A method for predicting the trans-splicing activity of peptide fragmentation and its application

Abstract

This invention discloses a method and its application for predicting the trans-splicing activity of fragmented inpeptides. The method takes the amino acid sequence of the fragmented inpeptide as input, extracts 1280-dimensional residue-level sequence embedding features using the ESM-2 protein language model, and uses AlphaFold2 for high-precision three-dimensional structure prediction to obtain the spatial coordinates of the residues. A graph structure is constructed with residues as nodes and residue pairs with a Cα atomic distance of less than or equal to 8 Å as edges. Node features include sequence embedding and chain affiliation information, while edge features include Cα distance and cross-chain connection identifiers. The graph structure is input into the DeepIntein deep learning model based on a graph attention network, and the splicing activity prediction probability is output after multi-layer graph attention convolution, global pooling, and fully connected layers. The prediction throughput of this invention is more than three orders of magnitude higher than that of the traditional Western blotting method, breaking through the limitations of homology sequence exploration space.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and protein engineering technology, specifically relating to a method for predicting the trans-splicing activity of fractured integuments based on graph attention networks (GAT), and the application of this method in the discovery of novel active fractured integuments. Background Technology

[0002] Cleavage inteins are key functional elements mediating protein transsplicing and have significant applications in protein engineering, synthetic biology, and targeted drug delivery. Efficiently mining cleavage inteins with both high splicing activity and sequence novelty from natural sequence resources is one of the core fundamental tasks driving the development of this field.

[0003] Currently, the identification of trans-splicing activity of fragmented inteins mainly relies on Western blotting (WB) experiments. Researchers typically select the fragmentation sites of natural inteins based on expert experience, construct co-expression systems, and identify splice products using specific antibodies. While this method has high detection specificity, it suffers from significant throughput limitations: its sample processing throughput is usually less than 20 cases per day, and a single detection cycle can take 24 to 48 hours. Due to these limitations, less than 2% of natural inteins have undergone trans-splicing activity identification, severely restricting the large-scale discovery of high-performance functional components.

[0004] To alleviate throughput limitations, researchers have recently attempted to introduce reporter gene-based pre-screening strategies: first, candidate sequences highly homologous to known active sequences are screened through multiple sequence alignment; then, a reporter gene system is used for initial activity screening, with only positive candidates undergoing Western blotting (WB) verification. However, this strategy is essentially a "blind screening" model relying on sequence homology, lacking a systematic analysis of the "sequence-structure-function" relationship of inteins. Its fundamental limitation is that it can only determine whether a specific sequence is active, but cannot reveal the key sequences and structural features that determine splicing activity. Due to the extremely high sequence diversity of natural inteins, many potential functional sequences with low similarity to known sequences but potentially excellent performance are systematically excluded from the screening scope because they cannot meet the homology threshold. Therefore, there is an urgent need to develop a high-throughput prediction and mining method that can overcome homology sequence limitations and systematically analyze the trans-splicing activity of inteins based on sequence and structural features. Summary of the Invention

[0005] This invention provides a method, a method for discovering trans-splicing activity of fragmented integrins, and a related apparatus, aiming to solve the technical problems of low throughput, reliance on sequence homology comparison, and difficulty in discovering novel non-homologous functional elements in existing screening methods. This invention constructs a graph representation of fragmented integrins by fusing deep evolutionary semantic information of protein sequences with three-dimensional spatial structural constraints, and adaptively captures key residue interactions that determine trans-splicing activity using a graph attention network, thereby achieving high-precision, high-throughput prediction of the trans-splicing activity of fragmented integrins.

[0006] Building upon this, the present invention also provides a novel method for discovering active fragmented integrins. This method combines activity prediction with a clustering-based hierarchical screening strategy based on structural similarity, overcoming the limitations of traditional homologous sequences. It can efficiently discover novel active fragmented integrin elements with diverse folding types from a structurally diverse candidate sequence space.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the trans-splicing activity of fragmented integuments. The core of this method is to fuse the sequence information and three-dimensional structural information of the fragmented integument into a graph structure representation, and to capture the spatial interaction relationships between residues through a graph attention network, so as to predict the trans-splicing activity probability of the fragmented integument in an end-to-end manner.

