System and method for pre-configuring training data to enable a neural network to recognize signals in NMR measurements

Abstract

A system, method and computer program product for training a neural network for signal analysis in an NMR spectrum. The system obtains a plurality of calculated NMR raw spectra, each raw spectrum being associated with a different NMR active molecule having a known number of protons. The system widens the peaks of the received raw spectra by convolution of each raw spectrum with one or more line shaping functions to produce a widened spectrum for each raw spectrum. The widening of the line widths follows a statistical distribution over a plurality of current spectra by sampling widening values ​​for various raw spectra from a statistical distribution. The system calculates an integral function for each widened spectrum to count the number of protons associated with the peak of the corresponding widened spectrum. The system identifies signal intervals as intervals in the widened spectrum, wherein the integral function is increased approximately by a multiple of the value associated with a single proton, so that the total number of counted protons matches the known number of protons of the associated molecule.

Description

Technical Field

[0001] The present invention relates generally to signal detection in NMR spectroscopy, and more particularly to the generation of training data for a neural network to enable the neural network to identify signal intervals in spectra obtained from real NMR spectroscopy experiments. Background Art

[0002] Nuclear magnetic resonance (NMR) spectroscopy is a spectral technique for observing molecular properties at the atomic level. When a sample is placed in a magnetic field, a local magnetic field is induced around the nucleus. These induced fields can be observed by applying radio frequency (RF) pulses to the sample and monitoring the response as NMR signals. The NMR signals are picked up by sensitive RF receivers. The local induced fields are specific to the molecular environment of the nucleus, thus providing access to the details of the electronic structure of the molecule and its individual functional groups. For example, NMR spectroscopy is used to identify organic compounds, proteins, and other complex molecules. In addition to identification, NMR spectroscopy also provides detailed information about the structure, dynamics, reaction state, and chemical environment of the molecule. Common types of NMR are proton and carbon-13 NMR spectroscopy, but it is applicable to any type of sample containing nuclear spins with a nuclear magnetic moment greater than zero. In the document, molecules that produce NMR signals are referred to as NMR active molecules or NMR active substances.

[0003] After the sample is excited by a radio frequency (typically, 60MHz to 1000MHz) pulse, a nuclear magnetic resonance response is obtained, which is called free induction decay (FID). FID is a very weak signal and requires a sensitive RF receiver to pick up. Fourier transform can be applied to extract the frequency domain spectrum from the original time domain FID. The spectrum from the signal FID typically has a low signal-to-noise ratio, so multiple FIDs are usually required, and it is averaged to obtain a spectrum with a better signal-to-noise ratio. The decay time (typically measured in seconds) of the response to the excitation depends on the effectiveness of relaxation, which is faster for lighter nuclei and in solids, and slower for heavier nuclei and in solutions, but it can be very long in gases.

[0004] The evaluation of NMR spectra is typically supported by human NMR experts, based on their experience of how to interpret certain peaks in the NMR spectra obtained from real NMR experiments. The main task is thus to identify which signal intervals in the obtained spectrum are associated with peaks that characterize the molecules in the NMR sample undergoing evaluation. Such characteristic signal intervals can be difficult to identify because the signal of a typical NMR experiment includes signal contributions caused by sources different from the active molecule itself (such as, impurities, phase shifts, baseline effects, noise, etc.). The identification of signal intervals is therefore an error-prone procedure that depends heavily on the subjective personal experience of the supporting expert. Summary of the invention

[0005] Therefore there is a need for systems and methods that provide more robust and objective identification of signal intervals in NMR spectra obtained from real NMR experiments. In addition, in order to enable computer-implemented algorithms to automatically interpret NMR spectra, for example, to verify the structure of a molecule based on its 1H NMR spectrum, reliable identification of signal regions is required in a first step.

[0006] Embodiments of the invention as claimed in the independent claims in the form of computer-implemented methods, computer systems and computer program products use the claimed features to solve the technical problems. It should be noted that the NMR response is typically analyzed in the frequency domain. The signal interval as used herein refers to a frequency interval centered around the peak frequency of the corresponding signal peak in the NMR spectrum, and characterizes a portion of the spectrum associated with the nuclear magnetic resonance response of the NMR active molecule in response to the radio frequency pulse. In the literature, such signal intervals are sometimes also referred to as signal regions. It should be noted that in the NMR spectrum, the relative distance of the center frequency of the resonance peak from an arbitrarily selected reference frequency (standard frequency) is called a chemical shift. A typical reference frequency commonly used is the frequency of the resonance peak of tetramethylsilane (TMS). The reference frequency is thereby assigned a chemical shift value of "zero". The chemical shift is independent of the fundamental frequency of the spectrometer and is measured in units of "ppm". In the following description and figures, the frequency values ​​in the NMR spectrum are also referred to as ppm values ​​on the corresponding ppm scale.

[0007] In one embodiment, a computer-implemented method generates a realistic training data set for training a neural network for signal analysis in NMR spectra, advantageously in liquid state 1H NMR spectra. Thus, an initial hurdle for signal analysis is identifying signal intervals (or signal regions). The data set provided by the claimed method is optimized to train a neural network to automatically identify such signal intervals in real NMR spectra with a high level of accuracy without any human intervention. In other words, the training data set can be considered realistic in the sense that it enables training of the network with calculated spectra that have a high similarity to NMR spectra obtained from real experiments.

[0008] In the initial step, a plurality of calculated NMR raw spectra are obtained. Each raw spectrum is thus associated with different NMR active molecules (hereinafter also referred to as active molecules) with a known number of protons. The plurality of calculated NMR raw spectra can accurately include one raw spectrum for each active molecule. There are available commercial tools that calculate specific raw spectra by obtaining the spin system of the associated NMR active molecules and the subsequent numerical solution of the Schrodinger equation for a given pulse sequence. An example of the tool is the PERCH spectrum simulator provided by PERCH Solutions Ltd. of Kuopio, Finland. The PERCH spectrum simulator is based on quantum mechanical operations. Use the X approximation to fill in large spin systems.

[0009] A computer system configured to perform the computer-implemented method may include the tool as an integrated component, or it may be communicatively coupled to the tool and receive the plurality of calculated NMR raw spectra through a corresponding interface. A molecular database may serve as a storage for molecular descriptions. An example of the database is PubChem, which is a database of chemical molecules and their activities for biological identification. The system is maintained by the National Center for Biotechnology Information (NCBI), which is a component of the National Library of Medicine, which is a part of the National Institutes of Health (NIH). PubChem can be freely accessed through a network user interface. From such molecular descriptions, the number of protons in the corresponding molecule is known. From the storage, a subset of NMR active molecules serves as input for a raw spectrum generator module implemented, for example, by the previously mentioned PERCH tool. The subset may include molecules having only structures that do not exceed a predetermined degree of complexity. Advantageously, for a neural network focused on analyzing NMR spectra of active molecules for pharmaceutical applications, the molecular weight of the associated NMR active molecules is less than or equal to 500 Daltons. Typically, in pharmaceutical applications, such smaller molecules are in focus.

