Whole genome prediction method based on chromosome specificity and related device

Abstract

The embodiment of the invention provides a whole genome prediction method based on chromosome specificity and a related device. The method comprises the following steps: acquiring a to-be-detected gene sequence, and segmenting the to-be-detected gene sequence according to preset chromosome length information to obtain a plurality of chromosome sequences; performing local feature extraction on each chromosome sequence to obtain a first chromosome feature corresponding to each chromosome sequence; generating a first global representation sequence corresponding to the to-be-detected gene sequence based on the first chromosome feature corresponding to each chromosome sequence; and performing phenotype prediction based on the first global representation sequence and the plurality of first chromosome features corresponding to the to-be-detected gene sequence to obtain a phenotype prediction result corresponding to the to-be-detected gene sequence. The method can improve the whole genome prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and in particular, relates to a chromosome-specific-based whole genome prediction method and related device. BACKGROUND

[0002] Whole genome selection is a breeding method based on whole genome data, which uses a large number of markers covering the entire genome to select individuals with desired traits. Whole genome prediction provides a method to estimate the phenotype value of an individual based on its genotype, that is, to establish a prediction model between genotype and phenotype for prediction. Whole genome selection can use the predicted phenotype value to guide actual breeding decisions, significantly improve genetic gain per unit time or cost, speed up the breeding process, and reduce the time and cost of traditional phenotype screening.

[0003] Deep learning-based whole genome prediction automatically identifies and extracts high-order features through a multi-layer neural network architecture in a data-driven manner, establishes the correlation between genotype and phenotype, and improves the prediction ability of genotype-based phenotype prediction.

[0004] However, the accuracy of the current deep learning-based whole genome prediction is not high. SUMMARY

[0005] The main purpose of the present application is to provide a chromosome-specific-based whole genome prediction method and related device for improving the accuracy of whole genome prediction.

[0006] To achieve the above-mentioned purpose, one aspect of the present application provides a chromosome-specific-based whole genome prediction method, comprising:

[0007] Obtaining a to-be-tested gene sequence, and dividing the to-be-tested gene sequence according to preset chromosome length information to obtain a plurality of chromosome sequences;

[0008] Respectively performing local feature extraction on each chromosome sequence to obtain a first chromosome feature corresponding to each chromosome sequence;

[0009] Generating a first global representation sequence corresponding to the to-be-tested gene sequence based on the first chromosome feature corresponding to each chromosome sequence;

[0010] Performing phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

[0011] Another aspect of the present application provides a chromosome-specific-based whole genome prediction device, comprising:

[0012] The acquisition unit is configured to acquire a to-be-tested gene sequence, and split the to-be-tested gene sequence according to preset chromosome length information to obtain a plurality of chromosome sequences.

[0013] The extraction unit is configured to respectively perform local feature extraction on each chromosome sequence to obtain a first chromosome feature corresponding to each chromosome sequence.

[0014] The generation unit is configured to generate a first global representation sequence corresponding to the to-be-tested gene sequence based on the first chromosome features corresponding to the respective chromosome sequences.

[0015] The prediction unit is configured to perform phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

[0016] Optionally, in some embodiments, the extraction unit includes:

[0017] The word segmentation subunit is configured to perform word segmentation processing on each chromosome sequence based on a plurality of preset base sequence fragments, and encode the obtained word segmentation result to generate an index sequence corresponding to the chromosome sequence.

[0018] The embedding subunit is configured to respectively perform embedding processing on each index sequence to obtain an embedding representation corresponding to the chromosome sequence.

[0019] The extraction subunit is configured to respectively perform local feature extraction on the embedding representation corresponding to each chromosome sequence to obtain the first chromosome feature.

[0020] Optionally, in some embodiments, the prediction unit includes:

[0021] The down-sampling subunit is configured to perform down-sampling processing on the first chromosome feature corresponding to each chromosome sequence to obtain a second chromosome feature corresponding to each chromosome sequence.

[0022] The first prediction subunit is configured to perform phenotype prediction based on the first global representation sequence and a plurality of second chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

[0023] Optionally, in some embodiments, the generation unit includes:

[0024] The generation subunit is configured to generate a global representation of each chromosome sequence based on the second chromosome feature corresponding to each chromosome sequence.

[0025] The splicing subunit is configured to splice the global representations corresponding to the respective chromosome sequences to obtain the first global representation sequence corresponding to the to-be-tested gene sequence.

[0026] Optionally, in some embodiments, the generating subunit comprises:

[0027] a determining module configured to determine a global representation quantity corresponding to each chromosome sequence based on a length of the second chromosome feature corresponding to the chromosome sequence and a preset proportion parameter;

[0028] a generating module configured to generate the global representation corresponding to each chromosome sequence based on the second chromosome feature and the global representation quantity.

[0029] Optionally, in some embodiments, the predicting unit comprises:

[0030] a first self-attention subunit configured to perform multi-head self-attention processing on the first global representation sequence to obtain a second global representation sequence;

[0031] a second self-attention subunit configured to splice the second global representation sequence with a plurality of second chromosome features corresponding to the to-be-tested gene sequence, and perform multi-head self-attention processing on the splicing result to obtain a third global representation sequence corresponding to the second global representation sequence and a third chromosome feature corresponding to the second chromosome feature;

[0032] an up-sampling subunit configured to perform up-sampling processing on the third global representation sequence to obtain a fourth global representation sequence with a total size same as that of the plurality of third chromosome features corresponding to the to-be-tested gene sequence;

[0033] a fusion subunit configured to add the fourth global representation sequence and the plurality of third chromosome features corresponding to the to-be-tested gene sequence bit by bit to obtain a fusion correlation feature;

[0034] a second predicting subunit configured to perform phenotype prediction based on the fusion correlation feature to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

[0035] Optionally, in some embodiments, the predicting unit further comprises:

[0036] a first position encoding subunit configured to perform position encoding on the first global representation sequence;

[0037] a second position encoding subunit configured to perform position encoding on the second chromosome feature corresponding to each chromosome sequence.

[0038] Another aspect of the present application provides an electronic device comprising:

[0039] a memory, a transceiver, a processor, and a bus system;

[0040] wherein the memory is configured to store a program;

[0041] the processor is configured to execute the program in the memory, including executing the method of the above aspects.

[0042] Bus systems are used to connect memory and processor to enable communication between them.

[0043] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0044] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0045] This method segments the gene sequence to be tested into chromosome sequences corresponding to different chromosomes. Local feature extraction is performed on each chromosome sequence, allowing for the extraction of specific information for that chromosome, resulting in the corresponding first chromosome feature. After generating a first global representation sequence in the gene sequence based on these first chromosome features, the first global representation sequence can be combined with the first chromosome features corresponding to all chromosomes to predict the phenotype. By combining chromosome-specific information with inter-chromosomal differences represented by the global representation, prediction accuracy can be improved. Attached Figure Description

[0046] Figure 1 This is a schematic flowchart of the chromosome-specific whole-genome prediction method provided in the embodiments of this application;

[0047] Figure 2 yes Figure 1 A flowchart illustrating step 102 in the middle section;

[0048] Figure 3 yes Figure 1 A flowchart illustrating step 104 in the middle section;

[0049] Figure 4 yes Figure 1 A flowchart illustrating step 103;

[0050] Figure 5 yes Figure 4 A flowchart illustrating step 401;

[0051] Figure 6 yes Figure 1 A flowchart illustrating step 104 in the middle section;

[0052] Figure 7 This is a schematic diagram of the preprocessing flow of the first global representation sequence and the second chromosome feature provided in the embodiments of this application;

[0053] Figure 8 This is a schematic diagram of the whole-genome prediction architecture provided in the embodiments of this application;

[0054] Figure 9is a phenotypic prediction process schematic diagram provided by an embodiment of the present application;

[0055] Figure 10 is another phenotypic prediction process schematic diagram provided by an embodiment of the present application;

[0056] Figure 11 is a structure schematic diagram of a whole genome prediction device based on chromosome specificity provided by an embodiment of the present application;

[0057] Figure 12 is a structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0059] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a manner different from the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0060] The word "exemplary" used herein means "serving as an example, embodiment or illustration". Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

[0061] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as processing circuitry or memory) or a combination thereof. Similarly, one processor (or multiple processors or memory) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0062] In addition, in order to better illustrate the present application, numerous specific details are given in the specific embodiments below. Those skilled in the art should understand that the present application can also be implemented without some specific details. In some examples, methods, means, elements and circuits that are well known to those skilled in the art are not described in detail, in order to highlight the main idea of the present application.

