DeepFLPS genome selection system based on CNN and multi-scale feature extraction

Abstract

The invention provides a DeepFLPS genome selection system based on CNN and multi-scale feature extraction, and belongs to the field of genome selection. The inherent limitations of single feature extraction scale and insufficient sparse feature utilization of the current deep convolutional neural network structure are overcome; the system comprises a deep feature learning genome selection DeepFLPS model, the model comprises an input layer, a multi-scale feature extraction module of a binding site interaction rule, a full connection layer and an output layer, and the multi-scale feature extraction module of the binding site interaction rule adopts a parallel convolution channel structure with a plurality of sliding windows increasing progressively. A hierarchical association relationship in genotype data is captured through a progressive expansion convolution kernel, and a hierarchical double-activation function mechanism is set; the novel CNN architecture provided by the invention not only can adaptively capture the interaction effect of different SNPs interval lengths, but also can more effectively retain sparse feature information in genotype data.

Description

Technical Field

[0001] This application relates to the field of genome selection technology, and in particular to a DeepFLPS genome selection system based on CNN and multi-scale feature extraction. Background Technology

[0002] Convolutional Neural Networks (CNNs) have demonstrated great potential in the field of genomic selection (GS) due to their powerful nonlinear feature extraction capabilities, significantly surpassing traditional methods based on linear assumptions such as GBLUP and rrBLUP, and representative models such as DeepGS and DNNGP. Through end-to-end learning, CNNs can automatically capture complex interactions between genotype SNPs, such as dominance effects, additive effects, and epistatic effects, thus achieving breakthrough progress in the phenotypic prediction of many plant and animal traits.

[0003] The DeepGS model, a successful example of applying CNNs to genetics, employs a single one-dimensional convolutional layer with eight 1×18 convolutional kernels to extract features from the input SNP genotype matrix. This is followed by dimensionality reduction using a 1×4 pooling layer and the introduction of a Dropout layer with inactivation rates of 0.2, 0.1, and 0.05 to prevent overfitting. Finally, the predicted phenotypic values ​​are output via a fully connected layer. Its core advantage lies in avoiding the prior assumptions about the distribution of marker effects found in traditional methods and effectively modeling complex genetic patterns through nonlinear transformations.

[0004] The DNNGP model shares a similar architecture with DeepGS, both being built upon a multi-layered convolutional neural network. DNNGP employs a three-layer convolutional architecture, with 64 1×4 convolutional kernels in each layer. It introduces batch normalization layers to accelerate training and improve stability, and utilizes a deeper network structure to capture more abstract, higher-order features. DNNGP also employs an early stopping strategy to monitor validation set loss, optimize the training process, and explore the potential of multi-omics data integration to some extent. A comparison of the model structures of DeepGS and DNNGP is shown below. Figure 1 As shown.

[0005] From a biological perspective of model application, within short-spaced sites, adjacent SNPs may be located in strongly linked disequilibrium blocks or jointly participate in local cis-regulation; there may also be SNPs that span several genes or regulatory elements, participating in the formation of transcription factor binding clusters or local chromatin interactions; at more distant sites, even across chromosomes, distant SNPs may mediate long-distance regulation through chromatin three-dimensional structures (such as topological association domains (TADs) and chromatin loops).

[0006] DeepGS's single 1×18 kernel and DNNGP's three-layer convolutional layer setup, all using 1×4 kernels, primarily rely on fixed-size kernels for feature extraction. Although they employ deep neural networks based on convolutional operations, this single-scale receptive field struggles to effectively model interactions between SNPs of varying intervals in genomic data. Specifically, a fixed small kernel (e.g., 1×4) fails to capture long-range associations, while a fixed large kernel (e.g., 1×18) may obscure important local fine-grained patterns, failing to simultaneously capture data features at different scales, especially considering the significantly different biological characteristics of interactions between SNPs at different intervals.

[0007] While deepening the network theoretically increases the receptive field, training deep networks is difficult, and the growth of the receptive field is fixed and cumulative layer by layer, lacking the ability to explicitly extract features at different scales in parallel. This results in insufficient ability of the model to capture interaction patterns on genomic data, limiting its accuracy in resolving complex genetic architectures.

[0008] Therefore, while DeepGS and DNNGP, based on deep convolutional neural network structures, have opened up applications of deep learning, especially CNNs, in genetic sequencing (GS) and achieved positive results, their inherent limitations—such as single feature extraction scale and insufficient utilization of sparse features—severely restrict the further improvement of current CNN-GS models in terms of their ability to mine complex genetic information across the entire genome and their prediction accuracy. Therefore, there is an urgent need to design and propose novel CNN architectures that can not only adaptively capture the interaction effects of different SNP interval lengths, but also more effectively preserve sparse feature information in genotype data, especially the features of recessive or repressive biological elements that present as weak negative genetic signals. Summary of the Invention

[0009] To address the aforementioned technical issues, this application proposes a DeepFLPS genome selection system based on CNN and multi-scale feature extraction.

[0010] The technical solution adopted in this application is: a DeepFLPS genome selection system based on CNN and multi-scale feature extraction, including a DeepFLPS model for deep feature learning genome selection. The DeepFLPS model includes an input layer, a multi-scale feature extraction module for binding site interaction patterns, a fully connected layer, and an output layer. The multi-scale feature extraction module for binding site interaction patterns adopts a parallel convolutional channel structure with multiple sliding windows increasing progressively. It captures the hierarchical associations in genotype data through progressively expanding convolutional kernels and sets up a hierarchical dual activation function mechanism.

