A circular RNA-disease association prediction method based on topological context alignment
Patent Information
- Application Number
- CN202611029761.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-10-09
AI Technical Summary
该类方法能够有效捕获局部拓扑模式，但消息传递过程天然受限于图结构中实际存在的连接路径，对于功能相关但结构距离较远的节点对，表示学习效果欠佳
（1）本申请通过将疾病核相似性特征与环状RNA序列语义特征分别编码后映射至统一特征空间，形成各节点的初始表征，并在后续拓扑上下文对齐模块的多次迭代更新中保持生物属性信息的持续参与，使生物属性先验与网络结构信息在深层编码过程中实现协同融合，克服了现有技术将生物属性仅作为静态输入而导致信息稀释的缺陷。
Smart Images
Figure CN122889068A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of bioinformatics technology, specifically relating to a method for predicting circular RNA-disease associations based on topological context alignment. Background Technology
[0002] Circular RNA (circRNA) is a class of covalently closed circular non-coding RNAs formed by backsplicing of precursor mRNA. It exhibits higher nuclease resistance and in vivo stability than linear RNA. Recent studies have shown that circRNAs play crucial regulatory roles in pathological processes such as tumorigenesis, neurodegenerative diseases, and cardiovascular diseases through multiple pathways, including microRNA adsorption, gene transcription regulation, and protein translation. Accurately identifying potential associations between circRNAs and diseases is of great significance for elucidating the molecular mechanisms of diseases, discovering early diagnostic biomarkers, and screening novel therapeutic targets. However, wet-lab-based association validation methods are limited by bottlenecks such as long experimental cycles, high costs, and low throughput, making it difficult to meet the growing demand for candidate association screening. Therefore, developing computational prediction methods to efficiently and reliably infer circRNA-disease associations has become a research hotspot in the field of bioinformatics.
[0003] Existing methods for predicting disease associations with circular RNA can be categorized into three main technical approaches.
[0004] The first category is based on matrix factorization or similarity propagation strategies. These methods use a known circular RNA-disease association matrix as a foundation, employing linear algebra tools such as nonnegative matrix factorization and singular value decomposition to decompose the association matrix into low-dimensional factor matrices. Alternatively, they utilize metrics such as Gaussian kernel similarity and cosine similarity to propagate functional information between nodes, thereby completing unknown associations. However, because known associations are extremely sparse in real biological networks, these methods face inherent difficulties due to insufficient information during decomposition and propagation. Furthermore, matrix factorization is essentially a linear transformation, making it difficult to characterize the complex nonlinear coupling relationship between circular RNA and disease, resulting in significant limitations in prediction accuracy.
[0005] The second category is based on shallow network representation learning paradigms. These methods embed circular RNA and disease nodes into a low-dimensional vector space using random walks, adjacency matrix factorization, or Skip-gram models. This preserves the network's connection structure as vector geometric relationships, and then evaluates the association strength through vector inner products or distance calculations. While these methods can utilize network topology to some extent, their learning objectives are usually limited to reconstructing the adjacency matrix or bringing neighboring nodes closer together. They fail to adequately utilize intrinsic attributes such as the sequence composition of circular RNA and the biological function of the disease. Furthermore, most methods rely on only a single network structure view, making it difficult to consider both local neighborhood features and global distribution patterns, thus limiting their representational capabilities.
[0006] The third category is based on deep graph neural network architectures. This type of method models the circular RNA-disease association network as a heterogeneous graph, employing graph convolutional networks, graph attention networks, or graph autoencoders. It aggregates multi-hop neighborhood information of nodes through multi-layer message passing mechanisms to update representations and performs link prediction based on the updated node embeddings. This type of method can effectively capture local topological patterns, but the message passing process is inherently limited by the actual connection paths existing in the graph structure. For functionally related but structurally distant node pairs, the representation learning effect is poor. Furthermore, existing deep graph models typically use biological attributes such as circular RNA sequence features and disease similarity features as static inputs. During multi-layer network propagation, there is a lack of continuous deep interaction between the aforementioned prior information and structural information, leading to the gradual dilution of biological attribute representations as network depth increases. In the final association determination stage, most methods simply concatenate or dot-product the representations of circular RNA and disease nodes, failing to fully characterize the differences and interactive responses between node pairs, resulting in insufficient robustness under sparse labeling conditions.
[0007] In summary, existing technologies have significant shortcomings in three aspects: deep synergistic fusion of biological attribute information and network topology information, unified modeling of multi-scale structural relationships, and discriminative construction of candidate association features. These shortcomings limit the accuracy and generalization ability of circular RNA-disease association prediction. Summary of the Invention
[0008] To address the issues of insufficient deep collaboration between biological attribute information and network topology information, inadequate unified modeling of local neighborhood and global context, and insufficient discriminative power of candidate pairing features, this application provides a circular RNA-disease association prediction method based on topological context alignment. Through alternating iterations of multi-perspective biological feature encoding, topology-aware neighborhood encoding, and context-dependent modeling, as well as composite pairing feature construction and joint optimization, high-precision association prediction is achieved.
