Improved method for identifying nucleotide modifications

Abstract

The present invention relates to identifying one or more abnormal monomers within a polymer. The present invention also relates to a method for preparing control data to be used in a method for identifying abnormal monomers within a polymer. The present invention also relates to an apparatus for carrying out said method, and a system comprising said apparatus and a nucleic acid sequencer.

Description

Technology Field

[0001] The present invention relates to a computer-implemented method for identifying one or more anomalous monomers in a polymer; and a computer program or computer storage medium comprising instructions for performing the method. The present invention also relates to an apparatus configured to perform the method; and a system comprising the apparatus and a nucleic acid sequencing device. Background Technology

[0002] Cells control their structure and function through complex mechanisms involving modifications to DNA, RNA, proteins, and other polymers. These modifications play a crucial role in many diseases.

[0003] Existing models for detecting polynucleotide modifications rely on prior knowledge of the modification of interest, which limits the identification of unknown modifications. However, it is believed that numerous modifications of these polymers exist that are currently unknown or impossible to observe in specific contexts.

[0004] Summary of the Invention

[0005] Methods currently used in the art focus on the detection of specific variations. The present invention detects all anomalies occurring relative to a baseline of normal observations, wherein the anomaly may be a variation not yet identified in context, or any observation result that differs from a canonical nucleotide call. The present invention enables the identification of the location of new and unknown variations present within a polymer.

[0006] In a first aspect, the present invention provides a computer-implemented method for identifying one or more abnormal monomers in a polymer, the method comprising the following steps:

[0007] a) a step of acquiring a stream of input data including measurement values ​​for a plurality of continuous monomers in a polymer;

[0008] b) a step of converting the stream of input data from step a) into multiple vector features;

[0009] c) Step to calculate the distance between (a) the following:

[0010] i) Each vector feature of step b); and

[0011] ii) Multiple vector features within the contrast data set,

[0012] Here, this control data set is generated by converting one or more streams of control data into one or more vector features to form a control data set, and this control data includes measurements for multiple continuous monomers within one or more polymers;

[0013] d) a step of determining, based on the distance calculated in step c), whether a monomer associated with one or more vector features within the polymer that is the input data is anomalous to the polymer of the control data.

[0014] In a further aspect, the present invention provides a computer-readable storage medium or a computer program comprising computer-executable instructions, and when the instructions are executed by a computing system, the computing system may perform the method.

[0015] In a further aspect, the present invention provides a computer-implemented method for preparing a control data set for use in a method for identifying one or more abnormal monomers in a polymer, the method comprising the following steps:

[0016] a) acquiring one or more streams of control data including measurements for a plurality of continuous monomers within one or more polymers; and

[0017] b) a step of converting one or more streams of this contrast data into one or more vector features to form a contrast data set, wherein the contrast data set is optionally normalized.

[0018] In a further aspect, the present invention provides an apparatus comprising a processing circuit configured to perform the method for identifying one or more abnormal monomers in a polymer.

[0019] In a further aspect, the present invention provides a system comprising the following:

[0020] a) nucleic acid sequencing device; and

[0021] b) A device of claim 19, wherein the device is operably connected to a sequencing device and configured to receive a stream of input data from the sequencing device. Brief explanation of the drawing

[0022] Figure 1 shows an oligonucleotide sequence. Figure 2 shows the electron current values ​​at positions 20 to 40 for three (color-coded) oligonucleotides from the control data set. For each position, a sub-sequence of five bases within the pore is indicated. Figure 3 illustrates the detection of RNA modifications in synthetic modified oligonucleotides. The Kolmogorov-Smirnov (KS) test is applied positionally to compare the distribution of anomalous scores in the modified dataset with a reference dataset without RNA modifications. P-values ​​(y-axis) are reported for each position (x-axis). Gray shaded areas indicate locations containing modified bases. Green dashed lines indicate locations where the modified base is at the center of the pore. P-values ​​highlighted in red correspond to local maxima (smaller peaks are removed until all consecutive peaks are separated by at least 9 p-values). Figure 4 illustrates the following: In the top plot, raw signals from contrast data of three different threads (or lead names) are shown. The bottom plot shows the signature projection of each thread by position, using UMAP for dimensionality reduction to visualize high-dimensional vectors in a single plot. Figure 5 illustrates three examples of the molecular-specific approach in the Oligos3 dataset. A curve is fitted to the correction score at a specific location to predict the score for the corresponding test molecule at that location. The resulting p-value (y-axis) is used to interpret the deformation at each location (x-axis). Blue bars represent non-deformed locations, and red bars represent locations where deformation is expected. A horizontal red line is added to emphasize that no non-deformed locations exceed this threshold. It is important to note that missing locations are common at the molecular level. Figure 6 shows the output scores of three different reads from the same region on chromosome 20 (locations 1,000,776 to 1,000,826). According to Nanopolish, only the read in 6a) (blue) is methylated. For the reads in plots 6b) and 6c), Nanopolish indicates no methylation. The gray shaded areas are locations where the CG base is inside the nanopore. FIG. 7 illustrates a flowchart of an exemplary method. The method (100) comprises, in block (102), obtaining a stream of input data containing measurements for a plurality of consecutive monomers in a polymer. In block (104), the stream of input data is converted into a plurality of vector features. In block (106), the distance between the vector features from the input data and the plurality of vector features in a comparison data set is calculated. Then, in block (108), one or more abnormal monomers are identified based on the calculated distance. This method may be implemented by the apparatus (200) described herein with reference to FIG. 8, but is not limited thereto. FIG. 8 shows a block diagram of an exemplary device having a control system (200). Figure 9 shows an additional flowchart of an exemplary method. Specific details for implementing the invention

[0023] Method for identifying one or more abnormal monomers within a polymer

[0024] This method is intended to identify one or more abnormal monomers within a polymer. In this context, an abnormality describes and is defined as any monomer not found in the control data. That is, an abnormality, which is the occurrence of an abnormal monomer, is defined relative to the composition of the polymer used in the control data and signifies a feature (which may be the presence or absence of a feature) that is not present in the control data. The control data may, in particular, contain variations, i.e., non-canonical monomers. Then, what is searched for in the input data is the abnormality compared to this control data—that is, the abnormality not present in the control data.

