Plant alternative splicing prediction method based on large model distillation and evidence constraint

Abstract

This invention relates to a plant alternative splicing prediction method based on large-model distillation and evidence constraints, comprising: acquiring plant genomic DNA sequences and their corresponding splicing site classification and annotation data; performing K-mer Tokenization and One-Hot Encoding in parallel on each DNA sequence, with Token IDs as input to the teacher model and tensors as input to the student model; constructing a joint loss based on the outputs of the teacher and student models for joint optimization, thereby transferring the semantic knowledge of the teacher model to the student model; using the trained student model to perform a sliding window scan on the plant genomic DNA sequence to be tested, and determining the splicing site of the plant genomic DNA sequence to be tested based on the site category corresponding to the highest probability output by the student model; and constructing a dual-model distillation structure to transfer the semantic representation capability of the pre-trained large DNA model to a lightweight convolutional network, thereby improving the stability and generalization ability of plant genome splicing site identification while balancing prediction performance and computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of plant splice site prediction, and more particularly to a method for predicting plant alternative splicing using large-scale model distillation and evidence constraints. Background Technology

[0002] The accurate identification of RNA splicing sites is of great significance in plant gene function research, complex trait analysis, and crop molecular breeding. Splicing abnormalities can lead to changes in exon composition, intron retention, or the formation of recessive splicing sites, resulting in changes in amino acid sequences and protein structure, which in turn affect crop stress resistance, yield, and environmental adaptability. Compared with animals, the proportion of intron retention in plant genomes is significantly higher, and splicing regulatory mechanisms are more complex and species-specific. Therefore, constructing high-precision, cross-species generalizable, and efficiently deployable plant splicing prediction models is a key technical issue in modern agricultural genomics research. Currently, the technical approaches for splicing site prediction mainly include the following categories: 1. Rule-based or traditional machine learning methods: These methods typically rely on manually constructed sequence features, such as k-mer frequency, GC content, and splicing motifs, combined with classification models like Support Vector Machines (SVM) and Random Forests (RF) for prediction. Essentially, these methods represent shallow modeling, relying on manually designed features and struggling to automatically learn complex high-order sequence dependencies. They cannot effectively characterize the impact of long-distance regulatory elements on splicing sites; their generalization ability is limited when splicing patterns differ significantly between species; and they lack stability for large-scale whole-genome scanning tasks. Therefore, these methods are insufficient to meet the high-precision splicing prediction requirements of complex plant genomes.

[0003] 2. Splicing prediction methods based on deep convolutional neural networks: These methods expand the receptive field through multi-layer convolutional structures and dilated convolutions, enabling them to capture long-distance sequence dependencies to a certain extent and achieve high prediction accuracy on human genome data. However, most of these methods are trained on human data, and the model parameters and structure design are not specifically optimized for plant splicing characteristics; the convolutional structure is essentially still mainly for local pattern extraction, lacking global semantic modeling capabilities; performance degrades significantly in cross-species or data distribution variation scenarios; and for plant genomes with a high intron retention rate, existing models struggle to reliably identify atypical splicing patterns. Therefore, directly applying existing deep convolutional models to plant genome prediction is unlikely to achieve reliable cross-species generalization.

[0004] 3. Pre-training methods based on large-scale DNA models: In recent years, DNA pre-trained models based on the Transformer architecture (such as AlphaGenome and DNABERT-2) have obtained sequence-level semantic representations through large-scale self-supervised learning, demonstrating strong expressive power in various downstream tasks. However, these models typically employ a full attention mechanism, whose computational complexity increases quadratically with sequence length, resulting in large model parameter sizes and high computational resource consumption. They are unsuitable for high-resolution base-by-base prediction tasks, have insufficient inference efficiency in whole-genome scanning scenarios, and are difficult to deploy efficiently in resource-constrained environments. Especially in plant genomes (such as polyploid crops), due to the massive genome size, directly using large-scale pre-trained models for site-by-site prediction presents a significant computational bottleneck.

[0005] In summary, existing technologies either rely on artificial features and have limited expressive capabilities, making them ill-suited to the complex splicing patterns of plants; or they are computationally expensive and difficult to deploy, failing to meet the demands for high-resolution predictions at the whole-genome scale. Furthermore, current analytical methods can only output results for gene variation and multiple transcript splicing events, failing to provide hierarchical analysis at the base sequence and biological pathway levels. On the one hand, traditional methods relying solely on predicting splicing sites struggle to accurately interpret base-level motif features and mine functional motifs, failing to elucidate the action sites and regulatory mechanisms of mutations and alternative splicing at the microscopic base level. On the other hand, existing methods struggle to integrate global biological pathways, gene regulatory networks, and molecular interactions to systematically analyze the connections between alternative splicing events and upstream / downstream regulatory mechanisms, and cannot generate natural language explanations for researchers. Therefore, current technologies lack a global biological pathway analysis mechanism based on large-scale AI models, making it difficult to connect molecular regulatory logic at the pathway network level and interpret prediction results. Given the diverse genomic structures, complex splicing regulatory mechanisms, and cross-species generalization requirements of plants, the key challenges that current technologies urgently need to address are how to maintain lightweight models and efficient inference capabilities while incorporating the semantic expression advantages of large-scale pre-trained models to achieve a balance between high accuracy and high generalization, and how to provide a unified and convincing explanation for complex splicing regulatory mechanisms. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing a plant alternative splicing prediction method based on large model distillation and evidence constraints. By constructing a dual-model distillation structure, the semantic representation capabilities of a pre-trained large DNA model are transferred to a lightweight convolutional network, thereby improving the stability and generalization ability of plant genome splicing site identification while balancing prediction performance and computational efficiency.