[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0064] First, let's analyze some of the terms involved in this application:

[0065] Chromosome specificity refers to the differences between different chromosomes or chromosome regions in terms of gene expression, function, or genetic effects at the genetic and molecular levels. This difference can be reflected in gene expression patterns, biological functions, and genetic effects, and helps to understand how genes affect the phenotype and disease of an organism. For example, some genes may only be expressed in specific tissues, or have different activities at different developmental stages, and the expression difference of these genes is a manifestation of chromosome specificity. In addition, certain regions on the chromosome may have a stronger association with certain genetic diseases, which is also a manifestation of chromosome specificity.

[0066] Genome-wide prediction (GWP) is a method that uses genetic markers such as single nucleotide polymorphisms (SNPs) in the entire genome to predict individual trait performance. This method attempts to identify genetic variations associated with specific traits by analyzing genetic information throughout the genome, and makes predictions accordingly. Genome-wide prediction has been widely used in agriculture, medicine, and ecology, such as predicting crop yield and disease resistance in agriculture, predicting disease risk and drug response in medicine, and predicting species adaptability to environmental changes in ecology.

[0067] Genomic selection (GS) is a breeding method based on whole genome data, which uses a large number of markers covering the entire genome to select individuals with desired traits. Compared with traditional breeding methods, GS can more comprehensively utilize genomic information, thereby improving the accuracy and efficiency of selection. By analyzing genetic markers throughout the genome, GS can predict the genetic value of individuals, thereby accelerating the breeding process and saving costs. Genome-wide prediction, as a key technical means of GS, the accuracy of the prediction model it establishes between genotype and phenotype directly affects the reliability of genetic evaluation and the effectiveness of breeding decisions. Deep learning-based genome-wide prediction uses a multi-layer neural network architecture to automatically identify and extract high-order features in a data-driven manner, establishing the relationship between genotype and phenotype, and improving the prediction ability of phenotype prediction based on genotype.

[0068] Traditional deep learning models such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) have been widely applied in genomics tasks. MLP is typically used in genomics to construct a deep neural network capable of learning nonlinear relationships in genomic data. MLPs, by stacking multiple fully connected layers, each containing a certain number of neurons, and introducing nonlinearity through activation functions, can fit complex patterns. However, MLP-based genomics models struggle to capture repetitive local structures or sequence information in the input data, limiting in-depth understanding of the functions of key gene fragments. Furthermore, their numerous and complex parameters make them prone to overfitting in crop breeding scenarios with limited sample sizes. CNNs, through their unique convolutional layers, can capture local features, such as identifying specific patterns or structural variations in the genome. RNNs, through variants such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), are designed with memory units to store information from previous time steps, enabling the network to capture local patterns and contextual relationships within a sequence. However, while CNNs and RNNs have certain advantages in extracting local features, they lack the ability to model global dependencies in extremely long genotype sequences.

[0069] Another type of deep learning model, attention-based models (such as Transformer), can transform genomic data into embedding vectors. The model's encoder learns to map each genetic marker to a high-dimensional space to capture richer feature representations. Then, the model computes attention scores for each marker to all other markers in parallel through multi-head self-attention layers; these scores reflect the importance of the relationships between markers. In this way, based on the ability to dynamically compute attention weights, it can capture complex long-range dependencies in the genome, and the model can identify the genetic markers most influential in predicting specific traits. However, it has a high computational cost when modeling whole-genome sequences, affecting its efficiency in long-sequence analysis. Furthermore, it does not focus on chromosome-specific properties and is insufficient in capturing cross-chromosomal interactions and establishing long-range spatial dependencies, resulting in low prediction accuracy.

[0070] To address the aforementioned issues, this application discloses a chromosome-specific whole-genome prediction method. This method includes acquiring a gene sequence to be tested and segmenting it according to preset chromosome length information to obtain multiple chromosome sequences; extracting local features from each chromosome sequence to obtain a first chromosome feature corresponding to each chromosome sequence; generating a first global representation sequence corresponding to the gene sequence to be tested based on the first chromosome features of each chromosome sequence; and performing phenotypic prediction based on the first global representation sequence and the multiple first chromosome features corresponding to the gene sequence to be tested to obtain the phenotypic prediction result corresponding to the gene sequence to be tested.

[0071] This method segments the gene sequence to be tested into chromosome sequences corresponding to different chromosomes. Local feature extraction is performed on each chromosome sequence, allowing for the extraction of specific information for that chromosome, resulting in the corresponding first chromosome feature. After generating a first global representation sequence in the gene sequence based on these first chromosome features, the first global representation sequence can be combined with the first chromosome features corresponding to all chromosomes to predict the phenotype. By combining chromosome-specific information with inter-chromosomal differences represented by the global representation, prediction accuracy can be improved.

[0072] The following describes the chromosome-specific whole-genome prediction method provided in the embodiments of this application.

[0073] Please see Figure 1 In some embodiments, the chromosome-specific whole-genome prediction method provided in this application includes, but is not limited to, steps 101 to 104.

[0074] Step 101: Obtain the gene sequence to be tested, and divide the gene sequence to be tested according to the preset chromosome length information to obtain multiple chromosome sequences;

[0075] Step 102: Extract local features from each chromosome sequence to obtain the first chromosome feature corresponding to each chromosome sequence;

[0076] Step 103: Based on the first chromosome features corresponding to each chromosome sequence, generate the first global representation sequence corresponding to the gene sequence to be tested;

[0077] Step 104: Perform phenotypic prediction based on multiple first chromosome features corresponding to the first global representation sequence and the gene sequence to be tested, and obtain the phenotypic prediction results corresponding to the gene sequence to be tested.

[0078] Steps 101 to 104, as illustrated in the embodiments of this application, involve splitting the gene sequence to be tested into independent chromosome sequences for local feature extraction, i.e., extracting specific information from each independent chromosome sequence to generate corresponding first chromosome features. Then, based on the first chromosome features of each chromosome sequence, a first global representation sequence is generated for the entire gene sequence to be tested. Finally, the phenotype is predicted based on the first chromosome features and the first global representation sequence. The global representation sequence can represent the association information of multiple chromosome sequences in the gene sequence to be tested, i.e., it integrates local information from each chromosome and global information across chromosomes to improve the accuracy of phenotype prediction.

[0079] In step 101 of some embodiments, the gene sequence to be tested can be the genome sequence of an organism that needs to be analyzed and evaluated; here, it can be the genome sequence for which a phenotype needs to be predicted. Here, phenotype can refer to an observable trait or characteristic of an organism, including appearance, physiological characteristics, behavior, etc.

