Method for constructing prediction model of nucleic acid chemical modification, prediction method and device

Abstract

The present application relates to the technical field of nucleic acid chemical modification detection, and discloses a nucleic acid chemical modification prediction model construction method, which comprises the following steps: obtaining a first nucleic acid sample, a second nucleic acid sample and a third nucleic acid sample of a nucleic acid sample prediction object, determining a first nucleic acid feature corresponding to the first nucleic acid sample, a second nucleic acid feature corresponding to the second nucleic acid sample and a third nucleic acid feature corresponding to the third nucleic acid sample, training based on the first nucleic acid feature, the second nucleic acid feature and the third nucleic acid feature, and obtaining a first prediction model and a second prediction model. The present application trains nucleic acid feature information through the first prediction model and the second prediction model in combination with a long read sequencing mode, obtains a prediction result of single-molecule level chemical modification and a prediction result of genome level chemical modification, can improve the prediction efficiency and prediction accuracy of nucleic acid chemical modification under low depth coverage, and reduces a large amount of labor and economic cost.

Description

Technical Field

[0001] This invention relates to the field of nucleic acid molecular chemical modification detection technology, specifically to a method for constructing a predictive model of nucleic acid chemical modification, a prediction method, and an apparatus. Background Technology

[0002] Covalent chemical modifications (such as methylation) of nucleic acid molecules such as deoxyribonucleic acid (DNA) or ribonucleic acid (RNA) are common epigenetic markers. Studies have shown that chemical modifications of DNA are important factors in regulating gene expression, controlling cell differentiation, and determining cell fate.

[0003] In related technologies, chemical reagents are typically used to transform or enrich nucleic acid samples, followed by high-throughput sequencing and specialized bioinformatics analysis to locate the sites and frequencies of chemical modifications. This involves numerous labor-intensive experimental steps and requires separate experiments, resulting in high detection costs. Furthermore, due to limitations in detection technology, these methods struggle to identify chemical modification sites over long distances with monomeric precision. Additionally, the enrichment efficacy of antibodies, the transformation efficiency of chemical reagents, and the fragmentation effect of chemical reagents on the original sample can also affect the accuracy of chemical modification detection results. Summary of the Invention

[0004] In view of this, the present invention provides a method for constructing a predictive model for nucleic acid chemical modifications, a prediction method, and an apparatus to solve the problems of related nucleic acid chemical modification identification methods, such as high labor costs, high experimental costs, low detection accuracy, and inability to identify chemical modification sites at monomeric accuracy.

[0005] In a first aspect, the present invention provides a method for constructing a predictive model for nucleic acid chemical modifications, the method comprising: The first, second, and third nucleic acid samples of the nucleic acid sample prediction object are obtained. The first nucleic acid sample is found to have undergone chemical modification through a first processing method, the second nucleic acid sample is found to have not undergone chemical modification through a second processing method, and the third nucleic acid sample is found to have undergone chemical modification based on a preset site through a third processing method. The first, second, and third nucleic acid samples are combined according to a preset sequencing template based on bioinformatics. The first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample are determined. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid sample prediction object recorded according to the preset sequencing depth in the long read sequencing method; The first and second prediction models are obtained by training based on the first, second and third nucleic acid features. The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variable of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

[0006] By implementing the above-described embodiments, this invention determines nucleic acid feature information based on a batch of first, second, and third nucleic acid samples using long-read sequencing. Then, by training the nucleic acid feature information using a first and second prediction model, it obtains prediction results for single-molecule-level chemical modifications and genome-level chemical modifications. This improves the prediction efficiency and accuracy of nucleic acid chemical modifications, reduces labor costs, eliminates the need for other experiments (such as pyrosequencing) on ​​the same nucleic acid sample, and lowers detection costs. Furthermore, it exhibits high accuracy under low-depth sequencing conditions.

[0007] In one optional implementation, obtaining the first, second, and third nucleic acid samples of the nucleic acid sample prediction target includes: Obtain various types of raw nucleic acid samples, including: non-wild-type DNA nucleic acid samples and / or wild-type DNA nucleic acid samples and / or wild-type RNA nucleic acid samples and / or non-wild-type RNA nucleic acid samples; Various types of raw nucleic acid samples were prepared based on a pre-defined preparation method; Nucleic acid sample sequencing libraries were created based on various types of prepared raw nucleic acid samples. Based on the nucleic acid sample sequencing library, the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample were obtained.

[0008] By implementing the above-described embodiments, this invention creates nucleic acid sample sequencing libraries from various types of original nucleic acid samples. For fragmented original nucleic acid samples, it is not necessary to create nucleic acid sample sequencing libraries for each sample individually. Therefore, it can reduce the cost of performing biological experiments on each original nucleic acid sample individually, thereby improving the enrichment efficiency of nucleic acid samples and the conversion efficiency of chemical reagents.

[0009] In one optional implementation, based on the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample, determining the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample includes: The first, second, and third nucleic acid samples were sequenced using a long-read sequencing method based on a preset sequencing depth. Feature extraction was performed on the first, second, and third nucleic acid samples after sequencing based on a biological sample sequencing platform. Based on the feature extraction results, the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample are obtained.

[0010] By implementing the above-described embodiments, this invention performs sequencing processing on the first, second, and third nucleic acid samples and extracts feature information based on long-read sequencing at a preset sequencing depth (high-depth or low-depth sequencing), which helps to improve the accuracy of sample features.

[0011] In one optional implementation, feature extraction is performed on the first, second, and third nucleic acid samples after sequencing processing using a biological sample sequencing platform, including: The first nucleic acid feature is determined based on the comparison results between the first nucleic acid sample and the first standard nucleic acid sample; the second nucleic acid feature is determined based on the comparison results between the second nucleic acid sample and the second standard nucleic acid sample; and the third nucleic acid feature is determined based on the comparison results between the third nucleic acid sample and the third standard nucleic acid sample.

[0012] By implementing the above-described embodiments, this invention compares the features of nucleic acid samples with those of a reference sample, thereby accurately obtaining specific information about the features of the nucleic acid samples.

[0013] In one alternative implementation, during the process of obtaining the trained first prediction model, the model parameters of the initial first prediction model are set based on the first input variable, and the initial first prediction model is gradually optimized based on the first output variable to obtain the trained first prediction model.

[0014] Through the above-described embodiments, this invention improves the accuracy of predicting nucleic acid chemical modifications at the single-molecule level by setting and optimizing the model parameters of the first prediction model.

[0015] In one alternative implementation, during the process of obtaining the trained second prediction model, the model parameters of the initial second prediction model are set based on the second input variable, and the initial second prediction model is gradually optimized based on the second output variable to obtain the trained second prediction model.

[0016] By implementing the above-described embodiments, this invention improves the accuracy of predicting nucleic acid chemical modifications at the genome level by setting and optimizing the model parameters of the second prediction model.

[0017] Secondly, the present invention provides a method for predicting nucleic acid chemical modifications, the method comprising: Obtain the nucleic acid sample to be predicted; Based on the nucleic acid sample to be predicted, the characteristics of the nucleic acid to be predicted corresponding to the nucleic acid sample are determined. The characteristics of the nucleic acid to be predicted include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to the preset sequencing depth in the long read sequencing method. The nucleic acid features to be predicted are input into the first prediction model and the second prediction model respectively for prediction, and the first prediction result output by the first prediction model is obtained, and the second prediction result output by the second prediction model is obtained; the first prediction model and the second prediction model are constructed by the nucleic acid chemical modification prediction model construction method in the first aspect or any embodiment of the first aspect; The probability of chemical modification of the nucleic acid sample to be predicted is determined based on the first prediction result, and the probability of chemical modification of the nucleic acid sample to be predicted is determined based on the second prediction result.

[0018] By implementing the above-described embodiments, this invention, based on a first prediction model (a single-molecule level chemical modification prediction model) and a second prediction model (a genome level chemical modification prediction model), can predict different types of nucleic acid samples to be predicted. This includes predicting nucleic acid samples with monomeric accuracy and predicting nucleic acid samples with monomeric accuracy, thereby improving the prediction efficiency and accuracy of nucleic acid chemical modifications.

[0019] In one optional implementation, obtaining the first prediction result corresponding to the nucleic acid chemical modification output by the first prediction model includes: When the first prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone nucleic acid chemical modification. When the first prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0020] By implementing the above methods, accurate prediction of nucleic acid samples to be predicted can be achieved based on single-molecule level nucleic acid molecules.

[0021] In one optional implementation, obtaining the second prediction result corresponding to the nucleic acid chemical modification output by the second prediction model includes: When the second prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone nucleic acid chemical modification. When the second prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0022] By implementing the above methods, accurate prediction of nucleic acid molecules at the genome level can be achieved for nucleic acid samples to be predicted.