[0025] The above could be a covalent modification. For example, it could be an RNA / DNA modification.

[0026] data stream

[0027] A data stream is data that is emitted by a process or generated in another way, and can be measured or recorded in a continuous, discrete, or incremental manner.

[0028] Input data

[0029] The stream of input data includes measurement values ​​for multiple consecutive measurements within the polymer. The stream of input data may include electrical signals and / or optical signals. For example, it may be ionic current, impedance, or fluorescence.

[0030] The input may be, for example, a raw electrical signal acquired at various sampling rates, or the raw signal may be a signal that has undergone preprocessing such as normalization and / or segmentation, which is typically applied in base reading software.

[0031] This method may further include a preliminary step of sequencing a polymer (e.g., DNA or RNA) to provide a stream of input data or control data.

[0032] Contrast data

[0033] One or more streams of contrast data may be used. These streams are selected to represent the diversity of data that the device user considers non-anomalous.

[0034] These are converted into one or more vector features (groups) to form a contrast data set. The measurements for the contrast data may be synthetic. Synthesis means that the measurements are not taken directly using physical or actual laboratory equipment, but instead (e.g., in silico( in silico It means that it was generated by ). In other words, if the measurement value is a composite measurement value, this data is a measurement value derived synthetically.

[0035] Alternatively, the control data used may have been electrical and / or optical signals, as with the input data above. Where actual measurements are taken, the polymer used for the control data measurement may be a natural or synthetic polymer.

[0036] One or more polymers used in the control data may be standard polymers (standard or most common sequences). For example, if the polymer is RNA, the RNA polymer used in the control data may not contain modifications (or, as further explained below, may specifically include only modifications that are not of interest, as it explores additional unknown modifications within the polymer).

[0037] "Non-modified" means that there is no additional modification, that is, the monomer unit is a standard monomer. For example, in the case of DNA, the non-modified monomer units are adenosine, thymidine, guanosine, and cytidine. The modified form may be a nucleotide in which any one of the following covalent chemical modifications is present, for example, 5-methyl-cytosine (5mC), 5-hydroxymethyl-cytosine (5hmC), and 6-methyl-adenosine (6mA).

[0038] In the case of RNA, the non-modifying monomer units are adenosine, uracil, guanosine, and cytidine.

[0039] Alternatively, since the above is relative to the polymer within the control data, a polymer containing monomers with known modifications can be used as control data to find unknown RNA modifications. This method then compares the input data with this control data and locates the positions of monomers with new and unknown RNA modifications that are not present in the control data.

[0040] Known RNA modifications include the following:

[0041] N6-methyladenosine (m6A)

[0042] Inosine(I)

[0043] 5-methylcytosine (m5C)

[0044] Pseudouridine (Ψ)

[0045] N6-Dimethyladenosine(m6,2A)

[0046] 1-Methylguanosine (m1G)

[0047] 2'-O methyladenosine (2-OMeA), which is also encompassed by 2'-O-methylation (Nm).

[0048] 7-Methylguanosine (m7G)

[0049] 5-Methyluridine (m5U)

[0050] N1-methyladenosine (m1A)

[0051] N4-acetylcytosine (ac4C)

[0052] 2-Thiouridine(s2U)

[0053] Our deal

[0054] Adenosine to Inosine (A-to-I) RNA Editing

[0055] dihydrouridine(D)

[0056] 5-Formylcytidine(C)

[0057] 2'-O-methylation (Nm)

[0058] Queuosine (Q)

[0059] 5-methylaminomethyl-2-thiouridine(mnm5s2U)

[0060] Wybutosine (yW)

[0061] 3-Methylcytidine (m3C)

[0062] 5-Methoxycarbonylmethyl-2-thiouridine (mcm5s2U)

[0063] 2-Thiocitidine(s2C)

[0064] N2,N2-dimethylguanosine(m2,2G)

[0065] N6-Isopentenyladenosine(i6A)

[0066] N6-Threonylcarbamoyladenosine (t6A)

[0067] 3-Methyluridine (m3U)

[0068] yosin (imG)

[0069] 1-Methylinosine (m1I)

[0070] monomer / polymer

[0071] The polymer comprises a sequence of monomers. The polymer may be a biopolymer. For example, the polymer may be a polynucleotide, e.g., DNA or RNA (e.g., mRNA or tRNA). Alternatively, the polymer may be a protein in which the monomer units are amino acids. Or the polymer may be a polysaccharide.

[0072] A monomer is an individual unit that constitutes a polymer. The data stream is measured for multiple continuous monomers. "Continuous" means adjacent to or neighboring each other within the polymer.

