Waveform analysis method and waveform analysis device

Abstract

To provide a waveform analysis device for analyzing an object waveform which is chromatogram or an optical spectrum.SOLUTION: A waveform analysis device 4 comprises: a waveform dividing part 54 for dividing an object waveform into a plurality of partial waveforms; a determining part 55 for using a learned model generated by machine learning in which, a plurality of sets of a plurality of partial waveforms generated by dividing a reference waveform whose peak part position is known, for determining whether, each of the plurality of partial waveforms of the object waveform is a peak part; and a classifying part 56 for, on the basis of a determination result by the determining part, classifying the object waveform into a peak region where the peak parts are continuous, and a non-peak region other than the peak region.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for analyzing the waveforms of chromatograms and optical spectra. [Background technology]

[0002] Chromatography is used to identify and quantify components contained in a sample. In a chromatograph, the components in a sample are separated in a column, and the components that flow out of the column are detected in sequence. A chromatogram is then created with the horizontal axis representing time and the vertical axis representing detection intensity, and the concentration or amount of the compound corresponding to the peak in the chromatogram is calculated from the area or height of the peak.

[0003] In order to obtain the area and height of a peak from a chromatogram, it is necessary to identify the start and end points of the peak rising from the baseline of the chromatogram. The task of identifying the start and end points of a peak in a chromatogram is called peak picking. Since the baseline and peaks of a chromatogram often contain noise, it is necessary to appropriately identify the start and end points of the peak taking this into consideration. There are various peak picking algorithms that take noise into consideration. For example, Patent Document 1 describes a technique for performing peak picking by performing filtering processing using a filter having a bandwidth and a cutoff frequency according to the degree of temporal fluctuation of the chromatogram data. Some chromatogram analysis software can execute multiple algorithms, and the user can select one of the algorithms and set the necessary parameters to perform peak picking (for example, Non-Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2009-8582 A [Non-patent literature]

[0005] [Non-Patent Document 1] "Intelligent waveform processing algorithms for efficient analysis", [online], [Retrieved August 16, 2019], Shimadzu Corporation, Internet<URL:http: / / www.an.shimadzu.co.jp / hplc / support / faq / faq8 / i-peakfinder_introduction.htm> [Non-Patent Document 2] Olaf Ronneberger and 2 others, "U-Net: Convolutional Networks for Biomedical Image Segmentation", [online], [Submitted on 18 May 2015], arXiv.org, Internet<URL: https: / / arxiv.org / pdf / 1505.04597.pdf> Summary of the Invention [Problem to be solved by the invention]

[0006] The shape of the peaks and baselines of a chromatogram and the state of noise vary depending on the type of sample and detector. Therefore, there is a problem that it is cumbersome to search for algorithms and parameters suitable for the sample that is the subject of the chromatogram and the detector used every time peak picking is performed. In addition, there is a problem that the analyst's subjectivity is easily involved in the setting of parameters, and therefore peak picking cannot always be performed with high accuracy.

[0007] Here, we have described the case of performing peak picking of a chromatogram obtained by separating and measuring the components in a sample using a chromatograph. However, the same problem as described above also occurs when performing peak picking of an optical spectrum obtained by measuring the components in a sample using a spectrometer.

[0008] The problem to be solved by the present invention is to provide a technique that enables peak picking with high accuracy without complicated operations. [Means for solving the problem]

[0009] One aspect of the present invention that has been made to solve the above problems is a waveform analysis method for analyzing a target waveform that is a chromatogram or a spectroscopic spectrum, comprising the steps of: A trained model is created that identifies peak portions contained in an input waveform by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the peak portions are known; Dividing the target waveform into a plurality of partial waveforms; Using the trained model, determine whether each of a plurality of partial waveforms of the target waveform is a peak portion; Based on the result of the judgment, the target waveform is classified into a peak region where the peak portions are successive and a non-peak region other than the peak region. It is something.

[0010] Another aspect of the present invention, which has been made to solve the above problems, is a waveform analysis device that analyzes a target waveform that is a chromatogram or a spectroscopic spectrum, comprising: a waveform division unit that divides the target waveform into a plurality of partial waveforms; a determination unit that determines whether each of a plurality of partial waveforms of a target waveform is a peak portion by using a trained model created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the peak portions are known; a classification unit that classifies the target waveform into a peak region where the peak portions are successive and a non-peak region other than the peak region based on the result of the judgment by the judgment unit; Equipped with. Effect of the Invention

