Determining HPLC method parameters using machine learning

Abstract

A computer-implemented method for predicting the HPLC retention time of one or more compounds using a machine learning model is disclosed. In particular, the present invention discloses the use of a machine learning model for identifying a set of HPLC process parameters suitable for separating two or more compounds in a composition. Related systems and articles are also described.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for predicting the HPLC retention time of one or more compounds using a machine learning model. In particular, the present invention relates to the use of a machine learning model to identify a set of HPLC method parameters suitable for separating two or more compounds in a composition, and to related systems and devices. Background Art

[0002] High performance liquid chromatography (HPLC) is the most commonly used analytical technique in the development of new drugs. From drug discovery and development to commercial manufacturing, HPLC is used for purity determination, stability studies, identification of degradation and byproducts, genotoxicity limit testing or preparative separations (Dong and Guillarme, 2013; Snyder et al., 2009).

[0003] During the development of HPLC methods, in order to achieve maximum separation between compounds, a wide range of parameters need to be optimized, such as the choice of mobile phase (eluent), the pH of the eluent, the gradient program, the flow rate, the column oven temperature, and the choice of the stationary phase (chromatographic column). Currently, a time-consuming and costly one-factor-at-a-time "trial and error" approach is often applied to find the optimal settings.

[0004] Therefore, it is desirable to provide methods that address the above-mentioned technical challenges. Summary of the invention

[0005] In summary, the present invention provides a machine learning model for predicting the HPLC retention time of one or more compounds and for identifying a set of HPLC method parameters suitable for separating two or more compounds in a composition.

[0006] Therefore, a first aspect provides a method for predicting the HPLC retention time of one or more compounds, the method comprising: (a) obtaining values ​​of one or more structural and / or physicochemical properties of the compounds and one or more HPLC method parameters; and (b) using a machine learning model to predict the retention time of each of the compounds when HPLC is performed using the one or more HPLC method parameters, wherein the machine learning model has been trained using a training data set, the training data set comprising molecular structural properties and / or physicochemical properties of one or more compounds, and one or more sets of chromatographic data for each compound, the one or more sets of chromatographic data comprising the retention time of the corresponding compound and the associated HPLC method parameters.

[0007] The inventors have determined that a machine learning model can be trained to predict with high predictive accuracy chromatographic data comprising retention times of corresponding compounds when separated using associated HPLC method parameters provided as input to the machine learning model. The inventors further recognized that these highly accurate predictions of HPLC methods can be used to identify HPLC method parameters suitable for separating two or more compounds in a composition.

[0008] Therefore, according to a second aspect, the present invention also describes a method for identifying a set of HPLC method parameters suitable for separating the compounds in a composition comprising two or more compounds, the method comprising: (a) performing the method of the first aspect for multiple sets of HPLC method parameters; (b) using the results of step (a) to calculate one or more separation performance metrics; and (c) identifying one or more sets of HPLC method parameters by applying one or more criteria to the separation performance metrics.

[0009] According to a third aspect, a method for providing a tool for predicting the HPLC retention time of one or more compounds is provided, the method comprising: (a) obtaining a training data set comprising values ​​of one or more structural and / or physicochemical properties of the one or more compounds, and one or more sets of chromatographic data for each compound, the one or more sets of chromatographic data comprising the retention time of the corresponding compound and associated HPLC method parameters; and (b) using the values ​​of one or more structural and / or physicochemical properties of the compound and the values ​​of the HPLC method parameters as input to train a machine learning model to predict the retention time of the compound when HPLC is performed using the one or more HPLC method parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of the HPLC system is shown.

[0011] Figure 2 depicts a flow chart of an example method for predicting HPLC retention times (A) and identifying a set of HPLC method parameters suitable for separating two or more compounds (B). Steps shown in dashed boxes are optional.

[0012] Figure 3 The gradient of SAM-0200368 is shown, with time points labeled tp_1, tp_2, and tp_3.

[0013] Figure 4 Is a star schema that represents the aggregated data.

[0014] Figure 5The definition and basic operation of the genetic algorithm according to the embodiment of the method of the present disclosure are shown. In the figure, Flow is the flow rate, Temp is the temperature, pH, Tp_1 is the time point 1, Tp_2 is the time point 2, Tp_3 is the time point 3, Ep_1 is the elution capacity 1, Ep_2 is the elution capacity 2, Ep_3 is the elution capacity 3, Length is the column length, inner_diameter is the column inner diameter, Particle size is the column particle size, Pore_size is the column pore size, H is the hydrophobicity, S is the spatial interaction, A is the hydrogen bond acidity, B is the hydrogen bond basicity, and c is the ion exchange capacity.

[0015] Figure 6 The basic operation of a genetic algorithm according to an embodiment of the method of the present disclosure is shown.

[0016] Figure 7 A prediction error plot of a machine learning model evaluated before feature selection using a random train-test-split in accordance with an embodiment of the disclosed method is shown.

[0017] Figure 8 A comparison of two splitting methods is shown. Left: Random training-testing-split (80 / 20), Right: Uniform training-testing-split (80 / 20). Y_train and y_test refer to the training data and the test data, respectively. With the "uniform" sampling method, an 80 / 20 ratio between training and test samples is maintained over the entire retention time range.

[0018] Fig. 9 The top 20 most important features of the XGB model using a uniform train-test-split are shown. The numbers are F-score values, which are also illustrated by the length of the bars.

[0019] Fig.10 Shown are the hyperparameters of the XGB models evaluated using different learning rates after a uniform train-test-split and feature selection. All models achieve good performance. Using learning_rate of 0.1 and 0.05, the models achieve good performance with only a few hundred estimators. The scores plotted are the average test scores of the 10-fold CV.

[0020] Fig.11 The prediction errors and residual plots of the final XGB model are shown. The residual plots show random errors indicating a good fit of the model over the entire RT range. The model performs best between 2min and 20min, probably due to the number of observations in this range.

[0021] Fig.12Prediction error and residual plots using molecular fingerprints and XGB are shown. The histograms show the distribution of the test set and the distribution of the predicted values.

[0022] Fig.13 Shown is a comparison of KMeans clustering with 30 clusters and 8 clusters using two ion exchange parameters (c). PCA was applied before clustering. "x" represents the cluster center.

[0023] Fig.14 Shown is a comparison of different eps values ​​for analyzing column clustering using DBSCAN and c_28.

[0024] Fig.15 Shown is a comparison of measured RTs to predicted RTs using the SAM method parameters and the XGB model trained specifically for the optimizer. The numbering of the peaks is defined by the RT from the Empower data (ie, measured RT, red peaks).

[0025] Fig.16 Shown is a comparison of predicted and measured results for a molecule in an example elution (labeled SAM-0114188) using method parameters from the optimizer.

[0026] Fig.17 Shown is a comparison of predicted and measured results for a molecule in an example elution (labeled SAM-0113392) using the method parameters recommended by the optimizer.

[0027] Fig.18 Screen shots of user interfaces for accessing the disclosed methods are shown. DETAILED DESCRIPTION

[0028] In describing the present invention, the following terminology will be employed and is intended to be defined as shown below.

[0029] As used herein, "and / or" will be considered as a specific disclosure of each of two specific features or components, with or without the other. For example, "A and / or B" will be considered as a specific disclosure of each of (i) A, (ii) B, and (iii) A and B, as if each were individually listed herein.

[0030] In addition, as used below, the terms "particularly", "more particularly", "specifically", "more specifically" or similar terms are used in conjunction with optional features without limiting the possibility of substitution. Therefore, the features introduced by these terms are optional features and are not intended to limit the scope of the claims in any way. As the skilled person will recognize, the present invention can be implemented by using alternative features. Similarly, the features introduced by "in one embodiment of the present invention" or similar expressions are intended to be optional features without any limitation on alternative embodiments of the present invention, without any limitation on the scope of the present invention, and without any limitation on the possibility of combining the features introduced in this manner with other optional or non-optional features of the present invention.

[0031] The compositions as described herein can be pharmaceutical compositions that additionally comprise a pharmaceutical carrier, diluent or excipient. The pharmaceutical composition can optionally comprise one or more additional pharmaceutically active polypeptides and / or compounds. Such formulations can be, for example, in a form suitable for intravenous infusion.

[0032] The compounds described herein can be small molecules (e.g., small molecule inhibitors, activators, cofactors, etc.) or macromolecules (e.g., biologics, therapeutic proteins or peptides such as antibodies or compounds derived therefrom, nucleic acids, etc.). The compounds can be organic compounds. The compounds can be pharmaceutically active agents (also referred to as drugs or therapeutic agents) or degradation products thereof.

[0033] The systems and methods described herein can be implemented in a computer system (including or in addition to the described structural components and user interactions). As used herein, the term "computer system" includes hardware, software and data storage devices for embodying the system or performing the method according to the above-mentioned embodiments. For example, a computer system may include a processing unit such as a central processing unit (CPU) and / or a graphics processing unit (GPU), an input member, an output member and a data storage device, which may be embodied as one or more connected computing devices. Preferably, the computer system has a display or includes a computing device with a display to provide a visual output display. The data storage device may include RAM, a disk drive or other computer-readable media. The computer system may include multiple computing devices that are connected via a network and can communicate with each other in the network. It is clearly envisioned that the computer system may be composed of or include a cloud computer.

[0034] The methods described herein may be provided as a computer program or as a computer program product or a computer-readable medium carrying a computer program, which is arranged to perform the methods described herein when executed on a computer. As used herein, the term "computer-readable medium" includes, but is not limited to, any one or more non-transitory media that can be directly read and accessed by a computer or computer system. The media may include, but is not limited to, magnetic storage media, such as floppy disks, hard disk storage media, and magnetic tapes; optical storage media, such as optical disks or CD-ROMs; electronic storage media, such as memory, including RAM, ROM, and flash memory; and hybrids and combinations of the above, such as magnetic / optical storage media.

[0035] High performance liquid chromatography (HPLC, also known as "high pressure liquid chromatography") is the most commonly used analytical technique in the development of new drugs. The advantage of HPLC is its almost universal applicability, as only a few samples cannot be measured using this technique. HPLC becomes even more important by combining it with other techniques such as mass spectrometry (LC-MS). From drug discovery and development to commercial manufacturing, HPLC is used for purity determination, stability studies, identification of degradation and byproducts, genotoxicity limit testing or preparative separations (Dong and Guillarme, 2013; Snyder et al., 2009). During the development of HPLC methods, a wide range of parameters need to be set. In order to achieve maximum separation performance between analytes, multiple parameters must be selected, such as the mobile phase (eluent), the pH of the eluent, the gradient program, the flow rate, the column oven temperature and the stationary phase (chromatographic column). Currently, determining the optimal parameters requires extensive manual optimization involving multiple trial and error steps, which is both time-consuming and leads to increased costs, although field available guidelines help to shorten this procedure to some extent.

[0036] In the context of the present invention, the term HPLC is also intended to encompass ultra high performance liquid chromatography, and HPLC performed alone or as part of an analytical or separation process, such as for example as part of LC-MS.