[0008] The prediction method includes the following steps: (1) Obtain the amino acid sequence of the fragmented inteptide to be predicted, wherein the fragmented intepteptide comprises an N-terminal fragment (IntN) and a C-terminal fragment (IntC). (2) Based on the amino acid sequence, extract sequence features and spatial structure features to construct graph structure data with amino acid residues as nodes and spatial contact relationships between residues as edges; wherein, the features of the nodes are constructed based at least on the sequence features and fragment chain information to which the amino acid residues belong, and the connection relationship of the edges is determined based on the spatial distance threshold between pairs of amino acid residues; (3) Input the graph structure data into the trained graph attention network model to obtain the prediction results of the trans-splicing activity of the fragmented peptide.

[0009] Further, step (2) of extracting the sequence features includes: using a language model pre-trained based on protein sequence data to extract features from the amino acid sequence; Preferably, the language model is the ESM-2 model, and the extracted feature vector has a dimension of 1280.

[0010] In the sequence feature extraction stage, the ESM-2 protein language model with 650M parameters is preferred. This configuration extracts a 1280-dimensional embedding vector that effectively captures the evolutionary conservation information of amino acid residues and long-range contextual semantic relationships. Compared to smaller parameter versions (e.g., 150M), the 650M version significantly improves activity prediction accuracy; compared to larger parameter versions (e.g., 3B), this configuration offers better computational efficiency while maintaining performance. As an alternative, the traditional one-hot encoding method can be used to characterize residue types, but ablation experiments show that feature representations based on one-hot encoding have significantly lower model prediction performance than ESM-2 embedding features due to the lack of contextual semantic information. Regarding sequence input, instead of extracting features independently from the N-terminal and C-terminal fragments, it is preferable to treat the fractured integument as a continuous sequence and input it into the model for feature encoding. This method more effectively preserves the co-evolutionary information between the two fragments, and experiments have verified its superior prediction performance.

[0011] Further, the extraction of the spatial structure features in step (2) includes: predicting the Cα atom spatial coordinates of each residue in the fractured peptide using a protein three-dimensional structure prediction tool; Preferably, the prediction tool is AlphaFold2.

[0012] Further, the connection relationship of the edge in step (2) is determined according to the following rules: calculate the Euclidean distance between any two Cα residue atoms, and if the Euclidean distance is less than or equal to a preset threshold, then establish an edge between the corresponding two nodes; The preset threshold value ranges from 6 angstroms to 10 angstroms, with a preferred value of 8 angstroms.

[0013] In the graph construction process, the preferred spatial distance threshold for determining whether a connection is established between nodes is 8 Å. This invention evaluated the model performance under three threshold settings: 6 Å, 8 Å, and 10 Å using an ablation experiment system. The results show that a 6 Å threshold leads to an overly sparse graph structure, easily losing some functionally relevant long-range spatial proximity information; a 10 Å threshold significantly increases graph density, introducing too many weakly correlated or even irrelevant noise connections; and an 8 Å threshold achieves the optimal balance between capturing direct contact between residues and effective spatial constraints, thus enabling the model to achieve the best predictive performance.

[0014] Furthermore, the edge features in the graph structure data include: the Cα atomic distance value between connected residues, and a cross-chain connection identifier used to identify whether the connected residues belong to the same fragment chain.

[0015] In the construction of edge features, this invention preferably introduces a cross-chain connection marker to distinguish between intra-chain interactions within the same peptide segment and inter-chain interface interactions spanning the N-terminus (IntN) and C-terminus (IntC). The trans-splicing activity of fragmented peptides is highly dependent on the spatial complementarity and interaction strength of the interfaces between the two segments. By introducing this marker, the graph attention network can explicitly perceive and learn the spatial relationship patterns of cross-chain interfaces. Ablation experiments show that removing the cross-chain marker feature reduces the accuracy of the model by approximately 4 to 6 percentage points when predicting candidate sequences with dense interface contacts, validating the effectiveness of this feature in improving prediction accuracy.

[0016] Furthermore, the features of the node in step (2) are formed by concatenating a first feature vector and a second feature vector, wherein the first feature vector is the sequence feature and the second feature vector is the chain attribution code used to distinguish between N-end segments and C-end segments.

[0017] Furthermore, the graph attention network model described in step (3) includes: (a) An input layer for receiving the graph structure data; (b) At least one graph attention convolutional layer, used to calculate the importance weights of neighboring nodes through an attention mechanism, and to aggregate the features of neighboring nodes based on the weights to update the node representation of the current node; (c) Global pooling layer, used to integrate the node representations of all nodes in the graph into a graph-level representation vector; (d) Output layer, used to output splice activity prediction probabilities via an activation function based on the graph-level representation vector.

