Prediction method and device for retention time
By predicting the retention time of compounds, using quantum chemocomputing and chemical informatics methods, combined with machine learning methods, the problem of high compound separation time consumption in liquid chromatography is solved, and efficient and accurate compound separation operations are achieved.
Patent Information
- Application Number
- CN202311070040.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-23
- Publication Date
- 2025-02-25
AI Technical Summary
Existing liquid chromatography cannot quickly and accurately analyze the compounds corresponding to impurity peaks, and requires multiple attempts to chromatographic columns and mobile phases to ensure separation effect, resulting in high-time consumption and the inability to achieve high-throughput parallel separation.
By predicting the retention time of compounds, using quantum chemistry-calculated pKa and molecular fingerprints of chemical informatics, combined with machine learning methods such as KNN regression, a retention time prediction model is constructed to achieve accurate prediction of compound retention time.
It improves the accuracy and scalability of compound retention time prediction, reduces the number of experimental attempts, saves experimental time, and realizes high-throughput parallelized separation operation.
Smart Images
Figure CN119510650A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of chemical analysis and separation, and in particular to a retention time prediction method and device. Background Art
[0002] Liquid chromatography is a commonly used chemical analysis and separation technology. The existing liquid chromatography has the following problems:
[0003] 1. It is impossible to quickly and accurately analyze the compounds corresponding to the impurity peaks in the liquid chromatography. 2. It is necessary to try various chromatographic columns and mobile phases based on experience to ensure the separation of impurities, by-products and target compounds. 3. It usually takes a lot of time to wait for the target compound to appear in the spectrum, and it is impossible to achieve high-throughput and parallel separation operations.
[0004] How to solve the above problems existing in liquid chromatography is one of the current research directions. Summary of the invention
[0005] The inventors found in the process of research that predicting the retention time of compounds in liquid chromatography is helpful to solve the problems existing in liquid chromatography. The present application provides a retention time prediction method and device, the purpose of which is to solve the problem of how to predict the retention time of compounds in liquid chromatography.
[0006] In order to achieve the above objectives, this application provides the following technical solutions:
[0007] The first aspect of the present application provides a retention time prediction method, comprising:
[0008] Predicting acid ionization equilibrium constants pKa of a compound to be predicted in multiple solutions, wherein the multiple pKas include pKas corresponding to multiple types of molecules matching the compound to be predicted, and the multiple types of molecules include acid molecules and / or base molecules divided based on ionizable groups;
[0009] Obtaining the lipid partition coefficient LogD of the multi-ionized form compound of the compound to be predicted based on multiple pKa;
[0010] Based on the retention time prediction descriptive quantity of the compound to be predicted and a preset retention time prediction model, the predicted retention time of the compound is obtained, and the retention time prediction descriptive quantity includes LogD.
[0011] Optionally, the lipid partition coefficient LogD of the multi-ionized form compound of the compound to be predicted is obtained based on multiple pKa, including:
[0012] Multiple pKa forms a pKa sequence;
[0013] Generate multiple compounds based on pKa sequence;
[0014] Calculate the lipid partition coefficient LogP of the non-ionized form of multiple compounds;
[0015] Calculate the proportion of multiple compounds in the solution based on the pKa sequence;
[0016] LogD is calculated based on the pKa sequence, the proportion, and P, where P is the raw exponential form of LogP.
[0017] Optionally, the acid ionization equilibrium constants pKa of the compound to be predicted in multiple solutions are predicted, including:
[0018] Selecting all acid molecules and / or base molecules matching the compound to be predicted from a preset pKa training set;
[0019] Obtain the pKa prediction description of the compound to be predicted, which includes the total molecular polarity, atomic description and electrophilic super delocalization SE;
[0020] The pKa prediction descriptor is input into the model for predicting pKa corresponding to all acid molecules and / or base molecules to obtain multiple pKa outputs of the model for predicting pKa.
[0021] Optionally, the pKa training set includes:
[0022] Information of a plurality of molecules, wherein the plurality of molecules are divided into a first number of acid molecules and a second number of base molecules according to the types of ionizable groups contained; the information of any one of the molecules includes: matching functional groups and key atoms.
