A method for analyzing complex evolutionary history based on deep learning

Abstract

The application provides an analysis method for complex evolutionary history based on deep learning, comprising: according to a preset species evolutionary history simulation sequence data, respectively as a model training set and a test set; determining the topological structure of the training set data, and labeling the training data in different topological structure proportions; constructing a convolutional neural network, training and testing the convolutional neural network with the data set, so that the error between the data prediction value and the label value is minimized; based on the trained convolutional neural network, analyzing real genomic sequence data, and combining other population genetics analysis methods, determining the evolutionary relationship and introgression site between different biological groups. The application uses comparative genome or population genome data, infers the topological structure between sequences through a deep learning algorithm, further evaluates the evolutionary relationship at the genome level, and identifies local introgression signals through the difference of the topological structure between different regions.

Description

Technical Field

[0001] This invention relates to multiple fields such as population genetics and deep learning, and in particular to a method for analyzing complex evolutionary history based on deep learning. Background Technology

[0002] Deciphering relationships between taxa is one of the fundamental tasks of evolutionary biology. Strictly dichotomous phylogenetic trees are widely used to describe the evolutionary history of species. Various algorithms, including distance-based, maximum likelihood, and Bayesian methods, can reconstruct phylogenetic relationships using nucleotide or protein alignment data. However, incomplete phylogenetic sorting or gene flow following species divergence can cause inconsistencies between gene trees and species trees. Furthermore, mounting evidence suggests that hybridization signals within the genome may be more prevalent than previously thought, significantly influencing the tree of life. Given the prevalence of hybridization, the strict dichotomy assumption struggles to represent the entire evolutionary history, and deciphering the complex relationships between species remains a formidable task.

[0003] With the rapid development of sequencing technology, researchers have developed various algorithms for inferring population history based on genomic sequence data. The hybrid characteristics of genomes have also been studied in various organisms, such as the sharing of beneficial alleles between crops and their wild ancestors, adaptive introgression between ancient and modern humans, and interspecific hybridization prevalent in butterflies. Specifically, these algorithms can be divided into three categories: 1) those that describe population history through statistical results using maximum likelihood or Bayesian methods, without focusing on local features, such as G-PhoCS, Treemix, and PhyloNet; 2) those that detect haplotype structure using large amounts of high-precision genomic data to compare the scale of linkage disequilibrium, such as HAPMIX, ELAI, and IBDmix; and 3) those that scan genomic windows and quantitatively describe the relationships between taxa based on allele frequencies of specific patterns, such as Patterson's D-statistic for four taxa. d Statistics and D for the five symmetric groups FOIL Compared to model testing and haplotype inference, allele frequency-based methods require less computation, but they ignore genotype location information and are only applicable to limited, predefined topologies.

[0004] With the rapid growth of computing power, deep learning has been widely applied in computer vision, speech recognition, natural language processing, and bioinformatics, such as medical image diagnosis and sequence feature recognition. Recently, deep learning algorithms have also shown potential in solving population genetics problems, such as detecting selective purges, infiltration between sister species, and inferring the topological structure of four-category taxa. Compared with traditional population genetics statistics, deep learning methods can better handle high-dimensional sequence data, comprehensively utilizing both positional and genotypic information.

[0005] Therefore, this invention proposes a method for analyzing complex evolutionary history based on deep learning. Summary of the Invention

[0006] This invention provides a method for analyzing complex evolutionary history based on deep learning. It utilizes comparative genomic or population genomic data to infer the topological structure of sequences through deep learning algorithms, thereby assessing evolutionary relationships at the genomic level and identifying local infiltration signals through differences in topological structures between different regions.

[0007] This invention provides a method for analyzing complex evolutionary history based on deep learning, comprising:

[0008] Step 1: Use the pre-set species evolutionary history simulation sequence data as the model training set and test set respectively;

[0009] Step 2: Determine the topological structure of the training set data and label the training data according to the proportion of different topological structures;

[0010] Step 3: Construct a convolutional neural network, train and test the convolutional neural network with the simulated dataset, so that the error between the predicted data value and the label value is minimized;

[0011] Step 4: Analyze real genome sequence data based on the trained convolutional neural network, and combine with other population genetics analysis methods to determine the evolutionary relationships and introgression sites between different biological groups.