[0010] The output of the raw spectrum generator is a corresponding raw spectrum for each NMR-active molecule, having peaks that are expected to be the nuclear magnetic resonance responses of the corresponding active molecule (after having been subjected to the radio frequency pulse sequence). For example, all raw spectra may be calculated with the same line width parameters. The line width parameters may be selected so that the resulting line width corresponds to the lower limit of a line width range that includes line widths typically obtained in a plurality of real NMR measurements.

[0011] In the step of broadening, the peak of the received original spectrum is broadened by convolution of each original spectrum with one or more spectral line shaping functions to produce a broadened spectrum as the current spectrum for each original spectrum. The broadening step is sometimes also referred to as spectral line broadening. Examples of typical spectral line shaping functions are Lorentzian and / or Gaussian functions. For example, the original spectrum can be convolved with Lorentzian and / or Gaussian functions. The convolution of the Lorentzian function with the Gaussian function is also called the Voigt function. Different spectral line broadening values ​​can be used for each original spectrum. For example, the range of spectral line broadening values ​​used can be equivalent to a range from 0.3 Hz to 0.6 Hz. Advantageously, different spectral line broadening values ​​are applied to multiple original spectra according to statistical distributions (such as gamma or beta distributions). A favorable example uses a gamma distribution with a shape parameter of 4 and a scale parameter of 0.15 for Gaussian and Lorentzian line broadening parameters. Other statistical distributions may also be used by those skilled in the art. The broadening value is approximately the value of the line width added to the original spectrum by the line broadening convolution. The line broadening values ​​applied to the various original spectra are thus sampled from a statistical distribution. In other words, the effects of Gaussian and Lorentzian line broadening on multiple current spectra follow the statistical distribution used. For example, when using the above-mentioned gamma distribution, 99% of the current spectrum will be broadened by line broadening values ​​between 0.10Hz and 1.65Hz (and, respectively, 75% will be broadened by line broadening values ​​between 0.28Hz and 0.95Hz). In other words, the statistical variation of the broadening values ​​applied to the original spectrum results in the statistical variation of the line width in the broadened spectrum. It is not necessary to generate multiple broadened spectra with different line widths for the same active molecule. On the contrary, it is sufficient to have a broadened spectrum for each active molecule, wherein the different broadened spectra generally cover the full range of line widths observed in real NMR experiments. The line broadening step results in a current spectrum that already exhibits a significantly higher similarity to the NMR spectrum obtained from a real NMR experiment than the originally calculated raw spectrum. For a raw spectrum simulated with a line width of 0.6 Hz, the typical spectrum broadening is 0.9 Hz, and thus the typical line width is 1.5 Hz (when using the above-mentioned gamma distribution for Gaussian and Lorentzian line broadening parameters).

[0012] Advantageously, for further processing, the received plurality of calculated NMR raw spectra may be pre-processed such that the number of data points in each raw spectrum approximately corresponds to the number of data points of a comparable realistic NMR spectrum obtained from a realistic NMR experiment.

[0013] For each broadened spectrum, the signal interval is now determined based on the calculation of the indefinite integral function for the corresponding broadened spectrum to count the number of protons associated with the peak of the corresponding broadened spectrum. The computer system identifies the signal interval as such an interval in the broadened spectrum, wherein the indefinite integral is roughly increased by a multiple of the value associated with a single proton. Because the broadened spectrum at the stage is not disturbed by impurities, noise or any other disturbance, each peak or peak cluster in the broadened spectrum is obviously associated with the expected NMR response of the associated active molecule. For all broadened spectra that are calibrated so that the definite integral over the entire spectrum matches the known number of protons of the associated molecule, the indefinite integral function is calculated. Finally, the interval initially identified is adjusted to at least cover the predetermined threshold of the corresponding known number of protons. Experiments have shown that a threshold of at least 95% of the number of protons leads to good results. In other words, the signal interval for a specific peak cluster is adjusted so that the start frequency and the end frequency of the interval are set symmetrically with respect to the peak maximum frequency, so that the integral value between the adjusted start frequency and the end frequency still includes an integral region covering a region equal to or greater than the predetermined threshold of the corresponding number of protons. For example, if the integral increases by 1.8 times the value associated with a single proton, the system recognizes that the increase should actually correspond to 2.0 times the value associated with a single proton. Assuming a 95% threshold, the system will adjust the size of the signal interval to include a region corresponding to 1.9 times the value associated with a single proton.

[0014] The identified signal intervals in a particular broadened (current) spectrum represent labels associated with the corresponding signal intervals for the current spectrum. Once such labels have been determined for all current spectra, the current spectrum together with the associated labels is provided as a training data set to a neural network to train the neural network for automatic identification of signal intervals in real NMR spectra.

[0015] It should be noted that the labeling is performed fully automatically by a computer system executing the described computer-implemented method without any need for human interaction. In addition, the start and end frequencies of the labels are determined by the system with very high accuracy, since the claimed integration method applied to the broadened spectrum is not affected by any perturbations (as they are found in real NMR spectra) and therefore allows highly accurate localization of the start and end of the corresponding signal peak or peak clusters. Due to the statistical variation of the line widths in multiple current spectra, the neural network learns to generalize across all kinds of line widths that occur in real NMR experiments and at the same time learns to objectively identify the start and end of the signal interval with high accuracy.

[0016] In an optional embodiment, the training data set can be further improved in the sense that perturbations present in the NMR spectrum obtained from real NMR experiments are also added to the calculated broadened spectrum. The term "current spectrum" as used herein always refers to the current state of the calculated spectrum reached after the original raw spectrum has passed through one or more post-processing steps. Thus, the broadened spectrum is the current spectrum because the original raw spectrum has passed through the broadening pre-processing step. The current spectrum can then be modified by additional post-processing steps.

[0017] After each post-processing step, a new (or modified) current spectrum is obtained. For reasons of keeping the language simple, any new (or modified) current spectrum is also referred to herein as the "current spectrum". Those skilled in the art will recognize the state of the current spectrum based on the post-processing steps that have been applied to the original raw spectrum. After each post-processing step, multiple current spectra are obtained, which can serve as training data sets.

