Gene signature-based developmental dynamic expression pattern prediction method, model, system and database structure

Abstract

The application discloses a gene feature-based developmental dynamic expression pattern prediction method, model, system and database structure, relates to the cross technical field of biological information and artificial intelligence, and comprises the following steps: determining a gene set of a developmental dynamic expression pattern to be detected; acquiring multi-dimensional gene feature information of the genes to be detected; inputting the multi-dimensional gene feature information into a developmental dynamic expression pattern prediction model, and outputting a developmental dynamic pattern corresponding to target genes. The application greatly expands the range of species involved in the training data of the gene expression related model, innovatively uses multi-dimensional gene features, and quickly, efficiently and at low cost predicts the dynamic expression pattern of genes in the development of biological tissues or organs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information and artificial intelligence, in particular to a development dynamic expression pattern prediction method, model, system and database structure based on gene characteristics. BACKGROUND

[0002] The entire genome and DNA sequences of genes held by different cells, tissues and organs of an organism are basically the same in the entire growth and development process, but the gene expression of different types of tissues and organs and different development stages is significantly different. In fact, the initiation and termination of biological development programs are largely mediated by changes in gene transcription, and timely and appropriate gene transcription events are necessary conditions to ensure normal biological processes. Many experimental results based on animal tissue or organ development show that genetic variations related to human diseases can affect phenotypes by changing the timing of gene expression. In plants, countless genes shape growth and development by modulating their temporal expression, especially the reproductive development process closely related to crop yield. Therefore, genes with significant expression changes along the development stage of organs or tissues (i.e., development dynamic genes) are an excellent entry point for exploring biological development mechanisms and related major issues behind them. To some extent, genes with development dynamic expression patterns are also more likely to be key functional molecules that regulate biological processes. Therefore, many research teams at home and abroad have invested a lot of resources and funds to obtain gene expression and development dynamic patterns in the development process of animal and plant organs by using next-generation sequencing (NGS) technology, third-generation sequencing technology (TGS), and single-cell sequencing technology (scRNA-seq) (Nature, 2019; Nature, 2019; Genome Biology, 2022; Cell, 2025). The research team of the laboratory has also analyzed the expression and identity of coding genes and non-coding genes in plant seed development by NGS and its analysis technology, and has achieved certain results (International Journal of Molecular Sciences, 2023; Journal of Plant Physiology, 2021; Journal of Plant Physiology, 2021). However, can we obtain the development dynamic pattern of gene expression simply and quickly without relying on complex experiments and expensive sequencing?

[0003] In recent years, with the continuous generation of biological big data and the vigorous development of artificial intelligence (AI), the application of machine learning (ML) and deep learning (DL) has made it possible to understand gene expression in biological genomes. Recently, a research team from the Arc Institute published a multi-modal AI model named Evo, which can not only generate almost real genomes and gene sequences, but also predict chromatin accessibility directly related to gene expression (Science, 2024; bioRxiv, 2025). In mammals such as humans, more examples are using ML or DL to predict gene expression levels. Early models were based on more indirect information to predict gene expression, such as chromatin immunoprecipitation data of transcription factors and histone modification data (PNAS, 2009). Deep convolutional neural networks as a popular DL scheme have reached an unprecedented level in predicting human and mouse gene expression, with two works published in Cell Reports and Nature Methods being representative. Xpresso can predict gene expression levels based solely on genomic DNA sequences and help infer transcriptional and post-transcriptional regulatory mechanisms (Cell Reports, 2020). The more well-known Enformer is the first DL model that can handle DNA sequences up to 100 kb, and it can also predict gene expression intensity (Nature Methods, 2021). By applying ML and DL, using cis-regulatory elements on genomic DNA can not only predict gene expression responses to temperature in corn, but also predict gene expression changes during tomato ripening (The Plant Cell, 2022). Last year, a work published in Nature Communications started from the DNA sequence of the flanking region of the gene and achieved an accuracy of more than 80% in predicting gene expression levels in four species of Arabidopsis thaliana, tomato, sorghum, and corn, while showing considerable cross-species performance (Nature Communications, 2024). It is not difficult to see that a series of DL models are indeed more powerful, but research results have shown that tree-based ML models are significantly better than DL in processing tabular data (Advances in Neural Information Processing Systems, 2022). Therefore, in the process of research, we should avoid blindly relying on DL and choose the appropriate model category for different types of data. In summary, using machine learning and deep learning architectures to "read" the expression patterns of genes in the genome not only provides a new perspective for a deep understanding of human diseases, but also provides a new strategy for intelligent plant breeding.

[0004] Although the research on gene expression with artificial intelligence technology is in full swing, the inventors noticed that all the published related excellent works at least have the following problems (Example Table 1 compares and illustrates by way of example). First, the species covered by the data training set in the currently published models is still very few, especially plants. Second, most models focus on using genome sequences to complete the prediction task related to gene expression, which seriously underestimates other valuable and meaningful information that can be used. Third, almost all research teams at home and abroad focus on the development of models for expression levels or expression differences, and the perspective of the problem is too one-sided and narrow. In the prior art, there is no report on the prediction of gene expression patterns during the development process, resulting in a long process, time-consuming and high cost to obtain the dynamic gene of the development.

[0005] In summary, the species covered by the current gene expression prediction model is not extensive enough, the type of training data used is too single, and it cannot complete the important task of dynamic prediction of gene expression during the development process, making it difficult to meet the actual needs of relevant researchers. SUMMARY

[0006] In order to overcome the limitations of the prior art, the technical problem to be solved by the present application is to provide a gene feature-based development dynamic expression pattern prediction method, model, system and database structure, which integrates genomics, transcriptomics and machine learning architecture to realize fast, effective and low-cost prediction of gene development dynamic expression patterns; to fill the technical gap of current gene development dynamic expression prediction, and to provide a new tool for life science research, disease mechanism analysis and intelligent crop breeding.

[0007] One aspect of the present application provides a gene feature-based development dynamic expression pattern prediction method, comprising the following steps:

[0008] S1, determining a gene set for detecting the development dynamic expression pattern;

[0009] S2, obtaining multi-dimensional gene feature information of the gene set to be detected;

[0010] S3, inputting the multi-dimensional gene feature information into a development dynamic expression pattern prediction model, and outputting a development dynamic pattern corresponding to the target gene, wherein the development dynamic expression pattern prediction model is trained according to a multi-dimensional gene feature information data matrix with a gene dynamic expression pattern label, and the gene dynamic expression pattern label is determined according to the gene expression matrix of the tissue or organ development process of the training sample.