[0037] Figure 1A schematic diagram of an HPLC system is shown. After the liquid sample is injected using an autosampler, the eluent flow from the pump carries the sample through the HPLC system. Once the sample enters the chromatographic column, the sample is separated into different components. The separation mode is mainly determined by the chromatographic column. There are many types of HPLC, such as reverse phase (RPLC), normal phase (NP-LC), hydrophilic interaction (HILIC), ion pair (IPC), size exclusion (SEC) and affinity chromatography. The method described herein is applicable to all types of HPLC. Currently, RP-LC is the most commonly used separation mode in the development process of new small molecule drugs, especially for water-soluble samples (Dong, 2006). In an embodiment, HPLC can be RP-LC. In RP-LC, the stationary phase (called a chromatographic column) is non-polar. The chromatographic column is a stainless steel tube filled with spherical particles made of porous modified silica. The most common modification in RP-LC is C-18. These C-18 groups are attached to the surface of each particle, providing the chromatographic column with its non-polar properties. The mobile phase consists of a polar mixture of an organic solvent (common examples include acetonitrile or methanol) and water. The different migrations ("travel speeds") of sample molecules (called analytes) lead to separation in the chromatographic column. The migration of the analyte is determined by the equilibrium process between the molecules of the same analyte present in the mobile phase and the stationary phase at any time. The stronger the binding of the analyte to the stationary phase, the stronger the retention of the stationary phase to it, and the longer its retention time (RT). In addition to the chemical structure of the analyte, the mobile phase, the stationary phase, and the temperature also affect the migration rate. The sample solvent does not bind to the chromatographic column due to its polarity and leaves the chromatographic column first. After being separated on the chromatographic column, the different molecules leave the chromatographic column and enter the detector. The most common type is the ultraviolet absorption detector (LC-UV). The UV detector measures the change in the fraction of light transmitted through the flow cell containing the sample over time. The relationship between the fraction of transmitted light and the concentration of the analyte in the sample can be described using the Beer-Lambert law. The output of the LC-UV experiment is a chromatogram, in which a graph of absorbance relative to retention time (RT) is plotted.

[0038] As a first aspect, the present invention provides a method for predicting the HPLC retention time of one or more compounds, the method comprising: (a) obtaining values ​​of one or more structural and / or physicochemical properties of the compounds and one or more HPLC method parameters; and (b) using a machine learning model to predict the retention time of each of the compounds when HPLC is performed using the one or more HPLC method parameters, wherein the machine learning model has been trained using a training data set, the training data set comprising molecular structural properties and / or physicochemical properties of one or more compounds, and one or more sets of chromatographic data for each compound, the one or more sets of chromatographic data comprising the retention time of the corresponding compound and the associated HPLC method parameters.

[0039] The method of this aspect may have any of the features described in relation to any of the other aspects.

[0040] The step of "obtaining" may include calculating the properties, receiving the properties from a user interface, or receiving the properties from a computing device or a database.

[0041] HPLC method parameters may include parameters related to the chromatography column (eg, size, particle size), parameters related to the eluent (eg, pH of the eluent), or parameters related to the chromatography procedure (eg, flow rate).

[0042] Figure 2AA schematic diagram is provided in . At step 100, one or more structures and / or physicochemical properties of one or more compounds are obtained. This may include obtaining the identification of the compound, for example, from a user at optional step 102. This may include an optional step 104 of determining the value of the structure / physicochemical property. Alternatively, these may be received from a user, a computing device, or a database. Properties may include molecular fingerprints and / or one or more molecular descriptors. In an embodiment, the structural molecular descriptor and / or molecular fingerprint is a 2D molecular descriptor or molecular fingerprint. In an embodiment, the molecular descriptor is selected from the group consisting of the following items: ABCIndex, AcidBase, AdjacencyMatrix, Aromatic, AtomCount, Autocorrelation, BCUT, BalabanJ, BaryszMatrix, BertzCT, BondCount, CarbonTypes, Chi, Constitutional, DetourMatrix, DistanceMatrix, EState, EccentricConnectivityIndex, ExtendedTopochemicalAtom, FragmentComplexity, Framework, Hydr In an embodiment, the parameters are calculated by the Mordred package. In an embodiment, the molecular descriptor is selected from the group consisting of SLogP, ATSC5v, ATSC8d, ATSC3Z, ABC, VSA EState4, ATSC5dv and ATSC6i.

[0043] At step 110, the value of one or more HPLC method parameters is obtained. These can be received from a user, a computing device or a database. For example, these can be received from an optimization algorithm performed on a processor. In an embodiment, the HPLC method parameter group is generated by a genetic algorithm. In an embodiment, the genetic algorithm is run until one or more stop criteria are applicable, and optionally wherein the stop criteria are selected from: i) a predetermined number of generations has been reached; and ii) the difference between the separation performance metric associated with one or more groups of HPLC method parameters of the current iteration and the separation performance metric associated with one or more groups of HPLC method parameters of the previous iteration is lower than a threshold value. In an embodiment, the genetic algorithm is initialized with a group of HPLC method parameters selected at random, wherein the genetic algorithm is initialized with a group of HPLC method parameters selected from a predetermined group or range of each HPLC method parameter, and / or wherein the genetic algorithm is initialized with a group of HPLC method parameters provided by a user. In an embodiment, one or more groups of HPLC method parameters are selected (e.g., by a genetic algorithm) from a corresponding predetermined group and / or range.

[0044] At step 120, the machine learning model is used to predict the retention time of each compound when subjected to HPLC using one or more HPLC method parameters. Optionally, the machine learning model further predicts a metric indicative of peak width, and one or more sets of chromatographic data in the training data set further include a metric indicative of the width of the peak of the corresponding compound.

[0045] At step 130 , the predicted value is presented to a user, such as through a user interface, to a computing device or a database.

[0046] In one embodiment, the machine learning model further predicts a metric indicative of peak width. In such embodiments, the machine learning model is trained using one or more sets of chromatographic data that further include a metric indicative of the width of a peak of a corresponding compound.

[0047] In one embodiment, the one or more compounds are individually selected from the group consisting of: pharmaceutically active agents and degradation products thereof.

[0048] In one embodiment, the HPLC method parameters are selected from the group consisting of: HPLC type, flow rate, temperature, pH, column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c-28), column ion exchange capacity (c-70), and one or more metrics defining the elution phase gradient. In one embodiment, the HPLC method parameters include one or more parameters selected from the group consisting of: column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c-28), column ion exchange capacity (c-70), and one or more metrics defining the elution phase gradient, and optionally one or more of the parameters selected from the group consisting of: HPLC type, flow rate, temperature, pH. In one embodiment, the HPLC method parameters include one or more of the parameters selected from the group consisting of HPLC type, flow rate, temperature, pH, and column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c-28), column ion exchange capacity (c-70), and one or more metrics defining the elution phase gradient. In one embodiment, the HPLC method parameters include one or more of the parameters selected from the group consisting of column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c-28), column ion exchange capacity (c-70), and one or more of the parameters optionally selected from the group consisting of HPLC type, flow rate, temperature, pH, and one or more metrics defining the elution phase gradient. In one embodiment, the HPLC method parameters include i) HPLC type, flow rate, temperature, one or more metrics defining the elution phase gradient, pH, and ii) one or more parameters selected from the group consisting of: column length, column inner diameter, column particle size, column pore size, column steric interactions, column hydrogen bonding acidity, column hydrogen bonding basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70).

[0049] In embodiments, the HPLC type can be, for example, normal phase (NP) chromatography, reverse phase (RP) chromatography, size exclusion chromatography, ion exchange chromatography, hydrophilic interaction chromatography (HILIC) and affinity chromatography. In embodiments, the HPLC type is RP or NP chromatography. Specifically, the HPLC type can be RP chromatography.

[0050] In one embodiment, the one or more metrics defining the elution phase gradient include:

[0051] (i) the elution capacity of each of the plurality of mobile phases; and / or

[0052] (ii) defining a metric for a time-dependent change in the ratio of a plurality of mobile phases during elution, optionally wherein the one or more metrics defining a time-dependent change in the ratio of a plurality of mobile phases during elution comprises values ​​at one or more time points corresponding to a change in the rate of change of the ratio of the mobile phases, and / or values ​​of the rate of change of the ratio of the mobile phases at one or more time points during elution, and / or a metric describing a function representing a change in the ratio of the mobile phases in one or more (e.g., each) of a plurality of regions in the gradient; and / or

[0053] (iii) the elution capacity of the mobile phase at two or more (e.g., three) different time points, optionally wherein the mobile phase comprises a plurality of mobile phases, and obtaining the elution capacity at one time point as the sum of the elution capacities of each of the plurality of mobile phases weighted by the corresponding proportion of each mobile phase.

[0054] In one embodiment, one or more physicochemical properties include one or more molecular descriptors and / or one or more molecular structural properties include molecular fingerprints and / or one or more structural molecular descriptors. In an embodiment, one or more physicochemical properties and / or molecular structural properties are calculated by PaDEL-Descriptor, BlueDesc, ChemoPy, PyDPI, Rcpi, Cinfony or Dragon software, preferably by Mordred package. Any package in the art that can be used to calculate the physicochemical properties and / or molecular structural properties of molecules, for example, from a molecular formula, a 2D or 3D structure, can be used.

[0055] In one embodiment the structural molecular descriptor and / or molecular fingerprint is a 2D molecular descriptor or molecular fingerprint.

[0056] In one embodiment, the molecular descriptor is selected from the group consisting of: ABCIndex, AcidBase, AdjacencyMatrix, Aromatic, AtomCount, Autocorrelation, BCUT, BalabanJ, BaryszMatrix, BertzCT, BondCount, CarbonTypes, Chi, Constitutional, DetourMatrix, DistanceMatrix, EState, EccentricConnectivityIndex, ExtendedTopochemicalAtom, FragmentComplexity, Framework, Hyd rogenBond, InformationContent, KappaShapeIndex, Lipinski, McGowanVolume, MoeType, MolecularDistanceEdge, MolecularId, PathCount, Polarizability, RingCount, RotatableBond, SLogP, TopoPSA, TopologicalCharge, TopologicalIndex, VdwVolumeABC, VertexAdjacencyInformation, WalkCount, Weight, WienerIndex, and ZagrebIndex. In an implementation scheme, the parameters are calculated by the Mordred package.

[0057] In one embodiment, the one or more structural and / or physicochemical properties include molecular descriptors selected from the group consisting of SLogP, ATSC5v, ATSC8d, ATSC3Z, ABC, VSA_EState4, ATSC5dv, and ATSC6i. In an embodiment, the parameters are calculated using the Mordred package.

[0058] In one embodiment, the elution phase gradient (in the training data or in the HPLC method for predicting retention time) is a fixed gradient. In an embodiment, the HPLC method represented in the training data and / or used to predict RT has a parameter selected from the following: a flow rate of 0.2 ml per minute, and an elution power (for the mobile phase or its components, e.g., determined as the summary value of all components of the mobile phase at the corresponding time points) of about 1.05, 1.50, and 1.95 at 3 time points (such as 0, 15, and 30 minutes).

