Automatic detection of boundary in mass spectroscopy data

Abstract

To provide an automatic detection method and system of a boundary in mass spectroscopy data.SOLUTION: In a system and method for automatic detection of the presence or absence of a quantity based on intensity expressed in terms of or derived from a frequency or time dependent data, intensity of mass spectroscopy is identified using a non-linear mathematical model, such as an artificial neural network trained to find out start and stop peaks of the intensity from which a present amount can be determined.SELECTED DRAWING: Figure 12

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to and the benefit of U.S. Provisional Patent Application No. 62 / 826,228, filed March 29, 2019, the entire contents of which are incorporated herein by reference in their entirety for all purposes.

[0002] FIELD OF THE DISCLOSURE This disclosure relates generally to the detection of physical entities using mass spectrometry, and more particularly to the automated detection of biological entities using multiple reaction monitoring mass spectrometry (MRM-MS). [Background technology]

[0003] Multiple reaction monitoring (MRM) is a technique utilized in tandem mass spectrometry (MS) that allows for sensitive and specific detection of proteins, lipids, and post-translational modifications (PTMs), among other analytes. This technique holds great promise for use in novel clinical analyses and diagnostics, but the availability of accurate analytical software remains a bottleneck.

[0004] Traditional analysis of MRM experiments involves manually selecting the start and end of chromatographic peaks and working from transitions previously characterized by data-dependent acquisition or other discovery techniques. A typical embodiment can be seen in Agilent's MassHunter software, where intensities are plotted over a pre-specified range of retention times (RT) for given precursor and product mass-to-charge ratios. Selecting the beginning and end of the peaks results in observed RTs, peak widths, and integrated abundance values. Individual variability in evaluation, human error, and large time investments make this method inadequate for use outside of research. In recent years, many software packages have attempted to fill this gap by automatically integrating peaks from a list of transitions.

[0005] Existing software packages have proven very useful in quantifying, for example, peptides when they have high relative abundance in mass spectrometry data, but these software packages are not accurate in integrating highly heterogeneous species at relatively low abundance, such as glycated peptides. Summary of the Invention

[0006] This disclosure describes methods and systems for the automated detection of the presence of quantities expressed in terms of or derived from time series data. This disclosure also describes methods and systems for the automated detection of the presence of quantities expressed in terms of or derived from frequency dependent data.

[0007] In one embodiment, the automatic detection is applied to a time series of intensity versus time. The intensity comes from a detector that detects the presence or absence of a quantity. For example, the automatic detection can be applied to mass spectrometry (MS), where the intensity comes from an MS detector that detects transition values ​​as a function of time, such as retention time. Thus, the automatic detection described herein can be applied to a wide variety of data inputs based on observed intensity values ​​as a function of time or as a function of frequency.

[0008] In one embodiment, automated detection may be applied to time series of mass-to-charge ratio intensity versus time generated from a process involving mass spectrometry (MS), hi another embodiment, automated detection is applied to time series of intensity versus time generated from a process involving liquid or gas chromatography.

[0009] According to another example, automatic detection may be applied to any of the following intensities expressed in terms of or derived from time or frequency data: X-ray diffraction resulting from an X-ray detector, e.g. a charge-coupled device (CCD), detecting the reflection of X-rays as a function of X-ray scan angle, where intensity is observed as a function of time and scan angle, e.g. 2θ. - Intensity, e.g. cm ―1 etc., Raman spectroscopy arises from infrared detectors that detect vibrational modes as a function of absorption or emission of light wavelengths. ·UV-VIS spectroscopy, where the intensity arises from a photodetector detecting the intensity of electronic transitions as a function of ultraviolet and visible wavelength or frequency. ·Fluorometric or phosphorescent spectroscopy, where the intensity of emitted fluorescent or phosphorescent light is detected as a function of the incident, i.e. absorbed or excited, wavelength. Nuclear Magnetic Resonance (NMR) spectroscopy, in which intensity results from a radio wave detector detecting nuclear transitions as a function of absorption or emission of a radiation frequency in a strong magnetic field. In NMR, the detection of intensity versus frequency can be converted to an equivalent intensity versus time, for example using a Fourier transform. The same principles applicable to NMR are also applicable to Magnetic Resonance Imaging (MRI).

[0010] In one particular embodiment, the automated detection is applied to a time series of mass-to-charge ratio intensities versus time generated from a process involving mass spectrometry. According to this embodiment, the automated detection involves using one or more mathematical models to detect where start and end peaks occur in the MS data. By providing accurate selection of these boundary points, a more accurate assessment of abundance is made.

[0011] According to one embodiment, a method of training a machine learning model to identify mass spectrometry peaks uses as a training set a sequence of points representing mass / charge ratios labeled with peak starts and peak stops, and the trained model is adapted to identify mass spectrometry start and stop peaks that indicate the presence of one or more glycopeptides or peptides, or fragments thereof, present in a biological sample.

[0012] According to another aspect, a method for identifying mass spectrometry peaks employs at least one trained neural network model that generates an output as a function of an input parameter, where the value of the input parameter may be a sequence of retention times of glycopeptides having m / z (mass-to-charge) ratios.

[0013] According to another aspect, the raw MS data is featured to generate a training set for a machine learning model for the purpose of training the model to identify peak start times and peak stop times in the MS data. The featurization may include discretizing analog samples of the MS data into a sequence of points, where retention times corresponding to peak areas in the analog samples are excluded from the sequence of points, and assigning labels to points in the sequence of points corresponding to peak start times and peak stop times to generate a sequence of labeled points. The labels may be found using human annotators trained to select peak start times and peak stop times using a display device and an input device.

[0014] According to another embodiment, the method calculates a distribution of probabilities over a sequence of points representing MS data for a given m / z ratio and retention time, the probabilities representing likely peak start or stop retention times. From the distribution of these probabilities, abundance statistics are calculated. The statistics may be averages.

[0015] According to another aspect, a method for diagnosing a disease includes quantifying mass spectrometry transitions, where the transitions are m / z intensity value peaks characteristic of a glycopeptide or peptide, or fragment thereof, including an amino acid sequence, and analyzing the quantification using a model to determine a probability that the sample falls or does not fall within a disease classification based on the quantification. In some examples, the glycopeptide or peptide comprising the amino acid sequence, or a fragment thereof, is selected from those sequences described in International Patent Application Publication No. WO2019 / 046814-Identification and use of glycopeptides as biomarkers for diagnosis and treatment monitoring, International Patent Application Publication No. WO201907639A1-Identification and use of biological parameters for diagnosis and treatment monitoring, U.S. Patent Application Publication No. US20190101544A1-Identification and use of glycopeptides as biomarkers for diagnosis and treatment monitoring, International PCT Patent Application No. PCT / US2020 / 0162861, filed January 31, 2020, and U.S. Provisional Patent Application No. 62 / 989,510, filed March 13, 2020. The entire contents of each of these patent applications are incorporated herein by reference in their entirety for all purposes. [Brief description of the drawings]