[0007] According to a first aspect of the present invention, a method for predicting plant alternative splicing based on large-scale model distillation and evidence constraints is provided, comprising: Step 1: Construct a dual-model distillation structure that includes a teacher model and a student model; Step 2, obtaining input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data; Step 3: Perform K-mer Tokenization and One-Hot Encoding in parallel on each DNA sequence to obtain Token IDs and B×L×4 tensors respectively; the Token IDs are the input of the teacher model, and the tensors are the input of the student model. Step 4: Construct a joint loss based on the outputs of the teacher model and the student model and perform joint optimization to realize the transfer of semantic knowledge from the teacher model to the student model; Step 5: Use the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determine the splicing site on the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model. Step 6: Calculate the base-level attribution score using the trained student model, and identify candidate sequence motifs based on the base-level attribution score; perform cluster analysis, enrichment analysis, and computer simulation mutation analysis on the candidate sequence motifs to characterize the splicing-related regulatory patterns. Step 7: Based on the integration of multiple transcript splicing events and recessive mutations, the effects on amino acid sequences, protein structures, and functional phenotypic variations are explained, a biological causal chain is constructed, and a natural language interpretation is provided based on a generative large language model.

[0008] Based on the above technical solution, the present invention can also be improved as follows.

[0009] Optionally, the teacher model constructed in step 1 includes: a sequence semantic encoder, an average pooling aggregation module, a linear classification head, and a probability output module with temperature regulation; the sequence semantic encoder is a pre-trained large genome model based on the Transformer architecture; the Transformer layer of the sequence semantic encoder is kept frozen during the training phase. The Transformer layer of the sequence semantic encoder generates a contextual semantic representation based on the input Token IDs; The average pooling aggregation module performs MeanPooling operations on the context semantic representation output by the Transformer layer, aggregating the base-level representation into sequence-level embeddings. The sequence-level embedding is input into the logistic regression head of the linear classification head to generate classification logits; The temperature-controlled probability output module generates a smooth soft probability distribution based on the classification logits through a Softmax layer with a temperature parameter τ. .

[0010] Optionally, the student model constructed in step 1 includes: an input layer, a local feature extraction layer, a classification head, and a Softmax output module; The input layer receives the input tensors and retains fine-grained information at the base level; The local feature extraction layer extracts local sequence features of the tensor through one-dimensional convolution; The classification head is used to input the encoded features into the classification layer to generate classification logits; The Softmax output module is used to perform Softmax calculation on the classification logits to obtain the student model prediction probability. .

[0011] Optionally, the classification categories of the splice site classification and labeling data in step 2 include: recipient, donor, and neither; Step 2 further includes: dividing the input data into a training set and a test set.

[0012] Optionally, the joint optimization process in step 4 includes: Step 401: Calculate the output of the student model. Soft labels of the teacher model The KL divergence between them results in distillation loss. ; Step 402: Calculate the cross-entropy loss between the classification logits of the student model and the true label. ; Step 403, calculate the total loss function as follows: , The weights are set; the teacher model and the student model are jointly optimized based on the total loss function to realize the transfer of teacher semantic knowledge to the student model; Step 404: During the training phase, only the parameters of the student model are updated, while the teacher model remains frozen; and during the training phase of the student model, the central bases are randomly masked.

[0013] Optionally, step 5 may be followed by: Step 6: Construct a splicing center window and obtain the base-by-base contribution distribution within the window; construct a reference sequence based on the original plant genomic DNA sequence, and calculate the base-by-base contribution degree at each position within the window based on the base-by-base contribution distribution and the reference sequence.

[0014] Optionally, step 6 includes: Step 601: Perform one-hot encoding on the input data in the window and record it as the actual input. L is the length of the splice center window; the target category will be... The output logit is denoted as By analyzing each position within the window The four-channel base contribution is calculated to obtain the base-by-base contribution distribution within the window range; Step 602: Fit a first-order Markov model to the original plant genomic DNA sequence set to generate a reference sequence; generate a reference sequence for each window. The reference sequence is obtained by one-hot encoding the reference sequence. n is a positive integer; Step 603: Under the gradient-based feature attribution algorithm rules, the true input is... With the Reference sequences By comparison, the channel-level attribution tensor is obtained. ;right The attribution tensor of the reference sequence is averaged to obtain ; using the actual input Mapping the hypothetical contributions to the actual contributions yields the base-by-base contribution scores for each position. .

[0015] Optionally, step 6 may be followed by: Step 7, based on the contribution vector at each location The absolute value threshold is used to screen continuous high-contribution segments, and the sequences corresponding to the continuous high-contribution segments are extracted as candidate splicing regulation motifs; the motifs are deredundant and a motif-original occurrence position mapping table is established; according to whether they conform to the typical AG / GT splicing core mode, they are divided into two categories: standard type and special type, and the distance matrix is ​​calculated and hierarchical clustering is performed to form a motif family structure to distinguish the core signal from atypical auxiliary regulation motifs.

[0016] Optionally, step 7 may be followed by: Step 8: Using the same window length, reference sequence generation strategy, and contribution calculation formula for different plant species, obtain the motif sets and family structures; by comparing the composition, branching and aggregation of Canonical / Special motif families in different species, evaluate whether the splicing regulation signals identified by the model maintain a stable pattern in cross-species scenarios.

[0017] According to a second aspect of the present invention, a plant alternative splicing prediction system based on large-scale distillation and evidence constraints is provided, comprising: a teacher model, a student model, a distillation training module, and a splicing site prediction module; The input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data; two parallel encoding methods, K-mer Tokenization and One-Hot Encoding, are performed on each DNA sequence to obtain Token IDs and B×L×4 tensors, respectively; the Token IDs are the input of the teacher model, and the tensors are the input of the student model; The distillation training module constructs a joint loss based on the outputs of the teacher model and the student model and performs joint optimization to realize the transfer of semantic knowledge from the teacher model to the student model; The splice site prediction module uses the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determines the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model.