[0011] Furthermore, the multi-scale feature extraction module adopts a parallel convolutional channel structure with three sliding windows that increase in size, namely a short-range convolutional channel, a medium-range convolutional channel, and a long-range convolutional channel;

[0012] Short-range convolution channel: Employs a convolution kernel with a width of 5 and configures 32 filters to focus on detecting local strong linkage effects among 5 adjacent SNP sites, thereby capturing the biological characteristics of short-range functions of strong linkage effects;

[0013] Mid-range convolution channel: Using a convolution kernel with a width of 10, configured with 64 filters, covering approximately 10 SNP sites, to identify the synergistic effects of mid-range regulatory elements in transcription factor binding clusters;

[0014] Long-range convolutional channels: Employ a 20-width convolutional kernel with 128 filters spanning approximately 20 sites to capture associations with distant regulatory sites.

[0015] Furthermore, all convolutional layers in the multi-scale feature extraction module slide with a stride of 1.

[0016] Furthermore, the layered dual activation function mechanism uses the Leaky ReLU activation function in the convolutional layer to alleviate feature loss, and the ReLU activation function in the fully connected layer to ensure gradient stability during model training.

[0017] Furthermore, two fully connected layers are set up, with 256 and 128 neurons respectively. The fully connected layers are gradually reduced in dimensionality, and the high-dimensional genotype data is safely mapped to phenotypic features through progressive compression.

[0018] Furthermore, the DeepFLPS model was used for genomic selection in pig breeding.

[0019] Furthermore, the DeepFLPS model is trained using the N-fold cross-validation method.

[0020] The advantages of this application compared to existing technologies are as follows: The parallel convolutional channels with incremental sliding windows in the multi-scale feature extraction module proposed in this application enhance the model's explanatory power for complex traits by capturing genetic information from short-range, medium-range, and long-range SNPs with different interval lengths. Compared to a single convolutional layer feature extraction strategy, PCC is improved by 21.8%, and R... 2 The MAE error was improved by 20.7%; the hierarchical dual activation strategy combined with the dual-connection layer dimensionality reduction setting retains sparse negative signals and stabilizes and strengthens positive correlations. Compared with the single-layer ReLU activation function strategy of the whole network, the MAE error was reduced by 14.4%, and it has been verified that it can better control the MAE error of the model prediction. Attached Figure Description

[0021] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0022] Figure 1 A schematic diagram comparing the model structures of DeepGS and DNNGP;

[0023] Figure 2 This is a schematic diagram of multi-scale feature extraction provided in an embodiment of this application;

[0024] Figure 3 A schematic diagram of scale-distinguishing convolution operations provided in the embodiments of this application;

[0025] Figure 4 A schematic diagram of the ReLU activation function and the Leaky ReLU activation function provided in the embodiments of this application;

[0026] Figure 5 A schematic diagram of the hierarchical structure of the DeepFLPS model provided in the embodiments of this application;

[0027] Figure 6 This is a comparison chart of the relative differences of Shannon entropy for each model provided in the embodiments of this application;

[0028] Figure 7 MAE plots of each model on different distribution intervals of LMD phenotypic values ​​provided in the embodiments of this application;

[0029] Figure 8 A comparison chart of PCC values ​​for each model within the top 50% distribution range of phenotypic values ​​provided in this application embodiment;

[0030] Figure 9 Performance charts of the DeepFLPS model at different learning rates provided in the embodiments of this application;

[0031] Figure 10 The graph shows the performance of the DeepFLPS model under different regularization combinations provided in the embodiments of this application. Detailed Implementation

[0032] like Figures 2 to 10 As shown, this application provides a DeepFLPS genome selection system based on CNN and multi-scale feature extraction, including a Deep Feature Learning for Phenotype Selection (DeepFLPS) model. This DeepFLPS model includes an input layer, a multi-scale feature extraction module for binding site interaction patterns, a fully connected layer, and an output layer. The overall structure of the DeepFLPS model is as follows: Figure 5 As shown.

[0033] Genome function exhibits a distinct spatial hierarchy, ranging from strongly linked disequilibrium blocks and local cis-regulatory units composed of closely spaced, adjacent SNPs, to transcription factor binding clusters and chromatin interaction regions spanning several genes or regulatory elements, and further to long-distance transdomain regulation mediated by the three-dimensional structure of chromatin. This ability to distinguish interaction effects across different SNP intervals necessitates that GS models possess flexible, multi-level feature extraction capabilities, enabling them to simultaneously perceive genetic association information of SNPs at different spatial intervals.

[0034] Although increasing the number of network layers can theoretically expand the final receptive field, this expansion is slow, fixed, and cumulative layer by layer. It lacks the ability to explicitly and in parallel extract features at different scales, resulting in existing models being insufficient in capturing different interaction patterns between genomic data, especially non-local interactions that span large spatial distances. This has become a bottleneck limiting their ability to resolve complex genetic structures and improve prediction accuracy.

[0035] The core idea of ​​the multi-scale feature extraction strategy proposed in this application is to abandon single-scale convolutional kernels and instead use a set of multiple convolutional layers with different receptive field sizes. In its core design, it captures hierarchical relationships in genotype data through progressively expanding convolutional kernels. The multi-scale feature extraction strategy is illustrated as follows: Figure 2 As shown in the diagram, the four main components represent the input, convolutions at different scales, generated feature maps, and feature fusion.

[0036] (1) Input (Channel=n): The module receives input data with n channels (features).

[0037] (2) Multiple convolutions of different scales: The input data is simultaneously fed into m independent convolutional layers, each using a convolutional kernel of a different size (filter1, filter2, ..., filter...). m Thus, they have different initial receptive fields.

[0038] (3) Generating feature maps: Each convolutional layer independently extracts features from the input at its specific receptive field scale, generating corresponding feature maps (Feature map1, Feature map2, ..., Feature map). m ).

[0039] (4) Feature Fusion: All m feature maps extracted at different scales are concatenated (Cat) along the channel dimension. The final output is a feature tensor that fuses spatial information from different convolutional perspectives, with dimensions (m, width). out height out ).

