CN101776671B - Real-time feature extraction method for analysis of complex ingredient of traditional Chinese medicine - Google Patents

Abstract

The invention provides a real-time feature extraction method for complex component analysis of traditional Chinese medicine, which consists of four modules: data communication module, two-dimensional feature chain detection, local noise and local baseline correction, and feature resolution, and sequentially analyzes the mass spectra collected by mass spectrometers data, to judge whether the data collected at the previous time point have continuous features, so as to dynamically complete the detection of the two-dimensional feature chain; using the mass-to-charge ratio and time information contained in the two-dimensional feature chain, the time dimension can be quickly removed. Noise and baseline, overcome the shortcomings of previous algorithms that are difficult to accurately estimate the baseline by simply using the time dimension; because the estimated noise and baseline have local characteristics, the local SNR is a feature of whether there are components in the feature chain, which simplifies the feature detection implementation. The method of the invention is reasonable in design, and the data processing system not only has real-time characteristics, but also has few user-defined parameters and fast operation speed, and is especially suitable for liquid chromatography mass spectrometry.

Description

A real-time feature extraction method for complex component analysis of traditional Chinese medicine

technical field

The invention belongs to the field of pharmacy and relates to a real-time online feature extraction method for complex component analysis of traditional Chinese medicines.

Background technique

Traditional Chinese medicine is the treasure of the Chinese nation. It has been clinically practiced for more than 2,000 years and has played an indelible role in the process of the Chinese nation's reproduction. With the country's continuous investment in Chinese medicine science and technology, gratifying progress has been made in the modernization of Chinese medicine. The curative effects of some medicines have been proved by scientific experiments again, and some even surpass chemical medicines. Today's western countries have an absolute advantage in the research and development of chemical medicines, and it is difficult to change them in a short period of time. Therefore, vigorously developing the cause of traditional Chinese medicine is of great significance to my country's pharmaceutical industry. However, the composition of traditional Chinese medicine is extremely complex, and the basic research has not been deep enough for a long time. Although there are historical reasons for this, the limitations of the existing technology are particularly prominent. Before liquid chromatography-mass spectrometry (LC-MS) technology matured, the study of traditional Chinese medicine needed to use phytochemical separation methods to separate monomer compounds from traditional Chinese medicine, and then go through four major spectral analysis to understand the compounds. structural information. However, LC-MS technology has completely changed the traditional traditional Chinese medicine substance basic research mode, and while improving the structure confirmation speed, it also makes it possible to identify trace components that could not be separated and extracted before. However, the analysis of existing LC-MS data is mainly done manually, which has become a bottleneck problem in current mass spectrometry applications, especially when a large number of samples of traditional Chinese medicine component libraries need to be analyzed. At present, the workstations provided by mainstream LC-MS manufacturers (such as Thermoelectric Group, Applied Bio, and Waters) can only perform simple one-dimensional data analysis after data collection is completed. Users need to set multiple parameters. The set of parameters can only be applied to specific samples, and different samples need to be adjusted accordingly, so data analysis has become the speed-limiting step for a large number of LC-MS applications.

The signal collected by LC-MS is composed of time dimension and mass dimension, while the signal collected by general liquid chromatography connected with ultraviolet detector (LC-UV) has only one time dimension. Usually, people refer to the change of the intensity of the compound elution over a period of time from LC-UV as "chromatographic peak"; while in two-dimensional LC-MS, when the compound is eluted, there is not only a time course, but also a mass distribution. We call the regions containing both two-dimensional information the "features" of compounds, and the algorithms used to find these regions are called feature extraction algorithms or methods. Due to the increase in the dimensionality of the data collected by LC-MS, it greatly increases the difficulty of extracting information from it. In the field of traditional Chinese medicine, there are few methods for studying LC-MS feature extraction, but in the field of bioinformatics, it is a very popular direction, which benefits from the need to process a large amount of LC-MS data in proteomics and metabolomics research driven by demand. Well-known open source tools include: XCMS, MZmine, etc.; commercial software includes: AnalyzerPro, ProTrawler, etc. These tools are only used for off-line analysis of data after LC-MS acquisition. The algorithm is based on the data within the entire analysis time. For example, XCMS must first combine the signals of a certain mass number range after the acquisition is completed. , and then peak detection can be performed from it, and these software need to set multiple parameters, some parameters have no actual physical meaning, such as wavelet scale, coefficients, etc., which are difficult for ordinary users to understand.

