N-terminal protein variant analysis method breaking through limitation of proteogenomics and de novo sequencing

Abstract

The invention provides a mass spectrum data analysis method of a new N-terminal protein variant, which breaks through the limitations of proteogenomics and de novo sequencing. A first specific property spectrum data analysis algorithm NovelNSeq specially aiming at a marker peptide fragment (a new N-terminal extension peptide fragment) of a new N-terminal protein is developed, and the algorithm does not depend on transcriptome data or de novo sequencing. By fully utilizing the peptide fragment coding logic, the accuracy is obviously higher than that of a mainstream de novo sequencing algorithm, and the method has the traceability of a peptide fragment coding mechanism. A plurality of new N-terminal protein variants coded by human genes are identified and verified in deep coverage mass spectrum data, the variants are not found in large-scale proteogenomics research of the same mass spectrum data, and NovelNSeq is proved to be capable of serving as an important supplement of a traditional method and is a powerful tool for finding new N-terminal proteins.

Description

Technical Field

[0001] This invention belongs to the field of protein analysis technology and relates to a method for the precise discovery of novel N-terminal protein variants. Specifically, it relates to a method for the precise discovery of N-terminal protein variants that overcomes the limitations of proteogenomics and de novo sequencing. Background Technology

[0002] The N-terminal sequence of proteins plays a crucial role in protein folding, stability, subcellular localization, and protein-protein interactions; variations in this sequence can lead to protein dysfunction and trigger various diseases. Our previous research found that the Nance-Horan syndrome-causing gene NHSL1 generates novel N-termini through alternative splicing, suggesting that alternative splicing is an important mechanism for the formation of novel N-termini. Recently, Rodriguez et al., through large-scale proteogenomic analysis, identified 192 5' untranslated region (5'UTR) translation events in the human genome, primarily producing protein variants with extended N-termini. These novel N-terminal protein variants exhibit low conservation across species and show significantly higher expression frequencies in cancer cells than in normal cells, suggesting a possible aberrant translation initiation regulatory mechanism in cancer cells. Since these N-terminal extensions mainly originate from highly expressed genes, and proteomics detection is more likely to capture high-abundance proteins, the currently identified N-terminal extensions may only represent a small fraction of the 5'UTR translation products.

[0003] Currently, using mass spectrometry data to search predicted protein libraries constructed from whole-genome or transcriptome data is the main strategy for discovering novel N-terminal proteins. However, transcriptome sequencing also has limitations in sensitivity and specificity. This proteogenomics strategy relies on databases and cannot resolve novel N-terminal proteins not included in the database but detected by mass spectrometry. De novo sequencing, which does not rely on databases, is an effective complementary strategy to overcome these limitations. In the past two decades, various de novo sequencing algorithms based on mathematical models such as graph theory and dynamic programming have been developed. In recent years, deep learning technology has driven the emergence of the current state-of-the-art de novo sequencing algorithms. However, the accuracy of de novo sequencing is highly dependent on the degree of peptide fragmentation. Even advanced de novo sequencing algorithms still have a peptide recall rate of less than 40% on omics-scale datasets, which is significantly lower than that of database searches. Furthermore, inferring the genome coding mechanism of de novo sequencing results containing erroneous sites still faces significant challenges.

[0004] To address the aforementioned issues, the inventors' research group developed the semi-de novo sequencing algorithm SpliceNovo, which significantly enhances the ability to discover novel protein variants through de novo sequencing. However, SpliceNovo, which integrates the independently developed de novo sequencing algorithm CNovo, still cannot avoid the limitations of low accuracy in de novo sequencing and difficulties in tracing peptide coding mechanisms. Summary of the Invention

[0005] This invention addresses the aforementioned challenges by developing NovelNSeq, a specific algorithm for resolving novel N-terminal extension peptides (NEPs), specifically targeting their marker peptides—novel N-terminal extension peptides (NEPs). This algorithm does not rely on transcriptome data but directly resolves NEPs from mass spectrometry data. NovelNSeq fully utilizes the coding logic of NEPs, achieving significantly higher accuracy than existing tools such as SpliceNovo, and also possesses the function of tracing coding mechanisms. Using NovelNSeq to resolve high-coverage mass spectrometry data, we successfully discovered and validated novel N-termini encoded by human genes CALM2, CAPNS1, and CPNE7, demonstrating that NovelNSeq is a powerful tool for discovering novel N-terminal proteins.

[0006] The specific technical solution of the present invention is as follows:

[0007] This invention provides a method for analyzing N-terminal protein variants that overcomes the limitations of proteogenomics and de novo sequencing.

[0008] Includes the following steps:

[0009] I. Construction of a database of theoretical N-terminal enzyme-digested peptides from the proteome

[0010] For an annotated human protein-coding sequence, if it consists of a single, continuous coding DNA sequence (CDS) on a chromosome, all theoretically cleavable N-terminal peptides of the protein can be predicted using theoretical enzymatic digestion methods. If an annotated human protein-coding sequence consists of multiple CDSs separated by introns on a chromosome, in addition to theoretical enzymatic digestion of the complete protein, theoretical enzymatic digestion of the amino acid sequence encoded by the remaining CDSs after removing the first CDS at the 5' end must also be considered. For the second CDS at the 5' end, there may be different translation initiation shifts, with values ​​of 0, 1, or 2, representing the start of translation from the first, second, or third base at the 5' end of the CDS, respectively.

[0011] In this study, a comprehensive database of N-terminal theoretical cleavage peptides was constructed based on all annotated human protein-coding sequences in NCBI, following the method described above. The considered N-terminal theoretical cleavage peptides ranged in length from 5 to 20 amino acids, and a maximum of two missed cleavage events were allowed. A dictionary was used to record each N-terminal theoretical cleavage peptide and all its corresponding protein information. The dictionary keys were N-terminal theoretical cleavage peptides, and the values ​​were sets containing information on all proteins that could have produced that peptide. Each protein information entry was represented as a string in the format: "chromosome|gene name|protein ID|positive or negative strand|position of the first 5' base of the first (or second) CDS on the positive strand of the chromosome|translation initiation shift of that CDS". This dictionary facilitates the tracing of peptide origins.

[0012] II. Acquisition of N-terminal peptides

[0013] Based on all annotated human protein-coding sequences in the NCBI database, a comprehensive database of N-terminal theoretically digestible peptides was constructed. For proteins encoded by a single CDS, theoretical digestion was performed directly to obtain the N-terminal peptide. For protein-coding sequences composed of multiple CDS, theoretical digestion was performed on the encoded amino acid sequence to obtain the N-terminal peptide, with and without the first CDS at the 5' end.

[0014] 3. Calculate the matching score between the secondary spectrum and the sequence in the N-terminal theoretical enzyme-digested peptide database, and screen out the top N incomplete peptides with the highest scores.

