Spectral estimation device, spectral estimation system, computer program, and spectral estimation method

Abstract

To provide a spectrum estimation device, a spectrum estimation system, a computer program, and a spectrum estimation method capable of improving the accuracy of estimating peak intensities.SOLUTION: A spectrum estimation device includes an observation spectrum signal acquiring unit for acquiring an observation spectrum signal on the basis of a detection signal on the analytical species of a target sample acquired using an observation filter, and an estimation unit for estimating the true spectrum signal on the basis of the sparsity on the true spectral signal of the analyzed species using the observed spectral signal as an objective variable, and an observation matrix, which has a bandwidth equal to or wider than the spectral bin width and whose elements are observation functions that represent the characteristics of the observation filter, as an explanatory variable.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a spectrum estimation device, a spectrum estimation system, a computer program, and a spectrum estimation method. [Background technology]

[0002] Spectral analysis is used as a qualitative or quantitative analysis method in various fields, such as industrial materials, food, the environment, and pharmaceuticals. Patent Document 1 discloses a mass spectrometer that performs mass analysis of components in a sample by ionizing a sample to be analyzed, separating the generated ions according to their mass-to-charge ratio using a quadrupole filter in an analysis chamber, and detecting only ions having a specific mass-to-charge ratio using a detector. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-64561 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the analyzer of Patent Document 1 performs observations using a narrow measurement window that allows only ions with a specific mass-to-charge ratio to pass through. While this has the advantage of enabling high-resolution observations, it also reduces the observed signal intensity, resulting in reduced accuracy and sensitivity of the peak intensities in the mass spectrum obtained as a result of the measurement.

[0005] The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a spectrum estimation device, a spectrum estimation system, a computer program, and a spectrum estimation method that can improve the estimation accuracy and measurement sensitivity of peak intensity. [Means for solving the problem]

[0006] The present application includes multiple means for solving the above-described problems. To cite one example, the spectral estimation device of the present embodiment includes an observed spectral signal calculation unit that calculates an observed spectral signal based on a detection signal related to an analyte of a target sample, acquired using an observation filter; a memory unit that stores an observation matrix having a bandwidth equal to or greater than the spectral bin width and whose elements are observation functions that represent the characteristics of the observation filter; and an estimation unit that uses the observed spectral signal as a dependent variable and the observation matrix read out from the memory unit as an explanatory variable, and estimates the true spectral signal based on prior knowledge including sparsity related to the true spectral signal of the analyte. [Effects of the Invention]

[0007] According to the present invention, it is possible to improve the estimation accuracy and measurement sensitivity of the peak intensity of a mass spectrum. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 illustrates an example of the configuration of a spectrum estimation system. [Figure 2] FIG. 1 illustrates an example of the configuration of a spectrum estimation device. [Figure 3] FIG. 2 is a diagram illustrating an example of an observation system model according to the present embodiment. [Figure 4] FIG. 10 is a diagram showing the movement of the observation function on the m / z axis. [Figure 5] FIG. 10 is a diagram illustrating an example of an observation system model of a comparative example. [Figure 6] FIG. 10 is a diagram showing the movement of the observation function on the m / z axis in the comparative example. [Figure 7] FIG. 1 is a diagram showing an example of an observed spectral signal and a genuine spectral signal (standard spectral signal) of an analyte of a standard sample. [Figure 8] FIG. 10 is a diagram showing an example of an observation function estimated using a standard sample. [Figure 9] FIG. 2 is a diagram illustrating a first example of the configuration of a spectrum estimation unit. [Figure 10] FIG. 10 is a diagram illustrating a second example of the configuration of the spectrum estimation unit. [Figure 11] FIG. 2 is a diagram showing an example of a mass spectrum. [Figure 12] FIG. 10 is a diagram showing a first example of an evaluation result of an estimation method. [Figure 13] FIG. 10 is a diagram showing a second example of the evaluation results of the estimation method. [Figure 14] FIG. 10 is a diagram showing a third example of the evaluation results of the estimation method. [Figure 15] FIG. 10 is a flowchart showing the processing procedure for constructing an observation matrix. [Figure 16] FIG. 10 is a flowchart showing a processing procedure for estimating a true spectral signal. DETAILED DESCRIPTION OF THE INVENTION

[0009] The present invention will be described below with reference to the drawings showing embodiments thereof. FIG. 1 is a diagram showing an example of the configuration of a spectrum estimation system. The spectrum estimation system includes a spectrum measurement device 10 and a spectrum estimation device 50. The spectrum measurement device 10 and the spectrum estimation device 50 may be configured as an integrated analysis device. In the following, a mass spectrometry system will be described as an example of the spectrum estimation system, but the spectrum estimation system is not limited to a mass spectrometry system.

[0010] The spectrum measuring device 10 includes a sample injection unit 11, a standard sample supply unit 12, a flow path switching valve 13, an ionization unit 14, an ion optical system 15, a mass separation unit 16, a detection unit 17, a voltage generation unit 18, an operation unit 19, and a control unit 20.

[0011] A target sample (also referred to as "sample") is injected into the sample injection unit 11. Target samples are used in various fields, such as industrial materials, food, the environment, and pharmaceuticals, and include, but are not limited to, low-molecular-weight compounds, proteins, amino acids, synthetic polymers, and inorganic ions. The sample injected into the sample injection unit 11 is introduced into the ionization unit 14 via a column (flow path) by a medium (liquid or gas). As the sample moves through the column, it is separated into analytes. The analytes refer to the components to be analyzed, the analytical sample, or the components targeted for analysis. A flow path switching valve 13 is provided midway through the column, and selectively switches between the analytes of the sample injected into the sample injection unit 11 and the analytes of the standard sample supplied to the standard sample supply unit 12, and introduces them into the ionization unit 14. The standard sample is a sample for which the peak positions of the spectrum (mass spectrum) related to the mass-to-charge ratio (m / z) of the ionized analytes are known in advance. Note that the standard sample supply unit 12 and the flow path switching valve 13 are optional. That is, both the target sample and the standard sample may be introduced into the ionization unit 14 from the sample injection unit 11 .

