Intelligent detection method of chromatographic peaks based on deconvolution algorithm

Abstract

The application discloses a kind of intelligent detection methods of chromatographic peak based on deconvolution algorithm, it is related to chromatographic peak data detection field, the application includes chromatographic signal modeling as the linear superposition and convolution of multiple-component peak signal;The single-component signal separated is centered on local maximum, extracts the super-sampling standardization area as feature image block, is input to CNN, and CNN model outputs three parallel results: whether it is the confidence score of true peak, peak center position and peak boundary box regression, the application realizes the full-process automation intelligent analysis from original chromatographic data to quantitative peak table.The method can be widely applied in the automatic processing of chromatography-mass spectrometry data in metabolomics, environmental analysis, drug analysis, food safety detection and the like.

Description

Technical Field

[0001] This invention relates to the field of chromatographic peak data detection, and specifically to an intelligent detection method for chromatographic peaks based on a deconvolution algorithm. Background Technology

[0002] Traditional chromatographic peak detection relies on first / second derivative algorithms, which are prone to misjudgment when affected by factors such as retention time drift, matrix effects, peak compression / broadening; Machine learning-based methods learn from a large amount of peak shape data to automatically identify chromatographic peak characteristics, which can overcome the limitations of traditional mathematical algorithms, reduce false positive rates, and efficiently handle overlapping peaks.

[0003] Therefore, there is an urgent need for an intelligent detection method for chromatographic peaks based on deconvolution algorithms. Summary of the Invention

[0004] This invention provides an intelligent detection method for chromatographic peaks based on a deconvolution algorithm, which solves the problems of existing technologies.

[0005] In a first aspect, the present invention provides an intelligent detection method for chromatographic peaks based on a deconvolution algorithm, comprising: The chromatographic signal is modeled as a linear superposition and convolution of multi-component peak signals; The separated single-component signals are centered on local maxima, and the oversampled normalized regions are extracted as feature image blocks and input into a CNN. The CNN model outputs three parallel results: a confidence score for whether it is a true peak, the peak center position, and the peak bounding box regression.

[0006] Furthermore, the step of modeling the chromatographic signal as a linear superposition and convolution of multi-component peak signals specifically includes: Chromatographic observation signal y ( t () represents the peak signals of each component. Linear superposition and convolution: ; in, K For grouping, h ( t () represents the instrument point spread function. n ( t This is additive noise; Peak signal for each component Represented using an exponentially modified Gaussian model: ; in Peak area To retain the time, For Gaussian broadening parameters, It is an exponential decay constant; in the blind deconvolution framework, h ( t )and All of these are unknown quantities, and they are jointly estimated through iterative optimization.

[0007] Furthermore, it also includes: Acquisition and preprocessing of raw chromatographic data: Raw profile-mode data, including retention times, were obtained from the chromatography-mass spectrometry (GC-MS) system. The three dimensions are mass-to-charge ratio and signal strength, with the mass-to-charge ratio measured in m / z. Preprocessing includes A / D conversion, signal normalization, and data format standardization, including: The instrument outputs raw binary data streams or vendor format files. File format parsing and data reading: ProteoWizard's msconvert tool is used to uniformly convert the vendor format to the open-source mzML format, and the RT-m / z-Intensity three-dimensional matrix is ​​extracted. , T For the number of scan points, M This refers to the number of m / z channels; Signal intensity normalization: The total ion current of each m / z channel is normalized by min-max or Z-score to eliminate sensitivity differences between different injections; The total ion current chromatogram or the extracted ion chromatogram within a specified m / z range is used as a one-dimensional time series signal for subsequent processing. y ( n )( n =1,2,..., N Outputs a standardized one-dimensional chromatographic time series signal. y ( n ) and the corresponding retention timeline.

[0008] Furthermore, this includes the standardization of the output chromatographic signal. y ( n Baseline estimation and trend removal are performed. Baseline drift and low-frequency trend components in chromatographic signals are removed using asymmetric least squares or the BEADS algorithm. Baseline correction allows the preservation of the true peak shape. The specific processing method is as follows: Baseline estimation is performed using asymmetric least squares smoothing: given smoothing parameters Asymmetric weight parameters p Baseline fitting was performed using iterative weighted least squares. b ( nFor data with severe gradient drift, the BEADS algorithm is used to jointly estimate the baseline, noise, and peak signal; the signal after baseline correction: y c ( n )= y ( n )- b ( n ); The objective function of asymmetric least squares smoothing (AsLS) is: ; Weighting rules: y ( n )> b ( n )hour = p ,otherwise =1- p Output: Baseline-corrected chromatographic signal y c ( n and baseline estimates b ( n ).

[0009] Furthermore, this includes detecting all local signal maxima in the baseline-corrected chromatographic signal as candidate points for subsequent deconvolution and peak identification. The specific processing method is as follows: right y c ( n Perform first-order difference d ( n )= y c ( n +1)- y c ( n ), detect zero crossover points d ( n )>0→ d ( n +1)≤0, mark as a candidate point of local maximum, apply signal-to-noise ratio threshold and minimum peak height constraint to the candidate points for initial screening, mark the candidate points below the threshold as noise extremes with unacceptable signal-to-noise ratio or too low intensity and remove them; Candidate point set: C={ n : d ( n- 1)>0 and d ( n )≤0}; Candidate point set after signal-to-noise ratio filtering: C′={ n ∈C: yc ( n )> H min and y c ( n ) / s noise > SNR min}, SNR min The minimum signal-to-noise ratio threshold. s noise Based on signal fluctuation estimation in the baseline region, the output is: a set of candidate local maxima locations C′={ p 1, p 2,..., p L} and its corresponding strength value { A 1, A 2,..., A L}