[0016] [Figure 1] 1 illustrates a process for calculating abundances from mass spectrometry data using machine learning models. [Diagram 2]Figure 1 shows the process for training the machine learning model. Model performance on test set A, bar plot of absolute error at boundary prediction and test set B, cumulative distribution of boundary prediction error. Note that 19039 (28.1%) samples are outside the range of the rule-based prediction (>6 s), 4824 (7.1%) samples are within the sequential NN, and 1470 (2.2%) samples are within the criteria-based NN. C, Scatter plot of peak abundance predictions / annotations on the test set. [Diagram 3] 3 shows a process for generating a sequence of points from an MS sample that is used as input to the machine learning model of FIG. 2. [Figure 4] A process for calculating abundance based on calculated probabilities of peak start retention times and peak stop retention times of an MS sample. [Figure 5A] A and B show the model architecture of the first machine learning model (Model A), which is a bidirectional recurrent neural network with long and short-term memory and an attention layer. [Figure 5B] A and B show the model architecture of the first machine learning model (Model A), which is a bidirectional recurrent neural network with long and short-term memory and an attention layer. [Figure 6A] We present the model architecture for the second machine learning model (Model B), which is a bidirectional recurrent neural network with long and short-term memory and an attention layer. [Figure 6B] We present the model architecture of the second machine learning model (Model B), which is a bidirectional recurrent neural network with long and short-term memory and an attention layer. [Figure 6C]The attention map of a sample using Model B is shown in Figure 6C. In Figure 6C, the sample peak and reference-based model predictions are for the MRM transition of SEQ ID No. 1 of International Patent PCT Application No. PCT / US2020 / 0162861, filed January 31, 2020, which is incorporated by reference in its entirety for all purposes, and its reference peak shown in the bottom right panel, and the attention map (extracted from the model) in the bottom left panel. [Figure 7A] Shown are sample labeled inputs and calculated probabilities (boundary predictions) for Model A and Model B. The star and triangle annotations represent the labeling of the MS data for training the models. The green curve represents the prediction using Model B, and the blue curve represents the prediction using Model A. [Figure 7B] Shown are sample labeled inputs and calculated probabilities (boundary predictions) for Model A and Model B. The star and triangle annotations represent the labeling of the MS data for training the models. The green curve represents the prediction using Model B, and the blue curve represents the prediction using Model A. [Figure 8A] We show the performance of Model A (blue, 2) and Model B (green, 3) against a known machine learning architecture trained on the same data (rule-based method - red, 1). In Fig. 10, each set of three columns is presented from left to right as: rule-based, MAE 4.79 s; sequential NN, MAE 2.33 s; and baseline-based NN, MAE 1.56 s. [Figure 8B] We show the performance of Model A (blue, 2) and Model B (green, 3) against a known machine learning architecture trained on the same data (rule-based method - red, 1). In Fig. 10, each set of three columns is presented from left to right as: rule-based, MAE 4.79 s; sequential NN, MAE 2.33 s; and baseline-based NN, MAE 1.56 s. [Figure 8C]We show the performance of Model A (blue, 2) and Model B (green, 3) against a known machine learning architecture trained on the same data (rule-based method - red, 1). In Fig. 10, each set of three columns is presented from left to right as: rule-based, MAE 4.79 s; sequential NN, MAE 2.33 s; and baseline-based NN, MAE 1.56 s. [Figure 9A] Additional sample labeled inputs and calculated probabilities (boundary predictions) for Model A and Model B are shown. The star and triangle annotations represent the labeling of the MS data for training the models. The green curve represents the prediction using Model B, and the blue collinear represents the prediction using Model A. Reference and sample peaks of a glycopeptide transition. A: For the reference peak, the grey line represents the input curve, the red stars and red triangles are the human annotated peak start and end, and the blue / green solid and dashed lines show the sequential / reference-based neural network model predictions of peak start and end probabilities. B: Sample peaks of the same transition with different input curves. Note that the reference-based model (green) outputs much more consistent predictions. [Figure 9B] Additional sample labeled inputs and calculated probabilities (boundary predictions) are shown for Model A and Model B, respectively. The star and triangle annotations represent the labeling of the MS data for training the models. The green curve represents the prediction using Model B, and the blue collinear represents the prediction using Model A. Reference and sample peaks of a glycopeptide transition. A: For the reference peak, the grey line represents the input curve, the red stars and red triangles are the human annotated peak start and end, and the blue / green solid and dashed lines show the sequential / reference-based neural network model predictions of peak start and end probabilities. B: Sample peaks of the same transition with different input curves. Note that the reference-based model (green) outputs much more consistent predictions. [Figure 10]Figure 10 shows the relative standard deviation of boundary point selection for human annotators (gray, 1) versus Model A (blue, 2) and Model B (green, 3). The vertical bar plots are the relative standard deviation (RSD) of 13 human annotators and the relative error (RE) of sequential / criterion-based neural network predictions. In Figure 10, each set of three vertical bars is presented from left to right as sequential NN, mean RE 0.034, criterion-based NN, mean R 0.019, and human annotators, mean = 0.025. [Figure 11A] Scatter plots comparing abundance calculations between Model A and Model B. Scatter plots of peak abundance predictions / annotations for sequential (A) and criterion-based (B) neural networks. For the sequential model predictions, weighted abundances (purple dots) show higher consistency with fewer outliers than the ground truth and maximum point set abundances (blue dots), and Pearson's r is higher at 0.9969. Note that weighted abundances are biased upwards in the low abundance range. For criterion-based predictions, the results are very similar. [Figure 11B] Scatter plots comparing abundance calculations between Model A and Model B, respectively. Scatter plots of peak abundance predictions / annotations for sequential (A) and criterion-based (B) neural networks. For the sequential model predictions, weighted abundances (purple dots) show higher consistency with fewer outliers than the ground truth and maximum point set abundances (blue dots), and Pearson's r is higher at 0.9969. Note that weighted abundances are biased upwards in the low abundance range. For criterion-based predictions, the results are very similar. [Figure 12] We show the confidence score for Model B, and the equation used to calculate the confidence score. Confidence score diagram: dashed blue line (*) - model predicted peak termination probability, dark red line (#) - maximum point set prediction, light red line (!) - other local maxima. The confidence score (0.634) is the area of ​​the shaded region divided by the area inside the black box. [Figure 13A]Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 13B] Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 13C] Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 13D]Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 13E] Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 13F] Histograms of prediction confidence, mean absolute error of boundary prediction, and scatter plots of bins of different confidence scores are shown for model A. A: Histogram of prediction confidence for sequential neural network. Abundance predictions for samples of four bins are shown in the following panels. B: The mean absolute error (MAE) of boundary prediction steadily decreases over samples with increasing certainty. C, D, E, F: Scatter plots of peak abundance predictions / annotations within bins of different confidence scores. Samples with higher certainty achieved better correlation scores. Note that the reported log-abundance Pearson's r values ​​are calculated for combinations of samples from different transitions. [Figure 14A]For model B, we show histograms of prediction confidence and mean absolute error of the boundary predictions. A: Histogram of prediction confidence for the criterion-based neural network. B: The mean absolute error (MAE) of the boundary predictions decreases steadily over the samples as confidence increases. [Figure 14B] For model B, we show histograms of prediction confidence and mean absolute error of the boundary predictions. A: Histogram of prediction confidence for the criterion-based neural network. B: The mean absolute error (MAE) of the boundary predictions decreases steadily over the samples as confidence increases. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] In the specification, reference numbers appearing in the drawings and the description designate corresponding or similar elements among the different figures.

[0018] definition As used herein, the following words and phrases are intended to have the meanings set forth below, unless the context in which they are used clearly indicates otherwise.

[0019] "Transition" refers to the time series of mass-to-charge ratio intensities in mass spectrometry. As used herein, the phrase "transition" or "multiple reaction monitoring (MRM) transition" refers to a mass-to-charge (m / z) peak or signal observed when a glycopeptide or a fragment thereof is detected by MRM-MS. MRM transitions are detected as precursor and product ion transitions. The time interval for transition detection depends on the feed rate of the tandem liquid chromatography instrument and the flow rate in the MRM analyzer.

[0020] "Biological amount" is understood to mean the amount of an analyte in a biological sample or the amount of a solute in a tandem chromatography instrument. Analytes and solutes include, but are not limited to, polysaccharides, peptides, glycopeptides, glycated peptides, proteins, glycoproteins, and fragments thereof.

[0021] As used herein, the phrase "biological sample" refers to a sample derived from an organism or from a fluid or tissue of an organism, obtained from an organism or from a fluid or tissue of an organism, generated from an organism or from a fluid or tissue of an organism, provided from an organism or from a fluid or tissue of an organism, taken from an organism or from a fluid or tissue of an organism, or removed from an organism or from a fluid or tissue of an organism. Biological samples include, but are not limited to, synovial fluid obtained by biopsy, whole blood, serum, plasma, urine, sputum, tissue, saliva, tears, spinal fluid, tissue section(s), cell(s) placed or adapted for tissue culture, sweat, mucous membrane, fecal material, gastric juice, peritoneal fluid, amniotic fluid, cyst fluid, peritoneal fluid, pancreatic juice, milk, lung lavage fluid, bone marrow, gastric acid, bile, semen, pus, aqueous humor, transudate, etc., and derivatives, portions, and combinations of the above. In some examples, biological samples include, but are not limited to, blood and / or plasma. In some examples, biological samples include, but are not limited to, urine or feces. Biological samples include, but are not limited to, saliva. Biological samples include, but are not limited to, tissue autopsies and tissue biopsies. Biological samples include, but are not limited to, any derivatives or portions of the above biological samples.