[0023] Optionally, obtain the predicted pKa description of the compound to be predicted, including:
[0024] Convert the compound to be predicted into multiple three-dimensional structures and select the molecular three-dimensional conformation with the lowest energy;
[0025] Based on the key atoms of the matching samples in the pKa training set, the atoms in the three-dimensional conformation of the molecule that need to extract the atomic description are determined as target atoms, and the matching samples are molecules with the same ionized groups as the compound to be predicted;
[0026] Extract the atomic descriptor of the target atom.
[0027] Optionally, before inputting the pKa prediction descriptor into the model for predicting pKa corresponding to all acid molecules and / or base molecules, the method further includes:
[0028] The pKa predicted descriptors were normalized and hyperbolic tangent transformed.
[0029] Optionally, the retention time prediction description also includes:
[0030] Physicochemical property descriptions, and molecular fingerprints.
[0031] Optionally, the molecular fingerprint acquisition process includes:
[0032] The atomic connection relationship of the preset radius value (such as the topological radius within 2) of the Morgan molecular fingerprint is represented by a third number (such as 2048) bits, and PCA is used to compress it into a fourth number (such as 20) description quantities.
[0033] A second aspect of the present application provides a retention time prediction device, comprising:
[0034] A pKa prediction module, used to predict the acid ionization equilibrium constant pKa of a compound to be predicted in multiple solutions, wherein the multiple pKas include pKas corresponding to multiple types of molecules matching the compound to be predicted, and the multiple types of molecules include acid molecules and / or base molecules divided based on ionizable functional groups;
[0035] A LogD prediction module, for obtaining the lipid partition coefficient LogD of a multi-ionized form of a compound to be predicted based on multiple pKa;
[0036] The retention time prediction module is used to input the retention time prediction description of the compound to be predicted into a preset retention time prediction model to obtain the predicted retention time of the compound output by the retention time prediction model, and the retention time prediction description includes LogD.
[0037] A third aspect of the present application provides an electronic device, characterized in that it includes:
[0038] Memory and processor;
[0039] The memory is used to store a computer program, and the processor is used to run the computer program to implement the retention time prediction method described in the first aspect of the present application.
[0040] The fourth aspect of the present application provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the retention time prediction method described in the first aspect of the present application.
[0041] The retention time prediction method and device provided in the present application can predict the retention time of the compound to be tested to solve some problems in liquid chromatography. In addition, the pKa of quantum chemistry dimension is used to obtain the LogD of electronic structure dimension, and then the retention time of the compound to be tested is predicted based on LogD. Compared with the descriptive quantity of chemical informatics, pKa can more effectively represent the electronic structure information of the ionization process, so it is conducive to improving the accuracy and scalability of the prediction results of the compound retention time. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0043] Figure 1 A flowchart of pKa training disclosed in the embodiments of the present application;
[0044] Figure 2 A flow chart for predicting pKa using a model for predicting pKa disclosed in an embodiment of the present application;
[0045] Figure 3 A flow chart of the calculation of LogD disclosed in the embodiments of the present application;
[0046] Figure 4 A flowchart of a training retention time prediction model disclosed in an embodiment of the present application;
[0047] Figure 5 Flowchart for predicting retention time using the retention time prediction model. DETAILED DESCRIPTION
[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to be used as limitations on the present application. As used in the specification and the appended claims of the present application, the singular expressions "one", "a kind", "said", "above", "the" and "this" are intended to also include expressions such as "one or more", unless there is a clear contrary indication in the context. It should also be understood that in the embodiments of the present application, "one or more" refers to one, two or more; "and / or" describes the association relationship of the associated objects, indicating that three relationships may exist; for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.
[0049] References to "one embodiment" or "some embodiments" etc. described in this specification mean that a particular feature, structure or characteristic described in conjunction with the embodiment is included in one or more embodiments of the present application. Thus, the phrases "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear at different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.
[0050] The multiple involved in the embodiments of the present application means greater than or equal to two. It should be noted that in the description of the embodiments of the present application, the words "first", "second", etc. are only used for the purpose of distinguishing the description, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying an order.
[0051] Retention time refers to the time from the injection of a sample component to the peak concentration of the component after the chromatographic column.
[0052] The inventors found in the course of their research that the prediction of retention time can solve three common problems in liquid chromatography for the following reasons:
[0053] 1. In chemical analysis, laboratory technicians usually analyze unknown impurity peaks in liquid chromatography. However, guessing the reaction mechanism alone cannot guarantee that it is the compound corresponding to the impurity peak. Other more accurate characterization methods (such as nuclear magnetic resonance spectroscopy) require a longer experimental cycle. The prediction of the retention time of the corresponding compound can assist the laboratory technician or the automated impurity prediction algorithm to judge the rationality of the guess.