[0040] Multiple hierarchical convolutional structures at different scales serve as the core functional modules in multi-scale feature extraction strategies, and their mechanism of action is as follows: Figure 3 As shown in the diagram, the cubes of different colors (red, orange, blue, and green) represent feature inputs at different scales. The convolutional operation transforms the dimensions and enhances the features of the input features in each branch's convolutional layer, generating a feature representation at that scale, corresponding to a color-coded feature map, thus capturing the data features at that scale. This can be achieved by adjusting the pooling stride, convolutional kernel size, or utilizing the outputs of different stages of the backbone. Each branch performs a convolutional operation on features at a specific scale, extracting discriminative features at that scale. For example, in the diagram, the red branch processes small-scale features, capturing fine details; the green branch processes large-scale features, obtaining the global context.

[0041] Current CNN-GS models commonly use standard ReLU as the activation function in convolutional and fully connected layers, largely ignoring the sparsity of genotype data and the importance of negative signals. While ReLU is computationally efficient and can alleviate the vanishing gradient problem, its core characteristic is that it directly sets all negative input values ​​to zero. From a biological perspective, genotype data, such as SNPs, which are encoded as 0 / 1 / 2, are inherently high-dimensional and extremely sparse. Many SNP sites have weak or zero effects, but some may contain negative effect sites, such as those carrying recessive harmful alleles (aa versus AA / Aa), sites that suppress gene expression regulation (e.g., variations in binding sites of certain transcriptional repressors), or sites associated with unfavorable phenotypes. The activation values ​​corresponding to these sites in the genotype matrix may be weakly positive or negative.

[0042] At this point, ReLU's "hard truncation" property will directly reduce these weak or negative signals to zero and discard them. This not only causes a loss of information, but more importantly, it erases the key information that affects the directionality of genetic effects.

[0043] LeakyReLU handles negative inputs by introducing a small non-zero slope, as shown in the figure. By preserving the information flow along the negative half-axis, LeakyReLU can more completely capture the complex regulatory relationships in genotypic data. Because negative values ​​can be set, the model has the potential to more accurately resolve allelic dominance and recessive relationships, identify repressive components in interactions, and characterize weak but important negative genetic associations. This not only helps improve prediction accuracy but also significantly enhances the model's interpretability for phenotypes affected by complex negative regulation, with the learned negative weight features more directly corresponding to biological repression or recessive effects.

[0044] The principle of the multi-scale feature extraction module for binding site interaction patterns proposed in this application is as follows:

[0045] To address the problem that existing convolutional neural network architectures lack the ability to distinguish scales in extracting interaction effects of SNPs with different interval lengths in genomic data, this application introduces a multi-scale feature extraction module and adopts a sliding window incremental parallel convolutional channel structure. Its core design uses progressively expanding convolutional kernels to capture hierarchical associations in genotype data. This module consists of three layers of parallel convolutional channels, with convolutional kernels of different widths and corresponding numbers of filters designed to capture label associations at different distances in genotype data. Specific settings include:

[0046] Short-range convolution channel: Employs a convolution kernel with a width of 5 and configured with 32 filters to focus on detecting local strong linkage effects between 5 adjacent SNP sites, used to capture biological features of short-range functions such as strong linkage effects within cis-regulatory elements.

[0047] Mid-range convolution channel: Using a convolution kernel with a width of 10, configured with 64 filters, covering approximately 10 SNP sites, it is used to identify the synergistic effects of mid-range regulatory elements such as transcription factor binding clusters.

[0048] Long-range convolutional channels: Employing a 20-width convolutional kernel with 128 filters spanning approximately 20 sites, these channels capture long-range regulatory site associations, such as those mediated by chromatin three-dimensional structure. The three channels extract features for short-range, mid-range, and long-range functions respectively, forming a multi-layered feature detection framework.

[0049] The number of filters in each channel increases from 32 to 64 and then to 128, and the number of convolutional kernels also increases from 5 to 20 to accommodate the gradual increase in the complexity of short, medium, and long-range features. All convolutional layers slide with a stride of 1 to ensure feature continuity. The multi-scale outputs are pooled and then fused to form a comprehensive feature representation that includes local details and global correlations. The mathematical expression of the feature output of each convolutional layer is given by formula (1).

[0050] (1);

[0051] Among them, W k ∈{5,10,20} represents the kernel width, and x represents the input genotype vector. For trainable weights, This is a bias term.

[0052] The principle of the hierarchical double activation function mechanism is as follows:

[0053] This paper addresses the issue of insufficient effective utilization of sparse genotype features in genomic data by existing convolutional neural network architectures. Conventional activation functions often lead to feature loss in sparse genotype data, affecting the model's ability to interpret the directionality of genetic effects. To balance genotype feature preservation with deep network stability, the proposed DeepFLPS model employs a hierarchical activation strategy. The convolutional layers use the Leaky ReLU activation function to mitigate feature loss, while the fully connected layers use the ReLU activation function to ensure gradient stability during model training.

[0054] Considering that the convolutional layer is the most forward-facing feature extraction stage, and the Leaky ReLU activation function can preserve weaker negative signals in the original data, and taking into account the sparse distribution of negative effect sites such as recessive alleles and repressive regulatory elements in genotype data, the mathematical expression of the Leaky ReLU activation function for the convolutional layer is Equation 2.

[0055] (2);

[0056] Among them, the negative slope coefficient The default value is usually set to 0.2. Compared to the hard truncation characteristic of traditional ReLU, the LeakyReLU activation function retains negative activation values, avoiding the loss of negative correlation signals at sparse sites, thus making the biological characteristics more complete.

[0057] Furthermore, because negative signals are preserved, the Leaky ReLU activation function is direction-sensitive and can distinguish between positive and negative genetic effects, such as synergistic or detrimental allele combinations. This functional localization for acquiring sparse biological features in convolutional layers has better adaptability than the traditional ReLU activation function.