[0018] At the model architecture level, this invention preferably employs a multi-head graph attention network (MNB) instead of a traditional graph convolutional network (GCN). This network stacks multiple layers of graph attention convolutional layers, dynamically calculating the importance weight coefficients of neighboring nodes within each layer using a self-attention mechanism, and then weighting and aggregating neighborhood information to update the node representation. After multiple updates, global average pooling or global attention pooling is used to aggregate all node features into a graph-level vector representing the entire peptide molecule, which is then output as an activity prediction probability via a fully connected layer and a sigmoid activation function. Compared to the aggregation method of graph convolutional networks, which treats all neighboring nodes equally, the graph attention mechanism can adaptively assign higher weights to key spatial neighbors, thereby focusing on residue interactions that have a decisive impact on splicing activity. Ablation experiments show that the AUC of the graph attention network on the test set is improved by approximately 3% to 5% compared to the graph convolutional network.

[0019] Furthermore, the graph attention network model is obtained through supervised training using a dataset containing both active positive and active negative samples, and the loss function used in the training process is the binary cross-entropy loss function.

[0020] In one specific embodiment, the prediction method includes the following steps: Step 1: Obtain the amino acid sequence of the fragmented inteptide to be predicted, wherein the fragmented intepteptide is composed of IntN and IntC; Step 2: Encode each amino acid residue in the sequence using the ESM-2 protein language model (650M parameters) to obtain a 1280-dimensional sequence embedding feature vector, capturing the evolutionary information and contextual semantics of the residues; Step 3: Use AlphaFold2 to perform high-precision three-dimensional structure prediction on the fractured peptide and obtain the three-dimensional spatial coordinates of the Cα atom of each residue; Step 4: Construct a graph structure using amino acid residues as nodes. Node features are the splicing of ESM-2 sequence embedding (1280-dimensional) and chain affiliation encoding; if the Euclidean distance between any two residue Cα atoms does not exceed 8 Å, an edge connection is established, and the edge features include the distance between Cα atoms and the cross-chain connection identifier; Step 5: Input the graph structure into the DeepIntein model. After multi-layer graph attention convolution to aggregate neighbor information and global pooling to generate graph-level representation vectors, the output layer outputs the splicing activity prediction probability (between 0 and 1) through the sigmoid function.

[0021] The core principle of the above technical solution lies in abstracting the fragmented integrities into a graph network composed of amino acid nodes and spatially adjacent edges. The deep evolutionary and semantic information of the sequence is captured using the protein language model ESM-2, and the true three-dimensional spatial constraints are obtained using the structure prediction tool AlphaFold2. The graph attention network can adaptively learn the importance weights of different spatially adjacent residues for splicing function, and pays particular attention to the cross-chain interactions between N-terminal and C-terminal fragments, thereby achieving accurate judgment of splicing activity.

[0022] Secondly, the present invention provides a novel method for mining active fragmented peptides, comprising the following steps: (1) Obtain a pool of candidate intima-peptide sequences; (2) Using the method for predicting the trans-splicing activity of the broken inteptide described in any of the above, calculate the predicted probability of trans-splicing activity of each candidate sequence in the candidate inteptide sequence pool. (3) Perform structure prediction on each candidate sequence in the candidate peptide sequence pool, and perform cluster analysis based on the similarity between the predicted structures to obtain multiple structural clusters; (4) Within each of the structural clusters, at least one representative sequence is selected based on the predicted probability of trans-splicing activity, and the selected representative sequences are combined to form a set of candidate active integrins.

[0023] When using the aforementioned prediction model for large-scale discovery of novel active integuments, a hierarchical screening strategy based on structure clustering is preferred over global single-score ranking. Specifically, firstly, tools such as Foldseek are used to perform structural alignment and clustering of all sequences in the candidate integument pool, grouping integuments with similar three-dimensional folding patterns into the same structural cluster. Subsequently, within each independent cluster, sequences are ranked according to predicted activity scores, and only the representative sequences with the highest scores are selected for inclusion in the final candidate set. Finally, the representative sequences selected across clusters are globally ranked according to activity scores, and the top-ranked sequences are selected for experimental validation. This strategy effectively avoids the problem of excessive concentration of screening results in a single structural family due to model bias, ensuring a high degree of diversity in sequence characteristics and three-dimensional structural folding types among the discovered candidate sequences, thereby maximizing the probability of discovering active elements with novel functional properties.