[0012] In one possible implementation, the simulated sequence data includes: simulated multiple sequence alignment data based on ancestry theory, wherein the sequence is a nucleotide sequence;

[0013] The simulated data in the simulated dataset have different topological structures, and the corresponding divergence times vary within a preset range. The simulated data also includes gene flows of different times and intensities between various non-sister groups.

[0014] In one possible implementation, the topological structure of the training set data is determined, and the training data is labeled according to the proportion of different topological structures, including:

[0015] Determine all possible topologies. For the case of m classes and one outclass, there are a total of (2m-3)!! possible rooted tree topologies.

[0016] According to preset rules, the topological structure of each simulated data is determined, and the proportion of different topological structures corresponding to the simulated data is quantified. The quantization method is a topological weighting method, which performs quantization weighting by calculating the number of subtrees in the pedigree tree that are consistent with a specific topological structure.

[0017] The quantification results are labeled with the data in the form of multidimensional vectors.

[0018] In one possible implementation, a convolutional neural network is constructed, and the network is trained and tested on the simulated dataset to minimize the error between the predicted data values ​​and the label values, including:

[0019] The original sequence information is encoded, wherein the encoding method is a one-hot code;

[0020] Construct a deep neural network for feature extraction;

[0021] The high-dimensional features obtained after multiple convolution operations are flattened into one dimension. After passing through multiple fully connected layers, the output is a multi-dimensional vector corresponding to each topological probability.

[0022] The mean absolute error between the predicted value and the label value is used as the loss function of the convolutional neural network, and the parameters are optimized and updated through gradient descent and backpropagation.

[0023] The formula for calculating the Mean Absolute Error (MAE) is as follows:

[0024]

[0025] Where n is the number of all possible topologies, v i It is the predicted value of topology i. is the label value of topology i, and MAE represents the mean absolute error.

[0026] In one possible implementation, real-world genome sequence data is analyzed based on a trained convolutional neural network, including:

[0027] Obtain the assembled genome sequences at the chromosome level for different taxa and obtain the multiple alignment results of the genome sequences;

[0028] Acquire genome resequencing data from different taxa and post them back to the reference sequence to identify population-level variation datasets;

[0029] The genome-level multiple alignment dataset or population variation dataset is analyzed using a pre-defined convolutional neural network, and the evolutionary relationships between different groups are inferred based on the obtained topological structure probabilities.

[0030] In one possible approach, determining the evolutionary relationships and infiltration sites between different taxa also includes:

[0031] Based on the aforementioned variant dataset, a phylogenetic tree is reconstructed according to a preset model;

[0032] Determine the group structure;

[0033] Assessing the evolutionary history of a population;

[0034] Neutral evolutionary regions are obtained by filtering protein-coding and repetitive sequences within the genome. Then, independent sites with a first preset length within these neutral evolutionary regions are used to determine population history and gene flow between species based on multiple independent Bayesian inferences.

[0035] Based on the population history simulation, a corresponding second preset length nucleotide sequence is generated, and the simulated sequence is analyzed using a preset convolutional neural network to distinguish between infiltration signals and incomplete phylogenetic sorting signals caused by population history.

[0036] In one possible implementation, determining the topology of the training set data includes: determining the topology of each piece of data in the training set, including:

[0037] Obtain all data in the training set and collect the nucleotide sequence for each data point;

[0038] Using a pre-defined data discrimination model, nucleotide sequences in the same data are distinguished to obtain a set of distinguishable sequences;

[0039] According to the preset sequence rules, determine the topological relationships between pairs of sequences in each distinguishing sequence set;

[0040] Based on the aforementioned topological relationships, establish the initial structure for each data item;

[0041] Select characteristic sequences and common sequences from each distinguishing sequence set, and label the sequence corresponding to the maximum sequence value and the sequence corresponding to the minimum sequence value in the characteristic sequences;

[0042] Determine the sequence distance values ​​between the labeled sequence and each sequence from the initial sequence to the final sequence in the common sequence;

[0043] Simultaneously, based on the initial structure, the first topological distribution of the characteristic sequence, the second topological distribution of the common sequence, and the difference distribution between the characteristic sequence and the common sequence are determined;

[0044] The first topological distribution, the second topological distribution, and the difference distribution are corrected for distance according to the sequence distance values, and then reconstructed to obtain the topological structure.