[0011] As a preferred technical solution of the present application, the gene set in step S1 is any one or any combination of the following:

[0012] (1) The complete set of genes within the newly sequenced and assembled genome;

[0013] (2) The set of differentially expressed genes obtained based on high-throughput sequencing;

[0014] (3) A set of genes that are specifically expressed in biological tissues or organs;

[0015] (4) A collection of species-specific or cross-species genes that exists within or between species;

[0016] (5) A collection of candidate genes compiled through literature review;

[0017] In step S2, multidimensional gene feature information of the gene to be detected is obtained. Step S2 specifically includes:

[0018] S2.1: Obtain numerical features of the set of genes to be detected based on gene sequences and annotations;

[0019] S2.2: Based on gene properties and annotations, obtain the categorization features of the set of genes to be detected;

[0020] S2.3: Integrate the numerical and categorical features of the gene set to be detected into a data matrix containing multidimensional gene feature information.

[0021] In step S3, as a preferred embodiment of the present invention, multidimensional gene feature information is input into the developmental dynamic expression pattern prediction model, and the developmental dynamic pattern corresponding to the target gene is output. Step S3 specifically includes:

[0022] S3.1: Load the data matrix containing multidimensional gene feature information into the developmental dynamic expression pattern prediction model. The developmental dynamic expression pattern prediction model performs mapping calibration on the input features to ensure that the order of the input features is consistent with the order of the developmental dynamic expression pattern prediction model in the pre-training process.

[0023] S3.2: For each gene, the developmental dynamic expression pattern prediction model traverses the set of trained decision trees, obtains the cumulative score of the prediction values ​​of all trees, and converts the cumulative score into a probability value.

[0024] S3.3: Manually set thresholds to determine the developmental dynamics of gene expression.

[0025] Another aspect of this invention discloses a series of related operations for constructing a developmental dynamic expression pattern prediction model, wherein the operations involved in constructing the developmental dynamic expression pattern prediction model include:

[0026] E1: Gene expression matrix representing the developmental process of organ systems in a species;

[0027] E2: obtaining gene developmental dynamic expression pattern tags;

[0028] E3: obtaining multi-dimensional gene feature information of whole genome;

[0029] E4: training method of developmental dynamic expression pattern prediction model;

[0030] E5: construction strategy of developmental dynamic expression pattern prediction model.

[0031] In operation E1, according to embodiments of the present disclosure, part of the model training data is obtained based on gene expression matrix representing the organ development process of the species, and operation E1 includes:

[0032] E1.1: obtaining transcriptome sequencing data representing the development process of tissues or organs of the species;

[0033] E1.2: filtering low-quality sequencing data and retaining high-quality sequencing data;

[0034] E1.3: obtaining genome data representing the species;

[0035] E1.4: based on the genome data, splicing the transcriptome representing the species, and completing the transcriptome assembly of the tissue or organ development;

[0036] E1.5: quantitatively analyzing the expression of genes and their transcripts, and generating a gene expression matrix of the organ development process.

[0037] In operation E2, according to embodiments of the present disclosure, part of the model training data is obtained based on gene developmental dynamic expression pattern tags, and operation E2 includes:

[0038] E2.1: manually sorting gene expression matrix and sample grouping information to ensure that the data format meets the requirements;

[0039] E2.2: constructing a regression design matrix for subsequent differential expression analysis;

[0040] E2.3: obtaining significantly differentially expressed genes by generalized linear model;

[0041] E2.4: filtering genes with significant expression changes in the development process, i.e., developmental dynamic genes, by stepwise regression.

[0042] In operation E3, according to embodiments of the present disclosure, the other aspect of the model training data comes from the multi-dimensional gene feature information of the whole genome, and operation E3 includes:

[0043] E3.1: obtaining numerical features of genes, such as length, GC content, and number of exons, according to gene sequences and annotations;

[0044] E3.2: Obtain the sub-type features of the genes according to the gene properties and annotations: whether it is a protein-coding gene; whether it is a gene that undergoes alternative splicing; whether it is a transposable element related gene; whether it is an evolutionarily conserved gene;

[0045] E3.3: Integrate the numerical features and sub-type features of the whole genome gene set into a matrix containing multi-dimensional gene features.

[0046] In operation E4, according to an embodiment of the present disclosure, a training method of a developmental dynamic expression pattern prediction model is developed, and operation E4 includes:

[0047] E4.1: Correspond the dynamic expression pattern of each gene to its respective gene features to form a data matrix of multi-dimensional gene feature information with gene dynamic expression pattern labels;

[0048] E4.2: Randomly divide the data matrix into a training set and a validation set, wherein the training set accounts for 80% and the validation set accounts for 20%, and use the grid search method to optimize the hyperparameters to obtain the optimal hyperparameter configuration under the current conditions;

[0049] E4.3: According to the expression pattern of the gene in the developmental process, map the "non-developmental dynamic" gene and the "developmental dynamic" gene into two types of labels, and train the model under the optimal hyperparameter configuration;

[0050] E4.4: Monitor the training error and validation error, and draw the error loss curve, use the early stopping method to determine the optimal training number of rounds, to avoid model overfitting, and save the trained prediction model;

[0051] E4.5: Evaluate the contribution of each feature to the developmental dynamic expression pattern;

[0052] E4.6: Input the test set data into the developmental dynamic expression pattern prediction model, and the model maps and calibrates the input features to ensure that the order of the input features is consistent with the order of the model pre-trained;

[0053] E4.7: For each gene in the test set, the model traverses the trained decision tree set, and combines the learning rate scaling to obtain the cumulative score of the predicted values of all trees;

[0054] E4.8: Convert the cumulative score into a probability value to determine whether the gene expression is developmental dynamic with a default threshold of 0.5, calculate the confusion matrix and the area under the ROC curve AUC, and draw the ROC curve;

[0055] E4.9: In view of the fact that the default threshold is not the best threshold for dividing the dynamic expression of genes, the sensitivity and specificity curves are drawn, and the intersection of the two is taken as the best threshold to optimize the sensitivity and specificity of the model, the confusion matrix and the area under the ROC curve are recalculated, and the optimized ROC curve is drawn.

[0056] In operation E5, according to the construction strategy of the developmental dynamic expression pattern prediction model of the embodiments of the present disclosure, operation E5 comprises:

[0057] E5.1: Integrate the multi-dimensional gene feature information of the whole genome genes of multiple species and the developmental dynamic expression pattern into multiple matrices, and construct multiple single-species models SSM through model training;

[0058] E5.2: Integrate the multi-dimensional gene feature information of the whole genome genes of multiple species and the developmental dynamic expression pattern into one matrix, and construct one cross-species model CSM through model training.

[0059] Another aspect of the present disclosure provides a developmental dynamic expression pattern prediction model, characterized in that: the model is constructed by the method for constructing the developmental dynamic expression pattern prediction model according to any one of the above.