[0009] To achieve the above technical objectives, this application specifically employs the following technical solution: In one aspect of this application, a method for predicting circular RNA-disease associations based on topological context alignment is provided, comprising the following steps: Step S1: Data acquisition and preprocessing, including constructing a circular RNA-disease bipartite graph, constructing a binary correlation matrix, and constructing a training sample set; Step S2: Multi-view biometric encoding to construct initial node representations, including: constructing a disease function embedding matrix based on the binary correlation matrix; extracting sequence semantic features from the original nucleotide sequence of circular RNA using a Transformer encoder and mapping them to the same feature space as the disease function embedding matrix; constructing a circular RNA sequence feature matrix; and vertically concatenating the circular RNA sequence feature matrix with the disease function embedding matrix in the row direction to obtain the initial representation matrix of circular RNA nodes and disease nodes in the circular RNA-disease binary network. Step S3: Topological context alignment representation learning. Construct and execute at least one topological context alignment module to iteratively update the initial representation matrices of the circular RNA nodes and disease nodes to obtain the final node representation that integrates topological structure and contextual semantics. The topological context alignment module includes two processing stages: topologically aware neighborhood encoding and contextual dependency modeling. The topology-aware neighborhood encoding stage performs neighborhood information propagation and aggregation on each node in the circular RNA-disease bipartite graph by stacking at least one graph convolutional layer to obtain a topology-enhanced representation; the context-dependent modeling stage treats the topology-enhanced representation as a node sequence, encodes the node sequence by stacking at least one Transformer layer, captures the global contextual dependencies between nodes, obtains a context-enhanced sequence representation, and uses this sequence representation as the output of the topology context alignment module; Step S4: Composite pairing modeling and association prediction, including: The embedding vectors of candidate circular RNA nodes and candidate disease nodes are extracted from the final representation of the nodes to construct a composite pairing feature vector. The composite pairing feature vector is then input into a perceptron classifier to output the predicted probability that there is an association between the candidate circular RNA node and the candidate disease node. The composite pairing feature vector is formed by sequentially concatenating the features of the circular RNA itself, the features of the disease itself, the absolute difference vector between the circular RNA and the disease, and the element-by-element interaction vector between the circular RNA and the disease. Step S5: Model training and optimization. Construct a joint optimization objective function, which includes classification loss and ranking loss. Use the training sample set to perform end-to-end joint optimization of all learnable parameters from Step S2 to Step S4. Step S6: New Sample Association Prediction. For the circular RNA-disease association data to be predicted, fix all the learnable parameters optimized in step S5, and execute steps S1 to S4 in sequence to obtain the predicted probability of the association between the circular RNA to be predicted and the disease.
[0010] In one implementation, in step S1, the circular RNA-disease dichotomy is represented as follows: ,in, It is a collection of circular RNA nodes. The number of circular RNA nodes. For disease node set, The number of disease nodes, Given a set of associated edges, representing a ring. With disease There is a correlation that can be verified experimentally, among which , ; The binary correlation matrix If it is ring-shaped With disease If there is a known association, then ,otherwise ; In the training sample set, positive samples are the known associated node pairs, and negative samples are selected from unknown associated node pairs.
[0011] In one implementation, in step S2, the disease functional embedding matrix is constructed as follows: Based on the binary correlation matrix For any two disease nodes and Define the associated vector , ,in ; Calculate the squared Euclidean distance between two disease nodes. ; Functional similarity between any two diseases is calculated using a Gaussian kernel function. The disease kernel similarity matrix is obtained. ,in For bandwidth parameters; For the disease kernel similarity matrix Perform eigenvalue decomposition , Eigenvector matrix A diagonal matrix composed of eigenvalues; a preset number of... Construct a disease functional embedding matrix using the largest eigenvalues and their corresponding eigenvectors. ,in For the front A matrix composed of eigenvectors, This is the corresponding eigenvalue diagonal matrix.
[0012] In one implementation, in step S2, the semantic features of the circular RNA sequence are constructed as follows: For each circular RNA The original nucleotide sequence Tokenization is performed and mapped to an embedding matrix through the embedding layer. ,in The length of the token sequence. For the embedded dimension; The embedding matrix The input is fed into a Transformer encoder, and the hidden state vector corresponding to the special marker [CLS] in the output of the Transformer encoder is extracted. ,in, This refers to the hidden dimension of the Transformer encoder; Through the linear projection layer Mapped to the same feature space as the disease functional embedding matrix:
[0013] in, The projection matrix; For all Perform the above operations on each circular RNA to obtain a circular RNA sequence feature matrix. , its first Behavior .
