emit light when the test subject and the light-emitting probe interact within the light-emitting layer. Various data can be obtained by acquiring the light emission information and current density of the light-emitting probe. A method for acquiring data using the organic EL element will be explained with reference to Figures 2 to 5.

[0029] In this method, first, a laminate comprising a transparent substrate 111, a transparent electrode 112 (e.g., an ITO film), a hole transport layer 113 (e.g., a polyaromatic diamine), and a receiving layer 114 (e.g., polystyrene) is prepared (Figure 2). The transparent substrate 111, transparent electrode 112, hole transport layer 113, etc., can be the same as those used in known organic EL devices. The receiving layer 114 only needs to be able to receive the subject and the light-emitting probe. The receiving layer 114 may also serve as the host layer.

[0030] Next, the sample layer 121 is applied to a desired area of ​​the receiving layer 114 of the laminate using any method (Figure 3).

[0031] Next, the luminescent probe is applied in a pattern to desired locations on the receiving layer 114 coated with the subject layer 121 using an inkjet method or the like (Figure 4). As a result, the areas coated with the luminescent probe become the luminescent layer 122. For example, by applying a component that dissolves the receiving layer 114 together with the luminescent probe, the applied luminescent probe, a part of the subject layer 121, and a part of the receiving layer 114 are mixed, and an integrated luminescent layer 122 is formed. In the luminescent layer 122, the subject and the luminescent probe interact. Although not shown in the figure, the luminescent probe may also be applied to areas where the subject layer 121 is not coated to obtain luminescence information from the luminescent probe alone.

[0032] Finally, a counter electrode layer 115, which is paired with the transparent electrode 112, is placed on the light-emitting layer 122 to fabricate an organic EL element (Figure 5).

[0033] In the fabricated organic EL element, the test subject and the light-emitting probe in the light-emitting layer 122 are excited by a conventional method, and emission spectral data is acquired.

[0034] When acquiring a light spectrum using an organic EL element, i.e., when using an organic EL element as a sensing device, fluorescent compounds, delayed-fluorescence compounds, and phosphorescent compounds can be used as light-emitting probes. Furthermore, different phosphorescent compounds may be used in combination as light-emitting probes, or phosphorescent compounds and fluorescent compounds may be used in combination. This allows for the acquisition of any desired emission color. Additionally, multiple light-emitting compounds with different emission colors may be combined to produce white light.

[0035] In this specification, "fluorescent compound" refers to a compound that emits fluorescence other than delayed fluorescence. "Fluorescence" refers to the light emitted when returning from a singlet excited state to the ground state. "Fluorescence other than delayed fluorescence" refers to fluorescence excluding the "delayed fluorescence" exemplified below. Examples of "delayed fluorescence" include "thermally activated delayed fluorescence (TADF)". Examples of "delayed fluorescence" include "triplet-triplet annihilation (TTA) delayed fluorescence".

[0036] In other words, in this specification, "fluorescent compound" does not include "delayed fluorescent compounds" such as "thermally activated delayed fluorescent compounds" and "triplet-triplet annihilation delayed fluorescent compounds." "Fluorescent compound" refers to a compound that does not undergo upconversion by reverse intersystem crossing from the lowest excited triplet energy level to the lowest excited singlet energy level.

[0037] Fluorescent compounds do not necessarily need to be heavy metal complexes like phosphorescent compounds. So-called organic compounds, composed of common combinations of elements such as carbon, oxygen, nitrogen, and hydrogen, can be used as fluorescent compounds. Other nonmetallic elements such as phosphorus, sulfur, and silicon may also be used as fluorescent compounds. Complexes of typical metals such as aluminum and zinc may also be used as fluorescent compounds. Known fluorescent compounds used in the light-emitting layers of organic EL devices may also be used as fluorescent compounds.

[0038] As used herein, the term "phosphorescent compound" refers to a compound that emits phosphorescence. Specifically, a "phosphorescent compound" is a compound that emits phosphorescence at room temperature (25°C), and has a phosphorescence quantum yield of 0.01 or more at 25°C. The phosphorescence quantum yield is preferably 0.1 or more.

[0039] "Phosphorescence" refers to light emitted when returning from a triplet excited state to the ground state. Known phosphorescent compounds used in the light-emitting layer of an organic EL device can be used as the phosphorescent compound.

[0040] As used herein, the term "delayed fluorescence compound" refers to a compound that emits delayed fluorescence. "Delayed fluorescence" refers to light emitted when returning from a singlet excited state to the ground state as a result of up-conversion due to reverse intersystem crossing from the lowest excited triplet energy level to the lowest excited singlet energy level. Known delayed fluorescence compounds used in the light-emitting layer of an organic EL device can be used as the delayed fluorescence compound.

[0041] When an organic EL device is used as a sensing device, for blue, green, and red with different emission colors, a plurality of data may be taken at intervals of several nm each. Also, the current density may be measured every 1V. Thereby, a large number of data can be obtained.

[0042] (Light-emitting dye molecule) In the method using the light-emitting dye molecule described below, microwells capable of forming a plurality of reaction fields on a plane are used. Different types of light-emitting dye molecules are prepared in advance on each well, and the sampled specimen is injected into each well to obtain emission spectrum data. Thereby, the state of the specimen is digitized from a plurality of emission spectrum data.

[0043] In this method, a plate in which a plurality of wells are regularly arranged is used. In a plate having such wells, the wells (reaction fields) are physically separated from each other by partition walls. Therefore, it is difficult for the specimen and the light-emitting dye molecule to mix in adjacent reaction fields, and accurate analysis is easy.

[0044] The luminescent dye molecule exhibits two or more types of luminescence selected from fluorescence, phosphorescence, excimer luminescence, exciplex luminescence, thermally activated delayed fluorescence, excited-state intramolecular proton luminescence, triplet-triplet annihilation delayed fluorescence, twisted intramolecular charge transfer luminescence, and aggregated organic luminescence with respect to a single excitation light.

[0045] The luminescent dye molecule can be used for analyzing the structure, state, etc. of a specific target substance. Specifically, when the luminescent dye molecule and the target substance interact with each other, the structure, electronic state, etc. in the chromophore or luminophore of the luminescent dye molecule change, and a complex luminescence behavior different from that of the luminescent dye molecule alone is obtained.

[0046] For example, as shown in FIG. 6 with respect to a single excitation light, a luminescent dye molecule that exhibits three different types of luminescence, namely fluorescence, phosphorescence, and excimer luminescence, is allowed to interact with a target substance. Due to the interaction between the target substance and the luminescent dye molecule, the processes by which fluorescence, phosphorescence, and excimer luminescence occur respectively change, and the wavelength or lifetime of each light changes. As a result, depending on the structure, state, etc. of the target substance, a complex and large number of data in which these lights are combined can be obtained, and from this data, the structure, state, etc. of the target substance can be grasped in great detail.

[0047] Hereinafter, the specific structure of the luminescent dye molecule according to this embodiment will be described.

[0048] The main chain of the luminescent dye molecule may have one or more structural units including a sugar structure derived from pentose or hexose and a phosphate ester bond bonded to the sugar structure. The main chain may contain only one of the above structural units, or may contain a plurality of them. That is, it may be a structure having one each of the above sugar structure and the phosphate ester bond bonded to the sugar structure, or may be a structure alternately containing the above sugar structure and the phosphate ester bond.