[0010] Furthermore, this includes: for each candidate point p i Neighborhood window signal y c ( n ), n ∈[ p i - W , p i + W Blind deconvolution analysis is performed to separate overlapping peaks into independent single-component signals. The specific processing method is as follows: Input candidate points p i and its neighboring window signals y local ( n ), n ∈[ p i - W , p i + W ], where the window half-width W Satisfy 3 w max ≤2 W ≤6 w max ,in, w max The expected maximum peak width is 0.5–2.0 min, which is preset by the column efficiency parameters and flow rate. Initialization: Set candidate group scoresK =1, initialize the initial value of the first component peak parameter with the retention time and intensity corresponding to the local maximum value. ; Model fitting and residual analysis: in the current group of scores K Below, the window signal is fitted using nonlinear least squares fitting. K The objective function is optimized by superimposing several EMG peaks: ; Where Θ K ={ i 1,…, i K}for K The parameter set of each peak For EMG model; Calculate the residuals after fitting. ; Local extremum detection is performed on the residuals: if a structured residual peak signal exceeding three times the noise level still exists in the residuals, the position of that residual peak is used as the initial value for the new component, and then... K ← K +1, return to the model fitting and residual analysis steps to refit; Model selection: The Akaike Information Criterion or the Bayesian Information Criterion will be used in the competition of models. K ={1,2,…, K max Choose the optimal group score from among them: AIC K )= N ln(RSS / N )+2*4 K； Where RSS is the sum of squared residuals. N This represents the number of data points in the window; each peak group contains 4 parameters. For high overlap, i.e., separation R s For values ​​less than 0.5, the PENDANTSS framework is invoked for joint detrending and blind deconvolution, and a generalized quasi-norm is used to improve separation accuracy compared to SOOT / SPOQ sparse penalty terms. The nonlinear least squares algorithm uses the Levenberg-Marquardt algorithm: Θ (m+1) =Θ (m) -[J T J+ α* diag(J T J)] -1 J T r, where the elements of the Jacobian matrix J are Damping parametersα Automatically adjusts as iterative convergence occurs; Output the estimated number of components within the neighborhood window and the signal of each separated single component and its peak parameters ( , , , ).

[0011] Furthermore, it also includes a step of extracting oversampled normalized regions as feature image blocks: centering on the local maxima of the separated single-component signals, normalized feature image blocks are extracted along the two dimensions of retention time and mass-to-charge ratio, transforming one-dimensional chromatographic information into a two-dimensional image representation, which is then processed by a CNN model. Specifically: by p k Centered on the RT dimension, the width is taken as... The expected range of isotope distribution in the m / z dimension [ m / z theo -0.5, m / z theo +3.0], the extracted region is resampled to a fixed image size using bilinear interpolation, and the intensity value of each pixel is standardized so that the overall image has a mean of 0 and a standard deviation of 1; Among them, oversampling interpolation: for the pixel coordinates of the target image ( i ′, j ′) corresponds to the original data coordinates ( t ′, m ′), interpolate from the four adjacent points of the original grid using distance weighting: ; in, w ab =(1-∣ t ′- t a ∣ / Δ t (1-∣) m ′- m b ∣ / Δ m The standardized output image is as follows: Output: Standardized feature image patch.

[0012] Furthermore, this includes: inputting standardized feature image patches into a pre-trained convolutional neural network (CNN), where the model automatically determines whether candidate locations are true chromatographic peak vertices, and simultaneously regresses and predicts the peak's center position and bounding box. The specific processing steps are as follows: Forward inference is performed using a pre-trained CNN model, outputting the results of three parallel branches: (i) peak / non-peak binary classification probabilities. P (peak); (ii) the offset of the peak center position (Δ) t R ,Δ m / z (iii) Bounding box offset (Δ) left ,Δ right ,Δ low ,Δ high ); Confidence score of output peak P (peak), precise positioning coordinates ( t R +Δ t R , m / z +Δ m / z ) and bounding box information.

[0013] Furthermore, this includes: for peaks identified as true peaks, but deconvolution residual analysis showing insufficient separation, i.e. Rs. For overlapping peak regions <0.8, the GCMSFormer model based on the Transformer architecture is used for end-to-end deep analysis to directly predict the purity mass spectra of each component from the signals in the overlapping peak regions. The specific operation is as follows: Signal regions identified as overlapping peaks by CNN y overlap ( n The system encodes the regional signals of overlapping peaks into sequence representations, which are then input into the self-attention encoder of the GCMSFormer model. A multi-head attention mechanism is used to capture long-range dependencies within the signal, automatically learning the interaction patterns between different components in the co-emission peaks. The decoder generates pure mass spectrometry sequences for each component, predicting the mass spectrometry fragment, m / z, and abundance pairs for each component token by token. An integrated orthogonal projection resolution module orthogonally projects and removes the resolved major components from the original mixed signal, thereby resolving low-concentration minor components. The Transformer model incorporates self-attention, multi-head attention, feedforward networks, layer normalization, and positional encoding. ; in, Q The query matrix represents the query vector emitted by each position of the currently focused target or to be encoded. In self-attention, Q A linear transformation derived from the input sequence, used to... K Matching to calculate attention weights; K The key matrix represents the key vector at each position in the sequence, used to calculate similarity and determine the importance of each position to the current query. V The Value matrix represents the value vector at each position in the sequence; attention weights will be used to... V Perform a weighted summation to obtain the final output; dk : The dimension of the key vector, used to scale the dot product result; softmax: Normalized exponential function; The OPR step, or Integrated Orthogonal Projection Resolution, is used to orthogonally project and remove the resolved principal components from the original mixed signal, leaving the residual signal with the unresolved minor components; specifically, the residual signal matrix Y... residual =(IS(S T S) -1 S T )y overlap Where S is the mass spectrum column vector of the resolved principal components, I is the identity matrix, and y is the mass spectrum column vector of the resolved principal components. overlap : Signal vector of the original overlapping peak region.

