Compound ionization efficiency prediction method based on context correction

Abstract

This invention discloses a method for predicting compound ionization efficiency based on context correction. The method includes extracting the mass concentration calibration slope of the target compound and converting it into a logarithmic slope under a molar concentration calibrator, which serves as the true ionization efficiency value. A stable feature set is constructed based on the molecular descriptors of the target compounds, with each sample representing a stable feature vector composed of stable descriptors for one target compound. The stable feature vectors are further nonlinearly corrected to obtain corrected feature vectors, and the corrected feature vectors of all target compounds are aggregated to construct a corrected feature set. Using the corrected feature set as input features and the true ionization efficiency as the label, a CatBoost regression model is trained using the corrected feature set. The trained CatBoost regression model is then used to predict the ionization efficiency of new compounds. This invention enables accurate and stable prediction of compound ionization efficiency.

Description

Technical Field

[0001] This invention relates to the field of compound mass spectrometry analysis technology, and in particular to a method for predicting compound ionization efficiency based on context correction. Background Technology

[0002] In quantitative / semi-quantitative analysis using liquid chromatography-high resolution mass spectrometry (LC-HRMS), environmental samples have complex matrix compositions and significant co-elution interference. The ionization efficiency (IE) of the same compound varies significantly under different sample systems and instrument detection conditions, resulting in a lack of stable correlation between the peak area detected by mass spectrometry and the true concentration of the compound. This severely limits the conversion of non-targeted detection results into semi-quantitative indicators such as concentration and pollution load that can support management decisions.

[0003] Currently, some quantitative analysis methods require the establishment of a separate standard curve for each target compound, along with internal standard correction and matrix matching to correct for response biases caused by fluctuations in ionization efficiency. This results in high experimental costs and makes it difficult to cover hundreds to thousands of untargeted analytes. Some studies use structurally similar compounds, homologous compounds, or empirical formulas to estimate the response factor, but these methods are essentially still indirectly correcting for differences in ionization efficiency. They are highly dependent on human experience, have limited applicability, and cannot stably correct for ionization efficiency fluctuations in cross-site and cross-matrix scenarios.

[0004] In recent years, machine learning methods based on molecular structure descriptors and structural fingerprints have directly predicted ionization efficiency, providing a new approach to solving the response bias problem. However, some shortcomings still exist: Inconsistent characterization of ionization efficiency: many methods use the slope of the mass concentration response as a label, failing to eliminate interference from molecular weight and concentration, thus limiting prediction accuracy; Lack of a dynamic correction mechanism that incorporates the instrument's actual response: existing prediction models mostly rely on static structure descriptors, failing to capture abrupt changes in ionization efficiency caused by minute structural differences or complex matrices in real samples, affecting the reliability of the prediction. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method for predicting the ionization efficiency of compounds based on context correction, so as to achieve accurate and stable prediction of the ionization efficiency of compounds.

[0006] Technical solution: To achieve the above objectives, the present invention provides a method for predicting compound ionization efficiency based on context correction, comprising the following steps:

[0007] S1, the slope for calibrating the mass concentration of the target compound;

[0008] S2. Convert the mass concentration calibration slope into a logarithmic slope under the molar concentration calibrator, and use it as the true value of the ionization efficiency of the target compound;

[0009] S3. Extract the molecular descriptors of the target compounds and construct a candidate feature matrix. In the matrix, rows correspond to different target compounds and columns correspond to different quantifiable molecular descriptors. Select stable descriptors from the candidate feature matrix and construct a stable feature set. Each sample in the set is a stable feature vector composed of stable descriptors of a target compound.

[0010] S4. Extract the contextual features of the target compound;

[0011] S5. Based on the contextual features of the target compound, the correction parameters of the stable feature vector of the target compound are generated using a correction neural network, and nonlinear correction is performed to obtain the correction feature vector of the target compound. The correction feature vectors of all target compounds are collected to construct a correction feature set.

[0012] S6. Using the corrected feature vector of the target compound as the input feature and the true value of ionization efficiency as the label, train the CatBoost regression model using the corrected feature set, and use the trained CatBoost regression model to predict the ionization efficiency of the new compound.

[0013] Preferably, for the target compound, the method for extracting the mass concentration calibration slope includes:

[0014] S11. Obtain primary mass spectrometry data of multi-concentration gradient standard solutions and blank solutions of the target compound;

[0015] S12. Based on the first-level mass spectrometry data of the target compound under multiple concentration gradients, extract the mass spectrometry response signal of the target compound at each concentration point. Based on the first-level mass spectrometry data of the blank solution, extract the background response signal. Subtract the background response signal from the mass spectrometry response signal to obtain the net response signal of the target compound at each concentration point.

[0016] S13. Using mass concentration as the independent variable and net response signal as the dependent variable, perform linear regression fitting through the origin, calculate the determination coefficient of the straight line obtained by linear regression fitting as the goodness of fit, and if the goodness of fit meets the preset threshold, extract the slope of the straight line as the mass concentration calibration slope of the target compound.

[0017] Preferably, the logarithmic slope logIE mol =log10(slope mol ), slope mol = slope ppb × MW ×10 6 Where MW is the molecular weight of the target compound, in g / mol, and slope ppb The slope is calibrated for the mass concentration of the target compound.

[0018] Preferably, the molecular descriptor includes:

[0019] Key ionization descriptors characterizing the true ionization state of a target compound molecule under given mobile phase pH conditions: acidic dissociation constant pKa acid Basic dissociation constant pKa base Deprotonation fraction alpha deprot And delta, used to characterize the difference between neutral hydrophobicity and charged state. logP,logD ;

[0020] Molecular physicochemical property descriptors characterizing the macroscopic spatial state of target compound molecules reflect the macroscopic physicochemical properties of target compound molecules during electrospraying related to droplet evaporation, interfacial enrichment, mass transfer, and desolvation: polarity TPSA, hydrophobicity such as LogP, molecular weight MolWt, number of hydrogen bonds HBA, and structural complexity parameters, including the number of rings, aromatic rings, heavy atoms, and molar refractive index MolMR;

[0021] Characterizing the local electronegativity distribution and elemental composition descriptors of the target compound molecule: representing the atomic count of the corresponding element in the molecule; representing the presence indication of whether the target compound molecule contains a certain type of element, if present, it is recorded as 1, otherwise it is recorded as 0.

[0022] Preferably, the method for selecting stable descriptors from the candidate feature matrix is ​​as follows:

[0023] First, the candidate feature matrix is ​​processed by filtering missing values, removing constant features, and filtering sparse / low occurrence rate. Then, correlation clustering is performed to remove redundancy, resulting in a representative feature matrix. The rows correspond to different target compounds, and the columns correspond to different representative descriptors. The representative descriptors are the descriptors retained after correlation clustering to remove redundancy.