[0049] Both ends in the main chain of the luminescent dye molecule are sugar structures, and the number of sugar structures is one more than the number of phosphate ester bonds. When the main chain contains a plurality of structural units, the plurality of structural units may be the same as each other or different from each other.

[0050] The number of structural units contained in the main chain of the luminescent dye molecule is appropriately selected according to the type of target substance, and is preferably in the range of 2 to 6. If the amount of structural units is large, the luminescent dye molecule is more likely to act specifically on the target substance. In this embodiment, it is preferable to obtain a large amount of data by having the luminescent dye molecule interact with various positions on the target substance. For this reason, it is preferable that the luminescent dye molecule and the target substance have a moderate degree of specificity, not excessively so, and the number of structural units is preferably 6 or less.

[0051] The main chain of the luminescent dye molecule may include structures other than the above-mentioned structural units, to the extent that it does not impair the purpose and effect of this embodiment. The structures at both ends of the main chain are not particularly limited and may include, for example, hydroxyl groups, alkoxy groups, etc.

[0052] Examples of pentoses include ribose, deoxyribose, and xylose. Examples of hexoses include allose, glucose, and mannose. Among these, it is particularly preferable that the sugar structure is derived from ribose or deoxyribose. This allows the main chain of the luminescent pigment molecule to have a structure similar to that of DNA or RNA, and the luminescent pigment molecule to readily interact with DNA or RNA.

[0053] When the above structural unit includes a structure derived from ribose or deoxyribose, the phosphate ester bond is preferably bonded to the carbon at position 3 and position 5 of ribose or deoxyribose. The chromophore or luminescent phore is preferably bonded to the carbon at position 1 of ribose or deoxyribose. In other words, the luminescent dye molecule according to this embodiment preferably includes a structure represented by the following general formula (1a) or (1b).

[0054]

[0055] In the general formulas (1a) and (1b), Y represents a chromophore or luminescent phose as described below.

[0056] The chromophore or luminescent phore of the luminescent dye molecule may have a structure that emits a predetermined type of light independently in response to a single excitation light, or emits a predetermined light through the action of multiple chromophores or luminescent phores. Preferably, the chromophore or luminescent phore is bonded to the sugar structure of the main chain such that the sugar structure is in its β-form.

[0057] In this embodiment, "chromophore" refers to a structure that absorbs light with a wavelength of 300 nm or more, and "luminescent phore" refers to a structure that absorbs light with a wavelength of 300 nm or more and emits light.

[0058] The number of chromophores or luminescent groups in a luminescent dye molecule may be as few as one, provided that the luminescent dye molecule is capable of exhibiting multiple types of luminescence. However, from the viewpoint of making it easier for the luminescent dye molecule to exhibit multiple types of luminescence, the number of chromophores or luminescent groups is preferably two or more, and more preferably in the range of three to six. When a luminescent dye molecule has multiple chromophores or luminescent groups, there may be only one type, or there may be two or more types.

[0059] In a luminescent dye molecule, it is preferable that one chromophore or luminescent phore is bound to one sugar structure in the main chain. Therefore, if a luminescent dye molecule has two or more chromophores or luminescent phore, it is preferable that there are also two or more sugar structures in the main chain. In other words, it is preferable that the number of chromophores or luminescent phore in a luminescent dye molecule is equal to or less than the number of sugar structures in the main chain.

[0060] If the number of chromophores or luminescent groups in a luminescent dye molecule is less than the number of sugar structures in the main chain, some sugar structures will not have chromophores or luminescent groups attached. Sugar structures that do not have chromophores or luminescent groups attached do not need to have other atomic groups attached, but they may have native bases attached to them to the extent that it does not impair the purpose and effect of this embodiment. In this specification, "native bases" refers to adenine, guanine, cytosine, thymine, and uracil.

[0061] The total number of native bases bound to the sugar structure is preferably 50% or less, and more preferably 25% or less, of the total number of sugar structures in the main chain. Having 50% or less native bases suppresses the association of luminescent dye molecules, making the interaction between the test subject and the luminescent dye molecules more dominant. Furthermore, it is preferable that the native and non-native bases are bound in such a way that the sugar structure is in the β-form.

[0062] Examples of fluorescent chromophores or luminescent phoses include structures derived from fluorescein, rhodamine, boron dipyromethene, etc. Examples of phosphorescent chromophores or luminescent phoses include structures derived from iridium complexes, platinum complexes, etc. Examples of excimer-luminescent chromophores or luminescent phoses include structures derived from pyrene, anthracene, perylene, etc. Examples of exciplex-luminescent chromophores or luminescent phoses include structures derived from pyrene-dimethylaniline, etc.

[0063] Examples of chromophores or luminescent molecules that emit thermally activated delayed fluorescence include structures derived from 4CzIPN, DABNA, etc. Examples of chromophores or luminescent molecules that emit excited-state intramolecular proton emission include structures derived from hydroxyphenylbenzoxazole, etc. Examples of chromophores or luminescent molecules that emit triplet-triplet annihilation delayed fluorescence include structures derived from 9,10-diphenylanthracene, rubrene, etc. Examples of chromophores or luminescent molecules that emit twisted intramolecular charge transfer emission include structures derived from diaminoanthracene, diaminonaphthalene, etc. Examples of chromophores or luminescent molecules that emit aggregated organic emission include structures derived from tetraphenylethene, hexaphenylsilole, etc.

[0064] In particular, the chromophore or luminescent phore preferably contains at least one structure selected from a structure that emits fluorescence, a structure that emits excimer emission, and a structure that emits exciplex emission. It is especially preferable that the chromophore or luminescent phore contains a structure that emits fluorescence. The fluorescence emitted by the luminescent dye molecule makes it easier to analyze with various measuring devices.

[0065] It is preferable that the luminescent dye molecule exhibits multiple types of emission when irradiated with light in the wavelength range of 300 to 400 nm. Because the luminescent dye molecule exhibits multiple types of emission, a special light source is not required when analyzing a subject, and the analysis is less likely to damage the subject.

[0066] The molecular weight of the luminescent dye molecule is appropriately selected depending on the type of chromophore or luminescent phose, the length of the main chain, etc. The molecular weight is preferably in the range of 500 to 10,000, and more preferably in the range of 100 to 4,000. A molecular weight of 10,000 or less results in moderately low specificity to the test subject, allowing the luminescent dye molecule to react non-specifically with multiple locations in the target substance within the test subject.

[0067] Luminescent dye molecules can be synthesized by the following procedure: Prepare monomers by bonding a chromophore or luminescent ester and a phosphate ester to a pentose or hexose. Polymerize these monomers in the desired sequence using the phosphoramidide method with a DNA / RNA synthesizer or the like.

[0068] In this synthesis method, as shown in Figure 7, for example, multiple types of monomers with different chromophores or luminescent phoses are prepared, and the desired number of monomers can be combined by changing the sequence of these monomers. In the example shown in Figure 7, three types of monomers with chromophores or luminescent phoses A, B, and C, respectively, are prepared, and the three monomers can be combined by changing the sequence of these monomers.