Contents of the invention

The invention aims at the deficiencies and defects of the prior art, and provides a real-time feature extraction method for complex component analysis of traditional Chinese medicine. This method is based on the two-dimensional feature information of the time dimension and quality dimension of LC-MS, and is realized through two-dimensional feature chain detection, local noise and baseline estimation, and feature resolution. . The present invention is realized through the following steps:

1. Mass spectrometry data collection: The complex samples of traditional Chinese medicine are firstly separated by the chromatographic unit, and then the mass spectrometer analyzes the fractions eluted by the chromatogram in a full-scan mode at a certain sampling frequency (f), and the collected data are represented by centroid (stick diagram ) format storage (this is the format supported by existing mass spectrometers). The data collected at each time point (integer multiple of 1/f) is a mass spectrum, corresponding to the data of the mass spectrum dimension; the data collected at different time points constitutes the chromatographic dimension information, such as the mass spectrum collected at each time point The intensities of all ions are added to obtain the response intensity at each time point, then the response intensities at all time points constitute the total ion current chromatogram. Chromatography in the present invention includes liquid chromatography (HPLC) and ultra-high pressure liquid chromatography (UPLC); Quadrupole mass spectrometry, triple quadrupole mass spectrometry, ion trap mass spectrometry or time-of-flight mass spectrometry;

2. Two-dimensional feature chain detection: BNN (minWidth, CC)

Every time the mass spectrometer collects a mass spectrum at a time point, it is sent to the BNN module for analysis. First, the mass-to-charge ratio and intensity information in the mass spectrogram are respectively assigned to the mass-to-charge ratio array MZ and the intensity array INTEN, and then the two-dimensional feature chain containing compound information is detected by the bidirectional nearest neighbor algorithm in time order, and the detected two-dimensional features The chain is stored in CC and can be obtained by other modules at any time;

3. Local noise and local baseline estimation: De_Noise_Baseline(minWidth)

With the increase of collected data, if the length N _k of a certain two-dimensional feature chain CC _k in CC is greater than minWidth, noise and baseline can be estimated for it. The two-dimensional feature chain contains dual information of chromatography dimension and mass spectrum dimension, which are composed of time, MZ and INTEN respectively. The response intensity information of the two-dimensional feature chain is linearly convolved with the high-pass filter, and the pulse signal is filtered out by applying 3 times the population standard deviation, which is the noise estimation of the chromatographic dimension. In order to estimate the baseline in the chromatographic dimension, the present invention designs the following algorithm based on the difference in mass fluctuation between the group partition and the zero group partition in the two-dimensional characteristic chain mass spectrum dimension:

(1) Find the time point with the highest intensity in the two-dimensional feature chain CC _k , and then calculate the average mass fluctuation (the difference between adjacent mass-to-charge ratios) mzMin of its adjacent area;

(2) With 5 times mzMin as the threshold, find all positions where the quality fluctuation is greater than this threshold, and define these positions and the first point of CC _k as key points;

(3) These key points also correspond to key points on the chromatographic dimension. On the chromatographic dimension, connecting these key points with a straight line is the estimation of the baseline B(x). If the last key point is not the last point of CC _k , then the line extending horizontally from the key point to the end is the baseline estimate of the corresponding area.