[0023] According to a third aspect, the present invention provides an apparatus for constructing a predictive model of nucleic acid chemical modifications, the apparatus comprising: The prediction sample acquisition module is used to acquire the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample of the nucleic acid sample prediction object. The first nucleic acid sample is known to have undergone chemical modification through the first sequencing method, the second nucleic acid sample is known not to have undergone chemical modification through the second sequencing method, and the third nucleic acid sample is known to have undergone chemical modification based on a preset site through the third sequencing method. The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are combined according to a preset sequencing template based on bioinformatics. The nucleic acid feature determination module is used to determine the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include the attribute information of the nucleic acid sample prediction object recorded according to the long read sequencing method based on the preset sequencing depth, the enzyme kinetic information, and the base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to the preset window. The prediction model training module is used to train based on the first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature, thereby obtaining the first prediction model and the second prediction model; The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variable of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

[0024] According to a fourth aspect, the present invention provides a device for predicting nucleic acid chemical modifications, characterized in that the device comprises: The sample acquisition module is used to acquire nucleic acid samples to be predicted. The feature determination module is used to determine the features of the nucleic acid to be predicted based on the nucleic acid sample to be predicted. The features of the nucleic acid to be predicted include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to a preset window, recorded according to the third sequencing method based on a preset sequencing depth. The sample prediction module is used to input the nucleic acid features to be predicted into the first prediction model and the second prediction model respectively, to obtain the first prediction result output by the first prediction model, and to obtain the second prediction result output by the second prediction model; the first prediction model and the second prediction model are constructed by the nucleic acid chemical modification prediction model construction method in the first aspect or any embodiment of the first aspect; The probability result determination module determines the probability of chemical modification of the nucleic acid sample to be predicted based on the first prediction result, and determines the probability of chemical modification of the nucleic acid sample to be predicted based on the second prediction result.

[0025] Fifthly, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method for constructing a predictive model for nucleic acid chemical modification as described in the first aspect or any corresponding embodiment.

[0026] In a sixth aspect, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method for predicting nucleic acid chemical modifications as described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating a method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating another method for constructing a predictive model of nucleic acid chemical modification according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating another method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention; Figure 4This is a schematic diagram of nucleic acid feature extraction according to an embodiment of the present invention; Figure 5 This is a schematic flowchart illustrating another method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention. Figure 6 This is a schematic flowchart of the method for predicting nucleic acid chemical modifications according to an embodiment of the present invention; Figure 7 This is a flowchart illustrating another method for predicting nucleic acid chemical modifications according to an embodiment of the present invention; Figure 8 This is a schematic flowchart illustrating another method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention. Figure 9 This is a structural block diagram of the device constructed based on a predictive model of nucleic acid chemical modification; Figure 10 It is a structural block diagram of a predictive device based on nucleic acid chemical modifications; Figure 11 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention; Figure 12 This is a parameter optimization performance diagram of an embodiment of the present invention; Figure 13 This is a diagram showing the impact of training data volume on an embodiment of the present invention; Figure 14 This is an ancestor influence test diagram according to an embodiment of the present invention; Figure 15 This is a benchmark test diagram of an embodiment of the present invention; Figure 16 This is a low-depth data flow diagram of an embodiment of the present invention; Figure 17 This is a graph illustrating the impact of the number of genomic CpG sites on the present invention. Figure 18 These are test diagrams of different models in embodiments of the present invention; Figure 19 This is a low-depth benchmark test image according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Common DNA chemical modifications include 5mC (5-methylcytosine), 5hmC (5-hydroxymethylcytosine), 5caC (5-carboxylcytosine), 5fC (5-formylcytosine), 6mA (N6-methyladenine), and 4mC (N4-methylcytosine). Among these, 5mC is the most common chemical modification in mammals, while 6mA is the most common in prokaryotes such as bacteria. Compared to DNA chemical modifications, RNA chemical modifications are more numerous, with approximately 170 types identified so far. However, their functions require further investigation. Common RNA chemical modifications include m6A (N6-methyladenosine) and m5C (5-methylcytidine).

[0031] In the application of nucleic acid molecular chemical modification detection technology, high-throughput sequencing technology is used to detect chemical modifications on DNA or RNA. This detection method generally involves two steps: First, using antibodies, restriction enzymes that recognize specific sequence patterns (e.g., using MspI to recognize the "CCGG" sequence pattern), or chemical conversion (e.g., treating DNA with bisulfite) to enrich, label, or transform chemically modified DNA or RNA fragments; subsequently, the treated DNA or RNA sample is subjected to high-throughput sequencing, and downstream bioinformatics analysis tools are applied to identify the location and frequency of chemical modifications on the DNA or RNA. For example, bisulfite sequencing (BS-seq) is commonly used to identify CpG sites on DNA that have undergone methylation (5mC or 5hmC). This method first treats wild-type DNA samples with bisulfite reagent. After treatment, all unmethylated cytosine C is converted to uracil U (sequencing readout T), while methylated cytosine C is retained (sequencing readout C). Subsequently, high-throughput sequencing (usually short-read sequencing) is performed on the processed samples. In the sequenced reads, cytosine C comes from the methylated DNA template fragment, while T comes from the unmethylated DNA template fragment. By calculating the ratio of C to C+T, methylation sites can be detected at single-base resolution. However, these methods involve complex experimental steps, and various factors can affect the final results. For example, the enrichment efficacy of antibodies, the conversion efficiency of chemical reagents, and the fragmentation effect of chemical reagents on the original sample can all bias the downstream analysis results. Furthermore, these methods require the separate construction of sequencing libraries, resulting in high experimental costs. Moreover, for regions with high CG content, the DNA double helix does not easily unwind, and the experimental sample may exhibit a bias towards enriched regions. Although Reduced Representation Bisulfite Sequencing (RRBS) can enrich promoters and CpG island regions through enzyme digestion, subsequent bisulfite conversion sequencing is still required. Finally, the current mainstream high-throughput sequencing technology is short-read sequencing, with read lengths around 100-300 bps. There are technical limitations to the alignment and haplotype typing of short-read sequences. These shortcomings mean that it is currently impossible to conduct large-scale, genome-wide detection of DNA or RNA chemical modifications at the population level.

[0032] In view of this, embodiments of the present invention provide a method, method and apparatus for constructing a prediction model for nucleic acid chemical modifications. By constructing two different types of prediction models, the prediction and identification of different types of chemical modifications can be achieved at the single-molecule level and the genome level, thereby improving detection efficiency and detection integrity.

[0033] According to embodiments of the present invention, a method, method and apparatus for constructing a predictive model for nucleic acid chemical modification are provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] This embodiment provides a method for constructing a predictive model for nucleic acid chemical modifications, which can be used on mobile terminals such as mobile phones and tablets (the executing entity is described in conjunction with the actual situation). Figure 1 This is a flowchart of a method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample of the nucleic acid sample prediction object. The first nucleic acid sample is found to have undergone chemical modification through the first processing method, the second nucleic acid sample is found to have not undergone chemical modification through the second processing method, and the third nucleic acid sample is found to have undergone chemical modification based on a preset site through the third processing method. The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are combined according to a preset sequencing template based on bioinformatics.

[0035] For example, the nucleic acid sample prediction target can be common biological samples of nucleic acid molecules such as DNA or RNA, including but not limited to DNA or RNA molecules extracted from materials such as cells, cell lines, blood, saliva, hair, tears, urine, feces, organoids, ground organ tissue samples, and ground tumor tissue samples. For example, nuclear DNA and non-nuclear DNA, including but not limited to mitochondrial DNA, chloroplast DNA, circulating tumor DNA, and other types of circulating free DNA, are all applicable to this embodiment.

[0036] For example, the first nucleic acid sample is a positive sample, the second nucleic acid sample is a negative sample, and the third nucleic acid sample is a wild-type sample, which can be a wild-type sample in nature that has not undergone experimental treatment. Wild-type samples, positive samples, and negative samples can come from the same sample or different samples; they can be different chromosomes from the same sample or different chromosomes from different samples. The first treatment method can be enzyme treatment or long-read sequencing; the second treatment method can be micro-DNA whole-genome amplification or long-read sequencing; and the third treatment method can be pyrosequencing or whole-genome bisulfite sequencing (WGBS). Bioinformatics methods mainly include sequence alignment of sequencing reads. For example, reads from the PacBioSMRT sequencing platform are subreads, which are reads from the same sequencing template. Subreads can be directly aligned to the reference genome; alternatively, bioinformatics software (such as CCS, DeepConsensus, etc.) can be used to merge subreads from the same sequencing template to obtain high-fidelity consensus sequences, i.e., HiFi Reads, which are then aligned to the reference genome.

[0037] Step S102: Determine the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid sample prediction object recorded according to the long read sequencing method based on the preset sequencing depth.

[0038] Specifically, the attribute information of the nucleic acid sample prediction object includes: the color and intensity of the nucleic acid sample prediction object; enzyme kinetic information includes: inter-pulse duration (IPD) and pulse width (PW); base sequence information includes: the base sequence content (A, T, C, G) of each site in the upstream and downstream range of 10 bps, 20 bps, or 50 bps.

[0039] For example, long-read sequencing can be used, such as long-read sequencing based on SMRT (Single Molecule, Real-Time) technology developed by PacBio (platforms mainly include Sequel, Sequel II / Sequel IIe, Revio, etc.), which can generate sequencing reads of 5K to 20K bps.

[0040] Furthermore, the preset sequencing depth in this embodiment can be a low-depth sequencing depth. For example, long-read sequencing based on a preset sequencing depth can improve nucleic acid sequencing efficiency, enabling the sequencing of a complete telomere-to-telomere (T2T) genome map in areas such as genome assembly, detection of genetic variations (especially structural variations), and haplotype typing.

[0041] Specifically, high-depth sequencing samples are down-sampling to different low-depth levels. For example, sampling can be done from 60x to 5x, 10x, 20x, 30x (in ZMW); real-world sequencing samples with coverage of 5x, 10x, 20x, 30x (in ZMW) can be used; sites with coverage of 5x, 10x, 20x, 30x (in ZMW) extracted from the same or different regions of the same or different samples can be used; or a combination of one or more of these methods can be employed. Therefore, this embodiment can enhance high-resolution, high-accuracy sequencing data by preset sequencing depth.