[0018] This invention provides a large-scale model distillation and evidence-constrained method and system for predicting plant alternative splicing, which significantly improves the accuracy of cross-species splicing site prediction and exhibits stronger comprehensive performance in plant splicing prediction tasks. It demonstrates good stability and generalization ability, maintaining stable prediction performance even without retraining for the target species, indicating that the method can effectively address differences in data distribution among species. It improves the interpretability of prediction results, constructing a splicing regulatory signal identification and functional interpretation system based on base-level contribution analysis. It achieves lightweight and efficient deployment and enhances cross-species transferability; in the inference stage, only a student model is needed to complete base-by-base prediction, which can be directly applied to unseen genomes without retraining or species-specific fine-tuning. This reduces computational resource consumption and verifies the cross-species representation transferability provided by the knowledge distillation strategy, supporting rapid scanning at the whole-genome scale. It supports downstream analysis of mutation effects; a generative large-scale language model integrating subtype sequences, structures, functions, and trait evidence obtained from various databases and tools links sequence variations with splicing results and structural consequences, thus moving beyond simple prediction towards evidence-based reasoning and enhancing the application value of the method in practical breeding and functional research. Attached Figure Description

[0019] Figure 1 A flowchart of a plant alternative splicing prediction method based on large-scale model distillation and evidence constraints provided in this embodiment of the invention; Figure 2 This is an overview of the splice site prediction model structure and a schematic diagram of the core modules of the SplicePlant provided in this embodiment of the invention. Figure 3 A schematic diagram of the process for mining multi-plant species splicing-specific motifs based on DeepLIFT provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the AI-assisted parsing process and its application in the decoding of splicing functions provided in an embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] This invention proposes a plant splice site prediction method (Splice-Plant) based on a knowledge distillation framework. This prediction method includes: Step 1: Construct a dual-model distillation structure that includes a teacher model and a student model.

[0022] Step 2, obtaining input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data.

[0023] Step 3: Perform two parallel encoding methods, K-mer Tokenization and One-Hot Encoding, on each DNA sequence to obtain Token IDs and B×L×4 tensors, respectively; Token IDs are the input of the teacher model, and the tensors are the input of the student model.

[0024] Step 4: Construct a joint loss based on the outputs of the teacher model and the student model for joint optimization, thereby transferring the semantic knowledge of the teacher model to the student model.

[0025] Step 5: Use the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determine the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model.

[0026] Step 6: Calculate the base-level attribution score using the trained student model, and identify candidate sequence motifs based on the base-level attribution score; perform cluster analysis, enrichment analysis, and computer simulation mutation analysis on the candidate sequence motifs to characterize the splicing-related regulatory patterns.

[0027] Step 7: Based on the integration of multiple transcript splicing events and recessive mutations, the effects on amino acid sequences, protein structures, and functional phenotypic variations are explained, a biological causal chain is constructed, and an AI big data model is used to integrate and provide a natural language interpretation.

[0028] This invention provides a plant alternative splicing prediction method based on large-model distillation and evidence constraints. By constructing a dual-model distillation structure, the semantic representation capabilities of a pre-trained large DNA model are transferred to a lightweight convolutional network. This improves the stability and generalization ability of plant genome splicing site identification while balancing prediction performance and computational efficiency.

[0029] Example 1

[0030] Embodiment 1 of this invention is an embodiment of a large-scale model distillation and evidence-constrained plant alternative splicing prediction method provided by this invention. This prediction method embodiment includes: Step 1: Construct a dual-model distillation structure that includes a teacher model and a student model.

[0031] In one possible embodiment, the teacher model constructed in step 1 includes: a sequence semantic encoder, an average pooling aggregation module, a linear classification head, and a probability output module with temperature regulation; the sequence semantic encoder is a pre-trained large genome model based on the Transformer architecture, and in one possible embodiment, the pre-trained large genome model is a pre-trained DNABERT-2 (approximately 117M parameters); the Transformer layer of the sequence semantic encoder remains frozen during the training phase and does not participate in gradient updates.

[0032] The Transformer layer of the sequence semantic encoder generates a contextual semantic representation based on the input Token IDs.

[0033] The average pooling aggregation module performs Mean Pooling operations on the context semantic representation output by the Transformer layer, aggregating the base-level representation into sequence-level embeddings.

[0034] Sequence-level embeddings are used to generate classification logits by inputting a logistic regression head into a linear classification head.

[0035] The probability output module with temperature control generates a smooth soft probability distribution based on classification logits through a Softmax layer with a temperature parameter τ. , used for monitoring subsequent distillation.

[0036] In one possible implementation, the student model constructed in step 1 includes: an input layer, a local feature extraction layer, a dilated residual encoder, a skip connection structure, an attention enhancement module, a classification head, and a Softmax output module.

[0037] The input layer receives the input tensor and retains fine-grained information at the base level; in specific implementations, the input tensor is a One-Hot Tensor.

[0038] The local feature extraction layer extracts local sequence features of tensors through one-dimensional convolution.

[0039] A hollow residual encoder is used to expand the receptive field and model the regulation dependence at different scales; the expansion rate (AR) is set to 1, 4, 10, and 20.

[0040] Skip connection structures are used to introduce skip connections between residual modules, enhancing gradient propagation stability and improving deep training performance.

[0041] The attention enhancement module is used to introduce an attention mechanism after high-level features to weight and enhance key regulatory regions, thereby improving the ability to identify complex splicing regulation patterns.