[0060] Another aspect of the present disclosure provides a system for predicting the developmental dynamic expression pattern of genes, comprising:

[0061] M1: Gene set determination module, for determining the gene set of the developmental dynamic expression pattern to be detected;

[0062] M2: Gene feature acquisition module, for acquiring multi-dimensional gene feature information of the genes to be detected;

[0063] M3: Developmental dynamic prediction module, for inputting the multi-dimensional gene feature information into the developmental dynamic expression pattern prediction model, and outputting the developmental dynamic pattern corresponding to the target gene; wherein the developmental dynamic expression pattern prediction model is trained and constructed based on the above E1-E5 operation steps.

[0064] Another aspect of the present disclosure provides an interactive expandable gene feature information database structure for developmental dynamic expression pattern prediction, which is constructed based on gene feature information and specially provided with expandable dimensions for gene feature information, and interactive or potential interactive for the gene set of the developmental dynamic expression pattern to be detected and the developmental dynamic expression pattern prediction model;

[0065] The main architecture of the gene feature information database structure is set as: the database structure is developed around the GENE_MASTER table, and meanwhile, it is provided with specific association relationship with other several tables, specifically including:

[0066] D1 : Core table element: GENE_MASTER; built as the core of the entire database structure, used to store basic information of genes, including basic identification, name, description or other basic gene data of genes;

[0067] D2: Association table and relationship element; including:

[0068] D2.1: NUMERIC_FEATURES table; relationship type setting: GENE_MASTER and NUMERIC_FEATURES are built as a one-to-many relationship, described by "contains"; data connotation setting: each record in the GENE_MASTER table allows multiple records in the NUMERIC_FEATURES table; the NUMERIC_FEATURES table is set to optionally store numerical features of genes;

[0069] D2.2: CATEGORICAL_FEATURES table; relationship type setting: GENE_MASTER and CATEGORICAL_FEATURES are set as a one-to-many "contains" relationship; data connotation setting: the CATEGORICAL_FEATURES table is used to store categorical features of genes;

[0070] D2.3: EXTENDED_FEATURES table; relationship type setting: GENE_MASTER and EXTENDED_FEATURES are a one-to-many "association" relationship; data connotation setting: the EXTENDED_FEATURES table optionally stores extended feature information of genes, including numerical and categorical features and other information;

[0071] D2.4: ANNOTATION_SOURCES table; relationship type setting: GENE_MASTER and ANNOTATION_SOURCES are set as a one-to-many "reference" relationship; data connotation setting: the ANNOTATION_SOURCES table stores the sources of gene annotation information;

[0072] The extension and interaction structure of D2.1-D2.4 follows the following data relationship: since the annotation information of a gene may have multiple different sources, a GENE_MASTER record is also set to allow reference to multiple NUMERIC_FEATURE, CATEGORICAL_FEATURE, EXTENDED_FEATURES, and ANNOTATION_SOURCES records respectively.

[0073] Compared with the prior art, the beneficial effects of the present application include: after determining the gene set of the development dynamic expression pattern to be detected, the present application obtains the multi-dimensional gene feature information of the genes to be detected, and inputs the multi-dimensional gene feature information into the development dynamic expression pattern prediction model to predict the development dynamic expression pattern of the target gene. On the one hand, the range of species involved in the training data of the gene expression related model is greatly expanded, and the multi-species information is integrated, which has considerable generalization ability and applicability. On the other hand, the multi-dimensional gene feature is innovatively used to quickly and efficiently predict the dynamic expression pattern of the gene in the development of biological tissues or organs. Compared with the cumbersome process of sampling the development stage samples and performing high-throughput transcriptome sequencing in the traditional method, the present application greatly reduces the experimental time, shortens the research cycle and greatly reduces the research cost. Therefore, the present application fills a technical blank in the current gene development dynamic expression prediction field, and at least partially overcomes the technical problems of long detection time and high detection cost of gene expression pattern detection in the development process in the related art. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is a flow framework diagram of a development dynamic expression pattern prediction method based on gene features provided by the present application.

[0075] Figure 2 is a database structure block diagram for predicting the development dynamic expression pattern of genes provided by the present application.

[0076] Figure 3 is a dynamic decision threshold division and ROC curve for predicting the development dynamic expression pattern of barley genome genes in grains using the cross-species model CSM provided by the present application.

[0077] Figure 4 is a system block diagram for predicting the development dynamic expression pattern of genes provided by the present application.

[0078] Figure 5 is a related operation framework diagram of a development dynamic expression pattern prediction method based on gene features provided by the present application.

[0079] Figure 6 is an error loss curve and importance feature ranking diagram in the CSM model construction process covering six species provided by the present application.

[0080] Figure 7 is a dynamic decision threshold division and ROC curve in the CSM model construction process covering six species provided by the present application.

[0081] Figure 8Error loss curve and importance feature ranking diagram of the single-species model SSM and the cross-species model CSM provided by the embodiment of the present application.

[0082] Figure 9 Dynamic decision threshold division and ROC curve in the construction process of the single-species model SSM and the cross-species model CSM provided by the embodiment of the present application.

[0083] Figure 10 Evaluation index column chart of the single-species model SSM and the cross-species model CSM for predicting the dynamic expression pattern of gene development provided by the embodiment of the present application. DETAILED DESCRIPTION

[0084] The application will be further described below in connection with the drawings. In the following description of the embodiments, specific details are set forth in order to provide a thorough understanding of the application. However, persons skilled in the art will understand that the application can be practiced without these specific details. In other instances, well-known systems, structures, circuits, and methods have not been described in detail in order to not obscure the application.

[0085] It should be understood that the term "comprises" as used in the specification and the appended claims indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should also be understood that the term "and / or" as used in the specification and the appended claims indicates any combination of one or more of the associated listed items and all possible combinations of those items.

[0086] As used in the specification and the appended claims, the term "if" can be interpreted as meaning "when" or "upon" or "in response to a determination" or "in response to a detection" depending on the context. Similarly, the phrase "if determined" or "if detected [the described condition or event]" can be interpreted as meaning "upon a determination" or "in response to a determination" or "upon a detection of [the described condition or event]" or "in response to a detection of [the described condition or event]" depending on the context.

[0087] The terms used in the present application, including technical terms and scientific terms, have the ordinary meanings understood by a person skilled in the art, unless otherwise specified. It should be noted that the terms used in the present application should be interpreted as having meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner. In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance. In the present application specification, the reference to "one embodiment" or "some embodiments" and the like means that the specific features, structures or characteristics described in connection with the embodiment are included in one or more embodiments of the present application. Therefore, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in further some embodiments" and the like appearing in different places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "include", "contain", "have" and their variants mean "include but are not limited to", unless otherwise specifically emphasized.