[0058] After feature extraction from the convolutional layers, the fully connected layers, serving as the feature compression and dimension mapping stage, need to enhance significant positively correlated signals and suppress noise. The ReLU activation function is used in the fully connected layers for high-order feature selection. The mathematical expression for the ReLU activation function in the fully connected layers is Equation (3).

[0059] (3).

[0060] Fully connected layers need to robustly map the fused multi-scale features—high-dimensional features—to a low-dimensional phenotypic space. Compared to the Leaky ReLU activation function, which preserves the slope calculation in the negative interval, the ReLU activation function improves the signal-to-noise ratio of phenotypic prediction by setting the negative interval value to 0 and having a derivative of 1 in the positive interval. Furthermore, due to its lower computational complexity, it reduces the computational resource consumption of the model and lowers the risk of overfitting.

[0061] The hierarchical structure of the DeepFLPS model is shown below. Figure 5 As shown in the figure, to more accurately extract genetic information from biological genotype data and capture the interaction effects between SNPs with different interval lengths in the genotype data, DeepFLPS, based on the model strategy of deep learning convolutional neural network architecture, adopts a multi-scale feature extraction strategy. It introduces parallel convolutional kernels with kernel widths of 5, 10, and 20, and sets up three convolutional layers with increasing sliding windows to extract SNP features with different interval lengths in parallel. A layered dual activation function approach is used. The convolutional layers use the Leaky ReLU function, which, unlike the ReLU function which directly sets the biological significance of negative effect sites to zero and disables them, better preserves the information of sparse genotype data. The fully connected layers use the ReLU activation function, pursuing higher computational efficiency and better gradient stability.

[0062] Furthermore, since the input to the fully connected layer is dense features after convolution and pooling, two fully connected layers are set to maintain the dimensionality change of the data in order to ensure the training stability of the deep network. The number of neurons is 256 and 128 respectively. The fully connected layers are gradually reduced in dimensionality, and the high-dimensional genotype data is safely mapped to phenotypic features through progressive compression.

[0063] The effectiveness of the DeepFLPS model in this application will be verified by specific experiments below.

[0064] 1. Experimental environment:

[0065] The model development and experimental setup in this embodiment are based on the Tensorflow deep learning framework. In order to provide an environmental benchmark for experimental reproduction, the hardware and software descriptions of the experimental environment cover the versions of hardware (CPU, GPU), software framework (TensorFlow), operating system (Linux / Ubuntu), computing platform (CUDA), and programming language (Python, R). The configuration details are shown in Table 1.

[0066] Table 1 Experimental Environment Configuration Table

[0067]

[0068] In the current experimental setup, during the DeepFLPS model training phase, cross-validation (CV) was set to 5 folds. In each fold of cross-validation, the model was trained using the training set, and performance was monitored on the validation set. To avoid overfitting, the maximum number of training iterations was set to 6000, the initial learning rate was 0.001, the Adam optimizer was used, and the momentum parameter β1 was configured to be 0.8, with a weight decay coefficient λ of 0.00001. The batch size was set to 64, achieving a balance between parameter update efficiency and gradient stability within the limits of GPU memory.

[0069] In addition, the experiment also introduced an early stopping strategy, setting the Patience to 50 for the optimized training process: when the validation set loss function (mean absolute error between phenotypic predictions and true values) does not improve for 20 consecutive epochs and tends to stabilize, the training is terminated early and the optimal model weights are saved.

[0070] 2. Comparative Experiment

[0071] To evaluate the performance of the DeepFLPS genomic selection model, this embodiment, based on the Duroc_LC01 dataset and using a rigorous five-fold cross-validation method (four training subsets and one validation subset), compared DeepFLPS with eight other genomic selection models in pig breeding: rrBLUP, LASSO, KAML, LightGBM, RFR, SPLS, and DeepGS. The results were analyzed to assess the performance of each model. The comparative experiment used Shannon entropy, PCC, and MAE to evaluate the predictive ability of each model, analyzing the superior performance of the DeepFLPS model in phenotypic prediction accuracy, prediction stability across different phenotypic value sample ranges, and identification of individuals with high phenotypic potential.

[0072] The mainstream genome selection models selected in this comparative experiment were comprehensively chosen based on four dimensions: model accuracy and performance, model stability, model application scenarios, and model timeliness.

[0073] rrBLUP, widely recognized as the gold standard in the field, can be applied to various genomic selection scenarios. Its robust linear model framework provides a reliable benchmark for calculating genetic power in numerous studies. The LASSO model, by applying L1 regularization, screens out key genes affecting phenotypes and serves as a mainstream control for measuring model stability. KAML, through kernel function approximation, captures complex nonlinear relationships and is suitable for genomic data with complex genetic structures, widely used in animal breeding. RFR, by constructing multiple decision trees to synthesize prediction results, not only boasts high prediction accuracy but also robust model stability. LightGBM and SPLS are emerging genomic selection model algorithms that have appeared in recent years. Combined with powerful nonlinear fitting capabilities, they excel in variable selection and dimensionality reduction when handling large-scale data, removing noisy variables while preserving high-dimensional information in genotype data, resulting in high prediction accuracy. DeepGS, representing the application of deep learning, especially CNN-GS structures, in genomic selection in recent years, embodies the potential and superior performance of deep neural networks in mining complex genotype data features.

[0074] The lean meat percentage (LMP) per 100 kg was used as the predicted phenotype in this comparative experiment. To systematically evaluate the DeepFLPS model's ability to resolve complex genotype-phenotype relationships, Shannon entropy was used as the evaluation index to intuitively reflect the distribution characteristics of the uncertainty in the model's prediction results and quantify the model's phenotypic prediction ability. The fluctuation of Shannon entropy prediction values ​​for each model under five-fold cross-validation is shown in Table 2.