[0015] Using the shape of The two-dimensional array is used as the input model, denoted as . First, two convolutional layers are used to perform a two-step convolution operation on the input array to learn the scoring weights of different types of ions and the weights of the total and average ion matching scores in recalling the correct peptides. Then, the SoftMax function is used to convolve the array obtained after the convolution operation. Convert to an output array with values ​​between 0 and 1 Based on a scoring function, the spectrum and incomplete peptide segments are compared. The matching results are scored. The specific process is as follows:

[0016] For positive and negative samples, the composition of the elements of the two-dimensional array is defined by the following formulas (9) and (10), respectively:

[0017]

[0018]

[0019] in, and ;

[0020] The input array is subjected to a two-step convolution operation using two convolutional layers: First, a one-dimensional convolutional kernel is used. and a bias term , input array Convert to a new two-dimensional array The convolution stride is 21, and the specific details of the convolution calculation are shown in the following formula (11):

[0021]

[0022] in, and ;

[0023] Then, use a one-dimensional convolution kernel. and bias terms , array Convert to a new array The convolution stride is 2, and the specific details of the convolution calculation are shown in formula (12):

[0024]

[0025] By summing the array Convert to a new one-dimensional array For positive and negative samples, These are represented by the following formulas (13) and (14), respectively:

[0026]

[0027]

[0028] Finally, the SoftMax function is used to convert the array... Convert to an output array with values ​​between 0 and 1 Assuming the spectrum and incomplete peptide segments The matching results are scored, and the scoring function designed in Model 2 is expressed by the following formula (15):

[0029] .

[0030] This invention compares this scoring algorithm with other scoring algorithms. Given the significant differences in the length of incomplete peptides, relying solely on the total score of daughter ion matching may result in longer erroneous peptides receiving higher total scores due to their length advantage. This second scoring algorithm considers the influence of peptide length, thus exhibiting a significant advantage in peptide recall.

[0031] Fourth, for each incomplete peptide segment, possible complete new N-terminal extended peptide segments (NEPs) are enumerated by a pre-constructed dictionary. Then, 12,000 bases are extended to the 5' end of the annotated CDS encoding the incomplete peptide segment to search for possible new CDSs. This ensures that the mass of the commonly encoded peptide segment is equal to that of the precursor ion in the spectrum. All possible NEPs corresponding to the secondary spectrum are obtained by this method.

[0032] 5. Perform peptide matching and scoring on all possible NEPs to obtain the best matching result, and provide the newly resolved peptide segment and the coordinates of the CDS encoding the new peptide segment on the chromosome.

[0033] The beneficial protections and effects of this invention are as follows:

[0034] This invention addresses the limitations of traditional proteogenomics methods in identifying novel N-terminal proteins not detected by transcriptome data, and the technical shortcomings of de novo sequencing in terms of accuracy and traceability of coding mechanisms. To address these limitations, the invention develops NovelNSeq, the first mass spectrometry data analysis algorithm specifically targeting marker peptides (novel N-terminal extension peptides) for novel N-terminal proteins. It does not rely on transcriptome data or de novo sequencing, fully utilizes peptide coding logic, and achieves significantly higher accuracy than mainstream de novo sequencing algorithms such as PEAKS, pNovo3, SpliceNovo, Casanovo, and InstaNovo, while also possessing traceability capabilities for peptide coding mechanisms.

[0035] Using NovelNSeq, we identified and validated novel N-terminal protein variants encoded by the human genes CALM2, CAPNS1, and CPNE7 in deep-coverage mass spectrometry data. These variants were not found in large-scale proteogenomic studies using the same mass spectrometry data, demonstrating that NovelNSeq can serve as an important complement to traditional methods. Furthermore, we revealed that the novel N-terminus in the recently reported human ATP9A gene, initiated by AAG, is actually generated by translation initiation via a classical ATG start codon and frameshift. This highlights the necessity of rigorous validation of non-classical start codon annotations in novel N-terminal protein variants, confirming that NovelNSeq is a powerful tool for discovering novel N-terminal proteins. Attached Figure Description

[0036] Figure 1The training models of two scoring algorithms for matching secondary spectra with incomplete peptides are shown: (a) a flowchart of generating positive and negative samples based on secondary spectra; enumerating the C-terminal amino acid sequence of the correct peptide, generating possible incorrect peptides by replacing the first three amino acids, defining (correct incomplete peptide, incorrect incomplete peptide) as a positive sample, and (incorrect incomplete peptide, correct incomplete peptide) as a negative sample; (b) a schematic diagram of the convolutional neural network model and training process of scoring algorithm one; (c) a schematic diagram of the convolutional neural network model and training process of scoring algorithm two.

[0037] Figure 2 This diagram illustrates four scenarios indicating the presence of a novel N-terminal sequence indicative of a protein, where translation is initiated by a new translation start site: n1 and n2 represent the two terminal positions of the novel CDS on the chromosome. a1 and a2 represent the two terminal positions of the first 5' CDS in the protein-coding sequence on the chromosome. a3 and a4 correspond to the two terminal positions of the second 5' CDS in the protein-coding sequence on the chromosome. The symbols "+" and "-" indicate that the protein-coding sequence is located on the positive and negative strands of the chromosome, respectively. In the amino acid sequence "ASQRVPATSQATAEFKFDKDG", the red and blue regions correspond to fragments encoded by the novel CDS and the annotated CDS, respectively. NEPs are specific protease-cleaved peptides co-encoded by the first or second annotated 5' CDS and the novel CDS at the positions shown in the diagram.

[0038] Figure 3 The diagram shows the principle of the NovelNSeq algorithm for resolving NEPs in mass spectrometry data.

[0039] Figure 4 The accuracy comparison of matching and scoring algorithms 1 and 2 for secondary spectra and incomplete peptides is shown. The cumulative recall curves of peptides ranked first to tenth in test sets 1-3 for both algorithms are also shown.

[0040] Figure 5The accuracy comparison between NovelNSeq and mainstream de novo sequencing algorithms in NEP identification is shown: (a) Comparison of peptide recall rates of the top-1 results reported by PEAKS, pNovo3, SpliceNovo, Casanovo, InstaNovo, and NovelNSeq in the three test sets SD1-3. (b) Comparison of amino acid recall rates of the top-1 results reported by PEAKS, pNovo3, SpliceNovo, Casanovo, InstaNovo, and NovelNSeq in the three test sets SD1-3. (c) Comparison of amino acid precision rates of the top-1 results reported by PEAKS, pNovo3, SpliceNovo, Casanovo, InstaNovo, and NovelNSeq in the three test sets SD1-3.

[0041] Figure 6 This demonstrates a quality control method for obtaining candidate novel N-terminal extended peptides using NovelNSeq mass spectrometry data.

[0042] Figure 7 This section demonstrates the validation of the N-terminus of a novel protein encoded by the human CALM2 gene: (a) Comparison of the original and synthesized spectra of the peptide “RPGMPATVVSQADQLTEEQIAEFK”; (b) Comparison of the resolution results of the original peptide spectra by different software. Amino acids highlighted in red indicate sites with incorrect resolution, and “AA recall rate” represents the ratio of correctly resolved amino acid sites to the actual peptide length; (c) The CDS composition of the novel N-terminal protein encoded by the CALM2 gene on the negative strand of chromosome 2 and its sequence alignment with homologous proteins from other species. The values ​​in the “Identities (%)” row represent the similarity of amino acid sequences within the aligned region.