[0059] In one embodiment, the HPLC method parameters used for predicting the retention time at step (b) are restricted to respective predetermined ranges and / or selected from corresponding predetermined value sets, and / or wherein the training data set comprises a chromatographic data set obtained using the HPLC method parameters within the corresponding ranges and / or the corresponding value sets, and the HPLC method parameters used for predicting the retention time at step (b) are within said ranges and / or selected from said value sets.

[0060] In one embodiment, the elution phase gradient is a non-fixed gradient.

[0061] In one embodiment, the machine learning model is an extreme gradient boosting model (e.g., extreme gradient boosting tree), a gradient boosting model (e.g., gradient boosting tree), a random forest model, a Lasso regression model, or a support vector machine, preferably an extreme gradient boosting model.

[0062] In one embodiment, the machine learning model is trained to minimize the difference between the predicted retention times and the corresponding retention times in the training data.

[0063] In one embodiment, the machine learning model has been trained using a dataset comprising data for at least 100, at least 200, at least 300, at least 400, at least 500, or at least 600 compounds.

[0064] In one embodiment, the data set comprises at least 2, 3, 4 or 5 sets of chromatographic data comprising retention times of corresponding compounds and associated HPLC method parameters.

[0065] In embodiments, training data includes multiple data points (e.g., at least 100, 200, 500, 1000, 1500), each data point includes the molecule / physicochemical properties of the compound, the corresponding retention time and the HPLC method parameters. In embodiments, data includes multiple data points for the same compound with the same HPLC method, multiple data points for the same compound with different HPLC methods and multiple compounds with the same HPLC method. Advantageously, the training data set includes data for different compounds (such as at least 100, at least 200, at least 300, at least 400, at least 500 or at least 600 compounds). Also advantageously, training data includes data points for different HPLC methods, such as at least 10, at least 20, at least 50, at least 100 or at least 150, at least 200 groups of different HPLC method parameters (wherein if at least one method parameter is different, the HPLC method parameter group is different). In embodiments, the training data set includes a measure of the width of the peak (i.e., the peak associated with the retention time of the corresponding compound) indicating the corresponding compound.

[0066] In one embodiment, HPLC is reverse phase chromatography, normal phase chromatography, size exclusion chromatography, ion exchange chromatography or hydrophilic interaction liquid chromatography, optionally wherein HPLC is reverse phase chromatography or normal phase chromatography.

[0067] In one embodiment, the machine learning model has been trained using a set of predictive features selected from a larger group by a feature selection process, and wherein the larger set of predictive features are each selected from: molecular structural properties or physicochemical properties (e.g., molecular descriptors or molecular fingerprints) and HPLC method parameters.

[0068] As a second aspect, the present invention provides a computer-implemented method for identifying a set of HPLC method parameters suitable for separating two or more compounds in a composition comprising said compounds, the method comprising:

[0069] (a) performing the method of any one of the embodiments of the first aspect for multiple sets of HPLC method parameters;

[0070] (b) using the results of step (a) to calculate one or more separation performance metrics; and

[0071] (c) identifying one or more sets of HPLC method parameters by applying one or more criteria to the separation performance metric.

[0072] Figure 2B A schematic diagram is provided in FIG. 140 . At step 140 , one or more sets of values ​​for HPLC method parameters are obtained. This may include obtaining the values ​​for one or more HPLC method parameters from an optimization algorithm. At step 150 , the retention time (and optionally a metric indicating peak width) is predicted by the method of the first aspect (e.g., Figure 2A 150). At step 160, one or more separation performance metrics are calculated using the results of step 150. At step 170, one or more sets of HPLC method parameters are identified by applying one or more criteria to the separation performance metrics. At step 180, the results are made available for output, such as by outputting them to a user, such as through a user interface, to a computing device or a database.

[0073] The method of this aspect may have any of the features described in relation to any of the other aspects.

[0074] In one embodiment, the separation performance metric includes one or more of the following: i) the average retention time difference between adjacent peaks corresponding to the predicted retention times of two or more compounds; ii) the gradient duration; and iii) a predicted metric indicative of peak width, optionally including both i) and ii).

[0075] In one embodiment, the separation performance metrics include: i) the average retention time difference between adjacent peaks corresponding to the predicted retention times of two or more compounds, wherein the criterion includes maximizing the average retention time difference; ii) gradient duration, wherein the criterion includes minimizing the gradient duration; and / or iii) predicted peak width, wherein the criterion includes minimizing the peak width.

[0076] In one embodiment, the one or more sets of HPLC method parameters in step (a) are generated by an optimization algorithm, optionally a genetic algorithm.

[0077] In one embodiment, step (a) comprises running the genetic algorithm until one or more stopping criteria apply, optionally wherein the stopping criteria are selected from: i) a predetermined number of generations has been reached; and ii) the difference between the separation performance metric associated with one or more sets of HPLC method parameters of the current iteration and the separation performance metric associated with one or more sets of HPLC method parameters of the previous iteration is below a threshold.

[0078] In one embodiment, the genetic algorithm is initialized with a randomly selected set of HPLC method parameters, wherein the genetic algorithm is initialized with a set of HPLC method parameters selected from a predetermined set or range for each HPLC method parameter, and / or wherein the genetic algorithm is initialized with a set of HPLC method parameters provided by a user.

[0079] In one embodiment, one or more sets of HPLC method parameters are selected (eg, by a genetic algorithm) from corresponding predetermined sets and / or ranges.

[0080] In one embodiment, the method includes presenting to the user one or more groups (e.g., 5 groups) of HPLC method parameters that meet one or more standards of the separation performance metric. Presenting to the user can be, for example, performed through a user interface, presented to a computing device or a database.

[0081] According to a third aspect, a method for providing a tool for predicting the HPLC retention time of one or more compounds is provided, the method comprising: (a) obtaining a training data set comprising values ​​of one or more structural and / or physicochemical properties of the one or more compounds, and one or more sets of chromatographic data for each compound, the one or more sets of chromatographic data comprising the retention time of the corresponding compound and associated HPLC method parameters; and (b) using the values ​​of one or more structural and / or physicochemical properties of the compound and the values ​​of the HPLC method parameters as input to train a machine learning model to predict the retention time of the compound when HPLC is performed using the one or more HPLC method parameters.

[0082] The method of this aspect may have any of the features described in relation to any of the other aspects.

[0083] In another aspect, a computer program product is disclosed that includes computer readable instructions that, when executed by one or more processors, cause the one or more processors to perform a method of any embodiment of any method described herein.

[0084] In another aspect, a non-transitory computer-readable medium having stored thereon computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform a method of any embodiment of any method is disclosed.

[0085] In another aspect, a system is disclosed that includes: at least one processor; and at least one non-transitory computer-readable medium containing instructions that, when executed by the at least one

[0086] The following are presented by way of example and should not be construed as limiting the scope of the claims.

[0087] Examples

[0088] Example 1

[0089] This study aimed to evaluate the use of ML for improving the HPLC method development process. To improve efficiency, a recommendation tool was developed to suggest appropriate parameters for a given sample mixture. These parameters should be used as a starting point for trained analysts to fine-tune the precise values ​​of various HPLC parameters. The tool utilizes a supervised ML model to predict the RT of analytes in samples. The model was trained using the following input data: i) molecules and their molecular descriptors, ii) HPLC method parameters, iii) HPLC column parameters from a hydrophobic subtraction model, and iv) chromatographic data.

[0090] The tool uses RT predictions to evaluate different combinations of method parameters to maximize the separation between analytes. A single-page web application was implemented for easy user access.

[0091] Another aim of this study was to reduce the number of different HPLC columns that would be needed during the HPLC method development process.Since different suppliers sell columns with similar properties, an unsupervised clustering algorithm was used to group columns with similar properties.

[0092] method

[0093] Technical implementation

[0094] All codes were written using Python version 3.9.7 (Rossum, 2021). Machine learning models were implemented using Scikit-Learn version 1.0.2 (Pedregosa et al., 2011). Extreme gradient boosting models were implemented using the Scikit-Learn wrapper interface of XGBoost version 1.5.0 (Chen and Guestrin, 2016). Mol files, chemical descriptors, and fingerprints were generated using RDKit version 2022.03.1 (“RDKit: Open Source Cheminformatics,” n.d.; https: / / www.rdkit.org). PyGAD version 2.17.0 was used to implement genetic algorithms (Fawzy, 2021).

[0095] Data collection and preprocessing

[0096] Chromatographic data

[0097] Collect chromatographic data from Empower chromatography data system (it is chromatography data software). Download chromatographic data in JavaScript object notation (JSON) format. Since analysts can process each injection multiple times, there are multiple results for each injection. Assume that in the chromatogram with the latest result_id, the peak is appropriately integrated and marked. Extract the RT and label of each integrated peak from these results and other relevant information such as method and store in CSV file for further processing.

[0098] After downloading each integrated peak, merge the data from the two Empower locations. Among the downloaded peaks, many peaks do not have names that can be associated to internal molecular identifiers (e.g., Roche numbers). This can be used to download molecular files that identify molecules. In order to associate abbreviations with Roche numbers, information from Roche's internal small molecule impurity database (SMIDB) and late-stage method database (Excel file) is used. The information from SMIDB can be obtained using the SMIDB API. The data from the late-stage method database must be manually merged before exporting the Excel file in CSV format. The two data sources are merged and matched with the peak names from the Empower data set.

[0099] HPLC instrumental method

[0100] The parameter settings that describe how to run an HPLC experiment are stored in the instrument method. These parameter settings directly affect the retention time of the analytes in the sample. Therefore, the parameters of the method are needed as input features for the ML model. Extract data corresponding to the following parameters: Method_id (SAM document number), chromatographic column, temperature, mobile phase A, mobile phase B, mobile phase C, mobile phase D, pH, gradient (time points and composition), and flow rate.

[0101] Mobile phase composition

[0102] The variable elution_power is introduced. The higher the elution power of the mobile phase composition, the shorter the expected retention time of the analyte. To calculate the "elution power" of the mobile phase, each component in the mobile phase and its proportions are extracted from the SAM document. Using this information, the mol% of each component in the mobile phase is calculated. To obtain the elution powers shown in Table 1, only the main solvents are considered, namely water, acetonitrile, methanol and isopropanol. The mol% of each solvent is multiplied by its elution strength to give the elution power. The elution strengths of the solvents are shown in Table 2. In NP, water has the highest elution strength (i.e. more water = shorter retention time), while in RP, water has the lowest elution strength (i.e. more water = longer retention time). Since only values ​​for NP are available (Trappe, 1940), it was decided to use the reverse NP value as the elution strength in RP-LC. The elution power of the mobile phase is the sum of the elution strengths of each component (e.g. the elution power of mobile phase A in SAM-0200368 is 0.98+0.04=1.02).

[0103] Table 1: Example for calculating the elution capacity of each component in SAM-0200368.