[0075] Table 2. Shannon entropy predictions from five-fold cross-validation for each model.

[0076]

[0077] As shown in Table 2, the Real value is 7.43891, representing the true average Shannon entropy of this phenotype. Clearly, the average Shannon entropy values ​​obtained by all model prediction methods are higher than this true value. This is because even the best-performing genomic selection models may fail to capture certain important genetic information from the genotype data, resulting in some omissions compared to the true data, thus leading to an increase in Shannon entropy.

[0078] During the experiment, DeepFLPS's predicted Shannon entropy in five-fold cross-validation was 7.43920, 7.43914, 7.43905, 7.43909, and 7.43915 in CV1 to CV5, with an average of 7.43912. The differences between the predicted and actual Shannon entropy values ​​were 0.00029, 0.00023, 0.00014, 0.00018, and 0.00024, with an average of 0.00019, which was the best among all models.

[0079] The performance of other models in the five cross-validations using Shannon entropy average prediction values ​​is as follows: DeepGS, with a prediction value of 7.43919, ranks among the best after DeepFLPS, with its entropy value differing from the actual value (Real) by only 0.00028. KAML and rrBLUP have Shannon entropy average prediction values ​​of 7.43924 and 7.43925 respectively, ranking third and fourth, with a difference of 0.00001 between them. Notably, LightGBM and RFR exhibit completely consistent prediction accuracy, both at 7.43935, which may be related to their tree-based similar feature extraction mechanism. In contrast, LASSO and SPLS, as representatives of linear methods, show prediction values ​​increasing by 0.00001 each time, forming the next tier of model performance. This gradient decay trend from nonlinear to linear models also confirms the importance of capturing the weight of complex genetic interaction nonlinear relationships on prediction accuracy at the model level.

[0080] Using the true value (Real) as the baseline zero line, after uniformly processing the predicted values ​​of each model, the average Shannon entropy difference is compared as follows: Figure 6 As shown.

[0081] The DeepFLPS entropy value, calculated relative to the baseline scale, is 2.1, which is the minimum among all control models. This indicates that its prediction results have less uncertainty and are superior to other models in this experiment. It can more effectively capture the complex association between genotype and phenotype and complete phenotype prediction.

[0082] It is worth noting that the genome selection model built based on deep learning achieved the best prediction accuracy in all five cross-validation rounds. In round CV1, DeepGS had the smallest difference between its predicted Shannon entropy value of 7.43912 and the true value of 0.00021. In rounds CV2 to CV5, DeepFLPS had the closest difference between its predicted Shannon entropy value and the true value. In particular, in round CV3, the predicted Shannon entropy value was 7.43905, with a difference of 0.00014 from the true value, which was only 45.2% of the rrBLUP Shannon entropy value of 0.00031, making it the best prediction performance among all trained models in this experiment.

[0083] Overall, the difference between DeepFLPS's predicted Shannon entropy and the true value fluctuated within the range of 0.00014-0.00029, while the fluctuation ranges for rrBLUP and LASSO models reached 0.00023-0.00037 and 0.00028-0.00051, respectively. This indicates that DeepFLPS can maintain relatively stable predictive performance even when the data distribution changes, avoiding the fluctuations caused by feature redundancy or overfitting in traditional models. These data quantifications reflect that genome selection models based on deep learning convolutional neural networks, especially the DeepFLPS model, benefit from its discriminative scale parallel convolutional units with progressively increasing sliding window sizes and multi-scale feature extraction modules, exhibiting good feature extraction capabilities and relatively stable performance in processing genotype data.

[0084] To evaluate the performance of each model in screening individuals with different phenotypic potential, and more importantly, to explore the predictive ability of DeepFLPS in screening individuals with high phenotypic development potential during the genomic selection process, this embodiment divides the 1940 samples in the dataset into three subgroups based on their LMD (100 kg eye muscle depth) phenotypic values: top 10% (approximately 195 samples), top 10%-50% (approximately 775 samples), and top 50%-100% (approximately 970 samples). The mean absolute error (MAE) is used as the evaluation criterion to compare and analyze the predictive ability of each genomic selection model in predicting individuals with different phenotypic distributions. Figure 7 This section compares the MAE of different models across different population distributions of the LMD phenotype.

[0085] Depend on Figure 7 It can be seen that DeepFLPS has a MAE of 2.42 when predicting the top 10% of high-potential individuals in phenotypic value distribution, which is the lowest among all comparative models. Specifically, compared with models such as DeepGS (2.58), KAML (2.45), rrBLUP (2.50), and RFR (2.53), DeepFLPS reduced the prediction error by 1.2%-6.2%, indicating that DeepFLPS has a strong ability to capture the genotype-phenotype association of individuals with high phenotypic values ​​and is more confident in identifying and predicting individuals with the top 10% of high potential phenotypes. Reducing the prediction bias of individuals with high phenotypic potential during genomic selection can often significantly improve the retention efficiency of superior genotypes. This characteristic is particularly important for the precise selection of core breeding populations.

[0086] In the top 10%-50% of the population with moderately high phenotypic values, DeepFLPS achieved a MAE of 1.21. In comparison, other models such as rrBLUP, DeepGS, KAML, and LASSO had corresponding MAE values ​​of 1.24, 1.28, 1.25, and 1.26 in this range, respectively, all higher than DeepFLPS's MAE. LightGBM and RFR also exceeded DeepFLPS's errors by 0.03 and 0.06, respectively. Furthermore, LASSO and SPLS also had higher error values ​​than DeepFLPS. These data indicate that DeepFLPS demonstrates high stability in phenotypic prediction for this population, outperforming other control models in the current experiment.