[0043] Figure 8 The validation of the N-terminus of the novel protein encoded by the human CAPNS1 gene is shown: (a) Comparison of the original spectrum and the synthetic spectrum of the peptide “ALVMLALLSEAAAQYNPEPPPPR”; (b) The top 5 alignment results of the peptide “ALVMLALLSEAAAQYNPEPPPPR” with the highest scores by NCBI BLASTP analysis; (c) The position of the coding sequence of the peptide “ALVMLALLSEAAAQYNPEPPPPR” on the positive strand of chromosome 19; (d) Comparison of the interpretation results of the original spectrum of the peptide “ALVMLALLSEAAAQYNPEPPPPR” by different software.

[0044] Figure 9The validation of the N-terminus of the novel protein encoded by the human CPNE7 gene is shown: (a) Comparison of the original spectrum and the synthetic spectrum of the peptide “LAGSGPSVRPAGPPNAGSMSAGSER”; (b) The top 5 alignment results of the peptide “LAGSGPSVRPAGPPNAGSMSAGSER” with the highest scores in NCBI BLASTP analysis; (c) The position of the coding sequence of the peptide “LAGSGPSVRPAGPPNAGSMSAGSER” on the positive strand of chromosome 16; (d) Comparison of the interpretation results of the original spectrum of the peptide “LAGSGPSVRPAGPPNAGSMSAGSER” by different software.

[0045] Figure 10 This demonstrates how translation frameshift mediates the formation of a new N-terminus in the human ATP9A protein.

[0046] Figure 11 The validation of the new peptide “LRAWPGAR” is shown: (a) comparison of the original spectrum of peptide “LRAWPGAR” with the spectrum of the synthesized peptide; (b) the top 5 comparison results of peptide “LRAWPGAR” with the highest scores by NCBI BLASTP analysis. Detailed Implementation

[0047] The following embodiments and experimental examples further illustrate the present invention and should not be construed as limiting the invention. The embodiments do not include a detailed description of conventional methods, which are well known to those skilled in the art and described in numerous publications.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as are familiar to those skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein may be applied to this invention, and the preferred embodiments and materials described in the specific implementation are for illustrative purposes only.

[0049] Unless otherwise specified, the experimental methods described in the following examples are generally performed under standard conditions or as recommended by the manufacturer.

[0050] I. Experimental Methods

[0051] 1. Construction of a database of theoretical N-terminal enzyme-digested peptides from the proteome

[0052] For an annotated human protein-coding sequence, if it consists of a single, continuous coding DNA sequence (CDS) on a chromosome, all theoretically cleavable N-terminal peptides of the protein can be predicted using theoretical enzymatic digestion methods. If an annotated human protein-coding sequence consists of multiple CDSs separated by introns on a chromosome, in addition to theoretical enzymatic digestion of the complete protein, theoretical enzymatic digestion of the amino acid sequence encoded by the remaining CDSs after removing the first CDS at the 5' end must also be considered. For the second CDS at the 5' end, there may be different translation initiation shifts, with values ​​of 0, 1, or 2, representing the start of translation from the first, second, or third base at the 5' end of the CDS, respectively.

[0053] In this study, a comprehensive database of N-terminal theoretical cleavage peptides was constructed based on all annotated human protein-coding sequences in NCBI, following the method described above. The considered N-terminal theoretical cleavage peptides ranged in length from 5 to 20 amino acids, and a maximum of two missed cleavage events were allowed. A dictionary was used to record each N-terminal theoretical cleavage peptide and all its corresponding protein information. The dictionary keys were N-terminal theoretical cleavage peptides, and the values ​​were sets containing information on all proteins that could have produced that peptide. Each protein information entry was represented as a string in the format: "chromosome|gene name|protein ID|positive or negative strand|position of the first 5' base of the first (or second) CDS on the positive strand of the chromosome|translation initiation shift of that CDS". This dictionary facilitates the tracing of peptide origins.

[0054] 2. Scoring of mass spectra matching incomplete peptide segments

[0055] 2.1 Algorithm 1 Design and Parameter Training

[0056] Given a mass spectrum, assume the correct peptide sequence is NPTQYDSTHGR. First, enumerate the C-terminal 5 amino acids (STHGR), 6 amino acids (DSTHGR), and so on, until... 1 amino acid (PTQYDSTHGR), of which This indicates the length of the correct peptide. Subsequently, for each enumerated incomplete peptide, such as STHGR, by replacing its first three amino acids with any combination of amino acids with a mass deviation within 20 parts per million, erroneous peptides such as MPPGR, HSTGR, and GNGPGR, which may be confused with STHGR during peptide matching, can be generated. The pairing of (correct incomplete peptide, erroneous incomplete peptide) is defined as a positive sample, and the pairing of (erroneous incomplete peptide, correct incomplete peptide) is defined as a negative sample. Based on this definition, several positive and negative samples can be generated from a single mass spectrum. Figure 1 a).

[0057] Positive samples can be represented as ( , Negative samples are represented as ( , In this notation, ( () represents the correct but incomplete peptide segment starting from the C-terminus. One amino acid residue, This is the length of the correct but incomplete peptide segment; similarly, ( () indicates an incorrect or incomplete peptide segment starting from the C-terminus. One amino acid residue, This represents the length of the erroneous, incomplete peptide segment. In the positive sample representation, Include The first fragmentation site. Specifically, the first... The fragmentation sites are located at and Between, among Furthermore, the first The fragmentation sites are located in the unknown prefix sequence and Between. Similarly, in the negative sample representation, Include The fragmentation site. Among them, the first... The fragmentation sites are located at and between, Furthermore, the first The fragmentation sites are located in the unknown prefix sequence and between.

[0058] For each fragmentation site, the theoretical mass-to-charge ratio of 21 fragment ions was calculated, including: b, y, a, b(2+), y(2+), a(2+), b-H2O, y-H2O, a-H2O, b-NH3, y-NH3, a-NH3 ions, and by-type internal ions with lengths from 2 to 10. Because and For the C-terminal peptide, the theoretical masses of the C-terminal series of daughter ions can be calculated precisely. However, the theoretical masses of the N-terminal series of daughter ions can only be estimated by combining the mass of the parent ion in the spectrum with that of a known sequence. Therefore, when searching for the closest matching peak to each daughter ion's mass in the secondary spectrum, the maximum allowable mass deviations for the C-terminal and N-terminal series of daughter ions were set to 20 ppm and 0.02 Da (empirically determined), respectively. The matching degree of each daughter ion in the spectrum was scored, and the score value was defined as... This can be expressed by Formula 1:

[0059]

[0060] in," "This indicates that the daughter ion has a matching peak in the spectrum." "Indicates that there is no matching peak. This indicates the intensity ranking of the matched peak among all spectral peaks in the spectrum. This represents the mass deviation between the daughter ion and the matching peak. Further, it defines... and They are respectively and The Middle The first fragmentation site fragment ions value.

[0061] By training a convolutional neural network for binary classification (denoted as Model 1), a scoring function 1 is obtained between the spectral image and the incomplete peptide segment. Figure 1 b). The input to Model 1 is a shape of A two-dimensional array, denoted as The first dimension represents 21 theoretical fragment ions. For positive and negative samples, the element composition of this two-dimensional array is represented by the following formulas 2 and 3, respectively:

[0062]

[0063]

[0064] in, , .