[0022] As used herein, the term "polysaccharide" refers to the carbohydrate residue of a complex carbohydrate, such as a glycopeptide, glycoprotein, glycolipid, or the carbohydrate portion of a proteoglycan.

[0023] As used herein, the term "glycopeptide" refers to a peptide having at least one polysaccharide residue attached thereto.

[0024] As used herein, the phrase "glycosylated peptide" refers to a peptide attached to a polysaccharide residue.

[0025] As used herein, the phrase "glycopeptide fragment" or "glycopeptide fragment" refers to a glycated peptide (or glycopeptide) having an amino acid sequence that is the same as part (but not all) of the amino acid sequence of a glycated protein. A glycated peptide is obtained from a glycated protein, for example, by digestion with one or more proteases or by fragmentation, such as, for example, ion fragmentation in an MRM-MS instrument. MRM refers to multiple reaction monitoring.

[0026] As used herein, the phrase "fragmenting a glycopeptide" refers to the process of ion fragmentation that occurs in an MRM-MS instrument. Fragmentation can produce a variety of fragments that have the same mass but vary in terms of their charge.

[0027] As used herein, the phrase "multiple reaction monitoring mass spectrometry (MRM-MS)" refers to a highly sensitive and selective method for targeted quantification of polysaccharides and peptides in biological samples. Unlike traditional mass spectrometry, MRM-MS is highly selective (targeted), allowing researchers to fine-tune the instrument to specifically look for specific peptide fragments of interest. MRM allows for greater sensitivity, specificity, speed, and quantification of peptide fragments of interest, such as potential biomarkers. MRM-MS involves the use of one or more of triple quadrupole (QQQ) mass spectrometers and quadrupole time-of-flight (qTOF) mass spectrometers.

[0028] As used herein, "peptide" is meant to include glycopeptides, unless otherwise specified.

[0029] "Retention time (RT)" is understood to mean the time it takes for a solute to pass through a chromatographic column. RT is calculated as the time from injection to detection. RT depends on several factors, including but not limited to gas flow rate, temperature difference between the oven and inside the column, column aging, column length, column type, and carrier gas.

[0030] Mass spectrometry (MS) data or data set is understood to mean a collection of spectra plotting the intensity of a transition as a function of transition number. The MS data for a given transition value is the mass (m) compared to charge (z) intensity level (m / z) over a period (t) which is the retention time (RT).

[0031] "Abundance" is understood to mean the calculated amount of any biological quantity represented in the mass spectrometry data. Abundance may be calculated by summing values ​​at discrete time points between the start and stop peak boundaries. Abundance may also be calculated by integration over discrete time steps between the start and stop peak boundaries. Integration may be performed using any standard method, such as Σ(t i+1 -t i )*(1 / 2)*[I(t i )+I(t i+1 )], i=1..n-1 (n is the number of discrete points between the boundaries, and I(t i ) is the retention time t i Unless otherwise specified, abundance refers to the relative abundance of one integrated MRM peak relative to another MRM peak.

[0032] Nonlinearity limits the mathematical models that may be developed to practice aspects of the invention to nonlinear mathematical models in the sense that there is no set of optimized functions / model parameters based on finding a minimum, but rather the optimized functions / model parameters reside only in local minima (there are no local maxima). Contemplated model architectures include, but are not limited to, artificial neural networks, graph models, and support vector machines.

[0033] A "machine learning model" or "learning model" is understood to mean any of a variety of mathematical models that have at least one nonlinear activation layer (in addition to one or more affine transformations). Machine learning models are further trained or optimized by minimizing one or more loss functions that are separate from the model itself (e.g., minimizing cross-entropy loss or negative log-likelihood). The training or optimization process seeks to optimize the model to reproduce known outcomes (low bias) and enable the model to make accurate predictions from unseen experience (low variance). The output of the model can be any of a variety of things related to the task, such as probabilities or statistics, classifications, sequences, etc. In one particular example, the data used to train and optimize the model is sequential data, such as time series data, where the target outcome is known, and the sequential data with known target output is taken to mean labeled sequential data or annotated sequential data.

[0034] One example of a well-known model suitable for making predictions based on labeled sequential data is a unidirectional or bidirectional recurrent neural network (RNN) with a long short-term memory (LSTM) unit or a gated recurrent unit (GRU), optionally with an attention mechanism. The attention mechanism (or attention layer) may retain more important information from past state variables than may be considered in the final LSTM unit or GRU, or combination of LSTM units or GRUs, of the sequence, along with the bidirectional RNN, which represents the accumulated knowledge from past states to influence future state predictions. The attention mechanism applies learned weights to each of the input states to highlight less important or more important features of the input MS data points that appear in the sequence when generating the output prediction. These learned weights are more or less directly coupled to the corresponding output state, or to each of the final output states, depending on the desired output.

[0035] Models trained and optimized using only labeled data are known as supervised models. Learning models may also be unsupervised or semi-supervised models. These models are trained and optimized using unlabeled data or a combination of labeled and unlabeled data. Models trained and optimized using a combination of labeled and unlabeled data or primarily unlabeled data are known as semi-supervised and unsupervised models, respectively. Learning models can be efficiently trained and optimized by the well-known technique of stochastic gradient descent (SGD). The hyperparameters of the learning model may be optimized by manual tuning during training and / or validation, or may be learned by a method that minimizes a surrogate loss function with respect to one or more of the hyperparameters in parallel with minimizing the global loss function of the learning model.

[0036] A "training set" is a data set used to train or tune parameters in a learning model to reproduce a known target output.

[0037] A "validation set" is a data set used to tune the model's predictive capabilities against new experience to reach an optimal balance between matching the labeled data (low bias) and predicting start / stop locations for unseen transitions (low variance). The validation set is applied after the model has been tuned based on the training set.

[0038] A "test set" (or left-out validation set) is a data set used to evaluate the predictive ability of an optimized model. The performance of the model against the test set is the metric used to evaluate the model's ability to make accurate predictions of new, unseen transitions.

[0039] "Featurization" is understood to mean the process associated with converting raw data representing samples from a distribution for use as input to a machine learning model. Examples of featurization include discretizing analog signals into digital signals, filtering or denoising or clipping portions of data (e.g., reducing the signal-to-noise ratio), signal normalization and signal standardization (e.g., centering a target peak in a window). Featurization may also include introducing noise into a dataset to validate the model or associated enhancements of the sample for generalization and / or to enhance the model's ability to make predictions from new (or unseen) experiences.

[0040] According to the present disclosure, the dataset of sequential data from which the training dataset, validation dataset, and test dataset are sampled may include at least labeled start and stop locations for transitions, each of which indicates a boundary point for computing abundances by integration between the boundary points. In a preferred embodiment, the targeted output is the probability of the start and stop locations of each point in the sequential data of the transition.

[0041] explanation As mentioned above, traditional analysis of MRM experiments involves manually picking the start and end of a chromatographic peak and working from transitions previously characterized by data-dependent acquisition or other discovery techniques. Software suites such as Agilent's MassHunter software plot signal intensity over a pre-specified range of retention times (RT) for a given precursor and product mass-to-charge ratio. Manual selection of the beginning and end of the peak results in observed RTs, peak widths, and integrated abundance values. However, personal variability in evaluation, human error, and large time investments make this method unsatisfactory for use outside of research.

[0042] While the above manual process of choosing boundary points to calculate abundances may be effective for transitions characterized by a single well-defined peak and a high signal-to-noise ratio, this method may not provide accurate results or produce consistent results when detecting abundances of biological masses characterized by a low signal-to-noise ratio.

[0043] Referring to Figure 1, in accordance with the present disclosure, a trained machine learning model is instead used to automatically select boundary points in a mass spectrum for a specified retention time window and mass / charge intensity. The input to the machine learning model is a sequence of points (samples of MS data). Each point represents a mass / charge intensity for a corresponding retention time.

[0044] Referring to Figure 2, according to one embodiment, the machine learning model is a neural network that generates probabilities of peak start and stop times for each corresponding point provided as input. The neural network is trained using mass spectra with labeled target peak start and stop retention times. Through a training and validation process, optimal model parameters are found. The labeled data may be obtained from human annotators who select peak start / stop times using a graphical user interface and an input device.