[0104] method_id Eluent Components unit quantity Mol-% elution_strength SAM-0200368 mobile_A water the remaining - 98.21 0.98 SAM-0200368 mobile_A ACN %v / v 5.00 1.78 0.04 SAM-0200368 mobile_A TFA %v / v 0.05 0.01 0.00 SAM-0200368 mobile_B CAN the remaining - 99.96 2.00 SAM-0200368 mobile_B TFA %v / v 0.05 0.04 0.00

[0105] Gradient information

[0106] The variables time_point_1, time_point_2, time_point_3 were introduced to describe the mobile phase at three different time points. For each instrument method, these time points were manually extracted together with the mobile phase composition at that time point. Figure 3The position of the time point of SAM-0200368 is shown. At each time point, the elution capacity is calculated by multiplying the portion of each mobile phase by its elution capacity and adding these values. For example, SAM-0200368: at time point 2, the mobile phase is composed of 60% A and 40% B. Therefore, the elution capacity at this time point is calculated by 0.60*1.02+0.40*2.00=1.41). In Table 3, the time points and elution capacities of SAM-0200368 are depicted. These six variables are used as features for machine learning models to represent gradients and mobile phases.

[0107] Table 2. Elution strengths of solvents used to calculate the elution capacity of the mobile phase. The inverse of the elution strength in NP-LC is used for RP-LC.

[0108] Solvents Elution Strength NP Elution strength RP Acetonitrile 0.5 2.0 Isopropyl alcohol 0.6 1.7 Methanol 0.7 1.4 water 1.0 1.0

[0109] Table 3. Examples of mobile phase compositions (tp = time point, ep = elution power).

[0110] method_id tp_1 tp_2 tp_3 ep_1 ep_2 ep_3 SAM-0200368 2.00 8.50 12.00 1.07 1.41 1.95

[0111] HPLC Column Information

[0112] Access ACDC database

[0113] The Analytical Column Database Compilation (ACDC) contains information about every HPLC and GC column in the PTDC. Using the ACDC's REST API, information such as column ID, manufacturer, stationary phase type, diameter, particle size, product number, USP code, pore size, and surface area is downloaded and stored locally in a CSV file.

[0114] Access USP Database

[0115] Since there is no REST API available, a web scraper was used to download the parameters of the hydrophobic subtraction model for more than 750 stationary phases from the United States Pharmacopeia (USP) web page. Since the data is continuously updated, manual download of the data is not advisable. Using Google Chrome's developer tools, the XMLHttpRequest (XHR) sent by the browser while loading the data of the requested web page was intercepted. A python script was used to mimic these XHR requests. The responses sent by the web page were stored locally in a CSV file.

[0116] Combines analytical column information from the ACDC database and the PQRI database

[0117] Information obtained from the ACDC and USP web pages was merged to produce a dataset containing stationary phase information for columns used in SAM methods, including characteristics such as column dimensions, particle size, pore size, PQRI parameters, and a number describing how often a column was used in SAM. Fuzzy string matching was used to merge data sources based on stationary phase name.

[0118] Chemical structure

[0119] For retention time prediction, molecular descriptors and molecular fingerprints are evaluated to represent the chemical structure of the molecule. The Integrated Roche Chemical Information (IRCI) database is a web application that stores the molecular structures of all registered molecules within Roche. Other databases of molecular structures, such as ChEMBL (https: / / www.ebi.ac.uk / chembl) can be used. The IRCI REST API is used to download the mol files for each peak in the cleaned Empower dataset (about 600 molecules). Using RDKit, the mol files are converted to specific RDKit mol files to calculate molecular descriptors using the Mordred package (Moriwaki et al., 2018). Use RDKit to generate fingerprints.

[0120] Combining data sources

[0121] The final step after collecting and preprocessing the various raw data sources is to integrate all the data to produce two different datasets ready to be used for training ML models. One dataset holds the molecular descriptors and the other the molecular fingerprints. Figure 4 A star schema representing the final data set is shown.

[0122] Retention time prediction

[0123] Different ML algorithms were evaluated using a molecular descriptor dataset. The general procedure was a random 75 / 25 training-test-split, followed by hyperparameter tuning using the GridSearchCV method provided by Scikit-learn and 10-fold cross validation on the training set. After defining the ideal values ​​for the hyperparameters, the new model was trained on the entire training set and its performance was evaluated using the test set. Only XGB was used to evaluate the use of molecular fingerprints for RT prediction (purely because of time constraints; other methods are also expected to be suitable).

[0124] Model Evaluation

[0125] Using the correlation coefficient (R 2 ), mean absolute error (MAE), and mean absolute percentage error (MAPE) to evaluate the performance of the model. 2In addition to the MAE (the most common metric for regression problems) (Müller and Guido, 2016), MAE is also calculated to provide a sense of how good the predictions are in minutes. MAPE is used to explain the magnitude of the RT. 2 Used to select optimal hyperparameters during grid search.

[0126] Random train-test-split vs. uniform train-test-split

[0127] Before training the ML model, the dataset must be split into a training set and a test set. To perform a random split, the random train_test_split method from Scikit-learn was used. The results of the model using random train-test-split showed that the prediction performance decreased with increasing retention time, and the random state defined in the method had an impact on the prediction score. Therefore, a unified train-test-split using an 80 / 20 split was implemented. To do this, the dataset was first sorted by retention time, and then every 5 observations were put into the test set. In addition, it was decided to limit the retention time to 30min because there were too few observations to achieve a 5:1 training-test ratio in the case of RT>30min.

[0128] Feature Scaling

[0129] For linear models and SVMs, it is important to scale the features before training the models. If scaling is not applied, features with higher values ​​have a greater impact than features with smaller values ​​(Müller and Guido, 2016). The MinMaxScaler provided by Scikit-learn is used to scale the input features. Normally, each feature is scaled individually, but since the time point and elution capacity features are related to each other (time points 1 to 3 and elution capacity 1 to 3), a special procedure is applied. For both time points and elution capacity, NumPy (Harris et al., 2020) arrays holding all values ​​from the various time points are created in order to fit the scaler. This way, all features have the same minimum and maximum values. After fitting the scaler, each feature is then transformed individually. In contrast to linear models and SVMs, tree-based models do not require scaling of the input data since each feature is processed individually (Müller and Guido, 2016).

[0130] Virtual Model

[0131] The dummy model acts as a baseline against which the results of the other models are judged. If the model achieves a higher score than the dummy regressor, then it can be said that the predictions are better than predicting random values ​​(Müller and Guido, 2016). The dummy regressor predicts the same output for each observation. In this case, the mean retention time (11.6min) is predicted from the observations in the test set.

[0132] Feature Selection XGB

[0133] Since feature importance is calculated in XGB, models with different numbers of features can be evaluated. Reducing the number of features helps prevent model overfitting and, therefore, improves prediction accuracy through better generalization (Müller and Guido, 2016). In addition, less complex models speed up both learning and prediction. The following number of features were evaluated: [5, 10, 20, 50, 75, 100, 125, 150, 175, 200, 225, 250, 500, 640]. To select the nth most important feature, the “weighted” importance measure from the XGB model was applied.

[0134] Clustering of analytical columns

[0135] Since different vendors sell columns with similar properties, clustering was evaluated to group columns with similar properties based on a hydrophobic subtraction model. The goal of this approach is to reduce the variety of columns in the laboratory. This can be achieved by having the optimizer predict a set of columns (clusters) where all columns share similar properties. For clustering, KMeans and DBSCAN were evaluated. Since the ion exchange capacity (c in the hydrophobic subtraction model) is pH-dependent, two clusterings must be performed.

[0136] Principal component analysis

[0137] Before applying the clustering algorithm, principal component analysis (PCA) was evaluated to reduce the number of features from five to two. By rotating the dataset, PCA transforms the dataset to produce uncorrelated features. After rotating the dataset, usually only a subset of features is selected, depending on their importance for explaining the data (Müller and Guido, 2016). Using two features allowed plotting the clustering results. All features were scaled using MinMaxScaler before applying PCA.

[0138] KMeans

[0139] With KMeans, the number of clusters has to be defined by the user. This is done iteratively by varying the number of clusters and plotting the results after performing PCA on the dataset. It is important to find the ideal number of clusters. If the number is too high, no reduction in columns will be achieved, while a small number will result in clusters containing columns that do not produce similar outputs. With the defined number of clusters, clustering is also applied to the dataset without performing PCA beforehand.

[0140] DBSCAN

[0141] In contrast to KMeans, the number of clusters is defined by the hyperparameters eps and min_samples. eps defines the maximum distance between data points that are considered to be “in the neighborhood” of each other. min_samples is the number of observations that must be in the same neighborhood to form a cluster. If the number of observations in the same neighborhood is less than min_samples, these observations are declared to be noise (Müller and Guido, 2016). Since each column should be represented by a cluster, min_samples is fixed to 0. The following eps values ​​are tested: [0.04, 0.06, 0.08, 0.1].

[0142] Recommended Tools

[0143] The purpose of the recommendation tool is to find suitable HPLC instrument parameters to separate a set of predefined molecules. To achieve this, the tool uses an ML model for retention time prediction to find the method parameter combination that maximizes the separation between the analytes. There are a total of 18 method parameters to be defined. These parameters are all features in the ML model: flow rate, temperature, pH, timepoint_1, timepoint_2, timepoint_3, elution_power_1, elution_power_2, elution_power_3, column_length, column_inner_diameter, column_article_size, column_pore_size, H (hydrophobicity), S (steric interactions), A (hydrogen bond acidity), B (hydrogen bond basicity), c (ion exchange capacity) (https: / / www.usp.org / resources / pqri-approach-column-equiv-tool). All other features, which are molecular descriptors used by the ML model, are based on the selected analytes.

[0144] Optimization methods

[0145] Evaluate two different optimization methods: brute force and evolutionary algorithms.

[0146] Using brute force, every possible combination of method parameters must be tested to find the one that achieves the best separation of the analytes. Using 18 parameters, some of which have more than 40 different values, results in a dataset that is too large for a computer to handle. It was evaluated whether the possible values ​​for each parameter could be restricted to reduce the size of the dataset, but there would still be millions of combinations to test. Therefore, two evolutionary algorithms were considered: differential evolution (DE) and genetic algorithms (GA). Evolutionary algorithms are metaheuristics (algorithmic frameworks) that are often inspired by nature. The goal of metaheuristics is to solve complex optimization problems (Bianchi et al., 2008). In this study, the optimization problem was to find 18 method parameters that maximize the separation. The advantage of metaheuristics is that they can efficiently explore a search space that is too large for other methods such as brute force (Blum and Roli, 2003). Both DE and GA are population-based methods, meaning that the algorithm maintains and improves multiple candidate solutions over time (Teghem, 2010). The search space can be restricted by defining boundaries using DE and GA. Unlike DE, GA allows different values ​​to be used to define the search space. Since different values ​​are a requirement for this study, it was decided to continue developing the recommendation tool using GA.