[0080] Preset chromosome length information refers to pre-defined data regarding chromosome length, which can be determined based on the length of each chromosome in a known genomic reference sequence. Using this preset chromosome length information, the entire gene sequence to be tested can be divided into multiple parts, each corresponding to one chromosome; that is, each part can be called a chromosome sequence.

[0081] In step 102 of some embodiments, local feature extraction may involve applying feature extraction techniques to each chromosome sequence to obtain the composition, structural features, or other genetic markers of the chromosome sequence. Local feature extraction can be performed using a convolutional neural network (CNN), a Long Short-Term Memory (LSTM) network, or a Recurrent Neural Network (RNN). This embodiment uses a CNN as an example. A CNN slides a convolutional kernel across the input sequence and calculates the dot product (inner product) between the kernel and the sequence segment at each position to generate a feature map. Here, each chromosome sequence can be processed by a separate convolutional kernel. The sliding kernel traverses the chromosome sequence, and the dot product between the kernel and the sequence segment is calculated to extract local features. These local features represent chromosome window markers, referred to in this application as the first chromosome feature (e.g., Figure 10 (As shown). In addition, for the purpose of dimensionality reduction, the first chromosome feature was downsampled in the embodiments of this application, and therefore the downsampled feature is referred to as the second chromosome feature.

[0082] In step 103 of some embodiments, the global representation sequence can refer to features that reflect the association or interaction between different regions in a biological sequence. Here, it can be constructed based on the global representation of multiple segments divided into each chromosome sequence (also called the vector marker of each segment), that is, representing the association or interaction of features between different chromosome sequences, capturing the long-range dependency between chromosomes. The global representation sequence here can be called the first global representation sequence (also called the vector marker sequence, such as...). Figure 10 (As shown).

[0083] In step 104 of some embodiments, phenotypic prediction can be a process of predicting the phenotypic characteristics of an organism using genetic, bioinformatics, and statistical methods. Here, the first chromosome features corresponding to each chromosome in the gene sequence to be tested can be integrated to form a genome feature representation. This feature representation can contain local information of all chromosomes in the gene sequence to be tested. Then, the local information of all chromosomes and the first global representation sequence representing the global interaction between chromosomes are combined to form a global label for phenotypic prediction. The phenotypic characteristics of the organism corresponding to the gene sequence to be tested can be obtained. This phenotypic characteristic is the phenotypic prediction result, which can be a continuous numerical value or a classification label, etc.

[0084] Please see Figure 2 In some embodiments, step 102 includes, but is not limited to, steps 201 to 202.

[0085] Step 201: Based on multiple preset base sequence fragments, each chromosome sequence is segmented into words, and the segmentation results are encoded to generate an index sequence corresponding to the chromosome sequence.

[0086] Step 202: Perform embedding processing on each index sequence to obtain the embedding representation corresponding to the chromosome sequence;

[0087] Step 203: Extract local features from the embedding representation corresponding to each chromosome sequence to obtain the first chromosome features.

[0088] In this embodiment, the base sequence fragment can be a short sequence in a DNA molecule. Here, based on word segmentation, each chromosome sequence can be divided into multiple short sequences of fixed length. The preset base sequence fragment can be the base sequence that the fixed-length short sequence may present.

[0089] The word segmentation process can be non-overlapping Kmer segmentation, which divides a long sequence into multiple consecutive sequence fragments of length K, where these fragments do not overlap. Here, each chromosome sequence can be segmented according to a fixed length K, forming multiple non-overlapping subsequences. For example, for the consecutive nucleic acid sequence "ATCTGC", when K=3, the sequence can be divided into two non-overlapping nucleic acid sequence fragments: "ATC" and "TGC".

[0090] An index sequence represents a unique identifier pre-constructed for each predetermined base sequence fragment. Each predetermined base sequence fragment can be mapped to a code, thus mapping each base sequence fragment represented by the segmentation result of the chromosome sequence to its corresponding code. This converts the chromosome sequence into a numerical form, which can be called an index sequence. For example, k-mer indexing is a method of mapping short sequences (k-mers) to unique integers. A k-mer refers to a nucleotide sequence fragment of length k. Taking k=3 as an example, it means that each k-mer consists of 3 consecutive bases (A, T, C, G). Since each base has 4 possibilities, a k-mer of length 3 has a total of 4... 3 =64 different combinations, AAA, AAC, AAG, ..., TTT can be numbered from 0 to 63. Assuming the sequence AACTG is segmented into AAC and GTG, its corresponding integer index sequence can be [1, 45].

[0091] Embedding processing can be used to learn the effective feature representation of each k-mer (a base segment of length k) in a chromosome sequence, capturing pattern and structural information in the sequence, and combining the feature representations of each k-mer in the chromosome sequence to construct the embedding representation corresponding to the chromosome sequence. In this way, by mapping discrete indices to a high-dimensional continuous space, it is convenient to capture higher-order relationships between genes, providing a foundation for models to capture higher-order relationships and interactions between genes that may affect the prediction of phenotypic traits.

[0092] Local feature extraction is performed on the embedding representation of each chromosome sequence, and further features of the chromosome sequence are extracted from the embedding representation. In this embodiment, the convolutional neural network includes convolutional layers and downsampling layers. A one-dimensional convolutional layer is used to extract features from the embedding representation to obtain the first chromosome feature. The first chromosome feature is then input into the downsampling layer to further shorten the sequence length and adjust the feature dimension to obtain the second chromosome feature.

[0093] By segmenting chromosome sequences into smaller index sequences and then performing embedding and local feature extraction, computational complexity can be significantly reduced.

[0094] Please see Figure 3In some embodiments, step 104 includes, but is not limited to, steps 301 to 302.

[0095] Step 301: Downsample the first chromosome feature corresponding to each chromosome sequence to obtain the second chromosome feature corresponding to each chromosome sequence;

[0096] Step 302: Based on the first global representation sequence and multiple second chromosome features corresponding to the gene sequence to be tested, perform phenotypic prediction to obtain the phenotypic prediction results corresponding to the gene sequence to be tested.

[0097] In this implementation, downsampling can be an operation that reduces the sequence length of chromosomal features based on the downsampling step size while retaining the most important information. This is used to compress sequence data and extract key genetic markers, specifically key information from the first chromosome feature. Downsampling can be implemented using various methods, such as max pooling, average pooling, or global average pooling. For example, max pooling can select the maximum value in each pooling window as a representative feature to extract the most significant features. Because downsampling reduces the sequence length of chromosomal features, phenotypic prediction based on the second chromosome feature can reduce the computational complexity during inference, improving processing speed and efficiency.

[0098] Please see Figure 4 In some embodiments, after locally extracting the first chromosome feature from each chromosome sequence and downsampling the first chromosome feature to obtain the second chromosome feature, step 103 includes, but is not limited to, steps 401 to 402.

[0099] Step 401: Generate a global representation of each chromosome sequence based on the second chromosome features corresponding to each chromosome sequence;

[0100] Step 402: Concatenate the global representations corresponding to each chromosome sequence to obtain the first global representation sequence corresponding to the gene sequence to be tested.

[0101] In this embodiment, global representation can refer to information extracted from biological sequences that can represent key attributes of chromosome sequences, and in this embodiment, it can also be referred to as a carrier marker for each chromosome sequence.

