JP7581158B2 - Chromatograph data processing device and chromatograph - Google Patents

Description

The present invention relates to chromatography technology such as liquid chromatography, and in particular to a data processing device for a chromatography, and a chromatograph.

In the field of high performance liquid chromatography (HPLC), one of the development themes is chromatogram analysis methods. This is because the chromatography data system (CDS) calculates the peak area by time integration, but there are problems with the accuracy and stability of the results. Conventional methods find the characteristic points of a peak based on the time derivative of the chromatogram. These are, for example, the apex, the start point, the end point, the valley point, and the shoulder-like inflection point. The area is calculated by cutting the peak waveform from the chromatogram in a straight line with scissors from one of these differential characteristic points to the next (see, for example, Patent Document 1). Alternatively, a method has been adopted in which a peak-like waveform similar to a normal distribution function (Gaussian) is fitted to the measured chromatogram by the kerf fitting method (see, for example, Patent Document 2).

JP 2020-153864 A JP 2006-177980 A

Simply put, one would aim for more accurate waveform processing or devise more stable waveform processing, but the objective of the present invention is slightly different. The inventors of the present application realized that there may be waveform processing methods that excel at determining peak areas accurately, and waveform processing methods that calculate peak areas with good reproducibility. From this perspective, it should be possible to consider optimizing each of them.

In other words, the focus shifts to selecting a waveform processing method suited to the purpose. The purposes referred to here are (1) improving the accuracy of peak areas, and (2) improving the reproducibility of peak areas. There is also (3) correct resolution of overlapping peaks, which can be considered as part of (1). The object of the invention is to provide users with a method for accurately calculating peak areas, or to provide users with a waveform processing method with good reproducibility. It is possible to provide users with each of these two options. For this reason, the present invention can also be called meta-waveform processing, as it is a waveform processing method based on a higher (super) perspective.

In order to achieve the above object, the present invention provides
A data processing device for a chromatography, which separates a measurement sample in a column, detects the separated components, processes data of a chromatogram obtained, and determines a peak area corresponding to the amount of the separated components,
a peak processing method selection unit that receives an instruction according to the accuracy of the obtained peak area and the priority level of reproducibility, and selects a peak processing method corresponding to the instruction;
a peak area calculation unit that processes data of a given chromatogram by a selected waveform processing method and calculates a peak area for the separated components;
The present invention is characterized by having the following.

This allows the user to specify the priority level, such as whether they want to place importance on determining the peak area accurately or on calculating the peak area with good reproducibility, or on the balance between accuracy and reproducibility, and the appropriate peak processing method is selected to determine the peak area without the user having to directly select the method.

Note that the peak area corresponds to the amount of a component, and this also includes using the peak height as the peak area.

The present invention makes it easy to select a waveform processing method with excellent accuracy or good reproducibility.

FIG. 2 is a block diagram showing a schematic configuration of a chromatograph. FIG. 13 is an explanatory diagram showing an example of a waveform processing method. FIG. 13 is an explanatory diagram showing an example of a waveform processing method. FIG. 13 is an explanatory diagram showing an example of a waveform processing method. FIG. 13 is an explanatory diagram showing a display example of a chromatogram. 11 is a table showing an example of display of a processing result. FIG. 13 is a block diagram showing a schematic configuration of another chromatograph.

(Schematic configuration of chromatograph 100)
As shown in FIG. 1, the chromatograph 100 includes a chromatograph unit 110 that separates components in a sample using a column to generate chromatogram data, and a data processor 120 that performs waveform processing on the generated chromatogram data to determine peak areas corresponding to the amounts of the separated components.

As described below, the data processing device 120 is provided with a waveform processing method learning and selection unit 121 that obtains and stores evaluation indices of the accuracy and reproducibility of the peak area obtained by multiple types of waveform processing methods for chromatogram data obtained by the chromatograph unit 110 for a specified sample specimen, and chromatogram data input from the outside by the chromatogram data input and generation unit 130 or generated by a specified data processing, and selects a waveform processing method according to an instruction from the user of the chromatograph 100. Examples of the multiple types of waveform processing methods include a vertical division method from a peak valley (Figure 2), a tangent division method from an inflection point, etc. (Figure 3), and a processing method using a method of decomposing overlapping peaks by curve fitting such as Gaussian (Figure 4).