[0147] Genetic Algorithm for Method Parameter Selection

[0148] Using PyGAD, a GA can be implemented by providing a fitness function to evaluate candidate solutions, defining the gene space (search space), and setting parameters to guide the optimization process. The parameters to be defined are, in particular, the number of generations, the number of parent matings, the solutions in each generation, and the mutation rate and crossover behavior. Figure 5 The important elements of the genetic algorithm are shown. The solution, also called chromosome, consists of genes. A single gene describes a feature, such as flow rate or temperature. Therefore, one solution holds all the information used to create an instrument method. In each generation, the genes of the solution (parent) that reached a high fitness value are exchanged through crossover to produce the next generation (offspring). These offspring are evaluated again using the fitness function. In addition to crossover, randomly selected genes are mutated. During the mutation process, the current value is exchanged for a new value from the gene space for that specific gene.

[0149] Figure 5The general procedure of the optimization process using GA is shown. The first step in the process is to define the molecules to be separated by providing Roche numbers. Using Roche numbers, mol files are downloaded from IRCI, and the molecular descriptors of each molecule are calculated. Next, GA generates the first generation of solutions by randomly selecting values ​​from the gene space of each gene to create solutions. The generated descriptors and method parameters are merged to give solutions, and the retention time of each analyte is predicted. Using the predicted retention time and equation 1, the fitness value of the current solution is evaluated. The fitness function considers the average retention time difference (Δi) and gradient duration (t) between adjacent peaks (n). The higher the fitness, the better the expected separation between all analytes. The gradient duration in the fitness function is added as a penalty term to limit the running time. Based on the results of the first generation, GA creates new candidate solutions by crossover and random mutation between the best solutions of the first generation. The process of calculating the fitness of all solutions of the second generation is repeated. If a solution in the new solution reaches a fitness higher than the best solution from the first generation, the solution is stored as the best solution. Repeat this process until a predefined number of generations is reached. best_solution is then used to set up the instrument method.

[0150] Equation 1:

[0151]

[0152] How Genetic Algorithms are implemented

[0153] The optimizer was set to run 100 generations, using 20 solutions in each generation. To create a new generation, the 10 lowest-fit solutions of the previous generation were replaced by 10 new solutions created by crossover and mutation of the 10 best-fit solutions. This behavior was achieved by using steady-state selection as the parent selection type in the parameters (M. Mitchell, 1996). The mutation type was set to random and the number of genes to mutate was set to one. Even though 18 method parameters had to be selected, the number of genes was set to 14. This is due to the fact that only certain combinations of H, S, A, B and c parameters, i.e. cluster centers, are allowed. If this constraint did not exist, the optimizer might choose parameters that would result in a non-existent column. The gene space for columns was set between 0 and 42, which corresponds to the number of unique stationary phases available in the data set. This approach allows evaluating the optimizer by comparing the predicted retention times with the measured retention times, since each number in the gene space corresponds to exactly one stationary phase. For future use, the tool can also suggest other columns that are in the same cluster as the predicted column. The analyst will then be able to select a chromatographic column that is already available in his or her laboratory. Depending on the number of clusters selected, the parameters H, S, A, B, and c are added to the solution. The c value (c_28, c_70) is selected depending on the pH of the solution (e.g., c_28 represents a pH less than 7, such as 2.8; and c_70 represents a pH of 7.0; see https: / / www.usp.org / resources / pqri-approach-column-equiv-tool). It was decided not to apply PCA before clustering so that the predicted parameters corresponded to exactly one chromatographic column. The gene space for column size was fixed to 150 mm x 4.6 mm. The particle size was set to 3.0, between the common 2.7 μm and 3.5 μm particles. The gene space for flow rate was set to [0.75, 1.0, 1.25] to avoid overpressure in the HPLC system combined with the defined column size. Two options for the gene space for defining the gradient (time point and elution capacity) are defined. The gradient can be fixed. The gradient extends from 5% B to 95% B in 30 min. If the gradient is not fixed, there are no restrictions and the GA will recommend the gradient.

[0154] Optimization constraints

[0155] Constraints are introduced to avoid the optimizer from suggesting impossible values. The following constraints are defined:

[0156] Time point 1 ≤ Time point 2 ≤ Time point 3: The running time can only increase with the passage of time.

[0157] • Elution capacity 1 ≤ Elution capacity 2 ≤ Elution capacity 3: The elution capacity must remain constant or increase during the gradient.

[0158] · RT latest eluting peak ≤ time point 3: The peak should elute during the gradient and not in the column wash step.

[0159] If time point 2 = time point 3, then elution capacity 2 = elution capacity 3: If the time point does not increase (ie only one step in the gradient), then the elution capacity cannot increase either.

[0160] For each violated constraint, the goodness-of-fit value is multiplied by 0.1.

[0161] Test Recommended Tools

[0162] To evaluate the optimizer, two instrument methods were randomly selected along with their corresponding standard suitability tests (SSTs). For evaluation, all data except the Roche number / method_id combination from the selected instrument method were used to train a new XGB model. The optimizer suggested parameters using two gradient options. When a column with standard dimensions was not available, the optimizer was rerun with adjusted parameters from an existing column to obtain correct retention time predictions. The analytes were measured using LC-MS (Agilent 1290LC and Agilent 6150MSD) using the optimized parameters. The results were evaluated by comparing the predicted retention times with the measured retention times. In addition, the retention times of all analytes were predicted using the parameters specified in the original method.

[0163] Web Application

[0164] The web application was implemented using Ploty Dash (“Collaborative Data Science at Ploty Technologies Inc.”, 2015). It allows the analyst to enter multiple analytes by providing information that identifies the compound (e.g., Roche number). In addition, the analyst can choose between two gradient options. The tool runs the optimization process multiple times (e.g., five times) because the results of GA (and other metaheuristics) can vary. This variation is caused by the randomly generated first generation and the crossover and mutation behavior of the algorithm. The trained analyst selects the run with the highest overall goodness of fit value or based on the available columns or predicted RT (e.g., shorter RT).

[0165] result

[0166] Data collection and preprocessing

[0167] Chromatographic data

[0168] A total of 523'038 injections (1'760'386 peaks) from BS10 and 152'330 injections (424'164 peaks) from BS12 were downloaded (accessed on 06 August 2022). After pre-processing of the chromatographic data, a total of 1'122'540 peaks with known SAM documents were retained. A Roche number could be assigned to 513'877 of these peaks. Since the datasets saved the same information multiple times (same molecule and same method_id), the data were grouped by Roche number and method_id, and the average retention time and standard deviation were calculated. This resulted in 1'800 unique Roche number / method_id combinations and a total of 600 unique Roche numbers

[0169] HPLC instrumental method

[0170] 240 different SAM files were processed, which included manually extracted mobile phase compositions and gradient programs.

[0171] HPLC Column Information

[0172] A total of 1993 unique columns (based on product number) were downloaded from ACDC. These columns were manually associated to SAM documents. 107 of these 1993 columns were used in one of the SAMs in the SAM. 757 different column types are available on the USP webpage (accessed May 2022). The downloaded parameters of the hydrophobic subtraction model were associated to the columns used in the SAM document. After association, 94 different columns (based on product number) were left in the data set, which contained both size and hydrophobic subtraction model parameters.

[0173] Chemical structure

[0174] All 1613 available 2D molecular descriptors from the Mordred package were calculated. Each descriptor that could not be calculated for each molecule was later discarded from the dataset. This resulted in a dataset with 622 different descriptors. RDKit was used to calculate the Morgan fingerprint with 1024 bits for each mol file.

[0175] Retention time prediction using molecular descriptors

[0176] Evaluation of different machine learning models

[0177] The linear lasso model achieved a prediction score of 0.663 (R 2 ): α: 0.001, tolerance: 1e-06, max_iterations: 500. SVR achieved an R of 0.743 using C: 1000 and γ: 0.01 2and a MAE of 2.4min. The tree-based model with the lowest predictive performance is the one with bootstrap:True, max_depth:16, min_samples leaf:1, num_estimators:175 with an R of 0.770 2 GB and XGB show R values ​​of 0.802 and 0.811, respectively. 2 Similar results were obtained for GB. For GB, a learning rate of 0.1 was used along with max_depth of 5, min_samples leaf of 100 and 1500 estimators. To train the XGB model, a learning rate of 0.05, max_depth of 4, max_leaves of 0, min_child weight of 4 and 750 estimators were used. The dummy regressor used to define the baseline score achieved a MAE of 5.9min by always predicting the mean retention time of the training data. Each evaluated model outperformed this baseline score. The results of the evaluated models are summarized in Table 4 and are presented in Figure 7 The prediction error plots are depicted in . While XGB and GB show good fit over the entire RT range, RF and Lasso overestimate the RT in the lower range and underestimate the RT for observations with RT>20min. SVR achieves similar performance over the entire RT range but achieves overall lower performance compared to XGB and GB. The results show that these two gradient boosting models have the highest prediction performance. Since model training using XGB is significantly faster compared to GB, it was decided to further optimize the XGB model.

[0178] Table 4. Predictive performance of ML models evaluated before feature selection using random train-test-split and molecular descriptors.

[0179] Model <![CDATA[R 2 ]]> MAE MAPE virtual -0.011 5.9 123.3 Lasso 0.663 2.9 47.4 SVR 0.743 2.4 37.5 RF 0.770 2.2 31.8 GB 0.802 2.0 27.3 XGB 0.811 1.9 26.5

[0180] Adjusting the XGB model

[0181] The performance of the XGB model is further improved by evaluating the unified train-test-split and applying feature selection.

[0182] Unified training-testing-split

[0183] Figure 8 The histograms of the two sampling methods before limiting the retention time are shown. Using this sampling method, R 2 The ,value ,increases from 0.811 to 0.830 while the MAE decreases from 1.9 min to 1.6 min . ,The same hyperparameter settings are used as described in Section 3.3.1. Fig. 9The 20 most important features of the XGB model when using a unified train-test-split are shown. In addition to the instrumentation method features, the most important features are:

[0184] • SLogP: Octanol-water partition coefficient (Wildman and Crippen, 1999).

[0185] ATSC: Autocorrelation of Topological Structure, also known as Centralized Moro-Brotto Autocorrelation. ATSC measures the distribution of atomic properties on a molecular graph (DR Todeschini and Consonni, 2020).

[0186] • ABC: An atom-bond connectivity index that shows excellent correlation with the heat of formation of alkanes (Estrada, 2008).

[0187] VSA_EState: A hybrid of the van der Waals surface area and the electrotopological state (Guha and Willighagen, 2012).

[0188] Feature Selection

[0189] Table 5 shows the performance of the model using different numbers of features. The 65 most important features were used to achieve the highest score. Since the required features particle_size and inner_diameter were not among the first 65, these two features were added manually. Using feature selection, the number of features was reduced from 640 to 67 while improving the prediction score. The following hyperparameters were used for evaluation: γ: 0, learning_rate: 0.05, max_depth: 4, 'max_leaves: 0, min_child_weight: 4, n_estimators: 750.