[0069] In other words, a wide variety of luminescent dye molecules can be synthesized from multiple types of monomers with different chromophores or luminescent phores. In the example shown in Figure 7, 27 different luminescent dye molecules can be synthesized. By changing the type of monomer used or the number of monomer bonds, a very large number of luminescent dye molecules can be synthesized.

[0070] The luminescent dye molecules according to this embodiment have a main chain structure similar to naturally occurring substances such as DNA and RNA, and can easily interact with various target substances contained in the sample. Therefore, by using luminescent dye molecules, data about the target substances can be easily and in detail. Furthermore, the luminescent dye molecules exhibit multiple types of emission when irradiated with light at a specific wavelength. Therefore, complex and large amounts of data can be obtained depending on the state of the target substance.

[0071] The following describes an example of a method for measuring and acquiring signal data using luminescent dye molecules, but the measurement method is not limited to this method.

[0072] Figure 8 shows a flowchart of the measurement method. A microwell plate is prepared for the interaction of the luminescent dye molecule and the test subject, and either the luminescent dye molecule or the test subject is placed in the reaction field (well) (first component placement step: step S101). The placed luminescent dye molecule or test subject is referred to as the first component.

[0073] The method for placing the first component in each reaction site is not particularly limited and can be appropriately selected depending on the type and physical properties of the first component. Examples of methods for placing the first component include coating by inkjet, coating by dispenser, placement of a carrier supporting the first component, and direct fixation of the first component to the reaction site. Among these, the inkjet method is preferred. By using the inkjet method, liquid first components can be efficiently placed in multiple reaction sites, and a large amount of data can be acquired.

[0074] If the first component is a luminescent dye molecule and the plate has multiple reaction fields, the same type of first component may be placed in all of the reaction fields, or multiple types of first components may be placed in the same reaction field. Alternatively, different types of first components may be placed in two or more reaction fields. By placing different types of first components in different reaction fields, multiple types of interactions occur between the luminescent dye molecule and the target substance, allowing for a more detailed analysis of the target substance.

[0075] The measurement is started, and first signal data is acquired from the plate on which the first component is placed (first signal data acquisition step: step S102).

[0076] A luminescent dye molecule or the subject, as the other component of the first component, is placed in the reaction field of the plate from which the first signal data was acquired (second component placement step: step S103). The placed luminescent dye molecule or subject is referred to as the second component.

[0077] The second signal data is acquired from the plate containing the second component (second signal data acquisition step: step S104), and the measurement is terminated.

[0078] (Communication Unit) The communication unit 12 transmits the data measured by the measurement unit 11 to the communication unit 23 of the analysis device 20. The communication unit 12 may use either wireless communication or wired communication.

[0079] (Control Unit) The control unit 13 is a processor that provides overall control over the operation of the measuring device 10. The control unit 13 includes a CPU (Central Processing Unit) that performs various calculations, and RAM (Random Access Memory) that provides the CPU with a working memory space and stores temporary data.

[0080] [Analysis Device] (Display Unit) The display unit 21 displays various information on the screen based on the display control signal received from the control unit 24. The display unit 21 is equipped with a display device. Examples of display devices include a display, a projector, etc. The display unit 21 displays the results of the analysis performed by the control unit 24 based on the data measured by the measurement unit 11 of the measuring device 10 and notifies the user.

[0081] (Operation Unit) The operation unit 22 accepts various inputs through user operation. The operation unit 22 is equipped with input devices. Examples of input devices include keyboards, mice, various switches, touchscreens, touchpads, etc.

[0082] (Communication Unit) The communication unit 23 receives data transmitted from the communication unit 12 of the measuring device 10. The communication unit 23 may use either wireless communication or wired communication.

[0083] (Control Unit) The control unit 24 is a processor that provides overall control over the operation of the analysis device 20. The control unit 24 includes a CPU (Central Processing Unit) that performs various calculations, and RAM (Random Access Memory) that provides the CPU with working memory space and stores temporary data.

[0084] The control unit 24 acquires data measured by the measurement unit 11 of the measuring device 10, that is, data on optical properties whose behavior changes in interaction with the subject. At this time, the control unit 24 functions as an acquisition unit. The control unit 24 generates a trained model based on the data acquired by the acquisition unit for known subjects. At this time, the control unit 24 functions as a learning unit. The control unit 24 inputs the data acquired by the acquisition unit for an unknown subject into the trained model generated by the learning unit and outputs the analysis results for the unknown subject. At this time, the control unit 24 functions as an analysis unit. Specifically, the CPU reads the stored program and loads it into RAM, and works with the program loaded into RAM to perform data acquisition, learning, and analysis.

[0085] [Subject] The type of subject in this embodiment is not particularly limited and may be a substance whose structure is known or a substance whose structure is unknown. The subject may be a mixture of various compounds, etc. The subject may belong to any field, such as the medical field, the industrial field, or the food field. The measurement unit 11 may measure the entire subject or measure a specific target substance within the subject.

[0086] Examples of target substances belonging to the medical field include proteins, antibodies, antibody-containing beads, and tumor markers. Examples of subjects containing these target substances include sweat and blood. Examples of target substances belonging to the industrial field include metal ions, metal nanoparticles, carbon nanotubes, magnetic fluids, nanosilica, and crystalline zirconia. Examples of subjects containing these target substances include rainwater, rivers, small ponds, tropical fish breeding water, soil, and wastewater. In other words, the analytical device of this embodiment can be applied to water quality management. Examples of target substances belonging to the food field include proteins, lipids, and carbohydrates. Examples of subjects containing these target substances include food products in general.

[0087] [Operation of the Analysis System] Figure 9 is a flowchart showing an example of the operation of the analysis system 100. When the analysis of the subject is started, the following operations are performed.

[0088] First, the control unit 24 instructs the measurement unit 11 of the measuring device 10 to measure the optical properties that change in behavior when interacting with a known subject (Step S1: Measurement step of a known subject).

[0089] Here, "known" means that at least one item of information about the subject that is to be analyzed is known. Information about the subject includes the structure and content of compounds contained in the subject. If the subject is food, information about the subject includes sensory evaluations such as flavor, texture, and smell.

[0090] The type of known subject does not necessarily have to be the same as the unknown subject being analyzed, but it is preferable that they be the same from the viewpoint of predictive accuracy and robustness. For example, when analyzing the type and content of metal ions contained in water, it is preferable that the known subject is water containing the same type of metal ions. There is no particular limit to the number of known subject samples, but a larger number of samples is preferable from the viewpoint of predictive accuracy and robustness. If there are two or more samples, the types of known subjects may be the same or different.

[0091] Next, the control unit 24 acquires data on the optical properties whose behavior changes when interacting with a known subject (Step S2: Data acquisition step for a known subject). At this time, the control unit 24 functions as an acquisition unit. The data to be acquired may be data measured by the measurement unit 11 of the measuring device 10, data measured by other measuring devices, or both.

[0092] The optical properties data acquired in the acquisition unit are complex and voluminous. By using this complex and extensive data, it becomes possible to analyze the sensory evaluation of subjects, which is generally difficult.

[0093] Next, the control unit 24 generates a trained model based on the data acquired by the acquisition unit for a known subject (Step S3: Training process). At this time, the control unit 24 functions as a learning unit. Details of the training process will be described later.