The data processing device 120 is also provided with a peak area calculation unit 122 that performs waveform processing using the waveform processing method selected by the waveform processing method learning and selection unit 121 on chromatogram data, such as that shown in FIG. 5, obtained by the chromatograph unit 110 for an actual sample, and calculates a peak area corresponding to the amount of separated components. The calculated peak area is displayed on the display unit 123 in a table format such as that shown in FIG. 6, together with the waveform of the chromatogram, such as that shown in FIG. 5. Although not particularly limited, when the chromatogram is displayed, the applied waveform processing method and the reliability score may be displayed together as shown in FIG. 5. The reliability score varies depending on the selected waveform processing method Xi and the components, but all the actual measurement data of the unknown sample show a common score. For example, the reliability scores in FIG. 5 and FIG. 6 are indicators that express the greater difference from the calculated peak area to both ends of the 95% confidence interval as a percentage. Other mathematical statistical indicators such as risk ratio and z score can also be displayed.

According to the chromatograph 100 described above, as described in detail below, the user can indicate whether they want to prioritize determining the peak area accurately or calculating the peak area with good reproducibility, and an appropriate waveform processing method can be selected to determine the peak area without the user having to directly select a waveform processing method.

(Key Points of the Present Embodiment)
The main points of this embodiment will be described below in detail.

(1) The user can specify peak area accuracy or reproducibility as the criterion for selecting the peak processing method, as described above. More specifically, the user interface allows selection of (a) improved peak area reproducibility or (b) improved peak area accuracy.

(2) Multiple peak processing methods are scored based on the evaluation criteria of peak area accuracy or reproducibility, and an appropriate peak processing method is selected.

(3) If the user specifies only peak area accuracy or reproducibility, the optimal waveform processing method is selected based on the above scores.

(4) When both peak area accuracy and reproducibility are specified by the user, the optimal waveform processing method is selected based on the overall score.

(5) When scoring is performed for the accuracy of the peak area, various chromatograms are input to the waveform processing method learning and selection unit 121. Both real data and simulated data can be applied, but the accuracy must be known. The same applies to the calculation of peak areas that evaluate the peak area ratio of overlapping peaks, such as by using a curve fitting method that uses a Gaussian-like peak shape.

(6) When scoring is performed for peak area reproducibility, multiple nearly identical chromatograms are input. Both real data and simulated data can be used.

(7) The specific form of the data processing device 120 may be a function installed in a stand-alone chromatograph 100 (CDS), or a network service that requests processing from a server PC.

(8) The above processing is suitable for using AI and big data, but can also be performed manually.

(Learning of waveform processing method, evaluation (learning) of chromatogram data input to the selection unit 121 and waveform processing method)
The multiple chromatograms input to the waveform processing method learning and selection unit 121 include general examples and individual special examples.

(1) For reproducibility evaluation, measured repeated injection data (multiple)
(2) Data obtained by combining multiple peak sizes measured (target: one component)
(3) Data showing overlapping peaks of known sizes of multiple analytes (subject: multiple components)
(4) The above three types of simulation data generated using random numbers (Monte Carlo method)
Since these are all intentionally synthesized, their true values are known and their accuracy can be evaluated (they become training data).

For the above chromatogram data, the peak areas are calculated using various waveform processing methods, such as the valley integration method, the truncation method that takes into account up to the second derivative, and the calculation of the peak area using a curve fitting method that uses a Gaussian-like peak shape, and the like. Scoring is then performed by determining the variance and error from the true value for each waveform processing method.

As an example of an optimal solution obtained as a result of the above search, the following waveform processing method is selected.

(1) Examples of waveform processing methods with high accuracy include the cutting method (scissors and balance integration) that takes into account up to the second derivative coefficient in waveform processing (Figure 2), or calculation of peak area using curve fitting method (Figure 3) that uses Gaussian-like peak shapes. (2) Examples of waveform processing methods with high reproducibility include the valley integration method in waveform processing (Figure 4). As described above, a waveform processing method is required that corresponds to an instruction on whether to prioritize the accuracy of the obtained peak area or reproducibility, and processing is performed using an appropriate waveform processing method by the user specifying the goal they want to prioritize.

(About each waveform processing method and processing parameters)
The processing steps and processing parameters of the waveform processing method learning and selection unit 121 that evaluates (learns) the accuracy and reproducibility of the waveform processing method will be described.

(1) Waveform processing method learning: The selection unit 121 imports multiple sheets of chromatogram data from the chromatograph unit 110 and the chromatogram data input and generation unit 130.

(2) Waveform processing method learning: The selection unit 121 obtains the score for each waveform processing method and selects the one that obtains the best score. The processing parameters used at this time include, for example, provisional automatic parameters and default parameters. Parameters related to the differential coefficient of each component include a noise threshold, a slope threshold, a shoulder threshold based on the second derivative, an S/N ratio, and a peak-valley ratio. Parameters for each component peak based on the curve fitting method include an initial value of the retention time, a standard deviation related to the peak width, and a time constant related to the asymmetry of the tailing peak.