[0024] Thirdly, the present invention provides a device for predicting the trans-splicing activity of fragmented inteptides, comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the prediction method described above.

[0025] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the prediction method described above, or implements the steps of the mining method described above.

[0026] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: (1) The predicted throughput and screening efficiency are significantly improved. This invention utilizes a deep learning model to rapidly infer the activity of fragmented integrins, with single sequence prediction taking less than a second. It also supports batch parallel processing of large-scale sequences, achieving a daily prediction throughput of tens of thousands of sequences. Compared to traditional characterization methods relying on Western blotting (WB) experiments (which typically process fewer than 20 samples per day), this invention improves throughput by more than three orders of magnitude, effectively overcoming the efficiency bottleneck in the large-scale mining of functional integrins.

[0027] (2) Break through the restriction of sequence homology and expand the space for exploring functional elements. This invention does not rely on homology alignment between candidate sequences and known active sequences. Instead, it directly learns the intrinsic mapping relationship between the "sequence-structure-function" relationship of integrins through graph attention networks. Therefore, for candidate sequences with low similarity to known active integrins and difficult to discover using traditional homology alignment methods, this invention can still provide high-confidence activity prediction results. This characteristic effectively expands the explorable sequence space of functional integrin elements, helping to discover a large number of "dark matter" integrins that were previously systematically missed due to significant sequence differences.

[0028] (3) Integrating sequence evolution information with three-dimensional structural constraints to improve prediction accuracy This invention incorporates sequence evolution embedding information extracted from protein language models (such as ESM-2) and residue-level three-dimensional spatial coordinate information provided by high-precision structure prediction tools (such as AlphaFold2) when constructing predictive features. By constructing a sequence-structure joint graph representation, the model can comprehensively utilize the contextual semantic associations and real spatial proximity relationships of amino acid residues, exhibiting higher prediction accuracy on independent test sets compared to traditional methods that rely solely on single sequence or single structural features.

[0029] (4) Explicitly model the cross-chain interface to enhance the perception of key areas of the splicing function. In constructing the graph structure, this invention explicitly distinguishes the interfacial interactions between the N-terminal and C-terminal fragments of the fractured inteptide by using cross-chain connection markers in the edge features. This design allows the graph attention network to focus on the spatial complementarity and interaction patterns at the interface between the two fragments, which is the key structural basis for determining whether the fractured inteptide can successfully undergo trans-splicing. Ablation experiments show that ( Figure 2 After introducing cross-chain identifier features, the model's prediction accuracy for candidate sequences with dense interface contact improved by about 2 to 3 percentage points, demonstrating the method's advantage in perceiving the core determinants of splicing activity.

[0030] (5) Verification of actual mining results: novel active integrins with diverse sequences and structures were obtained. Based on the method of this invention, virtual screening and experimental verification of fragmented integuments successfully identified five novel fragmented integuments with splicing activity from a pool of candidate sequences with diverse structural clusters. The maximum amino acid sequence identity among these active integuments ranged from 30% to 46%, and their sequence similarity to known highly active integuments (such as Npu DnaE) was relatively low. Furthermore, these five integuments belonged to multiple different three-dimensional structural folding types, with pairwise TM-scores ranging from 0.2 to 0.7. Given the significant differences between these sequences and known active sequences at both the sequence and structural levels, they would easily be filtered out in traditional homology-based screening processes. The results of this embodiment verify the effectiveness of the method of this invention in overcoming homology sequence limitations and expanding the boundaries of sequence and structural diversity of functional integument elements, demonstrating its application value in discovering novel functional molecules. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the method for predicting the trans-splicing activity of fragmented peptides in this embodiment of the invention.

[0032] Figure 2 This is a comparative schematic diagram of ablation experiment results in the embodiments of the present invention, showing the comparison of the predictive performance of the model under different feature combinations, different spatial distance thresholds, different graph neural network types, and with or without cross-chain edge features.

[0033] Figure 3 This is a schematic flowchart of the novel active fragmentation peptide mining method in an embodiment of the present invention.

[0034] Figure 4 This is a graph showing the activity verification results of the novel active fragmented in-cell peptide obtained through experiments in this embodiment of the invention; the upper part is a reporter gene fluorescence signal intensity graph, where the scatter dots represent different samples; the lower part is the corresponding protein immunoblot band graph.