[0045] Referring to Figure 3, there is a process flow for featurizing raw mass spectral data for input into the machine learning model of Figure 1. The featurization step may include (1) removing mass / charge intensities above a certain threshold amount, or instead choosing only those points that are mostly noise and close to the signal-to-noise transition region of the MS data, and (2) discretizing the analog signal into a set of points to generate sequential data.

[0046] Referring to FIG. 4, in one particular embodiment, the model calculates abundance amounts based on weighted sums of integrals. One of the most straightforward methods is to interpret the prediction results of the machine learning model as a probability distribution. In other words, let the model evaluate the probability of all possible pairs of points in the sequence that will serve as peak start / end markers. For a sequence of N points, this includes a total of N×(N-1) / 2 pairs after imposing order constraints. The probability of each pair can be easily calculated by multiplying the predicted values ​​of the peak start / end task by the two points. The probabilities are then used to calculate weighted sums for integrals between corresponding pairs, providing an average estimate instead of a single abundance value calculated from the two end points with the highest probability of the true boundary. The average abundance may provide a more robust abundance estimate.

[0047] Embodiment In some examples, described herein are methods of training a model to identify mass spectrometry peaks, including using a plurality of labeled sequential data, each of the plurality of labeled sequential data including a peak start label and a peak stop label representing a retention time value specific to a glycopeptide or peptide, or a fragment thereof, and training at least one machine learning model using the labeled sequential data, wherein the trained model is adapted to identify mass spectrometry start and stop peaks that are indicative of the presence of one or more glycopeptides or peptides, or fragments thereof, present in a biological sample.

[0048] In some examples, including any of the above, the method further includes quantifying the presence of one or more glycopeptides or peptides from the biological sample by integration over time between a start peak and a stop peak of the mass spectrometry.

[0049] In some examples, including any of the above, the machine learning model is a recurrent neural network model.

[0050] In some examples, including any of the above, the recurrent neural network is a bidirectional LSTM recurrent neural network model with an attention layer.

[0051] In some examples, methods are described herein that include calculating the abundance of glycopeptides or peptides using a machine learning model trained to perform the processes described herein.

[0052] In some examples, described herein is a system that includes a data repository of labeled sequential data, one or more processors coupled to the repository, and machine-executable code resident in non-transitory memory accessible to the one or more processors that, when executed by the one or more processors, performs the methods described herein.

[0053] In some examples, described herein are methods for identifying mass spectrometry peaks, comprising providing at least one trained neural network model that generates an output as a function of input parameters, where the values ​​of the input parameters include a sequence of retention time values ​​of glycopeptides having m / z (mass-to-charge) ratios, and providing mass spectrometry (MS) data comprising mass-to-charge (m / z) ratios for the sequence of mass spectrometry retention time values ​​of a biological sample as values ​​of the input parameters, where the output comprises locations of one or more peak start retention time values ​​and peak stop retention time values ​​indicative of the presence of one or more analytes or fragments thereof in the biological sample.

[0054] In some examples, including any of the above, the method further includes quantifying the presence of one or more glycopeptides or peptides by summing or integrating retention time values ​​between a start peak and a stop peak of a mass spectrometry method from the biological sample.

[0055] In some examples, including any of the above, the method further includes quantifying the presence of one or more glycopeptides or peptides by summing over retention time values ​​between a start peak and a stop peak of a mass spectrometry from the biological sample. In some examples, including any of the above, the method further includes quantifying the presence of one or more glycopeptides or peptides by integrating over retention time values ​​between a start peak and a stop peak of a mass spectrometry from the biological sample.

[0056] In some examples, including any of the above, the quantifying step includes applying a weighted smoothing function to the predicted start and stop peaks.

[0057] In some examples, described herein is a system that includes a data repository of mass spectrometry (MS) data, one or more processors coupled to the repository, and machine-executable code resident in non-transitory memory accessible to the one or more processors that, when executed by the one or more processors, performs the methods described herein.

[0058] In some examples, methods are described herein that include featurizing data from raw MS data, where the raw MS data includes a plurality of analog samples each representing a mass / charge intensity, and the featurizing includes, for each of the analog samples, centering the sample within a retention time window; discretizing the samples into a sequence of points representing intensity values; standardizing the intensity values; and assigning labels to points within the sequence of points corresponding to peak start times and peak stop times to generate a sequence of labeled points, where the sequence of labeled points is configured for training a machine learning model to predict the abundance of unseen MS data.

[0059] In some examples, including any of the above, centering includes placing the apex of the curve at the center of the retention window.

[0060] In some examples, including any of the above, normalizing the intensity values ​​excludes intensity values ​​greater than a threshold amount, which in some examples, including any of the above, is about 500.

[0061] In some examples, described herein is a system that includes a data repository of raw mass spectrometry (MS) data including a plurality of analog samples, one or more processors coupled to the repository, and machine executable code resident in a non-transitory memory accessible to the one or more processors that, when executed by the one or more processors, performs any of the methods. In some examples, described herein is a method that includes calculating a distribution of probabilities over a sequence of points representing mass spectrometry (MS) data of a given m / z ratio and retention time period, the distribution of probabilities representing likely peak start retention times or peak stop retention times, and calculating abundance statistics using the distribution of probabilities, the statistics indicating the presence of a biological sample in the MS data.

[0062] In some examples, including any of the above, the statistic is an average abundance calculated as a weighted sum of integrals using all possible pairs of start and end points to evaluate the integrals.

[0063] In some examples, including any of the above, the machine learning model is a sequential model (Model A) or a criterion-based model (Model B).

[0064] In some examples, methods are described herein that include using a plurality of labeled sequential data and training at least one machine learning model using the labeled sequential data, where the trained model is adapted to identify periods or frequency ranges in which start and stop peaks in intensity occur in unseen samples.

[0065] In some examples, including any of the above, the start and stop peaks fall within the noise-to-signal transition period.

[0066] In some examples, including any of the above, the machine learning model is a recurrent neural network with an attention layer.

[0067] In some examples, including any of the above, the samples represent a physical entity such that the physical entity is quantified by addition or integration over time or frequency from a start peak to a stop peak in intensity.

[0068] In some examples, including any of the above, the sample is a continuous series of mass-to-charge intensities over a specified time period, where the sum or integral represents the amount of a physical entity present.

[0069] In some examples, methods are described herein that include providing at least one trained neural network model that generates an output as a function of input parameters, where the values ​​of the input parameters are a time series or spectral series of intensity values ​​representing a quantity, and providing a value for each of the input parameters, where the values ​​represent a sequence of invisible intensities of the quantity, and the output includes one or more peak start and peak stop times or frequencies used to quantify the quantity.

[0070] In some examples, including any of the above, the quantity is quantified by summing or integrating over the time or frequency between the time or frequency of the peak onset and the time or frequency of the peak off, respectively.

[0071] In some examples, including any of the above, the intensities are obtained from a process selected from the set of mass spectrometry, MRM, liquid or gas chromatography, x-ray diffraction, Raman spectroscopy, UV-VIS spectroscopy, fluorimetry or phosphorimetry spectroscopy, nuclear magnetic resonance (NMR) spectroscopy, or magnetic resonance imaging (MRI).

[0072] Working Example The following examples describe studies conducted in accordance with several aspects of the present disclosure.

[0073] In this study, two learning models were trained, validated, and then tested to evaluate their accuracy in predicting the location of the peaks (or more precisely, the start / stop boundaries of the peaks to more precisely calculate the abundance of the associated transitions). The specific architecture chosen for the learning models was a bidirectional recurrent neural network with long- and short-term memory and attention layers (RNN-LSTM-AL). The model architectures and training methods are shown in Figure 5A, Figure 5B, Figure 6A, Figure 6B, and Figure 6C. The input parameters or input state variables of the models correspond to the time series representing the mass-to-charge ratio (m / z) intensity.

[0074] The study had access to a large dataset of manually labeled peaks from mass spectrometry (MS) transitions spanning a wide dynamic range, including both peptides and intact glycopeptides. The dataset contained a time series of intensities within the retention time (RT) of each transition. Each of these transitions was sequentially input into the RNN-LSTM-AL, and included with each transition were the associated human-annotated (or labeled) start RTs (indicating boundary points for integration to derive abundances), the start RTs and stop RTs being the labels used for supervised learning, validation, and testing, respectively.