[0014] Furthermore, this includes: the confidence score of the output peak. P (peak), precise positioning coordinates ( t R +Δ t R , m / z +Δ m / z After obtaining the peak area integration and bounding box information, the following steps are performed: Peak area integration, post-processing, and quantitative analysis. Peak integral: Based on the bounding box information output by the CNN, the peak area is numerically integrated on the baseline-corrected signal based on the EMG model parameters; Peak alignment: For peaks detected among multiple parallel samples, a retention time alignment algorithm is used to match peaks and eliminate the influence of retention time drift on qualitative analysis. False positive suppression: Chemical rule filtering based on isotope distribution matching, adduct ion regularity, and retention time rationality; Trapezoidal integral of peak area: ; Or based on the analytical integral of the EMG model fitting results: ; That is, the parameters in the EMG model Direct output; Compound identification was performed by two-way matching with a standard spectral library, and peaks with scores all exceeding 800 or matching factors > 85% / probability > 80% were selected for confirmation. Output: Quantitative peak table, including retention time, m / z, peak area, peak width, asymmetry factor, standard spectral library matching results, and confidence score.

[0015] The difference between this invention and existing technologies lies in the following: traditional mathematical algorithms (first / second derivative methods) are prone to failure when dealing with peak compression / broadening and baseline drift; pure machine learning methods (such as PeakBot) directly detect local maxima and use CNN classification, but do not perform blind deconvolution of the signal before recognition, resulting in limited ability to separate severely overlapping peaks; traditional chemometric deconvolution methods (such as MCR-ALS) rely on manual intervention for component quantity estimation. This invention achieves an organic integration of blind deconvolution separation and deep learning recognition at the feature level, systematically solving the problem of peak detection robustness under complex chromatographic conditions (peak drift, overlap, distortion, matrix interference).

[0016] This invention provides an intelligent peak detection method based on a deconvolution algorithm, realizing fully automated intelligent analysis from raw chromatographic data to a quantitative peak table. This method can be widely applied to the automated processing of chromatographic-mass spectrometry data in fields such as metabolomics, environmental analysis, drug analysis, and food safety testing. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention. In the drawings: Figure 1 The flowchart illustrates an intelligent detection method for chromatographic peaks based on a deconvolution algorithm, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention.

[0019] Technical concept of the present invention: This invention utilizes a classic blind deconvolution algorithm to separate overlapping chromatographic signals into independent single-component signals. A convolutional neural network is then employed to distinguish true / false peaks and regress peak parameters in the separated signals, enabling precise peak location and identification in complex chromatograms. This invention is the first to integrate blind deconvolution signal separation with a deep learning peak recognition system, forming a closed-loop intelligent detection process of "separation → discrimination → parameter regression," which differs from existing single-stage improvement schemes.

[0020] This invention provides an intelligent detection method for chromatographic peaks based on a deconvolution algorithm, which aims to solve the above-mentioned technical problems in the prior art.

[0021] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] Example 1: The following steps are performed to achieve this. This example includes two stages: The first stage is blind deconvolution separation: The chromatographic signal is modeled as a linear superposition and convolution of multi-component peak signals, specifically including: chromatographic observation signal y ( t () represents the peak signals of each component. Linear superposition and convolution: ; in, K The number of groups (unknown). h ( t () represents the instrument point spread function. n ( t This represents additive noise; the peak signal for each component... The following is represented using an Exponentially Modified Gaussian (EMG) model: ; in Peak area To retain the time, For Gaussian broadening parameters, It is an exponential decay constant; in the blind deconvolution framework, h ( t )and All of these are unknown quantities, and they are jointly estimated through iterative optimization.

[0023] Phase Two: Intelligent Peak Identification The separated single-component signals are centered on local maxima, and the oversampled normalized regions are extracted as feature image patches (retention-time vs. m / z), which are then input into a custom convolutional neural network (CNN). The CNN model outputs three parallel results: a confidence score for whether it is a true peak (binary classification), the precise location of the peak center (regression), and the peak bounding box regression, thereby enabling the screening of true chromatographic peaks from candidate local extrema and the elimination of baseline fluctuations and noise spurious peaks.

[0024] For cases of particularly severe overlapping peaks in GC-MS data, the GCMSFormer model based on the Transformer architecture is introduced to directly predict the pure mass spectra of each component from the original overlapping peak data, and orthogonal projection resolution (OPR) technology is integrated to resolve minor components.

[0025] The two-phase implementation method has the following steps, as follows: Figure 1 As shown; Step 1: Acquisition and preprocessing of raw chromatographic data; Raw profile-mode data, including retention time (RT), mass-to-charge ratio (m / z), and signal intensity, were obtained from chromatography-mass spectrometry (LC-MS, GC-MS). Preprocessing included A / D conversion, signal normalization, and data format standardization.

[0026] Input: Raw binary data stream or vendor format file (.d, .raw, .wiff, etc.) output by the instrument; Step 1 includes: Step 1a, File Format Parsing and Data Reading: Use ProteoWizard's msconvert tool to convert the vendor format to the open-source mzML format and extract the RT-m / z-Intensity three-dimensional matrix. , T For the number of scan points, M This refers to the number of m / z channels; Step 1b, Signal intensity normalization: The total ion current of each m / z channel is normalized by min-max or Z-score to eliminate sensitivity differences between different injections; Step 1c: Extract the total ion current chromatogram or the extracted ion chromatogram within a specified m / z range, as a one-dimensional time series signal for subsequent processing. y ( n )( n =1,2,..., N ); Output: Standardized one-dimensional chromatographic time series signal y (n ) and the corresponding retention timeline; Min-max normalization: ; Z-score standardization: ,in m y The mean, s y Standard deviation; Step 2: Baseline estimation and trend removal; Asymmetric least squares or the BEADS algorithm is used to remove baseline drift and low-frequency trend components from the chromatographic signal. Baseline correction allows the true peak shape of the chromatographic peaks to be preserved, avoiding peak area calculation errors caused by baseline rise or fall (especially under gradient elution conditions).