[0073] Vector features

[0074] Input data and contrast data are transformed into vector features. This means converting a stream of measurements, for example, an electrical signal, into multiple vectors. Through this, the signal can be represented as a vector. The transformation is chosen so that the vector features provide concise, informative, and contextual information regarding the signal shape around monomer locations within the stream. This contextual information is expressed in the form of vectors or tensors and includes features extracted from non-parameterized measurements. In other words, vectorization tracks or follows the signal. Accordingly, the present invention uses features extracted from signature transformations or other differential equations controlled by the signal. That is, the vector features reflect the shape of the curve or stream derived from the input data. Features derived from the shape of the stream are features derived from non-parameterized curves and are (generally) independent of any sampling of the curve. Vector features include systematic features derived from the local shape of the stream of input data. Therefore, the signature transform is a signature-based transform, and thus describes using the entire signature transform or the part of the signature transform most useful regarding features extracted from the signature transform—that is, what constitutes 'normality'—and applying a distance to it. These can be extracted from the signal, and the distance between these vectors can be calculated.

[0075] Therefore, the transformation step involves a signature-based transformation, rather than the use of other features from the signal, such as simple low-dimensional summaries like dwell time or median signal strength. A signature transformation is applied to a stream of input data and one or more streams of contrast data to form a signature-based transformation that extracts multiple vector features from the input data and contrast data, respectively, for example, a specific feature from the signature-based transformation. Then, the extracted vector features can be compared between the input data and the contrast data by calculating the distance between the input data and the contrast data.

[0076] Being unparameterized means that any two streams related by reparameterization produce the same vector feature. Being related by reparameterization means that each stream contains exactly the same set of values ​​in exactly the same order, but possibly asynchronously and possibly at different rates, and that one stream can be represented as the other stream by composition of the reparameterization function with the other stream.

[0077] Input data and contrast data are converted into vector features in a consistent manner. For example, one post-processing step typically applied to raw nanopore current signals consists of dividing the raw data series into overlapping 5-mer-corresponding signal segments and converting such segments into vector features. The resulting segmentation is illustrated by the vertical dashed line in Fig. 2. The windows in this segmentation may be referred to as positions.

[0078] distance

[0079] The distance is calculated by comparing each vector feature generated from the input data with a vector in the control data set. That is, it is calculated by comparing the location of a feature from the input data with a set of features from the control data set associated with it. If a measurement in the input data set is very different from a specific measurement in the control data set, the distance may be large for some of these comparisons. However, unless there is an anomaly, there will be some data in the control data set represented by vector features very similar to the vector features of the input data. Meanwhile, if an anomaly exists at a specific point in the polymer, the vector feature transformed from this data is likely to be significantly different from any vector feature in the control data set, and the difference calculated in claim 1(c) will be greater than the noise of natural variability calculated, for example, by thresholds. As a result, the monomer at this location, for example, a nucleotide, may be classified as an anomalous monomer, for example, an RNA monomer having a variation not present in the control data set. This can be accomplished by calculating the minimum distance (also referred to herein as a score) for the vector feature.

[0080] This distance can be described using the Mahalanobis distance (also referred to as the variance norm) for each element in the corpus (the distance from each vector generated from the input data to each vector in the contrast data set). This distance is invariant with respect to the scale of the underlying stream.

[0081] The monomer against the polymer in the control data Whether or not it is strange verdict

[0082] As explained above, the distance calculated in step c) of claim 1 can be used as a measure of the degree to which a vector in the input data corresponds to a corpus of normality calculated from the control data (represented by vector features obtained from the control data).

[0083] A correction data set may be used to correct the contrast data set. This may be a separate set from different contrast data or a subset of the contrast data. If a subset of the contrast data is used, it is extracted (i.e., partitioned) from the contrast data before the distance function is calculated.

[0084] Correction data can be used to test how accurately the vector features of the control dataset represent a normal corpus. Specifically, it calculates the density of the vectors in the control data with respect to distance and sets a threshold between normal and anomalous vector features.

[0085] The determination in step d) may include checking whether the calculated distance exceeds a threshold for anomalies, where the threshold for anomalies was determined by the following:

[0086] i) A step of obtaining correction data representing the control data;

[0087] ii) a step ensuring that the distance function of step c) is independent of the correction data;

[0088] iii) a step of converting one or more streams of the correction data into one or more vector features;

[0089] iv) a step of calculating the distance between each vector feature of step iii) and a plurality of vector features within the contrast data set;

[0090] v) a step of setting a threshold for anomalies in the input data using the distance calculated in step iv); and

[0091] vi) A step of applying the threshold obtained from step v) to the input data,

[0092] Optionally, the correction data here was extracted from the control data.

[0093] Correction can be performed by vectorizing the correction data in the same way as the input data, that is, by the following:

[0094] a) a step of converting one or more streams of correction data into one or more vector features;

[0095] Next, the step of calculating the distance between the following:

[0096] b) Each vector feature of step a) above; and

[0097] Multiple vector features within the contrast data set,

[0098] Here, the control data set is generated by converting one or more streams of control data into one or more vector features to form a control data set, and the control data includes measurements for multiple continuous monomers within one or more polymers.

[0099] For each vector feature formed from the calibration data, a score can be determined (in the same way as was done for the vectors in the input data). This score may be the minimum distance for that vector feature. Once determined, the set of scores across all points within the calibration data can be used to estimate the probability that, for a selected numerical value, a vector sampled from the calibration data will have a score exceeding this value, and

[0100] This can be performed by any one of the following methods:

[0101] - Calculate the aforementioned probability by fitting the parameter distribution to the calibration score and evaluating the complementary cumulative distribution function at the selected numerical value.