[0014] In one implementation, in step S3, during the topology-aware neighborhood encoding stage, the input of the topology context alignment module is used. As the initial propagation state Based on the binary correlation matrix Constructing symmetric normalized graph propagation operators Neighborhood information propagation and aggregation are achieved by stacking at least one graph convolutional layer. Each graph convolutional operation is defined as follows:
[0015] in, , This represents the total number of convolutional layers in the graph. For the first The learnable weight matrix of a graph convolutional layer It is a non-linear activation function; go through After the second propagation, a topological enhancement characterization is obtained. .
[0016] In one implementation, the symmetric normalized graph propagation operator The construction method is as follows: Build size is square array If circular RNA nodes Disease nodes There is a known association. ,but All other positions are set to 0; exist Add self-loop based on ,in It is the identity matrix; calculate degree matrix , Let be a diagonal matrix, and its first... diagonal elements For self-loop No. The sum of all elements in a row ; The symmetric normalized graph propagation operator is: .
[0017] In one implementation, step S3, the context dependency modeling stage, is as follows: The topology enhancement representation Treat it as a sequence of nodes to obtain the sequence input. ; Input the sequence Feed into at least one stacked Transformer layer, each Transformer layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer; After passing through at least one Transformer layer, a context-enhanced sequence representation is obtained. ,in This represents the total number of Transformer layers. Representing the sequence As the output of the current topology context alignment module .
[0018] In one implementation, the multi-head self-attention sub-layer in each Transformer layer is calculated as follows: definition The number of heads for multi-head attention is calculated as follows: , ,
[0019]
[0020] in, For the first The query, key, and value projection matrix of the head. For the first A matrix of input node sequences for each Transformer layer, where , The dimension of the key vector; The multi-head self-attention output is obtained by concatenating the outputs of all heads and projecting them through the output matrix.
[0021] in, To output the projection matrix; The calculation method for the feedforward neural network sublayer is as follows:
[0022] in, This is the weight matrix. Here, is the bias vector, and GELU is the activation function.
[0023] Residual connections and layer normalization are added after each sub-layer.
[0024] In one implementation, in step S4, the composite pairing feature vector The construction method is as follows:
[0025] in, Candidate circular RNA nodes Embedded vector, Candidate disease nodes Embedded vector, This represents element-wise multiplication. `absolute` indicates taking the absolute value of each element, and `Concat` indicates concatenating vectors. The perceptron classifier consists of at least one hidden layer and a sigmoid output layer. The learnable parameters of the perceptron classifier are jointly optimized end-to-end with the learnable parameters in steps S2 to S3.
[0026] In one implementation, the perceptron classifier employs a dual-hidden-layer perceptron structure:
[0027] Final output , representing candidate circular RNA nodes Disease nodes There are correlated predicted probabilities.
[0028] In one implementation, in step S5, the joint optimization objective function is:
[0029] in, For binary cross-entropy loss, For pairwise sorting loss, Hyperparameters for balancing the contributions of the two types of losses; For a single training sample The binary cross-entropy loss is:
[0030] in, These are real labels; 1 indicates a relationship exists, and 0 indicates no relationship exists. The predicted probability output by the perceptron classifier is obtained by averaging over all training samples. ; The pairwise sorting loss is:
[0031] in, For the positive sample set, Given a negative sample set, the margin is a preset boundary value. The predicted probability for positive samples. The predicted probability of a negative sample.
[0032] In one implementation, in step S6, for the circular RNA-disease association data to be predicted, all learnable parameters optimized in step S5 are fixed, all candidate node pairs are sorted in descending order of prediction probability, and the candidate node pairs with the highest ranking are selected as the prediction results output.
[0033] In one implementation, in step S2, the parameters of the Transformer encoder are initialized using a pre-trained model and optimized together with other learnable parameters during the end-to-end joint optimization process in step S5.
[0034] In one implementation, in step S5, the optimizer employs the Adam optimizer or a variant thereof.
[0035] In one implementation, in step S3, the number of stacked topology context alignment modules... As a preset positive integer, the topology-aware neighborhood encoding and the context-dependent modeling are coordinated and aligned in a unified deep framework through layer-by-layer stacking.
[0036] The beneficial effects of this application are as follows: (1) This application encodes the disease nuclear similarity features and the semantic features of the circular RNA sequence separately and maps them to a unified feature space to form the initial representation of each node. In the subsequent multiple iterations of the topological context alignment module, the biological attribute information is continuously involved, so that the biological attribute prior and network structure information are synergistically integrated in the deep encoding process. This overcomes the defect of the prior technology that treats biological attributes as static input and causes information dilution.