[0045] In one possible implementation, the process of labeling training data at proportions of different topologies also includes:

[0046] The process of recording and labeling the corresponding simulated data based on the determined results is described, and the labeling process is verified. The labeling process includes N labeling sub-items, and its steps include:

[0047] Obtain the first tag generated by the current tagged sub-item, and calculate the effective execution value Y of the determination result for the current sub-item according to the following formula;

[0048]

[0049] Where L represents the total number of tags in the first tag generated by the current tag sub-item; h l1 This represents the tag attribute value of the l1th first tag, and its value range is [1,3]; δ l1 β represents the label weight value of the l1th first label, and its value ranges from [0.1, 0.8]; l1 h' represents the correctness of the l1th first tag; it takes a value of 1 when correct and 0 when incorrect. h' represents the average attribute value of all first tags generated by the current tag sub-item.

[0050] When the effective value Y is greater than the preset effective value corresponding to the current tagging sub-item, the first tagging label is transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed;

[0051] Otherwise, obtain the tagging metrics of the current tagging sub-item, and based on the difference between the executed effective value and the preset effective value, filter the relevant correction parameters from the mapping database, and establish the correspondence between the correction parameters and the corresponding tagging metrics;

[0052] Each correspondence is then modified separately with the current marked sub-item to obtain the individual modification effect of each correspondence, as well as the risk events generated by each correspondence during the modification process and the probability of occurrence of the risk events.

[0053] At the same time, the combined support capacity of all individual support effects and the individual support capacity of each individual support effect were also determined.

[0054] Based on the comprehensive auxiliary capacity and the individual auxiliary capacity, the top n1 correspondences with the strongest auxiliary effects are selected from all the individual auxiliary effects. The current tagging sub-item is then corrected, and a new first tag is obtained. The new first tag is then transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed.

[0055] The present invention has the following beneficial effects:

[0056] The deep learning model trained in this invention is robust and universal, applicable to comparative and population genome data from different biological groups, quantifying evolutionary relationships at the genome level. Its multi-dimensional output retains more effective information than phylogenetic tree models, facilitating the study of evolutionary history in the presence of gene flow.

[0057] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0058] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a method for analyzing complex evolutionary history based on deep learning, as described in an embodiment of the present invention. Detailed Implementation

[0061] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0062] Example 1: Training and Testing of Deep Learning Networks

[0063] 1. Simulated Nucleotide Sequence Data. Sequence data for training and testing convolutional neural networks were generated using the simulation software ms and Seq-Gen, based on synaptic theory. Sequence alignments with four and five taxa were simulated, respectively, with each taxa consisting of eight 5000 bp nucleotide sequences. The four-taxa and five-taxa datasets contain 120,600 and 74,100 multiple sequence alignments, respectively. The datasets represent a series of evolutionary scenarios with different topologies, where divergence times vary from 0.2 to 2.7 time units (in units of 4N generations). Furthermore, the datasets include gene flow between various non-sister taxa.

[0064] The beneficial effects of the above technical solution are:

[0065] By simulating complex evolutionary history and population structure, the model's generalization and generalization capabilities are enhanced.

[0066] 2. The topological structure represented by the data is labeled using multi-dimensional vectors. Taking the case of four groups as an example, it contains three ingroups and one outgroup, and there are at most three rooted tree topological structures among the ingroups. Therefore, a topological weighting method developed from quartetsampling is used to quantify the relative strength of each topological structure, and the quantification results are labeled in the form of multi-dimensional vectors. For example, preferably, for four-group data containing three ingroups and one outgroup, a three-dimensional vector is used to describe the proportion of the three possible topological structures; more preferably, for five-group data containing four ingroups and one outgroup, a fifteen-dimensional vector is used to describe the proportion of the fifteen possible topological structures.