[0088] Embodiment 1

[0089] In the related art, the developmental stage sample is usually sampled and high-throughput transcriptome sequencing is performed to obtain the gene developmental dynamic expression pattern, but this method requires a long experimental time and is extremely high in cost, and is not suitable for researchers to apply.

[0090] Table 1 Comparison of prediction tools for gene expression related tasks

[0091]

[0092] Therefore, the present application provides a gene feature-based developmental dynamic expression pattern prediction method, the steps of which include determining a gene set of the developmental dynamic expression pattern to be detected; obtaining multi-dimensional gene feature information of the genes to be detected; inputting the multi-dimensional gene feature information into a developmental dynamic expression pattern prediction model, and outputting a developmental dynamic pattern corresponding to the target gene, wherein the developmental dynamic expression pattern prediction model is trained according to a multi-dimensional gene feature information data matrix with a gene dynamic expression pattern label, and the gene dynamic expression pattern label is determined according to a gene expression matrix of the training sample in the tissue or organ development process.

[0093] Figure 1 A flow framework diagram of a gene feature-based developmental dynamic expression pattern prediction method is shown, which includes S1-S3 steps:

[0094] Step S1, determining a gene set of the developmental dynamic expression pattern to be detected

[0095] Due to the differences in research objectives, research foundations, and research experimental designs of different researchers, the sources and screening criteria of gene sets will also be different. The following are the sources of common target gene sets:

[0096] (1) All gene set in newly sequenced and assembled genome:

[0097] For a newly sequenced and assembled species genome, only basic sequence and genome structure information can be obtained through genome annotation information, and often lack of understanding of its transcription level. This gene set usually contains all annotated coding genes and non-coding genes, and is suitable for comprehensive analysis of the dynamic pattern of gene expression in the development process of the newly sequenced species. In order to illustrate and demonstrate steps S1-S3 in detail, 32915 protein coding genes (PCG) and 4766 long intergenic noncoding genes (lincRNA genes) of the whole genome of barley (Hordeum vulgare) sorted as the target gene set to be detected, related experiments and tests are carried out.

[0098] (2) Differential expression gene set based on high-throughput sequencing:

[0099] Through high-throughput sequencing technologies such as RNA-seq, gene expression data in different drug treatments, different growth environments, or different genotypes of samples can be obtained. Using differential expression analysis tools (such as DESeq2, edgeR, etc.), a list of significantly differentially expressed genes is screened out. Further tracking the change pattern of such differentially expressed genes in tissue and organ development helps to find key regulatory genes.

[0100] (3) Gene set with specific expression in biological tissues or organs:

[0101] Some genes have specific expression patterns in specific tissues or organs. For example, plant seed development related genes may be highly expressed in endosperm, while root development related genes may be highly expressed in root meristem. Focusing on the expression dynamics of tissue-specific genes in their development is one of the important clues to understand organ development.

[0102] (4) Species-specific (or cross-species) gene set existing within (or between) species:

[0103] Through comparative genomics, species-specific genes or cross-species conserved genes can be screened. While species-specific genes may be related to the unique developmental characteristics and speciation mechanisms of a particular species, cross-species conserved genes may play a conserved core regulatory role in the development of multiple species. Exploring the expression variation of such gene sets in organ development helps to reveal the commonality and characteristics of developmental mechanisms in evolution.

[0104] (5) Candidate gene set through literature research:

[0105] In many cases, by searching relevant field literature on websites (such as China Knowledge Network, Google Scholar, PubMed, etc.), a gene set of interest to researchers can be quickly obtained. For example, some genes may have been reported to play an important role in multiple neurodegenerative diseases, and focusing on the expression pattern dynamics of these genes in organ development is helpful to grasp their molecular and biological functions.

[0106] Step S2, obtaining multi-dimensional gene feature information of the gene to be detected

[0107] S2.1: Obtain numerical features of the gene according to the gene sequence and annotation;

[0108] According to the ID, sequence, and annotation information of the gene to be detected, the length, GC content, and number of exons of the gene representative transcript are counted one by one. For transcript length, according to the genome sequence fasta file and genome annotation gtf / gff3 file, use bedtools or TBtools software to extract gene coordinates for calculation. For GC content, use BioPython, EMBOSS, SeqKit, or TBtools software to count the number of G and C bases in the transcript sequence, and then calculate the GC content according to the formula GC% = (G+C) / (A+T+G+C).

[0109] For the number of exons, according to the information in the genome annotation gtf file, use Gffread, bedtools, or TBtools software to obtain. The theoretical value of the length feature is a natural number greater than 1, the value range of the GC content feature is (0, 1), and the value of the exon feature is a natural number greater than or equal to 1.

[0110] S2.2: Obtain categorical features of the gene according to the gene properties and annotations;

[0111] (1) Protein coding feature acquisition:

[0112] According to the genome annotation gtf / gff3 file, if there is detailed annotation of the gene in the annotation (such as annotation: protein coding gene, lncRNA, rRNA, snoRNA, etc.), the feature information can be directly used.

[0113] If there is no detailed annotation of the gene category in the genome annotation file, or the gene is obtained by the researcher through cloning and other means and is not included in the genome annotation, the related website can be submitted to detect the coding ability of the gene and its transcript. CPAT software (https: / / cpat.readthedocs.io / en / latest / ) can predict the coding ability of the target gene sequence based on a logistic regression model, exons, and intron sequences of the genome gene. The CPC2 software (http: / / cpc2.cbi.pku.edu.cn / ) can analyze the Fickett score, ORF length, ORF integrity, and isoelectric point of the sequence to predict the ability of the sequence to have a reading frame and encode proteins. The LGC software (https: / / ngdc.cncb.ac.cn / lgc / ) can calculate the GC content, ORF length, and other indicators to determine the protein coding potential of the transcript. In addition, the BLAST method can be used to search for potential protein domains in the Pfam database by aligning the target gene transcript. In summary, combined with the current mainstream software, the target gene can be preliminarily determined to be a coding gene or a non-coding gene.

[0114] (2) Variable splicing (Alternative splicing, AS) feature acquisition:

[0115] According to the genome annotation gtf / gff3 file, if there is only one transcript at a gene site, there is no alternative splicing. However, most genes on the genome are prone to alternative splicing events, which usually present as exon skipping, exon exclusion, intron retention, alternative promoters, alternative terminators, alternative acceptor sites, and alternative donor sites. In short, if a gene site has multiple corresponding transcripts, it can be considered to have alternative splicing events. In addition, if you want to obtain the alternative splicing events of the whole genome gene or a large-scale gene set, you can use SUPPA, rMATS, AStalavista or ASprofile software to batch statistics and generate results.