[0027] Input: The standardized chromatographic signal output from step 1 y ( n ); Step 2 includes: Step 2a: Baseline estimation using asymmetric least squares smoothing (AsLS): given smoothing parameters Asymmetric weight parameters p Baseline fitting was performed using iterative weighted least squares. b ( n ); Step 2b: For data with severe gradient drift, use the BEADS (Baseline Estimation and Denoising with Sparsity) algorithm to jointly estimate the baseline, noise, and peak signal; Step 2c, the signal after baseline correction: y c ( n )= y ( n )- b ( n ); The objective function of asymmetric least squares smoothing (AsLS) is: ; Weighting rules: y ( n )> b ( n )hour = p ,otherwise =1- p .

[0028] Output: Baseline-corrected chromatographic signal y c ( n and baseline estimates b ( n ).

[0029] Step 3: Preliminary detection of local extreme points; All local maximum values ​​(LocalMaxima) in the baseline-corrected chromatographic signal were detected as candidate points for subsequent deconvolution and peak identification.

[0030] Step 3a, for y c ( n Perform first-order difference d ( n )= y c ( n +1)- y c ( n ); Step 3b: Detect zero-crossing points d ( n )>0→ d ( n If +1)≤0, mark it as a candidate point for a local maximum. Step 3c: Apply a signal-to-noise ratio threshold and a minimum peak height constraint to the candidate points for initial screening. Candidate points below the threshold are marked as noise extremes with unacceptable signal-to-noise ratio or too low intensity and removed. Wherein, the candidate point set: C={ n : d ( n -1)>0 and d ( n )≤0}; Candidate point set after signal-to-noise ratio filtering: C′={ n ∈C: y c ( n )> H min and y c ( n ) / s noise > SNR min}, SNR min The minimum signal-to-noise ratio threshold. s noise Estimated from signal fluctuations in the baseline region; Output: Set of candidate local maxima locations C′={p 1, p 2,..., p L} and its corresponding strength value { A 1, A 2,..., A L}

[0031] Step 4: Blindly deconvolution to separate overlapping peaks; For each candidate point p i Neighborhood window signal y c ( n ), n ∈[ p i - W , p i + W Blind deconvolution analysis is performed to separate the existing overlapping peaks into independent single-component signals; Input: Candidate points p i and its neighboring window signals y local ( n ), n ∈[ p i - W , p i + W ], where the window half-width W Satisfy 3 w max ≤2 W ≤6 w max ,in, w max The expected maximum peak width is 0.5–2.0 min, which is preset by the column efficiency parameters and flow rate. Step 4a, Initialization: Set candidate group scores K =1, initialize the initial value of the first component peak parameter with the retention time and intensity corresponding to the local maximum value. ; Step 4b, Model Fitting and Residual Analysis: In the current group of... K Below, the window signal is fitted using nonlinear least squares fitting. K The objective function is optimized by superimposing several EMG peaks: ; Where Θ K ={ i 1,…, i K}for K The parameter set of each peak For EMG model; Step 4c: Calculate the residuals after fitting. ; Step 4d: Perform local extremum detection on the residuals: If a structured residual peak signal exceeding three times the noise level still exists in the residuals, then the position of this residual peak is used as the initial value for the new component, and let... K ← K +1, return to step 4b to refit; Step 4e, Model Selection: The Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) is used in the competition of models. K ={1,2,…, K max Choose the optimal group score from among them: AIC K )= N ln(RSS / N )+2*4 K； Where RSS is the sum of squared residuals. N This represents the number of data points in the window; each peak group contains 4 parameters. Step 4f, PENDANTSS joint optimization: For high overlap, i.e., separation... R s For values ​​less than 0.5, the PENDANTSS framework is invoked for joint detrending and blind deconvolution, and a generalized quasi-norm is used to improve separation accuracy compared to SOOT / SPOQ sparse penalty terms. The nonlinear least squares algorithm uses the Levenberg-Marquardt algorithm: Θ (m+1) =Θ (m) -[J T J+ α* diag(J T J)] -1 J T r, where the elements of the Jacobian matrix J are Damping parameters α Automatically adjusts as iterative convergence occurs; Output: Estimated number of components within the neighborhood window and the signal of each separated single component and its peak parameters ( , , , ).

[0032] Step 5: Extraction of oversampled feature image patches; For each single-component signal separated in step 4, standardized feature image patches are extracted along the two dimensions of retention time and mass-to-charge ratio, centered on its local maximum value. The one-dimensional chromatographic information is transformed into a two-dimensional image representation and input into the CNN model for processing. Input: Separated single-component signals and the location of its local maximum value p k ; Step 5a, with p k Centered on the RT dimension, the width is taken as... (approximately ±3) s (Covering 99.7% of the peak area), taking the expected range of isotopic distribution in the m / z dimension [ m / z theo -0.5, m / z theo +3.0]; Step 5b: Use bilinear interpolation to resample the extracted region to a fixed image size (e.g., 64×64 pixels or 128×64 pixels). Step 5c: Standardize the intensity value of each pixel so that the overall image has a mean of 0 and a standard deviation of 1. Among them, oversampling interpolation: for the pixel coordinates of the target image ( i ′, j ′) corresponds to the original data coordinates ( t ′, m ′), interpolate from the four adjacent points of the original grid using distance weighting: ; in, w ab =(1-∣ t ′- t a ∣ / Δ t (1-∣) m ′- m b ∣ / Δ m The standardized output image is as follows: (Mean normalized to zero, variance normalized); Output: Normalized feature image patch .