[0075] 2.1 Dataset Two mass spectrometry data sets were used for the study. The raw Agilent QQQ MS output for the first experiment consisted of 716 transitions evaluated in 210 human serum samples. This resulted in 106355 valid sample peaks after filtering for positive abundance and RT shifts. This set was split into a training set with 572 transitions and a validation set with 144 peaks.

[0076] The second experiment used MS output generated from 135 human serum samples purchased from a commercial biorepository. 503 transitions were evaluated, resulting in 67,672 high-quality peaks that were used exclusively for testing. 313 test set transitions were not present in the training / validation sets. The 313 test set transitions were used to compare model output predictions.

[0077] 2.2 Baseline Methods The performance of the trained and validated RNN-LSTM-AL model was compared to two baseline predictors. The first of these baseline predictors is Skyline, a well-known open-source software tool used for data analysis of mass spectrometry applications including multiple reaction monitoring. The Skyline software calculates the peak areas (abundances) of the test samples. The second baseline predictor for comparison with the RNN model is a deterministic method of prediction, henceforth referred to as the "rule-based" method. The architecture of the rule-based method is described in Aoshima Ken, Takahashi Kentaro, Ikawa Masayuki, Kimura Takayuki, Fukuda Mitsuru, Tanaka Satoshi, Parry Howell Yuichiro, Fujita Akiyasu C Yoshizawa, Utsunomiya Shinichi et al. A simple peak detection and label free quantitation algorithm for chromatography-mass spectrometry, BMC bioinformatics, 15(1):376,2014. A rule-based method model was developed for this study by adjusting its model parameters to maximize performance on the training and validation sets, and these adjusted parameters were then applied during evaluation of the test set.

[0078] 2.3 Sequential model for peak integration The RNN-LSTM-AL architecture was chosen to predict from MS data-on-sequence in the study.

[0079] A visual block description of the training / prediction process and flow of the first of the two RNN-LSTM-AL models used in the study (hereafter, “Model A” or “Sequential Model”) is shown in Figure 5A and Figure 5B. The input (mass / charge ratio intensity as a function of time) is encoded through a bi-LSTM layer and attention mechanism to associate separate locations. Prediction is point-wise, meaning that for every input, the model predicts an output. The model was trained to generate a sequence of outputs with N × 2 dimensionality, where “N” is the number of input parameters of Model A (points corresponding to retention times where intensity values ​​were input to the model, ranging from 50 to 300 depending on the nature of the curve in the sample, see 5.3 below), the number of data points in the sequence, and the output value corresponding to each input value. The two (2) values ​​output for each of the N points are the probability (ranging from 0 to 1) that the corresponding point represents a peak start or peak stop. The point or neighborhood of points output by the model whose value is closest to 1 (high probability) for a given transition is then used to bound the interval over which the abundance integral (integrated over time) is calculated.

[0080] The output probabilities from the model thus form a distribution of boundaries. The model is trained end-to-end by gradient descent on cross entropy loss between the predicted boundaries and ground truth, or manually labeled peak start and peak stop. A minimum hyperparameter search, e.g., Bayesian optimization applied to an objective function minimized with respect to the model's hyperparameters, is applied to optimize performance on a leave-out validation or test set with the goal of achieving good generalization and avoiding overfitting on the training set. Featurization and hyperparameters are described in more detail below.

[0081] 2.4 Criterion-Based Sequential Model In some applications, due to the highly variable shape and width of the transition peaks, a unified predictive model such as Model A may not generalize well to unseen transitions. Based on the observation that the shapes of different patient samples at the same transitions are consistent, a second RNN-LSTM-AL model (hereafter "Model B" or "criterion-based model") was also developed.

[0082] A visual block description of the training / prediction process and flow of Model B is shown in Figure 6A, Figure 6B, and Figure 6C. Model B takes the query sample and the reference as input, respectively. Both encoded features are compared and merged in the sample-reference attention layer to generate predictions. Provided with the query peak features, the reference peak features, and the reference labels, the attention layer compares the query and the reference point by point, such that points in the query that are highly similar to the labeled start / stop points in the reference get a higher prediction probability. Given two consistent samples, the ideal attention map for this operation would be the "identity matrix" where points with the same surroundings are considered to be highly similar (see Figure 6C). A divergence term between the attention map and the retention time mapping matrix was included in the optimization process to achieve good regularization.

[0083] Note that in the context of criterion-based prediction (as opposed to query-based prediction), the task is no longer to identify noise-signal transition points as in model A, but rather to identify good coding mechanisms of points in a sequence and their context in order to optimize the similarity mapping. In application, criterion-based models operate similarly to few-shot learning models: they are trained on query / criterion pairs and aim to accurately map the criterion labels to predictions on query samples. This, of course, allows the model to perform better for unseen transition peaks, as long as a single (or a few) correctly labeled criterion is provided.

[0084] The adopted structure of Model B may be most applicable and cost-effective for high-throughput experiments on the same set of transitions, where manual annotation of several high-quality reference peaks is a reasonable prerequisite. For most discovery or small-scale experiments, a standalone reference-free predictor may be preferred.

[0085] 7A and 7B, each figure has two plots. The first plot is a plot of transitions with labeled start and stop times shown with stars and triangles, respectively. The bottom plot is a plot of the probability of the start / stop times predicted by the trained and validated model.

[0086] 3.1 Prediction performance of boundaries and abundances Leave-out or test set samples were used to evaluate the models. Several metrics were used to present a comprehensive picture of the problem being solved. Predictions are evaluated on two criteria: (1) accuracy in predicting boundaries (peak start / stop) and (2) accuracy in predicting abundance (peak area). While the former is directly related to the problem setting, the latter is much more important for downstream applications that utilize analyte abundance for analysis.

[0087] For prediction, the points in the curves with the maximum predicted probabilities for the two tasks, namely finding the peak start time and the peak end time, are used as markers for the boundary, which we call the "maximum point set boundary" in the following text. We calculated the mean absolute error (MAE) between the maximum point set boundary and the human annotation or label, along with the accuracy score defined as the proportion of samples whose boundary prediction is within an error threshold of 1.2 seconds (two points in a sequence) around the ground truth annotation (human-labeled point). Given that the peak duration is 20+ / -8 seconds (33+ / -14 points in the sequence) in the test set, the error threshold is small enough that any boundary prediction within the window does not significantly change the abundance.

[0088] For abundance assessment, first the baseline-adjusted integral between each boundary from the ground truth (human-labeled) and the model prediction is calculated, and then the Spearman rank correlation coefficient (Spearman's r) and the Pearson correlation coefficient (Pearson's r) between the two at each individual transition are evaluated. 15 ), Pearson's r is dominated by samples with large abundances, and here, to avoid bias, correlations calculated against log abundances are reported instead. The average score over all transitions is reported. In addition, for an indication of practical use, a precision score, defined as the proportion of samples whose predicted abundances are within + / - 5% of the annotated abundances, is also reported. There is one caveat to this precision metric: note that start / stop boundary prediction errors may cancel each other out, so that integrals calculated over intervals defined by boundary predictions that deviate from the annotated abundances may appear to have abundances very similar to the annotated abundances. That is, when intensity values ​​are summed in integrals, a 1% error in the predicted peak start time may cancel out a corresponding % error in the predicted peak stop time.

[0089] [Table 1]

[0090] [Table 2] * Percentage of samples for which the boundary prediction is within 1.2 seconds of error ** Proportion of samples with abundance predictions within 5% error

[0091] [Table 3]

[0092] Table 1 presents the performance scores of the three tested methods and Skyline predictions. Note that Skline only generated abundance predictions and was not evaluated and compared for the boundary prediction task. Both Model A and Model B demonstrated significant performance improvements in all tasks compared to the rule-based method and Skyline predictions. Model B achieved the highest score with additional baseline data. For boundary predictions, strong overfitting is observed since the rule-based method is tuned to optimize performance on the training / validation set (see Table 3). In comparison, thanks to better sample quality, Model A and Model B achieved MAEs of 2.33 and 1.56 seconds on the test set (with and without baseline), respectively, which are two times smaller than the rule-based method. Similarly, in accuracy score performance, Model B gained a 30% advantage over the rule-based method. Figures 8A and 8B show the bar graphs and non-normalized cumulative distributions of boundary prediction errors, respectively, with Model A and Model B presenting distributions with lower errors and fewer outliers.