[0065] Then, a one-dimensional convolution kernel is used. and a bias term , input array Convert to a new array The convolution stride is 21. Details of the convolution calculation are shown in Formula 4 below.

[0066]

[0067] By summing the array Convert to a new one-dimensional array For positive and negative samples, These can be represented by the following formulas 5 and 6 respectively:

[0068]

[0069]

[0070] Finally, the SoftMax function is used to process the array. Convert to an output array with values ​​between 0 and 1 (Formula 7), which is the probability that the sample is a positive sample.

[0071]

[0072] ConvNetSharp, a convolutional neural network framework https: / / github.com / cbovar / ConvNetSharp This was used to build and train Model 1. The goal of model training was to correctly distinguish between positive and negative samples. It was assumed that the model would differentiate between spectra and incomplete peptides. The matching results are scored, and the scoring function designed according to Model 1 is expressed by the following formula 8:

[0073]

[0074] 2.2 Design and Parameter Training of Algorithm 2

[0075] By training another convolutional neural network for binary classification (denoted as Model 2), a scoring function 2 is obtained between the spectral image and the incomplete peptide segment. Figure 1 c). Peptide sequence and and the corresponding matrix and Their definitions are consistent with those above. The input to Model 2 is a shape... A two-dimensional array, denoted as For positive and negative samples, the composition of the elements of this two-dimensional array is defined by the following formulas 9 and 10, respectively:

[0076]

[0077]

[0078] in, and .

[0079] The input array is subjected to a two-step convolution operation using two convolutional layers. First, a one-dimensional convolutional kernel is used. and a bias term , input array Convert to a new two-dimensional array The convolution stride is 21. The specific details of the convolution calculation are shown in Formula 11 below:

[0080]

[0081] in, and Then, a one-dimensional convolution kernel is used. and bias terms , array Convert to a new array The convolution stride is 2. The specific details of the convolution calculation are shown in Equation 12:

[0082]

[0083] By summing the array Convert to a new one-dimensional array For positive and negative samples, These are represented by the following formulas 13 and 14 respectively:

[0084]

[0085]

[0086] Finally, the SoftMax function is used to convert the array... Convert to an output array with values ​​between 0 and 1 Assuming the spectrum and incomplete peptide segments... The matching results are scored, and the scoring function designed in Model 2 is expressed by the following formula 15:

[0087]

[0088] 2.3 Performance Evaluation of Two Scoring Algorithms

[0089] Using pFind 22 The search parameters are detailed in Table 1. (Version 3.1.6) and PEAKS (Version PEAKS X) were used to search four sets of high-resolution HCD mass spectrometry data acquired by the Thermo Scientific Q-Exactive mass spectrometer.

[0090] Table 1 Database search parameters for pFind and PEAKS

[0091]

[0092] The peptide profiling matches (PSMs) jointly identified by the two software programs in each dataset were used as the baseline dataset. One set was used to construct the training set, and the other three sets were used to construct the test set (Table 2). Based on the baseline dataset, positive and negative samples were generated in the training and test sets according to the methods in Section 2 of the Materials and Methods section. Figure 1 a) In the test set, the correct and incorrect incomplete peptides corresponding to each spectrum are merged and matched for scoring. The recall rate of the top N peptides is defined as the proportion of the number of correctly recalled incomplete peptides within the top N to the total number of correctly recalled incomplete peptides in the test set, to evaluate the performance of the scoring algorithm.

[0093] Table 2. Mass spectrometry data used for the training and testing sets of the algorithm for matching and scoring secondary spectra and incomplete peptides.

[0094]

[0095] 2.4 Accuracy Assessment of NovelNSeq

[0096] Since a large-scale NEP mass spectrometry dataset does not exist in reality, this study uses simulated data to evaluate the accuracy of NovelNSeq. We assume that the three PSMs with IDs 2, 3, and 4 in Table 2 are from NEP, and their corresponding spectra are denoted as SD1, SD2, and SD3, respectively. The C-terminal six amino acids of the actual peptide in each spectra are denoted as PC, and the remaining N-terminal amino acid sequence is denoted as PN. We assume that PC is an incomplete C-terminal sequence encoded by a known CDS in NEP, and PN is an extended sequence encoded by a new CDS in NEP. When NovelNSeq resolves a spectra, it first resolves the incomplete C-terminal sequence by searching a database consisting of all theoretically digested N-terminal peptides and PC. Subsequently, the coding sequence of the lysine and PN combination sequence is randomly inserted into a 12,000-base-length nucleic acid sequence randomly selected from the human genome. If NovelNSeq can correctly resolve the PC, it further resolves the extended sequence from this nucleic acid sequence.

[0097] In addition to NovelNSeq, PEAKS, pNovo3, SpliceNovo, Casanovo, and InstaNovo were also used to resolve SD1, SD2, and SD3. All software used the same search parameters: a maximum allowable range of ±15 ppm for parent ion mass deviation; cysteine ​​alkylation was set as a fixed modification; and methionine oxidation was set as a variable modification. For the maximum allowable range of daughter ion mass deviation, PEAKS and SpliceNovo were set to ±0.02 Da, and pNovo3 and NovelNSeq were set to ±20 ppm. Casanovo and InstaNovo used the default parameters of their publicly available pre-trained models.

[0098] Three metrics were used to evaluate the accuracy of the software: peptide recall, amino acid recall, and amino acid precision. Peptide recall was the percentage of correctly recalled peptides out of the total number of peptides in the test set. Amino acid recall was the percentage of correctly recalled amino acids out of the total number of amino acids in the test set. Amino acid precision was the percentage of correctly recalled amino acids out of the total number of amino acids reported by the software.

[0099] II. Results and Discussion

[0100] 1. Identification of novel N-terminal extension peptides is key to revealing the novel N-terminus of proteins.

[0101] The N-terminal sequence of a protein usually refers to a number of amino acids starting from the N-terminus; however, there is no universally accepted standard for defining exactly how many amino acids constitute the N-terminal sequence. The novel N-terminal sequences discussed in this article specifically refer to the N-terminal sequences of proteins whose translation is initiated by newly discovered translation start sites in the genome. If a gene can encode a protein variant with a novel N-terminal sequence, it indicates the existence of at least one new CDS (Copyright Decision Group), whose relative position on the chromosome to the gene's already annotated coding sequence has multiple possibilities. For example, the novel CDS might be located at the 5' end of the first or second CDS in a known protein coding sequence, with a gap of at least 0 bases between them, and both encoding a continuous sequence within the novel N-terminal sequence (…). Figure 2 ).

[0102] In human genes with different translation initiation sites, 97.9% of genes produce protein variants with different N-terminal sequences due to the two situations mentioned above. Therefore, these two situations are important mechanisms for N-terminal protein variation. Of course, the new CDS mentioned here is not necessarily the first CDS at the 5' end of the coding sequence of the new protein variant. The N-terminal sequence of a new protein variant may also be encoded by multiple new CDS. However, the amino acid sequence co-encoded by multiple new CDS cannot directly prove that a known protein has undergone N-terminal variation. Figure 2 The NEPs co-encoded by the new CDS and the annotated CDS can directly demonstrate N-terminal variations in known proteins. Therefore, NEPs are marker peptides characterizing the presence of novel N-terminal protein variants. Enumeration was performed for every protein in the human proteome within a range of 12,000 bases. Figure 2 The number of new CDS with lengths of 9 to 60 bases will exceed 5 billion. Therefore, it is not feasible to use traditional database search algorithms to resolve NEPs in mass spectrometry data.