[0094] Next, the control unit 24 instructs the measurement unit 11 of the measuring device 10 to measure the optical properties that change in behavior when interacting with the unknown subject (Step S4: Measurement step of the unknown subject).

[0095] Next, the control unit 24 acquires data on the optical properties whose behavior changes when interacting with the unknown subject (Step S5: Data acquisition step for the unknown subject). At this time, the control unit 24 functions as an acquisition unit. The data to be acquired may be data measured by the measurement unit 11 of the measuring device 10, data measured by other measuring devices, or both.

[0096] Next, the control unit 24 inputs the data acquired by the acquisition unit for the unknown subject into the trained model generated by the learning unit, outputs the analysis results for the unknown subject (step S6: analysis process), and terminates the operation.

[0097] The details of the learning process are described below. (Learning Process) Figure 10 is a flowchart showing an example of the operation of the learning process S3 in the analysis system 100. When the learning process S3 is started, the following operations are performed. Note that the various pre-processing steps of the pre-processing step S32, feature selection step S34, and dimensionality reduction step S36 do not necessarily need to be performed and are performed as needed.

[0098] In the learning process S3, the control unit 24 first performs various preprocessing steps on the data acquired by the acquisition unit for known subjects to generate learning data. The explanatory variables for the learning data can include numerical values ​​representing the characteristics of the acquired data, and numerical values ​​calculated from them. If the acquired data is a spectral distribution, explanatory variables can include the intensity of light at each wavelength. The target variable can be appropriately selected according to the purpose of the analysis. The target variable is not limited to the structure or content of compounds contained in the subject; other variables related to the subject may be used, or it may be a sensory evaluation of the subject.

[0099] First, the control unit 24 determines whether or not to perform preprocessing on the data acquired by the acquisition unit for a known subject (step S31: preprocessing determination step). If the control unit 24 determines that preprocessing is necessary (step S31: YES), it performs preprocessing (step S32: preprocessing step). If the control unit 24 determines that preprocessing is not necessary (step S31: NO), it proceeds to step S33.

[0100] Data preprocessing is not particularly limited, but examples include outlier removal, standardization, normalization, type conversion, and handling of missing values ​​(removal or imputation). Preprocessing can improve prediction accuracy and robustness. However, as mentioned above, it is preferable to retain the variability of the subjects themselves, and to remove outliers caused by variability of the device itself.

[0101] The method for removing outliers is not particularly limited, and known methods can be used. Examples of methods for removing outliers include methods for identifying outliers based on quantiles, methods for identifying outliers based on robust estimation (Hubuer method, Cauchy method, quartile method, principal component analysis, etc.), K-nearest neighbors (K-Means cluster analysis), hierarchical cluster analysis, normal mixture analysis, latent class analysis, and variable clustering. Outliers can also be removed by standardizing the data.

[0102] Next, the control unit 24 performs intermediate processing on the data acquired by the acquisition unit for known subjects, as necessary. The intermediate processing includes a feature selection step S34 and a dimensionality reduction step S36. Either one of these may be performed, or both may be performed.

[0103] The control unit 24 determines whether to perform feature selection, which involves selecting some of the multiple features included in the acquired data (step S33: feature selection determination step). If the control unit 24 determines to perform feature selection (step S33: YES), it performs feature selection (step S34: feature selection step). If the control unit 24 determines not to perform feature selection (step S33: NO), it proceeds to step S35.

[0104] In this embodiment, "feature" refers to a numerical value that quantitatively represents the features or characteristics of the data acquired by the acquisition unit. Feature may be an explanatory variable or a dependent variable. If there are many types of features extracted from the acquired data, it is preferable to select only the significant features. By selecting features, multidimensional data can be visualized (converted to 2D), for example, and transformed into data that is easier to handle. Furthermore, by selecting features, the series of analysis processes can be sped up. In addition, by selecting features, overfitting can be reduced.

[0105] The method for selecting features is not particularly restricted, and publicly known methods can be used. Examples of feature selection methods include filtering, wrapping, and embedded methods. In filtering, multiple features are evaluated and ranked individually. In wrapping, multiple features are evaluated in combination. In embedded methods, learning and feature evaluation are performed simultaneously.

[0106] If the data acquired by the acquisition unit is a spectral distribution, for example, only the light intensity in the wavelength region (band) of interest may be selected as a feature from among many wavelength regions (bands).

[0107] The control unit 24 determines whether to perform dimensionality reduction, which reduces the number of dimensions of multiple features included in the acquired data (step S35: dimensionality reduction determination step). If the control unit 24 determines to perform dimensionality reduction (step S35: YES), it performs dimensionality reduction (step S36: dimensionality reduction step). If the control unit 24 determines not to perform dimensionality reduction (step S35: NO), it proceeds to step S37.

[0108] When there are many types of features extracted from the acquired data, i.e., when the number of dimensions is large, it is preferable to perform dimensionality reduction. By performing dimensionality reduction, multidimensional data can be visualized (reduced to 2D), for example, making it easier to handle. Furthermore, dimensionality reduction can speed up the analysis process. In addition, dimensionality reduction can reduce overfitting. Moreover, because the data is reduced in the dimensionality reduction process, less information contained in the data is lost compared to the feature selection process.

[0109] The method for performing dimensionality reduction is not particularly limited, and known methods can be used. Examples of dimensionality reduction methods include principal component analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA), non-negative matrix factorization (NMF), and non-negative matrix factorization with soft orthogonality constraints (NMF-so). The acquired data may be used as training data as is, or the data may be used as training data after performing one or more of the following: preprocessing, feature selection, and dimensionality reduction.

[0110] Next, the control unit 24 generates multiple weak learners from the training data using a neural network (step S37: weak learner generation step). At this time, the control unit 24 functions as a weak learner generation unit.

[0111] Figure 11 is an explanatory diagram showing the structure of the neural network 300. The neural network 300 has a hierarchical structure and includes an input layer 300a, a feature extraction layer 300b, and an output layer 300c.

[0112] A "neural network" is an information processing system that mimics the human nervous network. In the neural network 300, the optical neuron model corresponding to a nerve cell is denoted as neuron U. The input layer 300a, the feature extraction layer 300b, and the output layer 300c each have multiple neurons U.

[0113] The input layer 300a is typically a single layer. The data acquired by the acquisition unit is input to each neuron U in the input layer 300a. The input data is output directly from each neuron U in the input layer 300a to the feature extraction layer 300b. The feature extraction layer 300b extracts features from the output data and outputs them to the output layer 300c. The output layer 300c outputs analysis information of the subject using the features extracted by the feature extraction layer 300b.

[0114] Figure 12 is an explanatory diagram of neuron U. Neuron U uses a multi-input, single-output element. Signals are transmitted in only one direction, and the input signal xi (i = 1, 2, ..., n) is multiplied by a certain neuron weight (SUwi) and input to neuron U. The neuron weight can be changed through learning. The sum of each input value (SUwi × xi) multiplied by the neuron weight SUwi is transformed by the activation function f(X) and then output from neuron U.

[0115] In other words, the output value y of neuron U is expressed by the following equation (1). Also, X in equation (1) is expressed by the following equation (2).

[0116] Equation (1) y=f(X) Equation (2) X=Σ(SUwi×xi)

[0117] As activation functions, for example, the ReLU function (Rectified Linear Unit: ramp function) or the sigmoid function can be used.