[0054] 2. In chemical separation, the experimenter needs to try various chromatographic columns and mobile phases many times based on experience to ensure the separation of impurities, by-products and target compounds. The most critical part is the difference in retention time between the target compound and other components. The larger the difference, the higher the separation degree and the better the separation effect. Predicting the retention time of each component before the experiment can reduce the number of trials required, thereby saving experimental raw materials and accelerating the experimental process.
[0055] 3. In chemical separation, the experimenter usually spends a lot of time waiting for the target compound to appear in the spectrum, and cannot achieve high-throughput and parallel separation operations. The prediction of the retention time of the target compound can give the time range in which the target compound may appear, so as to assist the experimenter in arranging time and realize one-to-many docking of separation equipment.
[0056] The prediction method of compound retention time provided in the embodiment of the present application uses the physicochemical properties calculated by quantum chemistry and the chemical informatics description quantity processed by dimensionality reduction as the input of machine learning. The prediction of the acid ionization equilibrium constant (pKa) in the solution uses the conventional and special quantum chemical description quantities extracted by the quantum chemical semi-empirical method xTB to ensure the accuracy and scalability of the prediction. The molecular fingerprint as a chemical informatics parameter is processed by PCA data dimensionality reduction to obtain the most relevant linear combination description component to avoid the autocorrelation of the description quantity and the demand for a large amount of data. The machine learning part simultaneously ensures the accuracy and interpretability of the model through KNN regression. This model can theoretically obtain the most complete chemical information and an interpretable and scalable model.
[0057] The compound retention time prediction method provided in the application example specifically includes pKa training and prediction, calculation and prediction of the lipid partition coefficient (LogD) of multi-ionized form compounds considering ionization, and description of the construction of the quantity and the retention time prediction process.
[0058] The above processes will be described in detail below.
[0059] Figure 1 The pKa training process is shown, the purpose of which is to train a model for predicting the pKa of a compound, including the following steps:
[0060] S101. After receiving the compound SMILES sample, select an acid or base sample matching the compound SMILES from a preset pKa training set.
[0061] Simplified Molecular Input Line Entry Specification (SMILES) is a specification that uses an ASCII string to clearly describe the molecular structure. It is understood that SMILES is only an example, and the received compound may also be in other formats, which are not limited here.
[0062] In the training stage, the compound SMILES sample refers to the compound SMILES with known pKa, and the pKa of the compound SMILES sample is used as a label to adjust the parameters of the model.
[0063] The pKa training set includes information of multiple molecules. The multiple molecules are divided into a first number of acid molecules (containing acidic ionized groups) and a second number of base molecules (containing basic ionized groups) according to the types of ionized groups contained. The first number and the second number are not limited. In one example, the first number is 30 and the second number is 39.
[0064] The information of any molecule includes: matching functional groups and key atoms in SMARTS format. Key atoms refer to atoms that affect the degree of ionization, and their atomic properties such as charge are used as the basis for extracting the descriptive quantity for pKa prediction.
[0065] A molecule matching the received compound SMILES sample is selected from the pKa training set. For ease of explanation, the selected molecule is referred to as a matching sample. The matching sample uniquely corresponds to the received compound SMILES sample. Unique correspondence can be understood as the same ionized groups.
[0066] During the research, the inventors found that classifying molecules according to ionized groups based on experience and data distribution characteristics is conducive to achieving the following advantages: 1. Each category has sufficient real data for fitting; 2. There is a good linear relationship between the data within the category and the constructed descriptive quantity.
[0067] In some implementations, the types classified by ionizable groups include:
[0068] Acids: aromatic sulfonic acid OH, fatty sulfonic acid OH, phosphoric acid OH, phosphorous acid OH, aromatic carboxylic acid OH, fatty carboxylic acid OH, thiocarboxylic acid OH, boric acid OH, β,γ-unsaturated enol OH, β,γ-unsaturated ene thiol OH, active methylene / methine CH, imide NH, N-aryl aromatic sulfonamide NH, N-aliphatic aromatic sulfonamide NH, N-aryl fatty sulfonamide NH, N-aliphatic fatty sulfonamide NH, pyrrole NH, pyrazole NH, 2H-1,2,3-triazole, pyrrole-1-phenol OH, N-aliphatic hydroxylamine OH, thiol, thiol, phenol, alcohol, nitro-ortho CH, nitrosamine NH, aromatizable lactam NH, non-aromatizable lactam NH, diarylamine NH.