[0102] - Calculate (1-p)(probability) such that the empirical p-quantile of the correction score becomes the input score. This is the same as the first method, but the empirical cumulative distribution function is used (3).

[0103] In summary, calibration is achieved by performing the same operations on the calibration data set as on the input data. This generates a set of distances from the control data to the calibration data. Analysis of the distribution of these distances can indicate how stationary the control data set is. From this distribution, the probability that any distance is anomaly can be estimated.

[0104] Next, these normalized p-values ​​can be used to determine how abnormal or exceptional the scores for the input data stream are. Optionally, the p-values ​​may also be corrected for multiple tests using any method known in the field, for example, the false discovery rate.

[0105] Accordingly, step v) of claim 11 may further include: calculating a distribution of scores for calibration data. Then, using this distribution, calculating a probability estimate for scores (i.e., distances, e.g., minimum distances) for input data. Then, using these probability estimates, setting a threshold for anomalies and comparing input data distances (e.g., scores) with these thresholds.

[0106] This correction step allows the match score (the distance calculated from the correction set to the control data set) to be converted into a probability estimate. While scores across different segments may not be directly comparable (for example, a score of 0.05 may be extreme for one segment's correction set but not for another), the probability estimates can be compared.

[0107] This threshold determination can be predetermined, that is, it may not be calculated simultaneously with performing the method of claim 1. A comparison data set can also be predetermined. Accordingly, the method of claim 1 compares input data with these predetermined values ​​in the software.

[0108] The determination of a threshold (i.e., whether any vector feature has a value exceeding a threshold for anomalies) can be aggregated across one or more consecutive monomers.

[0109] Therefore, the output from this method is data indicating the location of one or more ideal monomers within the polymer.

[0110] Normalization of input data

[0111] Input data may be collected from a source different from the control data, for example, from a sequencing device. Therefore, the vector features of the input data can be normalized to have the same scale as the vector features from the control data.

[0112] nanopores

[0113] Nanopore technology utilizes nanopores embedded in a membrane. The movement of polymers through the nanopores can be used to provide continuous measurements that form a data set.

[0114] The membrane may be electrically resistive, and a current may be applied across the nanopores, and changes in electrical signals (e.g., current) from the polymer passing through the nanopores may be detected. Thus, the data output by nanopore sequencing data is an electrical signal that changes over time as different monomer units pass through the pores. Suitable nanopore systems include those described in Mackenzie and Argyropoulos (Micromachines 2023, 14, 459).

[0115] Preparation of a control data set to be used in a method for identifying abnormal monomers within a polymer

[0116] The features described above for Claim 1 apply equally to Claim 15, namely, the preparation of a control data set. For example, the control data set may be normalized and corrected as described above, i.e., by the following:

[0117] i) A step of obtaining correction data representing the control data;

[0118] ii) a step ensuring that the distance function of step c) is independent of the correction data;

[0119] iii) a step of converting one or more streams of the correction data into one or more vector features;

[0120] iv) a step of calculating the distance between each vector feature of step iii) and a plurality of vector features within the contrast data set; and

[0121] v) A step of setting a threshold for anomalies using the distance calculated in step iv).

[0122] computer programs and Non-temporary media

[0123] A computer program refers to machine-readable program instructions. These can be provided on ephemeral media, such as transmission media, or non-ephemeral media, such as storage media. These machine-readable instructions (computer program code) can be implemented in high-level procedural or object-oriented programming languages. However, if desired, the program may also be implemented in assembly or machine language. In either case, the language can be a compiled or interpreted language and can be combined with a hardware implementation. Program instructions can be executed on a single processor or in a distributed manner across two or more processors.

[0124] Accordingly, one or more non-transient computer-readable media, such as those exemplified in FIG. 7, are also included, storing machine-readable instructions that, when executed, cause one or more processors to perform any one of the methods of claims 1 through 13. Further description of a general process according to the present invention can also be found in FIG. 9.

[0125] device

[0126] The processing circuit may be a general-purpose processor circuit configured by program code to perform a specific processing function. The circuit may also be configured by modifications to the processing hardware. The configuration of the circuit to perform a specific function may be limited to hardware only, software only, or a combination of hardware modifications and software execution. Program instructions may be used to configure the logic gates of a general-purpose or special-purpose processor circuit to perform a processing function. The processing circuit is described further below.

[0127] This device includes a control system (200) (Fig. 8). The control system includes one or more processors (202) collectively configured to perform a method for identifying one or more abnormal monomers within a polymer. That is, it is for the following:

[0128] a) receiving a stream of input data including measurement values ​​for a plurality of continuous monomers in the polymer (204);

[0129] b) Convert the stream of input data from step a) into multiple vector features;

[0130] c) Calculate the distance between the following:

[0131] i) Each vector feature of step b); and

[0132] ii) Multiple vector features within the contrast data set,

[0133] Here, this control data set is generated by converting one or more streams of control data into one or more vector features to form a control data set, and this control data includes measurements for multiple continuous monomers within one or more polymers;

[0134] d) Based on the distance calculated in step c), determine whether a monomer associated with one or more vector features in the polymer of the input data is anomalous to the polymer of the control data, and output the determination (206).