4. Feature resolution: FeatureReslove(minWidth, minSN, feature_list)

After the local noise and baseline estimation of the two-dimensional feature chain CC _k is completed (referring to the current time, the feature chain may be extended in the subsequent time, and the corresponding noise and baseline will be re-estimated), feature discrimination can be performed. Due to the real-time nature of feature detection, generally only part of the features are eluted at that time. The purpose of feature resolution is to determine where the current time point is at the elution position of the chromatographic peak (feature): starting point, end point, etc. Subtracting the noise ε(x) and the baseline B(x) (where x is the time point) from the raw signal strength yields an approximate true signal estimate NS(x). If the feature discrimination is performed on CC _k for the first time, the feature detection state s=0 needs to be initialized. For the specific algorithm, refer to Embodiment 1. The detected features are saved in feature_list (feature list),

Define the signal-to-noise ratio at any point in CC _k as:

$\mathrm{<span\; class="notranslate">\; SN</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; LSD</span>}}$

Among them, LSD is the standard deviation around position x, the last point in CC _k is the data point collected currently, and its signal-to-noise ratio SN is calculated.

5. In the above four steps, each acquisition of a mass spectrum is a calculation cycle; each cycle only operates on the data that can enter the two-dimensional feature chain, and other signals are considered as noise; the maximum two-dimensional data processed at each time point The number of dimensional characteristic chains is the number of all ions in the previous mass spectrum, but in most cases it is far smaller than this value, which is one of the reasons for the fast operation of the algorithm. When all the mass spectrometry data are collected, the feature detection also ends accordingly, thereby realizing the real-time detection of features.

Advantage of the present invention is as follows:

(1) The two-dimensional feature chain fits the distribution characteristics of the chromatography-mass spectrometry data. Generally, the data volume of all two-dimensional feature chains in a data set only accounts for a small part (<1%) of the total data volume, which essentially improves the Efficiency of feature detection algorithms;

(2) The three-point high-pass filter designed by the present invention can accurately estimate the random noise in the chromatographic signal, and has the characteristic of constant variance;

(3) The baseline estimation method of the present invention utilizes the mass fluctuation information in the mass spectrum dimension, and overcomes the shortcoming that it is difficult to accurately estimate the baseline simply from the chromatographic dimension information;

(4) The algorithm designed by the present invention has few parameters and simple optimization, which has practical physical meaning, and a set of parameters can be applied to samples of different complexity;

(5) The algorithm designed in the present invention realizes the synchronization of sample collection and feature extraction, and is especially suitable for the analysis of a large number of samples in a digitized Chinese medicine component library.

Description of drawings

Figure 1 is a schematic diagram of LC-MS real-time feature extraction.

Figure 2 is the overlay of the analog signal (A) containing Gaussian white noise and different sampling frequencies (d), the signal after applying the high-pass filter (green line) and the original Gaussian white noise (blue line), where the dotted line is 3 times Standard deviation location (B).

Fig. 3 compares the estimation of the noise with the Savitzky-Golay smoothing algorithm of the present invention: when the A figure is that the sampling rate is from 1 to 20, the standard deviation of different algorithm estimates and the actual noise is compared; the B figure and the C figure are the sampling rate at 5 and 15, when the noise level is from 1% to 10%, the comparison of different algorithms; the blue line is the standard deviation of theoretical noise, the green line is the standard deviation estimated by the present invention, and the red line is the standard deviation estimated by Savitzky-Golay.

Figure 4 is an example of a two-dimensional feature chain from Weifuchun Tablets: Picture A is the time dimension of the two-dimensional feature chain, picture B is its quality dimension, and picture C is the relationship between quality fluctuation and time (the dotted line is 5 times mzMin) ; The red asterisks are key points, and the baselines are connected by green lines.

Figure 5 is the characteristic detection of naringenin and naringenin in Weifuchun Tablets: Figure A is the selected ion chromatogram of the quasi-molecular ions and isotope peaks of naringenin and naringenin; Figure B is the selected ion chromatogram of naringenin The two-dimensional feature regions of rutin and naringenin, the brown lines are two-dimensional feature chains, the detected "signatures" are indicated by green boxes, and the vertices are indicated by red asterisks.

Figure 6 is the total ion current chromatogram of Weifuchun Tablets (Panel A), the chromatogram reconstructed from the detected features (Panel B) and the chromatogram reconstructed from residual signal and noise (Panel C).

Figure 7 is a total ion chromatogram of Shuangdan granules.