[0069] Bases: pyridine N, pyridazine N, 1,2,3-triazine N, triaryl C substituted amine N, diaryl C-aryl N substituted amine N, monoaryl C diaryl N substituted amine N, diaryl C-aliphatic C substituted amine N, monoaryl C-aryl N-aliphatic C substituted amine N, diaryl C substituted hydrazine N, monoaryl C-aryl N substituted hydrazine N, monoaryl C dialiphatic C substituted amine N, monoaryl C-aliphatic C substituted hydrazine N, monoaryl C-aliphatic N substituted hydrazine N, trialiphatic C substituted polycyclic amine N, trialiphatic C substituted 7-8 membered monocyclic amine N, trialiphatic C substituted 5-6 membered monocyclic amine N, trialiphatic C substituted 3-4 membered monocyclic amine N, dialiphatic C-aliphatic N Substituted amine N, monoaliphatic C dialiphatic N substituted amine N, monoaryl C substituted imine N, monoaryl C substituted azo N, monoaliphatic C substituted imine N, monoaliphatic C substituted azo N, monoaryl N substituted imine N, monoaliphatic N substituted imine N, diaryl C substituted amine N, monoaryl C-aryl N substituted amine N, monoaryl C-aliphatic C substituted amine N, monoaryl C-aliphatic C substituted hydrazine N, dialiphatic C substituted acyclic / macrocyclic amine N, dialiphatic C substituted 5-6 membered cyclic amine N, dialiphatic C substituted 3-4 membered cyclic amine N, monoaliphatic C substituted hydrazine N, unsubstituted imine N, unsubstituted azo N, aromatic primary amine N, monoaromatic N substituted amine N, aliphatic primary amine N, unsubstituted hydrazine N.
[0070] S102. Convert the compound SMILES sample into multiple three-dimensional structures and select the molecular three-dimensional conformation with the lowest energy.
[0071] The three-dimensional structure can be understood as the topological structure, and the three-dimensional conformation of a molecule can be understood as the molecular coordinates.
[0072] In some implementations, RDKit (an open source chemical informatics software) is used to convert SMILES into a plurality of three-dimensional structures. An example of the plurality is ten.
[0073] In some implementations, after optimizing each three-dimensional structure using a Universal Force Field (UFF) model, the three-dimensional conformation of the molecule with the lowest energy among the optimized three-dimensional structures is selected to improve the accuracy of the selected three-dimensional conformation of the molecule.
[0074] Because the three-dimensional conformation of the molecule with the lowest energy is representative of all molecular conformations and is also the conformation that is most likely to occur, it can better represent the ionization properties of the molecule under real conditions. However, it is understandable that the lowest energy is only an example and not a limitation.
[0075] S103. Based on the selected three-dimensional conformation of the molecule, quantum chemical single-point energy calculation is performed to obtain the total molecular polarity, atomic coordination number, atomic charge, atomic dispersion coefficient C6, atomic polarizability and wave function Molden file.
[0076] In some implementations, the selected molecular three-dimensional conformation is first structurally optimized, and then the optimization result is subjected to quantum chemical single point energy calculation. An example of an algorithm used for structural optimization is structural optimization at the semi-empirical method GFN2-xTB level.
[0077] For the convenience of description, the descriptive quantity of the molecular dimension, that is, the total polarity of the molecule, is called dip (the variable is mol), and the descriptive quantity of the atomic dimension, that is, the atomic descriptive quantity including the atomic coordination number, atomic charge, atomic dispersion coefficient C6, and atomic polarizability, is collectively called prop (the variable is atom).
[0078] In some implementations, based on the key atoms (positions) of the matching samples in the pKa training set, the atoms in the three-dimensional conformation of the molecule from which props need to be extracted (i.e., the atoms in the same position as the key atoms, for example, if the positions of the key atoms in the matching samples are 1, 3, and 5, then the atoms at positions 1, 3, and 5 in the three-dimensional conformation of the molecule need to extract props), referred to as target atoms, are determined, and then the props of the target atoms are extracted. Specifically, the props of the target atoms are extracted through quantum chemical single-point energy calculations.
