Method for identifying mulberry variety by combining AI-driven PCA with random forest

Abstract

The invention discloses a method for identifying mulberry varieties by using AI-driven PCA and random forest combination, which comprises the following steps: firstly, obtaining mulberry seeds to be identified to construct a DNA library, and obtaining annotated SNP data through the steps of double-end sequencing, data preprocessing, comparison and the like; encoding the data, extracting a main SNP site in combination with PCA and singular value decomposition, training and screening out 225 core SNP sites through a machine learning algorithm, and constructing a variety prediction model; and finally, building a library for sequencing a to-be-identified sample, converting 225 SNP genotypes into effective codes, inputting the effective codes into a random forest algorithm for multiple predictions, and determining the variety type of the sample according to the predictions. The method can accurately and efficiently identify the mulberry variety.

Description

Technical Field

[0001] This invention relates to the field of mulberry variety identification technology, specifically to a method for identifying mulberry varieties using AI-driven PCA combined with random forest. Background Technology

[0002] The yield and quality of mulberry leaves largely depend on the mulberry variety.

[0003] Mulberry germplasm resources are extremely rich, but identifying varieties at the seed level is quite challenging. The market is flooded with inferior mulberry seeds, directly leading to reduced mulberry leaf yields and severely impacting silkworm farmers' income. Against this backdrop, establishing a method to accurately distinguish the seeds of six superior mulberry varieties from those of other varieties is crucial. This will not only help increase silkworm farmers' income but also construct an efficient classification, protection, and utilization mechanism, providing valuable references for taxonomic research on the mulberry section of plants.

[0004] Early variety identification relied primarily on phenotypic identification. However, with the widespread adoption of hybridization breeding, the phenotypes of individuals in hybrid populations after several generations are extremely similar to their parents, rendering phenotypic identification insufficient for the current identification needs of mulberry hybrid varieties. In recent years, DNA molecular marker technology has made significant progress, with techniques such as random amplified polymorphic DNA (RAPD), amplified fragment length polymorphism (AFLP), simple repeat sequence markers (SSR), inter-simple repeat sequence markers (ISSR), and single nucleotide polymorphisms (SNPs) widely applied in variety identification (Jiao et al., 2001; Fu et al., 2004; Agarwal et al., 2008; Yao et al., 2015; Chen et al., 2015). Among these, SNP marker technology, with its high throughput, high integration, miniaturization, and automation, has been widely used in variety identification (Wang et al., 2020; Fan et al., 2020). However, given the large number of SNP sites, identifying the most representative genetic markers is crucial.

[0005] Principal component analysis (PCA) is a commonly used dimensionality reduction technique in statistics. While it effectively reduces sample dimensionality, the number of SNP loci remains excessive for variety identification. Furthermore, PCA performs poorly in classifying nonlinear data and cannot assess the importance of selected SNP loci. Random forest models, proposed by Breiman and Cutler in 2001, are supervised machine learning algorithms based on classification trees (Wang et al., 2010). Molecular systematics of mulberry has been a research hotspot, but due to limitations in existing methods and insufficient information loci in DNA fragments, previous studies have struggled to completely distinguish all mulberry varieties. Currently, methods combining PCA and random forest to extract highly informative SNPs for variety identification are rarely reported in mulberry research.

[0006] Therefore, there is an urgent need for a method that combines AI-driven PCA and random forest to identify mulberry varieties quickly and accurately. Summary of the Invention

[0007] One object of the present invention is to address at least the aforementioned deficiencies and to provide at least the advantages that will be described later.

[0008] To achieve these objectives and other advantages of the present invention, a method for identifying mulberry varieties using AI-driven PCA combined with random forest is provided, comprising:

[0009] Mulberry seeds to be identified were obtained, DNA was extracted and subjected to random fragmentation, end repair, tailing, adapter addition, fragment screening, PCR amplification and purification to construct a library, thus obtaining a DNA library.

[0010] The integrity, insert size, and effective concentration of the DNA fragments in the library were detected. Qualified DNA fragments were screened and merged according to the effective concentration and the required amount of data to be sequenced. Then, paired-end sequencing was performed to obtain the raw data.

[0011] Data preprocessing: The raw data was filtered using FASTP software to remove sequencing reads containing adapter sequences, N content exceeding 10%, and low-quality bases exceeding 50%, thus obtaining high-quality and effective data.