[0067] The beneficial effects of the above technical solution are:

[0068] It can process individual or group data and, compared to simple dichotomous phylogenetic trees, can more effectively reflect the true relationships between taxa, especially for non-monophytic taxa or genomic regions with complex evolutionary histories.

[0069] 3. Training and Testing the Neural Network Model. The input nucleotide data is encoded in one-hot encoding, where G, T, A, and C are encoded as (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0), and (0, 0, 0, 1), respectively, while gaps ("-") or missing data ("N") are encoded as (0, 0, 0, 0). The convolutional neural networks for the four-class and five-class models alternately stack eight dense blocks and seven residual blocks, and nine dense blocks and eight residual blocks, respectively, to train deeper, more accurate, and more efficient neural networks. The high-dimensional vector obtained after multiple convolution operations on the initial data is flattened to one dimension. After passing through two fully connected layers and activation by the SoftMax function, the output is three or fifteen scores that sum to one, corresponding to the probability of each possible topology. The loss function is calculated based on the mean absolute error between the predicted and labeled values, and the parameters are optimized and updated during training using gradient descent and backpropagation to minimize the loss function.

[0070] With the training set parameters fixed, simulated datasets with different recombination rates, replacement rates, and sample sizes were further tested. For the four-class and five-class models, datasets with higher recombination rates (e.g., 0.01 and 0.1) had the same mean absolute error (Mann-Whitney U test P = 1, Bonferroni correction), while datasets with lower recombination rates (e.g., 0 and 0.001) had larger errors (Mann-Whitney U test P < 0.001, Bonferroni correction) and more outliers. For test sets with different replacement rates, both larger and smaller replacement rates led to an increase in error compared to the medium rate (Mann-Whitney U test P < 0.001, Bonferroni correction). The comparison of different sample sizes showed that the deep learning model is highly robust to different sample sizes. Although a slight decrease in accuracy was observed with decreasing sample size, there was no significant difference in error when the sample size varied between four and eight (Kruskal-Wallis test P = 0.08).

[0071] Example 2: Inferring Evolutionary Relationships and Adaptive Introgression Sites in Scaly Butterflies

[0072] 1. Obtain resequencing data of different populations of the Heliconius butterfly and post them to the reference genome sequence to identify population-level variation data. The Heliconius butterfly is known for its complex and varied Müller mimicry, and due to frequent hybridization during adaptive radiation, there are complex evolutionary relationships among species. In this example, resequencing data of three Heliconius butterfly populations (H. melapomee aglaope, Peru), H. melapomee amaryllis, and H. timareta thelxinoe, Peru) and one outgroup (H. ethilla, Brazil) were downloaded from the NCBI SRA database, with four individuals from each population (accession numbers PRJNA308754, PRJEB1749, PRJNA73595, PRJEB11772). After removing low-quality bases, the short read sequences were aligned to the H. melpomenev 2.5 reference genome (Davey et al., 2016) using the Bowtie2 alignment software (Langmead and Salzberg, 2012). Single nucleotide polymorphisms (SNPs) were identified using the "UnifiedGenotyper" command in GATK v3.7 (DePristo et al., 2011), and genotypes with quality values ​​less than 50 were removed in subsequent analyses.

[0073] 2. Analysis of interpopulation evolutionary relationships using deep learning methods. The four-group model established in Example 1 of this invention was used to analyze the population dataset, and the probabilities of three topologies were calculated in 5kb windows. Genomic-level statistical results show that, compared to *H. melpomene*, the two subspecies *H. melpomene*, *H. maglaope* and *H. mamaryllis*, are sister groups, consistent with their species divergence order. Notably, the Z chromosome showed the strongest support for the clustering of both species, likely due to the smaller effective population size on the Z chromosome, thus less affected by incomplete phylogenetic sorting. On the other hand, the region from 18700kb to 850kb on the chromosome showed the strongest support for the clustering of *H. melpomene* and *H. mamaryllis*. This region represents a known site controlling wing color patterns—B / D—which is shared through infiltration in mimicry pairs.