[0042] Step S103: Training is performed based on the first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature to obtain the first prediction model and the second prediction model.

[0043] The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variable of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

[0044] Specifically, the first prediction model is used to predict whether nucleic acids at the single-molecule level have undergone chemical modification, that is, a model used to identify whether a certain chemical modification exists at a base site on each sequencing template molecule, and the second prediction model is used to model the frequency of chemical modification at a certain site on the genome.

[0045] For illustrative purposes, the following describes the main process of training the first, second, and third nucleic acid features based on the PacBio SMRT sequencing platform. Taking the training of a predictive model for DNA methylation modification (5mC) as an example, similar methods can be applied to other types of DNA chemical modifications, including but not limited to 5mC (5-methylcytosine), 5hmC (5-hydroxymethylcytosin), 5caC (5-carboxylcytosine), 5fC (5-formylcytosine), 6mA (N6-methyladenine), and 4mC (N4-methylcytosine). First, an appropriate amount of DNA is obtained according to step S101. Then, wild-type samples, positive control samples (obtained by treatment with M.SssI methyltransferase), and negative control samples (obtained by micro-DNA whole-genome amplification) are prepared according to step S102. Subsequently, the three samples were processed using long-read sequencing methods (e.g., using the PacBio SMRT sequencing platform) to obtain the characteristic values ​​of each single molecule (HiFi-read or subreads from the same template) of wild-type DNA, positive sample DNA, and negative sample DNA. This included the base sequence content (A, T, C, G) and enzyme kinetic signals (IPD, PW) at each site within a preset CpG site and its upstream and downstream ranges of 10 bps, 20 bps, 30 bps, 40 bps, or 50 bps (the window size can be adjusted arbitrarily in this step, with the default being 10 bps upstream and downstream). Simultaneously, for the wild-type samples, pyrosequencing or whole-genome bisulfite sequencing (WGBS) was used to orthogonally obtain the frequency of 5mC methylation at each CpG site in the genome. On the PacBio SMRT sequencing platform, library template molecules are immobilized in ZMW wells (Zero-Mode Waveguides) through hybridization. These template molecules can be uniquely identified by the ZMW identifier, so the single-molecule model (first prediction model) can be called the ZMWmeth model. The genome-level model (second prediction model) can be called the Joint_CpG model because it combines information from the target site and its surrounding upstream and downstream CpG sites.When training a single-molecule level chemical modification prediction model (ZMWmeth model), the feature values ​​extracted from positive and negative samples are used as independent variables X, and the methylation probability at the single-molecule level is used as dependent variable Y (assuming complete conversion, the single-molecule methylation probability of positive samples is 1, and the single-molecule methylation probability of negative samples is 0). The relationship between X and Y is established, thereby generating a single-molecule level chemical modification prediction model.

[0046] For example, when training a genome-level chemical modification prediction model (Joint_CpG model), it is necessary to first use the ZMWmeth model mentioned above to predict all single molecules measured in a sample: that is, using ZMW as the processing unit (HiFi-reads or genome sequencing data subreads from the same ZMW), predict the probability of methylation at the covered CpG sites, and then combine the predicted probabilities of all molecules at 5, 10, 20, 30, 50, 60, 100 CpG sites upstream and downstream of each CpG site as the independent variable X, and use the methylation modification probability obtained from pyrosequencing or sulfite sequencing as the dependent variable Y to establish the relationship between X and Y.

[0047] The method for constructing a prediction model for chemical modifications of nucleic acid samples based on long-read sequencing technology provided in this embodiment utilizes the nucleic acid feature information determined by the first, second, and third nucleic acid samples under the long-read sequencing platform to train and obtain prediction models for chemical modifications at the single-molecule level and the genome level. When used to predict nucleic acid chemical modifications, it can improve the efficiency and accuracy of nucleic acid chemical modification prediction. Furthermore, it eliminates the need to conduct other experiments (such as pyrosequencing) on ​​the same nucleic acid sample, reducing a significant amount of labor and lowering detection costs.

[0048] This embodiment provides a method for constructing a predictive model for nucleic acid chemical modifications, which can be used on the aforementioned mobile terminals, such as mobile phones and tablets. Figure 2 This is a flowchart illustrating the specific execution of step S101 according to an embodiment of the present invention, as follows: Figure 2 As shown, the process includes the following steps: Step S201: Obtain various types of raw nucleic acid samples, including: non-wild-type DNA nucleic acid samples and / or wild-type DNA nucleic acid samples and / or wild-type RNA nucleic acid samples and / or non-wild-type RNA nucleic acid samples.

[0049] Specifically, wild-type DNA nucleic acid samples can be untreated wild-type samples found in nature, while non-wild-type DNA nucleic acid samples are treated wild-type samples found in nature. For details on how to obtain biological samples, please refer to step S101 above; these details will not be repeated here.

[0050] Step S202: Prepare various types of original nucleic acid samples based on a preset preparation method.

[0051] For example, the preset preparation method includes DNA sample fragmentation, end enzyme repair and completion, and the addition of sequencing adapters and primers.

[0052] Step S203: Based on the prepared original nucleic acid samples of various different types, create nucleic acid sample sequencing libraries.

[0053] Specifically, for the sake of clarity in this implementation, the original nucleic acid sample can be sequenced using the PacBio SMRT sequencing platform to achieve long-read sequencing of the nucleic acid sample.

[0054] Step S204: Based on the nucleic acid sample sequencing library, obtain the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample.

[0055] The method for constructing a predictive model for nucleic acid chemical modification provided in this embodiment creates nucleic acid sample sequencing libraries by combining various types of original nucleic acid samples with sequencing tools. For fragmented original nucleic acid samples, it is not necessary to build nucleic acid sample sequencing libraries one by one. Therefore, it can reduce the cost of performing biological experiments on original nucleic acid samples one by one, thereby improving the enrichment efficiency of nucleic acid samples and the conversion efficiency of chemical reagents.

[0056] This embodiment provides a method for constructing a predictive model for nucleic acid chemical modifications, which can be used on the aforementioned mobile terminals, such as mobile phones and tablets. Figure 3 This is a flowchart illustrating the specific execution of step S102 according to an embodiment of the present invention, as follows: Figure 3 As shown, the process includes the following steps: Step S301: Sequencing of the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample is performed according to the long read sequencing method based on the preset sequencing depth.

[0057] Specifically, the preset sequencing depth is a low-depth sequencing depth. The long-read sequencing method and preset sequencing depth have been described in detail in step S102 above, and will not be repeated here. The sequencing processing method can be the long-read sequencing processing method based on the PacBio SMRT sequencing platform.

[0058] Step S302: Based on the biological sample sequencing platform, feature extraction is performed on the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample after sequencing.

[0059] Step S303: Based on the feature extraction results, obtain the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample.

[0060] In some optional implementations, step S302 above involves: extracting features from the first, second, and third nucleic acid samples after sequencing processing using a biological sample sequencing platform, including: The first nucleic acid feature is determined based on the comparison results between the first nucleic acid sample and the first standard nucleic acid sample; the second nucleic acid feature is determined based on the comparison results between the second nucleic acid sample and the second standard nucleic acid sample; and the third nucleic acid feature is determined based on the comparison results between the third nucleic acid sample and the third standard nucleic acid sample.

[0061] like Figure 4 The diagram shown is a schematic diagram of feature extraction in this embodiment.

[0062] Specifically, in the sequence alignment file, the alignment reads that overlap with the study window (standard nucleic acid samples) are first located, then the study window is mapped onto each read (local study window, i.e. the nucleic acid sample under study), and finally the base sequence content (A, T, C, G) and enzyme kinetic signals (IPD, PW) of each site within the local study window are extracted.

[0063] For feature values ​​extracted from different subreads from the same sequencing template, they can be listed together according to the research window position number, and then the average value of each position can be calculated; alternatively, the accuracy of each subread can be preserved. The feature extraction method in this embodiment is relatively flexible, compatible with both subreads and HiFi reads, and can output raw signal values ​​or standardized signal values. Furthermore, the size of the upstream and downstream research windows can be flexibly adjusted, and the alignment quality of the selected reads, the sequencing quality of bases, and whether to extract features according to positive and negative strands can be set.

[0064] The biological sample sequencing platform is the PacBio SMRT sequencing platform. For example, using the PacBio SMRT sequencing platform, the nucleic acid molecules to be tested in the nucleic acid sample sequencing library are immobilized in the ZMW (zero-mode waveguide) pores through hybridization. Since these template molecules can be uniquely distinguished by the ZMW ID, feature extraction of the sample characteristics can be further performed.

[0065] For example, based on requirements, the user provides the location information of the target site to be identified on the reference genome (a preset standard nucleic acid genome), as well as the type of chemical modification to be identified. For instance, when identifying 5mC on the GRCh38 version of the human reference genome, the chromosome number (e.g., "chr1") and the location coordinate number of each CpG site (e.g., "1001") need to be provided. Then, using the LDLR-Meth (low-depth long-read sequencing software) developed in this embodiment, the base sequence content (A, T, C, G) and enzyme kinetic signals (IPD, PW) of each site within a 10bps, 20bps, 30bps, 40bps, or 50bps range upstream and downstream of the target site are extracted (the window size can be adjusted arbitrarily in this step, the default is 10bps upstream and downstream). Figure 3 In this context, the reference genome can be a common reference genome in the field, such as reference genomes hg19, GRCh38, mm10, or a genome sequence draft obtained through assembly.