[0102] By concatenating the global representations corresponding to each chromosome sequence into a higher-dimensional tensor, the local features of each chromosome segment can be aggregated. This higher-dimensional tensor can be called the first global representation sequence (carrier marker sequence), so that the local representations (second chromosome features) and global representations (carrier markers) corresponding to each chromosome sequence can be concatenated into a global marker for phenotypic prediction.

[0103] Please seeFigure 5 In some embodiments, step 401 includes, but is not limited to, steps 501 to 502.

[0104] Step 501: Based on the length of the second chromosome feature corresponding to each chromosome sequence and the preset ratio parameter, determine the number of global representations corresponding to each chromosome sequence;

[0105] Step 502: Generate a global representation for each chromosome sequence based on the second chromosome features and the number of global representations.

[0106] In this embodiment, the preset ratio parameter can be a fixed length value that divides the second chromosome feature into multiple parts. That is, based on the second chromosome feature of different lengths, it can be divided into a different number of feature segments based on the preset ratio parameter.

[0107] The number of global representations can refer to the total number of global representations extracted from each chromosome sequence. For example, when the preset ratio parameter is set to 16: for a sequence with a length of 325 for the second chromosome feature, 325 / 16 = 20 global representations will be generated (rounded down); for a sequence with a length of 160 for the second chromosome feature, 160 / 16 = 10 global representations will be generated.

[0108] Understandably, when the length of the second chromosome feature is not an integer multiple of the preset ratio parameter, the number of global representations is rounded down, and the length of the feature segment in the second chromosome feature corresponding to each global representation is re-determined based on the number of global representations when generating the global representation. The method for generating the global representation for each chromosome sequence based on the second chromosome features and the number of global representations can be as follows: determine the size and stride of the pooling window based on the number of global representations, apply a one-dimensional average pooling operation to each second chromosome feature, which can include dividing each second chromosome feature into multiple feature segments, each with a length equal to the pooling window size, and then calculating the average value of each segment; this average value can be called the global representation. A smaller number of features in the global representation can speed up data processing, and determining the number of global representations based on the actual length of the chromosome avoids losing important data information.

[0109] Please see Figure 6 In some embodiments, after downsampling the first chromosome feature to obtain the second chromosome feature, step 104 includes, but is not limited to, steps 601 to 605.

[0110] Step 601: Perform multi-head self-attention processing on the first global representation sequence to obtain the second global representation sequence;

[0111] Step 602: Concatenate the second global representation sequence with multiple second chromosome features corresponding to the gene sequence to be tested, and perform multi-head self-attention processing on the concatenation result to obtain the third global representation sequence corresponding to the second global representation sequence and the third chromosome feature corresponding to the second chromosome feature;

[0112] Step 603: Upsample the third global representation sequence to obtain a fourth global representation sequence with the same total size as the multiple third chromosome features corresponding to the gene sequence to be tested;

[0113] Step 604: Add the multiple third chromosome features corresponding to the fourth global representation sequence and the gene sequence to be tested positionally to obtain the fusion association feature;

[0114] Step 605: Perform phenotypic prediction based on fusion association features to obtain the phenotypic prediction results corresponding to the gene sequence to be tested.

[0115] In this implementation, multi-head self-attention can be an attention mechanism used in the Transformer architecture of deep learning models. It processes input data through multiple attention heads, each learning the relationships between different parts of the input sequence. This method captures dependencies between different positions in the sequence, thereby improving the model's understanding of the sequence data. Here, dependencies between different positions in the first global representation sequence can be captured to obtain the second global representation sequence.

[0116] A richer feature representation (which can be called a global label) is formed through splicing operations. This representation contains both intrachromosomal information and global information of the entire genome. Then, multi-head self-attention processing can be applied to the splicing result. The multi-head self-attention mechanism can learn different representations of the sequence in parallel across multiple subspaces, with each head focusing on different aspects of the sequence. By calculating attention scores between different features to identify and weight the most important features, it adaptively focuses on key parts of the spliced ​​features, thereby extracting higher-level feature representations and establishing connections between different positions within the sequence, thus capturing complex dependencies. The global representation sequence and chromosomal features in this self-attention processing result can be separated, resulting in a third global representation sequence corresponding to the second global representation sequence and a third chromosomal feature corresponding to the second chromosomal feature (e.g.,...). Figure 10 (As shown).

[0117] Understandably, in order to facilitate the integration of the third global representation sequence and the third chromosome features, this embodiment uses upsampling (such as interpolation or deconvolution) to adjust the length of the third global representation sequence to be the same as the total length of multiple third chromosome features, thereby generating the fourth global representation sequence.

[0118] Positional addition refers to adding the elements at the same positions of two feature vectors together to generate a new feature vector. Here, the fourth global representation sequence and the third chromosome feature are added element-wise and fused to obtain a fused associated feature, which can retain the unique information of the two features and strengthen the cooperative signal between them. In this embodiment, a hierarchical attention mechanism is used to strengthen long-distance and global dependencies, and element-wise addition achieves information complementarity, ultimately generating a more interpretable and predictive feature representation and improving prediction accuracy.

[0119] Please see Figure 7 In some embodiments, the preprocessing procedure for the first global representation sequence and the second chromosome feature includes, but is not limited to, steps 701 to 702.

[0120] Step 701: Perform position encoding on the first global representation sequence;

[0121] Step 702: Encode the position of the second chromosome feature corresponding to each chromosome sequence.

[0122] In this implementation, the multi-head self-attention mechanism in the Transformer architecture is not sensitive to the order of input elements. Introducing positional encoding helps the multi-head self-attention mechanism to more accurately consider the relationship between elements at different positions when calculating attention weights. That is, when focusing on the features of a certain position, it can simultaneously consider the distance information between that position and other positions, thereby more effectively capturing long-distance dependencies.

[0123] Positional encoding can be a method of assigning a fixed value or vector to each position in a sequence, enabling the model to recognize and utilize the order and positional relationships of the elements in the sequence. This can be done by adding positional information element by element or by positional embedding.

[0124] After downsampling (i.e., second chromosome features) based on the first chromosome features corresponding to each chromosome sequence, a first global representation sequence (vector marker sequence) corresponding to the gene sequence to be tested is generated. Then, position information can be added to the first chromosome features and the first global representation sequence by position encoding, so that after multiple second chromosome features and global representations (vector markers) are spliced ​​into a global marker, the multi-head self-attention mechanism can perceive the position information of the second chromosome features after position encoding and the position information of the global representation after position encoding in the global marker and perform self-attention processing.

[0125] By providing relative positional information of elements in a sequence through positional encoding, the multi-head self-attention mechanism can capture the dependencies between different positions in the sequence, thereby improving feature extraction efficiency.

[0126] In one example, step 104 can be phenotypic prediction based on a pre-trained phenotypic prediction model. The pre-training process of the phenotypic prediction model can be based on the sample gene sequence and the corresponding phenotypic label. The sample gene sequence is processed according to steps 101 to 103 to obtain the sample chromosome features corresponding to each sample chromosome sequence in the sample gene sequence, as well as the sample global representation sequence corresponding to the sample gene sequence. Then, based on the phenotypic prediction model, phenotypic prediction can be performed on the sample global representation sequence and multiple sample chromosome features to obtain the phenotypic prediction value. Then, the parameters of the phenotypic prediction model are updated based on the difference between the phenotypic prediction value and the phenotypic label.