[0093] Abundance prediction presented similar results: Models A and B achieved better and more robust performance when assessing peak abundance (Figure 8C), with an increase in accuracy of over 20%. Note that in this problem, boundary points usually contribute very little to the overall integral, thus greatly reducing the impact of boundary prediction errors. Correlation scores showed that Models A and B performed significantly better than current off-the-shelf mass spectrometry data analysis software, with Spearman correlation coefficients of 0.9975 and 0.9990 being nearly perfect. Interestingly, the rule-based method also achieved much higher correlation scores than Skyline predictions. Without wishing to be bound by any theory, this result could have been due to the complex composition of the test samples, which produced co-eluting peaks or compounds with close retention times. When the RT window was set closely around the desired transition (also the setting typically used by human annotators), Skyline could have mistakenly picked peaks from a different range, resulting in significantly worse performance.

[0094] Comparing Model A and Model B, because additional "reference data" were included during the training and validation phases of Model B, this model performed better than Model A, providing a 15% and 10% improvement in boundary and abundance prediction accuracy. Model B presented a better boundary error distribution with fewer outliers (2.2% vs. 7.1%, see Figure 8B), consistent with observations in the abundance correlation plot. The green points are more concentrated around the center line, while the blue points contain several outlying predictions (Figure 8C).

[0095] However, Model A may be preferred for some applications since a smaller training set is required. As the results show, Model A may provide superior accuracy / predicted abundances, so additional reference data used with Model B may not be required or may be practical for small-scale applications or experiments. Model B may be preferred for larger-scale experiments or when more accuracy is required.

[0096] In these experiments, model B offered another advantage not visible in the plots: greater consistency of predictions for the same transitions across multiple samples. Especially for transitions whose boundaries may vary due to shoulder and valley points, model B tends to follow the same predictive form of its criteria (Figure 9A), which is the primary objective of its model structure design. This is highly desirable in high-throughput experiments in that abundances are generated based on the very same standard, thus allowing for more accurate comparison and quantification. However, it should also be noted that in cases of retention time shifts or strong batch effects, strong prior assumptions in the model may undermine its performance.

[0097] 3.2 Comparison with human annotators To further validate the practical use of the neural network-based model, we compared the model predictions with a group of human annotators who calculated peak abundances.

[0098] Within the test set, a subset of 12 transitions was analyzed for all 135 serum samples, with 12 annotators independently marking peak starts / ends. Combined with the original test set labels, a set of 13 human annotators was prepared for this subset. We then calculated the relative standard deviation for each sample peak across the 13 annotations as a proxy for the variation between the different annotators, and compared it to the predicted relative error of Model A and Model B.

[0099] The results are summarized in Figure 10. A high consistency over the annotators is observed for the majority of samples, as evidenced by the grey vertical lines in Figure 10. The average RSD over all sample peaks for the human annotators is 2.5%. At the same time, both Model A and Model B perform well on the subset, achieving average relative errors of 3.5% and 2.1%, respectively, compared to the average of the human annotators. The distribution of errors is similar but slightly worse than the RSD across annotators, as shown by the blue and green vertical bars. Wilcoxon signed rank tests show no significance ($p=0.11, 0.50$) for both models, indicating that the differences are not significant.

[0100] Looking at the predictions of each individual annotator as a group, we then calculated the pairwise differences between the annotators, i.e., the average of the relative abundance differences at each sample peak. The results show an interannotator discrepancy of 0.025 + / - 0.009, consistent with the RSD value. At the same time, the difference between the predictions of model B and the 13 annotators is 0.026 + / - 0.007, overlapping with the interannotator difference. The corresponding value for model A is 0.039 + / - 0.005. Given that the model-to-human differences are not significantly large, or even closer in the criterion-based case, it is clear that neural network-based models can serve as a good substitute for manual annotation in this task. We also note that the human annotators had peaks selected whose start and stop served as the read criterion, making the task directly analogous to model B.

[0101] 3.3 Weighted peak start / end for abundance calculations For peaks with shoulder and valley points, the boundary determination suffers from much higher variability. Even in human annotations, variability exists among different annotators, so the output from Model A and Model B is expected to show a multi-modal shape due to averaging across training samples (Figure 9B, Figure 7B). This is somewhat severe even in the absence of a criterion (Model A), as there is no strong criterion prior, leading to indeterminate results when applying the maximum point set boundary. In these cases, methods were explored that could better accommodate full prediction instead of a single point estimate of the boundary.

[0102] One of the most straightforward ways is to interpret the prediction results as a probability distribution. In other words, have the model evaluate the probability of all possible pairs of points in the sequence that will serve as peak start / end markers. For a sequence of N points, this includes a total of N×(N-1) / 2 pairs after imposing order constraints. The probability of each pair can be easily calculated by multiplying the predicted values ​​of the peak start / end tasks by the two points. The probabilities are then used to calculate a weighted sum over the integrals between the corresponding pairs, providing an average estimate instead of a MAP estimate (this alternative method for calculating abundances is hereafter referred to as the "weighted sum method"). The MAP estimate calculates a single abundance value using the maximum point set value as the endpoint.

[0103] The abundance average may provide a more robust abundance estimate than using the maximum point set boundary, especially when a model architecture more similar to Model A than Model B is used. Because no explicit boundary is used, the method may avoid the problem of peak end predictions appearing before (or very close to) the peak start prediction. A comparison of abundance correlations for Model A and Model B is presented in Figures 11A and 11B, respectively.

[0104] Especially for model A, the weighted sum method helped to generate more accurate abundances and increase the abundance correlation from 0.9897 to 0.9969. The increase for model B was relatively small (0.9974 to 0.9977). The post-processing step mainly helped to avoid outliers whose maximum point set boundaries produced significantly lower abundances than the human annotations. These samples usually suffer from anomalously high boundary probabilities within the range of the peak (at the valley or shoulder point) that is then selected by the maximum point set operation. At the same time, with the weighted sum method, the overall prediction was biased upwards for low abundance peaks. This is mainly due to noisy signals and predictions. When the signal-to-noise ratio is low, the predicted boundary probabilities would fluctuate in a longer retention time range around the peak (Figure 7A and Figure 7B, and Figure 12), leading to a higher weighted sum. We considered this effect acceptable, since noisy peaks are less valuable for downstream applications.

[0105] From these rules, the confidence value was calculated as follows: For prediction on a sample curve with N points, calculate all the local maxima in the range. The maxima are labeled as

[0106]

number

[0107]

number

[0108] (If a maximum exists) the maximum to the left of it is

[0109]

number

[0110]

number

[0111]

number

[0112]

number

[0113]

number

[0114]

number

[0115]

number

[0116] 3.4 Prediction Reliability Another core advantage of the pointwise prediction structure is the ease with which it can estimate prediction certainty. Although the internal methods for quantifying the uncertainty of neural network predictions are still complex and fraught with limitations, for this task it is fairly straightforward to infer certainty / uncertainty by inspecting the model predictions against bounds. Most predicted probabilities take the form of Gaussian mixtures, whose quantity and height represent the model certainty. In the study, a heuristic method was adopted to quantify prediction confidence based on the following guidelines: Lower predicted value -> lower certainty Multiple maxima -> lower certainty Maximal values ​​spread out over items -> lower certainty

[0117] From these rules, a confidence value was calculated. Figure 12 provides an illustration along with an explanation of the calculations used to calculate the confidence level. The confidence score is proportional to the shaded area. We require the predicted value to be high (close to 1) and unimodal, or closed if multiple maxima appear.

[0118] Confidence scores were calculated for all predictions of Model A and Model B (Figures 13A and 14A). A trend in the change in performance can be seen across different confidence cutoffs. As shown in Figures 13B and 14B, the prediction error at the boundary steadily decreases with increasing confidence, dropping to nearly 1 second for the most confident samples.

[0119] Abundance correlations at different confidence intervals are shown in Figures 13A-F. Predictions in high abundance samples usually have higher certainty and better agreement with human annotations. Predictions in the top two bins achieved Pearson's r above 0.999. In contrast, samples with lower prediction confidence produced worse performance, with higher marginal mean absolute error (MAE) and lower abundance correlations. This is due, in part, to the poorer signal-to-noise ratio of the input samples indicated by low abundance. Overall, the certainty measure served as a good indicator of noise scale and model performance.

[0120] This post-processing step may offer versatility in performing workflow quality control. In real-world applications, a straightforward use would be to set a threshold confidence score below which predictions are filtered out or presented to a human expert for reassessment. This would also be useful for detecting anomalies in experimental data.