[0066] In one alternative implementation, during the process of obtaining the trained first prediction model, the model parameters of the initial first prediction model are set based on the first input variable, and the initial first prediction model is gradually optimized based on the first output variable to obtain the trained first prediction model.

[0067] For example, the first prediction model described in the above embodiments is used to detect nucleic acid molecules at the single-molecule level. This model is also known as the ZMWmeth model. The framework of the ZMWmeth model can be linear, generalized linear, mixed linear, based on neural network models (such as convolutional neural network models CNN, recurrent neural network RNN, long short-term memory network LSTM, Transformer, etc.), and other deep learning-based models. Different models have different architectures, and their frameworks and parameters need to be adjusted, but all can be trained using the methodology of this embodiment. The ZMWmeth model constructed in this embodiment is based on a CNN framework by default. The model's structure is as follows: First, it contains five convolutional modules that perform convolutions along the 5'-3' direction of the sequence on the genome. Each convolutional module includes: a one-dimensional convolutional layer (Conv1D) with a stride of 1, a convolution length of 5 (kernel_size), 256 kernels (filters), and ReLU (Rectified Linear Unit) activation function; a batch normalization layer; and a dropout layer with a random inactivation rate of 0.2. After convolution, the flattened convolution result is linked to the number of sequencing reads from the ZMW template's molecular sequencing data subreads. Finally, it passes through a ReLU-activated fully connected layer (Dense), a dropout layer with a random inactivation rate of 0.2, and three sigmoid-activated fully connected layers (Dense). The unit sizes of the first two fully connected layers are 128 and 32, respectively, ultimately yielding the output probability values. The loss function used during model training is the binary cross-entropy function, optimized using the Adam optimizer with default parameters in TensorFlow and Keras. X can be composed of 10%, 30%, 50%, 70%, or 100% CpG sites, or only CpG sites within CpG islands, etc. However, to achieve model saturation, at least 5 million different CpG sites should be selected for training. For CpG sites covered by multiple single molecules (e.g., covered by 30 genome sequencing HiFi reads), one or more single molecules can be randomly selected as the training set. The training set data can be single-molecule inputs of single-stranded or double-stranded DNA, which are then input into the ZMWmeth model for training. The LDLR-Meth included in this embodiment can be used to train the first predictive model for single-molecule level chemical modifications.

[0068] In one alternative implementation, the model parameters of the initial second prediction model are set based on the second input variable, and the trained second prediction model is obtained by progressively optimizing the initial second prediction model based on the second output variable.

[0069] For example, the second prediction model for detecting nucleic acids at the genome level, as described in the above embodiments, is the Joint-CpG model. This model first requires using the ZMWmeth described above to predict the methylation probability of all single molecules measured in a sample. Specifically, using ZMWs as processing units (HiFi reads from genome sequencing data or subreads from the same ZMW), the probability of methylation at the covered CpG sites is predicted. Then, the predicted probabilities of all molecules at 5, 10, 20, 30, 50, 60, and 100 CpG sites upstream and downstream of each CpG site are combined as independent variables X, and the methylation modification frequency obtained from pyrosequencing or sulfite sequencing is used as the dependent variable Y to establish the relationship between X and Y, thereby generating a genome-level chemical modification prediction model. The model framework here can also be linear, generalized linear, mixed linear, based on neural network models (e.g., Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Transformers, etc.), and other deep learning-based models. Different models have different architectures, requiring adjustments to their framework and parameters, but all can be trained using this embodiment. The Joint-CpG model built in this embodiment is based on a CNN framework by default, and its structure is as follows: First, it includes two one-dimensional convolutional layers (Conv1D), which convolve along the 5'->3' direction of the CpG site on the genome. Each Conv1D has a stride of 1, a convolution length of 4 (kernel_size), 96 kernels (filters), and a ReLU activation function. This is followed by a batch normalization layer, a dropout layer with a 0.3 inactivation rate, and a one-dimensional max pooling layer (MaxPooling1D). After convolution, the convolution result (Flatten) is unfolded and connected to three fully connected layers (Dense) activated by the ReLU function, with unit sizes of 30, 10, and 10 respectively. Finally, the predicted methylation level of the CpG site is obtained by outputting a fully connected layer (Dense) activated by a linear function. The loss function used during model training is the Mean Absolute Error, optimized using the 'sgd' optimizer in TensorFlow and Keras. The LDLR-Meth (Low-Depth Long Read Sequencing Software) included in this example can be used to train a genome-wide chemical modification prediction model.

[0070] Therefore, the method for constructing a predictive model for nucleic acid chemical modifications in this embodiment can simultaneously train both molecular-level and genomic-level chemical modification prediction models, or train either model individually. The nucleic acid samples (wild-type, positive, and negative samples) input into both models can come from the same sample or different samples; they can be different chromosomes from the same sample or different chromosomes from different samples. Based on the sequencing platform, the above model training is applicable to different data types, including single-molecule sequencing data subreads and HiFi reads. Because the feature values ​​of single-molecule sequencing data subreads and HiFi reads differ, the models for the two data types cannot be substituted for each other. Figure 5 The diagram shown illustrates the construction of a predictive model for nucleic acid chemical modifications. Figure 5 The gold standard test is whole-genome bisulfite sequencing (WGBS) or pyrosequencing.

[0071] According to an embodiment of the present invention, a method for predicting nucleic acid chemical modifications is also provided, which can be used on mobile terminals, such as mobile phones, tablet computers, etc. (the executing entity is described in conjunction with the actual situation). Figure 6 This is a flowchart of a method for constructing a predictive model for nucleic acid chemical modification according to an embodiment of the present invention, such as... Figure 6 As shown, the process includes the following steps: Step S601: Obtain the nucleic acid sample to be predicted.

[0072] For example, the nucleic acid sample to be predicted is a randomly obtained biological sample of nucleic acid molecules or genomes such as DNA or RNA. Typically, this biological sample is a wild-type biological sample, and the prediction is performed on this wild-type biological sample to determine whether it has undergone chemical modification. The method of obtaining this nucleic acid sample is the same as in steps S201-S204 of the above embodiments, and will not be described in detail here.

[0073] Step S602: Based on the nucleic acid sample to be predicted, determine the characteristics of the nucleic acid to be predicted corresponding to the nucleic acid sample to be predicted. The characteristics of the nucleic acid to be predicted include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to the preset sequencing depth recorded by the third sequencing method.

[0074] Specifically, the method for determining the characteristics of the nucleic acid to be predicted is the same as steps S301-S303 in the above embodiments, and will not be repeated here.

[0075] Step S603: Input the nucleic acid features to be predicted into the first prediction model and the second prediction model respectively for prediction, obtain the first prediction result output by the first prediction model, and obtain the second prediction result output by the second prediction model; the first prediction model and the second prediction model are constructed by the prediction model construction method of nucleic acid chemical modification.

[0076] Step S604: Determine the probability of chemical modification of the nucleic acid sample to be predicted based on the first prediction result, and determine the probability of chemical modification of the nucleic acid sample to be predicted based on the second prediction result.

[0077] Specifically, in Figure 5 In this study, the first prediction model is used to predict nucleic acids at the single-molecule level, and the second prediction model is used to predict nucleic acids at the genome level.

[0078] For example, for high-depth sequencing data, such as sequencing depths of 10X or higher (in HiFi-reads), using Subreads and HiFi-reads is essentially equivalent, with HiFi-reads being the preferred choice. The specific process is as follows: A single-molecule level chemical modification prediction model (ZMWmeth model) is used to predict the probability of chemical modification for each molecule; then, all molecules from multiple sites upstream and downstream of each research site (5, 10, 20, 30, 50, 60, 100, etc.) are combined as independent variables X, and a genome-level chemical modification prediction model (Joint-CpG model) is applied to predict the probability of chemical modification at each site at the genome level.

[0079] For example, for low-depth sequencing data, such as sequencing depths below 20X (in ZMW), it is recommended to use HiFi-reads and Subreads in combination with the second prediction model for low-depth sequencing (Joint-CpG model) included in this embodiment for prediction. For example, after sequence alignment and feature extraction using HiFi-reads and Subreads, the methylation probabilities of all molecules at 5, 10, 20, 30, 50, 60, and 100 CpG sites upstream and downstream of each CpG site are used as the independent variable X. The second prediction model for low-depth sequencing conditions at the genome level (Joint-CpG model) is then applied to predict the probability of chemical modification at each site at the genome level.

[0080] In one alternative implementation, such as Figure 7 As shown, step S603 above, obtaining the first prediction result output by the first prediction model, includes: S6031A, when the first prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone nucleic acid chemical modification.

[0081] S6032A, when the first prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0082] Specifically, the preset threshold can be flexibly set according to the actual situation. For example, to predict the chemical modification level of each site at the single-molecule level, the preset threshold can be 0.5. The state of whether each single molecule has chemical modification is judged (if the predicted value is greater than or equal to 0.5, it is judged that the molecule has chemical modification and counted as 1; otherwise, it is counted as 0). Then, the proportion of molecules predicted to have methylation at the site is calculated as the chemical modification level of the site ("count-based").