[0103] 2. NovelNSeq algorithm for analyzing NEPs in mass spectrometry data

[0104] This study developed an algorithm specifically for resolving NEPs in mass spectrometry data, namely NovelNSeq. Figure 3 The core principles were demonstrated. The algorithm consists of five key steps:

[0105] (a) Secondary spectrum search for N-terminal theoretical restriction peptide database. For protein coding sequences composed of multiple discontinuous CDS, theoretical restriction enzyme digestion was performed on the encoded amino acid sequence in two cases: retaining and removing the first CDS at the 5' end, to obtain the N-terminal peptide.

[0106] (b) Algorithm for matching secondary spectra with incomplete peptides. This algorithm constructs a two-dimensional array representing the theoretical daughter ion matching of all fragmentation sites, performs two-step convolution operations using two convolutional layers, and finally sums the output array to obtain the matching score.

[0107] (c) Screening out the top N incomplete peptides by score. Multiple entries under each peptide represent multiple potential sources of that peptide. For example, “Chr II | - | 47170764 | 0” indicates that the coding sequence of the peptide ADQLTEEQIAEFK is initiated by a CDS from the negative strand of chromosome II, the first base of which is located at coordinate 47170764 on the positive strand of the chromosome, and the translation initiation shift is 0.

[0108] (d) For each incomplete peptide, enumerate possible complete NEPs. Extend 12,000 bases to the 5' end of the annotated CDS encoding the incomplete peptide to search for possible new CDS, ensuring that the co-encoded peptide has the same mass as the precursor ion in the spectrum.

[0109] (e) Perform peptide matching and scoring on all possible NEPs to obtain the best matching result. The NovolNSeq algorithm not only provides the resolved new peptides, but also provides the coordinates of the CDS encoding the new peptides on the chromosome.

[0110] The specific analysis method is as follows:

[0111] First, based on all annotated human protein-coding sequences in the NCBI database, a comprehensive database of N-terminal theoretical enzyme-digested peptides was constructed. Figure 3 a) For proteins encoded by a single CDS, theoretical digestion was performed directly to obtain the N-terminal peptide. For protein-coding sequences composed of multiple CDS, theoretical digestion was performed on the encoded amino acid sequence to obtain the N-terminal peptide, with and without the first CDS at the 5' end. Specific operational details are provided in Section 1 of the Materials and Methods section. Digestion with Trypsin, GluC, and LysC yielded 59947, 38049, and 30892 theoretically digestible N-terminal peptides, respectively.

[0112] Then, the matching score between the secondary spectrum and the sequence in the N-terminal theoretical restriction peptide database is calculated. Figure 3 b), and select the top N peptides by score ( Figure 3 c). These peptides may be the C-terminal sequences of the correct peptides, but are defined as incomplete peptides due to the lack of the N-terminal region.

[0113] This paper proposes two scoring algorithms for matching secondary spectra with incomplete peptide segments, denoted as Scoring Algorithm 1 and Scoring Algorithm 2, respectively. Specific formulas and parameter training methods are detailed in the Experimental Methods section above. The following is a brief introduction to Scoring Algorithm 2: Secondary spectra do not undergo preprocessing steps such as removing isotope peaks and low-abundance noise peaks to prevent the loss of daughter ion signal peaks. For a length of... Incomplete peptide segments exist. There are 21 theoretical daughter ions, including the by-type internal ion, at each fragmentation site. The matching score of each theoretical daughter ion in the secondary spectrum is calculated using an empirical formula (Formula 1), which takes into account the role of ion intensity and mass deviation in reflecting the true matching.

[0114] The scoring algorithm 2 uses a convolutional neural network with two convolutional layers. The size of the input array is... Each row corresponds to a fragmentation site. The first 21 columns represent the matching scores of seed ions 1-21, and the last 21 columns represent the values ​​of these matching scores divided by the length of the incomplete peptide segment. This two-dimensional array represents the theoretical seed ion matching status of all fragmentation sites, and then performs a two-step convolution operation through two convolutional layers. The kernel size of the first convolutional layer is [size missing]. This is primarily used to learn the scoring weights for different types of daughter ions. The kernel size of the second convolutional layer is... This is primarily used to learn the weights of the total and average values ​​of the daughter ion matching scores in the recall of correct peptides. After two convolutional operations, the size of the output array is... Summing these values ​​yields the matching score between the secondary spectrum and the incomplete peptide segment. Figure 3 d). The size of the input array for scoring algorithm one is... Each row represents a fragmentation site, and columns 1-21 represent the matching scores of the 1st to 21st theoretical daughter ions, respectively. Through a... After the convolution kernel performs the convolution operation, the summation of the output array yields the matching score. Therefore, scoring algorithm two, compared to scoring algorithm one, additionally considers the impact of peptide length on the recall rate of the scoring algorithm.

[0115] The peptide recall rates of the two scoring algorithms were compared across three test sets. The results show that scoring algorithm 2 outperformed scoring algorithm 1 in recall rates for the top ten peptides. Specifically, the recall rate of the top peptide from scoring algorithm 2 was 8.2%, 7.2%, and 7.4% higher than that of scoring algorithm 1, respectively. For the top five peptides, scoring algorithm 2 was 1.3%, 0.8%, and 1.2% higher, respectively. Furthermore, in all three test sets, the recall rate of the top five peptides from scoring algorithm 2 was above 93%. Figure 4Given the significant differences in the length of incomplete peptides, relying solely on the total score for daughter ion matching may result in longer, erroneous peptides receiving higher total scores due to their length advantage. Scoring Algorithm 2 considers the influence of peptide length, thus exhibiting a significant advantage in peptide recall.

[0116] Next, for each selected N-terminal theoretically cleaved peptide, its possible origins are queried using a pre-constructed dictionary. If the N-terminal theoretically cleaved peptide is encoded by a CDS in a protein-coding sequence, the base sequence in that CDS that is not used to encode the peptide is defined as follows, based on the annotated translation initiation shift. Analysis of all CDS intervals in the human genome revealed that 90% are less than 12,000 bases. Therefore, extending 12,000 bases to the 5' end of this CDS sequence was used to search for all matching subnucleotide sequences. While longer extended sequences may help discover more candidate sites, this study empirically determined that a threshold of 12,000 bases achieves the optimal balance between discovery sensitivity and computational burden. and The new peptide is formed by combining the amino acid sequence from the combined translation with the theoretically N-terminal enzyme-cleaved peptide. neutral quality Should meet ,in It is the maximum permissible quality deviation ( Figure 3 d). In addition, sub-nucleic acid sequences The preceding codon must be a stop codon or encode a specific restriction enzyme site. By following these steps, all possible NEPs corresponding to the secondary spectrum can be obtained.