[0116] (3) Transposable element (Transposable element, TE) feature acquisition:

[0117] According to the genomic fasta sequence, BLAST in the transposable element database can detect whether there is a transposable element and its related fragments in the gene sequence. In addition, the genomic annotation gtf / gff3 file can also be converted into a bed file, and analyzed with the whole genome transposable element annotation of the target species using bedtools software to obtain the gene members related to the transposable element. If the whole genome transposable element annotation of the target species is not retrieved, the EDTA software (https: / / github.com / oushujun / EDTA) can be used to perform de novo transposable element annotation on the genome.

[0118] (4) Evolutionary conservation feature acquisition:

[0119] For the evolutionary conservation of genes, it can be understood from two aspects. On the one hand, the homology of nucleic acid sequences between species is used to determine whether the sequence of the gene is conserved; on the other hand, the genomic collinearity between species is used to determine whether the genomic position of the gene is conserved. If it is necessary to simultaneously consider the gene family division condition and the genomic collinearity condition between different species, OrthoFinder software (https: / / github.com / davidemms / OrthoFinder) or OrthoMCL software (https: / / orthomcl.org / orthomcl / app / ) and MCScanX (https: / / github.com / wyp1125 / MCScanX) can be used for analysis.

[0120] Note that these categorical features need to be recorded in one-hot encoding mode, that is, the feature exists as 1, and the feature does not exist as 0.

[0121] S2.3: Integrate the numerical features and categorical features of the target gene set into a multi-dimensional gene feature matrix.

[0122] The normalized numerical features and the encoded categorical features are integrated into a multi-dimensional gene feature matrix according to the rows (genes) and columns (features). The matrix is the input data of the developmental dynamic expression pattern prediction model. Each row of the matrix represents a gene, and each column represents a feature. For example: row, gene A, gene B, gene C; column, encoded gene, non-encoded gene, presence of alternative splicing, absence of alternative splicing. The example structure of the multi-dimensional gene feature matrix is given in Table 2:

[0123] Table 2 Example structure of multi-dimensional gene feature matrix

[0124]

[0125] The multi-dimensional gene feature information of the gene to be detected serves as a pivotal carrier of data processing, and a special data structure can be constructed based on the gene feature information ( Figure 2 ). The data structure is specially set to have an expandable dimension based on the gene feature information and an interactivity or potential interactivity of the gene set and the developmental dynamic expression pattern prediction model of the developmental dynamic expression pattern to be detected. The main architecture of the gene feature information database structure is set as follows: the database structure is developed around the GENE_MASTER table, and the GENE_MASTER table is specially associated with other several tables. Specifically, the main architecture includes core table elements, associated tables, and relationship elements.

[0126] D1: Core table elements:

[0127] GENE_MASTER: The core of the entire database structure, used to store basic information of genes, which can optionally include basic identifiers, names, descriptions, or other basic gene data of genes.

[0128] D2: Associated tables and relationship elements; including:

[0129] D2.1: NUMERIC_FEATURES table;

[0130] Relationship type setting: the GENE_MASTER and the NUMERIC_FEATURES are constructed as a one-to-many relationship, and the relationship is described by “contains”;

[0131] Data connotation setting: each record in the GENE_MASTER table can correspond to multiple records in the NUMERIC_FEATURES table; the NUMERIC_FEATURES table is set to optionally store some numerical features of genes, including but not limited to length, GC content, and number of exons; the expansion and interaction structure follows the following data relationship: one gene can have multiple numerical features in different aspects, so one GENE_MASTER record allows to contain multiple NUMERIC_FEATURES records.

[0132] D2.2: CATEGORICAL_FEATURES table;

[0133] Relationship type setting: the GENE_MASTER and the CATEGORICAL_FEATURES are set to have a one-to-many “contains” relationship;

[0134] ​Data content setting: The CATEGORICAL_FEATURES table is used to store the categorical features of the gene, which can optionally include the gene coding ability, variable splicing condition, transposable element condition, and evolutionary conservation. The expansion and interaction structure follows the following data relationship: Since a gene can have multiple different categorical attributes, the GENE_MASTER record is also set to contain multiple CATEGORICAL_FEATURES records.

[0135] D2.3: The EXTENDED_FEATURES table;

[0136] Relationship type setting: GENE_MASTER and EXTENDED_FEATURES are in a one-to-many "association" relationship;

[0137] Data content setting: The EXTENDED_FEATURES table can optionally store some extended feature information of the gene, including numerical and categorical features, and other information, which can optionally include chromosome position, gene start coordinate, gene end coordinate, transcription direction, and extended information of specific functional annotation of the gene, performance information of the gene in a specific environment, or other information. In coordination, the expansion and interaction structure follows the following data relationship: A gene can be associated with multiple extended features, so a GENE_MASTER record can be associated with multiple EXTENDED_FEATURES records.

[0138] D2.4: The ANNOTATION_SOURCES table;

[0139] Relationship type setting: GENE_MASTER and ANNOTATION_SOURCES are set to a one-to-many "reference" relationship;

[0140] Data content setting: The ANNOTATION_SOURCES table stores the source of gene annotation information. The expansion and interaction structure follows the following data relationship: Since the annotation information of a gene can have multiple different sources, a GENE_MASTER record is also set to reference multiple ANNOTATION_SOURCES records.

[0141] Step S3, inputting the multi-dimensional gene feature information into the developmental dynamic expression pattern prediction model, and outputting the developmental dynamic pattern corresponding to the target gene

[0142] S3.1: Load the multi-dimensional gene feature matrix into the developmental dynamic expression pattern prediction model. The model maps and calibrates the input features to ensure that the order of the input features is consistent with the order of the pre-trained model;

[0143] According to the target species, after viewing the appropriate model (single-species model SSM or cross-species model CSM), load manual_training_process_info.rds and xgboost_model_manual.model using readRDS and xgb.load, respectively, for subsequent prediction of the developmental dynamic expression pattern of the detected gene. xgboost_model_manual.model is the actual pre-trained XGBoost model, which contains the decision tree structure and related parameters. Use read_excel and as.data.table to read and convert the multi-dimensional gene feature matrix, and extract specific feature columns from it, which are the features used during model training. Convert the data to matrix format by as.matrix, so that XGBoost uses DMatrix as input. The model maps the input features, ensuring that the order of the input features is consistent with the order of the pre-trained features of the model.

[0144] S3.2: For each gene, the model traverses the trained decision tree set and obtains the cumulative score of all tree prediction values, and converts the cumulative score to a probability value;

[0145] Execute the predict function to predict the data using the trained developmental dynamic expression prediction model, and finally obtain a prediction value (user_predictions) for each gene, with a value range of (0, 1). In simple terms, the trained model is composed of multiple weak decision trees, and each data point will be input to multiple decision trees, and each tree will give a prediction score. Finally, these scores are weighted and accumulated to form the final prediction value for each detected gene.