[0012] The ionization section 14 ionizes the analyte, and the ion optical system 15 focuses the ionized analyte (ions) and introduces them into the mass separation section 16. The ion optical system 15 may also remove non-ionic particles.

[0013] The mass separator 16 functions as an observation filter. The mass separator 16 performs mass separation using a quadrupole electric field formed by, for example, four cylindrical (hyperboloidal inside) metal rods arranged at equal distances from the central axis. The quadrupole electric field can be generated by a DC voltage and a high-frequency AC voltage supplied from the voltage generator 18. The voltage values of the DC voltage and the high-frequency AC voltage can be adjusted by the control unit 20 as voltage parameters. By changing the ratio between the DC voltage and the high-frequency AC voltage, the range of mass-to-charge ratios of ions that can pass through the mass separator 16 (observation filter) (the bandwidth of the observation filter) can be expanded or contracted. For example, a first mode, a second mode, and a third mode can be provided as measurement modes, with the first mode, the second mode, and the third mode being selected in order of narrowest bandwidth (the width of the measurement window) of the observation filter.

[0014] Furthermore, by maintaining a constant ratio between the DC voltage and the high-frequency AC voltage and varying the DC voltage and the high-frequency AC voltage, the mass-to-charge ratio of ions that can pass through the mass separation unit 16 can be varied, and the mass-to-charge ratio of ions that pass through the observation filter can be scanned between a required lower limit and upper limit of the mass-to-charge ratio (on the m / z axis). That is, if the index of the position on the m / z axis of the observation filter is i, then i = 1 to M, and each time the index of the position is moved by one, the number (amount) of ions is detected (observed) by the detection unit 17. M is the number of observations.

[0015] The operation unit 19 can be used to set voltage parameters, operating conditions of the spectrum measuring device 10, and the like.

[0016] The detector 17 outputs to the spectrum estimation device 50 a detection signal corresponding to the number (amount) of ions that have passed through the mass separator 16 and reached the detector 17.

[0017] Although the mass separator 16 has been described as a quadrupole type, it is not limited to this. The mass separator 16 may be, for example, an ion trap type or a time-of-flight type. The ion trap type includes a pair of ring electrodes arranged opposite each other and a pair of end cap electrodes arranged on either side of the ring electrodes. Ions are introduced through one end cap electrode, and the trapped ions are output to the ion detector from the other end cap electrode. The time-of-flight type accelerates ions using an ion acceleration electrode and introduces them into a flight tube (field-free region), where they are detected by the ion detector. The mass-to-charge ratio of the ions can be identified based on the time it takes for the ions to reach the ion detector.

[0018] FIG. 2 shows an example of the configuration of a spectrum estimation device 50. The spectrum estimation device 50 may be configured as an information processing device such as a personal computer. The spectrum estimation device 50 includes a control unit 51 that controls the entire device, an observed spectrum signal calculation unit 52, an observation matrix calculation unit 53, a storage unit 54, a spectrum estimation unit 55, a display processing unit 56, and a standard spectrum signal calculation unit 57. A display device 60 is connected to the spectrum estimation device 50 via a wired or wireless connection. The display device 60 may be incorporated into the spectrum estimation device 50.

[0019] The control unit 51 can be configured with a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The storage unit 54 can be configured with a hard disk or a semiconductor memory, etc., and can store required information such as observation matrix data and a learning model, which will be described later.

[0020] The observed spectral signal calculation unit 52 calculates an observed spectral signal based on a detection signal related to the analyte of the target sample, acquired using the observation filter. Specifically, the observed spectral signal calculation unit 52 acquires the detection signal output by the spectrometer 10, calculates the signal intensity on the mass-to-charge ratio (m / z) axis, and obtains an observed spectral signal. In this process, the mass-to-charge ratio axis is appropriately divided into several intervals to discretize it, and an observed spectral signal is obtained for each interval (called a spectral bin) obtained by the discretization. The observation range (measurement range) of the mass-to-charge ratio can be, for example, but is not limited to, m / z = 300 to 1600. The width of the spectral bin can be, for example, but is not limited to, m / z = 0.1. Note that if the observed spectral signal can be acquired from an external analyzer, the control unit 51 can function as an observed spectral signal acquisition unit and may acquire the observed spectral signal. In this case, the observed spectral signal calculation unit 52 is not an essential component.

[0021] Although the mass-to-charge ratio of the standard material is known, the shape of the true spectral signal can change depending on the measurement conditions (method and apparatus). Therefore, it is necessary to obtain the true spectral signal (standard spectral signal) of the standard material under the measurement conditions used by some operation. Here, the true spectral signal is an ideal spectral signal that is assumed to be obtained when measurement is performed using a hypothetical sufficiently narrow measurement window. Therefore, the standard spectral signal calculation unit 57 calculates the standard spectral signal by performing processing such as peak detection on the observed spectral signal in the mode with the narrowest bandwidth of the observation filter. The obtained standard spectral signal can be used together with the observed spectral signal for the standard material to estimate the observation function.

[0022] The observation system model of the spectrum estimation system of this embodiment can be expressed by equation (1).

[0023]

number

[0024] In equation (1), y is the observed spectral signal calculated by the observed spectral signal calculation unit 52, and can be expressed by equation (2). In equation (2), yi is the value of the observed spectral signal obtained during the i-th observation, where i=1 to M, and M represents the number of observations. In other words, yi represents the value of the observed spectral signal calculated based on the detection signal detected when the index of the position on the m / z axis of the observation filter is i.