[0190] Table 5. Prediction performance of the XGB model as a function of the most important features. The lowest MAE was achieved using 65 features and more.

[0191] Number of features <![CDATA[R 2 ]]> MAE 5 0.753 2.2 10 0.799 1.8 20 0.807 1.8 50 0.818 1.7 65 0.824 1.6 75 0.827 1.6 100 0.827 1.6 150 0.827 1.6 200 0.827 1.6 250 0.827 1.6 640 0.827 1.6

[0192] Hyperparameter Tuning After Feature Selection

[0193] After significantly reducing the number of features in the model, the hyperparameters of XGB were evaluated again using the parameter grid depicted in the code listing below.

[0194] param_grid_feature_selection={

[0195] 'n_estimators':[1,10,20,50,100,200,500,750,1000,1250,1500],

[0196] 'learning_rate':[0.01,0.05,0.1],

[0197] 'max_depth':[3,4,5],

[0198] 'max_leaves': [0],

[0199] 'min_child_weight':[4],

[0200] 'gamma':[0]}

[0201] Fig.10 Results with different values ​​of n_estimators, max_depth, and learning_rate are shown. The following parameters were used to achieve the best R2: n_estimators: 750, learning_rate: 0.05, max_depth: 4, max_leaves: 0, min_child_weight: 4, and γ: 0. All other parameters used the default values. R2 was improved by unified training-test-split and feature selection followed by hyperparameter tuning. 2 From 0.811 to 0.827, MAE decreased from 1.9min to 1.6min, and MAPE decreased from 26.5% to 21.5%. In Table 6, the average test score of the 10-fold CV is shown as a function of the number of estimators. With more than 750 estimators, a small drop in the prediction score is observed, which is an indication of overfitting. Fig.11 Prediction error and residual plots of the final model are shown.

[0202] Table 6. Average test scores of the XGB model after uniform training-test-split and feature selection using different numbers of estimators. The highest score was achieved using 750 estimators. Performance degradation was observed with 1000 or more estimators.

[0203] n_estimators 100 250 500 750 1000 1250 1500 2000 Score 0.78059 0.81042 0.81604 0.81675 0.81670 0.81624 0.81613 0.81579

[0204] Retention time prediction using molecular fingerprints

[0205] exist Fig.12 The results of RT prediction using molecular fingerprints and random training-test-split are depicted in Figure 2. The model achieved an R of 0.810. 2, MAE of 1.9min, and MAPE of 26.4%. These values ​​are comparable to the model using molecular descriptors. The best parameters evaluated using GridSearchCV were learning_rate of 0.05, 3000 estimators, max_depth of 4, max_leaves of 0, min_child_weight of 4, and γ0.

[0206] Clustering of analytical columns

[0207] Clustering with PCA

[0208] As the features are reduced from five to two (the first and second principal components), approximately 85% of the variance of the dataset can be explained. Fig.13 The results of KMeans clustering after applying PCA are shown. By comparing the distribution of clusters in the two figures, the pH dependence of the ion exchange parameter "c" can be seen. As mentioned above, min_samples is fixed to 0 because each column should form a cluster when using DBSCAN. Fig.14 The clusters formed using DBSCAN combined with PCA using different eps values ​​are shown. It was decided to use the KMeans algorithm for the optimizer because KMeans can also be used to predict clusters for new observations (Pedregosa et al., 2011).

[0209] Optimizer

[0210] It was decided to omit the PCA step and test the optimizer using 43 clusters instead of 30. The optimizer uses 43 clusters (i.e. the number of different stationary phases in the data set) to evaluate the parameters of a specific column. This way, the predicted retention times can be compared to the measured retention times because the actual column parameters are used for the predictions rather than the parameters at the cluster center. To evaluate the optimizer, two SAM methods were randomly selected: SAM-0114188 and SAM-0113392.

[0211] Comparison of measured RT with predicted RT using SAM method parameters

[0212] The first test was to predict the RT for each molecule described in one of the two SAM documents. The model used for the optimizer was trained without a test set, but each Roche number / method_id combination from the SAM was discarded before training the model. Fig.15In the figure, the measured RT (from Empower data) of each molecule using the specified method and the predicted RT using the same method parameters are depicted. The overall accuracy of the prediction is good, with MAE of 0.10 min (SAM-0114188) and 1.09 min (SAM-0113392). The elution order of the predicted RT corresponds to the measured results of SAM-0114188. Changes in the elution order were observed for peaks 2 and 3 in SAM-0113392 ( Fig.15 b).

[0213] Compare predicted RTs to measured RTs using optimizer output

[0214] SAM-0114188

[0215] Table 7 shows the optimizer output for the molecules used in SAM-0114188 when fixed gradient is not used. The optimizer was run five times, and each solution had the same degree of fit. The method parameters between each solution were slightly different, but in the same range (e.g., the pH suggested in each solution was acidic). The suggested method parameters for the 5th run (index 4) were evaluated in the laboratory. Cluster 6 corresponds to Waters Symmetry C18 chromatographic column. Mobile phase A is composed of water + 0.1% TFA, and mobile phase B is composed of acetonitrile + 0.1% TFA.

[0216] Table 7. Optimizer output for molecules in SAM-0114188 without gradient limitation. In all outputs, flow rate is 0.75, tp1 is 0.0, tp3 is 26.0, ep1 is 1.15, ep2 is 1.25, column length is 150.0, column diameter is 4.6, particle size is 3.5, pore size is 120.0, and goodness of fit is 35.062.

[0217] index temperature pH Tp2 Ep3 cluster 0 25.0 3.0 10.0 1.30 32.0 1 25.0 2.0 10.0 1.30 32.0 2 20.0 3.0 10.0 1.35 32.0 3 20.0 3.0 6.0 1.30 32.0 4 25.0 2.0 10.0 1.35 6.0

[0218] Table 8 shows the output of the optimizer for SAM-0114188 using a fixed gradient. Cluster 6 corresponds to a Waters Symmetry C18 column. Mobile phase A is water + 0.1% TFA, and mobile phase B is acetonitrile + 0.1% TFA.

[0219] Table 8. Optimizer output for molecules in SAM-0114188 using fixed gradients, only evaluated solutions are shown.

[0220] index 4 Ep2 1.50 Flow rate 1.0 Ep3 1.95 temperature 20.0 Column length 150.0 pH 2.0 Column diameter 4.6 Tp1 0.0 Particle size 3.5 Tp2 15.0 Aperture 120.0 Tp3 30.0 cluster 6.0 Ep1 1.05 Goodness of fit 17.587

[0221] exist Fig.16 The comparison of the predicted RT with the measured RT is shown in . The proposed method parameters without fixed gradient ( Fig.16 a) results in good resolution between all peaks with the peaks eluting in the correct order. The difference between the predicted RT and the measured RT is not within the MAE of any peak. During the column cleaning step of the gradient, Peak 2 eluted when the %B reached 95. Another test was performed with a 5% increase in B at each time point of the predicted gradient. Although the RT difference for Peak 2 was still over 5 min (see Fig.16 b), but the difference between the predicted RT and the measured RT for each peak is reduced. Using the "fixed" gradient conditions, the measured peaks show good resolution and the correct elution order. Compared to the results without the fixed gradient, the analyte elutes too early, as shown in Fig.16 Depicted in c.

[0222] SAM-0113392

[0223] Table 9 shows the optimizer output for SAM-0113392 without gradient limiting. The solutions shown are those tested in the laboratory. Cluster 1 corresponds to a Phenomenex Kinetex Biphenyl column. A pH of 3 was achieved using the following mobile phases: A: water + 0.01% TFA, B: acetonitrile + 0.01% TFA. Fig.17 A comparison between the predicted RT of the optimizer and the measured RT is depicted in a. While peak 0 elutes too early compared to the predicted RT, peaks 2, 3, and 4 elute too late and do not show good resolution from each other.

[0224] Table 9. Optimizer output for molecules in SAM-0113392 without gradient restrictions. Only evaluated solutions are shown.

[0225] index 0 Ep2 1.45 Flow rate 0.5 Ep3 1.50 temperature 40.0 Column length 150.0 pH 3.0 Column diameter 3.0 Tp1 0.0 Particle size 2.6 Tp2 6.0 Aperture 120.0 Tp3 20.0 cluster 1 Ep1 1.30 Goodness of fit 10.120

[0226] Table 10 shows the optimizer output for SAM-0113392 using a fixed gradient. The solutions shown are those tested in the laboratory. Cluster 11 corresponds to Waters Acquity UPLC CSH C18. Mobile phase A consisted of 10 mM ammonium acetate in water adjusted to pH 6 with acetic acid. Mobile phase B consisted of acetonitrile + 0.01% acetic acid. Fig.17 Comparison between the predicted RT of the optimizer and the measured RT is shown in b. Peak 1 eluted significantly earlier than predicted, while peaks 3 and 4 changed elution order. Compared to the results without gradient limitation, the measured RT is more accurate, and good resolution between all peaks is observed.

[0227] Table 10. Optimizer output for molecules in SAM-0113392 using fixed gradients, only evaluated solutions are shown.

[0228] index 0 Ep2 1.50 Flow rate 0.2 Ep3 1.95 temperature 45.0 Column length 150.0 pH 6.0 Column diameter 2.1 Tp1 0.0 Particle size 1.7 Tp2 15.0 Aperture 120.0 Tp3 30.0 cluster 11 Ep1 1.05 Goodness of fit 10.872

[0229] Web Application

[0230] Fig.18 A graphical user interface of a web application prototype is shown. The web application was created using PlotlyDash ("Collaborative Data Science at Plotly Technologies Inc.", 2015) and allows users to define up to five molecules described as Roche numbers. The application allows users to choose between a fixed gradient, where only flow rate, temperature, pH and chromatographic column are optimized, and an unlimited option, where every parameter except column size is optimized. The optimization process can be started by clicking the Start button. Once the optimization is run, the solution appears on the result card. Trained analysts select the solution with the highest goodness of fit from the solutions produced. If the suggested gradient seems unreasonable, such as too flat, the analyst may choose another solution. Typically, many solutions will have the same goodness of fit value. In this case, a solution can also be selected based on the available chromatographic column.

[0231] in conclusion

[0232] In summary, a dataset of available HPLC data was used to train an ML model, which in turn was used to optimize analytical method development using GA. The developed tool can be used to suggest HPLC method parameter settings for a predefined set of analytes, which can be used as a starting point for analytical chemists. In order to evaluate the different solutions proposed by GA, the XGB model was used to predict the RT of the analytes. The XGB model was trained using n = 1191 observations, each consisting of 49 molecular descriptors and 18 method parameters. The model achieved R 2 =0.827 and MAE=1.6min. In order to provide access to the tool, a web application prototype was implemented. Based on the predicted RT using optimized parameters, the tool was able to find suitable settings for all evaluation cases. To the best of our knowledge, methods that take into account method parameters to predict RT have not been developed in other studies. Adding more data to the training data set is expected to improve the RT prediction accuracy of the ML model. Once the tool is deployed, it has the potential to improve efficiency by reducing the need for expensive and time-consuming one-factor-at-a-time methods during the development of new HPLC methods.