[0024] The representative descriptors in the representative feature matrix are sorted from high to low according to their univariate correlation strength with the true value of ionization efficiency, and the top-K representative descriptors are retained as stable descriptors; or only representative descriptors with a correlation strength not lower than a set correlation threshold are retained in the representative feature matrix as stable descriptors, so as to construct a stable feature set.

[0025] Preferably, the method for extracting the contextual features of the target compound includes:

[0026] S41. Obtain secondary mass spectrometry data of the precursor ions of the target compound to construct a fragmentation tree, and generate a tree context vector of the target compound based on the fragmentation tree;

[0027] S42. The continuous numerical descriptor, Morgan fingerprint and MACCS bond fingerprint in the molecular descriptor described in S3 are spliced ​​together and standardized to form the structural context vector of the target compound.

[0028] S43. Concatenate the tree context vector and the structural context vector according to the feature dimension, and perform standardization and dimensionality reduction to obtain fixed-dimensional context features.

[0029] Preferably, the correction neural network includes a context encoding module, a gated output head, a bias output head, a sample scaling output head, and an auxiliary prediction head. The context encoding module consists of two or more fully connected layers, with a non-linear activation function ReLU used between the two fully connected layers. The context encoding module processes the input context features h... i Mapped to implicit representation u i The gated output head, bias output head, and sample scaling output head respectively handle the implicit representation u i Generate correction parameters: multiplicative gating coefficient vector signed gate,i Additive bias coefficient vector bias i Sample scaling factor scale,i The correction parameters are used to adjust the stable feature vector x. i Perform nonlinear correction to obtain the intermediate correction eigenvector x. corr,i :

[0030] x corr,i = x i ⊙ [1 + α ×sample scale,i × signed gate,i ] + β × sample scale,i ×bias i , where ⊙ represents element-wise multiplication, and α and β are preset correction magnitude hyperparameters.

[0031] Preferably, a trust weight is introduced. i For the intermediate correction eigenvector x corr,i Further correction yields the corrected eigenvector x. mix,i :

[0032] x mix,i = (1-η eff ) × x i + η eff × x corr,i η eff,i = η × trust i γ ,

[0033] trusti = (0.75+0.25·r mol )·(0.65+0.35·r miss )·(0.80+0.20·r cnt )·(0.70+0.30·r cons ),

[0034] r mol = mol valid,i ;r miss = 1 - tree missing,i ;r cnt =min{log(1+tree count,i ) / log4,1};r cons = [clip(tree consistency,i ,-1,1)+1] / 2,

[0035] Where η is the fusion intensity hyperparameter, γ is the credibility modulation hyperparameter, and mol valid,i The quality factor that can be correctly normalized and resolved from the simplified molecular linear input canonicalization of the target compound; missing,i A quality factor characterizing whether a target compound has available fragmented tree information; tree count,i The mass factor for the number of fragmented trees can be used to characterize the target compound; tree consistency,i A quality factor characterizing the consistency among the statistical vectors of multiple fragmented trees of a target compound.

[0036] Preferably, the CatBoost regression model is a gradient boosting model composed of multiple regression decision trees integrated in an additive manner, and uses the correction feature vector as input features to output an ionization efficiency prediction.

[0037] Preferably, during the training process, the set of corrected features is divided into a training set, a validation set, and a test set based on the chemical spatial diversity of the compound structure.

[0038] Beneficial effects: The present invention has the following advantages: 1. The present invention eliminates the error caused by molecular weight and concentration unit by uniformly converting the mass concentration slope to molar concentration and performing logarithmic processing; at the same time, it introduces the contextual features of the compound, combines the static molecular structure features with the mass spectrometry contextual information reflecting the fragmentation behavior of the compound, corrects the input feature samples, overcomes the ionization mutation caused by small structural differences in complex samples, and realizes accurate and stable prediction of the ionization efficiency of the compound. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the process of constructing the true value of ionization efficiency in this invention;

[0041] Figure 3 A schematic diagram of the process for selecting stable feature sets.

[0042] Figure 4 A comparison chart of predicted and true values ​​on the CatBoost model training set;

[0043] Figure 5 This is a comparison chart of predicted and actual values ​​on the CatBoost model test set. Detailed Implementation

[0044] The technical solution of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0045] Example 1

[0046] like Figures 1-3 As shown, this embodiment provides a method for predicting compound ionization efficiency based on context correction, including the following steps:

[0047] S1. Extract the target compound TC i Mass concentration calibration slope.

[0048] The target compound TC described in this embodiment i (i=1~I, representing the indices of different target compounds, where I is the total number of target compounds) can be understood as analyzable chemical substances present in environmental water samples, wastewater samples, pharmaceutical standards, or pollutant standards systems, including pharmaceuticals and personal care products, pesticides, industrial additives, intermediates, metabolites, and other organic pollutants. The mass concentration calibration slope of the target compound can be understood as the change in net response signal corresponding to a unit change in the mass concentration of the target compound, where the net response signal is the mass spectrometry response signal of the target compound after subtracting the background signal.

[0049] For the target compound TC i Methods for extracting the mass concentration calibration slope include:

[0050] S11. Obtain first-order mass spectrometry data of multi-concentration gradient standard solutions and blank solutions of the target compound.

[0051] The multi-concentration gradient standard solution of the target compound is a series of solutions with a predetermined concentration gradient prepared by dissolving the target compound in the same solvent (such as HPLC-grade methanol). The blank solution is a solution in the same solvent as the above standard solutions but without the target compound.

[0052] In this embodiment, the multiple concentration gradients can be selected as follows: 1, 10, 20, 50, 100, 150, 200, 300, 400, 500 ppb. The blank solution serves as a reference sample with a standard solution concentration of 0.

[0053] Both the standard solution and the blank solution were detected under the same or equivalent liquid chromatography-high resolution mass spectrometry (LC-HDMS) analysis conditions to obtain primary mass spectrometry data. The primary mass spectrometry raw data is a set of ion response signals, including the mass-to-charge ratio (m / z), retention time (RT), and signal intensity of the ion signal.

[0054] The primary mass spectrometry data of the target compound standard solution include: the mass-to-charge ratio (m / z), retention time (RT), and signal intensity of each ion in the solution (target compound response ion and solvent background ion).

[0055] The primary mass spectrometry data of the blank solution include: the mass-to-charge ratio (m / z), retention time (RT), and signal intensity of each ion in the solution (background ions from the solvent).

[0056] S12. Based on the first-level mass spectrometry data of the target compound under multiple concentration gradients, extract the mass spectrometry response signal of the target compound at each concentration point. Based on the first-level mass spectrometry data of the blank solution, extract the background response signal. Subtract the background response signal from the mass spectrometry response signal to obtain the net response signal of the target compound at each concentration point.