[0025] In equation (1), β is a spectral signal (true spectral signal) that is assumed to be obtained when the analyte of the target sample is observed using a hypothetical sufficiently narrow measurement window in the spectrometer 10, and can be expressed by equation (3). In equation (3), β1, β2, ..., βN are the values of the true spectral signal in each of the spectral bins 1, ..., N on the m / z axis. Here, 1 to N are the indexes of the spectral bins, and N is the total number of bins. Furthermore, it is assumed that the true spectral signal β satisfies equation (4). The left side of equation (4) is the number of non-zero elements of the true spectral signal β.

[0026] In equation (1), ε can be expressed by equation (5), and represents the measurement noise for i=1 to M.

[0027] In equation (1), X is an observation matrix, which is a matrix with M rows and N columns and can be expressed by equation (6). Each row of the observation matrix X (x i j ), j=1 to N represent the characteristics of the observation filter. The observation matrix X has a bandwidth equal to or greater than the spectral bin width, and is a matrix whose elements are the values of the observation function that represents the characteristics of the observation filter. The observation matrix X is expressed as (x i j )i=1 to M, j=1 to N. i j ) represents the value of the observation function for each index i=1 to M, j=1 to N. The observation matrix X can be stored in the storage unit 54. If the observation matrix X can be acquired from an external device, for example, the control unit 51 may acquire the observation matrix X from the external device and use the acquired observation matrix X.

[0028] The observation system model expressed by equation (1) will now be specifically explained.

[0029] FIG. 3 is a diagram showing an example of an observation system model according to this embodiment. In the example of FIG. 3, for convenience, the observation matrix X is a 5-row by 7-column matrix. That is, the number of observations M is 5, and the total number of bins N is 7. The elements of the first row of the observation matrix X are (f1, f2, f3, 0, 0, 0, 0), and these elements correspond to the observation function that represents the characteristics of the observation filter. Note that f1, f2, and f3 are not zero. That is, the observation function is a function whose function value is f1, f2, f3, 0, 0, 0, 0 for each of the indexes j = 1 to 7. The elements of the second row of the observation matrix X are (0, f1, f2, f3, 0, 0, 0), which is the observation function in the first row shifted by one index j. This corresponds to moving the observation filter by one spectral bin on the m / z axis.

[0030] The true spectral signal β is a seven-dimensional vector, and the true spectral signal values for each spectral bin are represented by β1, β2, β3, β4, β5, β6, and β7. In this case, the observed spectral signal y based on equation (1) is a five-dimensional vector. The elements of the observed spectral signal y are y1 = f1·β1 + f2·β2 + f3·β3 + ε1, y2 = f1·β2 + f2·β3 + f3·β4 + ε2, y3 = f1·β3 + f2·β4 + f3·β5 + ε3, y4 = f1·β4 + f2·β5 + f3·β6 + ε4, and y5 = f1·β5 + f2·β6 + f3·β7 + ε5. y1 to y5 are the observed signal values observed each time the observation filter (observation function) is moved by one spectral bin on the m / z axis. ε1 to ε5 represent noise.

[0031] Figure 4 shows the movement of the observation function on the m / z axis. The observation function of this embodiment has a bandwidth equal to or greater than the spectral bin width. In the example of Figure 3, the observation function has non-zero function values f1, f2, and f3 across three consecutive spectral bins. That is, the observation filter passes only ionized analytes from the sample that have mass-to-charge ratios that span multiple consecutive spectral bins.

[0032] FIG. 5 shows an example of an observation system model of the comparative example. The observation system model of the comparative example can be expressed as y = Iβ + ε. The observation matrix I is a 7-row x 7-column matrix. The observation matrix I is expressed as a unit matrix. That is, in the comparative example, observation is performed using a narrow measurement window that allows only ions with a specific mass-to-charge ratio to pass. The observed spectrum signal y is a seven-dimensional vector. The elements of the observed spectrum signal y are y1 = β1 + ε1, y2 = β2 + ε2, y3 = β3 + ε3, y4 = β4 + ε4, y5 = β5 + ε5, y6 = β6 + ε6, and y7 = β7 + ε7. ε1 to ε7 represent noise. In the comparative example, because observation is performed using a narrow measurement window, the observable signal intensity decreases, resulting in reduced accuracy and measurement sensitivity of the peak intensity of the mass spectrum.

[0033] Figure 6 shows the movement of the observation function of the comparative example on the m / z axis. The observation function of the comparative example has a bandwidth equal to the spectral bin width (this can be considered high-resolution observation). The measurement window (observation function) of the comparative example allows ions whose mass-to-charge ratio falls within the spectral bin width to pass, but other ions cannot pass and are not observed. This reduces the observable signal intensity.

[0034] In contrast, in this embodiment, as shown in Figures 3 and 4, the measurement window is widened and observation signals are obtained by so-called low-resolution observation, thereby improving the measurement sensitivity of the mass spectrum. Furthermore, since the measurement window is widened during observation, information loss during observation can be reduced compared to the comparative example. For example, as shown in Figure 5, not only information on the true spectrum signal β1 but also information on the true spectrum signals β1, β2, and β3 can be obtained in a single observation, as shown in Figure 3.

[0035] Next, a method for constructing the observation matrix will be described. Note that if the properties of the measurement system are known and the observation matrix X is known in advance with sufficient accuracy, there is no need to construct the observation matrix again.

[0036] In what follows, we assume that the properties of the observation function are constant throughout the entire observation process. We define the observation function as a discrete function whose function value is 0 outside the finite interval -J≦j≦J (measurement window). Let N be the total number of spectral bins. We assume that M (=N-2J) observations are made for the standard sample by moving the position of the measurement window by 1 for each spectral bin index. The observation matrix X in this case can be expressed by equation (7).