[0079] S104, extracting orbital information based on the wave function Molden file, and calculating electrophilic superdelocalisability (SE) based on the orbital information.
[0080] SE (variable atom) is used to measure the electrophilicity of an atom.
[0081] S105. Combine the total molecular polarity, atomic coordination number, atomic charge, atomic dispersion coefficient C6, atomic polarizability and SE to obtain a combined description.
[0082] In some implementations, the merging may be understood as forming a description quantity from dip, prop, and SE, which is denoted as dp=[prop(atom), SE(atom), dip(mol)].
[0083] S106. Standardize and perform hyperbolic tangent transformation on the combined description quantity to obtain a transformed combined description quantity.
[0084] The purpose of normalization and hyperbolic tangent transformation is to eliminate extreme values. It can be understood that S106 is an optional step, the purpose of which is to improve the accuracy of the model.
[0085] At this point, the combined description corresponding to the matching sample is obtained.
[0086] S107. Perform LassoCV linear regression training under matching samples to obtain a model.
[0087] In some implementations, the training logic includes: inputting the transformed combined description into the model to be trained to obtain the pKa output by the model, and then adjusting the parameters of the model based on the pKa output by the model, the label of the compound SMILES sample, and the loss function.
[0088] It is understandable that after Figure 1 After the training process shown, a model for predicting pKa under matching samples is obtained.
[0089] In the SMILES samples with different compounds Figure 1 In the case of the process shown, as the matching samples of the compound SMILES samples are different, a model for predicting pKa corresponding to each type of molecule in the pKa training set can be obtained.
[0090] Figure 2 The workflow for predicting pKa using the model for predicting pKa includes the following steps:
[0091] S201, receiving compound SMILES.
[0092] S202, selecting all acid molecules and / or base molecules that match the SMILES of the compound from the pKa training set.
[0093] As mentioned above, each type of acid and base molecule corresponds to a model for predicting pKa.
[0094] S203, converting the compound SMILES into multiple three-dimensional structures, and selecting the molecular three-dimensional conformation with the lowest energy.
[0095] S204. Perform quantum chemical single-point energy calculation based on the selected molecular three-dimensional conformation to obtain the total molecular polarity, atomic coordination number, atomic charge, atomic dispersion coefficient C6, atomic polarizability and wave function Molden file.
[0096] S205, extracting orbital information based on the wave function Molden file, and calculating SE based on the orbital information.
[0097] exist Figure 1 and Figure 2 In the process shown, dip, prop and SE are collectively referred to as pKa prediction descriptors.
[0098] S206: Obtain the combined description amount.
[0099] S207, normalize and perform hyperbolic tangent transformation on the combined description quantity to obtain a transformed combined description quantity.
[0100] S208. Input the transformed combined description quantity into a model for predicting pKa to obtain pKa output by the model.
[0101] It can be understood that when there are multiple categories of acid molecules and / or base molecules matching the SMILES of the compound, there are multiple models for predicting pKa. Therefore, in this step, the transformed combined description quantity is input into all the models for predicting pKa, and each model outputs a pKa, and multiple models will obtain multiple pKa.
[0102] That is to say, the pKa prediction process is similar to the training process, but the molecule to be predicted may have multiple acidic and basic ionized groups, so all acids and bases are extracted in the test process. The description of all acids and bases can be obtained through a quantum chemical calculation of the whole molecule. Figure 1 The model obtained by the training process shown makes predictions under the corresponding acid molecules and / or base molecules respectively to obtain a series of pKa values of the molecules.
[0103] Figure 2 The specific implementation methods of each step in can be found in Figure 1 Follow the appropriate steps shown.
[0104] Figure 2 The process shown utilizes the descriptive quantities obtained from quantum chemical calculations, which can more effectively represent the electronic structure information of the ionization process than the descriptive quantities from cheminformatics, and is conducive to improving the accuracy and scalability of the prediction results of compound retention times.
[0105] Figure 3 The calculation process of LogD includes the following steps:
[0106] S301, sorting the predicted pKa from small to large to obtain a pKa sequence.
[0107] like Figure 3 As shown, the predicted pKa is Figure 2 The process shown results in a series of pKa values output by the model for predicting pKa.