[0035] Figure 5 These are sequence diversity and structural diversity analysis diagrams of the active fragmented peptides obtained in this embodiment of the invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0037] Example 1: Training and Evaluation of a Predictive Model for the Activity of Trans-splicing of Inteptides This embodiment details the construction, training, and performance evaluation process of a graph attention network model (named DeepIntein) used for predicting the splicing activity of fragmented inteptides. Figure 1 ).

[0038] 1.1 Construction of the training dataset The system collects and integrates training data on the activity of splicing integrins. The training data includes positive samples experimentally verified to have splicing activity, and negative samples experimentally verified to have no splicing activity or to have lost activity. Experimental verification methods include Western blotting (WB). Positive samples include naturally active splicing integrin sequences, and negative samples include naturally inactive splicing integrin sequences and inactive splicing integrin variant sequences that have been artificially mutated or truncated.

[0039] For each collected sequence, its source information, complete amino acid sequence, experimentally verified activity status, and corresponding literature source were labeled to construct a labeled training dataset. Among them, the activity status was labeled with binary values, with "1" for positive and "0" for negative.

[0040] To reduce data redundancy and prevent model overfitting, the dataset undergoes sequence de-redundancy processing. Redundant sequences with a similarity exceeding a preset sequence identity threshold are removed. In this embodiment, the sequence identity threshold is set to 80%. After de-redundancy processing, the dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0041] 1.2 Sequence Feature and Structural Feature Extraction (1) Sequence feature extraction: A pre-trained protein language model ESM-2 with 650M parameters was used to perform forward inference on each fragment containing peptide sequences (continuous sequences including N-terminal and C-terminal fragments) to extract the residue-level embedding vectors output by the last Transformer layer of the model. Each amino acid residue obtained a feature vector representation with a dimension of 1280. This extraction process was performed in batch mode on a graphics processing unit (GPU) to improve the efficiency of feature extraction.

[0042] (2) Structural feature extraction: Using the protein three-dimensional structure prediction tool AlphaFold2, the cofold structure of each fragment containing peptide sequences (input as continuous sequences of N-terminal and C-terminal fragments) was predicted. The model with the highest confidence in the prediction results (model_rank_1) was selected, and the three-dimensional spatial coordinates (x, y, z) of each amino acid residue Cα atom were extracted from its output PDB format coordinate file.

[0043] 1.3 Construction of Graph Structure For each fragment containing a peptide sequence, construct a graph structure according to the following rules. ,in Represents a set of nodes. Representing the set of edges: (1) Node definition: Each amino acid residue in the sequence is defined as a node in the graph, and the node set is... ,in This represents the total number of residues in the N-terminal and C-terminal segments.

[0044] (2) Node characteristics: Each node eigenvectors It is composed of two parts, namely .in, This is the 1280-dimensional sequence embedding vector of the residue extracted by the ESM-2 model in step 1.2; This encodes residue chain affiliation; when a residue belongs to the N-terminal segment, it is encoded as... When the residue belongs to the C-terminal segment, it encodes as The concatenated node feature vector has a dimension of 1282.

[0045] (3) Edge definition: Calculate any two residues and Euclidean distance between Cα atoms If and only if At the node and Establish an undirected edge connecting them, that is .

[0046] (4) Edge characteristics: Each edge corresponding feature vector .in, This represents the distance between two Cα residues. For cross-chain linkage identifiers, if residues With residues Belonging to N-terminal segments and C-terminal segments respectively, If the two residues belong to the same segment, then .

[0047] 1.4 Construction and Training of Graph Attention Network Model The DeepIntein model is implemented based on the PyTorch Geometric deep learning framework. The model takes the graph structure data constructed in step 1.3 as input, sequentially passing it through graph attention convolutional layers to extract residue-level spatial features, global pooling layers to aggregate it into chain-level representations, and the output layer to map it as a scalar. Finally, it is transformed into splicing activity probability values ​​between 0 and 1 using the Sigmoid function. The specific architecture and data transformation process are as follows: (1) Input layer: Receives the graph structure data constructed in step 1.3 It specifically includes three parts: (1.1) Node feature matrix :one A real matrix with 1282 rows and 1282 columns, where The total number of residues is represented by a 1282-dimensional feature vector for each residue in each row. (1.2) Edge feature matrix :one A real matrix with 2 rows and 2 columns, where The total number of edges is represented by a 2D feature (distance and cross-chain identifier) ​​for each edge. (1.3) Adjacency relationship :one OK, A real square matrix of columns, used to represent the connection relationships between nodes. If the nodes and If there is an edge, the corresponding position will have a non-zero value.