[0057] In this embodiment, the extraction method includes: peak extraction, retention time correction, alignment of characteristic peak groups, identification of target characteristic peak groups, and construction of peak alignment matrix for all primary mass spectrometry data; and extraction of mass spectrometry response signal and background response signal of target compound at each concentration point based on the peak alignment matrix.

[0058] Peak extraction includes: for each primary mass spectrometry data point, peak detection is performed within the two-dimensional signal space of mass-to-charge ratio (m / z) and retention time (RT). Preferably, ion signals that appear continuously along the retention time direction within a given mass-to-charge ratio (m / z) tolerance can be connected to form a local chromatographic curve, and effective peak intervals that meet a preset signal-to-noise ratio threshold (preferably S / N ≥ 10) and whose peak shape meets the requirements of a single peak or an approximately Gaussian distribution can be identified on the chromatographic curve. The effective peak interval is defined as a characteristic peak, and the representative mass-to-charge ratio (m / z), observed retention time (RT), and peak area of ​​the characteristic peak are recorded, wherein: the representative mass-to-charge ratio (m / z) can be taken as the mass-to-charge ratio corresponding to the peak apex or weighted centroid; the observed retention time can be taken as the peak apex time; and the peak area can be obtained by integrating the ion intensity within the peak boundary. Thus, each primary mass spectrometry data point forms a characteristic peak set consisting of multiple characteristic peaks.

[0059] Retention time correction and alignment of characteristic peak sets: Multiple characteristic peak sets are matched across samples. Anchor peaks are selected from all characteristic peak sets according to preset repeatability thresholds, mass-to-charge ratio deviation thresholds, peak area thresholds, and peak shape quality thresholds. The correspondence between these anchor peaks and each injection data (referring to injection data of target compound standard solutions or blank solutions, where the injection data refers to the data corresponding to the submitted samples for which corresponding primary mass spectrometry data was obtained) is used as the fitting basis for retention time correction. A retention time correction function is constructed based on the deviation between the observed retention time of the anchor peak and the reference retention time, and the retention time axis of each injection data is corrected. In some embodiments, the correction function is preferably a piecewise linear regression function or a locally weighted regression (LOESS) function.

[0060] After retention time correction, the residual retention time of the anchor peak in each injection data (the absolute value of the difference between the corrected observed retention time and the reference retention time) should not exceed 0.1 min. On the corrected retention time axis, a retention time matching tolerance ΔRT is set, preferably satisfying 0.1 min ≤ ΔRT ≤ 0.3 min; and combined with the mass-to-charge ratio matching tolerance, characteristic peaks from the same ion response event in different injection data are merged to form a group of aligned characteristic peaks.

[0061] Target characteristic peak group identification: After retention time correction and cross-sample grouping, based on the set mass-to-charge ratio m / z tolerance (error ±0.01 Da) and retention time tolerance (error ±0.1 min), the target response characteristic peak group related to the target compound is identified from the aligned characteristic peak groups of each standard solution injection data and blank solution injection data. For standard solution injection data, the identification results may include the target compound response characteristic peak group and the background characteristic peak group; for blank solution injection data, the identification results include the background characteristic peak group.

[0062] Peak alignment matrix construction: After the characteristic peaks are grouped across samples, a two-dimensional peak alignment matrix is ​​constructed for the same target compound, with the aligned characteristic peak group identifiers as rows and the injection data identifiers as columns. The aligned characteristic peak group identifiers are preferably determined by the average mass-to-charge ratio (Average Mz) and average retention time (Average Rt(min)) of the peak group; the matrix cell values ​​are the peak areas of the corresponding injection data at the aligned characteristic peak group.

[0063] Based on the peak alignment matrix, the peak area of ​​the standard solution injection data at the corresponding target alignment characteristic peak group is extracted as the original mass spectrometry response signal, and the peak area of ​​the blank solution sample at the corresponding alignment characteristic peak group is extracted as the background response signal.

[0064] In this embodiment, the background response signal is subtracted from the mass spectrometry response signal at each concentration gradient to obtain the target compound TC. i Net response signal at each concentration point.

[0065] In some embodiments, before subtracting the background response signal, it is determined whether the background response signal is a real background impurity or random noise from the instrument. Specifically, this includes: calculating the statistical representative value of the peak area of ​​the blank solution injection column as the blank response representative value in the peak alignment matrix constructed based on the blank solution primary mass spectrometry data. raw The median is preferred; in the peak alignment matrix constructed based on the primary mass spectrometry data of the standard solutions, the statistical representative value of the peak area of ​​the standard solution injection column is calculated as the standard response representative value sample. med The median is preferred. Set the blank gate threshold to thr = max(blank). abs , blank ratio × sample med ), where blank abs The absolute noise threshold, blank ratio As the background proportion weight, when the blank response represents the value blank raw If the background response signal is not lower than the blank gate threshold thr, it is determined that the background response signal has physical meaning and is recorded as the blank effective value blank. eff If the response is blank, it represents the value "blank". raw If the background response signal is less than the blank gate threshold thr, it is considered that the background response signal is more likely to originate from random noise or non-reproducible background, and no background subtraction is performed on the original mass spectrometry response signal. The background response signal is subtracted from the mass spectrometry response signal at each concentration gradient to obtain the effective response peak area after background interference is removed, i.e., the net response signal.

[0066] The deduction formula is: Area corr = max(Area raw - blank eff , blank floor ), where Area corr The net response signal after deduction, Area raw The original mass spectrometry response signal before subtraction, blank floor The lower limit of the response, which is set to prevent negative values, is preferably set to 0.

[0067] S13, targeting compound TC iWithin a preset mass concentration gradient range, net response signals at least five concentration points are acquired. Using mass concentration as the independent variable and net response signal as the dependent variable, a forced linear regression fitting is performed, assuming that when the target compound concentration is 0, the theoretical net response signal after deducting the effective background is 0. The coefficient of determination of the linear regression line is calculated as the goodness of fit. If the goodness of fit meets a preset threshold (e.g., greater than or equal to 0.95), the slope of the line is extracted as the mass concentration calibration slope for the target compound. ppb,i .

[0068] S2, Target compound TC i mass concentration calibration slope ppb,i Converted to logarithmic slope of molar concentration (logIE) mol,i TC, as the target compound i The true value of ionization efficiency.

[0069] In this embodiment, the logarithmic slope logIE mol =log10(slope mol ), where slope mol = slope ppb × MW × 10 6 MW represents the molecular weight of the target compound, in g / mol, and slope. ppb To calibrate the slope for the mass concentration of the target compound, under approximate aqueous phase conditions, 1 ppb can be equivalently written as 1 μg / L, i.e., 10 -6 g / L, 10 6 This is the conversion factor from ppb to g / L.