[0042] The classification head is used to input the encoded features into the classification layer to generate classification logits; in specific implementations, classification logits include: Acceptor, Donor, and Neither.

[0043] The Softmax output module is used to perform Softmax calculation on the classification logits to obtain the predicted probabilities of the student model. .

[0044] Step 2, obtaining input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data.

[0045] In one possible embodiment, the classification categories of the splice site classification and annotation data in step 2 include: receptor, donor, and neither, forming a base-level supervised learning dataset; step 2 also includes: dividing the input data into training and testing sets, while supporting cross-species training and cross-species testing scenarios, for evaluating the model's generalization ability.

[0046] Three-class Ground Truth labels are constructed based on the actual splice site annotations and uniformly converted into a supervised learning format for cross-entropy loss calculation.

[0047] Step 3: Perform two parallel encoding methods, K-mer Tokenization and One-Hot Encoding, on each DNA sequence to obtain Token IDs and B×L×4 tensors, respectively; Token IDs are the input of the teacher model, and the tensors are the input of the student model.

[0048] Step 4: Construct a joint loss based on the outputs of the teacher model and the student model for joint optimization, thereby transferring the semantic knowledge of the teacher model to the student model.

[0049] In one possible embodiment, the joint optimization process in step 4 includes: Step 401, calculate the output of the student model. Soft labeling of teacher models The KL divergence between them results in distillation loss. .

[0050] Step 402: Calculate the cross-entropy loss between the student model's classification logits and the true label. Implement a real-labeling supervision mechanism.

[0051] Step 403, calculate the total loss function as follows: , The weights are set; the teacher model and student model are jointly optimized based on the total loss function to realize the transfer of teacher semantic knowledge to the student model.

[0052] Step 404: During the training phase, only the parameters of the student model are updated, while the teacher model remains frozen, thereby achieving knowledge compression. Furthermore, during the training phase of the student model, the central bases are randomly masked to reduce the model's dependence on explicit local patterns and enhance its cross-species generalization ability.

[0053] Step 5: Use the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determine the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model.

[0054] In practice, after training, the teacher model is removed, and only the student model is retained for inference, achieving lightweight deployment. A sliding window scan is performed on the target plant genome to predict splice sites base by base. The predicted probability of each central base belonging to Acceptor, Donor, or Neither is output, and the category is determined based on the highest probability. It can be used for whole-genome splice site identification, variation impact assessment, and screening of anomalous splice sites.

[0055] Step 6: Calculate the base-level attribution score using the trained student model, and identify candidate sequence motifs based on the base-level attribution score; perform cluster analysis, enrichment analysis, and computer simulation mutation analysis on the candidate sequence motifs to characterize the splicing-related regulatory patterns.

[0056] In one possible embodiment, step 6 includes: A splicing center window is constructed to obtain the base-by-base contribution distribution within the window; a reference sequence is constructed based on the original plant genomic DNA sequence, and the base-by-base contribution of each position within the window is calculated based on the base-by-base contribution distribution and the reference sequence.

[0057] In one possible embodiment, the process of calculating the base-by-base contribution at each position within the window includes: Step 601: Perform one-hot encoding on the input data in the window and record it as the actual input. L is the length of the splice center window; the target category will be... The output logit is denoted as By analyzing each position within the window The four-channel base contributions are calculated to obtain the base contribution distribution within a window range, which is used to characterize the splice-related sequence signal learned by the student model.

[0058] In practice, a splice center window of length L can be set as the basic analysis unit, and the central base of the window can be the prediction target.

[0059] Step 602: A first-order Markov model is fitted onto the original plant genomic DNA sequence set to generate a reference sequence; its transition probability is defined as... Generate for each window The reference sequence is obtained by one-hot encoding the reference sequence. n is a positive integer, preferably 3.

[0060] This can avoid bias caused by a single reference sequence and prevent background structures from disrupting dinucleotide frequencies, thus ensuring that the statistical characteristics of the background sequence are consistent with the real sequence, making the contribution results more robust and reproducible.

[0061] Step 603: Under the gradient-based feature attribution algorithm rules, process the real input... With the Reference sequences By comparison, the channel-level attribution tensor is obtained. ;right The attribution tensors of the reference sequences are averaged to obtain Use real input Mapping the hypothetical contributions to the actual contributions yields the base-by-base contribution scores for each position. Thus, the length is obtained. Contribution vector This serves as a unified input for subsequent high-contribution region identification, motif extraction, and cross-species comparison.

[0062] In a preferred embodiment, the gradient-based feature attribution algorithm rule can be the rescale rule of WieldDeepLIFT.

[0063] In one possible embodiment, step 6 further includes: Contribution vectors based on each position The absolute value threshold is used to screen continuous high-contribution segments, and the sequences corresponding to the continuous high-contribution segments are extracted as candidate splicing regulation motifs. Redundancy is removed from the motifs and a motif-original occurrence position mapping table is established to retain traceability. Then, according to whether they conform to the typical AG / GT splicing core pattern, they are divided into two categories: standard type Canonical and special type. The distance matrix is ​​calculated and hierarchical clustering is performed to form a motif family structure to distinguish the core signal from atypical auxiliary regulation motifs.

[0064] In one possible embodiment, step 6 further includes: By applying consistent window lengths, reference sequence generation strategies, and contribution calculation formulas to different plant species, we obtained various motif sets and family structures. By comparing the composition, branching and aggregation of Canonical / Special motif families in different species, we evaluated whether the splicing regulation signals identified by the model maintained stable patterns in cross-species scenarios, thereby supporting the generalizable interpretable signals brought about by knowledge distillation.