[0108] The purpose of sorting from small to large is to simplify the implementation and is only an example and not a limitation.
[0109] 2 N The compounds include: neutral compounds (ie, the compounds SMILES received by S201), protonated and deprotonated ionic compounds, and zwitterionic compounds.
[0110] S302, generating ionization sites based on the pKa sequence, and obtaining multiple compounds based on the ionization sites.
[0111] All possible ionization sites are generated based on the pKa sequence, and then multiple compounds are obtained based on the combination of ionization sites. It can be understood that N ionization sites can obtain 2 N A compound.
[0112] S303. Calculate LogP of multiple compounds.
[0113] In some implementations, the LogP of the neutral compound is obtained by the LogP program built into RDKit. For protonated and deprotonated ionic compounds, the LogP is obtained based on the empirical difference and the LogP of the neutral compound, where the empirical difference represents the difference between the LogP of the neutral compound based on the empirical configuration. For zwitterionic compounds, the LogP of each pair of positive and negative ions is corrected by +4.3 based on the LogP of the protonated and deprotonated ionic compounds.
[0114] S304, calculate multiple (2 N The proportion of a compound in the solution.
[0115] The specific calculation method is shown in formula (1).
[0116] It is to be understood that the pKa described in the embodiments of the present application refers to the ionization coefficient in the solution, so the solution also refers to an aqueous solution.
[0117] S305. Calculate LogD.
[0118] The calculation method is shown in the following formula:
[0119] In formula (1), D represents the original exponential form of logD, P iRepresents each ionized component i (a total of 2 n The original exponential form of logP of a component). pH represents pH value. pK j(j+1) represents the ionization equilibrium constant from component j to component j+1, bin(i) j Represents the jth bit after i is binarized.
[0120] Representation 2 N The proportion of a compound in a solution. It can be seen that the proportion is determined by the pH value and a set of pKa values of the compound (i.e. Figure 2 The obtained series of pKa values) are jointly determined.
[0121] In formula (1), the pH range is [0, 14], and the step length is 0.5, that is, the interval is 0.5.
[0122] It can be understood that after D is calculated, a logarithm operation is performed on D to obtain LogD.
[0123] Figure 3 The process shown is equivalent to an innovative method for calculating the LogD of a compound. The advantage is that the pKa and LogP parts do not require external experimental information, and the estimation of LogP does not require all 2 N Formula (1) can be efficiently parallelized in the code and the calculation can be accelerated by constraining the difference between pKa and pH. Figure 3 The process shown can obtain LogD efficiently and quickly.
[0124] The retention time of a compound depends on many factors. In addition to the LogD value, the number of hydrogen bonds in the compound, the binding properties of the compound's special structure with the stationary phase chromatographic column, the shape and size of the compound's molecules, and other factors can affect the retention time of the compound in the mobile phase.
[0125] In order to control the parameter space to an appropriate size, the embodiment of the present application considers the retention time of different compounds under the same stationary phase and mobile phase, that is, the retention time is only related to the structure and properties of the compound itself. Based on this, the logarithm of the retention time t, log(t), is summarized into a linear model:
[0126] log(t)=aLogD+b...+c...(2)
[0127] Where a is the weight of LogD in the formula. Correspondingly, b and c are the weights of other descriptors. For the sake of distinction, the descriptor used to calculate log(t) is called the retention time prediction descriptor.
[0128] In the embodiments of the present application, three types of retention time prediction descriptors are used to predict retention time: LogD, physicochemical property descriptors, and molecular fingerprints, which correspond to the distribution of compounds in the dissolution process, the interaction with the solvent, and the functional group characteristics (which may affect the dissolution process), so as to better simulate the physical process of compound dissolution in chromatography, which is conducive to improving the readiness of retention time prediction results.
[0129] Based on the above technical orientation, Figure 4 The process for training a retention time prediction model includes the following steps:
[0130] S401. Obtain description quantities for training a retention time prediction model.