[0087] In the top 50%-100% phenotypic population with relatively low phenotypic values, DeepFLPS's MAE value was 0.91, maintaining an advantage over models such as LightGBM (0.93) and RFR (0.97), but not the optimal one. The phenotypic value distribution in this range is relatively wide, and DeepFLPS still exhibits relatively stable predictive performance. Although this population is not the primary target for breeding selection, the ranking among models is generally consistent with the specific high-potential populations; for example, DeepFLPS consistently ranks in the top two, while models such as RFR and SPLS rank lower. These data indicate that DeepFLPS's predictive performance maintains a high level across different phenotypic distribution ranges.

[0088] In summary, the MAE results on the predictive capabilities of various models across different distribution ranges of LMD phenotypic values ​​indicate that DeepFLPS can not only effectively identify the top 10% of high-potential individuals in phenotypic value selection, but also maintain good performance in predicting the next 50% of high-potential individuals. It also takes into account the global predictive stability of low-phenotypic groups, without sacrificing predictive ability for other phenotypic groups due to excessive focus on the prediction accuracy of high-phenotypic groups.

[0089] This cross-regional stability verification demonstrates the effectiveness of the DeepFLPS hierarchical dual activation function and dual-connection layer continuous dimensionality reduction feature value network structure in ensuring the stability of the model's prediction performance across the entire phenotypic value sample range.

[0090] To further validate DeepFLPS's ability to screen high-potential individuals with high phenotypic values ​​during the genomic selection process, this embodiment used backfat thickness (BF) (100 kg backfat thickness) as the target trait and selected individuals with the top 50% phenotypic values ​​as the subjects of study. Based on the BF phenotypic value distribution, the data were divided into two subgroups: top 10% (the top 10% of individuals with the highest phenotypic values) and top 10%-50% (individuals with phenotypic values ​​between the top 10% and 50%). The Pearson correlation coefficient (PCC), which better reflects the correlation between model predictions and actual phenotypes, was used as the evaluation index. Experimental results are as follows: Figure 8 As shown.

[0091] Depend on Figure 8 It is evident that DeepFLPS achieves superior PCC values ​​in both population groups. For the two populations with different phenotypic rankings, it performs better in the group with moderately high phenotypic values. In predicting the top 10% of individuals with high phenotypic values, DeepFLPS has a PCC of 0.052, outperforming traditional methods such as rrBLUP (0.042), KAML (0.038), and LASSO (0.035). For the top 10%-50% group, DeepFLPS has a PCC of 0.208, also higher than other models such as LightGBM (0.169), RFR (0.168), and SPLS (0.152).

[0092] It is worth noting that in predicting the phenotypes of the top 10%-50% of the population, DeepFLPS achieved a PCC value 20.2% higher than the widely used rrBLUP method (0.173), which represents prediction accuracy; and 12.4% higher than DeepGS (0.185), another deep learning model. This difference may stem from the fact that the DeepFLPS model focuses on nonlinear relationships such as gene-gene interactions and employs multi-scale feature extraction modules and the Leaky ReLU activation function, which are better at capturing complex genetic effects and negative features in genotype information.

[0093] In summary, combining MAE analysis of the predictive power of each model across different distribution ranges of LMD phenotypic values ​​and PCC analysis of the predictive power of each model across the distribution range of the top 50% of individuals with BF phenotypic values, the experimental results demonstrate that DeepFLPS performs well in accurately identifying samples with high phenotypic values. Whether predicting the highest potential individuals (top 10%) or predicting individuals with high phenotypic values ​​(top 50%) on a large scale, DeepFLPS outperforms other control models in the current experimental data sample, demonstrating good performance.

[0094] In deep learning modeling of genome selection, hyperparameter tuning is a crucial step in ensuring the model captures complex genetic regulatory patterns and significantly impacts model performance robustness. Among the hyperparameters used in model training, the learning rate controls the gradient update step size, Dropout1 prevents local feature overfitting when features extracted from convolutional layers are reduced to the first layer of the fully connected layer, and Dropout2 alleviates global feature overfitting in the second layer of the fully connected layer.

[0095] This experiment, conducted under a configured experimental environment, used the Duroc_LC01 dataset and selected the 100 kg lean meat percentage (LMP) phenotype prediction to explore the impact of three core hyperparameters—Learning rate, Dropout1 (first layer inactivation rate of fully connected layers), and Dropout2 (second layer inactivation rate of fully connected layers)—on the performance of the DeepFLPS model. PCC and R... 2 The performance evaluation index for the hyperparameter optimization experiment is established, and a detailed result analysis is performed.

[0096] (1) The impact of different learning rates on model performance

[0097] Experimental results on the impact of different learning rates on model performance are as follows: Figure 9 As shown.

[0098] As the learning rate increases from 0.0001 to 0.001, PCC and R 2 All values ​​showed an increasing trend. Specifically, the PCC value increased from the initial 0.418 to 0.514, and R... 2 The value increased from 0.526 to 0.621, indicating that the model's ability to capture the nonlinear relationship between genotype and phenotype gradually improved, with prediction accuracy and explanatory power optimizing simultaneously. However, when the learning rate was further increased to 0.1, the model performance deteriorated, with the PCC value dropping to 0.374 and R... 2 The value also decreased to 0.471. This phenomenon proves that an excessively high learning rate can indeed lead to uncontrolled gradient updates during model training, disrupting parameter convergence stability and thus weakening the model's generalization ability.

[0099] A comprehensive comparison shows that when the learning rate is 0.001, PCC and R... 2 All values ​​reached their peak values, specifically PCC at 0.514 and R... 2 The value is 0.621, which verifies that a learning rate of 0.001 performs well in balancing the convergence speed and accuracy of the model. Subsequent experiments will uniformly use 0.001 as the baseline learning rate.

[0100] (2) The impact of different combinations of regularization hyperparameters Dropout1 and Dropout2 on model performance

[0101] To systematically investigate the impact of regularization combinations of the first fully connected layer (Dropout1) and the second fully connected layer (Dropout2) on the performance of a genome selection model, this hyperparameter optimization experiment followed the principles of controlled variable and control investigation, setting up five different combination strategies for the values ​​of Dropout1 and Dropout2.