[0074] In summary, the embodiments of the present invention utilize a deep learning model to effectively obtain the evolutionary relationships between groups from real data through the analysis of the genome data of the butterfly population, and verify the known introgression signals.

[0075] Example 3: Detection of gene flow in Asian cultivated rice

[0076] 1. Reference genome sequences of *Oryza* species were obtained, and whole-genome sequence alignment was performed. Asian cultivated rice can be divided into two main groups—japonica rice (*O. sativa* ssp. japonica) and indica rice (*O. sativa* ssp. indica). While japonica and indica rice differ not only in morphology and genetics but also in their different wild ancestors, they share several key domestication genes, suggesting that gene flow may have played a crucial role in rice domestication. In this embodiment, chromosome-level genome assemblies of japonica, indica, and other wild rice species (*O. rufipogon*, *O. nivara*, *O. barthii*) were obtained from the OGE / IOMAP13-genome package (Stein et al., 2018), and whole-genome alignment data of *Oryza* species were generated according to a previously reported procedure (Zhang et al., 2019).

[0077] 2. Population History Inference of Rice Species. The Bayesian inference method G-PhoCS was used to analyze the divergence time and gene flow signals of the *Oryza* genus. G-PhoCS results showed that the divergence time between cultivated and wild rice in Asia was approximately 4.2 kya, slightly later than the archaeological record (approximately 9 kya), and strongly supports the migration from japonica to indica rice. While G-PhoCS analysis provides an evolutionary history at the genome level, it may miss adaptive introgression near protein-coding regions and cannot provide information on localized introgression because it only considers independent neutral sites.

[0078] 3. Identification of introgression signals within the genome using deep learning methods. The five-group model established in Example 1 of this invention was used to analyze the genome alignment data of the *Oryza* genus. Within the genome, topologies consistent with the species tree had the highest probability. Considering that incomplete phylogenetic sorting and gene flow can both cause changes in local topology, sequence simulations were performed based on the evolutionary history of rice, and the intensity of incomplete phylogenetic sorting in *Oryza* species was assessed accordingly. Compared with the modeling results without gene flow, the probabilities of the three topologies significantly increased, indicating the existence of gene flow between japonica and indica rice, and from wild rice (*O. rufipogon*) to indica rice within the genome, suggesting that gene introgression may have participated in the domestication process of indica rice. To further determine the introgression regions associated with domestication traits, sites strongly supporting the clustering of cultivated rice (topological probability greater than 0.4) were first extracted, and further analyzed using the absolute sequence divergence level (d... xyScreening was conducted to retain windows with divergence levels below the chromosome mean. A total of 71 potential introgression sites were identified, distributed across all 12 chromosomes, encompassing 1174 genes. Forty of these 71 sites overlapped with 19 previously reported selective scavenging regions associated with domestication traits such as panicle length, germination rate, grain color, stigma exposure, stigma color, tillering angle, and awn length (Huang et al., 2012). Several known domestication genes, such as OsSh1, PROG1, Bh4, OsC1, and Rc, were also located within the five introgression sites or their linkage disequilibrium ranges.

[0079] In summary, this invention, based on a deep learning algorithm, detected gene flow signals among Asian cultivated rice species at the genomic level through analysis of genome alignment data of Oryza species, and identified potential introgression sites between Japonica and Indica rice. Unlike Bayesian inference methods, this method does not require data to satisfy the assumption of neutral evolution, which is beneficial for finding adaptive introgression sites under selection pressure.

[0080] Example 4: Determining the topology of the training set data includes: determining the topology of each data point in the training set, including:

[0081] Obtain all data in the training set and collect the nucleotide sequence for each data point;

[0082] Using a pre-defined data discrimination model, nucleotide sequences in the same data are distinguished to obtain a set of distinguishable sequences;

[0083] According to the preset sequence rules, determine the topological relationships between pairs of sequences in each distinguishing sequence set;

[0084] Based on the aforementioned topological relationships, establish the initial structure for each data item;

[0085] Select characteristic sequences and common sequences from each distinguishing sequence set, and label the sequence corresponding to the maximum sequence value and the sequence corresponding to the minimum sequence value in the characteristic sequences;

[0086] Determine the sequence distance values ​​between the labeled sequence and each sequence from the initial sequence to the final sequence in the common sequence;

[0087] Simultaneously, based on the initial structure, the first topological distribution of the characteristic sequence, the second topological distribution of the common sequence, and the difference distribution between the characteristic sequence and the common sequence are determined;

[0088] The first topological distribution, the second topological distribution, and the difference distribution are corrected for distance according to the sequence distance values, and then reconstructed to obtain the topological structure.