[0135] The control system may include at least one electronic processor having an electrical input for receiving an input signal; and one or more controllers collectively comprising at least one memory device electrically coupled to the at least one electronic processor and having instructions stored therein; wherein the at least one electronic processor is configured to access the at least one memory device to execute the instructions. A stream of input data may be stored locally on a device or on a sequencer (within the system), for example, on a hard drive of either the device or the sequencer. Alternatively, input data for the method, device, and system may not be stored locally and may be stored, for example, in cloud storage.

[0136] System

[0137] The system includes the device described above and a nucleic acid sequencing device, for example, a nanopore sequencing device. The device is operably connected to a sequencer and configured to receive input data from the sequencer.

[0138] Throughout this specification, unless otherwise required by context, the terms 'comprise' or 'include', or variations such as 'comprises' or 'comprising', 'includes' or 'including', will be understood to imply that the method or kit includes the mentioned component or group of components but does not exclude any other component or group of components.

[0139] Each document, reference, patent application, or patent cited in the text is expressly incorporated by reference in its entirety into this specification, meaning that the reader should read and consider it as part of the text. The fact that documents, references, patent applications, or patents cited in the text are not repeated in the text is solely for the sake of brevity. References to cited materials or information contained in the text should not be construed as an acknowledgment that such materials or information were part of the general knowledge of the people or were known in any country.

[0140] Overview of the method

[0141] Prior to the examples, a more detailed description of this method is provided to illustrate one way of performing the above method. Any of the features of items 1 through 6 below may be combined with the general description above.

[0142] 1. Data Transformation: Sequencing data obtained from the polymer is transformed into an input data set containing multiple vectors. These vectors capture information reflecting the non-parameterized shape of the signal around each monomer within the polymer. The transformed vectors are expanded to provide concise and informative contextual information regarding changes in the shape of the polymer around each monomer. This contextual information is expressed in the form of vectors or tensors. This vectorization step is critical to this method because it effectively captures information reflecting, for example, the shape of each monomer within the polymer.

[0143] 2. Reference Data Set: We generate a reference data set known as the 'corpus of non-deformed feature data,' which represents the achievable output range by applying a data transformation process to non-deformed sequencing data. This reference data set can be generated through simulation or by transforming sequencing data from multiple non-deformed polymers.

[0144] 3. Variation Evaluation: Data from a single sequence or a set of sequences is compared with a corpus of non-variation feature data. This comparison can be performed at the level of individual vector features, specific positions, or monomers. This involves comparing individual vector features from the new sequence with equivalent vectors from the corpus of non-variation feature data. Considering the variability within the reference dataset (also referred to herein as the control dataset), an innovation or conformity score is determined for each position. Correction of this score is achievable.

[0145] 4. Calculation of Fit Scores: Using a corpus of non-deformable feature data in vector or tensor form, we establish a unique fit score function. This function evaluates new instances of vector features and provides a score that reflects the degree of innovation compared to the corpus. The corpus is divided into two parts: one for generating the score function, and the other to form a set of inliers used for score correction. Fit scores must be invariant to some extent to linear rescaling of features and changes in measurement units, and must depend on the proximity of the new vector feature to the corpus. For example, we can estimate the nearest neighbor distance from a new instance of a vector feature to the corpus of non-deformable feature data using the variance norm defined by the corpus.

[0146] 5. Scoring of positions: After identifying and correcting the matching scores for vector features at specific positions or monomers, whether from individual sequence data or aggregated data from multiple sequences, and by considering neighboring positions, we can indicate variations, groups of variations, or other segments of sequences where anomalies have occurred or are likely to occur in the case of multiple transformed sequences.

[0147] 6. Determination of Deformation: Information generated in the previous step regarding locations with a high probability of defects or anomalies is analyzed to determine the nature of the deformation as required for the specific application. This method identifies deformations based on the novelty of exemplary monomers and sequences within the reference dataset, thereby enabling the detection of previously unobserved complex deformations, such as deformation pairs, in addition to standard deformations. This step can be combined with contextual information to improve accuracy.

[0148] An additional overview of the process can be found in Fig. 9.

[0149] Examples

[0150] Now, aspects of the present invention will be described only by example and with reference to the following experiment.

[0151] Examples 1: Data preparation

[0152] data

[0153] We used publicly available data containing seven distinct RNA modifications in synthetic oligonucleotides. However, our method does not directly utilize the properties of these modifications and is applicable to any other type of sequencing data.

[0154] The designed oligonucleotides contain different variations across the same sequence.

[0155] Data sets were created from the oligonucleotides shown in Figure 1: the first includes Oligo1, which contains 2,324 different molecules with two m6A sites at distinct positions. The second data set, named Oligo2, includes I, m5C, and pseudoU(Ψ) variants with a total of 1,289 different molecules. The third data set, Oligo3, consists of 1,982 molecules with m62A, m1G, and 2'-OMeA variants. Additionally, a fourth data set, Oligo4, with unvaried sequences, was created as a control data set with 2,205 different molecules.

[0156] For each data set, namely Oligo1, Oligo2, Oligo3, and Oligo4 (control), multiple RNAs of this type of oligonucleotide are sequenced by Oxford Nanopore Technologies (ONT) direct sequencing. For each molecule, a univariate electrical signal is measured by a sensor while the oligonucleotide moves through the nanopore, as shown in Figure 2.

[0157] DRS electrical signals recorded discretely over time are in the form of a univariate time series (or stream). We The time series space for is represented as follows:

[0158]

[0159] Time series Given, we second It is called the length. A DRS dataset is a set of such streams, possibly having variable length.