[0233] References

[0234] All references cited herein are incorporated herein by reference in their entirety, and for all purposes, to the same extent as each individual publication or patent or patent application is specifically and individually indicated to be incorporated by reference in its entirety. The specific embodiments described herein are provided by way of example, rather than by way of limitation. Any subheadings herein are included only for convenience and should not be construed as limiting the disclosure in any way.

[0235] Bianchi, L., Dorigo, M., Gambardella, LM, and Gutjahr, WJ (2008). A survey of metaheuristics for stochastic combinatorial optimization. Nature Computing, 8(2), 239-287.

[0236] Blum, C. and Roli, A. (2003). Metaheuristics in combinatorial optimization: An overview and conceptual comparison. ACM Computing Surveys (CSUR), 35(3): 268-308.

[0237] Bouwmeester, R., Martens, L., and Degroeve, S. (2019). A comprehensive and empirical evaluation of machine learning algorithms for small molecule LC retention time prediction. Analytical Chemistry, 91(5), 3694–3703.

[0238] Chen, T.. and Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785-794.

[0239] Dong, MW (2006). Modern HPLC for the Practicing Scientist. John Wiley & Sons.

[0240] Dong, MW and Guillarme, D. (2013). Recent developments in HPLC impacting pharmaceutical analysis: a brief review. American Pharmaceutical Review, 16, 36–43.

[0241] Estrada, E. (2008). Atom-bond connectivity and the energetics of branched alkanes. Chemical Physics Letters, 463(4-6), 422–425.

[0242] Gad, A.F. (2021). Pygad: An intuitive Python library for genetic algorithms. CoRR, abs / 2106.06158.

[0243] Guha, R. and Willighagen, E. (2012). A survey of the quantitative description of molecular structure. Current Topics in Medicinal Chemistry, 12(18), 1946-1956.

[0244] Haddad, P. R., Taraji, M., and Szücs, R. (2021). Prediction of analyte retention times in liquid chromatography. Analytical Chemistry, 93(1), 228-256.

[0245] Harris, CR, Millman, KJ, van der Walt, SJ, Gommers, R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor, J., Berg, S., Smith, NJ, Kern, R., Picus, M., Hoyer, S., van Kerkwijk, MH, Brett, M., Haldane, A., del Río, JF, Wiebe, M., Peterson, P., .... Oliphant, TE (2020). Array programming with NumPy. Nature, 585(7825), 357-362.

[0246] Héberger, K. (2007). Quantitative structure-(chromatographic) retention relationships. Journal of Chromatography A, 1158(1-2): 273–305.

[0247] Hewitt, EF, Lukulay, P., and Galushko, S. (2006). Implementation of a rapid and automated high performance liquid chromatography method development strategy for drug candidates. Journal of Chromatography A, 1107(1-2):79-87.

[0248] Kaliszan, R. (1977). Correlation between retention index and connectivity index for alcohols and methyl esters with complex ring structures. Chromatographia, 10(9), 529–531.

[0249] Kaliszan, R. and Foks, H. (1977). Relationship between RM values ​​and connectivity index of pyrazine carboxysulfamide derivatives. Chromatographia, 10(7), 346-349.

[0250] Kensert, A., Bouwmeester, R., Efthymiadis, K., Broeck, P. V., Desmet, G., and Cabooter, D. (2021). Graph convolutional networks for improved prediction and interpretability of chromatographic retention data. Analytical Chemistry, 93(47), 15633-15641.

[0251] McKinney, W. (2010). Data structures for statistical computing in Python. In S. van der Walt and J. Millman (Eds.), Proceedings of the 9th Python for Science Conference (pp. 56-61).

[0252] Mitchell, M. (1996). Introduction to Genetic Algorithms (Volume 32). Polar Aprotic Solvents Selected by Hansen Solubility Parameters as Described by MIT Press. Mitchell, TM (1997). Machine Learning. McGraw-Hill.

[0253] Molnar, I. (2002). Computerized design of separation strategies for reversed-phase liquid chromatography: development of the DryLab software. Journal of Chromatography A, 965(1-2): 175-194.

[0254] Morgan, HL (1965). Generating unique machine descriptions for chemical structures - a technique developed by the Chemical Abstracts Service. Journal of Chemical Documentation, 5(2), 107–113.

[0255] Moriwaki, H., Tian, ​​Y.-S., Kawashita, N., and Takagi, T. (2018). Mordred: a molecular descriptor calculator. Journal of Cheminformatics, 10(1), 4.

[0256] Müller, A. and Guido, S. (2016). Introduction to Machine Learning with Python: A Guide for Data Scientists. O'Reilly Media. pandas development team, T. (2020). pandas-dev / pandas: Pandas (version 1.4.1). Zenodo.

[0257] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830.

[0258] Plotly Technologies, Inc. Collaborative Data Science. (2015). https: / / plot.ly

[0259] Probst, D. and Reymond, J.-L. (2018). Probabilistic molecular fingerprinting in a big data setting. Journal of Cheminformatics, 10(1), 66.

[0260] RDKit: Open Source Cheminformatics. (n.d.). https: / / www.rdkit.org /

[0261] Rogers, D. and Hahn, M. (2010). Extended connectivity fingerprints. Journal of Chemical Information and Modeling, 50(5), 742–754.

[0262] Rossum,Gv(2021). Python.https: / / docs.python.org / 3.9 /

[0263] Snyder, LR (2007). A new perspective on RPC column selectivity. Analytical Chemistry, 79(9), 3254-3262.

[0264] Snyder, LR, Kirkland, JJ and Dolan, JW (2009). Introduction to Modern Liquid Chromatography (Wiley, ed.; 3rd ed.).

[0265] Snyman, JA and Wilke, DN (2018). Preface to Practical Mathematical Optimization: Basic Optimization Theory and Gradient-Based Algorithms (pp. 3-40). Springer International Publishing.

[0266] Szucs, R., Brown, R., Brunelli, C., Heaton, J.C., and Hradski, J. (2021). Structure-driven prediction of chromatographic retention times: Application to pharmaceutical analysis. International Journal of Molecular Sciences, 22(8), 3848.

[0267] Teghem, J. (2010). Metaheuristics. From Design to Implementation, El-Ghazali Talbi. John Wiley & Sons Inc. (2009). XXI+593 pages, publication 978-0-470-27858-1. European Journal of Operational Research, 205(2), 486–487.

[0268] Todeschini, R. and Consonni, V. (2020a). Handbook of molecular descriptors. Methods and principles of medicinal chemistry.

[0269] Todeschini, R. and Consonni, V. (2020b). Molecular descriptors for cheminformatics. Methods and principles of medicinal chemistry.

[0270] Tome, T., N.、 Z. and Obreza, A. (2019). Development and optimization of liquid chromatography analytical methods using AQbD principles: Overview and recent advances. Organic Process Research & Development, 23(9), 1784–1802.

[0271] Trappe, W. (1940). Separation of biological lipids from natural mixtures by adsorption columns. II. Communication: Separation of phosphorus-free and nitrogen-free lipid fractions. Biochem. z, 305, 150–154.

[0272] Wildman, SA and Crippen, GM (1999). Predicting physicochemical parameters from atomic contributions. Journal of Chemical Information and Computer Sciences, 39(5), 868–873.

[0273] Wilson, N., Nelson, M., Dolan, J., Snyder, L., Wolcott, R., and Carr, P. (2002). Column selectivity in reversed-phase liquid chromatography I. General quantitative relationships. Journal of Chromatography A, 961(2), 171–193.

[0274] Implementation

[0275] Hereinafter, further specific embodiments of the present invention are listed. The specific embodiments described herein are provided by way of example, rather than by way of limitation. Any subheading herein is included only for convenience and should not be construed as limiting the disclosure in any way.

[0276] 1. In one embodiment, a computer-implemented method for predicting the HPLC retention time of one or more compounds is disclosed, the method comprising:

[0277] (a) obtaining values ​​of one or more structural and / or physicochemical properties of the compound and one or more HPLC method parameters; and

[0278] (b) using a machine learning model to predict the retention time of each of the compounds when subjected to HPLC using the one or more HPLC method parameters,

[0279] The machine learning model has been trained using a training data set, which includes molecular structural properties and / or physicochemical properties of one or more compounds, and one or more sets of chromatographic data for each compound, wherein the one or more sets of chromatographic data include retention times of corresponding compounds and associated HPLC method parameters.

[0280] 2. In one embodiment, a method of embodiment 1 is disclosed, wherein the machine learning model further predicts a metric indicative of peak width, and one or more sets of chromatographic data in the training data set further include a metric indicative of the width of the peak of the corresponding compound.

[0281] 3. In one embodiment, the method of embodiment 1 or 2 is disclosed, wherein the one or more compounds are individually selected from: a pharmaceutically active agent and a degradation product thereof.

[0282] 4. In one embodiment, a method according to any one of embodiments 1 to 3 is disclosed, wherein the HPLC method parameters are selected from the group consisting of: HPLC type, flow rate, temperature, pH, column length, column inner diameter, column particle size, column pore size, column steric interactions, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70), and one or more metrics defining the elution phase gradient.

[0283] 5. In one embodiment, a method of any one of embodiments 1 to 4 is disclosed, wherein the HPLC method parameters include one or more parameters selected from the group consisting of: column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70), and one or more metrics defining the elution phase gradient, and optionally one or more parameters selected from the group consisting of: HPLC type, flow rate, temperature, pH.

[0284] 6. In one embodiment, a method of any one of embodiments 1 to 5 is disclosed, wherein the HPLC method parameters include HPLC type, flow rate, temperature, pH, and one or more parameters selected from the group consisting of: column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70), and one or more metrics defining the elution phase gradient.

[0285] 7. In one embodiment, the method of any one of embodiments 1 to 6 is disclosed, wherein the one or more metrics defining the elution phase gradient include:

[0286] (i) the elution capacity of each of the plurality of mobile phases; and / or

[0287] (ii) defining a metric for a time-dependent change in the ratio of a plurality of mobile phases during elution, optionally wherein the one or more metrics defining a time-dependent change in the ratio of a plurality of mobile phases during elution comprises values ​​at one or more time points corresponding to a change in the rate of change of the ratio of the mobile phases, and / or values ​​of the rate of change of the ratio of the mobile phases at one or more time points during elution, and / or a metric describing a function representing a change in the ratio of the mobile phases in one or more (e.g., each) of a plurality of regions in the gradient; and / or

[0288] (iii) the elution capacity of the mobile phase at two or more (e.g., three) different time points, optionally wherein the mobile phase comprises a plurality of mobile phases, and obtaining the elution capacity at one time point as the sum of the elution capacities of each of the plurality of mobile phases weighted by the corresponding proportion of each mobile phase.