[0083] In another alternative implementation, in Figure 7 In step S603 above, obtaining the second prediction result output by the second prediction model includes: S6031B, when the second prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone nucleic acid chemical modification. S6032B, when the second prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0084] The same applies to predicting chemical modifications in low-depth sequencing of nucleic acids at the genome level, i.e., predicting the level of chemical modification at each site at the genome level. This preset threshold can also be 0.5. First, the presence or absence of chemical modification is determined for each single molecule (if the predicted value is greater than 0.5, the molecule is considered to have chemical modification and counted as 1; otherwise, it is counted as 0). Then, the proportion of molecules predicted to have methylation at that site out of all molecules at that site is calculated, and this proportion is taken as the level of chemical modification at that site.

[0085] like Figure 8 The diagram shown is a schematic representation of a method for predicting nucleic acid chemical modifications.

[0086] The method for predicting nucleic acid chemical modifications in this embodiment of the invention can be used to predict nucleic acid samples with haplotype accuracy, and is applicable to diploid, triploid, and other polyploid organisms. For ease of explanation, this embodiment uses DNA methylation (5mC) in diploid organisms as an example: First, it is necessary to obtain the genetic variations in the DNA of the test sample, such as single nucleotide sequence variations, small fragment insertion / deletion variations, etc. These genetic variations can be obtained through gene chips, short-read sequencing, long-read sequencing, Sanger sequencing, etc. For short-read sequencing, tools such as GATK, Samtools / BCFtools, and Freebayes can be used for this purpose; for long-read sequencing, tools such as deepvariant can be used to detect genetic variation sites. Subsequently, tools such as SHAPEIT and WhatsHap are used to phasing (distinguishing haplotypes) the genetic variations of the sample; haplotype phasing can also be performed using pedigree information. Typically, VCF and other file formats are used to store genetic variations, their corresponding haplotype blocks, and haplotype numbers. Genetic variant files such as VCFs (Vehicle Forms) labeled with haplotype information, along with alignment files from long-read sequencing (BAM format), are annotated with haplotype information (e.g., HP tags, PS tags) in the reads of the BAM files using tools like WhatsHap. Steps 1-4 are then performed to obtain haplotype-distinguishing feature values. Finally, steps S601-S604 output haplotype-level chemical modification prediction values. Genetic variants distinguishing haplotype phasing can be obtained in independent experiments or through steps S101-S102 in the above embodiments. Predicting chemical modification sites at haplotype precision can be applied to the auxiliary diagnosis of diseases caused by genetic imprinting.

[0087] Therefore, the nucleic acid chemical modification prediction method in this embodiment, based on the single-molecule chemical modification prediction model and the genome-level chemical modification prediction model, can predict different types of nucleic acid samples to be predicted, including the prediction of nucleic acid samples with monomeric precision. It can improve the prediction efficiency and accuracy of nucleic acid chemical modifications, and reduce a lot of labor, eliminating the need to conduct experiments on a large number of nucleic acid sample prediction objects one by one, thereby reducing experimental costs.

[0088] The above embodiments will now be explained and illustrated with several specific examples to fully demonstrate the predictive model construction method and the effect that the predictive method can achieve for nucleic acid chemical modification in the above embodiments.

[0089] Example 1: Training and Validation of a Single-Molecular-Level DNA Methylation Prediction Model This example demonstrates how sequencing data generated using the PacBio SMRT sequencing platform Sequel II was used to establish and apply a model to predict DNA methylation modifications (5mC) at the single-molecule level. Specific details include: generation of training samples and data, training and parameter optimization of the single-molecule level DNA methylation prediction model (i.e., the ZMWmeth model), and testing the model on independent samples.

[0090] Preparation and sequencing of PacBio SMRT sequencing libraries. DNA samples were cut using the Diagenode Megaruptor 2 instrument (Diagenode). Qualified DNA samples for library construction required the following quality controls: sample concentration ≥100 ng / ul (NanoDrop 2000 spectrophotometry and Qubit fluorescent dye method), sample purity NanoDrop 260 / 280 between 1.8 and 2.0, and main DNA fragment band ≥40 kb (detected by pulsed-field electrophoresis). SMRT bell libraries were prepared using the SMRTbellprep kit 2.0 (100-938-900, Pacific Biosciences) according to the manufacturer's standard protocol (101-693-800). The Sequel IIe system (101-986-400) / Sequel II system (101-447-000) is the sequencing platform, and the sequencing kit is the Sequel II Sequencing Kit 2.0. It uses a 30-hour movie / SMRT chip and preserves the enzymatic kinetics of DNA synthases during sequencing. The average product per SMRT cell is approximately 360 G subreads. Wild-type DNA samples collected from HG02080, HG01109, and HG02723 were each processed using two SMRT cells to achieve approximately 25x HiFi read coverage. Fully methylated samples treated with M.SssI and fully unmethylated WGA samples each used one SMRT cell to ensure sufficient read length / molecule count for model training and validation. In low-depth sequencing designs, a total of one SMRT cell was used for sequencing four samples, while in high-depth sequencing designs, three SMRT cells were used per sample.

[0091] Whole genome bisulfite sequencing was performed. 15 μg of wild-type DNA from HG02080, HG01109, and HG02723 cell lines was subjected to ultrasonic shearing followed by end repair and 3' adenylation. Fragments of 450–550 bp were then purified using CleanNGS (CNGS-0500) and bisulfite-transformed using the EZ-96 DNA Methylation-Lightning Kit (Zymo Research, catalog number D5032) to create libraries suitable for NGS sequencing. The number of libraries was determined using KAPA Library Quantification Kits. Paired-end sequencing with 150 bp reads was performed on the Illumina NovaSeq 6000 platform using the NovaSeq 6000 S4 kit v1.5 (300 cycles, 20028312). Three WGBS libraries were constructed for each wild-type cell line and subsequently sequenced, yielding 300 GB of raw data. In addition, the publicly available WGBS files of HG001, HG002, and HG005 B-LCL cell lines were downloaded from the EpiQC project website.

[0092] ZMWmeth model training. The ZMWmeth model detects DNA 5mC modifications using kinetic signals during single-molecule real-time (SMRT) sequencing, including pulse interval (IPD) and pulse width (PW), as well as sequence context (A, T, C, G). The model is named after the fact that individual molecules can be uniquely distinguished by the zero-mode waveguide (ZMW) identifier on each SMRT chip. Variants of the ZMWmeth model are trained for both HiFi-ZMW molecules (template molecules can be read as HiFireads) and non-HiFi-ZMW molecules, achieving both double-stranded and single-stranded resolution. The specific training process is described below: i. Feature Extraction and Formatting: From the BAM format alignment file, IPD, PW, and DNA sequence context (A, T, C, G) for each aligned read length are iteratively extracted, centered on the CpG site, ranging from ±10 bps (5'-3'). IPD and PW signals are stored in the ip and pw tags of the subreads and the fi, fp, ri, and rp tags of the HiFi reads. Sequence context is extracted from the SEQ tags and encoded using a one-hot encoding method (i.e., 'A' is (1,0,0,0), 'G' is (0,1,0,0), 'C' is (0,0,1,0), and 'T' is (0,0,0,1)) to represent these categorical variables. The signals of subreads with the same ZMW are averaged, and consistent sequences are obtained through alignment. To avoid proportional differences between different features, IPD and PW values ​​are normalized to Z-scores. The extracted features were then formatted into a (21, 13) matrix for double-strand integration, where columns 1-6 represent positive strand features, columns 7-12 represent negative strand features, and the last column represents the sub-read count. Each row stores ±10 bps relative to the central CpG site. Similarly, single-strand features were formatted into a (21, 7) matrix, where 7 represents the covered strand. Here, positive and negative strands are relative to the reference genome. The ground truth label "1" indicates a fully methylated training sample, and "0" indicates a fully unmethylated training sample. ±10 bps is used because... Figure 12 As tested, the information sites are roughly located within this region, and the prediction accuracy reaches a plateau of 0.89 when the window size is larger.

[0093] ii. ZMWmeth model construction: The CNN-based ZMWmeth model includes the following modules, implemented using Keras in the Python programming language.

[0094] • First, there are 5 one-dimensional convolutional (Conv1D) layers, each with 256 filters. Each filter has a kernel size of (5, 5) and uses ReLU as the activation function. • After convolution, there is a batch normalization layer, a dropout layer, and a MaxPooling1D layer, with a dropout rate of 0.2. • Finally, the convolution result is flattened into a one-dimensional vector and subreads are counted before being fed into fully connected layers; The ZMWmeth model consists of three fully connected layers. The first two layers have unit sizes of 128 and 32, respectively. The last fully connected layer is equipped with a sigmoid activation function to obtain output probabilities in the range [0,1].

[0095] The architecture and parameters of the ZMWmeth models selected above were determined by searching a predefined parameter space (as shown in Table 1 below) using a Bayesian optimization function that implements a Gaussian process to aid in testing hyperparameter combinations. In 50 trials, the second-best model was chosen as the default model for the software implementation because it was considered to strike a balance between prediction accuracy and the number of parameters. It has 134,593 parameters and achieved an accuracy of 0.86 on the training samples of HG002.

[0096] Table 1

[0097] iii. The ZMWmeth model training process involves extracting feature values ​​from whole-genome reads from fully methylated and fully unmethylated samples, formatting them, randomly shuffling them, and dividing them into training and validation sets in a 9:1 ratio. A total of 70 epochs are allowed. During iteration, binary cross-entropy is used as the loss function, and monitoring is performed on the validation set. The following guidelines are applied before terminating the iteration and generating the ZMWmeth model: • The loss function did not improve over 20 consecutive epochs; • If the first criterion above is not met, the training process will terminate after the maximum allowed 70 epochs have been reached.