[0160] Next, to vectorize these time series, we capture shape information using a signature transformation from rough path theory. As described below, we map time series of variable length, if possible, into fixed-dimensional vectors.

[0161] conversion

[0162] We, as explained in more detail below, each initial DRS stream second In stream We transform it into. In this way, we add a dimension (or dimensions) to the univariate data. This maps one data stream to another, allowing us to extract more information from the stream before applying the signature transformation, with the goal of extracting information relevant to the specific problem at hand. These transformations and combinations thereof can be used for optimization purposes.

[0163] These transformations offer various advantages. They can help reduce the dimensionality of time series, facilitate information extraction by preprocessing the time series before applying signature maps, and leverage the invariance of signatures to translation and / or reparameterization, which are described in more detail in references [1, 2]. While numerous pre-signature operations exist, we will provide a few examples. As mentioned earlier, we Let's assume we observe an arbitrary vector with increasing timestamps. Regarding this, we by the following time augmentation has:

[0164]

[0165] This basically adds a timestamp as an additional coordinate, and the length is n It is. of the d Dimensional stream of lead- lag conversion silver, length 2 n 2 having - 1 d A dimensional data stream is derived, which is defined as follows:

[0166] .

[0167] invisibility-reset transformation It adds a dimension as follows:

[0168] .

[0169] Vectorization

[0170] Rough path theory is a field of probabilistic analysis that provides a rigorous mathematical framework for modeling the interaction between streams and physical control systems. In particular, signature transformations are distinguished by their notable properties and serve as an effective feature mapping technique for capturing the non-parameterized shape of streamed data. An arbitrary continuous path of bounded variation , and arbitrary Regarding, The signature of is defined by the following, and

[0171] , (1)

[0172] In other words, it is an infinite set of (Riemann-Stieltjes) iterative integrals. k th Iterative integration is d k It can be viewed as a set of scalars. We use a dedicated, highly optimized Python library, stream The segment-by-segment linear interpolation of The first M of the signature (e.g., M ∈ Igo M We calculate the iterative integral for (≥1). We We obtain the scalar vector. We Calculating the signature of a segmented linear path having segments is Note that it has a time complexity of .

[0173] We select signatures due to their ability to capture the sequential order of events and effectively model their effects without the need for high-dimensional reconstruction of individual data points. The components of the signature form our feature set, enabling subsequent operations.

[0174] After illustrating the raw signals in Fig. 2, Fig. 4 shows the shape of the vectors after signature transformation for some of the raw signals. It is essential to note that since signatures generate high-dimensional vectors, dimensionality reduction is required for plotting. We use UMAP, which enables us to learn manifold embeddings that capture the signature features observed in the data. [https: / / umap-learn.readthedocs.io / en / latest / ] Figure 4 shows the corresponding projected representation of the signature at the bottom, along with a plot of the raw clean data for three molecules.

[0175] Example 2: Detection of alterations in polynucleotide data

[0176] In the previous section, we described a method for obtaining a convenient vector representation of the DRS signal. We explain how the time interval was selected for the experimental study.

[0177] Novelty score

[0178] Now, using a corpus of non-deformed feature data around the same location, we obtain vectorized DRS signals around the monomer location Explains how to assign fit scores by calculating the nearest neighbor Mahalanobis distance.

[0179] We can access non-transformed data such as the contrast data shown in Fig. 2 (also referred to above as the "reference data set"). We can access this data set Corpus It refers to this as... = It is represented as. Given a corpus, we have two types of statistics, namely sample mean class sample covariance matrix Calculate:

[0180]

[0181] input and corpus elements The Mahalanobis distance between them is defined as follows:

[0182] .

[0183] input Given, we calculate the Mahalanobis distance to every element of the corpus and use this set of distances to score It constitutes. Since the Mahalanobis distance includes a covariance matrix in its definition, it serves as an indicator that normalizes the distance based on the variability of the corpus. The fit score is defined by the nearest neighbor (NN) distance, i.e., as follows:

[0184]

[0185] Through this, we Data-based It comes to have a concept of distance.

[0186] The difference is more than Whether or not it indicates decision

[0187] Now, we explain how to construct a single-class classifier for anomaly detection—that is, how to determine whether a signal is associated with a deformed location given a score. To do this, we use a portion of the corpus of unused, undeformed data, and this calibration set It is referred to as and (similar to a corpus) represents the following:

[0188]

[0189] We score each element of the correction set We pre-calculate it. In this way, we define what is clean in our setup. The same method can include variations test set

[0190]

[0191] It applies to, and a score is calculated for each element of the test set below:

[0192]

[0193] Compilation results

[0194] Next, we analyze the correction and test score distributions by performing statistical tests to determine the differences between them, which are quantified by obtaining their respective p-values. This process involves comparing the two distributions and allows us to assess their similarity through rigorous statistical evaluation based on their distinct characteristics.

[0195] A set of pre-calculated correction scores and set of test scores Given, we perform a statistical test to compare their distributions. P( ) and P ( Let ) represent the probability distributions of the correction and test data sets, respectively. To evaluate the difference between these two distributions, we applied the Kolmogorov-Smirnov 2-sample test. This test evaluates the null hypothesis that the two distributions are identical. The obtained p-value quantifies the level of similarity between the two distributions.