[0011] In the waveform analysis method and the waveform analysis device according to the present invention, a trained model is created in advance to determine whether an input waveform is a peak portion by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the position of the peak portion is known. Here, the reference waveform used to create the trained model is typically a waveform obtained by the same type of measurement as the target waveform (chromatogram or spectroscopic spectrum), but one created by theoretical calculation or the like can also be used. In addition, the plurality of partial waveforms of the reference waveform are created by dividing the reference waveform at intervals narrower than the peak width of the peak contained in the reference waveform. In this way, a trained model that has learned various shapes of partial waveforms that constitute a part of a peak is created. Thereafter, similar to the division of the reference waveform, the target waveform is divided at intervals narrower than the peak width of the peak expected to be included in the target waveform to create partial waveforms. Then, using the trained model, it is determined whether each of the plurality of partial waveforms obtained by dividing the target waveform in which the position of the peak portion is unknown is a peak portion. As a result, multiple partial waveforms of the target waveform are classified into peak portions and non-peak portions, and based on this, regions with successive peak portions are classified into peak regions, and regions with successive non-peak portions are classified into non-peak regions. The waveform analysis method and waveform analysis device according to the present invention classify the target waveform into peak regions and non-peak regions by using a trained model that has learned the features of multiple reference waveforms, making it possible to perform peak picking with high accuracy without complicated operations. [Brief description of the drawings]

[0012] [Figure 1] 1 is a diagram showing the configuration of a main portion of a liquid chromatography mass spectrometry system including an embodiment of a waveform analysis device according to the present invention. [Diagram 2] 1 is a flowchart illustrating a procedure for creating a trained model in one embodiment of a waveform analysis method according to the present invention. [Diagram 3] 4 is a flowchart for explaining a procedure for analyzing a chromatogram in the waveform analysis method of the present embodiment. [Figure 4]FIG. 13 is a diagram for explaining the creation of a trained model with two-class output using the waveform analysis device of this embodiment. [Diagram 5] FIG. 13 is a diagram for explaining the results of analyzing a chromatogram with two-class output using the waveform analyzer of the present embodiment. [Figure 6] FIG. 13 is a diagram for explaining the creation of a trained model with five-class output using the waveform analysis device of this embodiment. [Figure 7] FIG. 13 is a diagram for explaining the results of analyzing a chromatogram with five-class output using the waveform analyzer of the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] An embodiment of a waveform analysis device and a waveform analysis method according to the present invention will be described below with reference to the drawings. The waveform analysis device of this embodiment is incorporated as a part of a chromatography mass spectrometry system. Note that the waveform analysis device of this embodiment does not necessarily have to be configured integrally with the chromatography mass spectrometry device, and can also be configured separately from the chromatography mass spectrometry device.

[0014] FIG. 1 shows the main components of the liquid chromatograph mass spectrometry system of this embodiment. The liquid chromatograph mass spectrometry apparatus of this embodiment is roughly composed of a liquid chromatograph 1, a mass spectrometer 2, and a control and processing unit 4 that controls their operations. The liquid chromatograph 1 includes a mobile phase container 10 in which a mobile phase is stored, a pump 11 that draws in the mobile phase and delivers it at a constant flow rate (or flow velocity), an injector 12 that injects a predetermined amount of sample liquid into the mobile phase, and a column 13 that separates components contained in the sample liquid over time. The sample liquid flowing out of the column 13 is introduced into an electrospray ionization probe 201 of the mass spectrometer 2. An autosampler 14 that introduces a plurality of liquid samples one by one into the injector 12 is connected to the liquid chromatograph 1.

[0015] The mass spectrometer 2 has a multi-stage differential pumping system configuration with a first intermediate vacuum chamber 21 and a second intermediate vacuum chamber 22, the degree of vacuum of which is increased stepwise, between an ionization chamber 20, which is at approximately atmospheric pressure, and an analysis chamber 23, which is at a high vacuum and is evacuated by a vacuum pump (not shown). An electrospray ionization probe (ESI probe) 201 is installed in the ionization chamber 20, which sprays the sample liquid supplied from the liquid chromatograph 1 while giving an electric charge. The ionization chamber 20 and the first intermediate vacuum chamber 21 are communicated with each other via a thin-diameter heated capillary 202. The first intermediate vacuum chamber 21 and the second intermediate vacuum chamber 22 are separated by a skimmer 212 having a small hole at the top, and the first intermediate vacuum chamber 21 and the second intermediate vacuum chamber 22 are respectively equipped with ion guides 211 and 221 for converging ions and transporting them to the subsequent stage. A quadrupole mass filter 231 and an ion detector 232 are installed in the analysis chamber 23.

[0016] The mass spectrometer 2 can perform selected ion monitoring (SIM) measurement and MS scan measurement. In SIM measurement, the mass-to-charge ratio of ions passing through the quadrupole mass filter 231 is fixed, and only ions of that mass-to-charge ratio are detected. In MS scan measurement, the mass-to-charge ratio of ions passing through the quadrupole mass filter 231 is scanned over a predetermined mass-to-charge ratio range, and ions within the predetermined mass-to-charge ratio range are detected for each mass-to-charge ratio.