[0098] Performance of the ZMWmeth model and comparison with other software. The ZMWmeth model was used with LDLR-Meth (low-depth long-read sequencing software) to determine the methylation status of CpG sites at the single-molecule level. This model is a CNN-based model trained on fully methylated positive and fully unmethylated negative DNA samples. First, model parameter optimization characterized factors affecting the parameter learning results, including the number of training CpG sites and the ancestry of the training samples. Specifically, ZMWmeth models were trained and tested with different numbers of CpG sites (from 300,000 to 30 million) as input. The prediction accuracy of the ZMWmeth model increased with the number of training CpG sites, converging to an accuracy of 0.89 after 5 million training CpG sites. Figure 13For training with fewer than 5 million CpG sites, the ZMWmeth model will fail to reach saturation and stability due to sampling bias during training and validation. Furthermore, for a given 1 million CpG sites, the accuracy of the ZMWmeth model increases with the number of reads / molecules per CpG site. Figure 13 The B result indicates that both the number of CpG sequence contexts and the read / molecule count played a role. In subsequent analyses, all reads / molecules were used to train the ZMWmeth model. To explore whether the ancestry of the training samples caused bias, this embodiment trained and tested ZMWmeth models with different ancestry, including East Asia, Southeast Asia, Europe, Africa, and the Americas. The test results show that the ZMWmeth model trained from one ancestry achieves similarly high accuracy across various ancestries. Figure 14 This indicates that the ancestry of the training samples is not a key influencing factor.

[0099] In addition to HiFi-ZMW molecules, this example also trains ZMWmeth model variants applicable to non-HiFi-ZMW molecules, including double-stranded and single-stranded resolutions. For example... Figure 15 As shown in Figure A, for HiFi-ZMW molecules, using HiFi reads or subreads as input can achieve an accuracy of 0.86-0.89 and an AUC of 0.94-0.95, with the accuracy using HiFi reads being slightly higher than 0.89. For non-HiFi-ZMW molecules, two distinct groups can be further distinguished: one group consists of non-HiFi-ClosedZMW molecules (where the template molecule cannot be read as a HiFi read, and both the forward and reverse strands of the template molecule are covered by the sequenced subreads), achieving an accuracy of 0.71; the other group consists of non-HiFi-OpenZMW molecules (where the template molecule cannot be read as a HiFi read, and only one strand of the template molecule is covered by the sequenced subreads), achieving an accuracy of 0.65.

[0100] This embodiment benchmarks LDLR-Meth (low-depth long-read sequencing software) against other available software, even though these software programs were originally designed to process specific types of molecules, rather than others. Figure 15As shown in Figure B, on HiFi reads, LDLR-Meth is comparable to ccsmeth, with accuracy and AUC of 0.89 and 0.95, respectively, and their performance is superior to primrose. In the software used for benchmarking, HKModel is capable of processing subread data. On HiFi-Closed ZMW molecules, LDLR-Meth and HKModel show similar performance, with an accuracy of 0.85 and an AUC of 0.93. On the same non-HiFi-Closed ZMW molecules reported by LDLR-Meth and HKModel, an accuracy of 0.75 and an AUC of 0.83 can be achieved; furthermore, LDLR-Meth can determine the modification state on a wider range of non-HiFi-Open ZMW molecules with an accuracy of 0.63 and an AUC of 0.67.

[0101] Example 2: Training and Validation of DNA Methylation Model under Low-Depth Sequencing This embodiment mainly demonstrates the prediction of methylation frequency at CpG sites at the genome level, including the training and parameter optimization of the Joint-CpG model, and in particular, solutions for low-depth sequencing coverage.

[0102] Training the Joint-CpG model. The Joint-CpG model predicts the frequency of 5mC modification by simultaneously considering the target CpG site and its surrounding neighboring cis-CpG sites. The training input features use the predicted methylation probabilities of all sequencing molecules that overlap with the target CpG site and its surrounding neighboring cis-CpG sites. Specifically, in the case of low-depth sequencing, a variant of the Joint-CpG model is generated by training with all types of sequencing template molecules, including those from HiFi-ZMW and non-HiFi-ZMW, thereby detecting the maximum number of CpG sites while maintaining high accuracy. The specific training process is described below.

[0103] The training input features for the Joint-CpG model are as follows: First, in wild-type samples, the methylation status of CpG sites on each individual molecule is predicted using the ZMWmeth model. The methylation probabilities of all individual molecules overlapping at the same CpG site are counted in a histogram vector ranging from [0,1], with each bin having a size of 10, and then L2 normalized. Specifically, in low-depth sequencing designs, this histogram vector can be spliced ​​after grouping HiFi-ZMW molecules and / or non-HiFi-ZMW molecules. Figure 16Then, for the target CpG site and its surrounding cis-CpG sites, their normalized vectors are stacked column-by-column in the 5'-3' direction to form a feature matrix, where each column corresponds to a CpG site. Finally, the distance weight from each surrounding cis-CpG site to the target CpG site is added to the last row of the feature matrix; this weight is calculated using the following function. ; in | Yes The absolute distance to the target CpG site (in base pairs). The distance weight is 1.0 for the target CpG and gradually decreases with increasing distance. To assess the impact of CpG sparsity on the Joint-CpG model, CpGs are further categorized into dense CpGs (average distance not exceeding 40 bp), medium-density CpGs (40-100 bp), and sparse-density CpGs (>=100 bp). The label for each feature matrix is ​​the methylation frequency of the target CpG estimated from WBGS experiments (>=10X coverage) or pyrosequencing.

[0104] The construction, training, and optimization of the Joint-CpG model. The Joint-CpG model, based on a convolutional neural network, has a similar structure to the ZMWmeth model, except that the loss function is mean absolute error, since the methylation frequency is a quantitative value ranging from [0, 1]. The training and optimization process is similar to that described in the ZMWmeth model and will not be repeated here. The Joint-CpG model includes 10 cis-CpGs surrounding the target CpG site, as more cis-CpGs do not contribute to the prediction accuracy based on metrics such as Pearson correlation coefficient (PCC) and mean absolute error (MAE) (Figure 17, A). After 50 trials, the optimal model was selected and implemented in the software as the default model, achieving the lowest MAE (0.074) on the training samples of wild-type HG002 (Table 2). Furthermore, to avoid CpG site bias within a specific DNA methylation frequency range, the CpG training dataset used for model training was collected in a balanced manner, stratified by different DNA m frequencies, through over-sampling or under-sampling. This example was benchmarked against other machine learning algorithms, such as the Transformer model (Figure 18).

[0105] Long read strategies for low-depth sequencing. In the case of low-depth sequencing, this example tested five strategies aimed at determining the maximum number of CpG sites recalled while maintaining accuracy. As shown below, these five strategies represent one or more sequencing molecules, which can be subreads or HiFi reads. Here, this example further distinguishes two subtypes of non-HiFi-ZMW molecules: one is a non-HiFi-ClosedZMW molecule (the template molecule cannot be read as a HiFi read; both the forward and reverse strands of the template molecule are covered by sequencing subreads), and the other is a non-HiFi-OpenZMW molecule (the template molecule cannot be read as a HiFi read; one strand of the template molecule is covered by sequencing subreads). Accordingly, the Joint-CpG model was trained and tested.

[0106] S1: Using ZMW molecules, in subreads format S2: Using Closed-ZMW molecules, in subreads format S3: Using HiFi reads S4: Combining HiFi reads and subreads formats for non-HiFi-Closed ZMW molecules S5: Combining HiFi reads and subreads formats with non-HiFi-ClosedZMW molecules and non-HiFi-OpenZMW molecules.

[0107] To evaluate the performance of the five strategies at different coverage depths, we generated training and testing datasets (at the molecule level) of 5x, 7x, 10x, 15x, and 20x by downsampling HG002 and two internal blood samples.

[0108] Cis-CpGs were used to predict 5mC modification frequencies at the genome level. LDLR-Meth predicted methylation frequencies at key CpG sites at the genome level using a convolutional neural network-based Joint-CpG model trained on wild-type DNA samples. The optimized Joint-CpG model achieved a mean absolute error (MAE) of 7.0 on the wild-type HG002 training sample. At least 500,000 CpG sites are required to generate a stable and saturated Joint-CpG model with a correlation coefficient (PCC) of 0.90 and an MAE of 10.0. Figure 17(B) Next, due to the uneven distribution of CpG across the human genome, the sparsity of cis-CpG was characterized in this example. The results show that the Joint-CpG model trained from dense CpG achieved a PCC of 0.92 and an MAE of 10.0 when testing with dense CpG. However, when used to predict medium-density and sparse-density CpG, the prediction accuracy decreased, with the PCC dropping below 0.90 and the MAE increasing to 13.0. Conversely, the Joint-CpG models trained from medium-density and sparse-density CpG were unaffected by these biases. Finally, the Joint-CpG model trained using whole-genome CpG remained stable across all CpG sparse groups in the tests, achieving an MAE of 11.0 and a PCC of 0.90 for whole-genome CpG. Therefore, the Joint-CpG model trained with whole-genome CpG is used by default.

