A large model-based soybean breeding gene optimization method and device

Abstract

The application discloses a soybean breeding gene optimization method and device based on a large model, which comprises the following steps: S1, obtaining a soybean germplasm resource dataset; S2, obtaining standardized breeding characteristics; S3, tokenizing and encoding genotype data in the standardized breeding characteristics; S4, inputting the obtained multi-modal input representation into a pre-trained gene large model; S5, performing multi-head self-attention feature extraction on the multi-modal input representation through the gene large model to generate an optimized candidate gene sequence; S6, inputting the candidate gene sequence into a trait prediction model for verification; and S7, performing legality verification on the optimized gene sequence that passes the verification, and outputting a gene optimization scheme containing variation site information, variation types and breeding suggestions. Compared with the prior art, the application has the advantages that a soybean breeding gene optimization method and device are provided, which realizes deep fusion of multi-modal breeding data, intelligent optimization of gene sequences and safety verification of optimization results.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method and apparatus for optimizing soybean breeding genes based on a large model. Background Technology

[0002] As an important grain and oil crop, improving the yield and quality traits of soybeans is of great significance to ensuring food security. Traditional soybean breeding mainly relies on field phenotypic selection and hybridization, which is time-consuming and inefficient.

[0003] With the development of high-throughput sequencing technology, molecular breeding methods based on genomic selection have been widely used. By integrating genotype and phenotypic data to build predictive models, the breeding cycle can be shortened to some extent.

[0004] However, existing technologies still have the following shortcomings: First, existing methods are mostly based on single-dimensional data (such as only genotype or only phenotype), which makes it difficult to effectively integrate multimodal data such as genome, transcriptome, metabolome and environment, resulting in limited prediction accuracy for complex traits (such as yield and stress resistance); Second, traditional machine learning models are not capable of extracting deep semantic features of gene sequences, making it difficult to capture the nonlinear mapping relationship between long-distance regulatory elements and target traits.

[0005] Therefore, there is an urgent need for a breeding gene optimization scheme that can integrate multimodal data, deeply understand gene sequences based on large models, and has a complete verification and validation mechanism. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the above-mentioned technical defects and provide a soybean breeding gene optimization method and device that realizes deep fusion of multimodal breeding data, intelligent optimization of gene sequences, and security verification of optimization results.

[0007] To solve the above-mentioned technical problems, the technical solution provided by this invention is: a soybean breeding gene optimization method based on a large model, comprising the following steps: S1: Obtain phenotypic, genotypic, and environmental data of soybean germplasm resources to construct a multimodal breeding dataset; S2: Perform quality control screening, genotype filling, and phenotypic standardization on the multi-dimensional breeding dataset to obtain standardized breeding characteristics; S3: Tokenize the genotype data in the standardized breeding characteristics, convert the base sequence into a gene token vector, and concatenate and fuse it with the phenotypic feature vector and environmental feature vector to construct a multimodal input representation; S4: Input the obtained multimodal input representation into the pre-trained large gene model to learn the mapping relationship between gene sequences and target traits; S5: Extract multi-head self-attention features from the multimodal input representation using a large gene model, and perform cross-attention fusion in conjunction with breeding requirement instructions to calculate the optimized probability distribution of each gene locus and generate optimized candidate gene sequences. S6: Input candidate gene sequences into the trait prediction model for validation, and calculate the degree of satisfaction between the predicted trait value and the breeding target value; S7: Verify the legitimacy of the validated optimized gene sequences, detect stop codon anomalies, frameshift mutations, and known harmful variants. After successful verification, output a gene optimization scheme containing variant site information, variant type, and breeding suggestions.

[0008] Preferably, the multimodal breeding dataset in S1 further includes transcriptome data, metabolome data, proteome data, and epigenetic modification data; The environmental data includes at least one of soil physicochemical parameters, climate temperature and humidity, sunshine duration, and precipitation.

[0009] Preferably, the quality control screening in S2 includes missing rate filtering, minimum allele frequency filtering, and removal of samples with abnormal heterozygosity. The genotype filling is based on a deep learning filling model; The phenotypic standardization is based on the best linear unbiased prediction method.

[0010] Preferably, the large gene model in S4 is based on the Transformer architecture and is obtained by self-supervised pre-training and fine-tuning with breeding instructions on the soybean genome database; The self-supervised pre-training includes: randomly masking part of the gene segments of the soybean genome sequence according to a preset masking ratio, having the model predict the masked segments, and using the difference between the prediction results and the actual segments as the loss function to iteratively train until convergence; The breeding instruction fine-tuning includes: taking the breeding requirement instruction text and the original gene sequence as input, and the known optimized gene sequence as the target output, and updating the model parameters by minimizing the cross-entropy loss.