[0118] The method for adjusting the neuron weights in the neural network 300, i.e., the learning method, is not particularly limited, and known learning methods can be used. An example of a learning method for the neural network 300 is backpropagation.

[0119] In backpropagation, first, the error is calculated using a predetermined error function based on the known information (ground truth values) of the subject and the analysis information output from the output layer 300c. Next, the neuron weights of the feature extraction layer 300b and the output layer 300c are adjusted using methods such as the steepest descent method to minimize this error.

[0120] The training data is input to the input layer 300a of the neural network 300 and output from the input layer 300a to the feature extraction layer 300b. Each neuron U in the feature extraction layer 300b performs calculations on the input data with neuron weights, and data representing the extracted features is output to the output layer 300c. Each neuron U in the output layer 300c performs calculations on the input data with neuron weights. As a result, predicted values ​​are output from the output layer 300c based on the extracted features.

[0121] The output value (predicted value) of the output layer 300c is compared with known information (ground truth value) of the subject, and an error (loss) is calculated using a predetermined error function. The neuron weights of the feature extraction layer 300b and the output layer 300c are sequentially changed and adjusted to minimize this error (backpropagation). This generates a learner with adjusted neuron weights in the feature extraction layer 300b and the output layer 300c.

[0122] The learner generated here has higher predictive accuracy than learners generally referred to as "weak learners," that is, learners that have predictive accuracy only slightly better than random guessing. However, since the learner generated here has lower predictive accuracy compared to the strong learner generated by further combining these learners, it is referred to as a "weak learner" in this specification.

[0123] Next, the control unit 24 generates a strong learner by combining multiple weak learners through boosting (step S38: strong learner generation step). At this time, the control unit 24 functions as a strong learner generation unit.

[0124] In ensemble learning, a strong learner with superior prediction accuracy and robustness than a single learner can be generated by combining multiple weak learners. In this embodiment, ensemble learning is performed by boosting.

[0125] Boosting involves sequentially training multiple weak learners and combining them to generate a strong learner. When training in the order of first weak learner, second weak learner, the second weak learner prioritizes training on data where the error between the predicted value output by the first learner and the correct value is large, so as to reduce that error. When prioritizing data, each data point may be weighted. Also, data where the error is considered to have been eliminated by the previously trained weak learner may be ignored by the next weak learner to be trained.

[0126] The predicted values ​​output by the first weak learner are compared with the correct values, and the error (loss) is calculated using a predetermined error function. The neuron weights are adjusted to minimize this error, and the first weak learner is trained. If there is data with a large error between the predicted value and the correct value in the trained first weak learner, the second weak learner prioritizes learning this data. All weak learners from the second weak learner onward are trained by repeating the same procedure, and a strong learner is generated.

[0127] Next, the control unit 24 generates a trained model from the strong learner (step S39: trained model generation step) and terminates its operation. At this time, the control unit 24 functions as a trained model generation unit.

[0128] The neuron weights of the feature extraction layer 300b and output layer 300c in the generated trained model are stored in the control unit 24 as trained parameters.

[0129] In this embodiment, the analysis device 20 comprises an acquisition unit (control unit 24), a learning unit (control unit 24), and an analysis unit (control unit 24). The acquisition unit (control unit 24) acquires data on optical properties whose behavior changes in interaction with the subject. The learning unit (control unit 24) generates a trained model for known subjects based on the data acquired by the acquisition unit (control unit 24). The analysis unit (control unit 24) inputs the data acquired by the acquisition unit (control unit 24) for unknown subjects into the trained model generated by the learning unit (control unit 24) and outputs the analysis results for the unknown subjects. The learning unit (control unit 24) includes a weak learner generation unit (control unit 24), a strong learner generation unit (control unit 24), and a trained model generation unit (control unit 24). The weak learner generation unit (control unit 24) generates multiple weak learners from training data based on data using a neural network. The strong learner generation unit (control unit 24) generates a strong learner by combining multiple weak learners through boosting. The trained model generation unit (control unit 24) generates a trained model from a strongly trained model. This improves the prediction accuracy and robustness of the analysis.

[0130] In this embodiment, it is preferable that the learning unit (control unit 24) selects a portion of the multiple features contained in the data to generate training data. This makes the training data easier to handle, speeds up the series of analysis processes, and reduces overfitting.

[0131] In this embodiment, it is preferable for the learning unit (control unit 24) to reduce the number of dimensions of multiple features included in the data and reduce the data to generate training data. This makes the training data easier to handle. Furthermore, it speeds up the series of analysis processes. In addition, it reduces overfitting.

[0132] In this embodiment, it is preferable that the learning unit (control unit 24) removes outliers from the data to generate training data. This improves prediction accuracy and robustness.

[0133] In this embodiment, it is preferable that the data include data on the optical properties derived from a luminescent probe whose luminescence behavior changes upon interaction with the subject. This allows for accurate data on the target substance contained in the subject, improving the predictive accuracy of the analysis.

[0134] In this embodiment, the data includes data on the optical properties of the organic electroluminescent element. The organic electroluminescent element preferably has a light-emitting layer containing a light-emitting probe between the electrodes (between the transparent electrode 112 and the counter electrode layer 115). This allows for the acquisition of various data on the target substance contained in the sample, improving the predictive accuracy of the analysis.

[0135] In this embodiment, it is preferable that the luminescent probe has a main chain having one or more structural units including a sugar structure derived from a pentose or hexose and a phosphate ester bond attached to the sugar structure, and one or more chromophores or luminescent phosphodiphores attached to the sugar structure. This allows the luminescent probe to easily interact with target substances having a main chain structure similar to that of DNA, RNA, etc. As a result, data about the target substance can be easily and in detail, and the predictive accuracy of the analysis can be improved.

[0136] In this embodiment, the analysis system 100 comprises an analysis device 20 and a measuring device 10 that measures optical properties whose behavior changes in interaction with the subject.

[0137] In this embodiment, the analysis method performed by the analysis device 20 includes acquisition steps S2 and S5, a learning step S3, and an analysis step S6. Acquisition steps S2 and S5 acquire data on optical properties whose behavior changes in interaction with the subject. Learning step S3 generates a trained model for known subjects based on the data acquired in acquisition step S2. Analysis step S6 inputs the data acquired in acquisition step S5 for an unknown subject into the trained model generated by learning step S3 and outputs the analysis result for the unknown subject. Learning step S3 includes a weak learner generation step S37, a strong learner generation step S38, and a trained model generation step S39. Weak learner generation step S37 generates multiple weak learners from data-based training data using a neural network. Strong learner generation step S38 generates a strong learner by combining multiple weak learners through boosting. Trained model generation step S39 generates a trained model from the strong learner. This improves the prediction accuracy and robustness of the analysis.

[0138] In this embodiment, the program causes the computer of the analysis device 20 to function as an acquisition unit (control unit 24), a learning unit (control unit 24), and an analysis unit (control unit 24). The acquisition unit (control unit 24) acquires data on optical properties whose behavior changes in interaction with the subject. The learning unit (control unit 24) generates a trained model for known subjects based on the data acquired by the acquisition unit. The analysis unit (control unit 24) inputs the data acquired by the acquisition unit for an unknown subject into the trained model generated by the learning unit and outputs the analysis results for the unknown subject. The learning unit (control unit 24) generates multiple weak learners from training data based on data using a neural network. The learning unit (control unit 24) generates a strong learner by combining multiple weak learners through boosting. The learning unit (control unit 24) generates a trained model from the strong learner. This improves the prediction accuracy and robustness of the analysis.