[0146] S3.3: Manually set the threshold to determine whether the gene expression is developmental dynamic or not;

[0147] Use the ifelse statement to manually convert the probability value to a category: if user_predictions > t, predict "developmental dynamic"; otherwise, predict "non-development dynamic". Wherein, t can be set to 0.5 by default, or manually find the optimal threshold through sensitivity and specificity curves (generally, the default 0.5 will not deviate too much). For example Figure 3As shown in FIG. 2a, the optimal threshold value for the barley gene example data is 0.54. Therefore, if user_predictions>0.54, it is "developmental dynamics"; otherwise, it is "non-development dynamics". Finally, among the 37681 genes, 16980 are predicted to have dynamic expression patterns in the grain development process. Through the above steps S1-S3, the prediction task of the developmental dynamic expression patterns of genes in the barley genome is completed. In order to investigate the effect of the model, a pre-experiment and trained CSM is used. It should be pointed out that the specified CSM does not contain any information and data related to the species of barley. The CSM is trained based on the grain development related data of six species of corn, millet, wheat, oat, rice and sorghum. Combined with the true results of the developmental dynamic expression patterns of genes in the genome derived from the real RNA-seq data in the grain development process, it can be known from the calculation of the confusion matrix that the area under the ROC curve (AUC) has reached 0.7477, the sensitivity is 0.6774, and the specificity is 0.6760 (as shown in FIG. 2b). Therefore, by using the prediction method provided by the present disclosure, the rapid, effective and low-cost prediction of the developmental dynamic expression patterns of genes is realized for the first time, and the limitations of the prior art are overcome. Figure 3 b). Therefore, by using the prediction method provided by the present disclosure, the rapid, effective and low-cost prediction of the developmental dynamic expression patterns of genes is realized for the first time, and the limitations of the prior art are overcome.

[0148] Therefore, based on the beneficial effects and database in the embodiments of the present disclosure, the present disclosure simultaneously provides a system for predicting the developmental dynamic expression patterns of genes (as shown in FIG. 3a): Figure 4

[0149] M1: a gene set determination module, used for determining a gene set of the developmental dynamic expression pattern to be detected;

[0150] M2: a gene feature acquisition module, used for acquiring multi-dimensional gene feature information of the gene to be detected;

[0151] M3: a developmental dynamic prediction module, used for inputting the multi-dimensional gene feature information into a developmental dynamic expression pattern prediction model, and outputting a developmental dynamic pattern corresponding to the target gene.

[0152] Embodiment 2

[0153] Figure 5 FIG. 2a shows a related operation framework diagram of a developmental dynamic expression pattern prediction method based on gene features provided by the embodiments of the present disclosure, which includes operations E1-E5.

[0154] ​Since the developmental dynamic expression pattern prediction model is trained according to the multi-dimensional gene feature information training sample with a gene dynamic expression pattern label, the gene dynamic expression pattern label is determined according to the gene expression matrix of the training sample of the tissue or organ development process. Therefore, first, the operations E1-E3 related to the acquisition of training data need to be explained. Here, the related data preparation for the construction of a cross-species model CSM based on data of six species of corn, millet, wheat, oat, rice and Zizania aquatica is taken as an example to help understand the operations E1-E3 in the embodiments of the present disclosure.

[0155] Operation E1, acquisition of a gene expression matrix representing the organ development process of a species

[0156] E1.1: Obtain transcriptome sequencing data representing the tissue or organ development process of a species. The NCBI (National Center for Biotechnology Information) in the United States and the CNCB (China National Center for Bioinformation) in China are important databases for storing most of the second-generation sequencing and third-generation sequencing data. In combination with relevant literature reports and database retrieval, RNA sequencing (RNA-seq) fastq data of different development stages of the tissues or organs of the species are sorted and downloaded. Among them, one tissue or organ of a species contains at least three samples of different development stages, and each stage contains at least two biological replicates.

[0157] E1.2: Filter low-quality sequencing reads. Preferably, fastqc and fastx_toolkit quality control software are used to filter sequencing reads, and high-quality reads in the sequencing data are retained.

[0158] E1.3: Obtain genome data representing a species. NCBI and Ensembl include most of the currently sequenced genomes of animals, plants and microorganisms and annotation data. The fasta genome sequence, gtf genome annotation information, CDS sequence, protein amino acid sequence and non-coding RNA (ncRNAs) sequence of the representative species can be obtained.

[0159] E1.4: Assemble the transcriptome of a representative species. Preferably, bowtie2, Tophat2 and cufflinks software are used to assemble the sequencing reads to the reference genome, and the transcriptome assembly of the tissue or organ development is completed.

[0160] E1.5: Quantification of gene and its transcript expression. Preferably, the transcript annotation files of different tissues or organs, different developmental stages are merged in each species respectively, and cuffdiff under cufflinks is used to quantify the expression of genes and their transcripts. The count value of genes and their transcripts is normalized using the FPKM (Fragments Per Kilobase of Transcript Per Million Fragments) method. Finally, the gene expression matrix of the seed development process of six species is obtained. Including (PCG + lincRNA gene): 39953 genes of maize, 317821 genes of millet, 126972 genes of wheat, 84987 genes of oat, 53801 genes of rice, and 32521 genes of sorghum.

[0161] Operation E2, acquisition of gene developmental dynamic expression pattern label

[0162] E2.1: Based on the generalized linear model (Generalized Linear Model, GLM), the maSigPro software can identify the gene expression pattern that presents significant difference in time sequence. Therefore, using maSigPro and related functions, taking the expression data of tissue or organ development as input, the developmental dynamic genes can be obtained by performing regression analysis in six species respectively. According to the input data format specified by the software, the gene expression profile and sample grouping information format are manually arranged.

[0163] E2.2: Constructing regression design matrix. Create a regression design matrix for fitting the regression model, and define the relationship between experimental groups and regression variables.

[0164] E2.3: Identifying significantly differentially expressed genes. Considering the data characteristics of NGS, the negative binomial distribution (Negative Binomial, NB) is selected as the distribution function used by the GLM model. The p.vector function is used to perform regression analysis on each gene, calculate the F statistic and p value of the regression model, and use the Benjamini-Hochberg (BH) correction to control the error rate of multiple hypothesis testing.