[0102] Experimental results on the impact of different combinations of regularization hyperparameters Dropout1 and Dropout2 on model performance are as follows: Figure 10 visible.

[0103] Combination 1 serves as the baseline, with Dropout1 at 0.2 and Dropout2 at 0.3, representing a standard model performance baseline. Combination 2 is the low-regularization convolutional layer group, with Dropout1 at 0.1 and Dropout2 at 0.3, where reducing the intensity of Dropout1 verifies whether it can enhance the sensitivity of local feature extraction. Combination 3 is the high-regularization fully connected layer group, with Dropout1 at 0.2 and Dropout2 at 0.4, where increasing the intensity of Dropout2 in the fully connected layer verifies whether it can suppress high-dimensional tables. The study explored whether overfitting noise in the correlation model could enhance its robustness; combination four was a symmetric fine-tuning group, where Dropout1 was 0.25 and Dropout2 was 0.35, fine-tuning the dropout rates of Dropout1 and Dropout2 to investigate whether local feature stability and global regularization could optimize fluctuation values; combination five was an asymmetric exploration group, where Dropout1 was 0.3 and Dropout2 was 0.25, to explore whether a design with high dropout1 and low dropout2 could achieve better performance.

[0104] The experimental data in the figure show that different Dropout combinations have varying impacts on model performance. The baseline group's PCC and R... 2 The values ​​are 0.526 and 0.638 respectively, clearly showing that the default parameters still have room for improvement in balancing data feature stability and generalization ability. Combination two suffers a comprehensive performance degradation due to Dropout1 (PCC=0.467, R0.638). 2 =0.594), indicating that excessive sensitivity to local features may lead to noise interference. However, appropriately increasing Dropout2 in combination three shows that it can optimize model performance (PCC=0.556, R0.594). 2 =0.667), becoming the optimal configuration. The performance of combination four is close to the benchmark group (PCC=0.521, R... 2=0.662), but it did not surpass the performance of combination three. The performance of combination five was not effectively improved either; on the contrary, due to the reduction of Dropout2, PCC and R... 2 All indicators decreased to some extent (PCC=0.463, R...). 2 =0.591).

[0105] Hyperparameter experiments have verified that Dropout2 plays a crucial role in suppressing overfitting. As can be seen from the results of combination three, within a reasonable range, Dropout2 significantly improves the model's performance metrics PCC and R². 2 It is helpful; however, Dropout1 should be kept moderate. Excessive reduction or reverse adjustment will destroy the stability of the model, as can be seen from the experimental results of combination two and combination five.

[0106] Therefore, subsequent experiments will use Dropout1 of 0.2 and Dropout2 of 0.4 as the regularization hyperparameter combination for the model to ensure the robustness of the model performance.

[0107] Ablation experiments can quantify the contribution of each experimental module in the model network structure to the phenotypic prediction accuracy and model stability. To verify the impact of the multi-scale feature extraction module and the hierarchical dual activation function mechanism in the DeepFLPS model structure on the model prediction performance, the experimental environment of the ablation experiments in this section is the same as that in the model performance comparison experiments. Based on the Duroc_LC01 dataset, with LMP (percentage of lean meat per 100 kg) as the predicted phenotypic, PCC and R... 2 Using MAE as the performance evaluation metric, five different ablation strategies were set up for ablation experiments. Two-fold cross-validation was also performed. Since the hyperparameter optimization experiments and discussions were conducted above, the optimal hyperparameter combination was set for this ablation experiment, with a learning rate of 0.001, a batch size of 64, Dropout1 of 0.2, and Dropout2 of 0.4. The specific design strategies for the five ablation experiment schemes are as follows:

[0108] (1) Option 1: Do not adopt multi-scale feature extraction design, only retain a single convolutional kernel (kernel_size=5);

[0109] (2) Scheme 2: Adopt a multi-scale feature extraction design, remove the short-range convolution channel (kernel_size=5), and retain the medium-range convolution channel (kernel_size=10) and the long-range convolution channel (kernel_size=20).

[0110] (3) Scheme 3: Adopt a multi-scale feature extraction design, remove the long-range convolution channel (kernel_size=20), and retain the short-range convolution channel (kernel_size=5) and the medium-range convolution channel (kernel_size=10).

[0111] (4) Scheme 4: A multi-scale feature extraction design is adopted, and short-range, medium-range, and long-range convolutional channels are all retained. However, a single activation function is used, and both convolutional layers and fully connected layers use the ReLU activation function;

[0112] (5) Option 5: Retain the complete network structure of the DeepFLPS model and adopt a multi-scale feature extraction design with parallel incremental convolutional kernels. Short-range, medium-range and long-range convolutional channels are all retained. A layered dual activation function mechanism is used. The convolutional layer uses the Leaky ReLU activation function to retain negative sparse features, and the fully connected layer uses the ReLU activation function to ensure gradient stability. In addition, the dual-connected layer design enables the safe dimensionality reduction output of high-dimensional feature data.

[0113] The results of the ablation experiment are shown in Table 3.

[0114] Table 3 Ablation Experiment Results

[0115]

[0116] As shown in Table 3, Scheme 1 retains only a single convolutional kernel, and its CV1 PCC is 0.452, R 2 The value is 0.578, and the MAE is 1.35; the PCC of CV2 is 0.481, and R... 2 The performance of this scheme is 0.61, and the MAE is 1.25. Compared with the subsequent schemes, the performance of this scheme is lower, indicating that a single convolutional kernel is difficult to effectively extract multi-scale features, which limits the predictive ability of the model.