[0037] (2) This application captures the local neighborhood topology pattern of each node by stacking graph convolutional layers in the topology-aware neighborhood encoding stage, and then globally encodes the entire node sequence by stacking Transformer layers in the context dependency modeling stage. The two stages are executed alternately and stacked layer by layer, and local structure aggregation and global long-range dependency modeling are completed simultaneously in a unified framework, which solves the problem that a single graph convolutional network is difficult to capture the potential association between nodes that are structurally far apart but functionally related.
[0038] (3) This application constructs a composite pairing feature vector composed of circular RNA features, disease features, absolute difference vector and element-wise interaction vector to characterize the relationship pattern between node pairs from multiple perspectives. Combined with the training objective of jointly optimizing classification loss and ranking loss, it improves the discriminative power and robustness of candidate association discrimination under sparse association labeling conditions.
[0039] (4) This application uses an end-to-end joint optimization strategy to enable the projection matrix, Transformer encoder parameters, graph convolution layer weights, Transformer layer weights and perceptron classifier weights in disease feature decomposition to converge collaboratively under the guidance of the same objective function, thereby avoiding suboptimal solutions caused by staged optimization and ensuring prediction accuracy. Attached Figure Description
[0040] Figure 1 This is a flowchart of the circular RNA-disease association prediction method based on topological context alignment in the embodiments of this application; Figure 2 A flowchart illustrating the construction of disease functional similarity features in the embodiments of this application; Figure 3 This is a flowchart illustrating the semantic features of circular RNA sequences in the embodiments of this application. Figure 4 This is a structural diagram of the topology context alignment module in an embodiment of this application; Figure 5 This is a diagram of the composite pairing feature structure in the embodiments of this application; Figure 6 This is a structural diagram of a dual-hidden-layer perceptron in an embodiment of this application; Figure 7(a) shows the comparison results of different downstream classifiers in the embodiments of this application on the CircR2Disease dataset; Figure 7(b) shows the comparison results between the TCA-CDA model of this application and existing methods. Detailed Implementation
[0041] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0042] Example like Figure 1 As shown, a method for predicting circular RNA-disease associations based on topological context alignment includes the following steps: Step S1: Data Acquisition and Preprocessing The data used in this example comes from a publicly available bioinformatics database, specifically the circRNADisease database. From this database, an experimentally validated circRNA-disease association pair dataset was obtained, along with the complete nucleotide sequences of all relevant circRNAs. Specifically, this is the circAtlas dataset.
[0043] S1.1: Constructing a circular RNA-disease bipartite graph Based on circular RNA-disease association data in the circAtlas dataset, a set of circular RNA nodes was constructed. (common (each node) and disease node set (common (number of nodes), and the known set of associated edges This forms a circular RNA-disease dichotomy. ,in , ; S1.2: Constructing a binary incidence matrix Based on the RNA-disease dichotomy described in S1.1 Constructing a binary correlation matrix Among them, if circular RNA With disease If a known association exists, then ,otherwise ; S1.3: Constructing the training sample set Among them, positive samples are known associated node pairs (i.e., corresponding Negative samples come from unknown associated node pairs (i.e. The positions are selected from the positive samples. To balance the training data, an equal number of unknown node pairs are randomly selected as negative samples, with 2576 of each pair, for a total of 5152 training samples. Step S2: Multi-view biometric encoding to construct initial node representations S2.1 Construction of disease functional similarity features, the construction process is as follows: Figure 2 ; S2.1.1: Define the disease node association vector; Based on the binary correlation matrix described in S1.2 For any two disease nodes and Define the associated vector , ,in , respectively, represent the association patterns between the corresponding diseases and all circular RNAs.
[0044] S2.1.2: Calculate the squared Euclidean distance between two disease nodes; The calculation formula is as follows:
[0045] in, This represents the L2 norm of a vector.
[0046] S2.1.3: Constructing the Disease Kernel Similarity Matrix The functional similarity between any two diseases can be calculated using the Gaussian kernel function, and the formula is as follows:
[0047] in, For bandwidth parameters, take all... The reciprocal of the mean. Based on this, a disease kernel similarity matrix is constructed. .
[0048] S2.1.4: Eigenvalue Decomposition of Disease Kernel Similarity Matrix For the disease kernel similarity matrix described in S2.1.3 The eigenvalue decomposition is calculated using the following formula:
[0049] in, The eigenvector matrix, It is a diagonal matrix formed by the eigenvalues.
[0050] S2.1.5: Constructing a Disease Functional Embedding Matrix Select preset quantity (where is a positive integer, representing the embedding dimension) largest eigenvalues and their corresponding eigenvectors are used to construct the disease functional embedding matrix:
[0051] in, For the front A matrix composed of eigenvectors, This is the corresponding eigenvalue diagonal matrix.