[0165] E2.4: Identifying significantly changed genes in the development process. Use the T.fit function to perform stepwise regression (Stepwise Regression) to select the most significant influencing factors, and output the p value for testing whether the entire regression model is significant and the goodness-of-fit R 2 of the model. It should be noted that the size of R 2 describes the explanatory power of the independent variables in the model to the change of the dependent variable, that is, by setting R 2threshold to divide whether the expression of a gene is dynamic or not in the process of organ development. 2 The value interval of the threshold is [0, 1], and 0.5 is set as the threshold to divide the dynamic expression genes. The final output of the development dynamic gene set is recorded and the data is saved. The development dynamic genes include 12799 in corn, 9811 in millet, 33849 in wheat, 16550 in oat, 11683 in rice, and 6874 in Zizania aquatica.

[0166] Operation E3, acquisition of multi-dimensional gene feature information of whole genome

[0167] The details of operation E3 are basically the same as step S2, and will not be described here.

[0168] Since the development dynamic expression pattern prediction model is trained according to the multi-dimensional gene feature information with gene dynamic expression pattern label training sample, the operations E4 and E5 related to the training and construction process of the model are described below. Here, the training and construction of the cross-species model CSM based on the data of corn, millet, wheat, oat, rice and Zizania aquatica are taken as examples to help understand operations E4 and E5 in the embodiments of the present disclosure.

[0169] Operation E4, training method of development dynamic expression pattern prediction model

[0170] The training process of the model can be at least The training can be carried out under the framework of R in the environment of i7-8550U CPU@1.80GHz 8G memory. The present disclosure integrates XGBoost (eXtreme Gradient Boosting) architecture to construct the development dynamic expression prediction model. XGBoost realizes machine learning algorithm under the framework of gradient boosting, and has the characteristics of high efficiency, flexibility and portability. This algorithm has been widely used in many machine learning tasks and has excellent performance, and when processing table data tasks, the effect even exceeds the deep learning architecture.

[0171] E4.1: Corresponding the dynamic expression pattern of each gene to its respective gene feature to form a data matrix of multi-dimensional gene feature information with gene dynamic expression pattern label;

[0172] E4.2: Randomly divide the matrix set into training set and validation set, where the training set accounts for 80% (146950 genes) and the validation set accounts for 20% (36740 genes). Using five-fold cross-validation, hyperparameter search is performed using grid search method, nrounds = c(1000), eta = c(0.001, 0.01, 0.1), max_depth = c(3, 5, 7), gamma = c(0.3, 0.5, 0.7), colsample_bytree = c(0.3, 0.5, 0.7), min_child_weight = c(1, 3, 5), subsample = c(0.3, 0.5, 0.7). The appropriate hyperparameter configuration of the model is determined according to the AUC value. Among the 729 hyperparameter combinations, the AUC value can reach the highest 0.7002 when the hyperparameters eta = 0.01, max_depth = 7, gamma = 0.5, colsample_bytree = 0.7, min_child_weight = 5, and subsample = 0.7 are used as input.

[0173] E4.3: According to the expression pattern of genes in the developmental process, the "non-development dynamic" label is mapped to the "development dynamic" label, and the value is 1. After extracting the features and labels, the data is converted to DMatrix format. The model is trained under the optimal hyperparameter configuration.

[0174] E4.4: Monitor the training error and validation error using watchlist, draw the error loss curve, and observe whether overfitting occurs. Early stopping method is used to determine the optimal training rounds, and if the loss value does not decrease for 10 consecutive rounds, the training is stopped in advance. As shown in FIG. a, the optimal training rounds are 529. The trained model is saved using xgb.save, saveRDS. The output file xgboost_model_manual.model is the trained CSM model of the six species, and the manual_training_process_info.rds file contains the relevant information of the CSM model training of the six species. Figure 6

[0175] E4.5: Use xgb.importance to evaluate the contribution of each feature to the developmental dynamic expression pattern. The obtained gain represents the average value of the loss reduction contributed by a feature as a splitting point in all trees and all nodes, which can be preliminarily understood as the contribution size of the feature to the prediction ability of the final model. Figure 6 b shows the contribution of 11 features to the model, among which, the contribution of conservation, length, and GC content to the prediction of developmental dynamic expression pattern is the largest.

[0176] ​E4.6: The test set data (36740 genes) is input into the developmental dynamic expression pattern prediction model xgboost_model_manual.model. The model maps the input features to ensure that the order of the input features is consistent with the order of the model pre-trained;

[0177] E4.7: For each gene in the test set, the model traverses the set of trained decision trees, combining the learning rate scaling to obtain the cumulative score of all tree prediction values;

[0178] E4.8: The cumulative score is converted into a probability value. With the default threshold (0.5), determine whether the gene expression is developmental dynamic, calculate the confusion matrix and the area under the ROC curve (AUC), and draw the ROC curve;

[0179] E4.9: Considering that the default threshold may not be the best threshold for dividing the dynamic expression of genes, by calculating the sensitivity and specificity at different thresholds, and taking the intersection of the two curves as the best threshold, the sensitivity and specificity of the model are optimized. As Figure 7 a shows that the optimal threshold should be adjusted to 0.54. Recalculate the area under the ROC curve, and draw the optimized ROC curve Figure 7 b). Through the final calculation of the confusion matrix of the test set, the AUC has reached 0.6986, and other indicators also indicate that the model has good prediction ability Figure 7 c).

[0180] Operation E5, the construction strategy of the developmental dynamic expression pattern prediction model

[0181] E5.1: Integrate the multi-dimensional gene feature information of the whole genome genes of multiple species with the developmental dynamic expression pattern into multiple matrices, and construct multiple single-species models (SSM) through model training. That is, only for the data of one species, according to the multi-dimensional gene feature information of the gene dynamic expression pattern label, train and construct SSM. According to the embodiments of the present disclosure, SSM is trained based on the grain development related gene expression data of seven species of corn, millet, wheat, barley, oat, rice and Zephyranthes grandiflora. By performing steps S1-S3 and operations E1-E4 as described above, seven SSM models are trained and constructed, and the error loss curve and importance feature ranking graph are shown in Figure 8 , and the dynamic determination threshold division and ROC curve in the SSM construction process are shown in Figure 9 . Figure 10 a shows that the experimental results show that the seven SSM models can perform excellently in the respective species gene developmental dynamic expression pattern prediction task. For example, the AUC in rice can even exceed 0.8.

[0182] E5.2: Integrate the multi-dimensional gene feature information of the whole genome genes of multiple species and the developmental dynamic expression patterns into a matrix, and construct a cross-species model (CSM) through model training. For mixed data of multiple species, train and construct a CSM according to the multi-dimensional gene feature information of the gene dynamic expression pattern labels. Perform steps S1-S3 and operations E1-E4 as described above to train and construct the CSM model. The error loss curve and the importance feature ranking graph are shown in FIG. 5, the dynamic determination threshold division in the CSM construction process and the ROC curve are shown in FIG. 6. Figure 8 Figure 9 Figure 10 b) As described above, the CSM trained based on the expression data of grain development related genes of corn, millet, wheat, oat, rice and Zephyranthes grandiflora has a good application ability in new unknown data (barley). Therefore, for plant species not covered by the present disclosure, the CSM can be directly used to predict the developmental dynamic expression pattern of the gene to be detected in the corresponding species of grain. In summary, the present application greatly expands the range of species involved in the training data of the gene expression related model, and a large amount of experimental data shown in the embodiments shows that the present application has considerable generalization ability and applicability.