[0065] In one possible embodiment, the present invention proposes a large-scale model distillation and evidence-constrained plant alternative splicing prediction method, which further includes population variation integration analysis: defining "functional motif regions" as genomic fragments whose DeepLIFT contribution exceeds the threshold set in the motif discovery step, and constructing gene background regions with matching length distributions as controls; calculating variation density using "number of SNPs per kb", and employing bootstrap resampling (e.g. (This is repeated) to obtain the background distribution and confidence interval, and use the normalized value of the variation density of functional motifs relative to the background mean as the relative mutation rate index; in addition, it can be binned according to DeepLIFT scores, and the average SNP density of different bins can be calculated to test the consistency relationship between "model importance-evolutionary constraints" and verify the biological rationality of high contribution signals at the population level.

[0066] In one possible embodiment, the present invention proposes a plant alternative splicing prediction method based on large-scale model distillation and evidence constraints, which further includes multiple transcript and structural impact interpretation: the predicted splicing change events are analyzed at the transcript level to form multiple transcript sequences and translated into protein sequences; at the structural level, existing structural resources (such as UniProt / structure library entries) are prioritized and the structures of missing entries are predicted, and then a weighted score is performed by combining structural quality / similarity assessment and functional annotation (such as domains, homology evidence, etc.) to form a set of structured evidence, providing a "verifiable evidence list" for subsequent interpretation generation.

[0067] Step 7: Based on the integration of multiple transcript splicing events and recessive mutations, the effects on amino acid sequences, protein structures, and functional phenotypic variations are explained, a biological causal chain is constructed, and an AI big data model is used to integrate and provide a natural language interpretation.

[0068] In one possible embodiment, this invention proposes a plant alternative splicing prediction method based on large-scale model distillation and evidence constraints, which further includes a generative large-scale language model-enhanced explanation module: "Splicing event type - amino acid change - structural change - functional / phenotypic clues" are organized as structured input, combined with the evidence list and scoring results obtained in the previous step, and a large-scale language model is invoked to generate standardized explanatory text; the output is given in the form of "biological causal chain + evidence list + reproducible experimental suggestions / verification path," thereby achieving a mechanistic inference report on the association between splicing changes and protein function and potential phenotypic factors within a limited evidence scope, improving the readability and systematicity of the explanation results.

[0069] Figure 1 A flowchart illustrating a large-scale model distillation and evidence-constrained plant alternative splicing prediction method is provided in this embodiment of the invention. Figure 1 In the given embodiments, splicing site prediction across more than 30 plant species is achieved through knowledge distillation based on a large-scale gene model. The trained model is used to calculate base-level attribution scores to identify candidate sequence motifs; these motifs are then subjected to cluster analysis, enrichment analysis, and computer-simulated mutation analysis to characterize splicing-related regulatory patterns. Multiple transcript splicing events and recessive mutations are explained through their impact on amino acid sequences, protein structures, and functional phenotypic variations, constructing a biologically plausible causal chain supported by automated evidence collection and reporting, which is then integrated by the large-scale AI model and explained in natural language.

[0070] Example 2

[0071] Embodiment 2 of the present invention is an embodiment of a plant alternative splicing prediction system based on large-scale distillation and evidence constraints provided by the present invention. The embodiment of the prediction system includes: a teacher model, a student model, a distillation training module, and a splicing site prediction module.

[0072] The input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data; two parallel encoding methods, K-mer Tokenization and One-Hot Encoding, are performed on each DNA sequence to obtain Token IDs and B×L×4 tensors, respectively; Token IDs are the input of the teacher model, and the tensors are the input of the student model.

[0073] The distillation training module constructs a joint loss based on the outputs of the teacher model and the student model for joint optimization, thereby realizing the transfer of semantic knowledge from the teacher model to the student model.

[0074] The splice site prediction module uses a trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determines the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the highest probability output by the student model.

[0075] Figure 2 This is an overview of the SplicePlant splice site prediction model structure and a schematic diagram of its core modules provided in this embodiment of the invention. Figure 2 In the given embodiment, 'a' represents a cross-species splice site prediction framework based on knowledge distillation. The teacher model encodes K-mer sequences using DNABERT-2, outputting a soft probability distribution after mean pooling and a classification head. The student model takes One-Hot encoding as input, generating prediction results through one-dimensional convolution, multi-scale dilated residual modules (AR = 1, 4, 10, 20), and an attention module. During the training phase, KL divergence and cross-entropy loss are combined for optimization to achieve knowledge transfer. 'b' represents a schematic diagram of the key sub-module structure, including the Transformer encoding layer structure, temperature softening calculation, logistic regression classification head, mean pooling implementation, and the multi-scale convolution and residual connection structure within the dilated residual module, demonstrating the feature transfer and calculation process within the model.

[0076] Figure 3 This is a schematic diagram of multi-plant species splicing-specific motif mining based on DeepLIFT provided in an embodiment of the present invention. Figure 3 This paper presents an interpretability framework based on DeepLIFT and visualizations of the contribution scores of randomly selected gene fragments at the nucleotide level. Nucleotides highly reflect their contribution to splice site identification, with positive values ​​representing splice-enhancing elements and negative values ​​representing splice-silencing elements. The x-axis represents the relative position on the genomic locus. The conservation and differentiation of splice regulatory motifs among the four species are also shown. Genome clustering analysis identified ubiquitous classical signals (AG / GT) and lineage-specific auxiliary regulatory motifs.