[0070] S3. Extract the molecular descriptors of each target compound and construct a candidate feature matrix F to characterize all target compounds. cand In this matrix, rows correspond to target compounds, and columns correspond to numerically quantifiable molecular descriptors, derived from the candidate feature matrix F. cand Stable descriptors are selected and a stable feature set is constructed. Each sample in the set is a stable feature vector composed of stable descriptors of a target compound.

[0071] S31. A molecular descriptor can be understood as various parameter features used to describe the ionization behavior, physicochemical properties, and elemental composition characteristics of a target compound molecule. In this embodiment, the molecular descriptor of the target compound is extracted based on the Simplified Molecular Linear Input Specification (SMILES), including:

[0072] 1) Key ionization descriptors used to characterize the true ionization state of a target compound molecule under given mobile phase pH conditions: acidic dissociation constant pKaacid Basic dissociation constant pKa base (Characterizing the acid-base dissociation ability of the target compound), deprotonation fraction alpha deprot (The proportion of target compounds that lose protons at a specified pH), and delta, used to characterize the difference between neutral hydrophobicity and charged state. logP,logD .

[0073] 2) Molecular physicochemical property descriptors used to characterize the macroscopic spatial state of target compound molecules, reflecting the macroscopic physicochemical properties of target compound molecules related to droplet evaporation, interfacial enrichment, mass transfer and desolvation during electrospraying: polarity TPSA, hydrophobicity such as LogP, molecular weight MolWt, number of hydrogen bonds HBA and structural complexity parameters, including the number of rings, aromatic rings, heavy atoms, and molar refractive index MolMR.

[0074] 3) Elemental composition descriptors used to characterize the local electronegativity distribution and elemental composition of the target compound molecule: elemental count characteristics, including N Hetero O Hetero S Hetero Cl Hetero F Hetero ,Br Hetero I Hetero , used to indicate the number of atoms of the corresponding element in a molecule; indicators of elemental presence, including has N has O has S has halogen A binary indicator bit is used to characterize whether the target compound molecule contains a certain type of element; if it is present, it is marked as 1, otherwise it is marked as 0.

[0075] S32, From the candidate feature matrix F cand Stable descriptors are selected, and a stable feature set is constructed, which includes:

[0076] Missing value filtering: Calculate the candidate feature matrix F cand For each column with a missing rate, columns with a missing rate greater than a preset missing threshold are removed (the preset missing threshold is preferably 0.8, that is, features with a missing rate > 0.8 are not included in the subsequent filtering).

[0077] Constant feature elimination: statistical candidate feature matrix F cand The number of unique values ​​in each column is counted, and columns with a unique value count of less than or equal to 1 are removed (a list with a unique value count of less than or equal to 1 indicates that the descriptor corresponding to that column does not carry distinguishing information, as shown in the candidate feature matrix F above). cand RD removed from SMR_VSA8 RD SlogP_VSA9RD fr_C_S (e.g., features without distinguishability)

[0078] Sparse / Low Occurrence Rate Filtering: Calculate the candidate feature matrix F cand The occurrence rate (non-zero proportion) of descriptors belonging to binary features, count features, and functional group hit position features is used to eliminate columns containing descriptors with an occurrence rate less than a preset non-zero proportion threshold. This is to avoid sparse positions, which only appear in a very small number of target compounds, interfering with the stability of subsequent model training. As shown in the above candidate feature matrix F... cand RD removed from fr_furan RD fr_quatN RD fr_thiophene (The preset non-zero ratio threshold is preferably 0.05).

[0079] Correlation clustering for redundancy removal: For the columns containing descriptors that have been filtered for missing values, constant features, and sparse / low occurrence rates, columns with an absolute Pearson correlation coefficient |r| ≥ 0.95 are grouped into the same collinear cluster to eliminate highly collinear redundant information, rather than simply deleting any one column; the correlation coefficient between each column and the true ionization efficiency is calculated as logIE. mol Univariate association strength scores (preferably F-regression scores, but mutual information scores can also be used) are used. For each collinear cluster, only the column with the highest score is retained as the representative descriptor of that cluster, forming a redundancy-free representative feature matrix F. dedup The rows correspond to the target compounds, and the columns correspond to the representative descriptors;

[0080] Stable feature set determination: In some embodiments, for the representative feature matrix F dedup The representative descriptors are sorted from high to low according to their univariate correlation strength with the true value of ionization efficiency. The top-K representative descriptors are retained as stable descriptors, forming a stable feature matrix F of fixed dimension. stable ∈R I×K In other embodiments, a univariate association strength threshold can also be set, for example, in the representation feature matrix F. stable Only representative descriptors with a correlation strength of at least 0.20 are retained as stable descriptors. A stable feature set is constructed based on these retained representative descriptors. Each sample in the stable feature set is a stable feature vector R composed of stable descriptors for one target compound. 1×k k is the vector dimension (k=K when the top-K representative descriptors are retained as stable descriptors), different samples correspond to different target compounds, and there are a total of I samples. For the target compound TC i The stable eigenvector is denoted as x. i .

[0081] In this embodiment, the stable descriptors in the stable feature subset that are ultimately retained after the above screening include: RD BertzCT RD SlogP_VSA8 RD EState_VSA3 RD SMR_VSA9 RD NumAromaticRings RD fr_Nhpyrrole RD HallKierAlpha RD HeavyAtomMolWt RD AvgIpc RD fr_piperzine RD MaxEStateIndex RD SMR_VSA3 RD MolLogP RD fr_aryl_methyl RD NumAromaticHeterocycles RD fr_NH2 RD NumHeteroatoms RD fr_Ar_N RD FpDensityMorgan1 ,RDf r_NH1 RD NumRotatableBonds RD RingCount RD NumAromaticCarbocycles RD MinEStateIndex RD SMR_VSA7 RD fr_NH0 ,RDf r_ArN pka, RD PEOE_VSA2 RD qed RD NumHeterocycles RD EState_VSA4 RD VSA_EState1 RD Chi3v RD SlogP_VSA6 RD PEOE_VSA4 RD PEOE_VSA8 The above are examples of the screening results in this embodiment, used to illustrate the specific sources of stable descriptors, and do not constitute a limitation on the stable descriptors that can be used.

[0082] In this embodiment, the aforementioned stable descriptor can be correlated with the true value of ionization efficiency, logIE. mol The stable correlation is formed because pKa and alpha deprot Directly determines the proportion of molecules in a deprotonable state under a given mobile phase pH condition; logD est With delta logP_logDThis corrects the characterization bias of traditional LogP in the distribution behavior of the real system by considering the ionization state; TPSA, HBA, HBD, and MolMR together reflect the polar surface area, hydrogen bonding ability, and polarizability of the molecule, which directly affect interfacial enrichment, solvation shell stripping, and gas-phase ion stability in electrospray droplets; MolWt, HeavyAtoms, Rings, AromaticRings, and FractionCSP3 reflect molecular size, rigidity, aromatic conjugation, and three-dimensional conformation characteristics, thereby indirectly affecting the exposure of the potential point, ion migration, and stabilization behavior; the element count variable N Hetero O Hetero S Hetero The halogen count variable characterizes the local electron pull effect and the ability to modulate acidic sites.