[0012] The high-quality valid data was compared with the mulberry genome using BWA-mem2 software, and repetitive sequences were removed using sambamba software to obtain the valid data after alignment.

[0013] GATK software was used to detect population SNPs in the valid data after alignment. Hard filtering and basic filtering were performed, and then ANNOVAR software was used to perform functional annotation on the SNPs to obtain annotated SNP data.

[0014] The annotated SNP data were encoded, and principal SNP loci were extracted using PCA and singular value decomposition. The optimal number of principal components was determined based on the variance contribution rate of the principal components, and some loci were selected. The selected loci were trained using machine learning algorithms to evaluate their importance. After two training sessions using random forest algorithms, 225 core SNP loci were obtained, and a variety prediction model was constructed.

[0015] Sample identification: The samples to be identified are constructed, sequenced, and SNPs are extracted. After quality control, they are compared with the reference genome to ensure that all indicators meet the requirements for annotation analysis. The genotypes of 225 SNPs are converted into effective codes and input into the random forest algorithm. The prediction is repeated multiple times, and the sample variety type is determined by the prediction value that appears most frequently.

[0016] Specifically, the above scheme uses an instrument platform developed by Advanced Analytical Technologies for nucleic acid analysis to detect the integrity of the DNA fragments in the library and the size of the inserted fragments, and uses real-time quantitative PCR to detect the effective concentration of the library. The BWA-mem2 software version is v2.2.1, with the parameter set to bwa-mem2 mem. The sambamba software version is v0.6.6, with the parameter set to markdup. The GATK software version is v4.5.0.0. The ANNOVAR software version is 2013-05-20.

[0017] By integrating standardized operations across the entire process of library construction, sequencing, SNP screening, and AI modeling, the subjective errors of manual SNP screening are avoided, and the identification accuracy is improved to over 99%.

[0018] PCA is used to extract main SNP sites, which solves the problem of model overfitting caused by tens of thousands of SNPs in traditional methods, reduces computational complexity, and reduces memory usage by 70%.

[0019] Two random forest training sessions combined with site importance assessments were conducted to select 225 core SNPs, ensuring that the model's generalization ability AUC for unknown species reached 0.998 on the test set.

[0020] By using multiple independent predictions and probability statistics to eliminate random errors in single predictions, the repeatability of the results reaches 98.5%.

[0021] Preferably, the mulberry seeds to be identified include the seeds of Sangteyou No. 1, Sangteyou No. 2, Guisang No. 5, Guisang No. 6, Guisangyou No. 12, and Guisangyou No. 62.

[0022] In the above scheme, the screening of core SNP sites was optimized based on the genetic characteristics of six commercially available mulberry varieties in Guangxi, which improved the sensitivity of the model in specific variety groups by 5%-8%; interference sites of non-target varieties were excluded to reduce SNP misjudgment caused by distant hybridization, with a false positive rate of <0.5%.

[0023] Preferably, the parameters for hard filtering are QD<2.0, FS>60.0, SOR>3.0, MQ<40.0, QUAL<30.0, MQRankSum<-12.5, and ReadPosRankSum<-8.0.

[0024] In the above scheme, by combining parameters such as quality-depth ratio and strand bias, sequencing errors such as systematic errors in Illumina sequencing are effectively filtered, increasing the median QUAL value of the SNP dataset from 45 to 62. By setting thresholds for symmetry ratio (SOR) and mapping quality (MQ), false positives are reduced while rare variants are preserved, resulting in a 12% improvement in recall.

[0025] Preferably, the basic filtering specifically includes: filtering out SNPs with a sequencing error rate greater than 1%; discarding multiple alleles; SNPs with a support number ≥ 3; and secondary alleles with a maf < 0.05.

[0026] In the above scheme, by further limiting the SNP validation rate and reducing the noise of the population structure, the clarity of principal component analysis clustering is improved.

[0027] Preferably, the formula for calculating the variance contribution rate is: The formula for calculating the number of optimal principal components, i.e., k values, is: Where, δ i p represents the eigenvalues ​​corresponding to the original index correlation matrix. j This represents the cumulative metric value of the first j principal components. The eigenvectors are the principal components.

[0028] In the above scheme, the number of principal components k is dynamically determined based on the cumulative variance contribution rate formula, which avoids information loss caused by fixed dimensionality reduction and improves information retention rate; principal components are screened by sorting by eigenvalues, which shortens the PCA calculation time.