[0037]

number

[0038] Furthermore, each element of the observed spectral signal y can be expressed by equation (8). From equation (8), the observed spectral signal y can be expressed by equation (9). B is a matrix constructed as shown in equation (10) based on the true spectral signals (standard spectral signals) βi, i = 1 to N of the analyte of the standard sample. f is a window function (a vector constructed by limiting the observation function to the interval -J≦j≦J corresponding to the measurement window). The window function f can be expressed by equation (11).

[0039] By observing a standard sample whose spectrum is known (i.e., the above-mentioned B is known), the window function f can be estimated (calculated) from the obtained observation signal y and equation (9) using, for example, the least squares method.

[0040] 7 shows an example of the observed spectral signal and the true spectral signal (standard spectral signal) of the analyte of a standard sample. The standard spectral signal calculation unit 57 detects peaks from the observed spectral signal of the standard sample and calculates the standard spectral signal by examining the signal intensity. In addition, the observed spectral signal near the peak can be used to estimate the window function f.

[0041] As described above, the standard spectral signal calculation unit 57 can calculate a standard spectral signal based on the observed spectral signal y related to the analyte of the standard sample, which is acquired using the observation filter. If the standard spectral signal can be acquired from an external analyzer, the control unit 51 can function as a standard spectral signal acquisition unit and may acquire the standard spectral signal. In this case, the standard spectral signal calculation unit 57 is not necessarily a required component.

[0042] The observation matrix calculation unit 53 has a bandwidth equal to or greater than the spectral bin width and calculates a window function f that represents the characteristics of the observation filter so as to minimize the error between the observed spectral signal y and the product (Bf) of a matrix B constructed from the standard spectral signal and the window function f, and can construct the observation matrix X based on the result. The observation matrix X constructed by the observation matrix calculation unit 53 can be stored in the storage unit 54. Note that the construction of the observation matrix may be performed by a device other than the spectrum estimation device 50, and the observation matrix X constructed by that device may be stored in the storage unit 54. Alternatively, the window function f may be stored in the storage unit 54 instead of the observation matrix X, and the window function f may be acquired from the storage unit 54 when the observation matrix X is used to construct the observation matrix X.

[0043] Figure 8 shows an example of an observation function estimated using a standard sample. The horizontal axis indicates the spectral bin index, and the vertical axis indicates the function value. The first, second, and third modes are measurement modes that specify the width of the measurement window, and the bandwidth of the observation function (width of the measurement window) increases in the order of the first, second, and third modes. The bandwidth of the observation function is wider than the spectral bin width, allowing so-called low-resolution observation. Furthermore, as can be seen from Figure 8, the wider the bandwidth of the mode, the larger the function value (corresponding to the amount of ions detected).

[0044] To estimate the observation function, we simply use the observed spectrum signal of the standard sample within a mass-to-charge ratio (m / z) range where the signal-to-noise ratio of the observed spectrum signal is sufficiently large. In the example shown in Figure 8, we used the observed spectrum signal for 1210 ≤ m / z ≤ 1235 and performed the least-squares method with J = 50. However, the mass-to-charge ratio (m / z) range and J value are not limited to these values. The full widths at half maximum of the observation function in the first, second, and third modes are 0.6, 0.7, and 2.3 [m / z], respectively. This indicates that all measurement modes provide low-resolution observations with a spectral bin width of 0.1 [m / z]. Furthermore, the function value increases with the measurement window width, indicating that widening the measurement window improves the ion transmittance through the observation filter.

[0045] If the window function f cannot be considered constant throughout the entire observation process, the spectral region is divided into ranges within which the window function can be considered constant, and the same processing as described above is performed for each divided spectral region. The window functions estimated for each spectral region can then be integrated to construct the observation matrix X. Furthermore, because spectral observation of a standard sample is a process necessary for calibrating the measurement device, constructing the observation matrix does not require the addition of new hardware or processing to the measurement system.

[0046] In this embodiment, a low-resolution observation is performed by using a measurement window (observation filter, observation function) with a width equal to or greater than the spectral bin width, and spectral signals across multiple consecutive spectral bins are observed simultaneously. Because the observed spectral signal obtained by low-resolution observation is a mixture of multiple spectral signal components, post-processing is required to estimate the original true spectral signal. Because observation is performed with a wide measurement window, the number (amount) of detectable ions increases, improving the estimation accuracy and measurement sensitivity of mass spectrum peak intensities. Post-processing can also achieve a spectral resolution equal to or less than the measurement window width (bandwidth).

[0047] Next, a method for estimating a true spectral signal will be described.

[0048] The spectral estimation unit 55 reads the observation matrix from the storage unit 54, and estimates the true spectral signal β using the observed spectral signal y calculated by the observed spectral signal calculation unit 52 as the objective variable, the observation matrix X as the explanatory variable, and prior knowledge of the properties expected to be possessed by the true spectral signal of the analyte. Examples of prior knowledge include the sparsity of the true spectral signal of the analyte of the target sample. Sparsity (sparse property) refers to the fact that, in data that depends on multiple estimation parameters, only a few estimation parameters contribute to explaining the data. Regarding the true spectral signal of the target sample, this property can be expressed as the fact that, among the spectra spanning multiple consecutive spectral bins, only a small number of spectral bins have non-zero signal values. Prior knowledge including sparsity can also include, for example, the tendency for spectral peaks to be adjacent to each other. Sparse estimation methods, such as Lasso regression, SCAD (smoothly clipped absolute deviation), and MCP (minimax concave penalty), can be used to estimate the true spectral signal. Below, we will explain using lasso regression as an example.

[0049] 9 is a diagram showing a first example of the configuration of the spectrum estimation unit 55. The spectrum estimation unit 55 includes a parameter selection unit 551. The spectrum estimation unit 55 estimates the true spectrum signal using equation (12).