[0121] 5.0 Experimentation Experimental parameters Chemicals and Reagents: Human serum, dithiothreitol (DTT), and iodoacetamide (IAA) were purchased from Sigma-Aldrich (St. Louis, MO). Sequencing grade trypsin was purchased from Promega (Madison, WI). Acetonitrile (LC-MS grade) was purchased from Honeywell (Muskegon, MI). Sample preparation: Serum samples were reduced with DTT, alkylated with IAA, and then digested with trypsin in a water bath for 18 h at 37° C. After incubation to cool the digestion, formic acid was added to each sample to a final concentration of 1% (v / v).

[0122] 5.1 LC-MS analysis The digested serum samples were injected into an Agilent 1290 Infinity UHPLC system and an Agilent 6490 triple quadrupole mass spectrometer equipped with an Agilent ZORBAX Eclipse Plus C18 column (2.1 mm x 150 mm ID, 1.8 um particle size). Separation of peptides and glycopeptides was performed using a 70 min binary gradient. Aqueous mobile phase A was 3% acetonitrile, 0.1% formic acid in water (v / v), and organic mobile phase B was 90% acetonitrile, 0.1% formic acid in water (v / v). The flow rate was set at 0.5 mL / min. Electrospray ionization (ESI) was used as the ionization source and was operated in positive ion mode. The triple quadrupole was operated in dynamic multiple reaction monitoring (dMRM) mode. The peptide and glycopeptide transitions used in training and testing were a selection of those published in Li et al., Site-Specific Glycsylation Quantification of 50 serum Glycoproteins Enhanced by Predictive Glcopeptidomics for Improved Disease Biomarker Discovery, Anal. Chem. 2019, 91, 5433-5445; DOI: 10.1021 / acs.analchem.9b00776, the entire contents of which are incorporated herein by reference in their entirety for all purposes.

[0123] 5.2 Preparing the Dataset The training and validation set peaks, as well as the test set peaks, were collected from two sets of mass spectrometry experiments. Set 1 consisted of human serum samples and transitions measured in Carlito Lebrilla's laboratory at UC Davis in late 2017, while set 2 was commercial serum from ovarian cancer and benign mass patients purchased from Indivumed GmbH in mid-2018. Both sets of samples underwent the experimental protocol outlined above. Post-run Agilient.D format files were converted to mzML via msConvert in the Proteowizard 3.0 software suite for bioinformatics processing.

[0124] To obtain human annotations, all transitions were labeled by a mass spectrometer expert for one reference serum sample (pooled Sigma serum), and these labels constituted the "reference" peak. For the remaining samples, annotators labeled peak start and peak end retention time marks with Agilent's MassHunter Qunatitative Analysis B5.0 based on the reference peak and taking into account any shifts in retention time.

[0125] From the two experiments, only peaks labeled with positive abundance were selected. Samples with large retention time shifts (>0.2 min) were also excluded. A total of 106355 (training / validation) and 67672 (test) peaks passed the sanity test and were used in this study. For each transition, we collected all signals within the retention time window from the reference retention time start -0.2 min to the reference time stop +0.2 min, allowing a mass-to-charge ratio window of + / -0.1 around the desired precursor and product m / z. Signals were summed along the mass-to-charge ratio window to generate extracted ion chromatograms (XICs). Both the sequential model and the human annotators were presented with XIC curves that were trimmed around the target transition (Figures 7A, 7B, and 9B show some example XIC inputs).

[0126] The test set used in the above section entitled "Comparison with human annotators" contained 1619 peaks. All peaks were further annotated independently by 12 annotators for consistency, but only positive abundance annotations were used to calculate the relative standard deviation per transition and pairwise discrepancy between annotators.

[0127] 5.3 Featurization The input was formulated as a sequence of points on the intensity-retention time plane, with no fixed start / stop or length. To generate a uniform representation, featurization was applied on each sample curve. The retention times of all points within a single sample were first centered at the apex position of the sample curve, then expanded by 128 equally spaced Gaussian bins, spanning from -1 min to 1 min, to produce a vector of 256 real numbers, F, representing the time points. RT was generated.

[0128] Since the mass-to-charge intensity values ​​have a wide range, the intensity values ​​are similarly discretized by applying a 256 Gaussian function to the range from 0 to 500, and the analog values ​​are converted into a vector of 256 real numbers representing the intensity, I RTUsing the assumption that any point with intensity greater than an upper threshold must be considered part of a peak, only those points located in the noise stage or in the noise-signal transition stage were featured. RT and I RT The complete input X (i.e., the input parameters or states to the RNN model) for each sample curve after forming was formed into a matrix of size N × 384. The variable N ranged from 50 to 300 and represented the length of the curve.

[0129] For Model B, each sample was paired with a reference. The input pair has the same featurization process as above, resulting in two matrices X and X of dimensions N × 384 and N′ × 384. ref Although N and N' must be close, they are not necessarily equal to each other.

[0130] The sample labeling was formulated as the same continuous structure where the human annotations of peak start / end have a value of 1 in the corresponding tasks. To enhance stability and robustness, further label smoothing was applied.

[0131]

number

[0132] 5.4 Training Setup Referring again to Figures 5A and 5B (Model A) and Figures 6A and 6B (Model B), Model A and Model B were implemented in Pytorch.

[0133] Model A contains two bi-LSTM layers with 128 features for the hidden state, and the LSTM output is then applied to a multi-head attention layer with 8 heads and 32 features for each head. The outputs from the heads are concatenated and passed through a fully connected hidden layer with 32 units, then mapped to two tasks, each with two classes, through an affine layer. A softmax is applied to each task and each point in the sequence to generate peak start / end probabilities. The model is trained by cross entropy loss between the smoothed labels and the softmax predictions on a training subset containing 86311 samples for 20 epochs with a batch size of 8 and a learning rate of 0.001. Early stopping is applied based on the performance on the validation set.

[0134] Model B contained two sets of sequential network structures identical to those above, except for the fully connected layers used to encode the samples and the references, respectively. Weights are not shared between the two structures. The sample-reference attention layer is applied with the sample encodings (as queries), the reference encodings (as keys), and the reference labels (as values), and its output is the same length as the sample encodings and 2 in the task dimension. This is used directly as the model prediction for peak start / end probability. Another loss term is computed as the KL divergence between the sample-reference attention map and a standard mapping that connects points with the same retention times of samples and references (similar to an identity matrix). The model is trained by cross-entropy loss with the same settings plus an additional divergence term on the same training set.

[0135] Additional definitions As used herein, "processor" should be interpreted to include one or more processors used to perform tasks such as executing SAS program code, processing data for a GUI, or reading and writing from memory, and includes general purpose computers.

[0136] As used herein, "user" should be interpreted to include an actual person who uses the application locally or accesses the application remotely using a web browser using a remotely located computer.

[0137] As used herein, a "computer" should be interpreted to include a computer having a hard disk drive for reading from and writing to a hard disk, a magnetic disk drive for reading from or writing to a removable magnetic disk, and an optical disk drive for reading from or writing to a removable optical disk, such as a CD ROM, DVD, or other optical media. The hard disk drive, magnetic disk drive, and optical disk drive are connected to the system bus by a hard disk drive interface, a magnetic disk drive interface, and an optical drive interface, respectively. The drives and their associated computer-readable media provide non-volatile (non-transient) storage of computer-readable instructions, data structures, programs, and other data for the computer system.

[0138] As used herein, a "database" should be interpreted to include a centralized database or a distributed database, including an SQL database.

[0139] As used herein, the term "data" should be interpreted to include any information contained in the machine-readable memory of a file system or used in a database, or other structured data, or unstructured data representations in temporary or non-transient memory.

[0140] As used herein, "input device" should be interpreted to include a keyboard, mouse, pointer, touchpad, voice recognition device, or other form of input device connected to a computer (with or without a display) for modifying, searching, accessing, or creating electronic content that may be stored in memory.

[0141] As used herein, a "display" is an LCD / LED type display or other type of display device that is connected to a computer and is capable of rendering and receiving input using a GUI or displaying other information generated by applications running on the computer executing machine-executable code.