[0127] In this context, the sample gene sequence and its corresponding phenotypic label can refer to the sample dataset used to train the phenotypic prediction model. The correspondence between the sample gene sequence and the phenotypic label can refer to the observable traits or characteristics of the organism corresponding to the provided sample gene sequence, which can be called the phenotypic label. The sample gene sequence and its corresponding phenotypic label can be obtained through the internet or input manually. Using preset chromosome length information, the entire sample gene sequence can be divided into multiple parts, each part corresponding to one chromosome; each part can then be called the sample chromosome sequence.

[0128] Each sample chromosome sequence is convolved by a convolution kernel. The convolution kernel is traversed through the sample chromosome sequence, and the dot product between the convolution kernel and the sequence segment is calculated to extract local features, which can be called sample chromosome features.

[0129] Based on the correlations between sample chromosome features of different sample chromosome sequences, long-range dependencies between chromosomes are captured, generating a global representation sequence for the sample. By combining the global representation sequence and the sample chromosome features of all chromosomes, phenotypic prediction can be performed to obtain the phenotypic prediction value of the organism corresponding to the sample gene sequence.

[0130] Understandably, phenotypic predictions are the model's inferences about the sample gene sequences, while phenotypic labels are the true labels of the sample gene sequences. Training a phenotypic prediction model can help it make its predictions increasingly closer to the true answers; that is, the smaller the gap between the model's predictions and the true answers, the more accurate the model's predictions. The parameters of the phenotypic prediction model can be adjusted by defining this gap. By continuously inputting sample gene sequences and phenotypic labels, calculating the gap, and adjusting the parameters, the model's guesses can be made closer and closer to the true answers, thus gradually improving the accuracy of the model's predictions.

[0131] The chromosome-specific whole-genome prediction method provided in this application will be described in detail below with a specific embodiment. Figure 8The diagram shown is a schematic of the whole genome prediction architecture provided in the embodiments of this application, including a gene sequence preprocessing module 801, a chromosome-specific convolution module 802, an adaptive chromosome vector marker generator 803, a hierarchical attention module 804, and a phenotype prediction head module 805.

[0132] Among them, the gene sequence preprocessing module 801 is used to perform Figure 2 Steps 201 and 202 involve receiving the input gene sequence to be tested. The sequence is segmented according to preset chromosome length information to obtain the corresponding chromosome sequences. Then, k-mer segmentation is applied to convert each chromosome sequence into a corresponding k-mer index. Finally, the index is mapped to an embedding representation based on the embedding layer. The segmentation operation shortens the sequence length through non-overlapping k-mer processing, reducing subsequent computational complexity. The embedding process maps the index to a high-dimensional space.

[0133] Specifically, the implementation process of the gene sequence preprocessing module 801 can be as described in steps S11 to S15.

[0134] Step S11: Receive the gene sequence X to be tested, which has the shape (B,L) total In this context, shape refers to the dimensional structure of a tensor, used to describe the structure of data in a deep learning model; B represents the batch size, i.e., the number of samples input into the model in parallel each time; and L... total This represents the total length of all chromosome sequences.

[0135] Step S12: Based on the preset chromosome length information C = [C1, C2, ..., C...] N The X chromosome is divided into N chromosome sequences X1, X2, ..., X. N Each chromosome sequence is C in length. i After segmentation, the start and end positions of each chromosome are recorded for subsequent processing.

[0136] Step S13: For each chromosome sequence X i Non-overlapping k-mer segmentation is applied to divide the sequence into consecutive subsequences of length k. Each k-mer subsequence is mapped to an integer index, generating a k-mer index sequence I. i The shape is (B, L′) i ),in:

[0137]

[0138] Step S14: I through the embedding layer i Mapped to embedded representation E iThe shape is (B, L′) i ,D), where D is the embedding dimension.

[0139] E i =Embedding(I i )

[0140] Step S15: Output the embedding representations E1, E2, ..., E for each chromosome. N .

[0141] Chromosome-specific convolution module 802 is used to perform Figure 2 Step 203 in the middle, and Figure 3 Steps 301 and 302 involve receiving the embedding representation corresponding to each chromosome sequence, then using convolution to extract features from the embedding vector data corresponding to each chromosome sequence. The extracted features are called the first chromosome features. Further downsampling is then used to shorten the sequence length and increase the feature dimension, outputting chromosome window markers, called the second chromosome features. Multiple independent convolutions and downsampling operations can efficiently extract local features of each chromosome while further reducing the sequence length.

[0142] Specifically, the implementation process of the chromosome-specific convolution module 802 can be as described in steps S21 to S24.

[0143] Step S21: Receive the embedding representation E of each chromosome sequence from the gene sequence preprocessing module 801. i The shape is (B, L′) i (D).

[0144] Step S22: For each chromosome sequence i, apply a one-dimensional convolutional layer Conv1D(D,D) out ), where the output feature dimension is D out Here, the corresponding first chromosome characteristics are obtained.

[0145] Step S23: Apply the downsampling layer Downsample1D(D out ,s down The first chromosome feature obtained from convolution is further shortened in sequence length and its feature dimensions are adjusted to generate the second chromosome feature W. i The shape is (B, L″) i ,D′), where:

[0146]

[0147] D′ represents the feature dimension after increasing the number of output channels based on the convolutional layer, s down To reduce the sampling step size.

[0148] Step S24: Output the second chromosome feature (W1, W2, ..., W) of each chromosome. N ).

[0149] The adaptive chromosome vector marker generator 803 is used to perform Figure 4 Step 401 and Figure 5 Steps 501 to 502 involve receiving the window marker (second chromosome feature) for each chromosome sequence, adaptively calculating the number Z of global representations (vector markers) for each chromosome sequence based on the length of the second chromosome feature and a preset ratio parameter, and then generating Z global representations (G) for each window marker through a pooling operation. i Z is G i The sequence length is determined, and finally, the global representations of all chromosomes are concatenated to output the first global representation sequence. The global representation (vector marker) learns the representation summary of each chromosome, providing a foundation for subsequent promotion of global information exchange between chromosomes and enhancing the model's ability to model cross-chromosomal dependencies.

[0150] Specifically, the implementation process of the adaptive chromosome vector marker generator 803 can be as described in steps 31 to 34.

[0151] Step S31: Receive window marker W for each chromosome sequence from the chromosome-specific convolution module 802. i The shape is (B, L″) i ,D′).

[0152] Step S32: Based on the second chromosome characteristic length L″ of each chromosome sequence i Calculate the global representation size ct_size using the preset ratio parameter r. i The calculation formula is:

[0153]

[0154] in, This indicates rounding down, meaning that regardless of the result of the division, the final ct_size will be the same. i Both will be the smallest integer that is not greater than or equal to the result of the division.

[0155] Step S33: Apply average pooling operation based on the second chromosome characteristic W of the chromosome sequence. i Generate a global representation G i The calculation formula is:

[0156]

[0157] Here, `AvgPool1D(·)` represents a one-dimensional average pooling operation, a downsampling method used to reduce the dimensionality of data. `kernel_size` represents the size of the pooling window. `stride` represents the step size by which the pooling window slides across the sequence, ensuring no overlap between pooling windows. The carrier label G... i The shape is (B, ct_size) i D ′ ).

[0158] Step S34: Put all G i The sequences are concatenated to generate the first global representation sequence CT = [G1, G2, ..., G]. N ], with shape (B, ∑ct_size i D ′ ).