[0196] By molecule

[0197] We also (each individual score It possesses the ability to detect variations at the individual molecular level (based on...). Unlike aggregate results that involve comparing two distributions, our molecule-specific approach uses correction scores at specific locations ( It involves fitting a curve to the distribution. Subsequently, based on this fitting, we perform a prediction for a single molecule test score at the same location to derive a p-value. This process involves comparing the calibration score distribution with individual test scores, providing an understanding of the probability that the calibration score is greater than the test score in a single molecule.

[0198] result:

[0199] The aggregated results are shown in Figure 3.

[0200] The results for each molecule are shown in Figure 5.

[0201] The results of the distortion detection in the DRS signal are illustrated in Fig. 3. Our scoring function, which relies on the contrast data, is applied to all locations and leads within the dataset. Then, for each location, we compare the dataset score distribution (Oligo1, Oligo2, or Oligo3) with the correction score distribution.

[0202] The molecular-specific results for the DRS signals are shown in Figure 5, representing three different molecules from the Oligos3 dataset. This is an exemplary example of how we can sometimes correctly identify variations at the molecular level, but it is essential to note that this may not always be the case when considering a significant number of molecules exceeding 2,000 per dataset. This also highlights the robustness of our aggregated results.

[0203] Applying a scoring function by position means the following: One post-processing step typically applied to ONT DRS signals is the raw data series x = (X1, ..., X ℓ It consists of dividing ) into 5-mer-corresponding signal segments. The resulting segmentation is illustrated by the vertical dashed lines in Fig. 2. We [use] the window in this segmentation position It is called. When current is recorded across that window / location, about 5 bases are present within the pore. For this experimental study, we calculate the signature across this window and score the signal.

[0204] The highest p-values ​​are consistently aligned with positions containing at least one modified base. In particular, distinct peaks are spread across the sequence and are aligned with the initial modified position. Outside of these specific positions in the aggregate results, No peaks are observed. It is important to note that this represents an improvement compared to other methods that require precise knowledge of the type of variation being sought. This demonstrates the sensitivity of our score function to variations.

[0205] By performing statistical tests, we can identify and highlight regions where anomalies occur, marking them as potentially deformed. This can lead to a more detailed analysis to determine the specific type of deformity. At this point, we do not classify the anomalies according to the deformity, but merely determine their location.

[0206] We present a semi-supervised approach that accurately pinpoints the location of specific variants without requiring prior knowledge of them. As a result, this method has the potential to identify novel or previously unidentified variants.

[0207] Examples 3: This method works to identify abnormal monomers in various polymers

[0208] To demonstrate that the above technique works on various types of polymers, we used the Nanopolish tool for identification.

[0209] We Oxford Nanopore sequencing 4 We downloaded publicly available DNA data processed using [tool / method]. Specifically, we obtained the first dataset, FAB39088. The data was from Nanopolish 5 It was processed with a tool that includes a module called eventalign, which is widely used to align events or "squiggles" to a reference genome. Nanopore sequencing is sensitive to base modifications, and Nanopolish provides a step-by-step tutorial for detecting DNA methylation, one of those modifications. Methylation is relatively simple to detect because it typically occurs in CpG contexts (where a G follows a C base).

[0210] Using Nanopolish, we obtained labels for chromosome 20 covering positions 1,000,000 to 60,000,000, which represent the log-likelihood ratio for each read name and position, representing Nanopolish’s belief regarding the location within the read name where a nucleotide was methylated. Nanopolish identifies variants using a hidden Markov model instead and, unlike the use of signature-based transformations used in our method described in Examples 1 and 2, requires that the variant be known.

[0211] To validate our method, we identified regions where Nanopolish reported some lead names having methylated modifications and others being clean. We used regions 1,000,776 to 1,000,826 (shown in Fig. 6). Three lead names were identified across these regions, and methylation was detected in only one of them. We compiled a list of kmers occurring in these regions.

[0212] For each of these kmers, we considered all lead names and matching clean kmers within the interval of 1,000,000 to 60,000,000 that were not located in that short area.

[0213] We randomly selected 3,000 location-ready-name pairs from these clean locations. We used the associated nanopore data as our stationary corpus for a given kmer. Note that these kmer read-ready-name pairs were sampled from various locations outside the tested interval.

[0214] For the locations within the identified regions and our three lead names, we calculated the fit of the nanopore data against the clean data corpus associated with the corresponding kmer at each location.

[0215] result:

[0216] Unique variant locations (k-mers within individual read-lines) exhibited low concordance, while all other locations within the region, as well as potential variant locations within two clean read names, demonstrated concordance. This demonstrates the broad capability of the methodology to identify variants in a corpus beyond the RNA context.

[0217] The shaded area in Fig. 6 corresponds to the region where the CG base was inside the nanopore and indicates potential methylation. We observe high output scores in this shaded area for methylated lead names identified by Nanopolish, and low output scores across the region in the case of no methylation.

[0218] This approach confirms that our method can localize methylation patterns in DNA data.

[0219] References

[0220] [1] James Morrill, Adeline Fermanian, Patrick Kidger, and Terry Lyons. A generalized signature method for time series . arXiv preprint arXiv:2006.00873, 2020

[0221] [2] Thomas Cochrane , Peter Foster , Terry Lyons , Imanol Perez Arribas . Anomaly detection on streamed data . arXiv preprint arXiv:2006.03487, 2020

[0222] [3] Laxhammar, Rikard. Conformal anomaly detection: Detecting abnormal trajectories in surveillance applications . Diss. University of Skϕvde, 2014.