[0050]

number

[0051] In equation (12), β (hat) is the estimated solution of the true spectral signal, y is the observed spectral signal, X is the observation matrix, β is a temporary estimate of the true spectral signal (estimated spectral signal), and λ is a regularization parameter. The regularization parameter λ is a hyperparameter that determines the magnitude of the penalty for the estimated spectral signal, and λ≧0. The first term on the right-hand side of equation (12) is the squared error, that is, the squared error between the observed spectral signal y and the product (Xβ) of the observation matrix and the estimated spectral signal (estimated true spectral signal).

[0052] The second term on the right-hand side of equation (12) is a term used to obtain a sparse solution (i.e., to utilize the sparsity of the true spectral signal of the target sample as prior knowledge), and is referred to as the regularization term, penalty term, or constraint. Here, the penalty term can be the L1 norm, and the penalty value is given as the value of the L1 norm. If the regularization parameter λ is set to 0, the situation is the same as when no constraint is given, but the larger λ is, the sparser the estimation result becomes. In other words, a sparser estimated spectral signal can be obtained.

[0053] The spectral estimation unit 55 estimates the true spectral signal β based on the sparsity of the true spectral signal β so as to minimize the sum of the penalty value and the squared error value. Note that, if a method based on gradient descent, for example, is used to estimate the true spectral signal, it is not necessary to specifically calculate the squared error value or the penalty value for the estimated spectral signal. In this way, the spectral estimation unit 55 performs post-processing to accurately estimate the original true spectral signal β from the observed spectral signal y, which is a mixture of multiple spectral signal components obtained by low-resolution observation.

[0054] The regularization parameter λ must be selected appropriately to obtain appropriate estimation results. For example, by using an information criterion such as AIC (Akai information criterion) or BIC (Bayesian information criterion) as a selection criterion for the regularization parameter λ, the regularization parameter can be selected with low computational cost. Methods such as CV (cross validation) may also be used. Furthermore, if an appropriate value for the regularization parameter is known in advance, there is no need to select the value again. The estimated solution of the true spectral signal calculated using the selected regularization parameter value is used as the final estimation result.

[0055] The parameter selection unit 553 selects the regularization parameter λ. In the following, a case where AIC for lasso is used as a criterion will be described, but the method for selecting the regularization parameter λ is not limited to this.

[0056] When AIC for lasso is used as the criterion, AIC lasso can be expressed by equation (13), and λ AIC can be expressed by equation (14). In equation (13), β λ (hat) is the estimated solution of the true spectrum signal obtained by Eq. (12) using the regularization parameter λ, and ||β λ (Hat) || 0 is the estimated solution β λThe specific process for solving equation (14) is to first find the number of non-zero elements in λ. min ≦λ≦λ max For multiple values of λ that satisfy λ (hat) is calculated, and the estimated solution β λ (Hat) to AIC using equation (13) lasso Calculate the multiple AICs calculated. lasso The smallest AIC among lasso When λ is obtained, AIC In addition, the estimated solution β for the value of λ λ (Hat) is calculated using Lasso regression, path following, and homotopy method (β λ (Hat) is piecewise linear as a function of λ, which can be used to solve the lasso regression as a function of λ.

[0057]

number

[0058] As described above, the parameter selection unit 553 can select the regularization parameter λ that provides an estimated solution that satisfies in a balanced manner two requirements: that the estimated solution of the true spectral signal of the analyte introduced into the observation filter whose characteristics are represented by the observation function well describes the observed spectral signal, and that the number of spectral bins whose values are non-zero in the estimated solution is reduced.

[0059] FIG. 10 illustrates a second example of the configuration of the spectrum estimation unit 55. The spectrum estimation unit 55 includes a learning model 555. The learning model 555 is a neural network model incorporating deep learning. The learning model 555 is composed of an input layer, an output layer, and multiple intermediate layers, and acquires information equivalent to prior knowledge regarding true spectral signals through learning or other means. The input layer receives an observed spectral signal of an analyte of a target sample. The input observed spectral signal is specifically a vector whose element count is the number of spectral bins of mass-to-charge ratio (m / z) and whose element value is the signal intensity at each spectral bin. In the example of FIG. 10, the spectral bins have mass-to-charge ratios in the range of J1 to J2. The input to the input layer may also include information regarding analytes that may be contained in the target sample. The output layer outputs an estimated true spectral signal of the analyte of the target sample. The output estimated true spectral signal is specifically a vector whose element count is the number of spectral bins of mass-to-charge ratio (m / z) and whose element value is the signal intensity at each spectral bin. In this way, when the observed spectral signal of a target sample is input to the learning model 555, the learning model 555 can output an estimate of the true spectral signal of the target sample. Note that if several analytes that may be contained in the sample can be identified in advance, the output may also include estimates of the amounts of those analytes.

[0060] As described above, the spectrum estimation device 50 may include a learning model 555 that, when inputted with an observed spectral signal calculated based on a detection signal related to an analyte of a target sample acquired using an observation filter, estimates a true spectral signal of the analyte, and an observed spectral signal calculation unit 52 that calculates the observed spectral signal based on the detection signal related to the analyte of the target sample acquired using the observation filter. The spectrum estimation device 50 can input the observed spectral signal calculated by the observed spectral signal calculation unit 52 to the learning model 555 to estimate the true spectral signal.

[0061] The learning model 555 can be configured by combining hardware such as a CPU (e.g., a multiprocessor having multiple processor cores), GPUs (Graphics Processing Units), DSPs (Digital Signal Processors), FPGAs (Field-Programmable Gate Arrays), etc. Furthermore, the learning model 555 is not limited to a neural network model, and may be other machine learning models.

[0062] The display processing unit 56 can output the estimation result (including, for example, the mass spectrum) based on the true spectrum signal estimated by the spectrum estimation unit 55 to the display device 60.