[0083] S4. Extract the target compound TC i Contextual features.

[0084] S41. Obtain the target compound TC i Precursor ion secondary mass spectrometry data; based on the secondary mass spectrometry data, a fragmentation tree is constructed; based on the fragmentation tree, the target compound TC is generated. i tree context vector v tree,i Specifically, it includes:

[0085] In this embodiment, the precursor ion of the target compound corresponds to the target compound in the standard solution in S1, and the secondary mass spectrometry data are obtained under the same or equivalent ion source polarity, adduct form and instrument conditions as in step S1.

[0086] The secondary mass spectrometry data are MS / MS records acquired by liquid chromatography-high resolution mass spectrometry (LC-HSMS). Each record includes at least: precursor ion information: precursor ion mass-to-charge ratio (m / z), charge number (z), adduct form, retention time, collision energy, and acquisition mode; fragment ion spectral data: a list of fragment ion peaks or a sequence of data points corresponding to the precursor ion, wherein the data points at least include the fragment ion mass-to-charge ratio (m / z) and peak intensity (I); the fragment ion spectral data are centroid data.

[0087] Construct a fragmentation tree, and generate the target compound TC based on the fragmentation tree. i tree context vector v tree,i Specifically, it includes:

[0088] First, each MS / MS record undergoes spectral preprocessing, preferably including peak extraction, denoising, isotope identification, and adduct identification, to obtain standardized MS / MS records. Then, based on a fragmentation tree construction algorithm, the standardized MS / MS records are converted into a directed acyclic tree with the parent ion as the root node, fragment ions as nodes, and neutral loss as edges. Specifically, nodes can be characterized by fragment ion m / z, peak intensity, mass deviation, and molecular formula candidate information; edges can be characterized by the mass difference between parent and child nodes, the corresponding neutral loss mass, and the loss rationality score.

[0089] For each fragmented tree, statistical features are extracted to form a tree context vector v. i,tree The tree context vector shall include statistics for at least the following categories:

[0090] Node and edge size statistics: such as the number of fragmented nodes and the number of lost edges;

[0091] Node m / z statistics: such as the mean, standard deviation, minimum, and maximum m / z of fragments;

[0092] Peak intensity statistics: such as mean intensity, standard deviation, maximum intensity, skewness, kurtosis, and spectral entropy;

[0093] Neutral loss statistics: such as the mean, standard deviation, and quantiles of neutral loss quality, and the mean and standard deviation of neutral loss scores;

[0094] Tree topology statistics: such as mean tree depth, maximum tree depth, proportion of leaf nodes, and proportion of branch nodes;

[0095] Mass error and molecular formula related statistics: such as the absolute value statistics of mass deviation ppm / Da, and the statistics of the number of atoms in the molecular formula;

[0096] Relative position statistics: such as root node m / z, relative m / z mean, intensity-weighted m / z mean.

[0097] If the same target compound corresponds to multiple MS / MS records or multiple fragmented trees, they can be merged according to the normalized SMILES of the target compound. The mean, standard deviation and inter-tree consistency of the statistics of each tree can be further calculated to form a compound-level aggregated tree context vector. Among them, the inter-tree consistency can be obtained from the mean similarity between the statistical vectors of multiple fragmented trees, and the number of fragmented trees can be used as additional confidence information.

[0098] S42, Generate the target compound TC i The structure context vector v mol,i Specifically, it includes:

[0099] Target compound TC iThe normalized SMILES are obtained by selecting continuously quantifiable descriptors, Morgan fingerprints, and MACCS bond fingerprints from the molecular descriptors described in S3, and concatenating them into a high-dimensional vector. After normalization, the structural context vector v of the target compound is formed. mol,i The structural context vector v mol,i This includes continuous features of the molecular descriptor, Morgan binary fingerprints, and MACCS binary bond information. In this embodiment, based on SMILES normalized representation, the two-dimensional topological information of the target compound molecule is converted into a fixed-length numerical structure vector through molecular structure encoding. The encoding methods include: RDKit physicochemical descriptor calculation, Morgan circular fingerprint generation, and MACCS structural key generation.

[0100] S43. Construct the target compound TC i Contextual features h i .

[0101] target compound TC i tree context vector v tree,i With structural context vector v mol,i The concatenated high-dimensional context features are then standardized and dimensionality reduced to obtain fixed-dimensional context features h. i ∈R m , where m represents the dimension of the context features. Preferably, dimensionality reduction can be achieved by performing standardized principal component analysis (PCA) on both the tree context and the structural context, and then combined with additional credibility features (such as tree...). count tree consistency tree missing mol valid The final contextual feature h is formed by splicing the features together. i .

[0102] S5, based on contextual features h i Stable feature vector x is generated using a bias correction neural network. i The correction parameters, and the stable eigenvector x i Perform nonlinear correction to obtain the corrected eigenvector x. corr,i The calibration feature vectors of all target compounds are collected to construct a calibration feature set.

[0103] S51. The CorrectorNet neural network includes a context encoding module, a gated output head, a bias output head, a sample scaling output head, and an auxiliary prediction head. The context encoding module consists of two or more fully connected layers, with the ReLU nonlinear activation function used between the two fully connected layers.

[0104] The context encoding module will input the context features h i Mapped to implicit representation u i The gated output head, bias output head, and sample scaling output head respectively address the implicit representation u. i Generate correction parameters: multiplicative gating coefficient vector signed gate,i Additive bias coefficient vector bias i Sample scaling factor scale,i , where, signed gate,i Additive bias coefficient vector bias i Sample scaling factor scale,i The same dimension as the stable feature vector; the correction parameter is used to adjust the stable feature vector x. i Perform nonlinear correction to obtain the intermediate correction eigenvector x. corr,i Subsequently, x will be combined with credibility weights. corr,i With the original stable eigenvector x i Further mixing yields the corrected feature vector x used for modeling in step S6. mix,i Assisted prediction head to correct feature vector x mix,i Input and output are auxiliary predictions of the true ionization efficiency, used for training supervision.

[0105] signed gate,i = tanh(W g ·u i + b g bias i = tanh(W b ·u i + b b ), sample scale,i =0.5 + sigmoid(W s ·u i + b s ); where W g W b and W s b are trainable weight parameters; g b b and b s These are trainable bias parameters; if the stable feature vector has dimension d and the latent representation u i If the dimension of W is q, then W g ∈R (d×q) W b ∈R (d×q) b g ∈R d b b ∈R d Ws ∈R (1×q) b s ∈R. signed gate,i The value of is in the range of −1 to 1, and is used to indicate the stable eigenvector x. i Should each feature dimension be adjusted up or down? (bias) i Used to supplement additive correction; sample scale,i This is a sample-level scaling factor used to control the stability of the feature vector x. i Overall correction range.