[0018] In general, a post-processing step that may be applied to the current spectra after the signal interval has been determined applies one or more statistical modifications to each current spectrum, wherein the statistical parameters of a particular statistical modification are statistically distributed over the plurality of current spectra. Such statistical modifications add perturbations to the current spectra that modify the calculated current spectra to become even more similar to real NMR spectra, thereby resulting in a higher degree of recognition of signal intervals by the neural network when trained with the perturbed current spectra.

[0019] In one embodiment, the disturbance applied is adding the effect of impurities to the current spectrum. That is, in addition to the active molecules to be analyzed, the real NMR sample always includes impurities. Such impurities also cause peaks in the real NMR spectrum obtained from the sample. Impurities can be statistically selected from a list of known impurities. Examples of such impurities are ethyl acetate or cyclohexane. Impurities can also be randomly constructed impurities. Typically, known impurities are associated with known patterns in the corresponding spectrum, while random impurities include regular and irregular patterns with a random number of spectral lines / peaks. The spectrum that will finally serve as the training spectrum for the neural network needs to focus on the relevant signals originating from the active molecules. It is not intended to train the network to identify impurities. Therefore, the concentration of the added impurities is kept relatively low so that the impurity ratio with the specific current spectrum as a whole is kept below, for example, an equivalent integer of 0.5 protons per impurity. As long as the integrated amplitude of the impurity-related peak remains below the equivalent value of preferably 0.5 protons, no signal area is generated because the threshold for interpreting the multi-value associated with a single proton is not reached. Parameters characterizing impurities include the number of impurities, the corresponding displacement and the amplitude. Such parameters vary statistically across a plurality of current spectra to which the impurity effect is added. That is, similar to the case of statistical variation for line broadening, each of the parameters varies according to a statistical distribution, resulting in current spectra with each spectrum reflecting a particular combination of parameters according to the statistical parameter distribution.

[0020] In one embodiment, the applied perturbation is adding a linear phase shift to each current spectrum, where the applied linear phase shift is statistically distributed over a plurality of current spectra. In the model for the phase shift, it is assumed to include only zero-order terms and first-order terms. The coefficients of the terms are again subject to a statistical distribution. For example, the coefficients may be selected to result in a typical phase shift of a maximum of four degrees over the full bandwidth. For each active molecule, a linear phase shift is calculated and applied according to a given statistical distribution. Again, due to the statistical variation of the coefficients over all current spectra, it is sufficient to have only one current spectrum for a given combination of coefficients in the training data set, since the neural network may have learned to generalize to the phase shift perturbation from the training data set.

[0021] In one embodiment, the applied disturbance is adding a baseline to each current spectrum. The baseline change for a particular current spectrum is calculated by a polynomial interpolation function over three to twelve sampling points (so-called knots), and the coordinates of the knots of the current spectrum are statistically uniformly distributed over a plurality of current spectra. For example, the interpolation function may be a cubic spline function. The number of sampling points may be determined by statistically selecting a number from a uniform distribution between a minimum of 3 and a maximum of 12. The x and y coordinates of the knots of the baseline change are taken from a uniform distribution of x and y coordinates over a predetermined range. Thus, the x coordinates are uniformly distributed over the entire spectral range of the corresponding current spectrum, and the y coordinates are uniformly distributed over a predetermined range including positive and negative values. The determined knots are then fitted with a cubic spline. Finally, the y axis is calibrated so that the mean absolute deviation of the baseline from the zero order follows a uniform distribution.

[0022] In one embodiment, the applied perturbation is adding noise to each current spectrum, where the noise amplitude follows a Gaussian distribution, and the standard deviation of the Gaussian distribution varies statistically across multiple current spectra.

[0023] Each perturbation transforms the current spectrum into one that becomes closer to the real spectrum, as it can be expected from real NMR experiments. It should be noted that the added perturbations can cover the broadened peaks of the broadened spectrum. However, the labels indicating the signal intervals are still associated with the modified current spectrum, and therefore, each current spectrum of the training data set provides information about the exact location of the signal interval in the perturbed spectrum.

[0024] In one embodiment, the provided current spectrum of the training data set is now actually used to train a neural network for signal analysis in NMR spectra to enable the neural network to identify signal intervals in real NMR spectra obtained from real NMR experiments. The neural network may be part of the same computer system used to generate the training data set, or it may be trained on a remote system communicatively coupled to the system. The neural network receives the generated current spectrum of the training data set, together with labels for the identified signal intervals. A supervised learning method is used to train the neural network with the received current spectrum of the training data set and with the associated label identifiers. Thus, the received current spectrum serves as input to the neural network, and the corresponding signal intervals as indicated in the associated labels are used as output.

[0025] In one embodiment, the trained neural network is then used for signal analysis in a real NMR spectrum. The trained neural network receives a real NMR spectrum obtained from a real NMR experiment as a test input. The trained neural network is then applied to the test input. As a result, the trained neural network provides as output one or more signal intervals as identified in the received test input.

[0026] It has been demonstrated that a neural network trained with a training data set according to any of the embodiments disclosed herein learns to generalize across a whole range of statistically varying parameters, even when only a single current spectrum per active molecule is included in the training data set. Even when all calculated raw spectra were simulated at a proton resonance frequency of 400 MHz, the trained neural network was able to identify signal intervals with a high level of accuracy in realistic NMR spectra obtained from an NMR spectrometer operating at a proton resonance frequency between 300 MHz and 500 MHz.

[0027] In one embodiment, a computer program product is provided for generating a realistic training data set for training a neural network for signal analysis in NMR spectroscopy. The computer program product has instructions that, when loaded into a memory of a computer system and when executed by at least one processor of the computer system, cause the computer system to perform method steps according to any one of the embodiments disclosed herein of a computer-implemented method for generating a training data set.

[0028] In one embodiment, a computer program product is provided for training a neural network for signal analysis in NMR spectra to enable the neural network to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment. The computer program product has instructions that, when loaded into a memory of a computer system and when executed by at least one processor of the computer system, cause the computer system to receive a training data set as generated according to the disclosure herein, and train the neural network using a supervised learning method with a received current spectrum of the training data set and an associated label identifier, wherein the training input to the neural network is the current spectrum of the training data set, and the output is the corresponding signal interval.

[0029] In one embodiment, a computer program product is provided for signal analysis in NMR spectra to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment. The computer program product has instructions that, when loaded into a memory of a computer system and when executed by at least one processor of the computer system, cause the computer system to receive a real NMR spectrum obtained from a real NMR experiment as a test input for a neural network as trained according to the training method disclosed herein, and to apply the trained neural network to the test input, and to provide one or more signal intervals as identified by the trained neural network as an output of the trained neural network.