[0011] Preferably, step S5 further includes a satisfaction threshold: If the satisfaction level does not reach the preset threshold, the candidate gene sequence is fed back to the large gene model for iterative optimization until the satisfaction level reaches the target or the maximum number of iterations is reached.

[0012] Preferably, the trait prediction model in S6 is any one of a graph neural network, a convolutional neural network, or a gradient boosting tree model; The satisfaction level is quantitatively assessed by calculating the root mean square error between the predicted trait value and the breeding target value.

[0013] Preferably, the legality verification in S7 includes sequence alignment with the reference genome, detection of the integrity of conserved structural domains in coding regions, comparison with a database of known harmful variants, and prediction of off-target effects; The gene optimization scheme is output in the form of a structured report, including chromosome coordinates of the variant site, variant type, functional annotation, and breeding priority ranking.

[0014] Another aspect of the present invention discloses a soybean breeding gene optimization device based on a large model, including a data acquisition module, a preprocessing module, a token encoding module, a gene large model module, a trait verification module, and a legality verification module; The data acquisition module includes a field phenotyping vehicle, a drone remote sensing platform, a multispectral camera array, a lidar scanner, a high-throughput gene sequencer, an SNP genotyping chip scanner, a PCR thermal cycler, an automatic nucleic acid extraction workstation, an automatic weather station, soil multi-parameter sensor nodes, and a breeding database server cluster. The preprocessing module includes an automated pipetting workstation, a nucleic acid purification system, a capillary electrophoresis apparatus, a bioanalyzer, a genotype filling computing node, and a phenotypic standardization processing server. The token encoding module includes a dedicated encoder for gene sequences, a vector splicing operation unit, an FPGA acceleration card, a multimodal feature fusion processor, and an encoding cache chip; The large gene model module includes a large model inference server array, GPU-accelerated computing nodes, a self-attention mechanism operation unit, a Transformer decoder chip, and a candidate sequence generation processor. The trait verification module includes an intelligent greenhouse, a light-controlled culture rack, a high-throughput phenotypic imaging chamber, an automated trait detection pipeline, and a feedback control and communication module. The legality verification module includes a gene editing verification platform, a regulatory compliance determination server, a compliance review terminal, a gene optimization scheme output device, and a data archiving and storage array.

[0015] Another aspect of the present invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the soybean breeding gene optimization method based on a large model as described above.

[0016] Another aspect of the present invention discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the soybean breeding gene optimization method based on a large model as described above.

[0017] The advantages of this invention compared with the prior art are as follows: by constructing a multimodal breeding dataset that includes genomic, transcriptomic, metabolomic, proteomic and environmental data, this invention achieves deep integration of multidimensional breeding information and overcomes the problem of limited prediction accuracy caused by the single data dimension in the prior art. The Transformer-based gene model, through self-supervised pre-training and breeding instruction fine-tuning, enables a deep understanding of gene sequence semantics and the learning of long-distance regulatory relationships, improving the accuracy and interpretability of gene optimization. By setting up a dual guarantee mechanism of trait prediction verification and legality check, it systematically screens for stop codon anomalies, frameshift mutations, and harmful variations, ensuring the biosafety and breeding feasibility of optimized gene sequences. This invention achieves dynamic tuning of gene optimization results and improves the achievement rate of breeding goals. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a soybean breeding gene optimization method based on a large model.

[0019] Figure 2 This is a schematic diagram of the frame structure of the device in this application.

[0020] Figure 3 This is a schematic diagram of the structure of the large gene model module. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings.