[0106] x corr,i = x i ⊙ [1 + α ×sample scale,i × signed gate,i ] + β × sample scale,i ×bias i Where ⊙ represents element-wise multiplication, and α and β are preset correction magnitude hyperparameters used to control the overall magnitude of the multiplicative correction term and the additive correction term, respectively. In a preferred embodiment, α can be 0.30 and β can be 0.10.

[0107] intermediate correction eigenvector x corr,i With stable eigenvector x i The dimensions are the same, but each feature value has been dynamically adjusted based on the context features. The original feature vector x is determined using context features. i The dimensions of the model should be magnified, reduced, or kept as unchanged as possible to achieve dynamic correction on a sample-by-sample and feature-by-feature basis.

[0108] S52. To prevent overcorrection of target compounds with missing fragmented tree information, incomplete molecular descriptors, or low context quality, this embodiment introduces a sample confidence weight, trust. i The trust i h is used to characterize the contextual features corresponding to the i-th target compound. i The reliability is calculated by combining the validity of structural information and the quality of fragmented tree information.

[0109] trust i = (0.75+0.25·r mol )·(0.65+0.35·r miss )·(0.80+0.20·r cnt )·(0.70+0.30·r cons ), and will trust i Crop to a preset range, such as [0.05, 1.00];

[0110] Where, rmol = mol valid,i ;r miss = 1 - tree missing,i ;r cnt =min{log(1+tree count,i ) / log4,1};r cons = [clip(tree consistency,i ,-1,1)+1] / 2.

[0111] mol valid,i Characterizing the target compound TC i Whether the SMILES can be correctly normalized and parsed as quality factors; tree missing,i Characterizing the target compound TC i Does a quality factor exist that provides available fragmented tree information? count,i Characterizing the target compound TC i The quality factor for the number of fragmented trees is available; tree consistency,i Characterizing the target compound TC i The quality factor for consistency among statistical vectors of multiple fragmented trees.

[0112] By letting r mol = mol valid,i ;r miss = 1 - tree missing,i ;r cnt =min{log(1+tree count,i ) / log4,1};r cons = [clip(tree consistency,i The above quality factors are mapped to the bounded interval [0,1] to eliminate the differences in the original dimensions and value ranges of different factors. After normalization, multiplicative fusion is performed to obtain trust. i .

[0113] Furthermore, the stable feature vector x i With intermediate correction eigenvector x corr,i Trust based on sample credibility weights i Weighted mixing yields the final corrected feature vector x. mix,i :

[0114] x mix,i = (1- η eff,i ) × x i + η eff,i × x corr,i η eff,i = η × trust iγ ,

[0115] Where η is the fusion intensity hyperparameter and γ is the confidence modulation hyperparameter. If the confidence weight of the stable feature vector of a target compound is low, then η eff,i It automatically decreases, thereby reducing the correction magnitude of the stable eigenvector.

[0116] Corrected eigenvector x mix,i With stable eigenvector x i Same dimensions.

[0117] In this step, an auxiliary prediction head is used to correct the feature vector x. mix,i The input is the same as the output, which is an auxiliary predicted value for the true ionization efficiency, and is compared with the true ionization efficiency by log IE. mol,i The comparison results in the main task loss.

[0118] S53. The training loss function of the correction neural network preferably consists of the regression main loss, reconstruction constraint, sparsity constraint, bias penalty, sample scaling penalty, and ranking consistency constraint: L = L task + λ recon ·L recon + λ sparse ·L sparse + λ bias ·L bias + λ scale ·L scale + λr ank ·L rank , where: L task The primary task loss is used; Huber loss is preferred for regression tasks, but mean squared error loss can also be used; L recon To reconstruct constraints, used to limit x corr The degree of deviation from x prevents the model from deviating from the original stable feature semantics; L sparse This is a gated sparsity constraint used to suppress unnecessary large-scale feature perturbations; L bias The bias penalty is used to control the magnitude of the additive correction term; L scale Sample scaling penalty, used to make the sample scale Try to fluctuate around 1; L rank This is a sorting consistency constraint used to improve the model's performance on logIE. mol The ability to preserve relative order. By minimizing the total loss function described above, the correction network can maintain a conservative correction of the stable feature vector of the target compound while ensuring prediction performance.

[0119] S6. Using the corrected feature vector of the target compound as the input feature and the true value of ionization efficiency as the label, train the CatBoost regression model and use the trained CatBoost regression model to predict the ionization efficiency of the new compound.

[0120] The S61 CatBoost regression model is an ensemble learning model based on gradient boosting regression decision trees, which includes multiple regression decision trees. For any input feature vector, the CatBoost model first provides an initial prediction value, and then each subsequent regression decision tree outputs an incremental correction value in turn. After all the incremental correction values ​​are summed with the initial prediction value, the final ionization efficiency estimate of the target compound is obtained.

[0121] When the corrected feature vector is input into the trained CatBoost regression model, it simultaneously enters all the pre-built regression trees within the CatBoost regression model. Within a single regression tree, the corrected feature vector starts from the root node, and at each level, a specific dimension of feature value is extracted and compared with a preset splitting threshold. Based on the comparison result, the corrected feature vector is directed to either the left or right child node, progressing layer by layer until it reaches the bottom leaf node, at which point the internal propagation ends. Since each bottom leaf node has been bound to a specific constant value during training, the regression tree ultimately outputs the constant value bound to the leaf node where the corrected feature vector falls.

[0122] During the training phase, the true ionization efficiency serves as a supervisory signal, determining the growth direction of the regression tree and the assignment of leaf nodes. Specifically, the CatBoost regression model first calculates the mean of the true ionization efficiencies for all target compounds as the initial bias term (i.e., the initial predicted value). Then, it calculates the difference between the true ionization efficiency for each target compound and the initial bias term, generating the initial residuals. The first regression tree is constructed using these initial residuals as the fitting objective: when selecting splitting features and setting thresholds for nodes, the criterion is minimizing the sum of squared residuals within the node; leaf nodes are assigned the mean of all initial residuals falling into that node to restore the initial residual information to the greatest extent possible.

[0123] Once the first regression tree is constructed, its output is superimposed onto the initial bias term to obtain an updated bias term. The CatBoost regression model then calculates the difference between the true ionization efficiency for each target compound and the updated bias term. This new round of residuals (i.e., the currently unexplained error) serves as the fitting target for the second tree. This mechanism iterates, with each subsequent newly added regression tree fitting the residuals that remain unexplained after the sum of all previous regression trees. Through this progressive fitting and additive superposition process based on the true label residuals, the CatBoost regression model can ultimately accurately capture the complex nonlinear mapping relationship between the correction feature vector and ionization efficiency.