[0052] S2.2 such as Figure 3 Construction of semantic features of circular RNA sequences S2.2.1: Constructing the embedding matrix For each circular RNA The original nucleotide sequence (A string consisting of the letters A, U, G, and C) is first tokenized (by a fixed length). (segmentation), and then mapped to an embedding matrix through an embedding layer. ,in The length of the token sequence is 512. The embedding dimension is 768.
[0053] S2.2.2: Input Transformer encoder The embedding matrix described in S2.2.1 The input is fed into a Transformer encoder. This encoder is initialized using a pre-trained model, DNABERT, and is jointly optimized with other parameters during the overall training process of this method. The Transformer encoder consists of multiple stacked multi-head self-attention layers, each modeling long-range dependencies between nucleotides through a scaled dot product attention mechanism. , ,
[0054]
[0055] in, For the first The query, key, and value projection matrix of the head. For the first A matrix of input node sequences for each Transformer layer, where , is the dimension of the key vector.
[0056] S2.2.3: Extract the hidden state vector Extract the hidden state vector corresponding to the special marker [CLS] in the Transformer encoder output described in S2.2.2. ( (The hidden dimension of the Transformer, specifically 768), is used as the circular RNA. The global sequence semantic representation.
[0057] S2.2.4: Mapping the hidden state vector Through a learnable linear projection layer, Mapped to the same feature space as the disease functional embedding matrix:
[0058] in, Let be the projection matrix. This is the dimension after projection, specifically 64.
[0059] S2.2.5: Constructing a circular RNA sequence feature matrix For all Perform the above operations on each circular RNA to obtain a circular RNA sequence feature matrix. , its first Behavior .
[0060] S2.3 Initial Node Representation Construction The circular RNA sequence feature matrix described in S2.2.5 With respect to the disease functional embedding matrix described in S2.1.5 By vertically concatenating the rows, we obtain the initial representation matrix of all nodes in the circular RNA-disease dichotomy network:
[0061] Among them, the former The initial representation of the corresponding circular RNA node, followed by The initial representation of the disease node corresponding to the row.
[0062] Step S3: Topological Context Alignment Representation Learning Build and execute A stack of Topology-Context Alignment Blocks (TCA-Blocks), structured as follows: Figure 4 For the initial representation matrix of S2.3 The process involves iterative updates to obtain the final node representation that integrates the topological structure and contextual semantics. .in, This is a preset positive integer representing the number of modules stacked, specifically 3.
[0063] For the TCA-Block ( Its input is the node representation of the previous layer. The output is the updated node representation. The processing of a single TCA-Block includes the following steps S3.1 to S3.2: S3.1 Topology-Aware Neighborhood Coding S3.1.1: Receive input , denoted as the initial propagation state .
[0064] S3.1.2: Constructing a symmetric normalized graph propagation operator First, build a size of square array Its element definition is as follows: If circular RNA node Disease nodes There is a known association (i.e.) ),but All other locations (including those between circular RNA nodes, between disease nodes, and between circular RNA-disease pairs where no known association exists) are set to [missing information]. .
[0065] Then Add a self-loop to the existing structure:
[0066] in, It is the identity matrix. Calculate... degree matrix (a diagonal matrix, whose first...) The diagonal elements are Then the symmetric normalization propagation operator is: .
[0067] S3.1.3: Neighborhood information propagation and aggregation are achieved by stacking at least one graph convolutional layer. Each graph convolutional operation is defined as follows:
[0068] in, , The total number of convolutional layers in the graph (a preset positive integer, specifically 2). For the first The first TCA-Block The learnable weight matrix of a graph convolutional layer It is a non-linear activation function (ReLU).
[0069] S3.1.4: After After the second propagation, the topological enhancement characterization is obtained:
[0070] This representation has aggregated the high-order neighborhood topology information of each node.
[0071] S3.2 Context Dependency Modeling S3.2.1: Characterize the topology enhancement described in S3.1.4 Viewing it as a sequence of nodes (each node's feature vector is an element in the sequence, arranged in node index order), we get the sequence input. .
[0072] S3.2.2: Input the sequence described in S3.2.1 Feed into at least one stacked Transformer layer. Each Transformer layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer.
[0073] For the Transformer layers ( , The number of Transformer layers is a preset positive integer, specifically 2, and its calculation process is as follows: (1) Multi-head self-attention sub-layer: First, define The number of heads receiving multi-head attention is a positive integer, specifically 8. Each head is calculated as follows: , ,
[0074]
[0075] in, For the first The query, key, and value projection matrix of the head. For the first A matrix of input node sequences for each Transformer layer, where , The dimension of the key vector; The outputs of all the heads are concatenated and then processed through the output projection matrix:
[0076] in, This is for outputting the projection matrix.
[0077] (2) Residual connectivity and layer normalization:
[0078] (3) Feedforward neural network sublayer: Each position independently applies a two-layer fully connected network:
[0079] in, This is the weight matrix. For bias vectors, This is the activation function.