[0077] Figure 4 This is a schematic diagram illustrating the AI-assisted parsing process provided in an embodiment of the present invention and its application in the decoding of splicing functions. Figure 4The diagram illustrates the three-stage process, including structure acquisition, annotation and scoring, and AI-driven functional analysis. Validation was performed using the Arabidopsis thaliana FLM gene. Stage 1: Top image: Gene structure model, labeling SNP sites 28,958,437 (orange arrows) and the resulting cryptic splicing events. Bottom image: Comparison of protein structures of wild-type (Col-0) and mutant (Can0), and AI-generated phenotypic predictions. Stage 3: Application in the upland cotton gene DC085. Top image: Gene structure model, showing intron retention events between transcripts DC085.1.1 (long exon 4) and DC085.1.2. Bottom image: Predicted protein structures of the two splicing isoforms, and corresponding biological hypotheses generated by the AI ​​analysis module.

[0078] It is understood that the plant alternative splicing prediction system based on large-scale distillation and evidence constraints provided by this invention corresponds to the plant alternative splicing prediction methods based on large-scale distillation and evidence constraints provided in the foregoing embodiments. The relevant technical features of the plant alternative splicing prediction system based on large-scale distillation and evidence constraints can be referred to the relevant technical features of the plant alternative splicing prediction methods based on large-scale distillation and evidence constraints, and will not be repeated here.

[0079] Example 3

[0080] Embodiment 3 of this invention is a first application embodiment of the plant alternative splicing prediction method based on large-scale model distillation and evidence constraints provided by this invention. This first application embodiment is an embodiment of the training and deployment method for a plant splicing site prediction method based on a knowledge distillation mechanism, including: 1. Data Acquisition and Preprocessing The training data used in this embodiment comes from publicly available plant genome data resources. The model training uses the DC085 dataset, which is a publicly available dataset containing complete plant genome DNA sequences and their corresponding GFF3 gene structure annotation files, and can be obtained through public database searches.

[0081] In addition to the DC085 training data, this method was validated and tested on the following specific reference genome versions: *Amborella trichopoda* (NCBI RefSeq accession number: GCF_001433935.1); *Arabidopsis thaliana* (NCBI RefSeq accession number: GCF_000001735.4); *Oryza sativa* (NCBI RefSeq accession number: GCF_000471905.2); and *Gossypium arboreum* (reference-level genome data published in Figshare public data entries).

[0082] All of the above data include chromosome-level DNA sequences (FASTA format) and standard GFF3 gene structure annotation files, which are used for splice site annotation construction and model evaluation.

[0083] 2. Construction of splice site annotations Based on the obtained GFF3 annotation file, exon coordinate information was parsed, and intron regions were deduced according to transcript structure. Intron start boundaries were marked as donor sites, intron termination boundaries as acceptor sites, and the remaining bases were uniformly marked as non-splicing sites, thus constructing a three-category supervisory tag system. This process was automated, ensuring that all tags originated from standard gene structure annotations.

[0084] 3. Sliding window sample construction A sliding window of 401 bp was used to segment the genome sequence, with a window step size of 1 bp. The central base of each window was used as the prediction target. Three-class samples were generated based on the label of the central base, and all samples were divided into training and test sets for model training and generalization performance evaluation.

[0085] 4. Dual-channel input encoding Two sets of input representations are generated for the same window sequence. The first set is used for the teacher model, where the DNA sequence is k-mer segmented at k=3 and mapped to an integer token ID sequence. The second set is used for the student model, where each base is converted into a one-hot vector of length 4, forming an L×4 tensor representation. During the training phase, the central bases in the student model input are randomly masked to enhance the model's adaptability to cross-species differences.

[0086] 5. Teacher Model Construction The teacher model uses a pre-trained DNABERT-2 model as the encoder, with a parameter size of approximately 117M. The input token ID sequence is processed through a Transformer encoding layer to obtain contextual feature representations. Mean aggregation is performed on effective positions to form a sequence-level embedding. This embedding is input to a linear classification layer to generate three-class logits, and a smooth probability distribution is obtained through a softmax operation with a temperature parameter τ. During training, the teacher model maintains its parameters frozen and does not participate in gradient updates; it is only used to provide distillation supervision signals.

[0087] 6. Student Model Construction The student model takes a one-hot tensor as input and first extracts local sequence features through a one-dimensional convolutional layer, mapping them to a high-dimensional representation space. It then enters an encoder structure consisting of 16 dilated convolutional residual modules, with dilation rates set in stages of 1, 4, 10, and 20 to expand the receptive field and model sequence dependencies at different scales. Residual connections are used to enhance gradient propagation stability. The encoder output features are projected and then enter an attention module, introducing learnable positional embeddings and modeling long-distance modulation relationships through a multi-head attention mechanism. The attention matrix is ​​calculated using Sinkhorn normalization. Finally, a three-class logits are output through a classification head, and the predicted probabilities of the student model are obtained through softmax.

[0088] 7. Joint Loss Training During training, a joint loss function is constructed, consisting of cross-entropy loss and distilled KL divergence loss. Cross-entropy loss constrains the student model's output to match the true labels, while distilled loss constrains the student model's output distribution to approximate the teacher model's soft label distribution. The total loss is a weighted sum of the two losses. Backpropagation updates are performed only on the student model parameters, while the teacher model remains frozen.

[0089] 8. Deployment of Reasoning After training, the teacher model is removed, and only the student model is retained for prediction. A sliding window scan is performed on the target species' genome, and the probability value of each central base belonging to Acceptor, Donor, or Neither is output to achieve genome-wide splice site prediction.

[0090] Example 4

[0091] Example 4 of this invention is a second application example of the plant alternative splicing prediction method based on large-scale model distillation and evidence constraints provided by this invention. This second application example is an embodiment of a splicing regulatory motif identification and population validation method based on attribution analysis, including: 1. Data Source This embodiment uses the reference genome of *Gossypium arboreum* as the analysis object, and its genomic data is derived from publicly available reference-level data resources published by Figshare. Population validation data comes from the NCBI SRABioProject: PRJNA626178, which contains whole-genome resequencing data from 216 *Gossypium arboreum* accessions. J24 was selected as the reference material during the analysis, and J103 and J106 were selected as comparison materials carrying candidate mutations. RNA-seq data also came from the corresponding sample data in this publicly available project.