[0117] Scheme 2 employs a multi-scale feature extraction design, but removes short-range convolution channels; its CV1 PCC is 0.512, and R1 is... 2 The value is 0.625, and the MAE is 1.3; the PCC of CV2 is 0.53, and R... 2 The performance of the first scheme is 0.641, and the MAE is 1.26, showing some improvement compared to the first scheme. However, the lack of short-range convolutional channels still limits the model's performance. Scheme 3 also uses multi-scale feature extraction, but removes long-range convolutional channels. Its CV1 PCC is 0.528, and R... 2 The value is 0.628, and the MAE is 1.32; the PCC of CV2 is 0.548, and R... 2The performance of scheme 4 is 0.642, with an MAE of 1.27, which is better than schemes 1 and 2, but does not reach the level of a complete model, indicating that long-range convolutional channels are irreplaceable in capturing long-distance genetic feature associations. Scheme 4 retains all convolutional channels for multi-scale feature extraction, but uses a single ReLU activation function, with a PCC of 0.543 and R0.642. 2 The value is 0.652, the MAE is 1.24; the PCC of CV2 is 0.569, and R... 2 The first scheme achieves a PCC of 0.682 and an MAE of 1.18, showing improvement in feature extraction. However, the single activation function fails to fully utilize the characteristics of each layer's features, resulting in a performance gap compared to the complete model. Scheme 5 retains the complete DeepFLPS model structure, employs a multi-scale feature extraction module, and sets up parallel convolution channels with incrementally increasing sliding window sizes and a hierarchical dual activation function mechanism. Experimental results show that CV1 achieves a PCC of 0.608 and R... 2 The coefficient of performance (COP) is 0.71, and the MAE is only 1.12; Scheme 5 achieves the best result in CV2, with a PCC of 0.634 and R². 2 The coefficient of performance (COP) was 0.736, and the MAE was only 1.04. Compared with the other four control schemes, Scheme 5 outperformed the others in cross-validation due to preserving the complete DeepFLPS model. Furthermore, although the Δ value representing performance fluctuation in the two cross-validations of Scheme 5 was not optimal, the PCC and R² values ​​were significantly higher. 2 The fluctuation is only 0.026, and the MAE fluctuation is only 0.08, indicating that its stability can be guaranteed.

[0118] Analyzing the performance fluctuations of other groups, although Scheme 4 retains the multi-scale feature extraction convolutional channel design, the lack of a layered dual activation function leads to a larger MAE error, and the PCC and R... 2 The results were also lower than those of Scheme 5, verifying the necessity of hierarchical dual activation functions and dual-connected layer dimensionality reduction for feature preservation. Schemes 2 and 3 showed higher MAE errors than Scheme 5, while PCC and R... 2 Compared to scheme five, both have been reduced, indicating that short-range and long-range convolutional channels have complementary effects in capturing the distance genetic effects of different sites in local and global genetic effect modeling, and parallel multi-range convolutional channels are indispensable.

[0119] In summary, ablation experiments demonstrate that the multi-scale feature extraction and hierarchical dual activation function mechanism of the DeepFLPS model are key to improving its phenotypic prediction performance. The parallel convolutional channels with an incremental sliding window in the multi-scale feature extraction module enhance the model's explanatory power for complex traits by capturing genetic information from short-, medium-, and long-range SNPs with different interval lengths. Compared to a single convolutional layer feature extraction strategy, PCC is improved by 21.8%, and RC is significantly higher. 2The MAE error was improved by 20.7%; the hierarchical dual activation strategy combined with the dual-connection layer dimensionality reduction setting retains sparse negative signals and stabilizes and strengthens positive correlations. Compared with the single-layer ReLU activation function strategy of the whole network, the MAE error was reduced by 14.4%, and it has been verified that it can better control the MAE error of the model prediction.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A DeepFLPS genome selection system based on CNN and multi-scale feature extraction, characterized by: This includes the DeepFLPS model for genome selection based on deep feature learning. The DeepFLPS model consists of an input layer, a multi-scale feature extraction module for binding site interaction patterns, a fully connected layer, and an output layer. The multi-scale feature extraction module for binding site interaction patterns adopts a parallel convolutional channel structure with multiple sliding windows that increase progressively. It captures hierarchical associations in genotype data through progressively expanding convolutional kernels and sets up a hierarchical dual activation function mechanism.

2. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to claim 1, characterized in that: The multi-scale feature extraction module adopts a parallel convolutional channel structure with three sliding windows that increase in size: short-range convolutional channel, medium-range convolutional channel, and long-range convolutional channel. Short-range convolution channel: Employs a convolution kernel with a width of 5 and configures 32 filters to focus on detecting local strong linkage effects among 5 adjacent SNP sites, thereby capturing the biological characteristics of short-range functions of strong linkage effects; Mid-range convolution channel: Using a convolution kernel with a width of 10, configured with 64 filters, covering approximately 10 SNP sites, to identify the synergistic effects of mid-range regulatory elements in transcription factor binding clusters; Long-range convolutional channels: Employ a 20-width convolutional kernel with 128 filters spanning approximately 20 sites to capture associations with distant regulatory sites.

3. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to claim 2, characterized in that: All convolutional layers in the multi-scale feature extraction module slide with a stride of 1.

4. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to claim 3, characterized in that: The layered dual activation function mechanism uses the Leaky ReLU activation function in the convolutional layer to alleviate feature loss, and the ReLU activation function in the fully connected layer to ensure gradient stability during model training.

5. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to claim 4, characterized in that: Two fully connected layers are set up, with 256 and 128 neurons respectively. The fully connected layers are progressively reduced in dimensionality, and the high-dimensional genotype data is safely mapped to phenotypic features through progressive compression.

6. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to any one of claims 1-5, characterized in that: The DeepFLPS model is used for genomic selection in pig breeding.

7. The DeepFLPS genome selection system based on CNN and multi-scale feature extraction according to claim 6, characterized in that: The DeepFLPS model was trained using the N-fold cross-validation method.

Images