[0089] In this embodiment, the data are all related to nucleotide sequences;

[0090] In this embodiment, the preset data differentiation model and the preset sequence rules are both pre-set;

[0091] In this embodiment, the characteristic sequence refers to a sequence that can represent the unique attributes of the data, such as the different rice varieties based on the nucleotide composition, and the common sequence refers to a sequence that represents the common attributes of the data, such as rice products composed based on the common sequence.

[0092] In this embodiment, the topological relationship can include the position and distribution of the sequence.

[0093] In this embodiment, the initial structure is based on topological relationships, which are also composed of positional and distributional relationships. However, the amount of data in the initial structure is much larger than that in the topological structure.

[0094] In this embodiment, the distinguished sequence set refers to a sequence in a certain region or space that has corresponding complete attributes, without splitting the sequence with complete attributes, which facilitates effective distance correction.

[0095] In this embodiment, the maximum and minimum sequence values ​​are used to effectively define the characteristic sequences and facilitate differentiation. The actual distance is obtained by determining the distance between the maximum and minimum sequence values ​​in the characteristic sequences and each sequence in the common sequence. The actual distance is judged by determining the first topological distribution, the second topological distribution, and the difference distribution. Finally, the reconstruction is performed by distance correction to avoid errors in subsequent analysis due to improper sequence distance values.

[0096] In this embodiment, the sequence distance value can be determined according to a standard unit distance, and the standard unit distance and the distance unit are preset.

[0097] The beneficial effects of the above technical solution are: by determining the topological relationship between pairs of sequences, it is easier to effectively establish the initial structure and improve the accuracy of the establishment; by correcting through sequence distance values, it is easier to reconstruct and obtain topological results, thereby improving their accuracy and providing an effective structural basis for subsequent changes in topological structure and sequence divergence levels.

[0098] Example 5: The process of labeling training data at different topological proportions also includes:

[0099] The process of recording and labeling the corresponding simulated data based on the determined results is described, and the labeling process is verified. The labeling process includes N labeling sub-items, and its steps include:

[0100] Obtain the first tag generated by the current tagged sub-item, and calculate the effective execution value Y of the determination result for the current sub-item according to the following formula;

[0101]

[0102] Where L represents the total number of tags in the first tag generated by the current tag sub-item; h l1 This represents the tag attribute value of the l1th first tag, and its value range is [1,3]; δ l1 β represents the label weight value of the l1th first label, and its value ranges from [0.1, 0.8]; l1 h' represents the correctness of the l1th first tag; it takes a value of 1 when correct and 0 when incorrect. h' represents the average attribute value of all first tags generated by the current tag sub-item.

[0103] When the effective value Y is greater than the preset effective value corresponding to the current tagging sub-item, the first tagging label is transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed;

[0104] Otherwise, obtain the tagging metrics of the current tagging sub-item, and based on the difference between the executed effective value and the preset effective value, filter the relevant correction parameters from the mapping database, and establish the correspondence between the correction parameters and the corresponding tagging metrics;

[0105] Each correspondence is then modified separately with the current marked sub-item to obtain the individual modification effect of each correspondence, as well as the risk events generated by each correspondence during the modification process and the probability of occurrence of the risk events.

[0106] At the same time, the combined support capacity of all individual support effects and the individual support capacity of each individual support effect were also determined.

[0107] Based on the comprehensive auxiliary capacity and the individual auxiliary capacity, the top n1 correspondences with the strongest auxiliary effects are selected from all the individual auxiliary effects. The current tagging sub-item is then corrected, and a new first tag is obtained. The new first tag is then transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed.