[0022] Combined with appendix Figure 1-3 As shown, this invention discloses a soybean breeding gene optimization method based on a large model, comprising the following steps: Phenotypic, genotypic, and environmental data of soybean germplasm resources were acquired to construct a multimodal breeding dataset. During data acquisition, transcriptomic, metabolomic, proteomic, and epigenetic modification data were also acquired simultaneously to enrich the data dimensions. Environmental data included at least one of soil physicochemical parameters, climate temperature and humidity, light duration, and precipitation, collected in real time using field IoT sensors and meteorological monitoring equipment. The multimodal breeding dataset underwent quality control screening, genotype imputation, and phenotypic standardization to obtain standardized breeding characteristics. Quality control screening included missing genotype filtering, minimum allele frequency filtering, and removal of heterozygous abnormal samples. Genotype imputation was based on a deep learning imputation model, utilizing convolutional neural networks to learn local haplotype structures and perform probabilistic inference and imputation of missing genotypes. Phenotypic standardization was based on the best linear unbiased prediction method, correcting for environmental random and fixed effects to obtain standardized phenotypic values ​​at the genetic level. Genotype data in standardized breeding characteristics are tokenized and encoded, the base sequence is converted into gene token vectors, and then concatenated and fused with phenotypic feature vectors and environmental feature vectors to construct a multimodal input representation; The obtained multimodal input representation is input into the pre-trained gene big model to learn the mapping relationship between gene sequences and target traits. The gene big model is based on the Transformer architecture and is obtained by self-supervised pre-training on the soybean genome database and fine-tuning with breeding instructions. During the self-supervised pre-training process, some gene segments of the soybean genome sequence are randomly masked according to a preset masking ratio. The model is then asked to predict the masked segments, and the difference between the prediction results and the actual segments is used as the loss function for iterative training until convergence. When fine-tuning breeding instructions, the model parameters are updated by minimizing the cross-entropy loss, using the breeding requirement instruction text and the original gene sequence as input and the known optimized gene sequence as the target output, so that the model has the ability to optimize the gene sequence according to the instructions. Multi-head self-attention feature extraction is performed on the multimodal input representation through a large gene model, and cross-attention fusion is performed in combination with breeding requirement instructions. The optimized probability distribution of each gene locus is calculated to generate optimized candidate gene sequences. Candidate gene sequences are input into the trait prediction model for validation. The satisfaction degree between the predicted trait value and the breeding target value is calculated. The trait prediction model can be any one of graph neural network, convolutional neural network or gradient boosting tree model. The satisfaction degree is quantitatively evaluated by calculating the root mean square error between the predicted trait value and the breeding target value. The validated optimized gene sequences are legally verified, and abnormal stop codons, frameshift mutations, and known harmful variants are detected. After the verification is passed, a gene optimization scheme containing variant site information, variant type, and breeding suggestions is output. The legitimacy verification includes sequence alignment with the reference genome, detection of the integrity of conserved domains in coding regions, comparison with known harmful variant databases, and prediction of off-target effects. The gene optimization plan is output in the form of a structured report, including chromosome coordinates of the variant site, variant type, functional annotation, and breeding priority ranking.

[0023] In one embodiment: A satisfaction threshold determination mechanism is also set up. When the satisfaction between the predicted trait value and the breeding target value does not reach the preset threshold, the candidate gene sequence is fed back to the gene big model for iterative optimization, the sampling strategy of the optimization probability distribution is adjusted, and the candidate sequence is regenerated until the satisfaction reaches the standard or the maximum number of iterations is reached, forming a closed-loop optimization. Another aspect of the present invention discloses a soybean breeding gene optimization device based on a large model, comprising a data acquisition module, a preprocessing module, a token encoding module, a gene large model module, a trait verification module, and a legality verification module, wherein each module is connected in sequence and works together to execute the above method steps; The data acquisition module acquires phenotypic, genotypic, and environmental data from soybean germplasm resources to construct a multimodal breeding dataset. The preprocessing module performs quality control screening, genotyping, and phenotypic standardization on the multimodal breeding dataset to obtain standardized breeding features. The token encoding module tokenizes the genotypic data in the standardized breeding features and concatenates and fuses it with phenotypic and environmental feature vectors to construct a multimodal input representation. The gene big model module performs multi-head self-attention feature extraction on the multimodal input representation and generates optimized candidate gene sequences based on breeding requirements. The trait verification module performs trait prediction verification on the candidate gene sequences. When the satisfaction level does not reach a preset threshold, the candidate sequence is fed back to the gene big model module for iterative optimization. The legality verification module verifies the legality of the verified optimized gene sequences and outputs the gene optimization scheme after successful verification.

[0024] In a specific implementation, the present invention can be implemented by an electronic device, which includes a processor and a memory. The memory stores a computer program, and when the processor executes the program, it implements the above-mentioned method steps. Alternatively, the computer program can be stored in a computer-readable storage medium, and when the processor executes it, it implements the above-mentioned method. By inputting breeding requirements, users can automatically complete the entire gene optimization process, from data acquisition, multimodal fusion, large model optimization, trait verification to safety verification, and output a structured gene optimization scheme that can guide actual breeding operations.

[0025] Working principle: This invention collects multimodal data on soybean germplasm resources, including phenotypic, genotypic, and environmental data. After quality control screening, standardization, and tokenization encoding, a multimodal input representation is constructed. The multimodal input is then fed into a pre-trained large-scale gene model. Deep features of gene sequences are extracted through multi-head self-attention, and cross-attention fusion is performed in conjunction with breeding requirement instructions to calculate the optimized probability distribution of each gene locus and generate candidate sequences. The candidate sequences are then input into the trait prediction model to verify their satisfaction. If the model does not meet the requirements, iterative optimization is performed. Finally, after verification of the legality of stop codons, frameshift mutations, and harmful variations, a structured gene optimization scheme is output.