[0030] In one embodiment, a computer system is provided for generating a realistic training data set for training a neural network for signal analysis in NMR spectroscopy. The system has modules adapted to perform corresponding steps when executing the above computer program product for generating a realistic training data set for training the neural network.

[0031] In one embodiment, a computer system is provided for training a neural network for signal analysis in NMR spectra to enable the neural network to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment. The system has modules adapted to perform corresponding steps when executing the above computer program product for training the neural network.

[0032] In one embodiment, a computer system is provided for signal analysis in NMR spectra to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment. The system has modules adapted to perform corresponding steps when executing the above computer program product for signal analysis in NMR spectra.

[0033] Further aspects of the invention will be realized and attained by means of the elements and combinations particularly described in the appended claims.It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention as described. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A block diagram of a computer system for generating a realistic training data set for training a neural network for signal analysis in NMR spectroscopy, the computer system being coupled to a spectrum generator tool and a neural network implementation, according to an embodiment;

[0035] Figure 2 is a simplified flow chart of a computer-implemented method for generating a realistic training data set with optional training and analysis steps according to various embodiments;

[0036] Figure 3 Examples of NMR-active molecules that served as the basis for raw spectrum generation are shown;

[0037] Figure 4 is an example of a calculated raw spectrum for an active molecule;

[0038] Figure 5 is an example of a widened spectrum according to an embodiment;

[0039] Fig. 6A illustrates an indefinite integral function applied to a widened spectrum according to an embodiment;

[0040] Figure 6B A schematic diagram of the indefinite integral function for an active molecule having 5 protons is shown;

[0041] Figure 7 illustrates a broadened spectrum with annotations of signal spacing labels according to an embodiment;

[0042] Figure 8 FIGURES illustrate the effect of impurities on the current spectrum according to an embodiment;

[0043] Fig. 9 FIGURES illustrate the effect of phase shift on the current spectrum according to an embodiment;

[0044] Fig.10 FIGURE 1 illustrates the effect of a baseline change on a current spectrum according to an embodiment;

[0045] Fig.11 FIGURE 1 illustrates the effect of noise on the current spectrum according to an embodiment;

[0046] Fig.12 illustrates examples of gamma and beta distributions that may be used by various embodiments;

[0047] and

[0048] Fig.13 is a diagram showing examples of a general computer device and a general mobile computer device that may be used with the techniques described herein. DETAILED DESCRIPTION

[0049] Figure 1 A block diagram of a computer system 100 for generating a real training data set 141 for training a neural network 230 for signal analysis in NMR spectroscopy is illustrated in accordance with an embodiment. Figure 1 The system 100 is used to generate a Figure 2 The following description is described in the context of a simplified flowchart of a computer-implemented spectrum method 1000 for a real training data set illustrated in FIG. Figure 1 and Figure 2 The system 100 is thus configured to perform the method 1000 when a corresponding computer program is loaded into a memory of the system and said program is executed with a processing means of the system.

[0050] exist Figure 1 In an example embodiment of , input to the computer system 100 is provided by a spectrum generator 101. The generator 101 may be implemented as an integral module of the computer system 100, or it may be provided as a separate tool provided on a remote computer device communicatively coupled to the computer system 100. In both cases, the data generated by the generator 101 is provided to the computer system 100 using standard communication means. The training data set 141 is provided to the neural network 230 which is part of the NMR spectrum analyzer 102. Also, the spectrum analyzer 102 may be an integral module of the computer system 100, or it may operate on a remote computer device communicatively coupled to the computer system 100.

[0051] The generator 101 is communicatively coupled to a molecular database (DB) 210. The molecular DB may be part of the generator itself, or more likely, it may be a remote database accessible by the generator 101 through a corresponding interface. The molecular DB provides a repository of molecular descriptions. An example of the database is the PubChem database mentioned above. From such molecular descriptions, the number of protons in the corresponding molecule is known. From the repository, a subset of NMR active molecules 211 serves as input to a raw spectrum generator module 212 implemented, for example, by the previously mentioned PERCH spectrum simulator. The subset may include molecules having only structures that do not exceed a predetermined complexity. Advantageously, the molecular weight of the associated NMR active molecules is less than or equal to 500 Daltons. Figure 3 An example of structural formula 30 of an NMR active molecule having 28 hydrogen atoms is shown (C 21 H 28 N6O4S2). The structural formula 30 is included in the PubChem database with the identifier CID 16587030. In order to train a reliable neural network, a large number of NMR active molecules that meet the above conditions are selected. For example, the number of selected molecules can be in the range of 100.000 to 10.000.000 NMR active molecules.

[0052] The raw spectrum generator module 212 can calculate a specific raw spectrum by estimating the spin system of the associated NMR active molecule and the subsequent numerical solution of the Schrödinger equation for a given pulse sequence. This allows, for example, simulation of the cyclosporine A spectrum (89 protons, 56 particles), and by the X-approximation, also simulation of much larger spin networks like testosterone (28 fully coupled protons). Figure 4 An example of a calculated raw spectrum 40 in the range from 0 ppm to 15 ppm is shown, wherein the calculated raw spectrum 40 is generated at a consumption of a spectral line width of 0.6 Hz. Each raw spectrum is associated with a different active molecule having a known number of protons #P. The number of protons #P is known from the corresponding structural formula. For example, as is apparent to one skilled in the art, structural formula 30 (see Figure 3 ) includes 28 protons. Figures 4 to 11 The current spectrum shown in the example diagram is actually for Figure 3 Structural calculations shown in .

[0053] The calculated NMR raw spectra 213, together with the corresponding number #P of known protons, are now obtained 1100 by the computer system 100, serving as input to the line broadening module 110 of the computer system 100. In one implementation, the line broadening module (or a separate pre-processing module) may perform an initial pre-processing 1150 of the obtained raw spectra so that the number of data points in each raw spectrum roughly corresponds to the number of data points of a fairly realistic NMR spectrum obtained from existing NMR experiments, e.g., a power of two, such as 512, 1024, 32768. In order to achieve a match between the number of data points of the test input spectrum and the number of data points of the current spectrum of the training data set, the test input may therefore be interpolated before being processed by the neural network.