Figure 8 is the characteristic detection of salvianolic acid E, salvianolic acid B and an unknown compound (m/z 719) in Shuangdan Granules: A is the selected ion chromatogram of m/z 719, and B is the selected ion chromatogram of m/z 718 Selected ion chromatogram, picture C is the selected ion chromatogram of m/z 717, picture D is the two-dimensional characteristic area of salvianolic acid E and salvianolic acid B, the brown line is the two-dimensional characteristic chain, and the detected " Features" are indicated by green boxes and vertices are indicated by red asterisks.

Fig. 9 is the total ion current chromatogram (Figure A) of Erigeron breviscapus injection, the base peak (base-peak) chromatogram

Figure (Panel B) and the chromatogram reconstructed from the detected features (Panel C).

Detailed ways

The present invention will be further described in conjunction with drawings and embodiments.

Embodiment 1 A kind of real-time feature extraction method for complex component analysis of traditional Chinese medicine of the present invention

1. Communication module: MS_Communication (acq_mode, cur_ms_data)

This function is responsible for communicating with the mass spectrometer. If the acquisition mode (acq_mode) is profile, when the current data is obtained from the mass spectrometer, it will be converted into centroid format using the watershed algorithm, and returned through the cur_ms_data parameter; if the acquisition mode is centroid, then Return the data directly. The parameter cur_ms_data is two-dimensional data containing the mass-to-charge ratio and its corresponding intensity.

2. Two-dimensional feature chain detection: BNN (minWidth, CC)

In the BNN module, by calling MS_Communication, the currently collected mass spectrum data can be obtained and assigned to the mass-to-charge ratio array MZ and the intensity array INTEN. For the data collected in sequence, use Bilateral Nearest Neighbor (BNN) to detect two-dimensional feature chains. The principle of the BNN algorithm is: sequentially take an ion MZ _{i, j} in the current mass spectrum (i is the scan number scan_number, which is equivalent to the i-th mass spectrum currently collected; j is the jth ion in MZ _i ), and then Find the ion MZ _{i-1, J} that is closest to its mass in the mass spectrum collected at the previous time point; if the ion closest to MZ _{i-1, J} in the current mass spectrum is also MZ _{i, j} , then connect MZ _i,j and MZ _i-1,J . With the increase of collected mass spectrometry data, some two-dimensional characteristic chains will be extended and some will be interrupted. Only the two-dimensional characteristic chains whose length len(CC _k ) is greater than minWidth will be considered as possibly containing real signals and stored in CC, otherwise, is considered noise. CC is a global variable that can be accessed by other modules.

3. Local noise and local baseline estimation: De_Noise_Baseline(minWidth)

When the length of a certain two-dimensional feature chain CC _k (k is the serial number of the detected feature chain) is greater than minWidth, the local noise and baseline can be estimated. The time dimension of a two-dimensional feature chain is equivalent to a chromatogram, which is generally considered to be composed of real signal, Gaussian white noise and baseline (F(x)=B(x)+NS(x)+ε(x)). Where Gaussian white noise ε(x) is estimated by linear convolution with the original signal and a three-point high-pass filter:

$\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CircleTimes;</span>\mathrm{<span\; class="notranslate">\; f</span>}$

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\sqrt{\mathrm{<span\; class="notranslate">\; 6</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; /</span>\sqrt{\mathrm{<span\; class="notranslate">\; 6</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\sqrt{\mathrm{<span\; class="notranslate">\; 6</span>}}<span\; class="notranslate">\; ]</span>$

From the simulated signal in Figure 2, it can be seen that when the sampling rate of the real signal is less than 5, some signals will remain in the chromatographic peak area, resulting in overestimation of the noise level in this area. The residual signal has the characteristics of impulsive noise, and its intensity is much larger than the overall standard deviation, so 3 times the overall standard deviation is used as the threshold, and the signal greater than this threshold is set to zero. The vector after the above convolution and threshold operation is the estimate of Gaussian white noise, which can accurately reflect the local variance of real white noise, as shown in Figure 2. By comparing different sampling rates and different noise levels, the noise estimation method of the present invention is very close to the actual value, which is better than the commonly used smoothing filtering method, and the corresponding results are shown in FIG. 3 .