[0048] This layer only encapsulates and validates the data format, without changing the data dimensions.

[0049] (2) Graph Attention Convolutional Layer: Three stacked GATConv graph attention convolutional layers, each with a hidden feature dimension of 256, employing an 8-head attention mechanism. Each graph convolution is followed by a batch normalization layer and a ReLU non-linear activation function. This module progressively abstracts the input 1282-dimensional node features into 256-dimensional high-order structural features, while utilizing edge features. The calculation of attention weights is used to encode the spatial interactions between residues.

[0050] (3) Global Pooling Layer: Global mean pooling is used to average the updated feature vectors (256 dimensions) of all nodes across the node dimension, and aggregate them into a single full-map level representation vector with a dimension of 256. This module completes the dimensional transformation from "residue-level information" to "protein chain-level information", enabling the model to perform activity discrimination on the basis of the entire sequence.

[0051] (4) Output layer: It contains two fully connected network layers. The first layer maps 256-dimensional features to 64-dimensional features, and the second layer maps 64-dimensional features to 1-dimensional features. Finally, the output layer is activated by the Sigmoid activation function and outputs a splicing activity prediction probability between 0 and 1. This module compresses the graph-level representation vector into the final prediction score, realizing a quantitative evaluation of the activity of the input sequence.

[0052] The model training uses the following hyperparameter configuration: the optimizer is Adam, and the initial learning rate is set to 1×10. -4Batch size set to 32; maximum training epochs set to 200; early stopping strategy adopted, training terminated when the validation set AUC metric showed no improvement for 20 consecutive training epochs; weight decay coefficient set to 1×10⁻⁶. -5 The loss function uses weighted binary cross-entropy loss, with the weights of positive and negative samples adjusted according to the ratio of positive to negative samples in the training set to alleviate the class imbalance problem.

[0053] 1.5 Ablation Experiments and Model Performance Validation To systematically evaluate the contribution of different feature combinations and model components to prediction performance, multiple ablation experiments were conducted on the validation set. The experimental results are as follows: Figure 2 As shown, it can be seen that: Comparing experimental group 1 (sequence only (one-hot)) and experimental group 2 (sequence only (ESM-2)), it is shown that the sequence embedding features extracted based on the pre-trained protein language model (ESM-2) can capture the evolutionary semantic information of amino acid residues more effectively than the traditional one-hot encoding method, and make a significant contribution to improving the model's prediction performance. Comparing experimental groups 2 (sequence only (ESM-2)) and 3 (structure only) with experimental group 4 (sequence + structure (6Å)), it is shown that fusing sequence evolution information with three-dimensional structural spatial constraints can further improve prediction performance; Comparing experimental groups 4 (sequence + structure (6 Å)), 5 (sequence + structure (8 Å)), and 6 (sequence + structure (10 Å)), it is shown that 8 Å is the optimal spatial distance threshold, which can achieve the best balance between preserving key contacts and suppressing noise. Comparing experimental group 7 (GAT) and experimental group 8 (GCN), it is shown that the graph attention mechanism is more effective than ordinary graph convolutional networks in focusing on the interactions of key residues that contribute significantly to activity prediction. Comparing experimental group 5 (sequence + structure (8Å)) with experimental group 7 (sequence + structure (8Å) + cross-chain edge), it is evident that the introduction of cross-chain edge features can further enhance the model's ability to perceive the interaction between the N-terminal and C-terminal interfaces.

[0054] 1.6 Model Performance Testing The fully configured DeepIntein model (Experiment 7) achieved a final performance of 0.921 AUC on the test set, indicating that the model has high prediction accuracy and generalization ability.

[0055] Example 2: Large-scale mining of novel active fragmented inteptides based on the DeepIntein model This embodiment illustrates the specific process and experimental results of using the DeepIntein model trained in Example 1 for virtual screening, and combining it with a structural clustering strategy to mine novel and structurally diverse active fragmented peptides. Figure 3 ).

[0056] 2.1 Construction and Structural Clustering of Candidate Inteptide Pools All fragmented inpeptide sequences included in the InBase inpeptide database were used as candidate sources. AlphaFold2 was used to perform batch 3D structure prediction on each sequence (in the form of continuous N-terminal and C-terminal fragments) to obtain the corresponding predicted structure model for each sequence. Figure 3 (Top left image in the image).