[0017] In addition to the memory unit 41, the control and processing unit 4 has the following functional blocks: a measurement control unit 51, a trained model creation unit 52, an analysis mode selection unit 53, a waveform division unit 54, a judgment unit 55, a classification unit 56, a noise value calculation unit 57, a noise removal unit 58, a peak separation unit 59, and a baseline estimation unit 60. The memory unit 41 is provided with a reference waveform storage unit 42, a measurement data storage unit 43, and a trained model storage unit 44. The substance of the control and processing unit 4 is a personal computer, and an input unit 6 and a display unit 7 are connected to it. In addition, a waveform analysis program is installed in the control and processing unit 4 in advance. By executing this waveform analysis program, the functions of the trained model creation unit 52, the analysis mode selection unit 53, the waveform division unit 54, the judgment unit 55, the classification unit 56, the noise value calculation unit 57, the noise removal unit 58, the peak separation unit 59, and the baseline estimation unit 60 are embodied.

[0018] The reference waveform memory unit 42 stores training data 421 and validation data 422. Both the training data 421 and validation data 422 are waveform (original waveform) data of a chromatogram obtained by measuring a sample containing various components with a chromatograph mass spectrometer (for example, a total ion chromatogram showing the time change in the total intensity of all ions of mass-to-charge ratios detected by MS scan measurement of components separated by a liquid chromatograph 1 with a mass spectrometer 2, or a mass chromatogram showing the time change in intensity of an ion of a specific mass-to-charge ratio by SIM measurement or MRM measurement), and the positions of the peaks have been specified in advance by peak picking. In addition, the original waveform data is represented by intensity values but The intensity scales are normalized in advance to be within a predetermined range (for example, ±1.0). By unifying a plurality of chromatograms with different intensity scales into a common intensity scale through normalization, the accuracy of the trained model described below can be improved. Here, chromatograms obtained by measuring real samples are used as the training data 421 and the validation data 422, but chromatograms created by simulation may also be used.

[0019] The original waveform of the chromatogram is divided into a predetermined number of partial original waveforms in the time axis direction. This predetermined number is, for example, 1,024 or 512, and is determined so that the width (length in the time axis direction) of each partial original waveform is at least smaller than the peak width. For example, it is determined based on the size of the peak width and the number of data points required to form one peak. As a specific example, a case will be described in which the minimum peak width in a chromatogram obtained by actual measurement is 0.2 min and the maximum peak width is 2.0 min. The analysis conditions and the state (degree of deterioration) of the column used for component separation may differ between the time when this measurement data is acquired and the time when the data to be analyzed is acquired. Therefore, taking these differences into consideration, the above-mentioned predetermined number is set so as to be able to handle even if the minimum peak width is narrowed to 0.1 min and the maximum peak width is widened to 3.0 min. Considering the reproducibility of the peak shape, one peak needs to include at least 20 points, more preferably 30 points. Considering these, in order to divide the chromatogram into 3.0-minute ranges and to divide the original waveform so that 30 points are included within a range of 0.1 minutes, which is the minimum expected peak width, the above-mentioned predetermined number must be set to at least 900. In the analysis example described below, in order to be able to handle cases in which the peak width is wider than the above-mentioned assumption, the chromatogram is divided into 5.0-minute time ranges, and the above-mentioned predetermined number is set to 1,024 based on the calculation: 5.0 minutes / 0.1 minutes x 20 points = 1,000 (≒ 1,024) points.

[0020] Each partial waveform data is associated with information (characteristic information) on the characteristics of the partial waveform. The characteristic information associated with the partial original waveform includes at least information on whether the partial waveform belongs to a peak region or a non-peak region. Other characteristic information will be described later.

[0021] The measurement data storage unit 43 stores various measurement conditions used for measuring a sample using the chromatograph mass spectrometer. In addition, chromatogram data obtained by measuring the sample is stored. The trained model storage unit 44 stores trained models for peak picking of a chromatogram. The trained model storage unit 44 stores trained models corresponding to each of the analysis modes described below.

[0022] Next, a method for analyzing the waveform of a chromatogram using the chromatographic mass spectrometry system of this embodiment will be described. In the chromatographic mass spectrometry system of this embodiment, by executing a waveform analysis program, it is possible to select either the creation of a trained model or the analysis of chromatographic data.

[0023] First, the procedure for creating a trained model will be described with reference to the flowchart in Figure 2.

[0024] When the user selects to create a trained model, the trained model creation unit 52 prepares an untrained trained model (step 1). Various trained models capable of executing semantic segmentation can be used for this trained model. Semantic segmentation is generally used to analyze images composed of pixel data distributed two-dimensionally, but in this embodiment, it is applied to the analysis of chromatogram waveforms composed of data arranged one-dimensionally along the time axis. For example, U-Net (see Non-Patent Document 2), SeGNet, PSPNet, etc. can be used as a trained model capable of executing semantic segmentation. In this embodiment, U-Net is used.