The mass dimension of the two-dimensional characteristic chain reflects the corresponding mass fluctuation (Figure 4). When the compound is eluted, that is, when the real signal is detected, the corresponding mass fluctuation tends to a minimum value mzMin (this value is the same as that of the mass spectrometer mass precision), and in the area where there is no real signal, the mass fluctuation becomes a random feature, which is much larger than mzMin; at the same time, the area with the smallest mass fluctuation is also the area with the largest response intensity. The specific baseline estimation method is as follows:

(1) Find the position with the greatest intensity in CC _k , and then the vicinity of its corresponding position ("near" in the present invention means the area with the specified position as the center and a width of minWidth, or the area with a width of minWidth in front of the specified position) The mass quality fluctuation of is mzMin;

(2) With 5 times mzMin as the threshold (Figure 4C), find all positions where the mass fluctuation is greater than this threshold, and define these positions and the first point of CC _k as key points;

(3) These key points also correspond to the key points on the chromatographic dimension. On the chromatographic dimension, connecting these key points with a straight line is the estimation of the baseline B(x) (Fig. 4A). 4. Feature resolution on the time dimension: FeatureReslove(minWidth, minSN, feature_list)

Subtract the ε(x) and B(x) estimated in step 3 from the original signal to obtain an approximate true signal estimate NS(x), which still contains some residual components of irregular baseline fluctuations. Define the signal-to-noise ratio at any point in CC _k as:

$\mathrm{<span\; class="notranslate">\; SN</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; LSD</span>}}$

Among them, LSD is the standard deviation around position x, the last point in CC _k is the data point collected currently, and its signal-to-noise ratio SN is calculated. Use the linear least squares method to fit the last minWidth points of CC _k , define its slope slope as the slope of the last point, and then make the following judgment (if it is the first time to perform feature resolution on CC _k , you need to initialize the feature detection state s=0):

(1) If slope*minWidth>minSN, and s=0, then this is the beginning of a "feature", recorded in feature_list (feature list);

(2) If slope<0, set s=1;

(3) If slope*minWidth>-minSN, and s=1, then this is the end of a "feature", recorded in feature_list, and set s=0.

5. The algorithm of the present invention has real-time characteristics, and the data collected by mass spectrometry is analyzed immediately by modules such as BNN, and the detection of the characteristic starting point may be delayed by minWidth/f (about a few seconds), while the chromatographic peaks eluted from the chromatographic column are generally It is much longer than this time and does not affect the detection of features. The present invention uses VC++6.0 to realize the system prototype, and the user only needs to provide two parameters with actual physical meanings, minWidth and minSN.

Example 2 Analysis of Complex Components in Weifuchun Tablets

A. Preparation of total extract of Weifuchun Tablets

Take 20 Weifuchun Tablets, remove the film coating, and grind into fine powder. Accurately weigh 0.5g and place it in a 50mL Erlenmeyer flask with a stopper, accurately add 10mL of methanol, and conduct ultrasonic extraction for 45 minutes. After the extraction, take out the Erlenmeyer flask, and make up the weight with methanol solution after cooling. The extract was shaken and centrifuged at 12000rpm for 15min, and the supernatant was filtered through a 0.45μm filter membrane for HPLC analysis.

B. Chromatographic and mass spectrometric conditions for LC-MS analysis

The liquid phase is an Agilent1100 high performance liquid chromatograph (Agilent, USA), equipped with a binary gradient pump, a DAD ultraviolet detector, a column thermostat, and an autosampler. Chromatographic column: ZORBAX SB-C ₁₈ chromatographic column (4.6mm×250mm, 5μm, Agilent), front Agilent C ₁₈ pre-column. Mobile phase: A phase: 0.05% formic acid in water; B phase: acetonitrile. Linear elution gradient (min/%B): 0/5, 15/20, 30/20, 55/30, 75/50, 90/95. Flow rate: 0.5mL/min; column temperature: 30°C; injection volume: 10μL. The mass spectrometer was a Finnigan LCQ-DECA XP Plus ion trap mass spectrometer (Thermo, USA), equipped with an electrospray ion source and Xcalibur1.3 control system, and was detected by ESI negative ion mode. Scanning range: 100-1500Da; spray voltage: 4.5kV; sheath gas and auxiliary gas are nitrogen, 30 and 10 units respectively.