[0131] In some implementations, LogD includes 29 values, and the 29 values are LogD with pH in the range [0, 14], with intervals of 0.5. The physicochemical property description uses 17 descriptions that come with RDKit, that is, 17 descriptions of the compound (target compound) are obtained: number of hydrogen bond donors, number of hydrogen bond acceptors, polar surface area, surface average hydrogen bond number, LogP, topological molecular polar surface area, molecular weight, molar refractive index, number of rotatable chemical bonds, number of aromatic rings, number of aliphatic rings, number of saturated aliphatic rings, sp3 hybridized carbon atom ratio, Chi0 index, Chi4n index, Kappa1 index, Kappa2 index. The molecular fingerprint uses MorganFingerprint, which uses 2048 bits to represent the atomic connection relationship within a topological radius of 2, and uses PCA to compress it into 20 descriptions. The above totals 66 descriptions.
[0132] S402: Based on the description quantity used to train the retention time prediction model, select N compounds from the retention time training set as retention time training samples.
[0133] The retention time training set includes multiple compound SMILES samples and their corresponding retention times.
[0134] The compound received by S201 is called the target compound (i.e., the compound sample used to train the retention time prediction model). Figure 2 as well as Figure 3 Following the process shown, the LogD of the target compound is obtained, and then the physicochemical property description quantity and molecular fingerprint of the target compound are obtained, a total of 66 description quantities.
[0135] In some implementations, the same approach (i.e. Figure 2 as well as Figure 3), obtain 66 description quantities of each compound in the retention time training set, and use the compounds in the retention time training set whose description quantities meet the conditions with the 66 description quantities of the target compound as retention time training samples. The conditions include: in the space spanned by the 66 description quantities, the distance (L1 or L2 norm) is the shortest. The distance is less than the first threshold, that is, the distance between the description quantity of the compound that meets the conditions and the description quantity of the target compound is close.
[0136] In this step, it is assumed that the number of selected retention time training samples is K, and an example of K is 5.
[0137] S403: Construct a KNN regression model based on the retention time training sample.
[0138] In some implementations, weights are configured for the K retention time training samples, and a KNN regression model is constructed based on the weights and the 66 descriptors of the K training samples. The weights of different retention time training samples can be the same (i.e., the KNN regression model is constructed by taking the average of the corresponding descriptors of the K retention time training samples) or different.
[0139] Figure 4 The training method shown in the figure, the KNN regression model ensures that the K compounds used for prediction have a certain degree of similarity with the molecule to be predicted in terms of electronic structure (LogD predicted by quantum chemistry), molecular structure (cheminformatics), and physicochemical properties.
[0140] Figure 5 The process of predicting retention time using the retention time prediction model includes the following steps:
[0141] S501. Predict multiple pKa values of the compound to be predicted.
[0142] In some implementations, the pKa values of all molecules matched by the compound to be predicted in the pKa training set are obtained, and the pKa values of any class of molecules matched by the compound to be predicted are based on Figure 2 The process shown is obtained.
[0143] S502 . Obtain LogD of the compound to be predicted based on multiple pKa values.
[0144] In some implementations, based on Figure 3 In the process shown, all the obtained pKa are sorted according to the preset sorting rules to obtain a pKa sequence. Based on the pKa sequence, LogD is obtained. In this step, logD is still taken as an example of 29 LogDs in the pH range of 0-14 with an interval of 0.5.
[0145] Retention time prediction descriptors include LogD.
[0146] In some implementations, in order to improve the accuracy of retention time prediction results, the retention time prediction descriptors also include the physicochemical property descriptors and molecular fingerprints of the compound to be predicted, as described above, plus 29 LogDs, for a total of 66 descriptors.
[0147] At this point, the predicted retention time description of the compound to be predicted is obtained.
[0148] S503 , inputting the description quantity of the compound to be predicted into the retention time prediction model to obtain the predicted retention time of the compound output by the retention time prediction model.
[0149] Figure 5 The retention time prediction method shown in the figure has a wide range of description quantities and a wider range of adaptability. And the machine learning part ensures the accuracy and interpretability of the model through KNN regression at the same time. This model can theoretically obtain the most complete chemical information and an interpretable and scalable model. In addition, the training and test sets in the above-mentioned papers all use a large number of relatively small molecules (molecular weight <200) or similar molecules, and the model has not been verified by multiple types and larger molecules in actual experimental production. The example model of the embodiment of the present application uses less than 1000 molecules with an average molecular weight of 555.7, and obtains an average relative deviation of only 15%.
[0150] If the functions described in the method of the embodiment of the present application are implemented in the form of software functional units and sold or used as independent products, they can be stored in a storage medium readable by a computing device. Based on this understanding, the part of the embodiment of the present application that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computing device (which can be a personal computer, server, mobile computing device or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.