[0108] In this embodiment, since the corresponding simulated data is labeled based on the determined results, in order to improve the accuracy of labeling and effectively measure the labeling process, firstly, the labeling process is recorded; secondly, the labeling process is verified; and finally, the accuracy of the labeling is determined.

[0109] In this embodiment, the tagging process is pre-defined, and the tagging sub-items are also pre-defined.

[0110] In this embodiment, the current tagging sub-item may include, for example, a highlighting mark for a sequence, a missing mark for a sequence, etc., which can generate corresponding tag labels. Since the weight values ​​and tag attribute values ​​of different tags in different sub-items are different, the effective execution value of the current sub-item is calculated according to the above formula.

[0111] In this embodiment, label verification is to verify whether the label of the previous sub-item is qualified, and label processing is to perform new label marking and other operations based on the previous sub-item.

[0112] In this embodiment, the marking index is such as the highlight index or the incomplete index. The difference is obtained in order to better filter and correct the parameters, facilitate subsequent correction, improve the effectiveness of the label, and establish a corresponding relationship. In fact, it is to establish a connection between the two.

[0113] In this embodiment, the auxiliary correction of the current marked sub-item is to make individual corrections based on each correction parameter corresponding to the indicator, and the result of the correction is the individual auxiliary effect. Risk events refer to the occurrence of events such as correction failure and their probability of occurrence.

[0114] In this embodiment, the principle of integrating auxiliary results is similar to that described above.

[0115] The beneficial effects of the above technical solution are: by calculating the effective value, it is easy to determine whether to perform label verification and label processing on the current labeled sub-item, saving workload; and based on the comprehensive auxiliary capacity and the individual auxiliary capacity, the top n1 correspondences with the strongest auxiliary effects are selected from all individual auxiliary effects and corrected until the work of N labeled sub-items is completed, which helps to improve the accuracy and effectiveness of labeling and ensure the rationality of labeling.

[0116] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for analyzing complex evolutionary history based on deep learning, characterized in that, include: Step 1: Use the pre-set species evolutionary history simulation sequence data as the model training set and test set respectively; Step 2: Determine the topological structure of the training set data and label the training data according to the proportion of different topological structures; Step 3: Construct a convolutional neural network, train and test the convolutional neural network with the simulated dataset, so that the error between the predicted data value and the label value is minimized; Step 4: Analyze real genome sequence data based on the trained convolutional neural network, and combine with other population genetics analysis methods to determine the evolutionary relationships and introgression sites between different biological groups; Step 2 includes: Determine all possible topologies, where for the case of m classes and one outclass, there are a total of (2m-3)!! possible rooted tree topologies; According to preset rules, the topological structure of each simulated data is determined, and the proportion of different topological structures corresponding to the simulated data is quantified. The quantization method is a topological weighting method, which performs quantization weighting by calculating the number of subtrees in the pedigree tree that are consistent with a specific topological structure. The quantification results are labeled with the data in the form of multidimensional vectors; Determining the topological structure of each data point in the training set includes: Obtain all data in the training set and collect the nucleotide sequence for each data point; Using a pre-defined data discrimination model, nucleotide sequences in the same data are distinguished to obtain a set of distinguishable sequences; According to the preset sequence rules, determine the topological relationships between pairs of sequences in each distinguishing sequence set; Based on the aforementioned topological relationships, establish the initial structure for each data item; Select characteristic sequences and common sequences from each distinguishing sequence set, and label the sequence corresponding to the maximum sequence value and the sequence corresponding to the minimum sequence value in the characteristic sequences; Determine the sequence distance values ​​between the labeled sequence and each sequence from the initial sequence to the final sequence in the common sequence; Simultaneously, based on the initial structure, the first topological distribution of the characteristic sequence, the second topological distribution of the common sequence, and the difference distribution between the characteristic sequence and the common sequence are determined; The first topological distribution, the second topological distribution, and the difference distribution are corrected for distance according to the sequence distance values, and then reconstructed to obtain the topological structure.