[0063] Figure 11 shows an example of a mass spectrum. A mass spectrum is a graph of the intensities of individual ions (ion intensities at each m / z) that are separated by utilizing the differences in the mass-to-charge ratio (m / z) of the ions produced after ionizing the analytes in a target sample. Because mass spectra vary depending on the analyte, they can be used for component analysis of target samples. In the example of Figure 11, spectral peaks appear at multiple mass-to-charge ratio (m / z) values, but Figure 11 is a schematic example, and the number of spectral peaks may be one.

[0064] Next, the evaluation results regarding the estimation of true spectral signals according to this embodiment will be described.

[0065] Using actual mass spectra registered in the mass spectrum database Massbank and observation matrices for each measurement mode (first mode, second mode, third mode) constructed based on the observation function shown in Figure 8, a noise-added linear observation simulation according to the above formula (1) was performed on a computer, and mass spectra were estimated from the obtained observation signals using the estimation method according to this embodiment and the estimation method according to the comparative example. The conventional mass spectrum estimation method is a technique in which (1) the desired mass spectrum is obtained by (2) setting an appropriate threshold for the observation signal itself in the third mode, and (3) performing peak detection for signals above the threshold. In the comparative example, in order to make this conventional method comparable to the estimation method according to this embodiment, (1) the observation matrix I is calculated for the observation spectrum signal y in the third mode. sub Assuming a linear observation model, (2)λ AIC (3) We performed lasso regression based on the above and performed peak detection on the estimated spectrum.

[0066] Observation matrix I sub can be expressed by equation (15), and the N-dimensional unit matrix I N is a submatrix of the observation matrix I sub f I can be expressed by equation (16), and f I (j) has a bandwidth equal to the spectral bin width (this can be considered a high-resolution observation), as expressed in equation (17).

[0067]

number

[0068] The numerical simulation included the maximum peak intensity K for 26 samples of mass spectrum data (50 ≦ m / z < 650, spectrum bin width = 0.1 [m / z], total number of bins N = 6000). max = 100. The noise was independent Gaussian noise, and the standard deviation of the noise was set to 1.0.

[0069] FIG. 12 shows a first example of the evaluation results of the estimation method. The indices used for the evaluation are MSE, λ AIC / (the square of the L2 norm of f), F1score was used. MSE (Mean Squared Error) is the mean squared error, which is the square of the error between the actual value β and the estimated value β (hat), averaged over the total number of bins. The numbers in the figure indicate the average, and the numbers in parentheses indicate the standard error.

[0070] The smaller the MSE value, the smaller the error. AIC / (the square of the L2 norm of f) is an index of the peak detection limit in lasso estimation. Here, f is a window function, and is a (2J+1)-dimensional vector that lists the function values of the observation function for -J≦j≦J. In lasso estimation based on the observation matrix constructed from the window function f, λ AIC Since / (the square of the L2 norm of f) can be interpreted as a soft threshold, it can be treated as an index of the peak detection limit.

[0071] The F1 score is an index used to evaluate binary classification methods. If the numbers of true positive, false negative, false positive, and true negative samples when using a binary classification method are nTP, nFN, nFP, and nTN, respectively, the F1 score is calculated as F1 score = 2 · Precision · Recall / (Precision + Recall). Here, Precision = nTP / (nTP + nFP) and Recall = nTP / (nTP + nFN). The F1 score is the harmonic mean of Precision and Recall. The problem of determining the presence or absence of a spectral peak can be viewed as a binary classification problem of determining whether the signal value in a spectral bin is zero or non-zero, so the F1 score can be used as an evaluation index.

[0072] As shown in Fig. 12, in this embodiment, the MSE is smaller than in the comparative example. That is, it is understood that the estimation accuracy of the spectral peak intensity is improved. Furthermore, in this embodiment, λ is smaller than in the comparative example. AIC / (the square of the L2 norm of f) is small. That is, it is clear that spectral peaks with lower intensity can be detected, and peak detection sensitivity is improved. Furthermore, in this embodiment, the F1 score is larger than in the comparative example. That is, it is clear that according to this embodiment, the peak / non-peak classification performance is improved.

[0073] 13 is a diagram showing a second example of the evaluation results of the estimation method. In FIG. 13, λ in each measurement mode of this embodiment is AIC As shown in Figure 13, the λ for the first mode, the second mode, and the third mode are compared. AIC / (the square of the L2 norm of f) is smaller. This shows that the third mode can detect a spectral peak with a lower intensity than the first and second modes.

[0074] FIG. 14 shows a third example of the evaluation results of the estimation method. In FIG. 14, the horizontal axis represents the mass-to-charge ratio (m / z), and the vertical axis represents the relative signal intensity. FIG. 14 illustrates a true spectral signal (corresponding to the true spectral signal), the estimation results of the true spectral signal according to the comparative example, and the estimation results of the true spectral signal according to this embodiment. In FIG. 14, the symbol O indicates the position where the estimated spectral signal correctly has a peak among the peak positions of the true spectral signal, and the symbol X indicates the position where the estimated spectral signal erroneously does not have a peak among the positions of the true spectral peak signal. The greater the number of symbols O, the more successfully detected the peaks of the true spectral signal. This demonstrates the superiority of the estimation method according to this embodiment over the comparative example.

[0075] FIG. 15 is a flow diagram showing the processing procedure for constructing an observation matrix. As mentioned above, if the properties of the measurement system are known and the observation matrix X is known in advance with sufficient accuracy, there is no need to reconstruct the observation matrix. For convenience, the following description will be given assuming that the main processing entity is the spectrum estimation system (hereinafter also referred to as the "system"). The system ionizes the analytes in the standard sample and introduces them into the mass separation unit 16 (S11). Based on user operation, the system adjusts voltage parameters to scan the measurement window (observation filter) along the m / z axis (S12).