[0159] The hierarchical attention module 804 is used for execution Figure 6 Steps 601 to 605 and Figure 7 Steps 701 and 702 involve receiving the second chromosome feature and the first global representation sequence. The first global representation sequence is first processed by a vector-labeled self-attention module. Then, a global-labeled self-attention module is used to achieve local and global representation interaction between the second chromosome feature and the first global representation sequence. The interacted second chromosome feature and the first global representation sequence are then fused to obtain the fused associated feature (global label). Through a hierarchical attention mechanism and downsampling operations, comprehensive modeling of local (intrachromosomal) and global (cross-chromosomal) features is achieved.

[0160] The specific implementation process of the hierarchical attention module 804 can be as described in steps 41 to 45.

[0161] Step S41: Receive the second chromosome features W = [W1, ..., W2] of all chromosomes from the chromosome-specific convolution module 802. N ], with a shape of (B,L ″ D ′ ), and the first global representation sequence CT from the adaptive chromosome vector marker generator 803, with shape (B,M,D) ′ ), where L ″ =∑L ″ i M = ∑ct_size i .

[0162] Step S42: Perform carrier labeling self-attention, first add location code P to CT. ct Subsequently, a second global representation sequence CT was obtained by applying the multi-head self-attention mechanism.′ The calculation formula is:

[0163] CT ′ =MultiHeadSelfAttention(CT+P ct )

[0164] Step S43: Perform global labeled self-attention. First, add positional encoding P to W. w Then the CT scan ′ and W+P w spliced ​​into [CT] ′ W+P w ], for [CT] ′ W+P w The third global representation sequence CT is obtained by applying the multi-head self-attention mechanism. ″ and third chromosome characteristic W ′ The calculation formula is:

[0165] [CT ″ W ′ ]=MultiHeadSelfAttention([CT ′ W+P w ])

[0166] Step S44: If the current hierarchical attention module 804 is a processing module that provides data to the phenotypic prediction head module 805, then the CT is upsampled. ″ The sequence length is adjusted to W ′ For the same length, the calculation process is as follows:

[0167] CT upsampled =Upsample(CT) ″ )

[0168] Then CT upsampled The fused correlation feature W″ is obtained by element-wise summation of W′ and W″. The calculation process is as follows:

[0169] W″=W′+CT upsampled

[0170] Step S45: If the current hierarchical attention module 804 is not a processing module that provides data to the phenotypic prediction head module 805, please refer to... Figure 9 The diagram shown illustrates the phenotypic prediction process. Figure 8 Based on this, the third global representation sequence CT, processed by the hierarchical attention module 804, is output. ″ and third chromosome characteristic W ′ Then, after downsampling layer 806 processing, the third global representation sequence CT is further processed.″ and third chromosome characteristic W ′ The sequence length is shortened, the feature dimension is doubled again, and then it passes through another adaptive chromosome vector marker generator 803 and hierarchical attention module 804 in sequence, that is, based on the output of the downsampling layer, steps S31 to S44 are executed again.

[0171] Phenotypic prediction head module 805 is used to perform Figure 1 Step 104 involves receiving the fused and correlated features from the hierarchical attention module 804, flattening the received features into vectors, and then performing a nonlinear transformation through a fully connected layer to output the phenotypic prediction value. By flattening the features and using a fully connected layer, the feature information extracted by the model is preserved to the greatest extent, thereby achieving accurate prediction from complex gene features to phenotypic prediction.

[0172] The specific execution process of the phenotypic prediction head module 805 can be described as in steps 51 to 54.

[0173] Step S51: Receive the fused correlation features W from the hierarchical attention module 804 ″ The shape is (B, L) ″ D ′ ).

[0174] Step S52: Place W ″ Flattened into a vector F, with shape (B, L) ″ ×D ′ ).

[0175] Step S53: Apply a fully connected Linear layer to map F to phenotypic predicted values. The shape is (B, 1). The calculation formula is:

[0176]

[0177] Step S54: Generate phenotypic predicted values This is used for subsequent evaluation of the Pearson correlation coefficient or root mean square error. The training process of the phenotypic prediction head module 805 can use the mean square error (RMSE) as the loss function to calculate the predicted values. The difference between the model parameter and the true value y is used to update the model parameters using the Adam optimizer, and the MSE loss is minimized through the backpropagation algorithm to iteratively optimize the model.

[0178] In one example, another flowchart for phenotypic prediction can be found here. Figure 10 As shown:

[0179] First, the gene sequence preprocessing module acquires the gene sequence to be tested. This sequence contains multiple chromosomes (chromosome 1, chromosome 2, chromosome 3, ..., chromosome n), typically represented in text form and containing the four bases A, T, C, and G. The SNP information on each chromosome is extracted from the gene sequence to obtain the single nucleotide polymorphism (SNP) sequence (chromosome sequence). Next, the converted SNPs undergo K-mer segmentation, which divides the sequence into fixed-length segments and embeds each segment into a numerical vector, obtaining the embedding representation of each segment.

[0180] Then, the chromosome-specific convolution module performs one-dimensional convolution (1D Conv 1) on the embedded representation of SNPs segments of each chromosome to extract chromosome-specific features, which can be called the first chromosome feature. Then, through downsampling, the second chromosome feature (or window label) is obtained.

[0181] Subsequently, the adaptive chromosome vector marker generator performs average pooling on each segment of the second chromosome feature to generate a global representation G corresponding to each segment. i The first global representation sequence is obtained (the first global representation sequence can be a set of global representations or the result of splicing together the global representations (carrier tag sequence). Here, the set of global representations is taken as an example).

[0182] Next, the hierarchical attention module concatenates the first global representation sequence encoded in 1D position, and processes the concatenation result using a multi-head self-attention mechanism through its internal carrier label self-attention module to obtain the second global representation sequence. The second global representation sequence is then concatenated with the second chromosome feature encoded in 1D position to obtain the global label. Entering the global label self-attention module, the model applies a multi-head self-attention mechanism to process the global label, resulting in a new round of global labels. This new round of global labels includes the third global representation sequence corresponding to the second global representation sequence and the third chromosome feature corresponding to the second chromosome feature. Finally, the third global representation sequence and the third chromosome feature are separated in the new global label.

[0183] Finally, the phenotypic prediction head upsamples the separated third global representation sequence to obtain a fourth global representation sequence, ensuring that the fourth global representation sequence has the same length as the total sequence of the separated third chromosome features. The fourth global representation sequence is then element-wise added to the separated third chromosome features to obtain the final fused association feature. This fused association feature is flattened and then linearly processed to obtain the phenotypic prediction result.

[0184] In one example, to verify the advantages of the chromosome-specific whole-genome prediction method of this application, two important crop datasets were used for experimental verification. The first dataset, called Rice3k, contains genomic data and corresponding phenotypic data of 3024 samples collected from 89 countries. Regarding the genomic data, the SNP sequence length of each sample is 404388. The Rice3k dataset covers 12 chromosomes, with chromosome lengths ranging from 23548 to 47497. This dataset includes various traits, such as ligule length (LL), spike length (PL), and seedling height (SH). Several phenotypes of significant value were selected for study. These phenotypes were provided by authoritative institutions such as the International Rice Research Institute (IRRI) and the Chinese Academy of Agricultural Sciences (CAAS). Specific information on the phenotypes can be found in Table 1. The second dataset, called Soybean, covers 2898 soybean germplasm resources and their corresponding phenotypic data. Regarding genomic data, the SNP sequence length of each sample was 390,215. The Soybean dataset contains 20 chromosomes, with chromosome lengths ranging from 12,971 to 38,06. This dataset also contains a variety of traits, and several phenotypes of research significance were selected for analysis. Detailed information is shown in Table 2.

[0185] Table 1. Rice3k dataset phenotypic details

[0186]

[0187]

[0188] Table 2. Detailed information on the Soybean dataset.