[0080] (4) Add residual connections and layer normalization again:
[0081] in, The representation layer normalization operation ultimately .
[0082] S3.2.3: After After several Transformer layers, a context-enhanced sequence representation is obtained:
[0083] S3.2.4: Represent the sequence described in S3.2.3 As the output of the current TCA-Block, denoted as Since the order of the node sequence is consistent with the index order of the nodes in the graph structure, the output can be equivalently regarded as the node feature matrix in the graph structure space.
[0084] S3.3 Stacked Execution Execute steps S3.1 to S3.2 sequentially to complete the calculation of one TCA-Block. Output the results. As the input for the next TCA-Block, repeat steps S3.1 to S3.2. This process is repeated layer by layer. Through this layered stacking, the local topological information captured by the topology-aware neighborhood encoding and the global contextual information captured by the context-dependent modeling achieve deep interaction and fusion in the stacked topological context alignment module via alternating iterations, ultimately outputting the final node representation.
[0085] Step S4: Composite Pairing Modeling and Association Prediction S4.1 Construction of composite pairing features, its structure diagram is as follows: Figure 5 For any candidate circular RNA node ( ) and disease nodes ( ), from the node's final representation Extract the corresponding embedding vector: :Pick The former The first in the row OK; :Pick After The first in the row OK.
[0086] Constructing composite paired feature vectors Its construction method is as follows:
[0087] in: This indicates vector concatenation; This represents element-wise multiplication; This indicates taking the absolute value of each element.
[0088] Specifically, this complex feature is composed of four parts assembled sequentially: the features of the circular RNA itself. Disease characteristics The absolute difference vector between the two Element-wise interaction vector between the two .
[0089] S4.2 Association Probability Prediction The composite paired feature vector described in S4.1 The input is fed into a perceptron classifier. The perceptron classifier consists of at least one hidden layer and a sigmoid output layer, and its learnable parameters are jointly optimized end-to-end with the parameters from steps S2 to S3.
[0090] As a specific implementation method, a dual-hidden-layer perceptron is used, and its structure diagram is as follows: Figure 6 :
[0091] Final output , representing the candidate circular RNA node Disease nodes There are correlated predicted probabilities.
[0092] Step S5: Model Training and Optimization Construct a joint optimization objective function:
[0093] in: This is the binary cross-entropy loss. For a single sample... :
[0094] These are the true labels (1 indicates a correlation exists, 0 indicates no correlation). The average of all training samples is used to obtain... . This is a pairwise ranking loss, used to encourage positive samples to have higher predicted scores than negative samples. The specific implementation is as follows:
[0095] in, For the positive sample set, For the negative sample set, This is a preset boundary value, specifically 1.0. The hyperparameter used to balance the contributions of the two types of losses is 0.1.
[0096] Using the training dataset, the backpropagation algorithm is employed to perform end-to-end joint optimization of all learnable parameters in steps S2 to S4 (including the projection matrix in disease feature decomposition, Transformer encoder parameters, weights of graph convolutional layers, weights of Transformer layers, weights of the perceptron classifier, etc.) to minimize the joint optimization objective function. The optimizer uses Adam.
[0097] Step S6: New Sample Association Prediction For a circular RNA-disease association dataset to be predicted, fix all parameters after model training and optimization in step S5, acquire and preprocess the data through S1.1 and S1.2, and then execute steps S2 to S4 to obtain the predicted probability of the association between the circular RNA and the disease to be predicted. Sort all candidate node pairs in descending order of predicted probability and select the top-ranked ones as the prediction results with high confidence.
[0098] This paper addresses the issue of insufficient integration of biological attribute information and network topology information. Existing graph neural network-based methods typically only use circular RNA sequence features and disease functional similarity as static initial node feature inputs. During multi-layer propagation, biological attribute priors are gradually dominated by neighborhood information, lacking continuous interaction with structural information. To address this deficiency, this application designs a multi-view biological feature encoding in step S2, mapping circular RNA sequence semantics and disease functional similarity to a shared feature space to form initial node representations. In the topology context alignment module (TCA-Block) of step S3, the alternating execution of topology-aware neighborhood encoding and context-dependent modeling ensures that biological attribute priors are collaboratively aligned with structural information in each layer of representation update. As shown in Table 1, the F1 score and MCC of the complete invention are significantly higher than those of the variants with "removed topology branches" or "removed context branches," demonstrating the effectiveness of the deep collaborative mechanism proposed in this application.