2. The analytical method as described in claim 1, characterized in that, The simulated sequence data includes: simulated multiple sequence alignment data based on ancestry theory, and the sequences are nucleotide sequences; The simulated data in the simulated dataset have different topological structures, and the corresponding divergence times vary within a preset range. The simulated data also includes gene flows of different times and intensities between various non-sister groups.

3. The analytical method as described in claim 1, characterized in that, Constructing a convolutional neural network, training and testing the convolutional neural network with the simulated dataset, and minimizing the error between the predicted data values ​​and the label values, includes: The original sequence information is encoded, wherein the encoding method is a one-hot code; Construct a deep neural network for feature extraction; The high-dimensional features obtained after multiple convolution operations are flattened into one dimension. After passing through multiple fully connected layers, the output is a multi-dimensional vector corresponding to each topological probability. The mean absolute error between the predicted value and the label value is used as the loss function of the convolutional neural network, and the parameters are optimized and updated through gradient descent and backpropagation. The formula for calculating the Mean Absolute Error (MAE) is as follows: in, n It is the number of all possible topologies. v i It is topology i The predicted value, v ^ i It is topology i The label value, MAE represents the mean absolute error.

4. The analytical method as described in claim 1, characterized in that, Analysis of real genome sequence data based on trained convolutional neural networks, including: Obtain the assembled genome sequences at the chromosome level for different taxa and obtain the multiple alignment results of the genome sequences; Acquire genome resequencing data from different taxa and post them back to the reference sequence to identify population-level variation datasets; The genome-level multiple alignment dataset or population variation dataset is analyzed using a pre-defined convolutional neural network, and the evolutionary relationships between different groups are inferred based on the obtained topological structure probabilities.

5. The analytical method as described in claim 4, characterized in that, Determining the evolutionary relationships and introgression sites between different taxa also includes: Based on the aforementioned variant dataset, a phylogenetic tree is reconstructed according to a preset model; Determine the group structure; Assessing the evolutionary history of a population; Neutral evolutionary regions are obtained by filtering protein-coding and repetitive sequences within the genome. Then, independent sites with a first preset length within these neutral evolutionary regions are used to determine population history and gene flow between species based on multiple independent Bayesian inferences. Based on the population history simulation, a corresponding second preset length nucleotide sequence is generated, and the simulated sequence is analyzed using a preset convolutional neural network to distinguish between infiltration signals and incomplete phylogenetic sorting signals caused by population history.

6. The analytical method as described in claim 1, characterized in that, The process of labeling training data according to different topological proportions also includes: The process of recording and labeling the corresponding simulated data based on the determined results is described, and the labeling process is verified. The labeling process includes N labeling sub-items, and its steps include: Obtain the first tag generated by the current tagged sub-item, and calculate the effective execution value Y of the determination result for the current sub-item according to the following formula; ； Where L represents the total number of tags in the first tag generated by the current tag sub-item; Indicates the first The tag attribute value of the first tag, and the value range is [1,3]; Indicates the first The label weight value of the first tag is set, and the value range is [0.1, 0.8]. Indicates the first The correctness of the first tag is determined by a value of 1 when it is correct and 0 when it is incorrect. This represents the average attribute value of all first-level tags generated by the current tag sub-item; When the effective value Y is greater than the preset effective value corresponding to the current tagging sub-item, the first tagging label is transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed; Otherwise, obtain the tagging metrics of the current tagging sub-item, and based on the difference between the executed effective value and the preset effective value, filter the relevant correction parameters from the mapping database, and establish the correspondence between the correction parameters and the corresponding tagging metrics; Each correspondence is then modified separately with the current marked sub-item to obtain the individual modification effect of each correspondence, as well as the risk events generated by each correspondence during the modification process and the probability of occurrence of the risk events. At the same time, the combined support capacity of all individual support effects and the individual support capacity of each individual support effect were also determined. Based on the comprehensive auxiliary capacity and the individual auxiliary capacity, the top n1 correspondences with the strongest auxiliary effects are selected from all the individual auxiliary effects. The current tagging sub-item is then corrected, and a new first tag is obtained. The new first tag is then transferred to the next tagging sub-item for tag verification and tag processing, until the work of N tagging sub-items is completed.

Images