[0092] 2. Candidate Window Filtering The student model trained in Example 1 was used to perform a 401 bp sliding window scan of the Asian cotton genome with a step size of 1 bp. Windows predicted as donors or recipients with confidence levels exceeding a threshold were selected for attribution analysis.

[0093] 3. DeepLIFT Attribution Calculation DeepLIFT attribution analysis was performed on the candidate windows. To reduce the impact of background bias, a first-order Markov model based on the Asian cotton genome was used to generate reference sequences. Three independent baseline sequences were generated for each window, and the attribution results were calculated separately. Finally, the contribution scores were averaged. Based on a window of optional length, splicing-specific base contribution interpretation can be supported for the whole genome.

[0094] 4. Motif Extraction and Clustering Continuous segments with absolute values ​​exceeding a set threshold in the contribution vector were screened, and segments with a length of at least 5 bp were retained as candidate splicing regulatory motifs. All candidate motifs were deduplicated, and their frequency in the genome was calculated. Motifs were classified into different categories based on whether they contained typical AG / GT splicing core signals, and similarity clustering analysis was performed.

[0095] 5. Population variation constraint verification Candidate high-contribution motif regions were cross-aligned with SNP coordinates in Asian cotton population resequencing data. The SNP density within the candidate regions was calculated and compared with length-matched random background regions. A bootstrap sampling method was used to construct the background distribution, and the relative mutation rate was calculated to assess whether high-contribution regions exhibited variation-constrained characteristics.

[0096] 6. Transcriptome validation RNA-seq data were compared between materials with candidate motif mutations and reference materials. Samples from the 4-day fiber development stage were uniformly selected, and differential splicing analysis was used to identify events such as exon skipping or intron retention, thereby verifying whether mutations caused changes in splicing patterns.

[0097] Example 5

[0098] Example 5 provided by this invention is the third application example of the plant alternative splicing prediction method based on large-scale model distillation and evidence constraints provided by this invention. This third application example is an embodiment of the training and deployment of a plant splicing site prediction method based on a knowledge distillation mechanism, including: 1. Sources of cross-species genomic data This embodiment uses the following specific reference genome versions: Asian cotton (Gossypium arboreum): reference-level genome data is published in the public data entry (Figshare), containing complete chromosome-level DNA sequences and corresponding GFF3 gene structure annotation files; *Cinnamomum camphora*: NCBI RefSeq GCF_001433935.1; *Arabidopsis thaliana*: NCBI RefSeq GCF_000001735.4; Rice: NCBI RefSeq GCF_000471905.2. All the above genome data are from the official NCBI RefSeq database and include the corresponding GFF3 annotation files.

[0099] 2. Unified sequence mining process The genomes of each species were scanned using the same 401 bp sliding window and DeepLIFT attribution rules as in the second application example, and candidate motifs were extracted using the same threshold and minimum length criteria to ensure comparability of results between different species.

[0100] 3. Verification using known splicing examples In Arabidopsis thaliana, reported FLM splicing regulatory genes were selected as control cases. Their gene regions were input into the trained student model for prediction and attribution analysis to verify whether the model can correctly identify known regulatory regions and their splicing-related signals.

[0101] This invention provides a large-scale model distillation and evidence-constrained plant alternative splicing prediction method and system, which significantly improves the accuracy of cross-species splicing site prediction and exhibits stronger comprehensive performance in plant splicing prediction tasks. It demonstrates good stability and generalization ability, maintaining stable prediction performance even without retraining for the target species, indicating that the method can effectively address differences in data distribution among species. It improves the interpretability of prediction results, constructing a splicing regulatory signal identification and functional interpretation system based on base-level contribution analysis. It achieves lightweight and efficient deployment and enhances cross-species transferability; in the inference stage, only a student model is needed to complete base-by-base prediction, which can be directly applied to unseen genomes without retraining or species-specific fine-tuning. This reduces computational resource consumption and verifies the cross-species representation transferability provided by the knowledge distillation strategy, supporting rapid scanning at the whole-genome scale. It supports downstream analysis of mutation effects; a generative large-scale language model integrating subtype sequences, structures, functions, and trait evidence obtained from various databases and tools links sequence variations with splicing results and structural consequences, thus moving beyond simple prediction towards evidence-based reasoning and enhancing the application value of the method in practical breeding and functional research.

[0102] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0103] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0108] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting plant alternative splicing based on large-scale model distillation and evidence constraints, characterized in that, The prediction method includes: Step 1: Construct a dual-model distillation structure that includes a teacher model and a student model; Step 2, obtaining input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data; Step 3: Perform K-mer Tokenization and One-Hot Encoding in parallel on each DNA sequence to obtain Token IDs and B×L×4 tensors respectively; the Token IDs are the input of the teacher model, and the tensors are the input of the student model. Step 4: Construct a joint loss based on the outputs of the teacher model and the student model and perform joint optimization to realize the transfer of semantic knowledge from the teacher model to the student model; Step 5: Use the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determine the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model. Step 6: Calculate the base-level attribution score using the trained student model, and identify candidate sequence motifs based on the base-level attribution score; perform cluster analysis, enrichment analysis, and computer simulation mutation analysis on the candidate sequence motifs to characterize the splicing-related regulatory patterns. Step 7: Based on the integration of multiple transcript splicing events and recessive mutations, the effects on amino acid sequence, protein structure, and functional phenotypic variations are explained, a biological causal chain is constructed, and a natural language interpretation is given based on a generative large language model. The student model constructed in step 1 includes: an input layer, a local feature extraction layer, a classification head, and a Softmax output module; The input layer receives the input tensors and retains fine-grained information at the base level; The local feature extraction layer extracts local sequence features of the tensor through one-dimensional convolution; The classification head is used to input the encoded features into the classification layer to generate classification logits; The Softmax output module is used to perform Softmax calculation on the classification logits to obtain the student model prediction probability. .