[0117] Finally, peptide matching scores were performed on all possible NEPs to obtain the best matching result. Figure 3 e). The scoring here uses the scoring function from the self-developed CNovo algorithm. Unlike traditional scoring algorithms that only provide the resolved peptides, NovolNSeq also provides the coordinates of the CDS encoding the peptides on the chromosome. Traditional de novo sequencing results often have some correct amino acids and some incorrect amino acids, making it difficult to infer the origin of the peptides. NovolNSeq effectively solves the problem of peptide origin tracing. NovolNSeq is developed in C# and is specifically designed for parsing high-resolution high-energy collision dissociation (HCD) spectra acquired in data-dependent acquisition (DDA) mode. This tool is available on GitHub (…). https: / / github.com / CProteomics / NovelNSeq ).

[0118] 3. NovelNSeq's accuracy in identifying NEPs is superior to mainstream de novo sequencing algorithms.

[0119] Given the lack of large-scale NEP mass spectrometry datasets in reality, this study uses simulated data to evaluate the NEP identification accuracy of NovelNSeq and mainstream de novo sequencing algorithms. Among the three simulated datasets (SD1, SD2, and SD3), compared with PEAKS... 11 pNovo3 13 SpliceNovo 21 Casanovo 18 and InstaNovo 19 In comparison, NovelNSeq has a higher recall rate at the peptide level ( Figure 5 a) Amino acid level recall ( Figure 5 b) and amino acid level accuracy ( Figure 5 In all aspects (c), NovelNSeq demonstrated superior performance. Specifically, in SD1, SD2, and SD3, NovelNSeq's peptide recall was 12.5%, 13.0%, and 12.2% higher than InstaNovo (the best-performing de novo sequencing software), respectively. These results strongly demonstrate that NovelNSeq has higher accuracy in NEP identification than traditional de novo sequencing algorithms.

[0120] 4. NovelNSeq identified 3 real human NEPs.

[0121] NovelNSeq was used to analyze the human proteome depth coverage mass spectrometry dataset with identifier PXD004452 in the ProteomeXchange database. Due to the high false positive rate of the obtained candidate NEPs, a rigorous quality control process was established in this study. Figure 6 The selection process was designed to identify genuine NEPs. Candidate NEPs consisted of two parts: annotated C-terminal sequences and newly extended N-terminal sequences. The preliminary screening criteria included: (1) both C-terminal and N-terminal sequences were ≥5 amino acids in length; (2) the proportion of unfragmented sites in both C-terminal and N-terminal sequences was ≤0.33; (3) PSM score was ≥0.25; and (4) the absolute value of the precursor ion mass deviation was ≤6 ppm. The PSM score was calculated by pFind software and was defined as the ratio of the intensity of the matched peak to the total peak intensity in a single spectrum.

[0122] Candidate NEPs that passed the initial quality control screening but did not exist in databases such as NCBI, UniProt, NextProt, PeptideAtlas, Massive, GPMDB, and PubMed were further subjected to manual spectral verification. The verification criteria included: (1) having a high signal-to-noise ratio and few impurity peaks, with at least three consecutive b / y ion matches; (2) excluding confusion with annotated peptides caused by isoweight sequences (such as I=L, Q[deamidation]=E, GG=N). Candidate NEPs that met the screening criteria were chemically synthesized, and the pBuild module built into the pFind software was used to perform matching analysis between the original spectrum and the synthesized peptide spectrum, and the cosine similarity score between the two was calculated. Finally, three real NEPs were successfully verified from the NovelNSeq analysis results: RPGMPATVVSQADQLTEEQIAEFK, ALVMLALLSEAAAQYNPEPPPPR, and LAGSSGPSVRPAGPPNAGSMSAGSER. Their Universal Spectrum Identifiers are shown in Table 3.

[0123] Table 3 Universal Spectrum Identifiers for the Secondary Spectra of the Four New Peptides

[0124]

[0125] 5. Novel N-terminal protein variants encoded by human genes CALM2, CAPNS1, and CPNE7

[0126] This study, through synthetic peptide verification, found that the cosine similarity between the original spectrum and the synthetic peptide spectrum of NEP "RPGMPATVVSQADQLTEEQIAEFK" was 0.99. Figure 7 a) This confirms the accuracy of the original spectral analysis results. NCBIBLASTP analysis showed that this peptide was the highest-scoring similar sequence in the human proteome, with a sequence coverage of only 58% (Table 4). Furthermore, this peptide was not found in databases such as UniProt, NextProt, PeptideAtlas, MassIVE, GPMDB, and PubMed, indicating that it is a novel peptide not annotated in the human genome.

[0127] Table 4. Top five alignment results of peptide "RPGMPATVVSQADQLTEEQIAEFK" with the highest scores in NCBI BLASTP.

[0128]

[0129] Sequencing results from PEAKS, pNovo3, SpliceNovo, and NovelNSeq of the raw NEP spectrum showed amino acid recall rates of 54%, 71%, 75%, and 100%, respectively. Figure 7 (b) Due to the lack of b and y ion fragments at the N-terminal 1st and 3rd sites of this NEP, PEAKS, pNovo3, and SpliceNovo, based on de novo sequencing algorithms, could not accurately determine its N-terminal sequence. NovelNSeq employs an innovative algorithm: first, it matches and scores the spectrum against a database of theoretical N-terminal restriction peptides to resolve the annotated C-terminal sequence ADQLTEEQIAEFK of this NEP; then, it enumerates all possible extension sequences at the 5' end of the DNA coding sequence of the annotated sequence, and determines the correct complete peptide by finely scoring the spectrum against the enumerated complete peptides. This algorithm is essentially a database search, capable of accurately resolving NEPs and simultaneously locating their DNA coding sequence on the chromosome, overcoming the problems of low accuracy and difficulty in tracing the origin of results in traditional de novo sequencing algorithms.

[0130] TBLASTN analysis revealed that the NEP “RPGMPATVVSQADQLTEEQIAEFK” is encoded by three discontinuous CDS sequences on the negative strand of human chromosome 2, with genomic coordinates of 47162655-47162662, 47170734-47170764, and 47176277-47176309. CDS (47170734-47170764) corresponds to the second CDS at the 5' end of CALM2 isoform 2, CDS (47162655-47162662) is located at the 5' end of the third CDS, and CDS (47176277-47176309) is located in the 5' untranslated region of the second CDS. This NEP is formed by extending the amino acid sequence “ADQLTEEQIAEFK”, encoded by the second and third CDS, to the N-terminus as “RPGMPATVVSQ”. The first annotated CDS at the 5' end of CALM2 isoform 2 contains only 3 bases (ATG), encoding a single methionine residue. This NEP discovery reveals the CALM2 gene's ability to encode the N-terminus of a novel protein. Figure 7 c).