[0109] Furthermore, this embodiment demonstrates the existence of tissue specificity among available tissues, including B-LCL cell lines, blood tissues, and tumor disease tissues. Joint-CpG models trained with a combination of multiple tissues typically exhibit lower MAE. However, Joint-CpG models trained using only a single tissue (e.g., using one blood sample, denoted as Blood (N=1)) show increased MAE compared to multi-tissue Joint-CpG models, even when more blood samples are included in the training process, such as Blood (N=3). This also applies to single tumor disease tissues. Further examination of the prediction spectrum of single-tissue Joint-CpG models reveals biased predictions for certain CpG groups. For example, Joint-CpG models trained from blood tissue show higher MAE predictions for CpGs with locus DNAm frequencies in the ranges of 60-80 and 80-100, indicating that these CpG groups are underestimated in the training tissue. Therefore, in subsequent analyses, we used a Joint-CpG model trained on multiple tissues to maintain a balance between sparsity and DNAm range in terms of CpG representativeness, avoiding biases associated with CpG representativeness. Finally, at different coverage depths (in HiFi reads), the Joint-CpG model outperformed counting-based methods overall, exhibiting higher PCC and lower MAE.

[0110] DNA methylation detection in low-depth long-read sequencing. In low-depth sequencing design, the goal is to accurately determine DNA methylation frequencies at the maximum number of CpG sites. LDLR-Meth increases the number of CpG sites reaching the desired depth (e.g., >= 4x) by utilizing all sequenced single molecules, while simultaneously considering surrounding CpG sites to maintain high predictive accuracy at the genome level. The methylation probability of single-molecule CpG sites on HiFi ZMWs, nonHiFi-Closed ZMWs, and nonHiFi-Open ZMWs was determined using the ZMWmeth model. To find the optimal trade-off between the number of CpG site reporters and predictive accuracy, this example presents five candidate strategies and trains and tests Joint-CpG model variants adapted to each strategy. Results show that with increasing sequencing depth (in single-molecule units), the accuracy of all tested strategies gradually improves, PCC increases, and MAE decreases. Furthermore, the Joint-CpG model trained with integrated data from all depth ranges (merging 5x, 7x, 10x, 15x, and 20x data) consistently reported the highest PCC and lowest MAE, and was selected as the default setting. The S1 and S5 strategies, using all molecules, had the highest CpG recall rates. The S3 strategy, using only HiFi reads, had the highest PCC, but recalled fewer CpG sites, clearly indicating that HiFi reads alone cannot achieve sufficient whole-genome coverage in low-depth sequencing designs. At 10x–20x single-molecule depths, the S4 and S5 strategies, using HiFi reads plus nonHiFi-ClosedZMW / nonHiFi-OpenZMW molecules from subreads, outperformed other strategies, achieving a PCC of 0.85 and recalling over 25 million CpG sites.

[0111] Currently, no dedicated software tools have been developed for processing low-depth long-read sequencing data with subreads. This example benchmarks existing software against LDLR-Meth (a low-depth long-read sequencing software). Specifically, a Joint-CpG model was retrained using single-molecule predictions from HKModel software as input. This hybrid combination of HKModel and LDLR-Meth demonstrates higher accuracy compared to HKModel (based on counting), showcasing the power and flexibility of LDLR-Meth. Figure 19As shown, across all tested coverage depths, LDLR-Meth outperformed other methods in terms of the number of recalled CpG sites, accuracy, and low MAE stratification based on the true DNA m frequency range. To further validate the performance of the low-depth solution in real-world scenarios, we performed low-depth CCS sequencing on four blood samples. The total data volume for these four samples was approximately 90 GB and 200 GB. Compared to the baseline DNA methylation frequency, the PCC was between 0.78 and 0.86, and the number of recalled CpG sites was approximately 23-26 million.

[0112] Example 3: Cross-platform testing of LDLR-Meth This embodiment demonstrates the performance of LDLR-Meth on different long-read sequencing platforms, including compatibility with the Nanopore platform and cross-platform testing between Nanopore and PacBio platforms.

[0113] Nanopore data compatibility. Publicly available Oxford Nanopore Technology (ONT) sequencing data for HG01109, HG02080, and HG02723 were downloaded from the T2T Diversity Panel, including FAST5 files (time-series signal data) and FASTQ files (sequence reads after base recognition). The methylation status of CpG sites at the read level was determined using Dorado software (dna_r9.4.1_e8_hac@v3.3 basecalling model). Then, the 5mC occurrence probability of CpG sites at the read level was extracted using Modkit software and formatted using LDLR-Meth. Subsequently, a Joint-CpG model was trained within the LDLR-Meth framework using CpG sites covered by 5x, 7x, 10x, 15x, and 20x ONT reads from the HG01109 and HG02723 samples. The Joint-CpG model was tested on HG02080.

[0114] Cross-platform testing results. In addition to the PacBio platform, LDLR-Meth also achieved low-depth coverage on the Oxford Nanopore Technology platform. At different coverage depths ranging from 2.5x to 20x, ONT reads showed higher prediction accuracy for PCC and MAE, although fewer CpG sites were detected at coverage depths of 2.5x–7.5x compared to the PacBio platform. On the PacBio platform, using HiFi reads resulted in higher accuracy compared to using the same number of ZMW molecules (including HiFi-ZMW and non-HiFi-ZMW), which is expected, as molecules with HiFi reads have higher single-molecule-level 5mC prediction accuracy than molecules without HiFi reads. Figure 15 A).

[0115] Example 4: Prediction at the Monotype Level This embodiment demonstrates the identification of DNA methylation sites at haplotype-resolved precision. High-depth sequencing subreads (>30X, in ZMW) of HG005 wild-type DNA were first used to generate HiFi-reads files using the bioinformatics method CCS. The HiFi reads were then aligned to the reference genome GRCh38 using pbmm2, resulting in an alignment BAM file. Subsequently, tools such as DeepVariant were used to detect genetic variation sites, which were stored as VCF files. Using the phase module of the WhatShap software, the HG005 genetic variation VCF file and the HiFi read alignment BAM file were taken as input, and the output was the phased VCF file, which distinguishes haplotypes. Using the haplotag module of WhatShap software, haplotype information was annotated on the HiFi reads using the aforementioned phased VCF file. Feature values ​​were then extracted and formatted to predict the methylation probability of each HiFi read. Finally, the DNA methylation level containing haplotype information was calculated.

[0116] This embodiment also provides a predictive model construction device for nucleic acid chemical modifications, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0117] This embodiment provides a device for constructing a predictive model of nucleic acid chemical modification, such as... Figure 9 As shown, it includes: The prediction sample acquisition module 901 is used to acquire a first nucleic acid sample, a second nucleic acid sample, and a third nucleic acid sample of the nucleic acid sample prediction object. The first nucleic acid sample is found to have undergone chemical modification through a first processing method, the second nucleic acid sample is found to have not undergone chemical modification through a second processing method, and the third nucleic acid sample is found to have undergone chemical modification of the nucleic acid molecule based on a preset site through a third processing method. The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are combined according to a preset sequencing template based on a bioinformatics method.

[0118] The nucleic acid feature determination module 902 is used to determine the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid sample prediction object recorded according to the long read sequencing method based on the preset sequencing depth.

[0119] The prediction model training module 903 is used to train based on the first nucleic acid feature, the second nucleic acid feature and the third nucleic acid feature, so as to obtain the first prediction model and the second prediction model. The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variable of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

[0120] In one optional implementation, the prediction sample acquisition module 901 includes: The nucleic acid sample acquisition submodule is used to acquire various types of raw nucleic acid samples, including: non-wild-type DNA nucleic acid samples and / or wild-type DNA nucleic acid samples or wild-type RNA nucleic acid samples and / or non-wild-type RNA nucleic acid samples; The nucleic acid sample preparation submodule is used to prepare various types of raw nucleic acid samples based on preset preparation methods. The sequencing library creation submodule is used to create nucleic acid sample sequencing libraries based on various types of prepared raw nucleic acid samples; The nucleic acid sample feature acquisition submodule is used to acquire the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample based on the nucleic acid sample sequencing library.

[0121] In one optional implementation, the nucleic acid sample feature acquisition submodule includes: The sequencing processing unit is used to sequence the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample according to the long read sequencing method and based on the preset sequencing depth; The feature extraction unit is used to extract features from the first, second, and third nucleic acid samples after sequencing based on the biological sample sequencing platform. The feature determination unit is used to obtain the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample based on the feature extraction results.

[0122] In one optional implementation, the feature extraction unit includes: The nucleic acid feature alignment subunit is used to determine the first nucleic acid feature based on the alignment result of the first nucleic acid sample and the first standard nucleic acid sample, to determine the second nucleic acid feature based on the alignment result of the second nucleic acid sample and the second standard nucleic acid sample, and to determine the third nucleic acid feature based on the alignment result of the third nucleic acid sample and the third standard nucleic acid sample.

[0123] In one alternative implementation, the prediction model training module 903 includes: The first prediction model optimization submodule is used to set the initial model parameters of the first prediction model based on the first input variable and to gradually optimize the initial first prediction model based on the first output variable to obtain the trained first prediction model during the process of obtaining the trained first prediction model.

[0124] In an optional implementation, the prediction model training module 903 further includes: The second prediction model optimization submodule is used to set the initial model parameters of the second prediction model based on the second input variable and to gradually optimize the initial second prediction model based on the second output variable to obtain the trained second prediction model during the process of obtaining the trained second prediction model.

[0125] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0126] In this embodiment, the predictive model building device for nucleic acid chemical modification is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0127] This embodiment also provides a device for predicting nucleic acid chemical modifications, such as... Figure 10 As shown, it includes: The sample acquisition module 1001 is used to acquire nucleic acid samples to be predicted.