[0139] Furthermore, the detailed configuration and operation of each device constituting the analysis apparatus can also be modified as appropriate, without departing from the spirit of this disclosure.

[0140] The following will provide specific examples, but this disclosure is not limited to these.

[0141] <Data Acquisition> Data was acquired using organic EL elements (OLED sensors) and luminescent dye molecules (luminescent DNA sensors) for 18 types of commercially available milk (Milk 1 to Milk 18), each used as a test subject.

[0142] [Data acquisition using an OLED sensor] (1) Fabrication of a blue OLED sensor A substrate was prepared on a 30 mm x 30 mm x 0.7 mm glass substrate with a 100 nm thick indium tin oxide (ITO) film deposited on it. After patterning the substrate, the substrate was ultrasonically cleaned with isopropyl alcohol and dried with dry nitrogen gas. The substrate was then cleaned with UV ozone for 5 minutes to obtain a transparent support substrate with an ITO transparent electrode (anode) attached.

[0143] An insulating polymer solution was obtained by diluting polystyrene (manufactured by ACROS ORGANICS Co., Ltd., molecular weight = 260,000) with n-propyl acetate as the solvent to 1.0% by mass. The insulating polymer solution was applied to the anode by spin coating at 500 rpm for 30 seconds. The coating film was then dried at 120°C for 30 minutes to obtain a laminate with a receiving layer 50 nm thick.

[0144] Commercially available milk 1 was used as the test subject. The laminate obtained above was immersed in a petri dish containing milk 1 for 1 minute, then the liquid was removed with an air gun and the laminate was dried.

[0145] n-propyl acetate was used as the solvent, and a luminescent probe (blue phosphorescent compound Ep1, described below) was mixed with the solvent at a concentration of 10 mg / mL. The resulting mixture was heated with ultrasound for 30 minutes, and then filtered through a 0.2 μm filter to remove aggregated components, thereby preparing a luminescent ink.

[0146]

[0147] The obtained luminescent ink was dropped onto the receiving layer of the laminate using an inkjet method under the following conditions to form a luminescent layer (detection region). The laminate was dried at 120°C for 30 minutes to evaporate the solvent.

[0148] The conditions for the inkjet method were as follows: The inkjet device used was the "IJCS-1" (manufactured by Konica Minolta, Inc.), and the inkjet head used was the "KM512" (manufactured by Konica Minolta, Inc.). The number of ejection shots was 2, the distance between the ejection nozzles from the head was 140 μm pitch, and printing was performed at a head scan speed of 90 mm / sec.

[0149] The laminate with the detection region formed was attached to the vacuum deposition apparatus. The vacuum chamber was 4 x 10 －４ The pressure was reduced to Pa, and an electron injection layer and a cathode were formed on the detection region of the laminate under the following conditions to obtain a blue OLED sensor A. The electron injection layer was formed by depositing potassium fluoride at a deposition rate of 0.1 Å / sec to a thickness of 2.0 nm. The cathode was formed by depositing aluminum at a deposition rate of 4 Å / sec to a thickness of 100 nm.

[0150] The resulting blue OLED sensor had four detection areas (2 x 2 mm) within a 30 x 30 mm area, with the detection area (inkjet area) being the same size as the electrode. The polystyrene receiving layer is insoluble in the milk sample. Therefore, when the laminate was immersed in the sample and dried, trace amounts of solid components of the milk sample were present on the surface of the polystyrene receiving layer. Subsequently, when the luminescent ink was applied to the receiving layer using an inkjet method, the solvent of the luminescent ink dissolved the polystyrene. The dissolved polystyrene, luminescent probe, and sample reached the lower electrode while forming a dispersed state. As a result, a detection area was formed in which the luminescent probe and sample were dispersed at the molecular level.

[0151] In this embodiment, 24 blue OLED sensors were prepared for each type of sample (Milk 1 to Milk 18), resulting in 4 x 24 = 96 measurement points for each sample. The anode and cathode were configured to allow voltage application via wiring.

[0152] (2) Preparation of green and red OLED sensors A green OLED sensor and a red OLED sensor were prepared using the same procedure as for the preparation of the blue OLED sensor, except that the light-emitting probe was changed to compound 1 (Green) or compound 2 (RED).

[0153]

[0154] (3) Data acquisition For the three types of OLED sensors (blue, green, and red) that were prepared, the spectral radiance spectrum [W・sr －１ ・m －２ nm －１ ] and the current density [mA / cm²] for each drive voltage ２ The following measurements were taken: The driving voltage was set in 1V increments from 1V to 10V, and the current at each voltage was measured. A spectroradiometer "CS-2000" (manufactured by Konica Minolta, Inc.) was used to measure brightness. A "6243 DC VOLTAGE CURRENT SOURCE / MONITOR" (manufactured by ADCMT Corporation) was used to measure current.

[0155] Using OLED sensors containing B (blue dopant), G (green dopant), and R (red dopant), 51 spectral data points were recorded at 5 nm intervals from 450 to 700 nm. The spectral data was normalized by setting the value at the wavelength with the highest radiance to 1.

[0156] [Data acquisition using a luminescent DNA sensor] (1) Synthesis of luminescent dye molecules 1 to 16: Unless otherwise specified, all reactions were carried out in oven-dried glassware under a nitrogen atmosphere. Chemical products not otherwise specified were purchased from Aldrich Co., Ltd., Tokyo Chemical Industry Co., Ltd., or Kanto Chemical Co., Ltd., and used as is without purification.

[0157] (1.1) Monomer 1 having a main chain containing a phosphate ester and a luminescent phosphodiol bonded to the main chain was synthesized via intermediates 1 to 6 of the synthesis of monomer 1.

[0158] Synthesized intermediate 1: Thymidine (15.0 g, 61.9 mmol) and imidazole (16.9 g, 248 mmol) were dissolved in dimethylformamide (124 mL). To the resulting solution, tert-butyldimethylsilyl chloride (19.6 g, 130 mmol) was added and the mixture was stirred at room temperature for 17 hours. Water was added to the resulting reaction mixture, and liquid-liquid extraction was performed with ethyl acetate. The resulting organic phase was dried over magnesium sulfate, and the solvent was removed by distillation to obtain the target intermediate 1 as a colorless solid (28.3 g, yield 97%).

[0159] - Synthesis of intermediate 2: Intermediate 1 (28.3 g, 60.1 mmol) and ammonium sulfate (12.7 g, 96.2 mmol) were dissolved in hexamethyldisilazane (314 mL, 1.50 mol). The resulting solution was heated under reflux for 3 hours. Water was added to the resulting reaction mixture, and liquid-liquid extraction was performed with ethyl acetate. The resulting organic phase was dried over magnesium sulfate, and the solvent was removed by distillation. The resulting crude product was purified by silica gel column chromatography to obtain the target intermediate 2 as a brown liquid (13.2 g, yield 64%).