C. Feature detection parameters, the minimum peak width (minWidth) is 9, and the minimum signal-to-noise ratio (minSN) is 4.

D. Feature detection results: In the analysis time of 90 minutes, a total of 1827 features were detected, which accounted for 96.1% of the total variance. In the characteristic regions of naringenin rutin (t _R = 38min) and naringenin (t _R = 42.3min) in Figure 5, it can be seen that the two-dimensional characteristic chain of the present invention covers all possible regions of compound characteristics, not only those with high intensity The quasi-molecular ion [MH] ^- (m/z 579) can be detected correctly, and even the isotope peak [M-H+3] ^- (m/z 582) with extremely low abundance can also be detected correctly, which shows the detection sensitivity of this method very high.

In order to more intuitively evaluate the feature detection effect of the present invention, all detected features are reconstructed into time-dimensional chromatograms, and compared with the total ion current chromatograms composed of all signals, and the chromatograms composed of signals in non-characteristic regions are Noise or residual chromatograms, as shown in Figure 6. From Figure 6, it can be seen that almost all real signals are detected correctly, while there are no obvious characteristic signals in the remaining residual chromatograms.

Example 3 Shuangdan Granule Complex Component Analysis

A. Preparation of Shuangdan Granule Samples

Precisely weigh 0.05 g of Shuangdan granules after fine grinding (Shandong Kongshengtang Pharmaceutical Co., Ltd., batch number: 040201, 031001), add Wahaha purified water 1mL, ultrasonically extract for 20min, then centrifuge at 10000rpm for 10min, take 0.5mL of the supernatant, wash with methanol - Dilute 1-fold with water-formic acid (50:50:1).

B. Chromatographic and mass spectrometric conditions for LC-MS analysis

Agilent 1100 liquid chromatography system, including binary high-pressure pump, automatic sampler, column oven and DAD detector. Chromatographic column: Agilent SB-C18 (2.1×250mm, 3.5m). Mobile phase: 0.1% formic acid acetonitrile (A)-0.1% formic acid water (B), phase A rises linearly from 10% to 20% in 0-5min, 40% in 5-7min, and 95% in 7-20min %; flow rate 0.3mL/min, column temperature 35°C. All analyzed samples were injected in 10 L.

Finnigan ion trap mass spectrometer (LCQ Deca XP plus, CA), equipped with ESI ionization source; negative ion detection, sheath gas and auxiliary gas are both N ₂ , flow rates are 30 and 10arb, spray voltage 4.5kV, source fragmentation voltage 15V, The temperature of the heating capillary is 350°C, the scanning method is a full-level scan, and the scanning range is 100-800Da.

C. Feature Detection Parameters

Minimum peak width minWidth=9, minimum signal-to-noise ratio minSN=4.

D. Feature detection results

After the sample in Example 2 was eluted with a gradient of 90 minutes, the main components were well separated. In this case, the feature detection is relatively easy; The rapid gradient elution of the algorithm artificially compresses the features of multiple components together, which greatly increases the difficulty of feature detection, so as to examine the application of the algorithm under extreme conditions. It can be seen from Figure 7 that the main components in Shuangdan granules accumulate in the area of retention time of 10 to 13 minutes. By applying the same detection parameters as in Example 2, a good feature detection result can be obtained. A total of 510 features were detected, accounting for 98.5% of the variance of all signals. The following examples illustrate how the present invention detects incompletely separated components in complex systems.