(3) For the waveform processing method with the best score, the parameters are successively changed to obtain the score in the same manner, and the optimal parameters that obtain the best score are successively set in accordance with the evaluation criteria.

(4) The waveform processing method learning and selection unit 121 determines and stores the optimization parameters.

(5) Here, multiple chromatogram data and their respective known true values are used to calculate the above score.

Once the optimized parameters have been determined in the manner described above, the user can then specify the waveform processing method for which the optimized parameters have been set for unknown samples.

In addition, instead of optimizing the processing parameters after determining the optimal waveform processing method using provisional automatic parameters as described above, it is also possible to directly determine the optimal combination of waveform processing method and parameters by calculating a score for each combination of each waveform processing method and multiple types of parameters.

(More detailed explanation of "criteria for improving peak area reproducibility" and "criteria for measuring peak area accurately")
To select the most suitable peak processing method from among multiple methods, the following criteria can be used for score evaluation. Examples include "criterion for improving the reproducibility of peak areas" and "criterion for measuring peak areas accurately."

1. Criteria for improving peak area reproducibility 1.1. As part of the validation of analytical methods, repeatability is calculated using the relative standard deviation (RSD) (%) of quantitative values or peak areas. In other words, a waveform processing method with high reproducibility is selected according to the variation in peak areas obtained by each waveform processing method based on multiple chromatograms obtained from the same sample. Although a smaller RSD is desirable, it is often the case that results differ depending on the waveform processing method and the parameters set, even when the same chromatogram is statistically processed. Users may adjust parameters to improve reproducibility.

In the case of multiple components, an RSD is calculated for each component a, b, c, etc., and the sum of the squares of each RSD becomes the overall score for all components. Alternatively, the maximum RSD from the RSDs, the function max(each RSD), can be selected to create an overall RSD. Depending on the purpose, the inverse function min(RSDa, RSDb, RSDc, etc.) may also be used.

1.2. The present invention aims to use the RSD of this evaluation index as a criterion for selecting a waveform processing method. As mentioned above, since the RSD also depends on the parameters, waveform processing methods A, B, and C with changed parameters are called A1, A2, ..., B1, B2, etc., respectively. The RSD of each waveform processing method X and set Xi of parameters i can be calculated, and the set that gives the smallest RSD can be selected.

1.3. For example, if B2 is selected, you can further tune parameters in the vicinity of that parameter.

1.4. If we estimate the factors that cause the RSD to fluctuate, we can assume that each waveform processing method depends on the S/N ratio. In this case, the S/N ratio means the ratio of the peak height to the baseline noise. For example, in the case of a 1 mg peak, the S/N ratio is sufficiently large and the difference between the waveform processing methods is not large, but in the case of 1 μg, it is expected that the S/N ratio will affect the response performance of the waveform processing methods, resulting in differences. As an example, when the S/N ratio is large and small, multiple sheets of chromatogram data for two groups are prepared and the RSD is calculated for each. A waveform processing method that is not affected by the level of the S/N ratio is selected. In other words, if there is almost no superiority or inferiority in the former, a waveform processing method that outputs a good RSD even for the latter difficult subject is selected. This preparation can be done by injecting an actual sample to prepare a chromatogram, but it is generally easier to use a simulation to create multiple chromatograms with different noises.

1.5. Another possible factor in RSD variation is when the peak is not isolated. If it is isolated, it is thought that the superiority or inferiority of the waveform processing method is not so clear. If it is not an isolated peak, it is difficult to find the start and end points of the peak, and the degree of influence on the RSD depends on the waveform processing method. Another difficult phenomenon is when the baseline is not a horizontal straight line. A method of inputting multiple such difficult chromatograms and evaluating the durability of the waveform processing method by RSD is an effective means for the present invention.

2. Criteria for measuring peak area accurately 2.1. As mentioned above, if the S/N ratio is large enough, the peak is isolated, and the baseline is horizontal and linear, there is no significant difference in the peak processing method in terms of accurately calculating the peak area. The issue is which peak processing method can obtain a more accurate peak area even when any of these are not the case, i.e., when the S/N ratio is small, there are adjacent peaks, or the baseline is inclined or curved.

2.2. In this case, evaluation is performed using chromatogram data with known true values. That is, each waveform processing method processes the waveform of a chromatogram with known peak areas to calculate the peak area. The waveform processing method whose calculated result is closest to the true value is selected as the superior method. As mentioned above, each waveform processing method is called Xi, which is a set with a parameter.