[0076] The system determines whether the scan has finished (S13), and if it has not finished (NO in S13), continues the processing from step S12 onwards. If the scan has finished (YES in S13), the system acquires a continuous signal on the m / z axis (S14). The system calculates an observed spectral signal for the continuous signal on the m / z axis and performs peak detection to obtain a true spectral signal (standard spectral signal) of the analyte in the standard sample (S15).

[0077] The system calculates an observation function and an observation matrix based on the standard spectral signal and the observed spectral signal of the analyte in the standard sample (S16). The observation function and the observation matrix can be calculated using equations (9) to (11). The system stores the calculated observation matrix in memory unit 54 (S17) and ends the process.

[0078] 16 is a flow diagram showing the processing procedure for estimating the true spectral signal of an analyte in a target sample. The system ionizes the analyte in the target sample and introduces it into the mass separation unit 16 (S21). Based on user operation, the system adjusts voltage parameters to scan the measurement window (observation filter) along the m / z axis (S22).

[0079] The system determines whether the scan has ended (S23), and if the scan has not ended (NO in S23), continues the processing from step S22 onwards. If the scan has ended (YES in S23), the system acquires continuous signals on the m / z axis (S24).

[0080] The system selects a regularization parameter λ (S25) and calculates an estimated solution β(hat) of the true spectral signal so as to minimize the sum of the squared error value (i.e., the squared error between the observed spectral signal and the product of the observation matrix and the estimated spectral signal) and the value obtained by multiplying the penalty value of the estimated spectral signal by the regularization parameter λ (S26). The penalty value of the estimated spectral signal corresponds to the L1 norm of β, the second term on the right-hand side of Equation (12). Equation (12) can be used to calculate the estimated solution β(hat) of the true spectral signal.

[0081] The system calculates the number of non-zero elements in the estimated solution β (hat) of the true spectral signal (S27), and determines whether the evaluation index of the regularization parameter satisfies a predetermined condition (S28). The evaluation index of the regularization parameter can be determined using Equation (13). Whether the predetermined condition is satisfied can be determined by whether the selected regularization parameter λ minimizes the evaluation index of Equation (13) using Equation (14).

[0082] If the evaluation index of the regularization parameter does not satisfy the predetermined condition (NO in S28), the system continues the processing from step S25 onwards. If the evaluation index of the regularization parameter satisfies the predetermined condition (YES in S28), the system outputs the estimated solution of the true spectral signal corresponding to that regularization parameter as the final estimated result of the true spectral signal (S29), and ends the processing.

[0083] The spectral estimation device 50 illustrated in FIG. 2 may be implemented, for example, by a personal computer. The spectral estimation device 50 may be configured with a CPU, ROM, RAM, GPU, and a recording medium reader. A computer program (program product) recorded on a recording medium (e.g., an optically readable disk storage medium such as a CD-ROM) may be read by a recording medium reader (e.g., an optical disk drive) and stored in RAM. Here, the computer program includes the processing procedures described in FIGS. 15 and 16. The computer program may be stored on a hard disk and then stored in RAM when the program is executed. By executing the computer program stored in RAM on a CPU, each process in the control unit 51, the observed spectral signal calculation unit 52, the observation matrix calculation unit 53, the spectral estimation unit 55, the display processing unit 56, and the standard spectral signal calculation unit 57 can be performed. Instead of being read by a recording medium reader, the computer program may be downloaded from another computer or network device via a network such as the Internet.

[0084] The spectral estimation device of this embodiment includes an observed spectral signal acquisition unit that acquires an observed spectral signal based on a detection signal related to an analyte in a target sample, acquired using an observation filter; and an estimation unit that uses the observed spectral signal as a response variable, and an observation matrix that has a bandwidth equal to or greater than a spectral bin width and whose elements are observation functions that represent the characteristics of the observation filter, as an explanatory variable, and estimates the true spectral signal based on sparsity related to the true spectral signal of the analyte.

[0085] In the spectrum estimation device of this embodiment, ionized analytes having mass-to-charge ratios (m / z) spanning a plurality of consecutive spectral bins are introduced into the observation filter.

[0086] In the spectral estimation device of this embodiment, the estimation unit estimates the true spectral signal so as to minimize the sum of the squared error between the observed spectral signal and the product of the observation matrix and the estimated true spectral signal, and the value obtained by multiplying the value of a penalty term based on the sparsity of the estimated true spectral signal by a regularization parameter.

[0087] The spectral estimation device of this embodiment includes a selection unit that selects the regularization parameter that reduces the number of spectral bins in which the value of the estimated solution of the true spectral signal of the analyte, which is introduced into the observation filter whose characteristics are represented by the observation function, is non-zero.

[0088] The spectrum estimation device of this embodiment includes a standard spectrum signal acquisition unit that acquires a standard spectrum signal based on a signal related to an analyte of a standard sample, the mass spectrum peak positions of which are known, acquired using an observation filter; and an observation matrix calculation unit that calculates a window function f so as to minimize an error between an observed spectrum signal y and the product (Bf) of a matrix B constructed from the standard spectrum signal and a window function f, and constructs an observation matrix X based on the result of the calculation. The estimation unit estimates the true spectrum signal using the observation matrix calculated by the observation matrix calculation unit.

[0089] The spectrum estimation device of this embodiment includes a learning model that, when an observed spectrum signal calculated based on a detection signal related to an analyte of a target sample acquired using an observation filter is input, estimates a true spectrum signal of the analyte; and an observed spectrum signal calculation unit that calculates the observed spectrum signal based on the detection signal related to the analyte of the target sample acquired using the observation filter. The observed spectrum signal calculated by the observed spectrum signal calculation unit is input to the learning model to estimate the true spectrum signal.