[0124] During the training of the CatBoost regression model, to prevent overfitting and determine the optimal tree combination, it is preferable to use the training set for fitting and the validation set for hyperparameter selection and early stopping mechanism. Specifically, the true value of the ionization efficiency of the target compound is used as the supervision label, the root mean square error (RMSE) is used as the optimization index, and the mean absolute error (MAE) and coefficient of determination (R²) are used as auxiliary evaluation indexes, such as... Figure 4 , 5 As shown. Further, explicit optimization is performed on the hyperparameters controlling the residual fitting process: explicit optimization is performed on the following hyperparameters: tree depth, number of tree iterations, learning rate, and leaf node L2 regularization coefficient (l2leaf). reg The minimum number of leaf node samples, and, if necessary, the random intensity or subsampling parameters. For example, this can be achieved with depth = 4~10, iterations = 300~1500, learning rate = 0.01~0.20, and l2leaf... reg Perform a grid search or equivalent traversal optimization within the candidate space of 1~10, and select the candidate with the smallest RMSE and the lowest MAE and R on the validation set. 2 The optimal parameter combination for overall performance.

[0125] After determining the optimal hyperparameter combination, the training and validation sets are merged back into complete training data. Under the control of the optimal parameters, the CatBoost regression model is re-driven to perform round-by-round fitting and additive superposition of the residuals, resulting in the final prediction model that includes the optimal number of trees and node splitting scheme. Finally, the test set samples are input into this final model, and the corresponding ionization efficiency prediction is output. The generalization prediction ability of the model for compounds with unknown structures is characterized by indicators such as R², RMSE, and MAE.

[0126] S622. Predict the ionization efficiency of new compounds using the trained CatBoost regression model. Specifically, this includes: extracting molecular descriptors for the new compounds according to step S3, constructing contextual features according to step S4, and generating a corrected feature vector x according to step S5. mix,new ; will x mix,new Input the trained CatBoost regression model and output the predicted ionization efficiency of the new compound.

[0127] S623. The training set, validation set, and test set mentioned in step S61 are obtained by dividing the calibration feature set based on the chemical spatial diversity of compound structures.

[0128] For each normalized SMILES of a target compound, a fixed-length Morgan fingerprint is generated according to the same or equivalent structural encoding rules as in step S42. The Morgan fingerprint is a circular fingerprint encoding method based on iterative expansion of the molecular local atomic environment, used to map the two-dimensional topological structure of a molecule into a fixed-length binary or counting vector, thereby characterizing the neighborhood composition features of the molecule in structural space. Based on the Morgan fingerprint, MaxMin diversity sorting is performed on all target compounds. Specifically, a target compound is randomly selected as the starting point and placed at the top. For each remaining unsorted target compound, based on the Morgan fingerprint, the phase structure similarity with all target compounds in the currently sorted sequence is calculated (e.g., using the Tanimoto similarity coefficient as a similarity measure between Morgan fingerprints). The target compound with the lowest similarity among all remaining unsorted target compounds is added to the end of the currently sorted sequence; this process is repeated until all target compounds have been traversed.

[0129] Based on the MaxMin diversity ranking results of all target compounds, the calibration feature set is divided into training, validation, and test sets according to a preset ratio. For example, if the ratio is 8:1:1, the calibration feature vectors corresponding to the top 80% of target compounds in the MaxMin diversity ranking are assigned to the training set, the calibration feature vectors corresponding to the bottom 10% of target compounds are assigned to the test set, and the calibration feature vectors corresponding to the remaining target compounds are assigned to the validation set.

[0130] Example 2

[0131] This embodiment uses a negative ion electrospray mode and an adduct ion type of [MH]- as an example to illustrate the application of the method of the present invention in non-targeted detection of real water samples. Specifically, it includes:

[0132] Twelve real water samples collected from the same batch were selected. After filtration, enrichment, concentration, and resolidation, the samples were non-targeted detected using liquid chromatography-high resolution mass spectrometry in negative ion electrospray ionization mode. Only compounds with adduct ion type [M−H]− were recorded to obtain a sample-level characteristic table. The sample-level characteristic table includes at least: retention time, parent ion mass-to-charge ratio, peak area, structural annotation results, SMILES, and adduct ion type.

[0133] For compound records with completed structural annotations in the sample-level feature table, molecular descriptors are extracted and stable feature vectors are constructed according to step S3 in Example 1. For records with available secondary mass spectrometry data, tree context features, structural context features, and correction feature vectors are further extracted according to steps S4 and S5. For records without secondary mass spectrometry data, a conservative feature input method consistent with the training phase is used for prediction.

[0134] The above-mentioned corrected feature vectors are input into the trained CatBoost regression model, which outputs the predicted ionization efficiency value (pred) for each compound record. logIE,mol For compound records requiring semi-quantitative estimation of mass concentration, then follow the pred... logIE,mol =pred logIE,ppb +log10(MW)+6 is recalculated to obtain the true ionization efficiency at the molar concentration level. The peak area Area from the sample-level characteristic table is then substituted into the mass concentration calculation formula C. ppb =Area / 10 (predlogIE,ppb) The estimated mass concentration of each analyte in a real water sample is obtained; for the water sample system, ppb can be approximated as μg / L. Therefore, high-throughput ionization efficiency prediction and semi-quantitative estimation of compounds in real water samples can be achieved without establishing separate standard curves for each analyte.

[0135] In this embodiment, a total of 20,925 real water sample characteristic records were imported, of which 20,607 records successfully output predicted values, with a prediction coverage rate of 98.48%; the resulting pred logIE,mol The values ​​range from 10.342 to 13.679, with a median of 12.207; the estimated mass concentration obtained by converting based on peak area ranges from 0.396 to 197408.933 μg / L, with a median of 52.449 μg / L.

[0136] This embodiment demonstrates that, under the conditions of negative ion electrospray mode and adduct ion type [MH]-, the method of the present invention can be directly applied to non-targeted feature records with completed structural annotation in real complex water samples, and output ionization efficiency prediction results and semi-quantitative estimation results, thereby providing technical support for pollutant screening, inter-sample comparison and risk ranking.