[0025] Next, the training data 421 and the verification data 422 are read out from the reference waveform storage unit 42 (step 2). Next, the trained model creation unit 52 sets the number of learning times (epoch) i to 1 (step 3) and inputs the training data 421 to the learning model (step 4). Based on the partial original waveform of the training data 421 and the characteristic information associated with the partial original waveform, the variables of the learning model are adjusted (step 5). In the U-Net used as the learning model in this embodiment, the weighting of the neural network is adjusted so that correct characteristic information can be obtained from the partial original waveform. After the input of the training data 421 and the adjustment of the variables of the learning model are completed, the training data 421 is stored in the storage unit 41 as the learning model (i) at the end of the number of learning times (epoch) i=1 (step 6). In addition, the verification data 422 is input to the learning model (i), and the accuracy rate of the characteristic information added by the learning model (i) by analyzing the partial original waveform of the verification data 422 is confirmed (step 7).

[0026] After checking the accuracy rate for the partial waveform of the verification data 422, it is determined whether i has reached a predetermined value (for example, 100) (step 8). Since i=1 at this point, it is determined that the predetermined value has not been reached (NO in step 8), 1 is added to i (i←i+1) (step 9), and the process returns to step 4. After returning to step 4, the same process as above is repeated until i reaches the predetermined value. Then, when i reaches the predetermined value (YES in step 8), an appropriate model is selected from the multiple learning models stored in the storage unit 41, and the learning model (i) is stored in the learned model storage unit 44 as a learned model (step 10), and a series of processes is terminated. The learned model is selected based on, for example, the highest accuracy rate for the verification data, no overlearning, and the like. In addition, the above-mentioned predetermined value is set to a value larger than the number of times that an appropriate learned model can be constructed. Such a value can be determined, for example, by referring to examples of past learning model construction, or by constructing a learned model preliminarily. In the above example, the learning model (i) is stored in the storage unit 41 for each learning round, but the learning model (i) may be stored for each predetermined number of times.

[0027] Next, the procedure for analyzing the waveform of an unanalyzed chromatogram will be described with reference to the flow chart of FIG.

[0028] When the user places a sample in the autosampler 14 and issues a command to start measurement, the measurement control unit 51 reads out the measurement conditions stored in the measurement data storage unit 43 and displays them on the screen of the display unit 7. When the user selects the measurement conditions to be used from the displayed ones (or makes appropriate changes) and issues a command to start measurement, the measurement control unit 51 executes chromatographic mass analysis of the sample to obtain a chromatogram. The measurement operation by the measurement control unit 51 is the same as in the past, so a detailed description will be omitted. Here, an example has been described in which a chromatogram is obtained by measuring a sample using the measurement control unit 51, but chromatogram data may also be obtained by, for example, reading in chromatogram data obtained in advance.

[0029] After acquiring chromatogram data by measuring a sample or reading out acquired data (step 11), when the user instructs analysis of the chromatogram data, the analysis mode selection unit 53 displays a screen on the display unit 7 inquiring of the user about the analysis mode of the chromatogram data. This screen displays analysis modes such as "two-class output (classification of peak region / non-peak region)", and the analysis mode is determined by the user selecting the desired one. Analysis modes other than the two-class output will be described later.

[0030] When an analysis mode is selected by the user (step 12), the waveform division unit 54 divides the waveform (original waveform) of the chromatogram into a predetermined number of partial waveforms (partial original waveforms) (step 13). However, the number is determined according to the length of the waveform (the length of the execution time of the chromatographic mass analysis) so that the width (length in the time axis direction) of each partial original waveform is at least smaller than the width of a peak predicted to be included in the chromatogram. This number of divisions may be the same as or different from the number of training data 421 and validation data 422. This number of divisions may also be determined appropriately from the above-mentioned viewpoint, and may be, for example, 512 or 1,024.

[0031] Next, the determination unit 55 reads out a trained model corresponding to the analysis mode selected by the user from the trained model storage unit 44 (step 14). Then, the partial original waveform is input to the trained model (step 15). The trained model determines whether or not the input partial original waveform belongs to a peak region (step 16). In this way, characteristic information (information on whether or not it belongs to a peak region) is attached to each partial original waveform.

[0032] When the determination unit 55 has finished determining each partial original waveform, the classification unit 56 classifies the partial original waveform based on the result and on the characteristic information (information on whether it belongs to a peak region or not) (step 17).

[0033] After classifying the partial original waveform, the noise value calculation unit 57 calculates a noise value based on the waveform of the non-peak region (step 18). The noise value can be, for example, the average value of the intensity of the partial original waveform belonging to the non-peak region. The noise removal unit 58 subtracts the calculated noise value from the original waveform to remove noise (step 19), and displays it on the screen of the display unit 7.

[0034] Next, an example of creating a trained model using actual chromatogram data and performing waveform analysis of the chromatogram (an example of two-class output) will be described. In this example, 13,359 pieces of chromatogram data were obtained by performing chromatographic mass analysis on a sample containing primary metabolites from a living organism, and the chromatogram data were manually peak-picked. Of these, 1,400 pieces of chromatogram data were used as validation data. In addition, 1,400 pieces of chromatogram data were used as test data to confirm the accuracy rate of the trained model. The remaining chromatogram data were used as training data. Furthermore, when creating the trained model, the number of times (the above-mentioned predetermined number of times) for inputting training data and checking the accuracy rate using validation data was set to 60, and the trained model that had been trained 60 times was selected as the trained model.