When the mass-to-charge ratios of compounds in a complex system are different, even if the retention time is the same, they still have different features on the two-dimensional projection plane of LC/MS, which can be correctly detected by the present invention, which is the same as the result that the components are completely separated ; If the mass-to-charge ratios of different compounds are the same, multiple features will overlap. Fig. 8 is the quasi-molecular ion m/z 717 of salvianolic acid B (11.3min) and salvianolic acid E (10.9min), and the characteristic region of isotopic ion m/z 718,719 thereof. It can be seen from Figure 8A that an unknown component m/z 719 is inserted between the isotope ions of salvianolic acid B and salvianolic acid E, making the three features overlap together. The present invention can still correctly resolve such overlapping features, which are resolved into 3 distinct features. In addition, the peak shape of salvianolic acid B is seriously tailed, and the signal fluctuates greatly. Between the peak apex and the complete elution, there are many spurious peaks, which are analyzed by the peak detection algorithm (Avalon) that comes with the mass spectrometry workstation. , the chromatographic peak of salvianolic acid B is divided into 7 peaks, and the algorithm of the present invention can correctly detect these features with only two parameters.

Example 4 Analysis of Complex Components of Erigeron Injection

A. Analytical sample preparation

Precisely draw 0.5ml of Erigeron breviscapus injection, put the sample on the waters OASIS HLB solid-phase column that has been activated (1ml of methanol, 1ml of 1% formic acid water), wash with 0.5ml of 1% formic acid, discard the washing solution, add Wash with 0.5ml methanol, collect the eluate, and set aside.

B. Chromatographic and mass spectrometric conditions for LC-MS analysis

Agilent 1100 liquid chromatography system, including binary high pressure pump, automatic sampler, column thermostat and DAD detector. Chromatographic column: YMC-C ₁₈ 250mm×4.6mm, 5m; mobile phase: phase A: 0.1% formic acid water; phase B: 0.1% formic acid acetonitrile, linear elution gradient: 0min: 10%B; 20min: 17.5%B ; 40min: 17.5%B; 80min: 45%B; 90min: 45%B. Split ratio: 1:3. Column temperature: 35°C. Injection volume: 10L.

The mass spectrometer was a Finnigan LCQ-DECA XP Plus ion trap mass spectrometer (Thermo, USA), equipped with an electrospray ion source and Xcalibur1.3 control system, and was detected by ESI negative ion mode. ESI source voltage: 4.5kV; sheath gas (N ₂ ) flow rate: 30arb; auxiliary gas (N ₂ ) flow rate: 10arb; capillary temperature: 350°C; capillary voltage: -15V(-), 19V(+); Scanning mode, scanning range m/z: 100~800.

C. Feature Detection Parameters

Minimum peak width minWidth=9, minimum signal-to-noise ratio minSN=4.

D. Feature detection results

This example analyzes traditional Chinese medicine injections, which mainly contain water-soluble phenolic acids. Due to the mobile phase additives, a lot of high background chemical noise is generated, so that many low-intensity signals are submerged, and even low-abundance signals cannot be seen in the base-peak (base-peak) chromatogram, as shown in Figures 9A and 9B shown. Using the same feature detection parameters as those in the previous examples 2 and 3, a total of 571 features were detected. From the chromatogram reconstructed from these features, it can be found that there is no interference of high background noise, and not only the signal with high intensity is Correctly detected, low intensity signals also appear. This shows that the present invention can not only filter out randomly distributed white noise, but also automatically filter out colored noise with obvious heteroscedasticity.

Claims (5)

1. real-time feature extraction method that is used for the Chinese medicine analysis of complex ingredient, this method detects through the two dimensional character chain based on the time dimension of LC-MS and the bidimensional characteristic information of quality dimension, and local noise estimates with baseline, and characteristic differentiates realization, and concrete steps are:

(1) mass spectrometric data collection: the Chinese medicine complex sample at first separates through the chromatogram unit; Mass spectrometer is under certain SF f then; Order is with stream part of full scan pattern analysis chromatography eluant, and the data of collection are with the bar graph format, is a mass spectrogram with the data of each time point collection of 1/f integral multiple; Corresponding to the data of mass spectrum dimension, the data that different time point is gathered constitute chromatogram dimension information;

(2) the two dimensional character chain detects: mass spectrometer whenever collects the mass spectrogram of a time point; Promptly passing to the BNN module analyzes; At first mass-to-charge ratio in the mass spectrogram and strength information are distinguished assignment and are given mass-to-charge ratio array MZ and intensity array INTEN, detect the two dimensional character chain that contains compound information with two-way nearest neighbor algorithm according to time sequencing then; Detected two dimensional character chain is stored among the CC, can be obtained at any time by other modules;