[0057] The Foldseek structural alignment tool was used to perform full-match structural similarity comparison and clustering on all predicted structures. A template modeling score (TM-score) ≥ 0.5 was used as the threshold for determining structural similarity. A hierarchical clustering algorithm was employed to group peptides with similar three-dimensional folding patterns into the same structural cluster. After clustering, multiple peptide clusters covering different folding types and exhibiting structural diversity were obtained, constructing a candidate peptide pool with structural heterogeneity. Figure 3 (The top right image in the image).

[0058] 2.2 Activity prediction and stratified diversity screening The DeepIntein model trained in Example 1 is invoked to perform batch activity prediction on all fragmented inteptide sequences in the candidate pool, and the splicing activity prediction probability value and corresponding prediction confidence of each sequence are obtained.

[0059] The screening strategy employs a hierarchical screening process of "first selecting the best candidate within each cluster, then ranking the overall results." (1) Within each structural cluster, the sequences with the highest scores in the cluster are selected as representative candidates based on the activity prediction probability from high to low. (2) Merge the representative candidates selected from all structural clusters and sort them globally according to the activity prediction probability; (3) Select the top k (k=6 in this example) sequences from the global ranking results as the final candidates for entering the experimental verification stage.

[0060] This screening strategy ensures that the final candidate set simultaneously meets the conditions of high activity prediction probability and high sequence / structure diversity, effectively avoiding the screening results from being overly concentrated in a single structural family.

[0061] 2.3 Experimental Verification and Results The candidate fragmented integuments obtained from screening were subjected to gene synthesis and in vivo expression activity verification (see [link]). Figure 3 (The following is the procedure). The experimental steps are as follows: (1) Construct fusion expression vectors of N-terminal fragments of each candidate fragment containing peptides and first reporter proteins (such as protein A), and fusion expression vectors of C-terminal fragments and second reporter proteins (such as protein B); (2) Co-express the corresponding N-terminal fragment expression vector and the C-terminal fragment expression vector in the Escherichia coli expression system; (3) First, the co-expression products are screened for activity using reporter signal reconstruction experiments (such as mCherry fluorescence reconstruction). The candidate inteptide is preliminarily judged based on the fluorescence signal intensity. (4) For candidates who test positive in the reporter signal test, the protein immunoblotting technique is used to detect the generation of splicing products (i.e., the fusion protein of the first reporter protein and the second reporter protein) using specific antibodies to verify and confirm the splicing activity.

[0062] Experimental results showed that 5 out of the 6 candidate fragmented integrins were confirmed to have splicing activity through the above-mentioned dual verification (see [link to experimental results]). Figure 4 The positive rate was 83%. The amino acid sequences of the five active fragments containing peptides are shown in Table 1, and their encoding DNA sequences are shown in Table 2.

[0063] Table 1. Amino acid sequences, cleavage sites, and linker sequences of five novel bioactive fragmented peptides. Table 2. DNA sequences of five novel active fragmented peptides. Sequence and structure alignment analysis revealed that the amino acid sequence similarity among the five active fragment peptides ranged from 30% to 46% (see [link to analysis]). Figure 5 (See the image above), and they belong to multiple different structural folding types, with TM-score values ​​between each pair ranging from 0.2 to 0.7 (see [reference]). Figure 5 (See the figure below). Given that the novel inteptides discovered exhibit significant differences in both sequence and structure, these sequences are easily missed if traditional screening strategies based on sequence homology comparison (i.e., using only known active sequences as templates for similarity retrieval) are employed.

[0064] The mining results of this embodiment verify the effectiveness of the method of the present invention in overcoming homologous sequence restrictions and expanding the boundaries of peptide sequence diversity and structural diversity within active fragments.

[0065] Example 3: Software encapsulation and calling interface of the DeepIntein model To facilitate the application of the method of this invention in protein engineering and synthetic biology research, this embodiment encapsulates the DeepIntein model as an independent Python software module and provides a standardized application programming interface.

[0066] The Python module provides the following calling functions: (1) Single sequence prediction interface: It receives the N-terminal fragment sequence and C-terminal fragment sequence of the fragmented peptide as input, automatically performs feature extraction, graph structure construction and model inference process, and returns the activity prediction probability value and prediction confidence.

[0067] (2) Batch prediction interface: It supports inputting a large number of candidate sequences in the form of sequence list files, using parallel computing mechanisms for batch processing, and outputting the prediction results corresponding to each sequence. It is suitable for large-scale virtual screening scenarios.