[0029] Preferably, the two random forest algorithm training steps include: first training: building an initial model based on the main SNP sites selected by PCA and calculating the site importance score; second training: further selecting 225 core SNP sites based on the importance score and optimizing the model parameters.

[0030] In the above scheme, the model accuracy is improved by using two random forest algorithms, the unchecked collinearity is reduced, and the interpretability of the model is enhanced.

[0031] Preferably, the prediction is repeated 100 times, and the prediction result must have a confidence level of ≥95%.

[0032] Preferably, sample identification specifically includes: constructing a library and sequencing the sample to be identified, extracting SNPs, and comparing it with a reference genome after quality control to ensure that all indicators meet the requirements for annotation analysis; inputting the genotype codes of 225 core SNP loci into a trained random forest model, performing N independent and repeated predictions, and recording the variety type and corresponding probability distribution of each prediction, where N≥100; calculating the entropy value of the prediction results to quantify uncertainty, wherein the entropy value formula is: Where C represents the total number of candidate varieties, P iLet be the predicted probability of the i-th variety; based on the Bootstrap resampling method, 1000 samplings with replacement are performed on 225 core SNP loci to generate 95% confidence intervals for the probabilities of each variety; the final variety must meet the following conditions simultaneously: the predicted probability of a certain variety appears ≥80% in repeated predictions, the 95% confidence interval of the variety does not overlap with the confidence intervals of other varieties, and the entropy value H≤0.5; if the above conditions are not met, it is marked as a low confidence result and the manual review process is triggered.

[0033] In the above scheme, the proportion of fuzzy judgments is reduced by combining entropy value with the principle of non-overlapping confidence intervals; then, 95% CI is generated by Bootstrap resampling to ensure the statistical power of variety differences (p<0.001); finally, manual review triggering conditions, such as frequency <80%, are used to avoid error propagation and achieve a 100% misjudgment correction rate.

[0034] The present invention has at least the following beneficial effects:

[0035] First, this invention utilizes an AI-driven method combining PCA and random forest to identify mulberry varieties. By combining AI technology, core SNP loci are screened using PCA and random forest algorithms to construct an accurate variety prediction model, which greatly improves the accuracy of mulberry variety identification.

[0036] Secondly, this invention utilizes an AI-driven method combining PCA and random forest to identify mulberry varieties. The entire identification process is highly automated, reducing the complexity and errors of manual operation.

[0037] Furthermore, the present invention utilizes an AI-driven PCA combined with random forest method to identify mulberry varieties, which reduces unnecessary experimental steps and reagent consumption compared to traditional identification methods, thereby lowering costs.

[0038] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0039] Figure 1 This is a density map of the original SNP sites in this invention;

[0040] Figure 2 The principal component contribution value of the PCA in this invention;

[0041] Figure 3 The results of this invention are the performance metrics for evaluating the model using four key indicators: F1-score, Precision, Recall, and Accuracy.

[0042] in, Figure 1This is a density map of the original SNP sites, i.e., the distribution of each SNP on the chromosome. The horizontal axis represents the base position on each chromosome, and the vertical axis represents the chromosome number. The color change represents the number of SNPs within 20k. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0044] Example

[0045] A method for identifying mulberry varieties using AI-driven PCA combined with random forest includes the following steps:

[0046] The first step is to obtain mulberry seeds to be identified, extract DNA, and construct a library by randomly fragmenting, end repair, tailing, adapter addition, fragment screening, PCR amplification, and purification to obtain a DNA library.

[0047] The second step involves detecting the integrity, insert size, and effective concentration of the DNA fragments in the library. Qualified DNA fragments are selected, merged according to the effective concentration and the required amount of data for sequencing, and then subjected to paired-end sequencing to obtain the raw data.

[0048] Step 3: Data preprocessing: Use FASTP software to filter the raw data, removing sequencing reads containing adapter sequences, N content exceeding 10%, and low-quality bases exceeding 50%, to obtain high-quality and effective data.

[0049] The fourth step involves using BWA-mem2 software to align high-quality, valid data with the mulberry genome, and then using sambamba software to remove repetitive sequences to obtain the aligned valid data.

[0050] Step 5: Use GATK software to detect population SNPs in the valid data after alignment, perform hard filtering and basic filtering, and then use ANNOVAR software to perform functional annotation on the SNPs to obtain annotated SNP data.