[0035] FIG. 4 is a graph showing the transition of the accuracy rate during the creation process of the trained model. In FIG. 4, "main / accuracy" is the accuracy rate of the training data 421, and "validation / main / accuracy" is the accuracy rate of the validation data 422. It can be seen that the accuracy rate of both the training data 421 and the validation data 422 improves little by little as the number of inputs increases. However, the accuracy rate does not improve as the number of inputs of the training data 421 and the validation data 422 increases, and if the number of inputs increases too much, a phenomenon called overlearning occurs and the accuracy rate decreases. Therefore, it is preferable to end the input of the training data 421 and the validation data 422 before overlearning occurs while checking the transition of the accuracy rate. When the accuracy of the two-class output of this trained model was verified using test data, the accuracy rate was 98%.

[0036] Figure 5 shows the results of analyzing the original waveform using the trained model. In Figure 5, the correct answers for the two-class output are shown as -1.25 (peak region) and -1.5 (non-peak region), and the judgment results using the trained model are shown as -1.75 (peak region) and -2.0 (non-peak region). These are also displayed together with the original waveform. As shown in Figure 5, it can be seen that the peak and non-peak regions of the original chromatogram waveform are correctly classified.

[0037] The above analysis mode (2-class output) is an example of classification with the minimum number of classes. Next, some preferable additional configurations that enable more detailed analysis of the chromatogram waveform will be described. The configurations described below can be appropriately combined with the above embodiment. In addition, multiple additional configurations can be combined as long as they do not perform contradictory processing.

[0038] One of the preferred additional configurations is a three-class output in which classification of peak boundaries is added to the two-class output. In this case, a trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing training data 421 and validation data 422 (hereinafter, collectively referred to as "reference waveforms"), in which the positions of peak boundaries are known. Then, the peak region, which is one of the characteristic information in the two-class output, is classified into two, a peak boundary portion and a non-peak boundary portion (others).

[0039] In chromatography, the components contained in a sample are separated by a column, but in a sample containing many components, it is not always possible to completely separate all of the components. As a result, multiple peaks that are superimposed on one another (superimposed peaks) may appear on the chromatogram. The process of calculating the concentration or content of a compound from the peak area or height is based on the premise that there is no overlap between the peaks. Therefore, when multiple peaks overlap, a process to split the peaks is required. When performing such a process, the peaks can be easily separated by using the above three-class output.

[0040] In addition, in the above-mentioned three-class output configuration, a four-class output configuration can be adopted in which the peak boundary portion, which is one of the characteristic information, is classified into two, the peak start point and the peak end point. In this case, a trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the peak start point and end point are known.

[0041] Conventionally, methods for splitting overlapping peaks include tailing, which splits the target peak into two peaks by treating the start point and end point of the target peak as one peak and superimposing another peak on the target peak; complete separation, which connects the start point, minimum point, and end point of the target peak in order to separate the peak into two peaks; and vertical separation, which separates the two peaks with a perpendicular line passing through the minimum point of the target peak. For example, Patent Document 1 describes software that can automatically pick peaks when an analyst selects tailing, complete separation, or vertical separation and inputs the required parameters. In addition, peaks may be separated by fitting using a model function such as a Gaussian function.

[0042] In the above tailing process, the start and end points of two peaks are arranged in the order of the first peak start point and the second peak start point from the side with the shortest retention time, followed by the first and second peak end points. On the other hand, in complete separation and vertical splitting, the first peak start and end points are arranged in the order of the shortest retention time, followed by the second peak start and end points. In the above four-class output configuration, the peak boundary is distinguished into the peak start point and peak end point, so the user can check the order and then consider an appropriate peak splitting method.

[0043] Furthermore, a configuration can be adopted in which the peak region, which is one of the characteristic information, is classified into a superimposed peak region and a single peak region. In this case, a trained model can be created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including a single peak, both of which have known positions, and a superimposed peak, which is a combination of multiple peaks. By adopting this configuration, it becomes possible for the user to easily determine the region in which peak separation is required.

[0044] Furthermore, in this configuration, the superimposed peak regions can be classified into three types, namely vertical division, complete separation, and tailing processing, to form a five-class output configuration. In this case, a trained model can be created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including superimposed peaks that have been peak-separated by tailing processing, complete separation, and vertical division.

[0045] Which method is suitable for splitting (separating) overlapping peaks contained in the chromatogram to be analyzed depends on the shape of the baseline drift and the shape of the peaks contained in the chromatogram, and is subjective to the user's judgment. With the above 5-class output configuration, a trained model can be used to automatically determine the method suitable for splitting overlapping peaks, making it possible to split peaks objectively and accurately without the user's subjective judgment.