[0054] The line broadening module 110 generates a broadened spectrum 111 as a current spectrum for each original spectrum by broadening 1200 the peaks of the original spectrum (by convolution of each original spectrum 213 with one or more line shaping functions). The one or more line shaping functions may be Lorentz or Gaussian functions, or a combination thereof (which is equivalent to line broadening by Voigt functions). The broadening of the line width follows a statistical distribution over a plurality of original / current spectra. In other words, different line broadening parameters are used to broaden different original spectra, wherein a specific line broadening parameter for the broadening of a specific original spectrum is selected from a predetermined parameter range based on a statistical distribution having a supported positive lower limit, such as a gamma distribution, a beta distribution, a log-normal distribution, or another suitable distribution having a supported non-negative lower limit. Figure 5 The original spectrum 40 is shown (see Figure 4 ) is convolved by a Lorentzian function (with a width of 0.6 Hz) and a Gaussian function (with a width of 0.3 Hz) resulting in a broadened spectrum 50-2 (see Figure 5 ). For example, in the case of an original spectrum with a line width of 0.6 Hz, a Lorentzian function with a line width of 0.4 Hz and a Gaussian function with a line width of 0.3 Hz can be used for line broadening. The line broadening parameters applied in the example are therefore 0.4 Hz and 0.3 Hz, respectively.

[0055] In its entirety, the original spectrum 50-1 appears identical to the widened spectrum 50-2, since the difference between the two spectra is too small to be discerned by the human eye in a black and white representation without applying a zoom. Figure 5In the upper part of the diagram, a zoomed portion of the spectrum is illustrated in the range from 3.15 ppm to 3.35 ppm, wherein the spectrum has been shifted apart in the vertical axis for better visibility. In the zoomed view, the effect of the broadening step becomes clearly visible. Especially when looking at the peaks with low amplitudes 50-1a, 50-1b, the original spectrum 50-1 has significantly sharper peaks, which almost disappear in the broadened spectrum 50-2. Due to the broadening of the spectral lines, the peak-to-valley ratio is reduced. The broadened spectrum 111 is already more comparable to the real NMR spectrum than the originally calculated original spectrum 213.

[0056] Of course, since the number of protons #P is known for the original spectrum, it is also known for the derived broadened spectrum. The broadened spectra 111 together with their corresponding number of protons #P serve as input for a signal interval (SI) identifier module 120. The SI identifier 120 calculates 1310 its integral function for each broadened spectrum 111 to count the number of protons associated with the peak of the corresponding broadened spectrum. Fig. 6A The integral function 60 calculated for the broadened spectrum 50-2 is illustrated. The SI identifier 120 now analyzes the integral function 60 and detects intervals in the broadened spectrum, wherein the integral function 60 increases approximately by a multiple of the value associated with a single proton. The dashed lines (e.g., 60-1a, 60-1b, 60-2a, 60-2b) on the ppm axis represent the counts of the number of protons associated with the corresponding peaks (e.g., 60-1, 60-2). In other words, the peak integral of peak 60-1 corresponds to the equivalent of two protons. The peak integral of peak 60-2 also corresponds to the equivalent of two protons. The integral function 60 is calibrated over the entire spectrum to result in a total number of protons that match the count of the known number of protons #P of the associated molecule. The signal intervals are now identified 1320 as such intervals in the broadened spectrum 111, wherein the indefinite integral increases approximately by a multiple of the value associated with a single proton. Thus, the identified intervals 211 are adjusted to at least cover a predetermined threshold of the corresponding peak intervals. For example, the predetermined threshold may be defined such that 95% of the corresponding peak integrals are covered. Higher coverage values ​​may also be used.

[0057] Figure 6BThe schematic diagram of the integral function 65 calculated for the widened spectrum of the active molecule with 5 protons is illustrated. The y-axis shows the value of the indefinite integral in the equivalent (#P) ​​of the proton. In the example embodiment, a three-step calculation is used to determine the signal interval. In the first step, the widened spectrum is convoluted with a rectangular window function with a width of 7Hz. The interval that exceeds the predetermined threshold value in the convolved (further widened) spectrum 66 defines the intermediate signal interval. Such intermediate signal intervals are illustrated for the following: two peaks (with intermediate intervals A1, A2) on the left side of the spectrum associated with 2 protons, the next two peaks (intermediate intervals B), the four peaks (intermediate intervals C1, C2) and a peak (intermediate intervals D) on the right side that follow each other associated with a proton. It should be noted that the multiple peaks associated with a single proton typically characterize the liquid state 1HNMR spectrum, thereby making the identification of the signal interval a difficult and error-prone task in the NMR spectrum analysis background technology solution. In the second step, in cases where the indefinite integral does not approach a multiple of a single proton anywhere between the intermediate signal intervals, adjacent intervals are merged. Figure 6B In the example of , this is the case for the two intermediate intervals A1 and A2 (merged into A) and the two intermediate signal intervals C1 and C2 (merged into C) on the left. In the third step, these modified intermediate signal intervals are adjusted to cover, for example, 95% of the integration area for the corresponding peak. The embodiment allows reliable identification of a single interval, even in the case where the signal peak does not return to zero due to overlap with adjacent peaks. It should be noted that the calculated signal peak never returns to the zero line, but actually has an infinite tail. As a result, the determined integral value does not show the full multiples of the proton equivalent, but only approaches those multiples. For example, for interval D, the calculated integral value in the corresponding initially determined intermediate interval segment may only total 1.8. When the value corresponds to approximately two protons, the merged interval is then adjusted to cover, for example, 95% of the area corresponding to two proton equivalents (i.e., 1.9 proton equivalents).

[0058] It should be noted that the signal intervals 121 are determined without any further perturbation of the broadened spectrum.The identified signal intervals 121 associated with the respective broadened spectrum 111 are indicated by corresponding labels. Figure 7 An example of a broadened spectrum 70 and its associated labels 70-1 to 70-n is illustrated. Thus, each label comprises information about the respective starting ppm value and ending ppm value of the signal region associated with the corresponding peak.

[0059] The broadened spectrum 111 , which is the current spectrum at the point in time, is then provided 1500 to the neural network 230 as a training data set 141 together with the associated labels for the identified signal intervals 121 .

[0060] In one embodiment, the computer system 100 further includes a statistics modifier module 130 that provides one or more statistics modifier functions. Each identifier function involves the modification of a disturbance (e.g., noise, baseline, etc.) that is typically observed as a corresponding spectrum in a real NMR spectrum. Once the signal interval has been determined for the broadened spectrum, the resulting current spectrum can be transformed into an even more realistic spectrum by applying 1400 a calculated disturbance corresponding to such disturbance that occurs in a real NMR experiment. To this end, the statistics modifier 130 applies 1400 a statistical modification to each current spectrum. Thus, statistical parameters of a specific statistical modification related to a corresponding characteristic of the real NMR spectrum are statistically distributed over a plurality of current spectra. Typical distributions that can be used for statistical distributions for line broadening are gamma distributions and beta distributions. Possible choices for disturbances of line broadening parameters are shown in Fig.12 For other types of modifications, other statistical distributions may be advantageously used, such as, for example, a uniform distribution in the case of baseline or phase shift modifications. In other words, a particular current spectrum is modified by adding the calculated effect of the corresponding perturbation to the current spectrum. In order to generate a training data set from which the neural network 230 can be generalized, the modifications to a particular parameter applied to the various current spectra are varied across all spectra according to a predetermined statistical distribution suitable for reflecting the full range of parameter values ​​that may be expected in real NMR experiments obtained by a variety of different NMR spectrometers.