[0160] - Synthesis of Intermediate 4: A mixture of intermediate 2 (10.1 g, 29.3 mmol), 1-bromopyrene (8.24 g, 29.3 mmol), tris(dibenzylideneacetone)dipalladium(0) (671 mg, 733 μmol), tritert-butylphosphonium tetrafluoroborate (850 mg, 2.93 mmol), dicyclohexylmethylamine (9.35 mL, 44.0 mmol), and 1,4-dioxane (100 mL) was heated at 90°C for 1 hour. Water was added to the heated mixture to stop the reaction, and liquid-liquid extraction was performed with ethyl acetate. The resulting organic phase was dried over magnesium sulfate, and the solvent was removed by distillation. The crude product containing the obtained intermediate 3 was used directly in the next reaction.

[0161] To the crude product containing intermediate 3, tetrahydrofuran (100 mL), 1 M tetrabutylammonium fluoride / tetrahydrofuran solution (117 mL, 117 mmol), and acetic acid (6.74 mL, 117 mmol) were added, and the mixture was stirred at 40°C for 2 hours. Water was added to the resulting reaction product to stop the reaction, and liquid-liquid extraction was performed with ethyl acetate. The resulting organic phase was dried over magnesium sulfate, and the solvent was removed by distillation. The resulting crude product was purified by silica gel column chromatography to obtain the target intermediate 4 as a light brown solid (5.87 g, yield 63%).

[0162] - Synthesis of Intermediate 5: Sodium triacetylborate (11.8 g, 55.8 mmol) and acetic acid (7.87 mL, 138 mmol) were dissolved in acetonitrile (93 mL). The resulting solution was cooled to 0°C, and a solution of intermediate 4 (5.87 g, 18.6 mmol) dissolved in tetrahydrofuran (62 mL) was added dropwise. After the addition was complete, the resulting reaction mixture was heated to room temperature, stirred for 15 minutes, and then water was added to stop the reaction. Liquid-liquid extraction was performed with ethyl acetate. The resulting organic phase was dried over magnesium sulfate, and the solvent was removed by distillation. The resulting crude product was purified by silica gel column chromatography and reverse-phase chromatography to obtain the target intermediate 5 as a colorless solid (3.44 g, yield 58%).

[0163] - Synthesis of intermediate 6: Intermediate 5 (3.44 g, 10.8 mmol), 4,4'-dimethoxytrityl chloride (4.40 g, 13.0 mmol), ethyl diisopropylamine (2.82 mL, 16.2 mmol), and anhydrous pyridine (54 mL) were mixed. The resulting mixture was stirred at room temperature for 4 hours. Methanol was added to the resulting reaction solution to stop the reaction, and the solvent was removed by distillation. The resulting crude product was purified by silica gel column chromatography to obtain the target intermediate 6 as a colorless viscous solid (5.71 g, yield 85%).

[0164] Synthetic intermediate 6 of monomer 1 (5.71 g, 9.20 mmol) was mixed with ethyl diisopropylamine (6.42 mL, 36.8 mmol) and anhydrous dichloromethane (92 mL). 2-Cyanoethyldiisopropyl chlorophosphoramidito (3.08 mL, 13.8 mmol) was added dropwise to the resulting mixture at 0°C. The resulting reaction mixture was heated to room temperature and stirred for 3 hours, after which the solvent was removed by distillation. The crude product was purified by silica gel column chromatography to obtain the target monomer 1 as a colorless solid (4.64 g, yield 61%).

[0165] (1.2) Preparation of Monomer 2 Monomer 2 was purchased from Glen Research Co., Ltd. (Sterling, Virginia) as a reagent with the structure shown below.

[0166] (1.3) Synthesis of Luminescent Dye Molecules 1-16 Following the standard method, 16 oligonucleotides (Seq1-16) with mixed sequences of monomer 1 and monomer 2 were synthesized as shown in Table I below. DNA synthesis reagents were purchased from Glen Research, Inc. (Sterling, Virginia). All oligonucleotides were synthesized using the DNA / RNA synthesizer "NTS T-series" (manufactured by Nippon Techno Service Co., Ltd.) using the standard protocol for phosphoramidite-based coupling methods.

[0167] Each luminescent dye molecule carrier obtained by automated synthesis was reacted with ammonium water at room temperature for 2 hours. The particulate carrier was excised, and the solvent was dried using a centrifugal dryer. Ultrapure water was added to the dried particulate carrier to obtain first components 1 to 16 containing each luminescent dye molecule 1 to 16. It was confirmed that these luminescent dye molecules 1 to 16 emit fluorescence and excimer emission when exposed to specific excitation light (light with a wavelength of 350 nm).

[0168]

[0169] (2) Data Acquisition (2.1) Arrangement of Luminescent Dye Molecules A 96-well microplate was prepared in which 12 rows of 8 wells with an opening diameter of 7 mm were arranged at intervals of 9 mm. 100 μL each of the above-mentioned luminescent dye molecules 1 to 16 were placed in the 96-well microplate to form multiple reaction fields. An automated dispensing device, "NichiMart CUBE" (manufactured by NICHIRYO Co., Ltd.), was used to arrange the luminescent dye molecules.

[0170] (2.2) Acquisition of first signal data A 96-well microplate containing the above-mentioned luminescent dye molecules was irradiated with excitation light (wavelength 350 nm), and the resulting fluorescence spectra were acquired as first signal data.

[0171] (2.3) Placement of the samples After acquiring the first signal data described above, 20 μL of milk 1 was placed in each of the reaction fields containing the luminescent dye molecules 1 to 16 in the 96-well microplate. For milk 2 to milk 18, the reaction fields were formed using the same procedure as for milk 1, except that the type of milk was changed. An automated dispensing device, "NichiMart CUBE" (manufactured by NICHIRYO Co., Ltd.), was used to place the milk.

[0172] (2.4) Acquisition of second signal data A 96-well microplate containing the above-mentioned luminescent dye molecules and milk as the test subject was irradiated with excitation light (wavelength 350 nm), and the resulting fluorescence spectra were acquired as second signal data.

[0173] <Human Sensory Evaluation> 100 panelists conducted a sensory evaluation of Milk 1 through Milk 18. The sensory evaluation was based on sweetness, rated on a 5-point scale. The score was expressed as the arithmetic mean, calculated by dividing the sum of each panelist's score by the number of panelists.

[0174] [Test A] Data was acquired for milk 1 through milk 18 using an OLED sensor. Data was acquired twice, on day 1 and day 2. On day 1, data was acquired for all milk 1 through milk 18, and machine learning and validation 1 were performed using the acquired data. On day 2, data was acquired for milk 12 through milk 18, and validation 2 was performed using the acquired data.

[0175] <Machine Learning> Training data was generated from 2 / 3 of the data acquired on the first day, i.e., data from 12 types of milk (Milk 1-12). One-third of the data acquired on the first day, i.e., data from 6 types of milk (Milk 13-18), was used as validation data.

[0176] First, data acquired from OLED sensors for 12 types of milk (Milk 1-12) were used as explanatory variables, and the results of sensory evaluation were used as the dependent variable. As shown in Table II below, various preprocessing steps were performed on the acquired data to generate training data, and then trained models were generated using various algorithms.