[0051] Step 6: Encode the annotated SNP data, extract principal SNP loci using PCA and singular value decomposition; determine the optimal number of principal components based on the principal component variance contribution rate, and screen some loci; train the screened loci using machine learning algorithms, evaluate the importance of loci, and obtain 225 core SNP loci through two training sessions using random forest algorithms, and construct a variety prediction model.

[0052] Step 7, Sample Identification: The sample to be identified is constructed, sequenced, and SNPs are extracted. After quality control, it is compared with the reference genome to ensure that all indicators meet the requirements for annotation analysis. The genotypes of 225 SNPs are converted into effective codes and input into the random forest algorithm. The prediction is repeated multiple times, and the sample variety type is determined by the prediction value that appears most frequently.

[0053] Among them, the mulberry seeds to be identified include seeds of Sangteyou No. 1, Sangteyou No. 2, Guisang No. 5, Guisang No. 6, Guisangyou No. 12, and Guisangyou No. 62.

[0054] Furthermore, the parameters for hard filtering are QD<2.0, FS>60.0, SOR>3.0, MQ<40.0, QUAL<30.0, MQRankSum<-12.5, and ReadPosRankSum<-8.0.

[0055] Furthermore, the basic filtering specifically includes: filtering out SNPs with a sequencing error rate greater than 1%; discarding multiple alleles; SNPs with a support number ≥ 3; and secondary alleles with a maf < 0.05.

[0056] Furthermore, the formula for calculating the variance contribution rate is as follows: The formula for calculating the number of optimal principal components, i.e., k values, is: Where, δ i p represents the eigenvalues ​​corresponding to the original index correlation matrix. j This represents the cumulative metric value of the first j principal components. The eigenvectors are the principal components.

[0057] Furthermore, the two training iterations of the random forest algorithm include: the first training: building an initial model based on the main SNP sites selected by PCA and calculating the importance score of the sites; the second training: further selecting 225 core SNP sites based on the importance score and optimizing the model parameters.

[0058] Furthermore, the prediction is repeated 100 times, and the prediction result must have a confidence level of ≥95%.

[0059] Furthermore, sample identification specifically includes: constructing a library and sequencing the samples to be identified, extracting SNPs, and comparing them with a reference genome after quality control to ensure that all indicators meet the requirements for annotation analysis; inputting the genotype codes of 225 core SNP loci into a trained random forest model, performing N independent and repeated predictions, and recording the variety type and corresponding probability distribution of each prediction, where N≥100; calculating the entropy value of the prediction results to quantify uncertainty, wherein the entropy value formula is: Where C represents the total number of candidate varieties, P iLet be the predicted probability of the i-th variety; based on the Bootstrap resampling method, 1000 samplings with replacement are performed on 225 core SNP loci to generate 95% confidence intervals for the probabilities of each variety; the final variety must meet the following conditions simultaneously: the predicted probability of a certain variety appears ≥80% in repeated predictions, the 95% confidence interval of the variety does not overlap with the confidence intervals of other varieties, and the entropy value H≤0.5; if the above conditions are not met, it is marked as a low confidence result and the manual review process is triggered.

[0060] Experiments and Analysis

[0061] Building a prediction model

[0062] Six mulberry varieties—Sangteiyou 1, Sangteiyou 2, Guisang 5, Guisang 6, Guisangyou 12, and Guisangyou 62—were compared with a reference genome. The alignment rate and genome coverage were statistically analyzed. The alignment rate reflects the similarity between the sample sequencing data and the reference genome, while the coverage depth and coverage directly reflect the uniformity of the sequencing data and its homology with the reference sequence. The alignment rate range is shown in Table 1.

[0063] Table 1 Statistics of Sequencing Depth and Coverage

[0064]

[0065] Coverage refers to the amount of genome data covered by the measured data. 1× means 1x coverage, 5× means 5x coverage, and 10× means 10x coverage.

[0066] According to the data in Table 1, the comparison rate of the six mulberry varieties ranged from 92.87% to 97.34%, all of which met the requirements for subsequent annotation analysis.

[0067] SNP Detection Statistics and Annotations

[0068] After alignment with the reference genome, population SNPs were detected, and the obtained SNPs were subjected to GATK hard filtering, yielding 5,985,961 SNPs. Further basic filtering was performed by filtering out SNPs with sequencing error rates greater than 1% and discarding multiple alleles. Finally, the detected SNPs were annotated, and the detection and annotation statistics are shown in Table 2.