[0026] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A soybean breeding gene optimization method based on a large model, characterized in that: Includes the following steps: S1: Obtain phenotypic, genotypic, and environmental data of soybean germplasm resources to construct a multimodal breeding dataset; S2: Perform quality control screening, genotype filling, and phenotypic standardization on the multi-dimensional breeding dataset to obtain standardized breeding characteristics; S3: Tokenize the genotype data in the standardized breeding characteristics, convert the base sequence into a gene token vector, and concatenate and fuse it with the phenotypic feature vector and environmental feature vector to construct a multimodal input representation; S4: Input the obtained multimodal input representation into the pre-trained large gene model to learn the mapping relationship between gene sequences and target traits; S5: Extract multi-head self-attention features from the multimodal input representation using a large gene model, and perform cross-attention fusion in conjunction with breeding requirement instructions to calculate the optimized probability distribution of each gene locus and generate optimized candidate gene sequences. S6: Input candidate gene sequences into the trait prediction model for validation, and calculate the degree of satisfaction between the predicted trait value and the breeding target value; S7: Verify the legitimacy of the validated optimized gene sequences, detect stop codon anomalies, frameshift mutations, and known harmful variants. After successful verification, output a gene optimization scheme containing variant site information, variant type, and breeding suggestions.

2. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: The multimodal breeding dataset in S1 also includes transcriptome data, metabolome data, proteome data, and epigenetic modification data; The environmental data includes at least one of soil physicochemical parameters, climate temperature and humidity, sunshine duration, and precipitation.

3. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: The quality control screening in S2 includes missing rate filtering, minimum allele frequency filtering, and removal of samples with abnormal heterozygosity. The genotype filling is based on a deep learning filling model; The phenotypic standardization is based on the best linear unbiased prediction method.

4. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: The large gene model in S4 is based on the Transformer architecture and is obtained by self-supervised pre-training and fine-tuning with breeding instructions on the soybean genome database. The self-supervised pre-training includes: randomly masking part of the gene segments of the soybean genome sequence according to a preset masking ratio, having the model predict the masked segments, and using the difference between the prediction results and the actual segments as the loss function to iteratively train until convergence; The breeding instruction fine-tuning includes: taking the breeding requirement instruction text and the original gene sequence as input, and the known optimized gene sequence as the target output, and updating the model parameters by minimizing the cross-entropy loss.

5. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: S5 also includes a satisfaction threshold: If the satisfaction level does not reach the preset threshold, the candidate gene sequence is fed back to the large gene model for iterative optimization until the satisfaction level reaches the target or the maximum number of iterations is reached.

6. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: The trait prediction model in S6 is any one of graph neural network, convolutional neural network or gradient boosting tree model; The satisfaction level is quantitatively assessed by calculating the root mean square error between the predicted trait value and the breeding target value.

7. The soybean breeding gene optimization method based on a large model according to claim 1, characterized in that: The legality verification in S7 includes sequence alignment with the reference genome, detection of the integrity of conserved structural domains in coding regions, comparison with a database of known harmful variants, and prediction of off-target effects. The gene optimization scheme is output in the form of a structured report, including chromosome coordinates of the variant site, variant type, functional annotation, and breeding priority ranking.

8. A soybean breeding gene optimization device based on a large model, using a soybean breeding gene optimization method based on a large model as described in any one of claims 1-7, characterized in that: It includes a data acquisition module, a preprocessing module, a token encoding module, a large gene model module, a trait verification module, and a legality verification module; The data acquisition module includes a field phenotyping vehicle, a drone remote sensing platform, a multispectral camera array, a lidar scanner, a high-throughput gene sequencer, an SNP genotyping chip scanner, a PCR thermal cycler, an automatic nucleic acid extraction workstation, an automatic weather station, soil multi-parameter sensor nodes, and a breeding database server cluster. The preprocessing module includes an automated pipetting workstation, a nucleic acid purification system, a capillary electrophoresis apparatus, a bioanalyzer, a genotype filling computing node, and a phenotypic standardization processing server. The token encoding module includes a dedicated encoder for gene sequences, a vector splicing operation unit, an FPGA acceleration card, a multimodal feature fusion processor, and an encoding cache chip; The large gene model module includes a large model inference server array, GPU-accelerated computing nodes, a self-attention mechanism operation unit, a Transformer decoder chip, and a candidate sequence generation processor. The trait verification module includes an intelligent greenhouse, a light-controlled culture rack, a high-throughput phenotypic imaging chamber, an automated trait detection pipeline, and a feedback control and communication module. The legality verification module includes a gene editing verification platform, a regulatory compliance determination server, a compliance review terminal, a gene optimization scheme output device, and a data archiving and storage array.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the soybean breeding gene optimization method based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the soybean breeding gene optimization method based on a large model as described in any one of claims 1 to 7.

Images