[0046] FIG. 6 is an example of a trained model used in the above analysis mode (five-class output). As in FIG. 4, which shows an example of two-class output, "main / accuracy" in FIG. 6 is the accuracy rate of the training data 421, and "validation / main / accuracy" is the accuracy rate of the validation data 422. As in FIG. 4, in this example, the accuracy rate of both the training data 421 and the validation data 422 improves little by little as the number of inputs increases. When the accuracy of the five-class output of this trained model was verified using test data, the accuracy rate was 97%.

[0047] Figure 7 shows an example of the original waveform of a chromatogram analyzed with five-class output using the trained model. In Figure 7, the correct answers for the five-class output are shown as -1.1 (peak end point), -1.2 (peak start point), -1.3 (vertical split peak), -1.4 (single peak), and -1.5 (non-peak area), and the outputs by the trained model are shown as -1.6 (peak end point), -1.7 (peak start point), -1.8 (vertical split peak), -1.9 (single peak), and -2.0 (non-peak area). These are also displayed together with the original waveform. As shown in Figure 7, it can be seen that the partial waveforms of the chromatogram are correctly classified.

[0048] In the configuration for classifying superimposed peaks and single peaks, the peaks can also be separated by fitting using a predetermined model function (for example, Gaussian fitting or EMG (Exponential Modified Gaussian) fitting). In this case, the model function to be used is stored in advance in the storage unit 41. In addition, a peak separation unit 59 is used that separates peaks by fitting a portion determined to be a superimposed peak region by the trained model using a predetermined model function.

[0049] In the configuration for classifying superimposed peaks and single peaks, classification according to the number of peaks included in the superimposed peaks can be further added. In this case, a trained model can be created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including a plurality of superimposed peaks in which different numbers of peaks overlap. This allows the user to easily grasp the number of peaks included in the superimposed peaks.

[0050] Although the above preferred embodiment mainly adds classification related to peak separation, it is preferable to add a function of estimating the baseline included in the chromatogram. In this case, a baseline estimation unit 60 is used to estimate the baseline based on a partial waveform determined to be a non-peak region. The baseline estimation unit 60 estimates the baseline of the entire target waveform by fitting the partial waveform of the non-peak region with a model function, etc. This can be useful for considering which of vertical division, complete separation, and tailing processing is more suitable when separating superimposed peaks, for example. Furthermore, in a configuration in which superimposed peaks are classified using a trained model, the user can easily check the validity of the classification result using the trained model.

[0051] Furthermore, it is advisable to add a peak area calculation unit that calculates the area of ​​a peak region (or a peak height calculation unit that calculates the height of a peak) after classifying the original waveform by the trained model to the functional block in the control / processing unit 4 of the chromatographic mass spectrometer of the above embodiment. This makes it possible to easily quantify the component corresponding to the peak.

[0052] The above-mentioned embodiments are merely examples, and may be modified appropriately in accordance with the spirit of the present invention. In the above-mentioned embodiments, the configuration is that the device is incorporated into a part of a chromatographic mass spectrometry system, but the device may be configured as a waveform analyzer independent of the chromatographic mass spectrometry device. In that case, chromatogram data previously acquired by the chromatographic mass spectrometry device may be read and analyzed. In the above-mentioned embodiments, the processing of the waveform of a chromatogram obtained by chromatographic mass spectrometry is described as an example, but a chromatogram obtained by a chromatograph having a detector (spectrophotometer) other than a mass spectrometer or a gas chromatograph may also be analyzed in the same manner. Furthermore, the subject of the analysis is not limited to a chromatogram, and for example, a spectroscopic spectrum (a waveform showing a change in detection intensity with respect to a wavelength or wave number axis) obtained by measurement using a spectrophotometer may also be analyzed in the same manner as described above.

[0053] [Aspects] It will be appreciated by those skilled in the art that the exemplary embodiments described above are examples of the following aspects.

[0054] (Section 1) A waveform analysis method according to one aspect of the present invention includes: A waveform analysis method for analyzing a target waveform which is a chromatogram or a spectroscopic spectrum, comprising: A trained model is created that identifies peak portions contained in an input waveform by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the peak portions are known; Dividing the target waveform into a plurality of partial waveforms; Using the trained model, determine whether each of a plurality of partial waveforms of the target waveform is a peak portion; Based on the result of the judgment, the target waveform is classified into a peak region where the peak portions are successive and a non-peak region other than the peak region. It is something.

[0055] (Section 10) A waveform analysis device according to another aspect of the present invention includes: A waveform analyzer for analyzing a target waveform which is a chromatogram or a spectroscopic spectrum, comprising: a waveform division unit that divides the target waveform into a plurality of partial waveforms; a determination unit that determines whether each of a plurality of partial waveforms of a target waveform is a peak portion by using a trained model created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the peak portions are known; a classification unit that classifies the target waveform into a peak region where the peak portions are successive and a non-peak region other than the peak region based on the result of the judgment by the judgment unit; Equipped with.