[0069] Table 2 SNPs detection statistics

[0070]

[0071]

[0072] According to the data in Table 2, a total of 1,163,291 high-quality SNPs were obtained. These SNPs are located in the Upstream, Exonic, Intronic, Splicing, Downstream and Intergenic regions, with the most SNPs in the ts region (752,132) and the fewest in the Stop loss region (141).

[0073] SNP marker screening and model building

[0074] The original variant sites underwent further rigorous filtering based on the following criteria: 1. Markers were evenly distributed across the genome; 2. Marker read coverage depth was greater than or equal to 3; 3. Markers contained no missing sites, i.e., 100% site integrity; 4. Sites with an LD value greater than 0.2 were discarded. Ultimately, 97,750 variant sites were identified as candidate markers, and SNP sites were further extracted using PCA. The principal component contribution values ​​of the PCA were obtained as follows: Figure 2 When PC k It is a PC k+1 If the value is more than twice the value of k, then that value is the optimal k value; otherwise, the default value is k=3.

[0075] according to Figure 2 The data shows that no PCk is more than twice PCk+1, so the default value is 3. Then, the score for each SNP marker under the first three principal components is calculated. All SNP sites are ranked based on their scores, and the top 10,000 sites are selected for machine learning training and site selection.

[0076] The number of SNP loci after PCA pre-selection is still large, making it impossible to assess the importance of the selected SNP loci. A machine learning algorithm was used to train the PCA-selected SNPs to obtain the importance value of each SNP. After the initial selection of important SNPs, a random forest algorithm was used for training, ultimately resulting in a table of importance values ​​for 225 core SNP loci, as shown in Table 3. Four metrics (F1-score, Precision, Recall, and Accuracy) were used to evaluate the model's performance, and the results are as follows: Figure 3 .

[0077] Table 3: Importance Scale of 225 Core SNPs (Continued Table)

[0078]

[0079]

[0080]

[0081]

[0082] Quality control of samples to be identified

[0083] Six mulberry cultivar samples—Sangteiyou 1, Sangteiyou 2, Guisang 5, Guisang 6, Guisangyou 12, and Guisangyou 62—were numbered 1 to 6. The cultivar information for each sample was unknown to the predictor, but the cultivar information was known. Library construction, sequencing, and SNP extraction were performed on all six samples to evaluate the accuracy of the aforementioned model identification. After quality control, a clean base of 23.53 G was obtained, with each sample exceeding 3.13 G and the Q30 base percentage exceeding 90%. The clean bases of each sample were compared with the reference genome. The quality control statistics of the samples to be identified are shown in Table 4.

[0084] Table 4 Quality Control Results of Test Samples

[0085]

[0086] Wherein, Sample is the sample name; Raw reads is the number of sequencing Raw reads; Clean reads is the number of Clean reads obtained after filtering; Raw Base is the total number of bases in the raw sequencing data, which is calculated by multiplying the number of Raw reads by the sequencing length, in G units; Clean Base is the total number of bases in the filtered data, in G units; Errorrate is the average base sequencing error rate; Q20 and Q30 are the percentages of bases with Phred values ​​greater than 20 and 30, respectively, out of the total number of bases; GC is the percentage of G / C bases out of the total number of bases.

[0087] According to the data in Table 4, the alignment rate ranges from 95.67% to 97.3%, which meets the requirements for subsequent annotation analysis.

[0088] Variety identification of samples to be identified

[0089] For six submitted samples, variety prediction was performed based on 225 selected SNPs. During the prediction process, the genotypes of the 225 SNPs were first converted into one-hot values, i.e., 0-1-2 (Ref-Het-Alt) form. Then, the one-hot values ​​of the corresponding locus for each sample were input into a random forest algorithm in matrix form. During algorithm execution, the matrix of each input sample was transformed into a decision tree, thus obtaining the predicted variety value for each decision tree; this predicted value is the variety type value. To ensure the reliability of the results, this prediction process was repeated 100 times. After all predictions were completed, all variety type values ​​were integrated. By comparing the predicted values ​​with preset variety type values ​​(1-2-3-4-5-6) one by one, the variety type prediction value that appeared most frequently was determined as the final variety type. The expected variety type results for each sample are shown in Table 5.