[0189] Trait Abbreviation Trait Full Name Unit Trial Location Year SL Seed Length mm Region 4 2015 ST Seed Thickness mm Region 4 2015 VBN Number of Effective Branches - Region 4 2015 LL Leaf Length cm Region 4 2015 LW Leaf Width cm Region 4 2015 SDI Seed Damage Index - Region 4 2015

[0190] The methods in this application can also be comprehensively compared with other current algorithms. These algorithms are divided into three categories to cover different modeling methods and theoretical foundations. The first category is the GBLUP (Genomic Best Linear Unbiased Prediction) algorithm, which is based on a mixed linear model. As a representative method based on the theory of best linear unbiased prediction, GBLUP has shown stable performance in various whole-genome prediction tasks. The second category is methods based on Bayesian theory, including BayesA, BayesB, BayesCpi, and BayesR algorithms. These methods adopt different prior assumptions for the variance and effect of each molecular marker. The third category is deep learning algorithms. Among them, the DNNGP algorithm, based on convolutional neural networks, combines a convolutional neural network architecture and introduces batch normalization and dropout mechanisms to enhance the stability and generalization ability of the model; while the Transformer_IJCAI algorithm consists of the original Transformer Encoder and a simple prediction head, which effectively captures long-range dependencies in gene sequences using a self-attention mechanism. To objectively evaluate the predictive performance of the embodiments of this application, the Pearson correlation coefficient was used as the main indicator, which can effectively measure the linear correlation between the predicted value and the actual phenotypic value.

[0191] To comprehensively evaluate the performance of the proposed whole-genome prediction model, the Rice3k dataset was used for 10-fold cross-validation. The Pearson correlation coefficient was used as the evaluation metric to compare the performance of the method described in this application (Ours) with benchmark algorithms such as GBLUP, BayesA, BayesB, BayesCpi, BayesR, DNNGP, and Transformer_IJCAI. Specific results are shown in Table 3. Overall, the method described in this application achieved excellent predictive performance on most phenotypes. For example, in traits such as stem length (CL), thousand-grain weight (TGW), grain length (GL), grain width (GW), leaf curl (LRI), ligule length (LL), spike length (PL), and seedling height (SH), the Pearson correlation coefficients of the method in this application are 0.7850, 0.7416, 0.6761, 0.7545, 0.6667, 0.6568, 0.6791, and 0.4871, respectively, all higher than or close to the highest values ​​of other methods, indicating its effectiveness in capturing complex gene-phenotype relationships. In traits that are difficult to predict, such as leaf curl (LRI) and ligule length (LL), the method in this application significantly outperforms other algorithms (such as BayesR's 0.3641 and Transformer_IJCAI's 0.6026) with correlation coefficients of 0.6667 and 0.6568, respectively.

[0192] Table 3 shows the Pearson correlation coefficient results for the ten-fold cross-validation of the Rice3k dataset.

[0193] CL CN DTH FLL TGW GL GBLUP 0.7662 0.4808 0.7553 0.5881 0.7301 0.6452 BayesA 0.7697 0.4831 0.7616 0.5860 0.7335 0.6655 BayesB 0.7707 0.4837 0.7642 0.5789 0.7242 0.6837 BayesCpi 0.7666 0.4822 0.7572 0.5884 0.7315 0.6589 BayesR 0.7680 0.4500 0.7568 0.5830 0.6785 0.4981 DNNGP 0.7677 0.4837 0.7497 0.5819 0.7276 0.5813 Transformer_IJCAI 0.7384 0.4491 0.7430 0.5750 0.7388 0.6414 Ours 0.7850 0.5017 0.7616 0.5981 0.7416 0.6761 GW LRI LL PL PN PH GBLUP 0.7307 0.6061 0.6296 0.6460 0.6444 0.7901 BayesA 0.7545 0.5717 0.6290 0.6417 0.6450 0.7960 BayesB 0.7618 0.5934 0.6264 0.6359 0.6408 0.8007 BayesCpi 0.7545 0.6046 0.6295 0.6457 0.6447 0.7913 BayesR 0.7296 0.3641 0.6164 0.6256 0.6420 0.7916 DNNGP 0.6811 0.6543 0.6253 0.6363 0.6349 0.7881 Transformer_IJCAI 0.7334 0.6610 0.6026 0.6227 0.6400 0.7851 Ours 0.7545 0.6667 0.6568 0.6791 0.6458 0.7966

[0194] Similarly, the performance of the method in this application embodiment and the aforementioned benchmark algorithm was evaluated using ten-fold cross-validation on the Soybean dataset, and the specific results are shown in Table 4. Using Pearson correlation coefficient as an indicator, the method in this application embodiment performed excellently on multiple phenotypes. For example, for traits such as stem length (SL_region 4_2015), thousand-grain weight (ST_region 4_2015), VBN predicted values ​​(VBN_BLUP and VBN_region 4_2015), leaf length (LL_BLUP and LL_region 4_2015), grain width (LW_BLUP), and stem diameter index (SDI_region 4_2015), the Pearson correlation coefficients of the method in this application embodiment were 0.5321, 0.6928, 0.7971, 0.3897, 0.5452, 0.6403, 0.7360, and 0.5069, respectively, all higher than or close to the highest values ​​of other methods, demonstrating its competitiveness in predicting complex phenotypes. In particular, the method of this application significantly outperforms other algorithms (such as BayesR's 0.2825 and 0.3726) in terms of VBN predicted value (VBN_region 4_2015) and stem diameter index (SDI_region 4_2015), with correlation coefficients of 0.3897 and 0.5069, respectively.

[0195] Through 10-fold cross-validation experiments on the Rice3k and Soybean datasets, the method of this application demonstrates excellent performance in predicting diverse phenotypes. These results fully demonstrate the good generalization ability of the method of this application on different crop datasets.

[0196] Table 4. Pearson correlation coefficient results for 10-fold cross-validation on the Soybean dataset.

[0197]

[0198]

[0199] Reference Figure 11 In some embodiments, this application also provides a chromosome-specific whole-genome prediction device 1100, which includes:

[0200] The acquisition unit 1101 is used to acquire the gene sequence to be tested and to segment the gene sequence to be tested according to the preset chromosome length information to obtain multiple chromosome sequences;

[0201] The extraction unit 1102 is used to extract local features from each chromosome sequence to obtain the first chromosome feature corresponding to each chromosome sequence.

[0202] The generation unit 1103 is used to generate a first global representation sequence corresponding to the gene sequence to be tested based on the first chromosome features corresponding to each chromosome sequence.

[0203] The prediction unit 1104 is used to perform phenotypic prediction based on multiple first chromosome features corresponding to the first global representation sequence and the gene sequence to be tested, and to obtain the phenotypic prediction result corresponding to the gene sequence to be tested.

[0204] Optionally, in some embodiments, the extraction unit 1102 includes:

[0205] The word segmentation subunit is used to segment each chromosome sequence based on multiple preset base sequence fragments, and encode the segmentation results to generate an index sequence corresponding to the chromosome sequence.

[0206] The embedding subunit is used to perform embedding processing on each index sequence to obtain the embedding representation corresponding to the chromosome sequence;

[0207] Extraction subunits are used to extract local features from the embedding representation corresponding to each chromosome sequence to obtain the first chromosome feature.

[0208] Optionally, in some embodiments, the prediction unit 1104 includes:

[0209] The downsampling subunit is used to downsample the first chromosome feature corresponding to each chromosome sequence to obtain the second chromosome feature corresponding to each chromosome sequence.