[0061] The calculated perturbations may include adding 1410 the effects of impurities 131 to each current spectrum, applying 1420 a linear phase shift 132 to each current spectrum, adding 1430 a baseline change to each current spectrum, and adding 1440 noise 134 to each current spectrum.

[0062] Figure 8 An example of the effect of impurities on the current spectrum according to an embodiment is shown. The current spectrum 80 is the spectrum obtained after adding the calculated effect of impurities. The effect is Figure 8The resolution used in the present invention is substantially invisible to the human eye. For this reason, the calculated effect diagram is shown in the enlarged difference spectrum 80-1 under the modified current spectrum 80. The difference spectrum 80-1 thus shows the difference between the original widened spectrum and the current spectrum 80 generated by the addition of impurity disturbances. The impurities can be statistically selected from a list of known impurities and from unknown / constructed impurities. In the example, in addition to 12 random multiplets, tetramethylsilane (TMS) with the structural formula Si(CH3)4 is used as an impurity. In the case of known impurities, the NMR response of such impurities is well known and can be retrieved from the corresponding database. In the case of random impurities, a small perturbation signal line cluster is inserted into the current spectrum of the position (frequency) in the spectrum that can be randomly selected. In order to calculate the effect of impurities to the current spectrum, the corresponding displacement and amplitude of the perturbation signal are statistically varied on multiple current spectra. For each specific current spectrum 111', the impurity ratio to the whole is maintained at an equivalent integer below 0.5 protons. In the described manner, the addition of impurity effects is avoided to cause new signal intervals.

[0063] Fig. 9 The effect of the phase shift disturbance on the current spectrum 90 according to the embodiment is shown. The current spectrum 90 is the spectrum obtained after adding the effect of the phase shift. Fig. 9 At the resolution used in , the effect is barely visible to the human eye. For this reason, the effect of the phase shift is illustrated in an amplified difference spectrum 90-1 below the amplified current spectrum 90, which illustrates the difference between the current spectrum before applying the phase shift perturbation and the current spectrum 90 resulting from adding the phase shift perturbation.

[0064] In the applied phase shift model, it is assumed to include only zero-order and first-order terms. The coefficients of the terms are again subject to a statistical distribution that typically results in a maximum phase shift of four degrees over the entire spectral range. The real part of the phased NMR signal can be described by:

[0065] S(δ)=cos(α)*A(I(δ))-sin(α)*D(I(δ))

[0066] where A is the absorption spectrum (ie the real part of the spectrum before applying the phase shift) and D is the dispersion spectrum (ie the imaginary part, respectively). The phase α depends linearly on the chemical shift δ: α=a+b*δ, where a and b are the phase shift parameters sampled from the uniform distribution.

[0067] For each active molecule, and therefore for each associated current spectrum, a linear phase shift is calculated and applied according to a given statistical distribution (e.g., a uniform distribution). Due to the statistical variation of the coefficients over all current spectra, it is sufficient to have only one current spectrum for a given combination of coefficients in the training dataset, since the neural network can have learned to generalize to phase shift perturbations from the training dataset. This avoids an exponential growth in the number of training spectra. The statistical variance arises from the distribution of the phase shifts over multiple current spectra.

[0068] For example, the statistical variation may be based on a uniform distribution with a zero-order phase shift of [-2, +2] degrees / ppm and a first-order phase shift of [-1 / 5, +1 / 5] degrees / ppm. That is, in the extreme case, when the zero-order phase shift is equal to +2 degrees / ppm and the second-order phase shift is +1 / 5 degrees / ppm, the phase is 2 degrees at 5ppm (the chosen pivot point) and 2+(15-5)*1 / 5=4 degrees at 15ppm. In other words, in the example, the maximum possible phase shift that can occur in the spectrum is 4 degrees.

[0069] Fig.10 The effect of a baseline change on a current spectrum 10 according to an embodiment is shown. The current spectrum 10 is the spectrum obtained after adding 1430 the effect of the baseline change. Fig.10 At the resolution used in , the effect is barely visible to the human eye. For this reason, the calculated effect is illustrated in an amplified baseline spectrum 10-1 under the modified current spectrum 10. The baseline spectrum corresponds to the difference between the current spectrum before the amplification is applied and the modified current spectrum 10 after the amplification. In the example, the baseline change of the current spectrum 10 is calculated by a spline function with four sampling points. The coordinates of the sampling points of the spline functions of all current spectra are statistically uniformly distributed over a plurality of current spectra. Thus, the x-coordinate is distributed over the entire spectral range, and the y-coordinate is distributed over a predetermined range with positive and negative values. The cubic spline as an example of a piecewise polynomial function is then used for interpolation according to the distribution by the sampling points assigned to the corresponding current spectra.

[0070] Fig.11 The effect of noise on the current spectrum 11 according to an embodiment is illustrated. In the example, the noise spectrum 11-1 is added 1440 to the current spectrum and results in a modified current spectrum 11. The noise amplitude follows a Gaussian distribution across all current spectra, and the standard deviation of the Gaussian distribution varies statistically across multiple current spectra.

[0071] It should be noted that the accuracy of the neural network 230 for correct identification of signal intervals in realistic NMR spectra improves significantly when at least noise perturbation is applied to the broadened spectrum. Further significant improvement can be achieved by adding baseline perturbation.

[0072] Fig.12 10 ...

[0073] Return to Figure 1 and Figure 2 , the computer-implemented method for generating a training data set 141 may be continued with yet another computer-implemented method 1600 for training a neural network 230 for signal analysis in an NMR spectrum. The trained neural network 230 is thereby enabled to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment. To this end, a spectrum analyzer 102 including a neural network 230 to be trained receives 1610 a training data set 141 having labels 121 of signal intervals for identification. The neural network is then trained 1620 using a supervised learning method with the received training data set and the associated label identifiers 121. Thus, the training input to the neural network 230 is the received current spectrum 111' of the training data set 141, and the output is the corresponding signal interval. The output may be compared with the received labels 121, and the weights in the neural network adjusted accordingly.