[0131] Within the terminator-to-terminator region of the newly extended sequence “RPGMPATVVSQ”, a unique start codon, ATG, is found upstream, suggesting that the translation of the novel N-terminal protein encoded by CALM2 may have been initiated from this region. Therefore, the CDS genomic coordinates encoding the novel N-terminus of CALM2 are 47176277–47176321. CALM2 encodes human calmodulin 2, a key calcium-binding protein that participates in physiological processes such as muscle contraction, neurotransmitter release, and cellular metabolism by regulating calcium signaling pathways. Mutations in this protein are associated with diseases such as arrhythmias and neurodevelopmental disorders. According to the latest annotations of the human genome, there are four known transcriptomic variants of CALM2. Following human genome nomenclature conventions, this newly discovered N-terminal protein is named CALM2 isoform 5, with a sequence length of 165 amino acids. The existence of this novel N-terminal protein is also supported by homologous proteins from closely related species: CALM2 isoform 5 is completely identical to the CALM2 isoform 1 sequence annotated in cynomolgus monkeys (Macaca fascicularis), and the amino acid sequences encoded by each CDS separated by introns are also identical. This evolutionary conservation further validates the credibility of the novel N-terminal protein.

[0132] In addition, NEPs “ALVMLALLSEAAAQYNPEPPPPR” and “LAGSGPSVRPAGPPNAGSMSAGSER” also passed the synthetic peptide verification, with cosine similarity of 0.98 between their original spectra and the synthetic peptide spectra. Figure 8 a) and 0.99 ( Figure 9 a). NCBI BLASTP analysis showed that their sequence coverage of the highest-scoring similar sequences in the human proteome was 65% ( Figure 8 b) and 40% Figure 9 b). These NEPs were not found in other databases, indicating that they are novel, unannotated peptides in the human genome.

[0133] TBLASTN analysis showed that the NEP “ALVMLALLSEAAAQYNPEPPPPR” is encoded by three discontinuous CDS on the positive strand of human chromosome 19, with genomic coordinates of 36133756-36133781, 36142300-36142333, and 36142652-36142660. Figure 8c). This NEP is generated by extending the amino acid sequence “EAAAQYNPEPPPPR”, encoded by the second and third CDS at the 5' end of the human CAPNS1 isoform 3, to the N-terminus as “ALVMLALLS”. The first annotated CDS at the 5' end of CAPNS1 isoform 3 contains only 5 bases (ATGGG) and encodes two amino acids, “MG”. The CAPNS1 gene generates a new protein through a novel alternative splicing technique, replacing the original N-terminal “MG” with a new protein.

[0134] The NEP “LAGSGPSVRPAGPPNAGSMSAGSER” is encoded by two consecutive CDS on the positive strand of human chromosome 16, with genomic coordinates of 89575844-89575897 and 89575898-89575918, respectively. Figure 9 c). This NEP is generated by extending the amino acid sequence “MSAGSER” encoded by the first CDS at the 5' end of human CPNE7 isoform 1 to the N-terminus as “LAGSGPSVRPAGPPNAGS”.

[0135] CAPNS1 is an essential small regulatory subunit that, upon binding to the large catalytic subunits (CAPN1 or CAPN2), forms functional calpain-1 or calpain-2 heterodimers, respectively. These complexes are crucial for calcium-dependent proteolytic activity. CPNE7 is a calcium-dependent membrane-bound protein involved in calcium signaling and cell membrane translocation, and may be associated with tumor suppressor function. Due to a lack of reference information such as homologous proteins, the complete sequences of the novel N-terminal proteins encoded by the CAPNS1 and CPNE7 genes cannot currently be determined. Sequencing results of the original NEP spectra from PEAKS, pNovo3, SpliceNovo, and NovelNSeq (…) Figure 8 d and Figure 9 d) shows that only NovelNSeq correctly resolved the spectra, further confirming the advantages of NovelNSeq in NEP sequencing.

[0136] 6. Reassessment of non-classical translation initiation sites in novel N-terminal protein variants

[0137] Although a recent large-scale proteogenomics study reported 192 translation initiation events within the human 5' UTR, it failed to identify novel N-terminal protein variants from the CALM2, CAPNS1, and CPNE7 genes. Further analysis revealed a common characteristic among the genes reported in this study with novel N-terminal coding potential: the terminator-to-terminator region of the first CDS at the 5' end typically lacks a classic ATG start codon upstream of the CDS. This characteristic led to the assignment of non-classical start codons to 157 potential novel N-terminal proteins. However, our analysis suggests that, as demonstrated in subsequent case studies, such annotations may be significantly unreliable.

[0138] The NEP “AGGAADMTDNIPLQPVR” reported in the literature was also identified by NovelNSeq. This peptide is formed by extending “MTDNIPLQPVR” to “AGGAAD” from the N-terminus of the protein encoded by the ATP9A gene on the negative strand of human chromosome 20. Since there is no classical start codon ATG upstream of the terminator region in the reading frame containing this peptide, the literature annotated the predicted non-classical start codons AAG (51768388-51768390) as the translation initiation site of the new N-terminus of the protein. However, based on homologous protein alignment analysis in the closely related species *Pongo pygmaeus*, it is speculated that the new N-terminus of the protein encoded by the human ATP9A gene may be generated through a translation frameshift mechanism. Three-frame translation analysis of the DNA sequence upstream of the coding sequence “AGGAADMTDNIPLQPVR” showed that the new N-terminus of the protein consists of “MVRPCPSVGPRGR” encoded by the first reading frame and “LRAWPGAR…” encoded by the third reading frame, and is achieved through a one-base frameshift translation. Figure 10 ).

[0139] Further verification revealed the peptide “LRAWPGAR” in trypsin-specific enzyme digestion mass spectrometry data, with a cosine similarity of 1.00 between the original spectrum and the synthesized peptide spectrum. Figure 11 a), and this peptide is a novel peptide not annotated in the human genome ( Figure 11 (b) It is worth noting that this peptide only exhibits pancreatic enzyme specificity under the frameshift mechanism of translation.

[0140] Enzyme cleavage characteristics strongly support the theory that novel N-termini are generated through a translational frameshift mechanism. Based on the above evidence, the translation initiation site for the novel N-terminus of the protein encoded by the human ATP9A gene should be the classical start codon ATG, rather than the non-classical start codon AAG. Whether translational frameshift is a universal mechanism for the generation of novel N-termini in proteins warrants further investigation.

[0141] III. Conclusion

[0142] NovelNSeq is the first algorithm in proteomics that specifically resolves novel N-terminal protein marker peptides (novel N-terminal extension peptides) generated by alternative splicing, independent of transcriptome data. By fully considering the coding logic of these novel N-terminal extension peptides, NovelNSeq can not only accurately resolve them in mass spectrometry data but also trace their coding mechanism. Compared to mainstream proteogenomics methods, NovelNSeq can effectively resolve novel N-terminal proteins that were not predicted by transcriptome data but were independently detected by mass spectrometry, thus serving as an important supplement to existing methods.

[0143] Currently, NovelNSeq is only suitable for identifying novel N-terminal proteins generated by typical alternative splicing mechanisms. This limitation stems from the restriction on marker peptide characteristics in the algorithm design. Future work needs to expand the algorithm's functional modules to improve its ability to analyze complex splicing products. Furthermore, the novel N-termini encoded by the human CALM2 and CAPNS1 genes replace extremely short initial CDS coding sequences (such as "M" and "MG"), suggesting that regions with shorter initial CDS may be more likely to generate novel N-termini through alternative splicing. This requires further verification. Published literature... The annotation of translation start sites for the 157 novel N-terminal proteins reported in the literature as non-classical promoters may contain a common error. Their true translation start sites and the generation mechanism of the novel N-terminus deserve further investigation.