[0142] As used herein, a "module" or "component" should be interpreted to include a software or source code file, or a set of machine-executable instructions that may or may not be included in a source code file and reside in a non-transitory or transitory memory medium. A module may represent a program that exists as a stand-alone program or that is called by another program. A "component" is a part of a "module" in the sense that the module calls the component to perform a task or part of the tasks of the module.

[0143] As used and interpreted, "data repository" or "repository" may include a database, a data store, a data warehouse, a data lake, or any combination thereof.

[0144] As used herein, "memory" should be interpreted to include either machine-readable media, transient computer-readable storage media, or non-transient computer-readable storage media. In various embodiments, the volatile (transient) portion of the memory may be implemented using any suitable memory technology, such as static random access memory (SRAM), synchronous dynamic RAM, or any other type of memory. Although the illustrated and described memory examples may refer to a single machine-readable medium, the memory should be interpreted to include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers, as well as various storage media such as processor registers, memories, and storage devices) that store one or more sets of machine-executable code or data. Additionally, the term "machine-readable medium" should be interpreted to include any medium that can store, encode, or carry a set of instructions for execution by a machine, cause a machine to perform any one or more of the methods herein, or store, encode, or carry data structures utilized by or associated with such set of instructions. The term memory shall be interpreted accordingly to include tangible media, such as solid-state memory, and optical and magnetic media.

[0145] As used herein, a "module" should be construed to include units of different functions that can provide information to and receive information from other modules. Thus, the described modules may be considered to be communicatively coupled. Modules may also initiate communication with input or output devices and act on resources (e.g., collections of information, such as databases). Modules may include hardware circuits, optical components, single or multi-processor circuits, memory circuits, software program modules and objects, and combinations thereof, as necessary for the particular implementation of various embodiments. The term "module" may include identifiable portions of code, data, or computational objects to accomplish a particular function, operation, process, or procedure.

[0146] In some examples, methods for diagnosing a disease are described herein that include quantifying mass spectrometry transitions, where the transitions are peaks of m / z intensity values ​​characteristic of a glycopeptide or peptide that includes an amino acid sequence, and analyzing the quantification using a model to determine a probability that the sample is within or not within a disease classification based on the quantification. In some of these examples, the transitions, glycopeptides, peptides, amino acid sequences, or any combination thereof may be those described in International PCT Patent Application No. PCT / US2020 / 0162861, filed January 31, 2020.

[0147] The above description of illustrated embodiments of the invention, including what is described in the Abstract, is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Although specific embodiments and examples of the invention have been described herein for illustrative purposes, various modifications are contemplated within the scope of the invention, as those skilled in the relevant art will recognize.

[0148] These modifications can be made to the invention in light of the above detailed description. The terms used in the claims should not be construed to limit the invention to the specific embodiments disclosed in the specification.

Claims

1. 1. A method of training a model to identify mass spectrometry peaks, comprising: using a plurality of labeled sequential data, each of the plurality of labeled sequential data including a peak start label and a peak stop label representing a retention time value characteristic of a glycopeptide or peptide, or a fragment thereof; and training at least one machine learning model using the labeled sequential data, wherein the trained model is adapted to identify mass spectrometry start and stop peaks that indicate the presence of one or more glycopeptides or peptides, or fragments thereof, present in a biological sample.

2. 2. The method of claim 1, further comprising quantifying the presence of the one or more glycopeptides or peptides from the biological sample by integration over time between a start peak and a stop peak of the mass spectrometry.

3. The method of claim 1 , wherein the machine learning model is a recurrent neural network model.

4. 4. The method of claim 3, wherein the recurrent neural network is a bidirectional LSTM recurrent neural network model with an attention layer.

5. 2. A method comprising calculating abundances for glycopeptides or peptides using a machine learning model trained by the process of claim 1.

6. 1. A system comprising: a data repository of labeled sequential data; one or more processors coupled to the repository; and machine executable code resident in non-transitory memory accessible to said one or more processors which, when executed by said one or more processors, performs the method of any one of claims 1 to 5.

7. 1. A method for identifying a mass spectrometry peak, comprising: providing at least one trained neural network model that generates an output as a function of input parameters, the values ​​of the input parameters including a sequence of retention time values ​​of glycopeptides having m / z (mass to charge) ratios; providing mass spectrometry (MS) data values ​​of said input parameters including mass-to-charge (m / z) ratios across said sequence of retention time values ​​from a mass spectrometry analysis of a biological sample; The method, wherein the output includes the location of one or more peak start retention time values ​​and peak stop retention time values ​​indicative of the presence of one or more analytes or fragments thereof in the biological sample.

8. The method of claim 7, further comprising quantifying the presence of the one or more glycopeptides or peptides by summing or integrating retention time values ​​between a start peak and a stop peak of the mass spectrometry from the biological sample.

9. The method of claim 8 , wherein the step of quantifying comprises applying a weighted smoothing function to the predicted start and stop peaks.

10. 1. A system comprising: a data repository for mass spectrometry (MS) data; one or more processors coupled to the repository; and machine executable code resident in a non-transitory memory accessible to said one or more processors which, when executed by said one or more processors, performs the method of any of claims 7 to 9.

11. 1. A method comprising: Featurizing data from raw MS data, the raw MS data comprising a plurality of analog samples each representing a mass to charge intensity, the featurizing comprising: centering the sample within a retention time window; discretizing the samples into a sequence of points representing intensity values; normalizing the intensity values; assigning labels to points in the sequence of points corresponding to peak start times and peak stop times to generate a sequence of labeled points; The method, wherein the sequence of labeled points is configured to train a machine learning model to predict abundance in unseen MS data.

12. The method of claim 11 , wherein the centering comprises placing an apex of a curve at a center of the retention window.

13. The method of claim 11 , wherein the normalization of the intensity values ​​excludes intensity values ​​that are greater than a threshold amount.

14. 14. The method of claim 13, wherein the threshold amount is about 500.

15. 1. A system comprising: a data repository of raw mass spectrometry (MS) data comprising a plurality of analog samples; one or more processors coupled to the repository; and machine executable code resident in non-transitory memory accessible to said one or more processors which, when executed by said one or more processors, performs the method of any of claims 11 to 14.

16. 1. A method comprising: calculating a distribution of probabilities over a sequence of points representing mass spectrometry (MS) data for a given m / z ratio and retention time, the probabilities representing likely peak start retention times or peak stop retention times; and calculating an abundance statistic using the distribution of probabilities, the statistic indicating a presence of a biological sample in the MS data.

17. 17. The method of claim 16, wherein the statistic is an average abundance calculated as a weighted sum of integrals using all possible pairs of start and end points to evaluate the integral.

18. 12. The method of claim 1, 7, or 11, wherein the machine learning model is a sequential model (Model A) or a criterion-based model (Model B).

19. 1. A method comprising: Using multiple labeled sequential data; and training at least one machine learning model using the labeled sequential data, wherein the trained model is adapted to identify periods or frequency ranges in which start and stop peaks in intensity occur in unseen samples.

20. The method of claim 19 , wherein the start and stop peaks are included in a noise-to-signal transition period.

21. 20. The method of claim 19, wherein the machine learning model is a recurrent neural network with an attention layer.

22. 20. The method of claim 19, wherein the samples represent a physical entity such that an addition or integration of the intensity over time or frequency from the start peak to the stop peak quantifies the physical entity.

23. 20. The method of claim 19, wherein the sample is a continuous series of mass-to-charge intensities over a specified time period, and the sum or integral represents the abundance of the physical entity.

24. 1. A method comprising: providing at least one trained neural network model that generates an output as a function of input parameters, the values ​​of the input parameters being a time series or a spectral series of intensity values ​​representing a quantity; providing a value for each of the input parameters, the value representing a sequence of non-visible intensities of quantities; The method, wherein the output includes one or more of a time or frequency of a peak start and a time or frequency of a peak stop, which are used to quantify the quantity.

25. 25. The method of claim 24, wherein the quantity is quantified by addition or integration over the time or frequency between the time or frequency of the peak onset and the time or frequency of the peak off, respectively.

26. 23. The method of claim 24 or claim 22, wherein the intensities are obtained from a process selected from the set of mass spectrometry, MRM, liquid or gas chromatography, x-ray diffraction, Raman spectroscopy, UV-VIS spectroscopy, fluorimetry or phosphorimetry spectroscopy, nuclear magnetic resonance (NMR) spectroscopy, or magnetic resonance imaging (MRI).

Images