[0223] [4] https: / / github.com / nanopore-wgs-consortium / NA12878 / blob / master / Genome.md

[0224] [5] https: / / github.com / jts / nanopolish / tree / master

Claims

Claim 1 A computer-implemented method for identifying one or more anomalous monomers within a polymer, comprising the following steps: a) acquiring a stream of input data containing measurements for a plurality of continuous monomers within the polymer; b) converting the stream of input data of step a) into a plurality of vector features; c) calculating a distance between: i) each vector feature of step b); and ii) a plurality of vector features in a control data set, wherein the control data set is formed by converting one or more streams of control data into one or more vector features to form a control data set, and the control data contains measurements for a plurality of continuous monomers within one or more polymers; and d) determining, based on the distance calculated in step c), whether a monomer associated with one or more vector features in the input data polymer is anomalous with respect to the polymer of the control data. Claim 2 A computer-implemented method according to claim 1, wherein the measurement value for the comparison data is a synthetic measurement value and / or a non-synthetic measurement value. Claim 3 A computer implementation method according to claim 1 or 2, wherein the polymer is a biopolymer, optionally the biopolymer is a polynucleotide, and optionally the polynucleotide is DNA or RNA. Claim 4 A computer-implemented method according to any one of claims 1 to 3, wherein the method identifies a polymer comprising monomers having more than one type of anomalous modification, optionally wherein the anomalous modification is a covalent modification. Claim 5 A computer implementation method in paragraph 3, wherein the polynucleotide is RNA and the above is a covalent modification. Claim 6 A computer-implemented method according to any one of claims 1 to 5, wherein one or more non-defective polymers used to generate a control data set comprise: a) a polymer composed of an unmodified monomer; or b) a polymer comprising one or more monomers comprising any one or more of 5-methylcytosine, 5-hydroxymethyl-cytosine, inosine, pseudouridine, N6-dimethyladenosine, 1-methylguanosine, 2'-O methyladenosine, 7-methylguanosine, or N6-methyladenosine. Claim 7 A computer-implemented method according to any one of claims 1 to 6, wherein a stream of data is measured from the polymer during the movement of the polymer through the nanopore. Claim 8 A computer-implemented method according to any one of claims 1 to 7, wherein the measured value includes an electrical signal or an optical characteristic. Claim 9 A computer implementation method according to claim 8, wherein the input data is an ion current. Claim 10 A computer-implemented method according to any one of claims 1 to 9, wherein a vector feature obtained from input data is normalized to the same scale as a vector feature of a comparison data set. Claim 11 A computer-implemented method according to any one of claims 1 through 10, comprising determining whether the distance calculated in step d) exceeds a threshold for an anomaly, wherein the threshold for an anomaly is determined by: i) obtaining correction data representing the comparison data; ii) ensuring that the distance function of step c) is independent of the correction data; iii) converting one or more streams of the correction data into one or more vector features; iv) calculating the distance between each vector feature of step iii) and a plurality of vector features within the comparison data set; v) setting a threshold for an anomaly for the input data using the distance calculated in step iv); and vi) applying the threshold obtained from step v) to the input data, optionally wherein the correction data is extracted from the comparison data. Claim 12 A computer-implemented method according to claim 11, further comprising the step of aggregating threshold determinations associated with successive monomers to form a new threshold determination. Claim 13 A computer implementation method according to any one of claims 1 to 12, wherein the vectorization of the input data in step b) is consistent with the vectorization of the contrast data in step c). Claim 14 A computer-implemented method according to any one of the aforementioned claims, wherein steps b) and c) each extract vector features from a signature-based transformation of input data and / or contrast data. Claim 15 A computer-implemented method for identifying one or more abnormal monomers within a polymer, comprising the following steps: a) acquiring a stream of input data containing measurements for a plurality of consecutive monomers within the polymer, wherein the stream of input data is an ion current measured from the polymer during the movement of the polymer through a nanopore; b) converting the stream of input data of step a) into a plurality of vector features; c) calculating a distance between: i) each vector feature of step b); and ii) a plurality of vector features in a control data set, wherein the control data set is formed by converting one or more streams of control data into one or more vector features to form a control data set, wherein the control data includes measurements for a plurality of consecutive monomers within one or more polymers; steps b) and c) each include extracting vector features from signature-based conversions of the input data and / or control data; and d) determining, based on the distance calculated in step c), whether a monomer associated with one or more vector features in the polymer of the input data is abnormal with respect to the polymer of the control data, wherein the polymer is RNA or DNA. Claim 16 A computer-implemented method according to any one of the aforementioned claims, further comprising the step of outputting from this method data indicating the location of one or more ideal monomers. Claim 17 A computer-readable storage medium or computer program comprising computer-executable instructions, wherein when the instructions are executed by a computing system, the computing system causes the computing system to perform a method according to any one of the aforementioned claims. Claim 18 A computer-implemented method for preparing a control data set for use in a method for identifying one or more abnormal monomers in a polymer, comprising the following steps: a) acquiring one or more streams of control data comprising measurements for a plurality of continuous monomers in one or more polymers; and b) converting one or more streams of the control data into one or more vector features to form a control data set, wherein the control data set is optionally normalized. Claim 19 A device comprising a processing circuit configured to perform the method of any one of claims 1 to 16. Claim 20 A system comprising: a) a nucleic acid sequencing device; and b) a device of claim 19, wherein the device is operably connected to the sequencing device and configured to receive a stream of input data from the sequencing device.

Images