[0056] In the waveform analysis method described in paragraph 1 and the waveform analysis device described in paragraph 10, a trained model is created in advance to determine whether an input waveform is a peak portion or not by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the position of the peak portion is known. Here, the reference waveform used to create the trained model is typically a waveform obtained by the same type of measurement as the target waveform (chromatogram or spectroscopic spectrum), but one created by theoretical calculation or the like can also be used. In addition, the plurality of partial waveforms of the reference waveform are created by dividing the reference waveform at intervals narrower than the peak width of the peak contained in the reference waveform. In this way, a trained model that has learned various shapes of partial waveforms that constitute a part of a peak is created. Thereafter, similar to the division of the reference waveform, the target waveform is divided at intervals narrower than the peak width of the peak expected to be included in the target waveform to create partial waveforms. Then, using the trained model, it is determined whether each of the plurality of partial waveforms obtained by dividing the target waveform in which the position of the peak portion is unknown is a peak portion or not. As a result, multiple partial waveforms of the target waveform are classified into peak portions and non-peak portions, and based on this, regions with successive peak portions are classified into peak regions, and regions with successive non-peak portions are classified into non-peak regions. In the waveform analysis method and waveform analysis device according to the present invention, the target waveform is classified into peak regions and non-peak regions by using a trained model that has learned the features of multiple reference waveforms, so that peak picking can be performed with high accuracy without complicated operations.

[0057] (Section 2) In the waveform analysis method according to claim 1, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform having known peak boundary positions; Identifying a boundary of a peak included in the partial waveform determined to be the peak portion; The peak regions are divided into peak borders and non-peak borders.

[0058] (Section 11) 11. The waveform analyzer according to claim 10, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform having known peak boundary positions; The determination unit identifies a boundary of a peak included in the partial waveform determined to be the peak portion, The classification unit further classifies the peak regions into peak boundary portions and non-peak boundary portions.

[0059] The waveform analysis method described in item 2 and the waveform analyzer described in item 11 make it possible to easily separate overlapping peaks when quantifying sample components, etc.

[0060] (Section 3) 3. The waveform analysis method according to claim 2, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the start and end points of peaks are known, determining whether the boundary of the peak is a peak start point or a peak end point; The peak boundaries are further classified into peak starts and peak ends.

[0061] (Section 12) 12. The waveform analyzer according to claim 11, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform in which the positions of the start and end points of peaks are known, The determination unit determines whether the boundary of the peak is a start point of the peak or an end point of the peak, The classification unit further classifies the boundaries of the peaks into peak start points and peak end points.

[0062] In the waveform analysis method described in paragraph 3 and the waveform analysis device described in paragraph 12, the peak boundary portions are distinguished into the peak start point and the peak end point, so that the user can check the arrangement and then consider an appropriate peak division method.

[0063] (Section 4) In the waveform analysis method according to any one of claims 1 to 3, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including a single peak, each of which has a known position, and a superimposed peak in which a plurality of peaks are superimposed; determining whether a peak included in a peak region, which is a region in which the peak portions are continuous, is a single peak or a superposed peak; The peak regions are further classified into single peak regions and overlapping peak regions.

[0064] (Section 13) In the waveform analysis device according to any one of claims 10 to 12, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including a single peak, each of which has a known position, and a superimposed peak in which a plurality of peaks are superimposed; the determining unit determines whether a peak included in a peak region, which is a region in which the peak portions are continuous, is a single peak or a superimposed peak; The classification unit further classifies the peak regions into a single peak region and a superimposed peak region.

[0065] In the waveform analysis method described in item 4 and the waveform analyzer described in item 13, the user can easily determine the area where peak separation is required.

[0066] (Section 5) In the waveform analysis method according to claim 4, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including superimposed peaks that have been peak-separated by a tailing process that separates the region from the start point to the end point of the superimposed peak into a single peak with another peak superimposed on the single peak, a complete separation that separates the peak with a line connecting the start point, minimum point, and end point of the superimposed peak in order, and a vertical division that separates two peaks with a perpendicular line passing through the minimum point of the superimposed peak; A suitable one for separating the multiple peaks contained in the overlapping peak region is determined.

[0067] (Section 14) 14. The waveform analyzer according to claim 13, the trained model is created by machine learning using a plurality of sets of partial waveforms created by dividing a reference waveform including superimposed peaks that have been peak-separated by a tailing process that separates the region from the start point to the end point of the superimposed peak into a single peak with another peak superimposed on the single peak, a complete separation that separates the peak with a line connecting the start point, minimum point, and end point of the superimposed peak in order, and a vertical division that separates two peaks with a perpendicular line passing through the minimum point of the superimposed peak; The determining unit determines which is suitable for separating the plurality of peaks included in the overlapping peak region.

[0068] The waveform analysis method described in paragraph 5 and the waveform analysis device described in paragraph 14 can automatically determine a method suitable for splitting overlapping peaks using a trained model, making it possible to split peaks objectively and accurately without the user's subjectivity being involved.

[0069] (Section 6) In the waveform analysis method according to claim 4, A predetermined model function is used to separate the multiple peaks contained in the overlapping peak region.

[0070] (Section 7) 7. The waveform analysis method according to claim 6, A plurality of peaks contained in the overlapping peak region are separated by Gaussian fitting or Exponential Modified Gaussian fitting.