[0068] (3) Candidate sorting and filtering interface: sort the batch prediction results in descending order according to the activity prediction probability, support users to set probability thresholds to filter low confidence prediction results, and output a list of high activity candidate sequences that meet the conditions.

[0069] (4) Active learning iteration interface: Reserved interface for connection with active learning iteration framework, supporting the feedback of newly obtained experimental verification results sample data to the training set, and incrementally updating and training the model to achieve continuous optimization and iteration of model performance.

[0070] The Python modules described above are compatible with mainstream protein engineering computing environments and can run on local computing devices equipped with CPUs or GPUs. They also support deployment on high-performance cloud computing platforms and have good portability and ease of use.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various improvements and modifications to the technical solutions of the present invention without departing from the spirit and principles thereof, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the trans-splicing activity of peptides containing fragments, characterized in that, Includes the following steps: (1) Obtain the amino acid sequence of the fragmented inteptide to be predicted, wherein the fragmented intepteptide comprises an N-terminal fragment and a C-terminal fragment; (2) Based on the amino acid sequence, extract sequence features and spatial structure features to construct graph structure data with amino acid residues as nodes and spatial contact relationships between residues as edges; wherein, the features of the nodes are constructed based at least on the sequence features and fragment chain information to which the amino acid residues belong, and the connection relationship of the edges is determined based on the spatial distance threshold between pairs of amino acid residues; (3) Input the graph structure data into the trained graph attention network model to obtain the prediction results of the trans-splicing activity of the fragmented peptide.

2. The prediction method according to claim 1, characterized in that, Step (2) involves extracting the sequence features, including: Feature extraction of the amino acid sequence is performed using a language model pre-trained on protein sequence data; The language model is the ESM-2 model, and the extracted feature vector has a dimension of 1280.

3. The prediction method according to claim 1, characterized in that, Step (2) involves extracting the spatial structure features, including: The spatial coordinates of the Cα atoms of each residue in the fractured peptide were predicted using a protein three-dimensional structure prediction tool. The prediction tool is AlphaFold2.

4. The prediction method according to claim 3, characterized in that, The connection relationships of the edges mentioned in step (2) are determined according to the following rules: Calculate the Euclidean distance between any two Cα residues. If the Euclidean distance is less than or equal to a preset threshold, then establish an edge between the corresponding two nodes. The preset threshold value ranges from 6 angstroms to 10 angstroms.

5. The prediction method according to claim 4, characterized in that, The edge features in the graph structure data include: The Cα atomic distance value between the connected residues, and the cross-chain linkage identifier used to identify whether the connected residues belong to the same fragment chain.

6. The prediction method according to claim 1, characterized in that, The features of the node described in step (2) are formed by concatenating the first feature vector and the second feature vector, wherein, The first feature vector is the sequence feature, and the second feature vector is the chain attribution code used to distinguish between N-terminal segments and C-terminal segments.

7. The prediction method according to claim 1, characterized in that, The graph attention network model mentioned in step (3) includes: (a) An input layer for receiving the graph structure data; (b) At least one graph attention convolutional layer, used to calculate the importance weights of neighboring nodes through an attention mechanism, and to aggregate the features of neighboring nodes based on the weights to update the node representation of the current node; (c) Global pooling layer, used to integrate the node representations of all nodes in the graph into a graph-level representation vector; (d) Output layer, used to output splice activity prediction probabilities via an activation function based on the graph-level representation vector.

8. The prediction method according to claim 7, characterized in that, The graph attention network model is obtained through supervised training using a dataset containing both active positive and active negative samples. The loss function used in the training process is the binary cross-entropy loss function.

9. A novel method for mining active fragmented peptides, characterized in that, Includes the following steps: (1) Obtain a pool of candidate intima-peptide sequences; (2) Using the method for predicting the trans-splicing activity of the broken inteptide according to any one of claims 1 to 8, calculate the predicted probability of trans-splicing activity of each candidate sequence in the candidate inteptide sequence pool; (3) Perform structure prediction on each candidate sequence in the candidate peptide sequence pool, and perform cluster analysis based on the similarity between the predicted structures to obtain multiple structural clusters; (4) Within each of the structural clusters, at least one representative sequence is selected based on the predicted probability of trans-splicing activity, and the selected representative sequences are combined to form a set of candidate active integrins.

10. A device for predicting the trans-splicing activity of peptides containing fragments, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the prediction method as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the prediction method as described in any one of claims 1 to 8, or the steps of the mining method as described in claim 9.

Images