[0099] Table 1
[0100] To address the issue of a single perspective in structural modeling, this application achieves continuous collaborative modeling of local topological patterns and global contextual dependencies. Existing graph convolution-based methods are limited to fixed-order neighborhood aggregation, making it difficult to capture long-range semantic relationships between arbitrary nodes in the graph; while transformer-based methods, although capable of modeling global dependencies, may lose the inherent prior knowledge of the graph's neighborhood structure. To address this deficiency, this application designs an alternating execution mechanism of topology-aware neighborhood encoding (stacked graph convolutional layers) and contextual dependency modeling (stacked Transformer layers) within the TCA-Block, and stacks multiple TCA-Blocks in the depth direction, aligning local neighborhood patterns and global long-range dependencies layer by layer within the same framework. As shown in Figure 7(a), using the representation learning framework of this application, high performance can be achieved even with a simple classifier; and Figure 7(b) further shows that the AUC of this application reaches 0.9296 and the MCC reaches 75.99% on the CircR2Disease dataset, both significantly outperforming existing methods that rely solely on single-structure modeling.
[0101] To address the issue of insufficient discriminative power of candidate pairing features, this application constructs composite pairing features rich in both difference and interaction relationships. Combined with a joint supervision strategy, this significantly improves discriminative power and ranking stability. Existing methods often use vector concatenation or dot product similarity as pairing features, failing to explicitly model the difference information (such as vector distance) and affinity interaction information (such as element-wise multiplication) between node pairs. Furthermore, training often employs a single cross-entropy loss, neglecting the ranking constraints between positive and negative samples. To address this deficiency, this application constructs composite pairing features in step S4. And in step S5, a joint optimization objective is introduced. As shown in Figure 7(b), the classification performance of the double hidden layer perceptron (which can make full use of composite features) is significantly better than that of the dot product or single linear classifier. It has been verified by the literature that more than 80% of the top 20 candidate circular RNAs predicted on three datasets have been verified, which proves the robust ranking ability of this application under sparse data.
[0102] The method exhibits excellent overall prediction performance and strong generalization ability. Five-fold cross-validation results on three public benchmark datasets—circAtlas, CircR2Disease, and circRNADisease—(as shown in Table 2) demonstrate that the average AUCs of this application reach 0.9145, 0.9296, and 0.9244, respectively, and the F1 scores reach 85.04%, 88.09%, and 87.06%, respectively. All these indicators outperform existing representative methods.
[0103] Table 2
[0104] This application effectively addresses the shortcomings of existing technologies, such as insufficient fusion of biological attributes and structural information, single perspective of structural modeling, and insufficient discriminative power of candidate pairing features, by employing topological context-aligned representation learning, multi-perspective biomarker encoding, composite pairing features, and joint supervision strategies. This results in more accurate and robust prediction of circular RNA-disease associations.
[0105] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.
Claims
1. A method for predicting circular RNA-disease associations based on topological context alignment, characterized in that, Includes the following steps: Step S1: Data acquisition and preprocessing, including constructing a circular RNA-disease bipartite graph, constructing a binary correlation matrix, and constructing a training sample set; Step S2: Multi-view biometric encoding to construct initial node representations, including: constructing a disease function embedding matrix based on the binary correlation matrix; extracting sequence semantic features from the original nucleotide sequence of circular RNA using a Transformer encoder and mapping them to the same feature space as the disease function embedding matrix; constructing a circular RNA sequence feature matrix; and vertically concatenating the circular RNA sequence feature matrix with the disease function embedding matrix in the row direction to obtain the initial representation matrix of circular RNA nodes and disease nodes in the circular RNA-disease binary network. Step S3: Topological context alignment representation learning. Construct and execute at least one topological context alignment module to iteratively update the initial representation matrices of the circular RNA nodes and disease nodes to obtain the final node representation that integrates topological structure and contextual semantics. The topological context alignment module includes two processing stages: topologically aware neighborhood encoding and contextual dependency modeling. Step S4: Composite pairing modeling and association prediction, including: The embedding vectors of candidate circular RNA nodes and candidate disease nodes are extracted from the final representation of the nodes to construct a composite pairing feature vector. The composite pairing feature vector is then input into a perceptron classifier to output the predicted probability that there is an association between the candidate circular RNA node and the candidate disease node. Step S5: Model training and optimization. Construct a joint optimization objective function, which includes classification loss and ranking loss. Use the training sample set to perform end-to-end joint optimization of all learnable parameters in steps S2 to S4. Step S6: New Sample Association Prediction. For the circular RNA-disease association data to be predicted, fix all the learnable parameters optimized in step S5, and execute steps S1 to S4 in sequence to obtain the predicted probability of the association between the circular RNA to be predicted and the disease.
2. The method according to claim 1, characterized in that, In step S1, the circular RNA-disease dichotomy is represented as follows: ,in, It is a collection of circular RNA nodes. The number of circular RNA nodes. For disease node set, The number of disease nodes, Given a set of associated edges, representing a ring. With disease There is a correlation that can be verified experimentally, among which , ; The binary correlation matrix If it is ring-shaped With disease If there is a known association, then ,otherwise ; In the training sample set, positive samples are the known associated node pairs, and negative samples are selected from unknown associated node pairs.