2.3. For example, the true value of the peak area is At, the calculated peak area is Ac, and the absolute value (%) of |Ac - At|/At is used as the score, with the smaller being judged as good. This is a comparison of ratios, but in some cases comparing the absolute value of the peak area |Ac - At| (V·s) is more preferable for evaluating accuracy, as it is not affected by the size of At. In other words, the latter absolute value comparison is an evaluation method in which, even if there is a certain degree of deviation in |Ac - At| for relatively small peaks, if there is a large deviation for even one large peak, it is judged to be poor.

For example, if there is a waveform processing method X and multiple chromatograms i = 1, 2, ... are input, the peak area errors a1, a2, a3 ... for a certain component a are accumulated. Generally, the sum of squares is used.

When evaluating not only component a but also b, c, etc. simply, if the sum of squares of the errors for group a and the sum of squares of the errors for group b are added together in succession, the larger peaks will be emphasized in comparison. In such cases, if you want to evaluate accuracy more fairly from small peaks to large peaks, it is preferable to evaluate using an index such as the error rate obtained by dividing each error by its true value. In that case, the evaluation will be done using the sum of squares of the error rates of a1, a2, a3... b1, b2, b3... c1, c2, c3...

2.4. Chromatograms of overlapping input peaks with known true peak areas can be obtained in several ways:

2.4.1. Measured chromatogram obtained by mixing multiple components with known concentrations in a specified amount 2.4.2. Measure each chromatogram of a single component with known concentration and convert it into numerical data. Next, convert each component to an arbitrary amount by stretching or shrinking along the vertical axis, and generate a single chromatogram by synthesizing each chromatogram. Repeat the same operation to obtain multiple composite chromatograms. The true value of the peak area is known from the synthesis process. This method can be called a semi-measured composite chromatogram.

2.4.3. There is also the Monte Carlo method, which uses random numbers to generate a simulation. First, a single peak chromatogram of components a, b, ... is generated using Gaussian. EMG (Exponentially Modified Gaussian) equivalent to a unit amount such as 1.0 μg may also be used. Next, the blending amounts determined using random numbers are converted to the vertical axis direction, and finally, they are synthesized into one chromatogram. Noise and drift components can also be added randomly. Since the synthesis process of this Monte Carlo chromatogram is also known, the true value of the peak area is known. Only the Monte Carlo chromatogram has the characteristic that random numbers can be applied to the time axis on the horizontal axis. In other words, the retention time is also changed randomly, and the true value of the peak area is still known. However, it is generally preferable to set rules during generation, such as not placing two components close to each other for a certain period of time, or setting the peak area ratio to 100:1 or more.

2.5. When inputting a large amount of chromatogram data, such as 100 or 1,000 sheets, it may be difficult to actually measure each one, so in that respect, semi-measured synthetic chromatograms or Monte Carlo type chromatograms are advantageous.

2.6. The waveform processing method learning and selection unit 121 operates as follows:

2.6.1. Each waveform processing method uses provisionally automatically set parameters, which are denoted as automatic parameters 0. The waveform processing method learning and selection unit 121 takes in multiple pieces of input chromatogram data and selects the one with the smallest deviation score from A0, B0, .... The search method that uses these multiple chromatograms and the true values of their peak areas is the first screening of the waveform processing method.

2.6.2. Next, the parameters of the selected waveform processing method, for example, method B, are changed successively to B0, B1, B2, etc. The waveform processing method learning and selection unit 121 then searches for the waveform processing method Bn that minimizes the deviation score using an experimental design method. This is the second screening.

2.6.3. Using the optimized parameter method Bn, the waveform processing method learning and selection unit 121 thereafter instructs the peak area calculation unit 122 to process the analysis of the unknown sample using the Bn method when the user instructs the accuracy of the peak area to be prioritized.

(Other matters)
In addition, the waveform processing method corresponding to the instruction according to the accuracy of the obtained peak area and the priority level of reproducibility is not limited to the waveform processing method with the highest score, but may be determined within a range in which the accuracy score and/or the reproducibility score satisfy a predetermined relationship. Specifically, for example, the waveform processing method with the highest reproducibility may be determined among the waveform processing methods with an accuracy equal to or higher than a predetermined threshold, or conversely, the waveform processing method with the highest accuracy may be determined among the waveform processing methods with a reproducibility equal to or higher than a predetermined threshold. In addition, when there are multiple waveform processing methods with accuracy or reproducibility equal to or higher than a predetermined threshold, they may be presented to the user as candidates so that the user can select from them.