[0177] <Verification 1: Confirmation of Prediction Accuracy> Using the data for Verification 1 from six types of milk (milk 13-18), we confirmed whether overfitting had occurred (prediction accuracy). The data acquired from the OLED sensor in the Verification 1 data was input into the trained model, the predicted values ​​were output, and the R-squared value was calculated.

[0178] <Verification 2: Confirmation of Robustness (Reproducibility)> Robustness (reproducibility) was confirmed using data for Verification 2 from six types of milk (Milk 13-18). The measurement unit, including the OLED sensor, is susceptible to the influence of the installation environment. Therefore, in Verification 2, we confirmed whether the trained model, which was generated based on data acquired on the first day, could accurately output predicted values ​​even when data acquired on the second day, when the installation environment of the measurement unit had changed slightly, was input. The data acquired by the OLED sensor in the Verification 2 data was input into the trained model, predicted values ​​were output, and the R-squared value was calculated.

[0179] [Test B] Data obtained from a luminescent DNA sensor was used as explanatory variables, and the test was performed using the same procedure as in Test A.

[0180] [Test C] For 18 types of milk (Milk 1-18), the combination of milk types used for training data and validation data was changed, and all combinations were tested using the same procedure as in Test A.

[0181] Table II below shows the details of the learning process and the results of the analysis process. The relationship between the predicted value and the correct value (the actual sensory evaluation result) is expressed by the coefficient of determination R-squared. The closer the R-squared value is to 1, the closer the predicted value is to the correct value.

[0182] In the preprocessing steps shown in Table II, the preprocessing involved standardizing the data and removing outliers. A "-" in Table II indicates that no intermediate processing was performed. As mentioned earlier, "neural boosting" involves generating weak learners using a neural network and combining these weak learners through boosting to create a strong learner.

[0183] The R-squared values ​​in Table II represent the results of tests A to C. For example, in the learning process of Example 1, tests A to C all had R-squared values ​​of 0.9 or higher; in other words, the maximum R-squared value in tests A to C was 0.9. Similarly, in the analysis process of Example 1, verification 1 showed that tests A to C all had R-squared values ​​within the range of 0.7 to 0.9; in other words, the maximum R-squared value in tests A to C was 0.9. Verification 2 showed that tests A to C all had R-squared values ​​of 0.4 or less; in other words, the maximum R-squared value in tests A to C was 0.4.

[0184]

[0185] Regarding the R-squared value in Verification 1, the comparative example had a value of 0.7 or less, while the embodiment had a value within the range of 0.7 to 0.9, indicating improved prediction accuracy in the embodiment. Furthermore, regarding the R-squared value in Verification 2, the comparative example had a maximum value of 0.3, while the embodiment had a maximum value of 0.4 or more, indicating improved robustness in the embodiment of this model.

[0186] Although Table II shows PLS regression as the comparative algorithm, the R-squared values ​​in Verification 2 were 0.3 or less for other commonly used algorithms as well.

[0187] We input training data into the pre-trained model and created graph G1 showing the relationship between the outputted predicted values ​​and the actual sensory evaluation results. We input validation 1 data into the pre-trained model and created graph G2 showing the relationship between the outputted predicted values ​​and the actual sensory evaluation results. We input validation 2 data into the pre-trained model and created graph G3 showing the relationship between the outputted predicted values ​​and the actual sensory evaluation results. Both graphs G2 and G3 show that the predicted values ​​and the ground truth values ​​(actual sensory evaluation results) are relatively close, indicating that prediction accuracy and robustness can be improved.

[0188] A comparison of the comparative examples and the examples shows that using the neural boosting algorithm described above can improve the accuracy and robustness of predictions. Furthermore, it can be seen that robustness can be improved by performing intermediate processing, particularly dimensionality reduction, as a preprocessing step for generating training data during machine learning.

[0189] This disclosure can improve the predictive accuracy and robustness of analysis in the subject.

[0190] 10 Measuring device 11 Measuring unit 12 Communication unit 13 Control unit 20 Analysis device 21 Display unit 22 Operation unit 23 Communication unit 24 Control unit 100 Analysis system 111 Transparent substrate 112 Transparent electrode 113 Hole transport layer 114 Receptor layer 115 Counter electrode layer 121 Subject layer 122 Light-emitting layer 300 Neural network 300a Input layer 300b Feature extraction layer 300c Output layer

Claims

An acquisition unit that acquires data on optical properties whose behavior changes in interaction with the subject, A learning unit that generates a trained model based on the data acquired by the acquisition unit for a known subject, The system comprises: an analysis unit that inputs the data acquired by the acquisition unit for an unknown subject into a trained model generated by the learning unit, and outputs the analysis results for the unknown subject; The aforementioned learning unit, A weak learner generation unit that generates multiple weak learners from training data based on the aforementioned data using a neural network, A strong learner generation unit that generates a strong learner by combining multiple weak learners through boosting, An analysis apparatus comprising a trained model generation unit that generates a trained model from the aforementioned strong learner. 　 The analysis apparatus according to claim 1, wherein the learning unit selects a portion of a plurality of features contained in the data to generate training data. 　 The analysis apparatus according to claim 1, wherein the learning unit reduces the number of dimensions of multiple features included in the data and reduces the data to generate training data. 　 The analysis apparatus according to claim 1, wherein the learning unit generates training data by removing outliers included in the data. 　 The analysis apparatus according to claim 1, wherein the data includes data on optical properties derived from a light-emitting probe whose light emission behavior changes upon interaction with the subject. 　 The aforementioned data includes data on the optical properties of an organic electroluminescent element, The analytical apparatus according to claim 5, wherein the organic electroluminescent element has a light-emitting layer containing the light-emitting probe between electrodes. 　 The light-emitting probe is A main chain having one or more structural units including a sugar structure derived from a pentose or hexose, and a phosphate ester bond attached to the sugar structure, The analytical apparatus according to claim 5, comprising one or more chromophores or luminescent phosphonates bonded to the sugar structure. 　 The analysis apparatus described in claim 1, An analysis system comprising a measuring device for measuring optical properties whose behavior changes in response to interaction with a subject. 　 In the analysis method performed by the analysis device, An acquisition process to obtain data on optical properties whose behavior changes in interaction with the subject, A learning step that generates a trained model based on the data acquired in the acquisition step for a known subject, The analysis step includes inputting the data acquired in the acquisition step for an unknown subject into a trained model generated in the training step, and outputting the analysis results for the unknown subject. The aforementioned learning process is, A weak learner generation step that generates multiple weak learners from training data based on the aforementioned data using a neural network, A strong learner generation process that generates a strong learner by combining multiple weak learners through boosting, An analysis method comprising a pre-trained model generation step of generating a pre-trained model from the aforementioned strong learner. 　 The computer of the analysis device, An acquisition unit that acquires data on optical properties whose behavior changes in interaction with the subject, A learning unit that generates a trained model based on the data acquired by the acquisition unit for a known subject, A program that functions as an analysis unit that inputs the data acquired by the acquisition unit for an unknown subject into a trained model generated by the learning unit, and outputs analysis results for the unknown subject, The aforementioned learning unit, Using a neural network, multiple weak learners are generated from training data based on the aforementioned data. By boosting, multiple weak learners are combined to generate a strong learner. A program that generates a trained model from the aforementioned strong learner.

Images