[0074] Once the neural network has been trained 1600, it may be used by the spectrum analyzer 102 to perform yet another computer-implemented method 1700 for signal analysis in NMR spectroscopy. Initially, the spectrum analyzer 102 receives 1710 a real NMR spectrum 202 obtained from a real NMR experiment as a test input for a trained neural network 230. The spectrum analyzer 102 applies 1720 the trained neural network 230 to the test input 202. The output of the trained neural network 230 includes one or more signal intervals 203 as identified by the trained neural network 230. The output is then provided 1730 by the signal analyzer 102 to a user or to another analysis module for further evaluation.

[0075] Fig.13 900 and a general mobile computer device 950 that can be used with the techniques described herein. In some embodiments, the computing device 900 can be used with the system 100 (see Figure 1). Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, and other similar computing devices. In the context of the present disclosure, computing device 950 may provide I / O components for a user to interact with computing device 950. In other embodiments, the entire system 100 may be implemented on mobile device 950. The components, their connections and relationships, and their functions shown here are intended to be exemplary only, and are not intended to limit implementations of the inventions described and / or claimed in this document.

[0076] The computing device 900 includes a processor 902, a memory 904, a storage device 906, a high-speed interface 908 connected to the memory 904 and a high-speed expansion port 910, and a low-speed interface 912 connected to a low-speed bus 914 and the storage device 906. Each of the components 902, 904, 906, 908, 910, and 912 is interconnected using various buses and can be mounted on a common motherboard or otherwise appropriately mounted. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input / output device (such as a display 916 coupled to the high-speed interface 908). In other implementations, multiple processors and / or multiple buses can be used appropriately, along with multiple memories and memory types. In addition, multiple computing devices 900 can be connected, with each device providing a portion of the necessary operations (for example, as a server group, a blade server group, or a multi-processor system).

[0077] The memory 904 stores information within the computing device 900. In one implementation, the memory 904 is one or more volatile memory units. In another implementation, the memory 904 is one or more non-volatile memory units. The memory 904 may also be another form of computer-readable media, such as a magnetic disk or an optical disk.

[0078] The storage device 906 can provide mass storage for the computing device 900. In one implementation, the storage device 906 can be or contain a computer-readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid-state memory device, or an array device, including a device in a storage area network or other configuration. A computer program product can be tangibly embodied in an information carrier. The computer program product can also contain instructions that, when executed, perform one or more methods (such as the methods described above). The information carrier is a computer or machine-readable medium, such as the memory 904, the storage device 906, or a memory on the processor 902.

[0079] The high-speed controller 908 manages bandwidth-intensive operations for the computing device 900, while the low-speed controller 912 manages lower bandwidth-intensive operations. The allocation of functions is exemplary only. In one implementation, the high-speed controller 908 is coupled to the memory 904, the display 916 (e.g., through a graphics processor or accelerator), and to the high-speed expansion port 910 that can accept various expansion cards (not shown). In the implementation, the low-speed controller 912 is coupled to the storage device 906 and the low-speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet), can be coupled to one or more input / output devices, such as a keyboard, a pointing device, a scanner, or a networking device (such as a switch or router), for example, through a network adapter.

[0080] The computing device 900 can be implemented in many different forms, as shown in the figure. For example, it can be implemented as a standard server 920, or multiple times in a group of such servers. It can also be implemented as part of a rack server system 924. In addition, it can be implemented in a personal computer such as a notebook computer 922. Alternatively, components from the computing device 900 can be combined with other components in a mobile device (not shown) such as device 950. Each of such devices can contain one or more of the computing devices 900, 950, and the entire system can be composed of multiple computing devices 900, 950 in communication with each other.

[0081] The computing device 950 includes a processor 952, a memory 964, an input / output device such as a display 954, a communication interface 966, and a transceiver 968, in addition to other components. The device 950 may also be provided with a storage device such as a micro hard drive or other device to provide additional storage. Each of the components 950, 952, 964, 954, 966, and 968 are interconnected using various buses, and several of the components may be mounted on a common motherboard or otherwise appropriately mounted.

[0082] The processor 952 can execute instructions within the computing device 950, including instructions stored in the memory 964. The processor can be implemented as a chipset including a single or multiple analog and digital processors. The processor can provide, for example, coordination of other components of the device 950, such as control of a user interface, applications run by the device 950, and wireless communications performed by the device 950.

[0083] The processor 952 can communicate with the user through the control interface 958 and the display interface 956 coupled to the display 954. The display 954 can be, for example, a TFT LCD (thin film transistor liquid crystal display) or an OLED (organic light emitting diode) display, or other suitable display technology. The display interface 956 may include suitable circuits for driving the display 954 to present graphics and other information to the user. The control interface 958 can receive commands from the user and convert them for submission to the processor 952. In addition, an external interface 962 in communication with the processor 952 can be provided to enable near-area communication of the device 950 with other devices. The external interface 962 can provide wired communication in some implementations, or wireless communication in other implementations, for example, and multiple interfaces can also be used.

[0084] The memory 964 stores information within the computing device 950. The memory 964 may be implemented as one or more of one or more computer-readable media, one or more volatile memory units, or one or more non-volatile memory units. An expansion memory 984 may also be provided and connected to the device 950 via an expansion interface 982, which may include, for example, a SIMM (single in-line memory module) card interface. The expansion memory 984 may provide additional storage space for the device 950, or may also store applications or other information for the device 950. Specifically, the expansion memory 984 may include instructions to perform or supplement the processes described above, and may also include security information. Thus, for example, the expansion memory 984 may act as a security module for the device 950, and may be programmed with instructions that permit secure use of the device 950. In addition, a security application may be provided via a SIMM card, along with additional information, such as placing identification information on the SIMM card in an invulnerable manner.

[0085] The memory may include, for example, flash memory and / or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product contains instructions that, when executed, perform one or more methods (such as those described above). The information carrier is a computer or machine readable medium, such as memory 964, expansion memory 984, or memory on processor 952, which may be received, for example, via transceiver 968 or external interface 962.

[0086] The device 950 may communicate wirelessly via a wireless interface 966, which may include digital signal processing circuitry, if necessary. The communication interface 966 may provide for communication in various modes or protocols, such as GSM voice calls, SMS, EMS or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. The communication may occur, for example, via a radio frequency transceiver 968. In addition, short range communication may occur, such as using Bluetooth, WiFi, or other such transceivers (not shown). In addition, a GPS (global positioning system) receiver module 980 may provide additional navigation and location related wireless data to the device 950, which may be used appropriately by applications running on the device 950.