[0033] Step 6: CNN model inference, true / false peak discrimination; Input: The normalized 64×64 feature image patch output from step 5 The initial peak parameter values ​​from the deconvolution output in step 4 are fed into a pre-trained convolutional neural network (CNN). The model automatically determines whether the candidate position is the true chromatographic peak vertex and simultaneously regresses and predicts the peak's center position and bounding box. The specific processing steps are as follows: the pre-trained CNN model is used for forward inference, outputting the results of three parallel branches: (i) peak / non-peak binary classification probability. P (peak); (ii) the offset of the peak center position (Δ) t R ,Δ m / z (iii) Bounding box offset (Δ) left ,Δ right ,Δ low ,Δ high ); CNN network architecture design: The CNN uses the ResNet residual architecture, containing four residual modules. Each module contains two 3×3 convolutional layers (batch normalization + ReLU activation), for a total of approximately 20 layers. The input layer accepts a 64×64×1 single-channel image. After feature extraction by the four residual modules, it is compressed into a one-dimensional feature vector using global average pooling, and finally branches into three fully connected output heads. Classification head (2 neurons, Softmax activation): Output [ P (noise) P (peak)]; Localization head (2 neurons, linearly activated): Output (Δ) t R ,Δ m / z ); Bounding box head (4 neurons, linear activation): Output (Δ) left ,Δ right ,Δ low ,Δ high ); The total loss function of CNN uses a combination of multiple tasks: ; in L CE Cross-entropy classification loss, L MSE The mean squared error regression loss, weighting coefficients ( l 1, l 2, l 3) = (1.0, 0.5, 0.3). The model training uses at least 100 reference features to learn the peak shape feature, and can achieve a discrimination accuracy of 0.99 on the test set; Output: Peak confidence score P (peak), precise positioning coordinates ( t R +Δ t R , m / z +Δ m / z ) and bounding box information.

[0034] Step 6.1: Deep learning-based secondary analysis of overlapping peaks; For peaks identified as true peaks, but deconvolution residual analysis shows insufficient separation, i.e. R s For overlapping peak regions <0.8, the GCMSFormer model based on the Transformer architecture is used for end-to-end deep analysis to directly predict the purity mass spectra of each component from the signals in the overlapping peak regions. The specific operation is as follows: Step 6.1a: Encode the regional signals of overlapping peaks into sequence representations and input them into the self-attention encoder of the GCMSFormer model. Use the multi-head attention mechanism to capture the long-range dependencies within the signal and automatically learn the interaction patterns between different components in the co-eluent peaks.

[0035] Step 6.1b: The decoder generates pure mass spectrum sequences for each component, i.e., predicts the mass spectrum fragment (m / z and abundance pair) for each component token by token.

[0036] Step 6.1c: Integrate the Orthogonal Projection Resolution (OPR) module to orthogonally project and remove the resolved major components from the original mixed signal, and then resolve the low-concentration minor components to avoid the masking effect of the major component spectral peaks on the trace component signals.

[0037] The Transformer model incorporates self-attention, multi-head attention, feedforward networks, layer normalization, and positional encoding. ; OPR step, or orthographic projection resolution integration: Y residual =(IS(S T S) -1 S T )y overlap Where S is the mass spectrum column vector of the resolved principal components; Output: Pure mass spectra of each component and corresponding chromatographic peak shape parameters.

[0038] Step 7: Peak area integration, post-processing, and quantitative analysis; For each peak confirmed as true in step 6 (and further analyzed in step 6.1), the peak area is calculated by precise integration, and peak alignment, compound identification and quantitative analysis are completed; Input: All confirmed chromatographic peaks, including precise location coordinates, bounding box information, and peak parameters; Step 7a, Peak Integration: Based on the bounding box information output by the CNN, numerically integrate the peak area on the baseline-corrected signal based on the EMG model parameters; Step 7b, Peak Alignment: For peaks detected among multiple parallel samples, a retention time alignment algorithm is used to match peaks and eliminate the influence of retention time drift on qualitative analysis. Step 7c, False Positive Suppression: Based on isotope distribution matching, adduction ion regularity, and retention time rationality, chemical rule filtering is performed to further reduce the system's false positive rate.

[0039] Trapezoidal integral of peak area: ; Or based on the analytical integral of the EMG model fitting results: ; That is, the parameters in the EMG model Direct output; Compound identification was performed by two-way matching with a standard spectral library, and peaks with scores all exceeding 800 or matching factors > 85% / probability > 80% were selected for confirmation. Output: Quantitative peak table, including retention time, m / z, peak area, peak width, asymmetry factor, standard spectral library matching results, and confidence score.

[0040] Table 1. Range of Key Values

[0041] Example 2: Based on the method of Example 1, in metabolomics research, intelligent detection of overlapping peaks in serum metabolomics UHPLC-QTOF data is performed. Due to the complex matrix of biological samples, polar small molecules are prone to co-elution peaks on reversed-phase chromatography. At the same time, the concentration differences of endogenous metabolites are huge, and high-abundance peaks often mask low-abundance substances.