3. The method according to claim 2, characterized in that, In step S2, the disease functional embedding matrix is constructed as follows: based on the binary correlation matrix. For any two disease nodes and Define the association vector , ,in ; Calculate the squared Euclidean distance between two disease nodes. ; Functional similarity between any two diseases is calculated using a Gaussian kernel function. The disease kernel similarity matrix is obtained. ,in For bandwidth parameters; For the disease kernel similarity matrix Perform eigenvalue decomposition , Eigenvector matrix A diagonal matrix composed of eigenvalues; a preset number of... Construct a disease functional embedding matrix using the largest eigenvalues and their corresponding eigenvectors. ,in For the front A matrix composed of eigenvectors, This is the corresponding eigenvalue diagonal matrix.
4. The method according to claim 2, characterized in that, In step S2, the semantic features of the circular RNA sequence are constructed as follows: For each ring The original nucleotide sequence Tokenization is performed and mapped to an embedding matrix through the embedding layer. ,in The length of the token sequence. For the embedded dimension; The embedding matrix The input is fed into a Transformer encoder, and the hidden state vector corresponding to the special marker [CLS] in the output of the Transformer encoder is extracted. ,in, This refers to the hidden dimension of the Transformer encoder; Through the linear projection layer Mapped to the same feature space as the disease functional embedding matrix: in, The projection matrix; For all Perform the above operations on each circular RNA to obtain a circular RNA sequence feature matrix. , its first Behavior .
5. The method according to claim 2, characterized in that, In step S3, during the topology-aware neighborhood encoding stage, the input of the topology context alignment module is used. As the initial propagation state Based on the binary correlation matrix Constructing symmetric normalized graph propagation operators Neighborhood information propagation and aggregation are achieved by stacking at least one graph convolutional layer. Each graph convolutional operation is defined as follows: in, , This represents the total number of convolutional layers in the graph. For the first The learnable weight matrix of a graph convolutional layer It is a non-linear activation function; go through After the second propagation, a topological enhancement characterization is obtained. .
6. The method according to claim 5, characterized in that, The symmetric normalized graph propagation operator The construction method is as follows: Build size is square array If circular RNA nodes Disease nodes There is a known association. ,but All other positions are set to 0; exist Add self-loop based on ,in It is the identity matrix; calculate degree matrix , Let be a diagonal matrix, and its first... The diagonal elements are ; The symmetric normalized graph propagation operator is: .
7. The method according to claim 5, characterized in that, In step S3, the context dependency modeling stage is as follows: The topology enhancement representation Treat it as a sequence of nodes to obtain the sequence input. ; Input the sequence Feed into at least one stacked Transformer layer, each Transformer layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer; After passing through at least one Transformer layer, a context-enhanced sequence representation is obtained. ,in This represents the total number of Transformer layers. Representing the sequence As the output of the current topology context alignment module .
8. The method according to claim 7, characterized in that, In each of the Transformer layers, the multi-head self-attention sublayer is calculated as follows: definition The number of heads for multi-head attention is calculated as follows: ， , in, For the first The query, key, and value projection matrix of the head. For the first A matrix of input node sequences for each Transformer layer, where , The dimension of the key vector; The multi-head self-attention output is obtained by concatenating the outputs of all heads and projecting them through the output matrix. in, To output the projection matrix; The calculation method for the feedforward neural network sublayer is as follows: in, This is the weight matrix. Here, is the bias vector, and GELU is the activation function. Residual connections and layer normalization are added after each sub-layer.
9. The method according to claim 2, characterized in that, In step S4, the composite paired feature vector The construction method is as follows: in, Candidate circular RNA nodes Embedded vector, Candidate disease nodes Embedded vector, This represents element-wise multiplication. `absolute` indicates taking the absolute value of each element, and `Concat` indicates concatenating vectors. The perceptron classifier consists of at least one hidden layer and a sigmoid output layer. The learnable parameters of the perceptron classifier are jointly optimized end-to-end with the learnable parameters in steps S2 to S3.
10. The method according to claim 2, characterized in that, In step S5, the joint optimization objective function is: in, For binary cross-entropy loss, For pairwise sorting loss, Hyperparameters for balancing the contributions of the two types of losses; For a single training sample The binary cross-entropy loss is: in, These are real labels; 1 indicates a relationship exists, and 0 indicates no relationship exists. The predicted probability output by the perceptron classifier is obtained by averaging over all training samples. ; The pairwise sorting loss is: in, For the positive sample set, Given a negative sample set, the margin is a preset boundary value. The predicted probability for positive samples. The predicted probability of a negative sample.