[0210] The first prediction subunit is used to perform phenotypic prediction based on the first global representation sequence and multiple second chromosome features corresponding to the gene sequence to be tested, and to obtain the phenotypic prediction result corresponding to the gene sequence to be tested.

[0211] Optionally, in some embodiments, the generation unit 1103 includes:

[0212] The generation subunit is used to generate a global representation of each chromosome sequence based on the second chromosome feature corresponding to each chromosome sequence;

[0213] The splicing subunit is used to splice together the global representations corresponding to each chromosome sequence to obtain the first global representation sequence corresponding to the gene sequence to be tested.

[0214] Optionally, in some embodiments, generating sub-units includes:

[0215] The determination module is used to determine the number of global representations corresponding to each chromosome sequence based on the length of the second chromosome feature corresponding to each chromosome sequence and a preset ratio parameter;

[0216] The generation module is used to generate a global representation for each chromosome sequence based on the second chromosome features and the number of global representations.

[0217] Optionally, in some embodiments, the prediction unit 1104 includes:

[0218] The first self-attention subunit is used to perform multi-head self-attention processing on the first global representation sequence to obtain the second global representation sequence;

[0219] The second self-attention subunit is used to splice the second global representation sequence with multiple second chromosome features corresponding to the gene sequence to be tested, and to perform multi-head self-attention processing on the splicing result to obtain the third global representation sequence corresponding to the second global representation sequence and the third chromosome feature corresponding to the second chromosome feature.

[0220] The upsampling subunit is used to upsample the third global representation sequence to obtain a fourth global representation sequence with the same total size as the multiple third chromosome features corresponding to the gene sequence to be tested;

[0221] The fusion subunit is used to add up multiple third chromosome features corresponding to the fourth global representation sequence and the gene sequence to be tested positionally to obtain the fusion association feature;

[0222] The second prediction subunit is used to predict the phenotype based on fusion association features and obtain the phenotype prediction results corresponding to the gene sequence to be tested.

[0223] Optionally, in some embodiments, the prediction unit 1104 further includes:

[0224] The first position encoding subunit is used to perform position encoding on the first global representation sequence;

[0225] The second positional encoding subunit is used to positionally encode the second chromosome feature corresponding to each chromosome sequence.

[0226] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described chromosome-specific whole-genome prediction method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0227] Please see Figure 12 , Figure 12 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0228] The processor 1201 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0229] The memory 1202 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1202 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1202 and called and executed by the processor 1201 to execute the chromosome-specific whole-genome prediction method of the embodiments of this application.

[0230] The input / output interface 1203 is used to implement information input and output;

[0231] The communication interface 1204 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0232] Bus 1205 transmits information between various components of the device (e.g., processor 1201, memory 1202, input / output interface 1203, and communication interface 1204);

[0233] The processor 1201, memory 1202, input / output interface 1203 and communication interface 1204 are connected to each other within the device via bus 1205.

[0234] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described chromosome-specific whole-genome prediction method.

[0235] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0236] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0237] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0238] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0239] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0240] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0241] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0242] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0243] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0244] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0245] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0246] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A chromosome-specific based whole genome prediction method, characterized by, The method comprises the following steps: obtaining a to-be-tested gene sequence, and segmenting the to-be-tested gene sequence according to preset chromosome length information to obtain a plurality of chromosome sequences; locally extracting features of each chromosome sequence to obtain first chromosome features corresponding to each chromosome sequence; generating a first global representation sequence corresponding to the to-be-tested gene sequence based on the first chromosome features corresponding to each chromosome sequence; performing phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

2. The method of claim 1, wherein, The locally extracting features of each chromosome sequence to obtain first chromosome features corresponding to each chromosome sequence comprises the following steps: performing word segmentation processing on each chromosome sequence based on a plurality of preset base sequence fragments, and encoding the obtained word segmentation result to generate an index sequence corresponding to the chromosome sequence; performing embedding processing on each index sequence to obtain an embedding representation corresponding to the chromosome sequence; locally extracting features of the embedding representation corresponding to each chromosome sequence to obtain the first chromosome features.

3. The method of claim 1, wherein, The performing phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence comprises the following steps: performing down-sampling processing on the first chromosome features corresponding to each chromosome sequence to obtain second chromosome features corresponding to each chromosome sequence; performing phenotype prediction based on the first global representation sequence and a plurality of second chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

4. The method of claim 3, wherein, The generating a first global representation sequence corresponding to the to-be-tested gene sequence based on the first chromosome features corresponding to each chromosome sequence comprises the following steps: generating a global representation of each chromosome sequence based on the second chromosome features corresponding to each chromosome sequence; splicing the global representations corresponding to each chromosome sequence to obtain the first global representation sequence corresponding to the to-be-tested gene sequence.

5. The method of claim 4, wherein, The generating a global representation of each chromosome sequence based on the second chromosome features corresponding to each chromosome sequence comprises the following steps: determining the number of global representations corresponding to each chromosome sequence based on the length of the second chromosome features corresponding to each chromosome sequence and a preset proportion parameter; generating a global representation corresponding to each chromosome sequence based on the second chromosome features and the number of global representations.

6. The method of claim 3, wherein, The performing phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence comprises the following steps: performing multi-head self-attention processing on the first global representation sequence to obtain a second global representation sequence; The second global representation sequence is spliced with a plurality of second chromosome features, and multi-head self-attention processing is performed on the spliced result to obtain a third global representation sequence corresponding to the second global representation sequence and a third chromosome feature corresponding to the second chromosome feature; The third global representation sequence is up-sampled to obtain a fourth global representation sequence with a total size same as that of the plurality of third chromosome features corresponding to the to-be-tested gene sequence; The fourth global representation sequence and the plurality of third chromosome features corresponding to the to-be-tested gene sequence are added bit by bit to obtain a fusion correlation feature; Phenotype prediction is performed based on the fusion correlation feature to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

7. The method of claim 6, wherein, Before the multi-head self-attention processing is performed on the first global representation sequence to obtain a second global representation sequence, the method further includes: Position encoding is performed on the first global representation sequence. Before the second global representation sequence is spliced with a plurality of second chromosome features and multi-head self-attention processing is performed on the spliced result, the method further includes: Position encoding is performed on the second chromosome feature corresponding to each chromosome sequence.

8. A chromosome-specific based whole genome prediction apparatus, comprising: The method includes: An acquisition unit is configured to acquire a to-be-tested gene sequence, and split the to-be-tested gene sequence according to preset chromosome length information to obtain a plurality of chromosome sequences; An extraction unit is configured to respectively perform local feature extraction on each chromosome sequence to obtain a first chromosome feature corresponding to each chromosome sequence; A generation unit is configured to generate a first global representation sequence corresponding to the to-be-tested gene sequence based on the first chromosome feature corresponding to each chromosome sequence; A prediction unit is configured to perform phenotype prediction based on the first global representation sequence and a plurality of first chromosome features corresponding to the to-be-tested gene sequence to obtain a phenotype prediction result corresponding to the to-be-tested gene sequence.

9. An electronic device, comprising: The method includes: A memory, a transceiver, a processor, and a bus system; The memory is configured to store a program; The processor is configured to execute the program in the memory, including executing the method in any one of claims 1 to 7; The bus system is configured to connect the memory and the processor to enable the memory and the processor to communicate.

10. A computer-readable storage medium, characterized in that, The instructions, when executed on a computer, cause the computer to perform the method in any one of claims 1 to 7.

Images