[0042] Instrument: Agilent 1290 Infinity II U HPLC tandem with 6545 QTOF mass spectrometer; Column: ACQUITYUPLCHSST3 (2.1×100mm, 1.8μm); Mobile phase: A = water (0.1% formic acid), B = acetonitrile (0.1% formic acid); gradient: 0-2 min 5% B, 2-20 min 5% → 95% B; Mass spectrometry: ESI positive ion mode, scan range m / z 50-1000, profile mode; Sample: Supernatant from serum protein precipitation in healthy individuals; The instrument's original .d file was converted to mzML format using ProteoWizard, and the total ion current chromatogram (TIC) was extracted. The TIC signal was then analyzed. y ( n Z-score standardization is performed. y norm ( n )=[ y ( n )- m y ] / s y ,in m y =1.2×10 5 counts s y =8.5×10 4 counts. Output length is N A one-dimensional signal with 12000 points (retention time 0-20 min, sampling rate 10 Hz) was analyzed using the asymmetric least squares (AsLS) method with appropriate smoothing parameters. l b =10 6 asymmetric weights p =0.001. After 15 iterations and convergence, the baseline estimate is obtained. b ( n ) and correction signal y c ( n Approximately 5000 counts of baseline drift caused by gradient elution were successfully subtracted. First-order difference was calculated. d ( n )= y c ( n +1)- y c ( n Extracting what satisfies d ( n -1)>0 and d ( n Points with a signal-to-noise ratio (SNR) ≤ 0 are considered candidates. A SNR threshold is set. S / N ≥3.0, noise standard deviation s noise The peakless interval of 0.5–1.0 min was estimated (approximately 120 counts). 342 candidate local maxima points were initially obtained. A window was then taken centered on each candidate point.W =6 s wide ( s wide The preset timeframe is 0.1 min, corresponding to a peak width of approximately 0.4 min. The actual window size is ±0.3 min, with a total of 360 data points. An EMG model is used: initial component count... K =1, parameter i 1=[ A 1= y c ( p i ), t R1 = t ( p i ), s 1 = 0.05 min, =0.04min]. Using the Levenberg-Marquardt algorithm for nonlinear least-squares fitting, the initial damping coefficient... α =0.01, maximum iterations 200. Calculate the residuals after fitting. r ( n If the residuals still contain values ​​exceeding 3, s noise The peak-like structure adds a new component at the residual maximum. K +1, repeat fitting. Finally, for K The fitting results for =1,2,3 were used to calculate AIC, as shown in Table 2. Table 2. AIC calculation table based on fitting results

[0043] Select the minimum AIC corresponding to K =2, confirming the presence of two overlapping peaks within the window. Separation effect: The original signal in t R At 5.24 min, a single peak was observed, and component A was separated after deconvolution. t R1 =5.22min, A 1 = 6.8 × 10 4 (peak width 0.18 min) and component B ( t R2 =5.30min, A 2 = 1.3 × 10 4 The peak width was 0.17 min, and the retention time difference between the two was only 0.08 min, indicating good resolution. R s =0.47. For the signal of the separated component A, the retention time... t R1and precise quality m / z Centered at 188.0706 (confirmed by mass spectrometry bar chart), ±3 is taken in the RT dimension. σRT =±0.15min, taking [ in the m / z dimension] m / z -0.3, m / z [+1.2], bilinear interpolation is used to resample the image to 64×64 pixels, followed by mean normalization and variance normalization. The 64×64 image is then fed into a pre-trained 20-layer ResNet network. This network achieves 0.991 AUC on a training set containing 50,000 manually labeled peak and noise samples; the network outputs three branches: classification probability... P (peak) = 0.987 (≥ 0.5 threshold, determined to be a true peak), center offset Δ t R =-0.002min,Δ m / z =+0.0003, bounding box offset: left boundary -0.06min, right boundary +0.07min; baseline correction signal within the boundary range. y c ( n Trapezoidal integration yields an area of ​​component A of 6.78 × 10⁻⁶. 4 Component B has an area of ​​1.28 × 10⁻⁶. 4 All sample peaks were aligned using the OWarp algorithm and matched with the HMDB spectral library (forward score 856, reverse score 892), identifying them as hippuric acid and phenylacetylglycine. Using this method, the false positive rate decreased from 8.7% using the traditional derivative method (Chromeleon) to 1.8%, and the detection rate of overlapping peaks increased from 61% to 94%.

[0044] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or systems. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0046] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0047] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

[0048] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the invention disclosed herein in the specification and examples. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the foregoing claims.

[0049] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A smart detection method for chromatographic peaks based on a deconvolution algorithm, characterized in that, It includes two stages: The first stage is blind deconvolution separation: The chromatographic signal is modeled as a linear superposition and convolution of multi-component peak signals. Phase Two: Intelligent Peak Identification The separated single-component signals are centered on local maxima. The oversampled and normalized regions are extracted as feature image blocks and input into a customized convolutional neural network. The CNN model outputs three parallel results: a confidence score for whether it is a true peak, the precise location of the peak center, and the peak bounding box regression. Specifically, the steps include the following: Step 1: Acquisition and preprocessing of raw chromatographic data; Input: Raw binary data stream or vendor format file output by the instrument; Raw profile-mode data were obtained from the chromatography-mass spectrometry system, including three dimensions: retention time (RT), mass-to-charge ratio, and signal intensity. Output: Standardized one-dimensional chromatographic time series signal and corresponding retention time axis; Step 2: Baseline estimation and trend removal; Input: The standardized chromatographic signal output from step 1; Use asymmetric least squares or the BEADS algorithm to remove baseline drift and low-frequency trend components from chromatographic signals; Output: Baseline-corrected chromatographic signal and baseline estimate; Step 3: Preliminary detection of local extreme points; In the baseline-corrected chromatographic signal, all local signal maxima are detected as candidate points for subsequent deconvolution and peak identification. The output is: a set of candidate local maxima locations and their corresponding intensity values. Step 4: Blindly deconvolution to separate overlapping peaks; Input: Candidate points and their neighborhood window signals; Blind deconvolution analysis is performed on the neighborhood window signal of each candidate point to separate the existing overlapping peaks into independent single-component signals. Output: The estimated number of components within the neighborhood window, and the signal of each isolated single component and its peak parameters; Step 5: Extraction of oversampled feature image patches; For each single-component signal separated in step 4, standardized feature image patches are extracted along the two dimensions of retention time and mass-to-charge ratio, centered on its local maximum value. The one-dimensional chromatographic information is transformed into a two-dimensional image representation and input into the CNN model for processing. Output: Standardized feature image patch; Step 6: CNN model inference, true / false peak discrimination; Input: The standardized feature image patch output from step 5 and the initial peak parameter values ​​output from deconvolution in step 4 are fed into a pre-trained convolutional neural network (CNN). The model automatically determines whether the candidate position is the true chromatographic peak apex and regresses to predict the center position and bounding box of the peak. Output: The purity mass spectra of each component, the confidence score of the peak, the precise location coordinates, and the bounding box information. Step 7: Peak area integration, post-processing, and quantitative analysis; Input: All confirmed true chromatographic peaks, including precise location coordinates, bounding box information, and peak parameters; For each peak confirmed true in step 6, perform precise integration to calculate the peak area, and complete peak alignment, compound identification, and quantitative analysis; Output: Quantitative peak table, including retention time, m / z, peak area, peak width, asymmetry factor, standard spectral library matching results, and confidence score.

2. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 1, characterized in that, The method of modeling chromatographic signals as a linear superposition and convolution of multi-component peak signals specifically includes: The chromatographic observation signal y(t) represents the peak signals of each component. Linear superposition and convolution: ； Where K is the number of components, h(t) is the instrument point spread function, and n(t) is the additive noise; Peak signal for each component Represented using the Exponentially Modified Gaussian (EMG) model: ； in Peak area To retain the time, For Gaussian broadening parameters, It is the exponential decay constant; In the blind deconvolution framework, h(t) and All of these are unknown quantities, and they are jointly estimated through iterative optimization.

3. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 2, characterized in that, Also includes: Acquisition and preprocessing of raw chromatographic data: Raw profile-mode data were obtained from the chromatography-mass spectrometry system, including three dimensions: retention time, mass-to-charge ratio, and signal intensity. The mass-to-charge ratio was measured in m / z. Preprocessing includes A / D conversion, signal normalization, and data format standardization, including: The instrument outputs raw binary data streams or vendor format files. File format parsing and data reading: ProteoWizard's msconvert tool is used to convert the vendor format to the open-source mzML format, extracting the RT-m / z-Intensity 3D matrix. T is the number of scan points, and M is the number of m / z channels; Signal intensity normalization: The total ion current of each m / z channel is normalized by min-max or Z-score to eliminate sensitivity differences between different injections; Extract the total ion chromatogram or the extracted ion chromatogram within a specified m / z range as a one-dimensional time series signal y(n) for subsequent processing, n=1,2,...,N, and output the standardized one-dimensional chromatographic time series signal y(n) and the corresponding retention time axis.

4. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 3, characterized in that, This includes baseline estimation and trend removal of the output standardized chromatographic signal y(n): Baseline drift and low-frequency trend components in chromatographic signals are removed using asymmetric least squares or the BEADS algorithm. Baseline correction allows the preservation of the true peak shape. The specific processing method is as follows: Baseline estimation is performed using asymmetric least squares smoothing: given smoothing parameters Asymmetric weight parameter p is used to fit the baseline b(n) using iterative weighted least squares; for data with severe gradient drift, the BEADS algorithm is used to jointly estimate the baseline, noise, and peak signal; the signal after baseline correction: y c (n) = y(n) - b(n); The objective function of asymmetric least squares smoothing (AsLS) is: ； Weight setting rule: when y(n) > b(n) =p, otherwise =1-p, Output: Baseline-corrected chromatographic signal y c (n) and baseline estimate b(n).

5. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 4, characterized in that, include: In the baseline-corrected chromatographic signal, all local signal maxima are detected as candidate points for subsequent deconvolution and peak identification. The specific processing method is as follows: For y c (n) Perform a first-order difference d(n) = y c (n+1)-y c (n), detect zero crossover points d(n)>0→d(n+1)≤0, mark them as local maximum candidate points, apply signal-to-noise ratio threshold and minimum peak height constraint to candidate points for initial screening, mark candidate points below the threshold as noise extremes with unacceptable signal-to-noise ratio or too low intensity and remove them; Candidate point set: C = {n: d(n-1) > 0 and d(n) ≤ 0}; Candidate point set after signal-to-noise ratio filtering: C′={n∈C:y c (n)>H min And y c (n) / σ noise >SNR min }, SNR min σ is the minimum signal-to-noise ratio threshold. noise Based on signal fluctuation estimation in the baseline region, the output is: a set of candidate local maxima locations C′={p1,p2,...,p L } and their corresponding intensity values ​​{A1,A2,...,A} L }, where H min This represents the minimum peak height.

6. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 5, characterized in that, include: For each candidate point p i The neighborhood window signal y c (n), n∈[p i -W,p i Blind deconvolution analysis is performed using +W] to separate the existing overlapping peaks into independent single-component signals. The specific processing method is as follows: Input candidate point p i and its neighborhood window signal y local (n), n∈[p i -W,p i +W], where the window half-width W satisfies 3w max ≤2W≤6W max , where w max The expected maximum peak width is 0.5–2.0 min, which is preset by the column efficiency parameters and flow rate. Initialization: Set the candidate group score K=1, and initialize the initial value of the first group peak parameter θ1=(A1, t) with the retention time and intensity corresponding to the local maximum. R1 ,σ1, 1); Model Fitting and Residual Analysis: Given the current number of components K, the window signal is fitted to the superposition of K EMG peaks using nonlinear least squares fitting, and the objective function is optimized. ； Where Θ K ={θ1,…,θ K } represents the parameter set of K peaks. For EMG model; Calculate the residuals after fitting. ; Local extremum detection is performed on the residuals: if there are still structured residual peak signals in the residuals that are more than 3 times the noise level, the position of the residual peak is used as the initial value of the new component, K←K+1 is set, and the model fitting and residual analysis steps are returned to refit. Model selection: The Akaike Information Criterion or the Bayesian Information Criterion is adopted in the competing model K={1,2,…,K max Choose the optimal group score from among them: AIC(K) = Nln(RSS / N) + 2 * 4K; Where RSS is the residual sum of squares, N is the number of data points in the window, and each peak group contains 4 parameters; For high overlap, i.e., separation degree R s For values ​​less than 0.5, the PENDANTSS framework is invoked for joint detrending and blind deconvolution, and the generalized quasi-norm is used to improve separation accuracy compared to SOOT / SPOQ sparse penalty terms. The nonlinear least squares algorithm uses the Levenberg-Marquardt algorithm: Θ (m+1) =Θ (m) -[J T J+α*diag(J T J)] -1 J T r, The elements of the Jacobian matrix J are: The damping parameter α is automatically adjusted as the iteration converges; Output the estimated number of components within the neighborhood window and the signal of each separated single component and its peak parameters (A) k , t Rk , σ k , ).

7. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 6, characterized in that, It also includes a step of extracting oversampled normalized regions as feature image blocks: using the local maximum value p of the separated single-component signal. k Centered on this model, standardized feature image patches are extracted along two dimensions: retention time and mass-to-charge ratio. This transforms the one-dimensional chromatographic information into a two-dimensional image representation, which is then processed by a CNN model. Specifically: p k Centered on the RT dimension, the width is 6. RT, taking the expected range of isotope distribution in the m / z dimension [m / z theo -0.5,m / z theo +3.0], the extracted region is resampled to a fixed image size using bilinear interpolation, and the intensity value of each pixel is standardized so that the overall image has a mean of 0 and a standard deviation of 1; Oversampling interpolation involves interpolating the target image pixel coordinates (i′, j′) with the corresponding original data coordinates (t′, m′) using distance-weighted interpolation from the four adjacent points of the original grid. ； Among them, w ab =(1-∣t′-t a ∣ / Δt)(1-∣m′-m b | / Δm), the standardized output image: Output: Standardized feature image patch.

8. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 7, characterized in that, include: The standardized feature image patch is input into a pre-trained convolutional neural network (CNN). The model automatically determines whether the candidate position is the true chromatographic peak vertex, and simultaneously regresses and predicts the peak center position and bounding box. The specific processing procedure is as follows: Forward inference is performed using a pre-trained CNN model, outputting the results of three parallel branches: (i) peak / non-peak binary classification probability P(peak); (ii) the offset of the peak center position (Δt). R (iii) Bounding box offsets (Δleft, Δright, Δlow, Δhigh); The confidence score P(peak) of the output peak, and the precise location coordinates (t) R +Δt R (m / z+Δm / z) and bounding box information.

9. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 8, characterized in that, include: For overlapping peak regions identified as true peaks but showing insufficient separation (Rs < 0.8) in deconvolution residual analysis, the GCMSFormer model based on the Transformer architecture is used for end-to-end deep analysis to directly predict the pure mass spectra of each component from the signals in the overlapping peak regions. The specific operation is as follows: The signal y in the region identified as an overlapping peak by CNN overlap (n) encodes the regional signals of overlapping peaks into a sequence representation, inputs it into the self-attention encoder of the GCMSFormer model, and uses a multi-head attention mechanism to capture the long-range dependencies within the signal and automatically learns the interaction patterns between different components in the co-eluent peak. The decoder generates pure mass spectrometry sequences for each component, i.e., predicts the mass spectrometry fragment, m / z and abundance pairs for each component token by token; the integrated orthogonal projection resolution module orthogonally projects the resolved major components from the original mixed signal to remove them, and then resolves the low-concentration minor components. The Transformer model incorporates self-attention, multi-head attention, feedforward networks, layer normalization, and positional encoding. ； Where Q: Query matrix, representing the query vector emitted by the current target of attention or each position to be encoded. In self-attention, Q is derived from the linear transformation of the input sequence and is used to match K to calculate attention weights. K: Key matrix, representing the key vector at each position in the sequence, used to calculate similarity and determine the importance of each position to the current query; V: Value matrix, representing the value vector at each position in the sequence. Attention weights will be used to sum V in a weighted manner to obtain the final output. d k : The dimension of the key vector, used to scale the dot product result; softmax: Normalized exponential function; An integrated orthogonal projection resolution step is used to orthogonally project and remove the resolved principal components from the original mixed signal, so that the residual signal retains the unresolved secondary components; Specifically: residual signal matrix ; Where S is the mass spectrum column vector of the resolved principal components, I is the identity matrix, and y overlap : Signal vector of the original overlapping peak region.

10. The intelligent detection method for chromatographic peaks based on deconvolution algorithm according to claim 8, characterized in that, include: The confidence score P(peak) of the output peak, and the precise positioning coordinates (t) R +Δt R After obtaining the peak area integration (m / z + Δm / z) and bounding box information, the following steps are performed: Peak area integration, post-processing, and quantitative analysis. The processing procedure is as follows: Peak integral: Based on the bounding box information output by the CNN, the peak area is numerically integrated on the baseline-corrected signal based on the EMG model parameters; Peak alignment: For peaks detected among multiple parallel samples, a retention time alignment algorithm is used to match peaks and eliminate the influence of retention time drift on qualitative analysis. False positive suppression: Chemical rule filtering based on isotope distribution matching, adduct ion regularity, and retention time rationality; Trapezoidal integral of peak area: ； Or based on the analytical integral of the EMG model fitting results: ； That is, the parameters in the EMG model Direct output; Compound identification was performed by two-way matching with a standard spectral library, and peaks with scores all exceeding 800 or matching factors > 85% / probability > 80% were selected for confirmation. Output: Quantitative peak table, including retention time, m / z, peak area, peak width, asymmetry factor, standard spectral library matching results, and confidence score.

Images