Claims

1. A method for predicting compound ionization efficiency based on context correction, characterized in that, Includes the following steps: S1, the slope for calibrating the mass concentration of the target compound; S2. Convert the mass concentration calibration slope into a logarithmic slope under the molar concentration calibrator, and use it as the true value of the ionization efficiency of the target compound; S3. Extract the molecular descriptors of the target compounds and construct a candidate feature matrix. In the matrix, rows correspond to different target compounds and columns correspond to different quantifiable molecular descriptors. Select stable descriptors from the candidate feature matrix and construct a stable feature set. Each sample in the set is a stable feature vector composed of stable descriptors of a target compound. S4. Extract the contextual features of the target compound; S5. Based on the contextual features of the target compound, the correction parameters of the stable feature vector of the target compound are generated using a correction neural network, and nonlinear correction is performed to obtain the correction feature vector of the target compound. The correction feature vectors of all target compounds are collected to construct a correction feature set. S6. Using the corrected feature vector of the target compound as the input feature and the true value of ionization efficiency as the label, train the CatBoost regression model using the corrected feature set, and use the trained CatBoost regression model to predict the ionization efficiency of the new compound.

2. The method for predicting compound ionization efficiency according to claim 1, characterized in that, For the target compound, methods for extracting the mass concentration calibration slope include: S11. Obtain primary mass spectrometry data of multi-concentration gradient standard solutions and blank solutions of the target compound; S12. Based on the first-level mass spectrometry data of the target compound under multiple concentration gradients, extract the mass spectrometry response signal of the target compound at each concentration point. Based on the first-level mass spectrometry data of the blank solution, extract the background response signal. Subtract the background response signal from the mass spectrometry response signal to obtain the net response signal of the target compound at each concentration point. S13. Using mass concentration as the independent variable and net response signal as the dependent variable, perform linear regression fitting through the origin, calculate the determination coefficient of the straight line obtained by linear regression fitting as the goodness of fit, and if the goodness of fit meets the preset threshold, extract the slope of the straight line as the mass concentration calibration slope of the target compound.

3. The method for predicting compound ionization efficiency according to claim 1, characterized in that, The logarithmic slope logIE mol =log10(slope mol ), slope mol = slope ppb × MW × 10 6 Where MW is the molecular weight of the target compound, in g / mol, and slope ppb The slope is calibrated for the mass concentration of the target compound.

4. The method for predicting compound ionization efficiency according to claim 1, characterized in that, The molecular descriptor includes: Key ionization descriptors characterizing the true ionization state of a target compound molecule under given mobile phase pH conditions: acidic dissociation constant pKa acid Basic dissociation constant pKa base Deprotonation fraction alpha deprot And delta, used to characterize the difference between neutral hydrophobicity and charged state. logP,logD ; Molecular physicochemical property descriptors characterizing the macroscopic spatial state of target compound molecules: polarity TPSA, hydrophobicity such as LogP, molecular weight MolWt, number of hydrogen bonds HBA, and structural complexity parameters, including the number of rings, aromatic rings, heavy atoms, and molar refractive index MolMR. Characterizing the local electronegativity distribution and elemental composition descriptors of the target compound molecule: representing the atomic count of the corresponding element in the molecule; representing the presence indication of whether the target compound molecule contains a certain type of element, if present, it is recorded as 1, otherwise it is recorded as 0.

5. The method for predicting compound ionization efficiency according to claim 1, characterized in that, The method for selecting stable descriptors from the candidate feature matrix is ​​as follows: First, the candidate feature matrix is ​​processed by filtering missing values, removing constant features, and filtering sparse / low occurrence rate. Then, correlation clustering is performed to remove redundancy, resulting in a representative feature matrix. The rows correspond to different target compounds, and the columns correspond to different representative descriptors. The representative descriptors are the descriptors retained after correlation clustering to remove redundancy. The representative descriptors in the representative feature matrix are sorted from high to low according to the univariate correlation strength with the true value of ionization efficiency, and the top-K representative descriptors are retained as stable descriptors. Alternatively, only representative descriptors with a correlation strength not lower than a set correlation threshold can be retained in the representative feature matrix as stable descriptors to construct a stable feature set.

6. The method for predicting compound ionization efficiency according to claim 1, characterized in that, Methods for extracting contextual features of target compounds include: S41. Obtain secondary mass spectrometry data of the precursor ions of the target compound to construct a fragmentation tree, and generate a tree context vector of the target compound based on the fragmentation tree; S42. The continuous numerical descriptor, Morgan fingerprint and MACCS bond fingerprint in the molecular descriptor described in S3 are spliced ​​together and standardized to form the structural context vector of the target compound. S43. Concatenate the tree context vector and the structural context vector according to the feature dimension, and perform standardization and dimensionality reduction to obtain fixed-dimensional context features.

7. The method for predicting compound ionization efficiency according to claim 1, characterized in that, The correction neural network includes a context encoding module, a gated output head, a bias output head, a sample scaling output head, and an auxiliary prediction head. The context encoding module consists of two or more fully connected layers, with a non-linear activation function ReLU used between the two fully connected layers. The context encoding module processes the input context features h... i Mapped to implicit representation u i The gated output head, bias output head, and sample scaling output head respectively handle the implicit representation u i Generate correction parameters: multiplicative gating coefficient vector signed gate,i Additive bias coefficient vector bias i Sample scaling factor scale,i .

8. The method for predicting compound ionization efficiency according to claim 7, characterized in that, The correction parameters are used to stabilize the feature vector x. i Perform nonlinear correction to obtain the intermediate correction eigenvector x. corr,i : x corr,i = x i ⊙ [1 + α ×sample scale,i × signed gate,i ] + β × sample scale,i ×bias i , Where ⊙ represents element-wise multiplication, and α and β are preset correction magnitude hyperparameters; Introducing trust weight i For the intermediate correction eigenvector x corr,i Further correction yields the corrected eigenvector x. mix,i : x mix,i = (1 - n) eff,i ) × x i + the eff,i × x corr,i ,or eff,i = η × trust i γ , trust i = (0.75+0.25·r mol )·(0.65+0.35·r miss )·(0.80+0.20·r cnt )·(0.70+0.30·r cons ), r mol = mol valid,i ;r miss = 1 - tree missing,i ;r cnt =min{log(1+tree count,i ) / log4,1};r cons =[clip(tree consistency,i ,-1,1)+1] / 2, Where η is the fusion intensity hyperparameter, γ is the credibility modulation hyperparameter, and mol valid,i The quality factor that can be correctly normalized and resolved from the simplified molecular linear input canonicalization of the target compound; missing,i A quality factor characterizing whether a target compound has available fragmented tree information; tree count,i The mass factor for the number of fragmented trees can be used to characterize the target compound; tree consistency,i A quality factor characterizing the consistency among the statistical vectors of multiple fragmented trees of a target compound.

9. The method for predicting compound ionization efficiency according to claim 1, characterized in that, The CatBoost regression model is a gradient boosting model composed of multiple regression decision trees integrated additively, and it takes the correction feature vector as input features and outputs an ionization efficiency prediction.

10. The method for predicting compound ionization efficiency according to claim 1, characterized in that, During the training process, the set of corrected features is divided into a training set, a validation set, and a test set based on the chemical spatial diversity of the compound structure.

Images