(3) local noise and local baseline are estimated: along with increasing of image data, if certain the two dimensional character chain CC among the CC _kLength N _kGreater than minWidth; Then it is carried out the estimation of noise and baseline; The two dimensional character chain comprises chromatogram peacekeeping mass spectrum dimension double-point information, is made up of time and MZ and INTEN respectively, and the response intensity information and the Hi-pass filter of two dimensional character chain carried out linear convolution; And use 3 times of population standard deviations and filter out pulse signal; The noise that is the chromatogram dimension estimates that actual signal, white gaussian noise and baseline constitute: F (x)=B (x)+NS (x)+ε (x), and wherein white gaussian noise ε (x) carries out the linear convolution estimation with original signal and 3 Hi-pass filters:

$\mathrm{\&epsiv;}\left(x\right)=F\left(x\right)\&CircleTimes;f$

$f=[-1/\sqrt{6},2/\sqrt{6},-1/\sqrt{6}];$

(4) characteristic is differentiated:

As two dimensional character chain CC _kLocal noise and baseline estimate to accomplish after, be meant the current time, promptly carry out characteristic and differentiate, detected characteristic is kept at feature list,

Definition CC _kIn arbitrarily any signal to noise ratio (S/N ratio) be:

$\mathrm{SN}\left(x\right)=\frac{F\left(x\right)-B\left(x\right)-\mathrm{\&epsiv;}\left(x\right)}{\mathrm{LSD}}$

Wherein LSD is near the standard deviation the x of position, CC _kIn last point be the data point of current collection, calculate its signal to noise ratio (S/N ratio) SN;

(5) detect in real time: in above four steps, mass spectrogram of every collection is an execution cycle, and the phase is only carried out computing to the data that can get into the two dimensional character chain weekly, and other signals are considered to noise; The maximum two dimensional character chain number that each time point is processed is the number of all ions in the last mass spectrogram, when all mass spectrometric data finishing collecting, and the also corresponding end of feature detection, thus realized the real-time detection of characteristic.

2. a kind of real-time feature extraction method that is used for the Chinese medicine analysis of complex ingredient according to claim 1; It is characterized in that; Step (3) according to the difference of component district in the two dimensional character chain mass spectrum dimension with zero component district quality fluctuation, designs following algorithm in order to estimate the baseline in the chromatogram dimension:

(a) at two dimensional character chain CC _kIn find the maximum time point of intensity, calculate the average quality fluctuation of its close region then, i.e. the difference mzMin of adjacent mass-to-charge ratio;

(b) be threshold value with 5 times of mzMin, find the position of all quality fluctuations, these positions and CC greater than this threshold value _kFirst point be defined as key point;

(c) these key points, connect these key points on the chromatogram dimension also corresponding to the key point on the chromatogram dimension with straight line, are the estimation of baseline B (x), if last key point is not CC _kLast point, then this key point horizontal-extending line to the end baseline of being the corresponding region is estimated.

3. a kind of real-time feature extraction method that is used for the Chinese medicine analysis of complex ingredient according to claim 1; It is characterized in that; Step (4) is because the real-time of feature detection; Generally had only Partial Feature at that time by wash-out, the purpose that characteristic is differentiated judges that promptly current point in time is in the beginning or end position of chromatographic peak wash-out.

4. a kind of real-time feature extraction method that is used for the Chinese medicine analysis of complex ingredient according to claim 1 is characterized in that, the described Hi-pass filter of step (3) is made up of three data points, 3 and be that 0, three quadratic sum is 1.

5. a kind of real-time feature extraction method that is used for the Chinese medicine analysis of complex ingredient according to claim 1; It is characterized in that; Used chromatogram comprises liquid chromatography and UHV (ultra-high voltage) liquid chromatography, and mass spectrum comprises substance level Four bar mass spectrum, triple level Four bar mass spectrum, ion trap mass spectrometry and flight time mass spectrum.

Images