[0090] The spectrum estimation system of this embodiment includes the above-described spectrum estimation device and a spectrometer that outputs a detection signal related to an analyte in a target sample, the detection signal being acquired using an observation filter. The spectrum estimation device estimates a true spectrum signal based on the detection signal output by the spectrometer.

[0091] The computer program of this embodiment causes a computer to execute a process of acquiring an observed spectral signal based on a detection signal related to an analyte of a target sample, acquired using an observation filter, using the observed spectral signal as a response variable, and using an observation matrix, which has a bandwidth equal to or greater than the spectral bin width and has elements each of which is an observation function that represents the characteristics of the observation filter, as an explanatory variable, and estimating the true spectral signal based on the sparsity related to the true spectral signal of the analyte.

[0092] The spectral estimation method of the present embodiment acquires an observed spectral signal based on a detection signal related to an analyte in a target sample, acquired using an observation filter; uses the observed spectral signal as a dependent variable; uses an observation matrix, which has a bandwidth equal to or greater than the spectral bin width and whose elements are observation functions that represent the characteristics of the observation filter, as an explanatory variable; and estimates the true spectral signal based on the sparsity related to the true spectral signal of the analyte.

[0093] As described above, the estimation method of this embodiment applies a linear observation model to observed spectral signals obtained by low-resolution observation of an observation target with a wide measurement window, and estimates the target spectrum using prior knowledge, including sparsity, expected properties of the observation target. By utilizing prior knowledge about the spectrum of the observation target, a high-resolution spectrum can be reconstructed with high accuracy and sensitivity from the observed signal obtained by low-resolution observation, which contains a mixture of multiple spectral components. Furthermore, the estimation method of this embodiment improves the accuracy of spectral peak intensity estimation, as well as spectral peak detection sensitivity and peak classification performance. Furthermore, this embodiment can achieve a spectral resolution equal to or less than the width of the measurement window, depending on the measurement conditions. It can also be implemented by simply adding software processing without modifying the hardware of the measurement system.

[0094] Although the above embodiment has been described with respect to mass spectra in mass spectrometry, the present embodiment is not limited to the field of mass spectrometry. For example, the present embodiment can also be applied to an observation system (such as spectroscopic analysis) in which spectral signals relating to multiple consecutive spectral bins are simultaneously observed through low-resolution observation. [Explanation of symbols]

[0095] 10 Spectral measurement device 11 Sample injection section 12 Standard sample supply section 13 Flow path switching valve 14 Ionization section 15 Ion optics 16 Mass separation section 17 Detector 18 Voltage generation section 19 Control section 20 Control Unit 50 Spectral Estimation Device 51 Control section 52 Observation spectrum signal calculation unit 53 Observation matrix calculation unit 54 Storage section 55 Spectral estimation unit 551 Parameter Selection Section 555 Learning Model 56 Display processing section 57 Standard spectrum signal calculation unit 60 Display device

Claims

1. an observation spectrum signal acquisition unit that acquires an observation spectrum signal based on a detection signal related to an analyte of a target sample, the detection signal being acquired using an observation filter; an estimation unit that uses the observed spectral signal as a response variable, an observation matrix having a bandwidth equal to or greater than a spectral bin width and having elements each of which is an observation function representing a characteristic of the observation filter as an explanatory variable, and estimates the true spectral signal based on sparsity regarding the true spectral signal of the analyte; Equipped with Spectral estimator.

2. the observation filter introduces ionized analytes having mass-to-charge ratios (m / z) spanning a plurality of contiguous spectral bins; The spectral estimation device according to claim 1 .

3. The estimation unit the true spectral signal is estimated so as to minimize the sum of a square error between the observed spectral signal and a product of the observation matrix and the estimated true spectral signal, and a value obtained by multiplying a value of a penalty term based on the sparsity of the estimated true spectral signal by a regularization parameter.

3. The spectral estimation device according to claim 1 or 2.

4. a selection unit that selects the regularization parameter that reduces the number of spectral bins in which the value of an estimate of the true spectral signal of the analyte that is introduced into the observation filter whose characteristics are represented by the observation function is non-zero; The spectral estimation device according to claim 3 .

5. a standard spectrum signal acquisition unit that acquires a standard spectrum signal based on a signal related to an analyte of a standard sample, the mass spectrum peak position of which is known, acquired using an observation filter; an observation matrix calculation unit that calculates a window function so as to minimize an error between an observed spectrum signal and a product of a matrix constructed from the standard spectrum signal and a window function, and constructs an observation matrix based on the result of the calculation; Equipped with the estimation unit estimates the true spectral signal using the observation matrix calculated by the observation matrix calculation unit. The spectral estimation device according to any one of claims 1 to 4.

6. A spectral estimation device according to any one of claims 1 to 5; a spectrometer that outputs a detection signal related to the analyte of the target sample obtained using an observation filter; Equipped with the spectrum estimation device estimates a true spectrum signal based on the detection signal output by the spectrum measurement device; Spectral estimation system.

7. On the computer, obtaining an observed spectrum signal based on a detection signal relating to an analyte of the target sample obtained using the observation filter; the observed spectral signal is used as a response variable, an observation matrix having a bandwidth equal to or greater than a spectral bin width and having elements each of which is an observation function representing a characteristic of the observation filter is used as an explanatory variable, and the true spectral signal is estimated based on sparsity related to the true spectral signal of the analyte. A computer program that executes a process.

8. obtaining an observed spectrum signal based on a detection signal relating to an analyte of the target sample obtained using the observation filter; the observed spectral signal is used as a response variable, an observation matrix having a bandwidth equal to or greater than a spectral bin width and having elements each of which is an observation function representing a characteristic of the observation filter is used as an explanatory variable, and the true spectral signal is estimated based on sparsity related to the true spectral signal of the analyte. Spectral estimation methods.

Images