[0087] Device 950 may also communicate audio using audio codec 960, which may receive spoken information from a user and convert it into usable digital information. Audio codec 960 may likewise generate audible sounds for the user, such as through a speaker, for example, in a microphone of device 950. The sounds may include sounds from voice phone calls, may include recorded sounds (e.g., voice messages, music files, etc.), and may also include sounds generated by applications operating on device 950.

[0088] The computing device 950 may be implemented in many different forms, as shown in the figure. For example, it may be implemented as a cellular phone 980. It may also be implemented as part of a smart phone 982, a personal digital assistant, or other similar mobile device.

[0089] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system that includes at least one programmable processor, which can be special purpose or general purpose, coupled to receive data and instructions from a storage system, at least one input device, and at least one output device, and to transmit data and instructions to the storage system, at least one input device, and at least one output device.

[0090] These computer programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and may be implemented in a high-level procedural and / or object-oriented programming language and / or in assembly / machine language. As used herein, the terms "machine-readable medium," "computer-readable medium," refer to any computer program product, apparatus, and / or device (e.g., a disk, an optical disk, a memory, a programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0091] To provide interaction with a user, the systems and techniques described herein may be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other kinds of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, verbal, or tactile input.

[0092] The systems and techniques described herein may be implemented in a computing device that includes a back-end component (e.g., as a data server) or includes a middleware component (e.g., an application server) or includes a front-end component (e.g., a client computer with a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described herein), or any combination of such back-end, middleware, or front-end components. The components of the system may be interconnected by any form of digital data communication or digital data communication media (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.

[0093] Computing devices may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

Claims

1. A computer-implemented method (1000) for generating a training data set (141) for training a neural network (230) for signal analysis in NMR spectroscopy, comprising: obtaining (1100) a plurality of calculated NMR raw spectra (213), each raw spectrum being associated with a different NMR active molecule (211) having a known number of protons (#P); widening (1200) the line widths of the original spectra (213) by convolution of the original spectra with one or more line shaping functions to produce widened spectra (111) as current spectra for each original spectrum, wherein the widening of the line widths follows a statistical distribution over the plurality of current spectra by sampling corresponding widening values ​​for the various original spectra from the statistical distribution; calculating (1310) for each broadened spectrum (111) an integral function thereof to count the number of protons associated with a peak of the corresponding broadened spectrum; identifying (1320) signal intervals as intervals in the broadened spectrum (111), wherein the integral function is increased by approximately a multiple of the value associated with a single proton such that the total number of counted protons matches the known number of protons (#P) ​​of the associated molecule, wherein the identified intervals (211) are adjusted to encompass at least a predetermined threshold of corresponding peak intervals; wherein counting the total number of protons is accomplished by calculating an indefinite integral function for the entire broadened spectrum scaled such that the definite integral over the entire spectrum matches the known number of protons of the associated molecule; The current spectrum (111, 111') with associated labels (121) for the identified signal intervals is provided (1500) to the neural network as the training data set (141).

2. The method of claim 1, wherein the associated NMR-active molecule has a molecular weight less than or equal to 500 Daltons.

3. The method according to claim 1, further comprising: After the plurality of calculated NMR raw spectra have been received, the received raw spectra are preprocessed (1150) so that the number of data points in each raw spectrum approximately corresponds to the number of data points of a comparable real NMR spectrum obtained from a real NMR experiment. 4 . The method of claim 1 , wherein the one or more line shaping functions are selected from the group consisting of: a Lorentz function, a Gaussian function, and a Voigt function. 5 . The method of claim 1 , wherein the statistical distribution used for broadening is selected from the group consisting of: a gamma distribution, a beta distribution, a log-normal distribution. The method according to claim 5 , wherein the line broadening value is equivalent to a range from 0.3 Hz to 0.6 Hz.

7. The method according to claim 1, further comprising: After the signal intervals have been identified, one or more statistical modifications are applied (1400) to each current spectrum, wherein a statistical parameter of a particular statistical modification is statistically distributed across the plurality of current spectra.

8. The method according to claim 1, further comprising: After the signal interval has been identified, the effect of the impurities is added (1410) to the current spectrum (111), wherein the number of impurities, corresponding displacements and amplitudes are statistically varied across the multiple current spectra so that the impurity contribution to the particular current spectrum (111') as a whole remains below an equivalent integer of 0.5 protons per impurity.

9. The method according to claim 1, further comprising: After the signal interval has been identified, a linear phase shift is applied (1420) to each current spectrum (111), wherein the applied linear phase shift is statistically distributed across the plurality of current spectra.

10. The method according to claim 1, further comprising: After the signal interval has been identified, a baseline variation is added (1430) to each current spectrum (111), wherein the baseline variation for a particular current spectrum is calculated using a term-by-term polynomial interpolation function over three to twelve sampling points called knots, and wherein the coordinates of the knots of the current spectrum are statistically uniformly distributed over the multiple current spectra.

11. The method according to claim 1, further comprising: After the signal interval has been determined, noise is added (1440) to each current spectrum (111), wherein the noise amplitude follows a Gaussian distribution and a standard deviation of the Gaussian distribution varies statistically across the multiple current spectra.

12. A computer-implemented method for training (1600) a neural network for signal analysis in NMR spectra to enable the neural network to identify signal intervals in a real NMR spectrum obtained from a real NMR experiment, the method comprising: receiving (1610) a training data set as produced according to any of the preceding claims having labels for said identified signal intervals; as well as The neural network (230) is trained (1620) using a supervised learning method with the received current spectrum (111') of the training data set (141) and the associated label identifier (121), wherein the training input to the neural network is the current spectrum (111') of the training data set and the output is the individual signal intervals.

13. A computer-implemented method (1700) for signal analysis in NMR spectroscopy, comprising: receiving (1710) a real NMR spectrum (202) obtained from a real NMR experiment as a test input for a neural network (230) trained according to claim 12; applying (1720) the trained neural network (230) to the test input (202); One or more signal intervals (203) as identified by the trained neural network (230) are provided (1730) as outputs of the trained neural network (230).

14. A computer program product for generating a realistic training data set for training a neural network for signal analysis in NMR spectroscopy, comprising instructions which, when loaded into a memory of a computer system and when executed by at least one processor of the computer system, cause the computer system to perform the method steps according to any one of claims 1 to 11.

15. A computer system (100) for generating a realistic training data set for training a neural network for signal analysis in NMR spectroscopy, the system comprising modules adapted to perform the steps of the computer-implemented method according to any one of claims 1 to 11.

Images