In addition, instead of using waveform processing to determine the peak area as described above, it is also possible to use machine learning to obtain the peak area and peak area ratio directly from the chromatogram data. That is, as shown in FIG. 7, for example, chromatogram data and known peak areas are input to a peak area calculation model 140 using a neural network or the like, and the correspondence between them is machine-learned as so-called supervised learning, and the peak area can be determined based on the machine learning results and the actually measured chromatogram data.

As a simple example of supervised learning, we present a problem of finding the peak area ratio of two overlapping peaks. Multiple chromatograms are input to the input layer, and the correct area ratio of the front peak is shown as teacher data to the output layer. This input data is not two-dimensional image data, but one-stroke time series data (numerical data showing the signal strength at each time). It is defined in advance that the area ratio of the rear peak added to the front peak is 1. The various chromatograms presented to the input layer can be formed by simulation as mentioned above, and 10,000 or more are preferable. Since they are formed by simulation, the correct area ratio is naturally known. In so-called deep learning, a significantly multi-layered intermediate layer is set in the neural network. The weight coefficients in the neural network are tuned by assigning multiple chromatograms and their corresponding teacher data. Backpropagation technology is used for this tuning. In addition, convolutional neural network technology can also be used for preprocessing if necessary. Once these weighting coefficients have been obtained, a chromatogram of unknown overlapping peaks can be input and the peak area ratio can be output. This process is exactly the resolution of overlapping peaks. Strictly speaking, this function for obtaining area ratios does not require the peak area itself. It is merely a ratio. For this reason, the total peak area is required to convert to peak area. In other words, the idea of this example is to draw a baseline for the two peaks together and then calculate the total peak area as their sum separately. Another feature of this method using a neural network is that it may not be able to answer at all how the two peaks overlap. In this respect, it differs greatly from the curve fitting method.

It is conceivable to apply the present invention to increase the number of peaks from two to three or four, but applicability becomes increasingly difficult. In the first place, the number of hidden peaks may be unknown, and humans will have to estimate the number by trial and error. It is possible to train deep learning to estimate the number, but this is also a new challenge. In addition to calculating peak area ratios, supervised learning is also possible, in which the peak areas are directly calculated using simulated chromatograms. It is expected that supervised learning will result in high accuracy.

On the other hand, the use of machine learning including reinforcement learning is expected to achieve high reproducibility. For example, in a task of searching for optimal parameters for a certain scissors-and-balance integration method, the relative standard deviation RSD of the peak area is used as an evaluation index, and the method searches for optimal parameters that minimize the RSD. In this case, the input can be multiple unknown chromatograms that have been actually measured. Normally, scissors-and-balance integration is performed with predetermined parameters, but in this example, the idea is to search for parameters that minimize the RSD, which is the evaluation index. The optimized parameters obtained in this way are then used to calculate the peak area.

Reference Signs List 100 Chromatograph 110 Chromatograph unit 120 Data processing device 121 Waveform processing method learning and selection unit 122 Peak area calculation unit 123 Display unit 130 Chromatogram data input and generation unit 140 Peak area calculation model

Claims (4)

A data processing device for a chromatography, which separates a measurement sample in a column, detects the separated components, processes data of a chromatogram obtained, and determines a peak area corresponding to the amount of the separated components,
a peak processing method selection unit that receives an instruction according to the accuracy of the obtained peak area and the priority level of reproducibility, and selects a peak processing method corresponding to the instruction;
a peak area calculation unit that processes data of a given chromatogram by a selected waveform processing method and calculates a peak area for the separated components;
A chromatographic data processing device comprising: 2. The data processing device of claim 1,
4. A data processing device for a chromatography, wherein the waveform processing method includes at least one of a valley integration method, a clipping method, and a curve fitting method. A data processing device for a chromatograph according to claim 1;
a chromatographic unit that separates and measures components contained in a sample;
A chromatograph comprising:
The data processing device is a chromatograph characterized in that it selects the waveform processing method in accordance with the error from the true value of the peak area obtained by each of a plurality of waveform processing methods based on a plurality of chromatograms in which the true values of the peak areas are known. A data processing device for a chromatograph according to claim 1;
a chromatographic unit that separates and measures components contained in a sample;
A chromatograph comprising:
The data processing device is a chromatograph characterized in that it selects the waveform processing method in accordance with the variation in peak areas obtained by a plurality of waveform processing methods based on a plurality of chromatograms obtained by the chromatographic unit for the same sample.