[0151] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
Claims
1. A retention time prediction method, characterized in that: include: Predicting acid ionization equilibrium constants pKa of a compound to be predicted in multiple solutions, wherein the multiple pKas include pKas corresponding to multiple types of molecules matching the compound to be predicted, wherein the multiple types of molecules include acid molecules and / or base molecules divided based on ionizable groups; Obtaining the lipid partition coefficient LogD of the multi-ionized form compound of the compound to be predicted based on the multiple pKa; Based on the retention time prediction descriptor of the compound to be predicted and a preset retention time prediction model, the predicted retention time of the compound is obtained, and the retention time prediction descriptor includes the LogD.
2. The method according to claim 1, characterized in that The step of obtaining the lipid partition coefficient LogD of the multi-ionized form compound of the compound to be predicted based on the multiple pKa comprises: forming the plurality of pKa into a pKa sequence; generating a plurality of compounds based on the pKa sequence; Calculating the lipid partition coefficient LogP of the non-ionized form of the plurality of compounds; Calculating the proportion of the plurality of compounds in the solution based on the pKa sequence; The LogD is calculated based on the pKa sequence, the ratio, and P, where P is the original exponential form of the LogP.
3. The method according to claim 2, characterized in that The method of predicting the acid ionization equilibrium constants pKa of the compound to be predicted in multiple solutions comprises: Selecting all acid molecules and / or base molecules matching the compound to be predicted from a preset pKa training set; Obtaining a pKa prediction description of the compound to be predicted, wherein the pKa prediction description includes total molecular polarity, atomic description, and electrophilic super delocalization SE; The pKa prediction descriptor is input into the model for predicting pKa corresponding to all the acid molecules and / or base molecules to obtain a plurality of pKa outputs of the model for predicting pKa.
4. The method according to claim 3, characterized in that The pKa training set includes: Information of a plurality of molecules, wherein the plurality of molecules are divided into a first number of acid molecules and a second number of base molecules according to the types of ionizable groups contained; the information of any one of the molecules includes: matching functional groups and key atoms.
5. The method according to claim 4, characterized in that The step of obtaining the predicted pKa value of the compound to be predicted comprises: Converting the compound to be predicted into multiple three-dimensional structures, and selecting the three-dimensional conformation of the molecule with the lowest energy; Based on the key atoms of the matching samples in the pKa training set, the atoms in the three-dimensional conformation of the molecule whose atomic description quantity needs to be extracted are determined as target atoms, and the matching samples are molecules with the same ionized groups as the compound to be predicted; The atomic descriptor of the target atom is extracted.
6. The method according to claim 3, characterized in that Before inputting the pKa prediction descriptor into the model for predicting pKa corresponding to all the acid molecules and / or base molecules, the method further includes: The pKa predicted descriptors were normalized and hyperbolic tangent transformed.
7. The method according to any one of claims 1 to 6, characterized in that: The retention time prediction description also includes: Physicochemical property descriptions, and molecular fingerprints.
8. The method according to any one of claims 1 to 6, characterized in that: The molecular fingerprint acquisition process includes: The topological radius of the Morgan molecular fingerprint is represented by a third number of bits, and the atomic connection relationship within a preset radius value is represented by a third number of bits, and PCA is used to compress the third number of bits into a fourth number of descriptive quantities, which is smaller than the third number.
9. A retention time prediction device, characterized in that: include: A pKa prediction module, used to predict the acid ionization equilibrium constant pKa of a compound to be predicted in multiple solutions, wherein the multiple pKas include pKas corresponding to multiple types of molecules matching the compound to be predicted, wherein the multiple types of molecules include acid molecules and / or base molecules divided based on ionizable functional groups; A LogD prediction module, for obtaining the oil partition coefficient LogD of the multi-ionized form compound of the compound to be predicted based on the multiple pKa; The retention time prediction module is used to input the retention time prediction description of the compound to be predicted into a preset retention time prediction model to obtain the predicted retention time of the compound output by the retention time prediction model, wherein the retention time prediction description includes the LogD.
10. An electronic device, characterized in that: include: Memory and processor; The memory is used to store a computer program, and the processor is used to run the computer program to implement the retention time prediction method according to any one of claims 1 to 9.