[0090] Table 5: Identification Results of Six Sample Varieties

[0091]

[0092] Based on the data in Table 5, the predicted variety type of each sample was completely consistent with the actual result, which fully verified the accuracy and reliability of the model in mulberry variety identification.

[0093] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. A method for identifying mulberry varieties using AI-driven PCA combined with random forest, characterized in that, include: Mulberry seeds to be identified were obtained, DNA was extracted and subjected to random fragmentation, end repair, tailing, adapter addition, fragment screening, PCR amplification and purification to construct a library, thus obtaining a DNA library; The integrity, insert size, and effective concentration of the DNA fragments in the library were detected. Qualified DNA fragments were screened and merged according to the effective concentration and the required amount of data to be sequenced. Then, paired-end sequencing was performed to obtain the raw data. Data preprocessing: The raw data was filtered using FASTP software to remove sequencing reads containing adapter sequences, N content exceeding 10%, and low-quality bases exceeding 50%, in order to obtain high-quality and effective data. High-quality valid data were compared with the mulberry genome using BWA-mem2 software, and repetitive sequences were removed using sambamba software to obtain the valid data after comparison. GATK software was used to detect population SNPs in the aligned valid data, and hard filtering and basic filtering were performed. Then, ANNOVAR software was used to perform functional annotation on the SNPs to obtain annotated SNP data. The annotated SNP data is encoded, and principal SNP sites are extracted by combining PCA and singular value decomposition. The optimal number of principal components is determined based on the variance contribution rate of the principal components, and some sites are selected. Machine learning algorithms were used to train the selected loci and evaluate their importance. After two training sessions using random forest algorithms, 225 core SNP loci were obtained, and a variety prediction model was constructed. Sample identification: The samples to be identified are constructed, sequenced, and SNPs are extracted. After quality control, they are compared with the reference genome to ensure that all indicators meet the requirements for annotation analysis. The genotypes of 225 SNPs are converted into effective codes and input into the random forest algorithm. The prediction is repeated multiple times, and the sample variety type is determined by the prediction value that appears most frequently.

2. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, The mulberry seeds to be identified include those of Sangteyou No. 1, Sangteyou No. 2, Guisang No. 5, Guisang No. 6, Guisangyou No. 12, and Guisangyou No.

62.

3. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, The parameters for hard filtering are QD<2.0, FS>60.0, SOR>3.0, MQ<40.0, QUAL<30.0, MQRankSum<-12.5, and ReadPosRankSum<-8.

0.

4. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, Basic filtering specifically includes: SNPs with a sequencing error rate greater than 1% were filtered out; Discard multiple alleles; The number of SNPs is ≥3; Secondary allele maf < 0.

05.

5. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, The formula for calculating the variance contribution rate is: The formula for calculating the number of optimal principal components, i.e., k values, is: Where, δ i p represents the eigenvalues ​​corresponding to the original index correlation matrix. j This represents the cumulative metric value of the first j principal components. The eigenvectors are the principal components.

6. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, The two random forest algorithm training sessions included: First training: Build an initial model based on the main SNP sites selected by PCA and calculate the site importance score; Second training: Based on importance scores, 225 core SNP sites were further selected to optimize model parameters.

7. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, The predictions are repeated 100 times, and the prediction results must have a confidence level of ≥95%.

8. The method for identifying mulberry varieties using AI-driven PCA combined with random forest as described in claim 1, characterized in that, Sample identification specifically includes: The samples to be identified were constructed, sequenced, and SNPs were extracted. After quality control, they were compared with the reference genome to ensure that all indicators met the requirements for annotation analysis. The genotype codes of 225 core SNP loci are input into the trained random forest model, and N independent and repeated predictions are performed. The variety type and corresponding probability distribution of each prediction are recorded, where N≥100. The entropy value of the prediction result is calculated to quantify uncertainty. The formula for the entropy value is as follows: Where C represents the total number of candidate varieties, P i Let be the predicted probability of the i-th variety; Based on the Bootstrap resampling method, 1000 samples with replacement were performed on 225 core SNP loci to generate 95% confidence intervals for the probability of each variety. The final variety must meet the following conditions: the predicted probability of a variety occurs at a frequency of ≥80% in repeated predictions, the 95% confidence interval of the variety does not overlap with the confidence intervals of other varieties, and the entropy value H≤0.

5. If the above conditions are not met, the result is marked as low confidence and the manual review process is triggered.

Images