[0128] The feature determination module 1002 is used to determine the features of the nucleic acid to be predicted based on the nucleic acid sample to be predicted. The features of the nucleic acid to be predicted include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to a preset window, recorded according to the long read sequencing method based on a preset sequencing depth.

[0129] The sample prediction module 1003 is used to input the nucleic acid features to be predicted into the first prediction model and the second prediction model respectively for prediction, obtain the first prediction result output by the first prediction model, and obtain the second prediction result output by the second prediction model; the first prediction model and the second prediction model are constructed by the prediction model construction method of nucleic acid chemical modification.

[0130] The probability result determination module 1004 determines the probability of chemical modification of the nucleic acid sample to be predicted based on the first prediction result, and determines the probability of chemical modification of the nucleic acid sample to be predicted based on the second prediction result.

[0131] In one optional implementation, the sample prediction module 1003 includes: The first prediction submodule is used to determine that the nucleic acid sample to be predicted has undergone nucleic acid chemical modification when the first prediction result is greater than or equal to a preset threshold. The second prediction submodule is used when the first prediction result is less than a preset threshold, indicating that the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0132] In one optional implementation, the sample prediction module 1003 includes: The second prediction submodule is used to indicate that the nucleic acid sample to be predicted has undergone nucleic acid chemical modification when the second prediction result is greater than or equal to a preset threshold. The second prediction submodule is used when the second prediction result is less than a preset threshold, indicating that the nucleic acid sample to be predicted has not yet undergone nucleic acid chemical modification.

[0133] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0134] In this embodiment, the nucleic acid chemical modification prediction device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0135] This invention also provides a computer device having the above-described nucleic acid chemical modification prediction apparatus. Please refer to... Figure 11 , Figure 11 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 11 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 11 Take a processor 10 as an example.

[0136] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0137] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0138] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0139] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0140] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0141] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0142] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0143] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for constructing a predictive model for nucleic acid chemical modifications, characterized in that, The method includes: The method involves obtaining a first nucleic acid sample, a second nucleic acid sample, and a third nucleic acid sample for nucleic acid sample prediction. The first nucleic acid sample is identified as having undergone chemical modification through a first processing method, the second nucleic acid sample is identified as not having undergone chemical modification through a second processing method, and the third nucleic acid sample is identified as having undergone chemical modification at a preset site through a third processing method. The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are combined according to a preset sequencing template using a bioinformatics approach. The first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample are determined. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include attribute information, enzyme kinetic information, and base sequence information of the nucleic acid sample prediction object recorded according to a preset sequencing depth using a long read sequencing method. Training is performed based on the first nucleic acid feature, the second nucleic acid feature and the third nucleic acid feature to obtain the first prediction model and the second prediction model; The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variables of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

2. The method according to claim 1, characterized in that, The first, second, and third nucleic acid samples for obtaining the nucleic acid sample prediction object include: Obtain various types of raw nucleic acid samples, including: non-wild-type DNA nucleic acid samples and / or wild-type DNA nucleic acid samples and / or wild-type RNA nucleic acid samples and / or non-wild-type RNA nucleic acid samples; The various types of original nucleic acid samples were prepared based on a preset preparation method; Based on the prepared raw nucleic acid samples of various different types, nucleic acid sample sequencing libraries were created. Based on the nucleic acid sample sequencing library, the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are obtained.

3. The method according to claim 1, characterized in that, Based on the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample, determine the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample, including: The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample were sequenced using a long-read sequencing method based on a preset sequencing depth. Feature extraction was performed on the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample after sequencing processing, based on a biological sample sequencing platform. Based on the feature extraction results, the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample are obtained.

4. The method according to claim 3, characterized in that, The feature extraction based on the biological sample sequencing platform for the first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample after sequencing processing includes: The first nucleic acid feature is determined based on the comparison result between the first nucleic acid sample and the first standard nucleic acid sample; the second nucleic acid feature is determined based on the comparison result between the second nucleic acid sample and the second standard nucleic acid sample; and the third nucleic acid feature is determined based on the comparison result between the third nucleic acid sample and the third standard nucleic acid sample.

5. The method according to claim 1, characterized in that, In the process of obtaining the first prediction model after training, the initial model parameters of the first prediction model are set based on the first input variable, and the initial first prediction model is gradually optimized based on the first output variable to obtain the first prediction model after training.

6. The method according to claim 1, characterized in that, In the process of obtaining the trained second prediction model, the initial model parameters of the second prediction model are set based on the second input variable, and the initial second prediction model is gradually optimized based on the second output variable to obtain the trained second prediction model.

7. A method for predicting nucleic acid chemical modifications, characterized in that, The method includes: Obtain the nucleic acid sample to be predicted; Based on the nucleic acid sample to be predicted, the characteristics of the nucleic acid to be predicted corresponding to the nucleic acid sample to be predicted are determined. The characteristics of the nucleic acid to be predicted include the attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to a preset window, recorded by long read sequencing method based on a preset sequencing depth. The nucleic acid features to be predicted are respectively input into a first prediction model and a second prediction model for prediction, and a first prediction result output by the first prediction model is obtained, and a second prediction result output by the second prediction model is obtained; the first prediction model and the second prediction model are constructed by the nucleic acid chemical modification prediction model construction method according to any one of claims 1 to 6; The probability of chemical modification of the nucleic acid sample to be predicted is determined based on the first prediction result, and the probability of chemical modification of the nucleic acid sample to be predicted is determined based on the second prediction result.

8. The method for predicting nucleic acid chemical modifications according to claim 7, characterized in that, The step of obtaining the first prediction result corresponding to the nucleic acid chemical modification output by the first prediction model includes: When the first prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone the nucleic acid chemical modification phenomenon; When the first prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone the nucleic acid chemical modification phenomenon.

9. The method for predicting nucleic acid chemical modifications according to claim 7, characterized in that, The step of obtaining the second prediction result corresponding to the nucleic acid chemical modification output by the second prediction model includes: When the second prediction result is greater than or equal to the preset threshold, the nucleic acid sample to be predicted has undergone the nucleic acid chemical modification phenomenon; When the second prediction result is less than the preset threshold, the nucleic acid sample to be predicted has not yet undergone the nucleic acid chemical modification phenomenon.

10. A device for constructing a predictive model for nucleic acid chemical modifications, characterized in that, The device includes: The prediction sample acquisition module is used to acquire a first nucleic acid sample, a second nucleic acid sample, and a third nucleic acid sample of the nucleic acid sample prediction object. The first nucleic acid sample is found to have undergone chemical modification through a first processing method, the second nucleic acid sample is found to have not undergone chemical modification through a second processing method, and the third nucleic acid sample is found to have undergone chemical modification of the nucleic acid molecule based on a preset site through a third processing method. The first nucleic acid sample, the second nucleic acid sample, and the third nucleic acid sample are combined according to a preset sequencing template based on a bioinformatics method. The nucleic acid feature determination module is used to determine the first nucleic acid feature corresponding to the first nucleic acid sample, the second nucleic acid feature corresponding to the second nucleic acid sample, and the third nucleic acid feature corresponding to the third nucleic acid sample. The first nucleic acid feature, the second nucleic acid feature, and the third nucleic acid feature include attribute information, enzyme kinetic information, and base sequence information of the nucleic acid sample prediction object recorded according to a preset sequencing depth in a long read sequencing method. The prediction model training module is used to train based on the first nucleic acid feature, the second nucleic acid feature and the third nucleic acid feature, thereby obtaining the first prediction model and the second prediction model; The training process of the first prediction model includes: taking the first nucleic acid feature and the second nucleic acid feature as the first input variables of the initial first prediction model, taking the feature result of the training output as the first output variable, and obtaining the trained first prediction model based on the correspondence between the first input variable and the first output variable. The training process of the second prediction model includes: using the third nucleic acid feature as the second input variable of the initial second prediction model, using the feature result of the training output as the second output variable, and obtaining the trained second prediction model based on the correspondence between the second input variable and the second output variable.

11. A device for predicting nucleic acid chemical modifications, characterized in that, The device includes: The sample acquisition module is used to acquire nucleic acid samples to be predicted. The feature determination module is used to determine the predictable nucleic acid features corresponding to the predictable nucleic acid sample based on the predictable nucleic acid sample. The predictable nucleic acid features include attribute information, enzyme kinetic information, and base sequence information of the nucleic acid molecule arranged upstream and downstream of each site according to a preset window, recorded according to a third sequencing method based on a preset sequencing depth. The sample prediction module is used to input the nucleic acid features to be predicted into a first prediction model and a second prediction model respectively, to obtain a first prediction result output by the first prediction model, and to obtain a second prediction result output by the second prediction model; the first prediction model and the second prediction model are constructed by the nucleic acid chemical modification prediction model construction method according to any one of claims 1 to 6; The probability result determination module determines the probability of the nucleic acid sample to be predicted undergoing chemical modification based on the first prediction result, and determines the probability of the nucleic acid sample to be predicted undergoing chemical modification based on the second prediction result.

12. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the steps of the method for constructing a predictive model of nucleic acid chemical modification as claimed in any one of claims 1 to 6, or the steps of the method for predicting nucleic acid chemical modification as claimed in any one of claims 7 to 9.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for constructing a predictive model of nucleic acid chemical modification as described in any one of claims 1 to 6, or the steps of the method for predicting nucleic acid chemical modification as described in any one of claims 7 to 9.

Images