2. The prediction method according to claim 1, characterized in that, The teacher model constructed in step 1 includes: a sequence semantic encoder, an average pooling aggregation module, a linear classification head, and a probability output module with temperature regulation; the sequence semantic encoder is a pre-trained large genome model based on the Transformer architecture; the Transformer layer of the sequence semantic encoder is kept frozen during the training phase; The Transformer layer of the sequence semantic encoder generates a contextual semantic representation based on the input Token IDs; The average pooling aggregation module performs Mean Pooling operation on the context semantic representation output by the Transformer layer, aggregating the base-level representation into sequence-level embedding; The sequence-level embedding is input into the logistic regression head of the linear classification head to generate classification logits; The probability output module with temperature regulation is based on the classification logits via a temperature parameter. τ The Softmax layer generates a smooth soft probability distribution. .

3. The prediction method according to claim 1, characterized in that, The classification categories of the splice site classification and labeling data in step 2 include: recipient, donor, and neither; Step 2 further includes: dividing the input data into a training set and a test set.

4. The prediction method according to claim 1, characterized in that, The joint optimization process in step 4 includes: Step 401: Calculate the output of the student model. Soft labels of the teacher model The KL divergence between them results in distillation loss. ; Step 402: Calculate the cross-entropy loss between the classification logits of the student model and the true label. ; Step 403, calculate the total loss function as follows: , The weights are set; the teacher model and the student model are jointly optimized based on the total loss function to realize the transfer of teacher semantic knowledge to the student model; Step 404: During the training phase, only the parameters of the student model are updated, while the teacher model remains frozen; and during the training phase of the student model, the central bases are randomly masked.

5. The prediction method according to claim 1, characterized in that, Step 6 includes: A splicing center window is constructed, and the base-by-base contribution distribution within the window is obtained; a reference sequence is constructed based on the original plant genomic DNA sequence, and the base-by-base contribution degree at each position within the window is calculated based on the base-by-base contribution distribution and the reference sequence.

6. The prediction method according to claim 5, characterized in that, The process of calculating the base-by-base contribution at each position within the window includes: Step 601: Perform one-hot encoding on the input data in the window and record it as the actual input. L is the length of the splice center window; the target category will be... The output logit is denoted as By analyzing each position within the window The four-channel base contribution is calculated to obtain the base-by-base contribution distribution within the window range; Step 602: Fit a first-order Markov model to the original plant genomic DNA sequence set to generate a reference sequence; generate a reference sequence for each window. The reference sequence is obtained by one-hot encoding the reference sequence. n is a positive integer; Step 603: Under the gradient-based feature attribution algorithm rules, the true input is... With the Reference sequence By comparison, the channel-level attribution tensor is obtained. ;right The attribution tensor of the reference sequence is averaged to obtain ; using the actual input Mapping the hypothetical contributions to the actual contributions yields the base-by-base contribution scores for each position. .

7. The prediction method according to claim 6, characterized in that, Step 6 also includes: The contribution vector based on each position The absolute value threshold is used to screen continuous high-contribution segments, and the sequences corresponding to the continuous high-contribution segments are extracted as candidate splicing regulation motifs; the motifs are deredundant and a motif-original occurrence position mapping table is established; according to whether they conform to the typical AG / GT splicing core mode, they are divided into two categories: standard type and special type, and the distance matrix is ​​calculated and hierarchical clustering is performed to form a motif family structure to distinguish the core signal from atypical auxiliary regulation motifs.

8. The prediction method according to claim 7, characterized in that, Step 6 also includes: Step 8: Using the same window length, reference sequence generation strategy, and contribution calculation formula for different plant species, obtain the motif sets and family structures; by comparing the composition, branching and aggregation of Canonical / Special motif families in different species, evaluate whether the splicing regulation signals identified by the model maintain a stable pattern in cross-species scenarios.

9. A plant alternative splicing prediction system based on large-scale model distillation and evidence constraints, characterized in that, The prediction system includes: a teacher model, a student model, a distillation training module, and a splice site prediction module; The input data includes: plant genomic DNA sequences and their corresponding splice site classification and annotation data; two parallel encoding methods, K-mer Tokenization and One-Hot Encoding, are performed on each DNA sequence to obtain TokenIDs and B×L×4 tensors, respectively; the Token IDs are the input of the teacher model, and the tensors are the input of the student model; The distillation training module constructs a joint loss based on the outputs of the teacher model and the student model and performs joint optimization to realize the transfer of semantic knowledge from the teacher model to the student model; The splice site prediction module uses the trained student model to perform a sliding window scan on the genomic DNA sequence of the plant to be tested, and determines the splice site of the genomic DNA sequence of the plant to be tested based on the splice site category corresponding to the maximum probability output by the student model. The student model includes: an input layer, a local feature extraction layer, a classification head, and a Softmax output module; The input layer receives the input tensors and retains fine-grained information at the base level; The local feature extraction layer extracts local sequence features of the tensor through one-dimensional convolution; The classification head is used to input the encoded features into the classification layer to generate classification logits; The Softmax output module is used to perform Softmax calculation on the classification logits to obtain the student model prediction probability. .

Images