[0071] (Section 15) 14. The waveform analyzer according to claim 13, further comprising: a peak separation section for separating the plurality of peaks included in the overlapping peak region using a predetermined model function; Equipped with.

[0072] (Section 16) 16. The waveform analyzer according to claim 15, The peak separation unit separates peaks by Gaussian fitting or Exponential Modified Gaussian fitting.

[0073] In the waveform analysis method described in item 6 or 7 and the waveform analyzer described in item 15 or 16, overlapping peaks can be automatically separated by a model function.

[0074] (Section 8) In the waveform analysis method according to any one of claims 1 to 7, A noise value is calculated based on the target waveform classified into the non-peak region.

[0075] (Section 17) 17. The waveform analyzer according to any one of claims 10 to 16, further comprising: A noise value calculation unit that calculates a noise value based on the target waveform classified into the non-peak region. Equipped with.

[0076] In the waveform analysis method described in paragraph 8 and the waveform analysis device described in paragraph 17, the noise value contained in the target waveform can be automatically calculated.

[0077] (Section 9) 9. The waveform analysis method according to any one of claims 1 to 8, A baseline is estimated based on the target waveform classified into the non-peak region.

[0078] (Section 18) 18. The waveform analyzer according to any one of claims 10 to 17, further comprising: A baseline estimation unit that estimates a baseline based on the target waveform classified into the non-peak region. Equipped with.

[0079] The waveform analysis method described in paragraph 9 and the waveform analyzer described in paragraph 18 can be useful in determining whether vertical division, complete separation, or tailing processing is more appropriate when separating overlapping peaks, for example. In addition, in a configuration in which overlapping peak regions are classified using a trained model, a user can easily check the validity of the classification results obtained using the trained model. [Explanation of symbols]

[0080] 1. Liquid chromatograph 2...Mass spectrometer 4. Control and processing section 41...Storage section 42...Reference waveform storage section 421...Training data 422…Verification data 43...Measurement data storage unit 44…Model memory section 51...Measurement control section 52…Model Creation Section 53...Analysis mode selection section 54…Waveform division section 55…Judgment section 56...Classification section 57...Noise value calculation section 58…Noise reduction section 59...Peak separation section 60...Baseline Estimation Section 6. Input section 7…Display section

Claims

1. A waveform analysis method for analyzing a target waveform composed of a target data string, which is a one-dimensional data string, using a computer, comprising: A trained model is created that uses machine learning to identify, by semantic segmentation, the parts that constitute peaks in a waveform that is composed of an input data string, using multiple reference data strings that are one-dimensional data strings that constitute a reference waveform in which the positions of the peak parts are known; inputting the target data string into the trained model to identify a portion of the target waveform that constitutes a peak; The target waveform is classified into a peak region corresponding to the portion that constitutes the peak and a non-peak region other than the peak region. A waveform analysis method for performing processing.

2. The trained model classifies each of the plurality of data constituting the input data string into data that belong to the portion that constitutes the peak and other data, thereby identifying the portion that constitutes the peak in the waveform formed from the input data string. The waveform analysis method according to claim 1 .

3. The trained model is executing a first downsampling process for generating a first feature data string having a smaller number of data than the input data string by performing a convolution process on the input data string; executing a second downsampling process for generating a second feature data string having a smaller number of data than the first feature data string by performing a convolution process on the first feature data string; performing a first upsampling process on the second feature data sequence to generate a first reconstructed data sequence having the same size as the first feature data sequence; performing a second upsampling process on the first reconstructed data sequence to generate a final reconstructed data sequence having the same data size as the target data sequence; In the first upsampling process, each data constituting the second feature data string is restored to a position in the first restored data string corresponding to a position in the first feature data string before being compressed into the second feature data string; In the second upsampling process, each data constituting the first restored data string is restored to a position in the final restored data string corresponding to a position in the target data string before being compressed into the first feature data string. By doing so, a portion constituting a peak is identified in a waveform formed from the input data string. The waveform analysis method according to claim 1 .

4. The trained model is In the first upsampling process, data other than the restored data in the first feature data string is set to 0; In the second upsampling process, data other than the restored data in the second feature data string is set to 0. The waveform analysis method according to claim 3 .

5. The waveform analysis method according to claim 1 , wherein the trained model is composed of any one of U-Net, SeGnet, and PSPNet.

6. A waveform analysis device that analyzes a target waveform that is configured from a target data string that is a one-dimensional data string, comprising: a classification unit that inputs a target data sequence into a trained model that is created by machine learning using a plurality of reference data sequences, which are one-dimensional data sequences that constitute a reference waveform in which the positions of peak portions are known, and that identifies, by semantic segmentation, portions that constitute peaks in a waveform constituted by an input data sequence, identifies portions that constitute peaks in the target waveform, and classifies the target waveform into peak regions that correspond to the portions that constitute the peaks and other non-peak regions; A waveform analysis device comprising:

Images