[0144] The core idea of ​​NovelNSeq—designing specific resolution algorithms based on the sequence characteristics of marker peptides—is not only applicable to the identification of novel N-terminal proteins generated by alternative splicing, but also provides a general framework for exploring N-terminal variations in proteins driven by other mechanisms. Future research will focus on exploring the functions of novel N-terminal proteins in diseases such as cancer, potentially providing new scientific evidence for disease diagnosis, therapeutic target discovery, and precision medicine research.

[0145] The undescribed parts of this invention are the same as or implemented using existing technology. The applicant declares that this invention is illustrated through the above embodiments, but the invention is not limited to the above detailed methods, i.e., it does not mean that the invention must rely on the above detailed methods to be implemented. Those skilled in the art should understand that any improvements to this invention, equivalent substitutions of raw materials for the product of this invention, additions of auxiliary components, and selection of specific methods all fall within the protection and disclosure scope of this invention.

Claims

1. A method for analyzing N-terminal protein variants that overcomes the limitations of proteogenomics and de novo sequencing, characterized by: Includes the following steps: A. Construction of a database of N-terminal theoretical enzyme-digested peptides Based on all annotated human protein-coding sequences in the NCBI database, a comprehensive database of N-terminal theoretical enzyme digestion peptides was constructed. For proteins encoded by a single CDS, theoretical digestion was performed directly to obtain their N-terminal peptides. For protein-coding sequences composed of multiple CDS, theoretical digestion was performed on the encoded amino acid sequences to obtain N-terminal peptides, with and without the first CDS at the 5' end. B. Calculate the matching score between the secondary spectrum and the sequence in the N-terminal theoretical restriction peptide database. Secondary spectra are not preprocessed to prevent the loss of daughter ion signal peaks; for peaks of length... Incomplete peptide segments exist. There are 10 fragmentation sites, each considering 21 theoretical daughter ions, including the by-type internal ion; the size of the input array is 1000. Each row corresponds to a fragmentation site. The first 21 columns represent the matching scores of the 1st to 21st seed ions, and the last 21 columns represent the values ​​of these matching scores divided by the length of the incomplete peptide segment. This two-dimensional array represents the theoretical daughter ion matching of all fragmentation sites, followed by a two-step convolution operation through two convolutional layers: the kernel size of the first convolutional layer is... It is mainly used to learn the scoring weights of different types of daughter ions; the kernel size of the second convolutional layer is... It is mainly used to learn the weights of the total and average values ​​of the daughter ion matching scores in the recall of correct peptides. After two convolution operations, the size of the output array is [size missing]. Summing these values ​​yields the matching score between the secondary spectrum and the incomplete peptide segment, where the spectrum matches the incomplete peptide segment. The scoring function that assigns scores based on the matching results is as follows: Where H1 and H2 refer to the convolution kernel, W k Refers to the convolution kernel, T i,k B1 and B2 refer to the matrix after the first convolution operation, respectively; C. Screen out the N incomplete peptide segments with the highest scores; D. For each incomplete peptide segment, enumerate possible complete new N-terminal extended peptide segments (NEPs) using a pre-constructed dictionary, extend 12,000 bases to the 5' end of the annotated CDS encoding the incomplete peptide segment, search for possible new CDS, and ensure that the mass of the commonly encoded peptide segment is equal to that of the precursor ion in the spectrum. All possible NEPs corresponding to the secondary spectrum are obtained by this method. E. Perform peptide matching and scoring on all possible NEPs to obtain the best matching result, and provide the newly resolved peptide segment and the coordinates of the CDS encoding the new peptide segment on the chromosome.

2. The analytical method according to claim 1, characterized in that, In step A, the method for constructing the N-terminal theoretical cleavage peptide database is as follows: For an annotated human protein coding sequence, if it consists of a continuous coding DNA sequence (CDS) on a chromosome, all N-terminal theoretical cleavage peptides of the protein can be predicted by theoretical cleavage methods. If an annotated human protein-coding sequence consists of multiple CDSs separated by introns on a chromosome, In addition to theoretical enzymatic digestion of the complete protein, theoretical enzymatic digestion of the amino acid sequence encoded by the remaining CDS after removing the first CDS at the 5' end must also be considered. For the second CDS at the 5' end, its value is defined as 0, 1, or 2 to indicate that translation begins from the first, second, or third base at the 5' end of the CDS, respectively, in case there are different translation initiation shifts.

3. The analytical method according to claim 1, characterized in that, The positive and negative samples are generated as follows: Based on the correct peptide sequence, the 5, 6, and up to N-1 amino acids at the C-terminus of the peptide are enumerated sequentially. Then, for each enumerated incomplete peptide, its first three amino acids are replaced with any combination of amino acids with a mass deviation within 20 parts per million, generating multiple erroneous peptides that may be confused with the initial incomplete peptide during peptide spectrum matching. The pairing of a correct incomplete peptide and an erroneous incomplete peptide is defined as a positive sample, and the pairing of an erroneous incomplete peptide and a correct incomplete peptide is defined as a negative sample. Several positive and negative samples can be generated from a single mass spectrum.

4. The analytical method according to claim 1, characterized in that, The positive samples are represented as ( , ), negative samples represent ( , ), In step B, the process of calculating the matching score between the secondary spectrum and the sequence in the N-terminal theoretical restriction peptide database is as follows: For positive and negative samples, the composition of the elements of the two-dimensional array is defined by the following formulas (9) and (10), respectively: in, and ; The input array is subjected to a two-step convolution operation using two convolutional layers: First, a one-dimensional convolutional kernel is used. and a bias term , input array Convert to a new two-dimensional array The convolution stride is 21, and the specific details of the convolution calculation are shown in the following formula (11): in, and ; Then, use a one-dimensional convolution kernel. and bias terms , array Convert to a new array The convolution stride is 2, and the specific details of the convolution calculation are shown in formula (12): By summing the array Convert to a new one-dimensional array For positive and negative samples, These are represented by the following formulas (13) and (14), respectively: Finally, the SoftMax function is used to convert the array... Convert to an output array with values ​​between 0 and 1 Assuming the spectrum and incomplete peptide segments The matching results are scored, and the scoring function designed in Model 2 is expressed by the following formula (15): 。 5. The analytical method according to claim 1, characterized in that, Secondary spectra are not preprocessed by removing isotope peaks and low-abundance noise peaks to prevent the loss of daughter ion signal peaks.

6. The analytical method according to claim 1, characterized in that, In step D, the dictionary is constructed as follows: the length of the theoretical N-terminal cleavage peptides considered is between 5 and 20 amino acids, and a maximum of two missed cleavage events are allowed. A dictionary is used to record each theoretical N-terminal cleavage peptide and all the corresponding protein information. The key of the dictionary is the theoretical N-terminal cleavage peptide, and the value is a set containing all the protein information that may produce the peptide. Each piece of protein information is represented in the form of a string, with the format: chromosome / gene name / protein ID / positive or negative strand / position of the first base at the 5' end of the first or second CDS on the positive strand of the chromosome / translation start position of the CDS.

7. The analytical method according to claim 1, characterized in that, In step E, the scoring function in the CNovo algorithm is used to score the peptide profiles of NEPs.

Images