[0289] 8. In one embodiment, a method according to any one of embodiments 1 to 7 is disclosed, wherein the one or more physicochemical properties comprise one or more molecular descriptors and / or wherein the one or more molecular structural properties comprise a molecular fingerprint and / or one or more structural molecular descriptors.

[0290] 9. In one embodiment, a method according to embodiment 8 is disclosed, wherein the structural molecular descriptor and / or molecular fingerprint is a 2D molecular descriptor or molecular fingerprint.

[0291] 10. In one embodiment, a method of any one of embodiments 1 to 9 is disclosed, wherein one or more structural and / or physicochemical properties include molecular descriptors selected from the group consisting of SLogP, ATSC5v, ATSC8d, ATSC3Z, ABC, VSA_EState4, ATSC5dv and ATSC6i.

[0292] 11. In one embodiment, a method according to any one of embodiments 6 to 10 is disclosed, wherein the elution phase gradient is a fixed gradient.

[0293] 12. In one embodiment, a method according to any one of embodiments 1 to 11 is disclosed, wherein the HPLC method parameters used for predicting the retention time at step (b) are restricted to a corresponding predetermined range and / or selected from a corresponding predetermined value set, and / or wherein the training data set comprises a chromatographic data set obtained using the HPLC method parameters within a corresponding range and / or a corresponding value set, and the HPLC method parameters used for predicting the retention time at step (b) are within said range and / or selected from said value set.

[0294] 13. In one embodiment, a method according to any one of embodiments 6 to 10 is disclosed, wherein the elution phase gradient is a non-fixed gradient.

[0295] 14. In one embodiment, a method of any one of embodiments 1 to 13 is disclosed, wherein the machine learning model is an extreme gradient boosting model (e.g., an extreme gradient boosting tree), a gradient boosting model (e.g., a gradient boosting tree), a random forest model, a Lasso regression model, or a support vector machine, preferably an extreme gradient boosting model.

[0296] 15. In one embodiment, a method of any one of embodiments 1 to 14 is disclosed, wherein the machine learning model is trained to minimize the difference between the predicted retention time and the corresponding retention time in the training data.

[0297] 16. In one embodiment, a method of any one of embodiments 1 to 15 is disclosed, wherein the machine learning model has been trained using a dataset comprising data for at least 100, at least 200, at least 300, at least 400, at least 500, or at least 600 compounds.

[0298] 17. In one embodiment, the method of embodiment 16 is disclosed, wherein the data set comprises at least 2, 3, 4 or 5 sets of chromatographic data, the chromatographic data comprising retention times of corresponding compounds and associated HPLC method parameters.

[0299] 18. In one embodiment, a method according to any one of embodiments 1 to 17 is disclosed, wherein HPLC is reverse phase chromatography, normal phase chromatography, size exclusion chromatography, ion exchange chromatography or hydrophilic interaction liquid chromatography, optionally wherein HPLC is reverse phase chromatography or normal phase chromatography.

[0300] 19. In one embodiment, a method of any one of embodiments 1 to 18 is disclosed, wherein the machine learning model has been trained using a set of predictive features selected from a larger group through a feature selection process, and wherein the larger set of predictive features are each selected from: molecular structural properties or physicochemical properties (e.g., molecular descriptors or molecular fingerprints) and HPLC method parameters.

[0301] 20. In one embodiment, a computer-implemented method for identifying a set of HPLC method parameters suitable for separating two or more compounds in a composition comprising said compounds is disclosed, the method comprising:

[0302] (a) performing the method of any one of embodiments 1 to 19 for multiple sets of HPLC method parameters;

[0303] (b) using the results of step (a) to calculate one or more separation performance metrics; and

[0304] (c) identifying one or more sets of HPLC method parameters by applying one or more criteria to the separation performance metric.

[0305] 21. In one embodiment, a method of embodiment 20 is disclosed, wherein the separation performance metric includes one or more of the following: i) an average retention time difference between adjacent peaks corresponding to predicted retention times of two or more compounds; ii) gradient duration; and iii) a predicted metric indicative of peak width, optionally including both i) and ii).

[0306] 22. In one embodiment, a method of embodiment 20 or 21 is disclosed, wherein the separation performance metrics include: i) the average retention time difference between adjacent peaks corresponding to the predicted retention times of two or more compounds, wherein the criterion includes maximizing the average retention time difference; ii) gradient duration, wherein the criterion includes minimizing the gradient duration; and / or iii) predicted peak width, wherein the criterion includes minimizing the peak width.

[0307] 23. In one embodiment, a method according to any one of embodiments 20 to 22 is disclosed, wherein the one or more sets of HPLC method parameters in step (a) are generated by an optimization algorithm, optionally a genetic algorithm.

[0308] 24. In one embodiment, a method of embodiment 23 is disclosed, wherein step (a) comprises running the genetic algorithm until one or more stopping criteria apply, optionally wherein the stopping criteria are selected from: i) a predetermined number of generations has been reached; and ii) the difference between the separation performance metrics associated with one or more sets of HPLC method parameters of the current iteration and the separation performance metrics associated with one or more sets of HPLC method parameters of the previous iteration is below a threshold.

[0309] 25. In one embodiment, a method of embodiment 23 or 24 is disclosed, wherein the genetic algorithm is initialized with a randomly selected set of HPLC method parameters, wherein the genetic algorithm is initialized with a set of HPLC method parameters selected from a predetermined group or range of each HPLC method parameter, and / or wherein the genetic algorithm is initialized with a set of HPLC method parameters provided by a user.

[0310] 26. In one embodiment, a method according to any one of embodiments 1 to 25 is disclosed, wherein one or more sets of HPLC method parameters are selected (eg by a genetic algorithm) from respective predetermined sets and / or ranges.

[0311] 27. In one embodiment, a method of any one of embodiments 20 to 26 is disclosed, wherein the method comprises presenting to a user one or more sets (e.g., 5 sets) of HPLC method parameters that satisfy one or more criteria regarding the separation performance metric.

[0312] 28. In one embodiment, a computer program product is disclosed, comprising computer-readable instructions which, when executed by one or more processors, cause the one or more processors to perform the method of any one of embodiments 1 to 27.

[0313] 29. In one embodiment, a non-transitory computer-readable medium is disclosed, having computer-readable instructions stored thereon, which, when executed by one or more processors, cause the one or more processors to perform the method of any one of embodiments 1 to 27.

[0314] 30. In one embodiment, a system is disclosed, comprising:

[0315] at least one processor; and

[0316] At least one non-transitory computer-readable medium comprising instructions that, when executed by the at least one processor, cause the at least one processor to perform the method of any one of embodiments 1 to 27.

Claims

1. A computer-implemented method for predicting the HPLC retention time of one or more compounds, the method comprising: include: (a) obtaining values ​​of one or more structural and / or physicochemical properties of the compound and one or more HPLC method parameters; as well as (b) using a machine learning model to predict the retention time of each of the compounds when subjected to HPLC using the one or more HPLC method parameters, The machine learning model has been trained using a training data set, which includes molecular structural properties and / or physicochemical properties of one or more compounds, and one or more sets of chromatographic data for each compound, wherein the one or more sets of chromatographic data include retention times of corresponding compounds and associated HPLC method parameters.

2. The method of claim 1, wherein the machine learning model further predicts a metric indicative of peak width, and the one or more sets of chromatographic data in the training dataset further include a metric indicative of the width of the peak of the corresponding compound.

3. The method according to claim 1 or 2, wherein the HPLC method parameters are selected from the group consisting of: HPLC type, flow rate, temperature, pH, column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70) and one or more metrics defining the elution phase gradient, for example, wherein the HPLC method parameters include one or more of the parameters selected from the group consisting of: column length, column inner diameter, column particle size, column pore size, column steric interaction, column hydrogen bond acidity, column hydrogen bond basicity, column ion exchange capacity (c_28), column ion exchange capacity (c_70) and one or more metrics defining the elution phase gradient.

4. A method according to any one of claims 1-3, wherein one or more physicochemical properties include one or more molecular descriptors and / or wherein one or more molecular structural properties include molecular fingerprints and / or one or more structural molecular descriptors, optionally wherein the one or more structural and / or physicochemical properties include molecular descriptors selected from the group consisting of: SLogP, ATSC5v, ATSC8d, ATSC3Z, ABC, VSA_EState4, ATSC5dv and ATSC6i.

5. The method according to any one of claims 1 to 4, wherein the HPLC is reverse phase chromatography, normal phase chromatography, size exclusion chromatography, ion exchange chromatography or hydrophilic interaction liquid chromatography, optionally wherein the HPLC is reverse phase chromatography or normal phase chromatography.

6. A method according to any one of claims 1-5, wherein the machine learning model has been trained using a set of predictive features selected from a larger group by a feature selection process, and wherein the larger set of predictive features are each selected from: molecular structural properties or physicochemical properties (such as molecular descriptors or molecular fingerprints) and HPLC method parameters.

7. A computer-implemented method for identifying a set of HPLC method parameters suitable for separating two or more compounds in a composition comprising said compounds, said method include: (a) performing the method according to any one of claims 1 to 6 for multiple sets of HPLC method parameters; (b) using the results of step (a) to calculate one or more separation performance metrics; as well as (c) identifying one or more sets of HPLC method parameters by applying one or more criteria to the separation performance metric.

8. A method according to claim 7, wherein the separation performance metric includes one or more of the following: i) an average retention time difference between adjacent peaks corresponding to the predicted retention times of the two or more compounds; ii) gradient duration; and iii) a predicted metric indicative of peak width, optionally including both i) and ii).

9. The method according to claim 7 or 8, wherein the separation performance metric include: i) an average retention time difference between adjacent peaks corresponding to the predicted retention times of the two or more compounds, wherein the criterion includes maximizing the average retention time difference; ii) a gradient duration, wherein the criterion includes minimizing the gradient duration; and / or iii) a predicted peak width, wherein the criterion includes minimizing the peak width.

10. The method according to any one of claims 7 to 9, wherein the one or more sets of HPLC method parameters in step (a) are generated by an optimization algorithm, optionally a genetic algorithm.

11. The method of claim 10, wherein step (a) comprises running the genetic algorithm until one or more stopping criteria apply, optionally wherein the stopping criteria are selected from: i) a predetermined number of generations has been reached; and ii) a difference between a separation performance metric associated with one or more sets of HPLC method parameters of a current iteration and a separation performance metric associated with one or more sets of HPLC method parameters of a previous iteration is below a threshold.

12. The method according to any one of claims 7 to 11, wherein the method comprises presenting to a user one or more sets (e.g., 5 sets) of HPLC method parameters that satisfy one or more criteria regarding the separation performance metric.

13. A computer program product comprising computer readable instructions which, when executed by one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 12.

14. A non-transitory computer-readable medium having computer-readable instructions stored thereon, which, when executed by one or more processors, cause the one or more processors to perform the method according to any one of claims 1-12.

15. A system, wherein include: at least one processor; as well as At least one non-transitory computer-readable medium comprising instructions which, when executed by the at least one processor, cause the at least one processor to perform the method according to any one of claims 1-14.

Images