[0183] Example 3, code and its description

[0184] The following is a gene feature-based developmental dynamic expression pattern prediction method disclosed by the present application, and the code corresponds to steps S1-S3:

[0185]

[0186]

[0187] The following is a training method of a gene feature-based developmental dynamic expression pattern prediction model disclosed by the present application, and the code corresponds to operation E4:

[0188]

[0189]

[0190]

[0191]

[0192] ​​

[0193]

[0194]

[0195]

[0196] In the above embodiments, the description of each embodiment has its own focus, and the parts not described or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0197] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for predicting developmental dynamic expression patterns based on gene characteristics, characterized in that, Includes the following steps: S1. Determine the set of genes whose developmental dynamic expression patterns are to be detected; S2. Obtain multidimensional gene feature information of the gene set to be detected; S3. Input multidimensional gene feature information into the developmental dynamic expression pattern prediction model and output the developmental dynamic pattern corresponding to the target gene. The developmental dynamic expression pattern prediction model is trained based on a multidimensional gene feature information data matrix with gene dynamic expression pattern labels. The gene dynamic expression pattern labels are determined based on the gene expression matrix of the tissue or organ development process of the training sample. The operations involved in constructing the developmental dynamic expression pattern prediction model include: E1: Gene expression matrix representing the developmental process of organ systems in a species; E2: Acquisition of gene developmental dynamic expression pattern tags; E3: Acquisition of multidimensional genetic feature information of the whole genome; E4: Training method for developmental dynamic expression pattern prediction model; E5: Construction strategy for predicting developmental dynamic expression patterns; Operation E1 includes: E1.1: Obtain transcriptome sequencing data representing the developmental process of tissues or organs in a species; E1.2: Filter low-quality sequencing data and retain high-quality sequencing data; E1.3: Obtain genomic data representing the species; E1.4: Based on genomic data, the transcriptome representing the species is assembled to complete the transcriptome assembly for tissue or organ development; E1.5: Quantitative analysis of gene and transcript expression to generate a gene expression matrix for organ development process; Operation E4 includes: E4.1: Correspond the dynamic expression pattern of each gene to its respective gene characteristics to form a data matrix of multidimensional gene feature information with gene dynamic expression pattern labels; E4.2: The data matrix is ​​randomly divided into a training set and a validation set, with the training set accounting for 80% and the validation set accounting for 20%. The grid search method is used to perform hyperparameter tuning to obtain the optimal hyperparameter configuration under the current conditions. E4.3: Based on the expression patterns of genes during the developmental process, "non-developmental dynamic" genes and "developmental dynamic" genes are mapped to two types of labels, and the model is trained under the optimal hyperparameter configuration. E4.4: Monitor training and validation errors, plot error loss curves, use early stopping to determine the optimal number of training rounds to avoid model overfitting, and save the prediction model after training is completed. E4.5: Assess the contribution of each feature to the dynamic expression pattern of development; E4.6: Input the test set data into the developmental dynamic expression pattern prediction model. The model performs mapping calibration on the input features to ensure that the order of the input features is consistent with the order of the model's pre-training. E4.7: For each gene in the test set, the model traverses the set of trained decision trees and, combined with learning rate scaling, obtains the cumulative score of the predictions of all trees. E4.8: Convert the cumulative score into a probability value, determine whether gene expression is developmentally dynamic with a default threshold of 0.5, calculate the confusion matrix and the area under the ROC curve (AUC), and plot the ROC curve. E4.9: Considering that the default threshold is not the optimal threshold for dividing gene expression dynamics, the sensitivity and specificity of the model are optimized by plotting sensitivity and specificity curves and using the intersection of the two as the optimal threshold. The confusion matrix and the area under the ROC curve are recalculated, and the optimized ROC curve is plotted.

2. The method for predicting developmental dynamic expression patterns based on gene characteristics according to claim 1, characterized in that, The gene set mentioned in step S1 can be any one or any combination of the following: (1) The complete set of genes within the newly sequenced and assembled genome; (2) The set of differentially expressed genes obtained based on high-throughput sequencing; (3) A set of genes that are specifically expressed in biological tissues or organs; (4) A collection of species-specific or cross-species genes that exists within or between species; (5) A collection of candidate genes compiled through literature review; Step S2 specifically includes: S2.1: Obtain numerical features of the set of genes to be detected based on gene sequences and annotations; S2.2: Based on gene properties and annotations, obtain the categorization features of the set of genes to be detected; S2.3: Integrate the numerical and categorical features of the gene set to be detected into a data matrix containing multidimensional gene feature information.

3. The method for predicting developmental dynamic expression patterns based on gene characteristics according to claim 1, characterized in that, Step S3 specifically includes: S3.1: Load the data matrix containing multidimensional gene feature information into the developmental dynamic expression pattern prediction model. The developmental dynamic expression pattern prediction model performs mapping calibration on the input features to ensure that the order of the input features is consistent with the order of the developmental dynamic expression pattern prediction model in the pre-training process. S3.2: For each gene, the developmental dynamic expression pattern prediction model traverses the set of trained decision trees, obtains the cumulative score of the prediction values ​​of all trees, and converts the cumulative score into a probability value. S3.3: Manually set thresholds to determine the developmental dynamics of gene expression.

4. The method for predicting developmental dynamic expression patterns based on gene characteristics according to claim 1, characterized in that, Operation E5 includes: E5.1: Integrate the multidimensional gene feature information and developmental dynamic expression patterns of the whole genome of multiple species into multiple matrices, and construct multiple single-species models (SSM) through model training; E5.2: Integrate the multidimensional genetic feature information and developmental dynamic expression patterns of the whole genome of multiple species into a matrix, and construct a cross-species model CSM through model training.

5. A developmental dynamic expression pattern prediction model, characterized in that: The model is constructed by the method described in any one of claims 1 or 4.

6. A system for predicting dynamic gene expression patterns during development, used to perform the method of claim 1, characterized in that, include: M1: Gene set determination module, used to determine the set of genes to be tested for developmental dynamic expression patterns; M2: Gene feature acquisition module, used to acquire multidimensional gene feature information of the gene to be detected; M3: Developmental dynamics prediction module, used to input the multidimensional gene feature information into the developmental dynamics expression pattern prediction model and output the developmental dynamics pattern corresponding to the target gene.

Images