Method for predicting multi-type association in miRNA-LncRNA-disease network through simplified input strategy

Abstract

The invention discloses a method for predicting multi-type association in a miRNA-LncRNA-disease network through a simplified input strategy, and relates to the technical field of biomolecular association prediction, and the method comprises the following steps: data acquisition and preprocessing, heterogeneous network construction, deep feature extraction, feature fusion, multi-task joint prediction and result output. Innovation and breakthrough of correlation prediction of multiple types of biomolecules are achieved by simplifying an input strategy and a depth feature fusion mechanism, a verified binary correlation matrix is adopted to replace a traditional similarity network, redundant calculation noise can be effectively avoided, and the correlation prediction of multiple types of biomolecules is achieved by stacking a depth matrix decomposition module. Low-dimensional features are gradually extracted by adopting an encoder-decoder cascade structure, known associated signals can be enhanced, false negative interference can be inhibited, and by introducing a dynamic weighted feature fusion and multi-task joint prediction framework, the collaborative reasoning capability of three types of associations can be synchronously optimized, and the data collection complexity is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of biomolecular association prediction technology, and in particular to a simplified input strategy for predicting multi-type associations in miRNA-LncRNA-disease networks. Background Technology

[0002] In the current field of bioinformatics, association prediction methods based on molecular similarity networks have been extensively studied. These methods typically integrate multi-source data, such as miRNA sequence similarity, lncRNA expression profile similarity, and disease phenotype similarity, to construct complex network topologies among biomolecules. Typical technical approaches include: using Gaussian interaction spectral kernel functions to calculate molecular similarity matrices and combining them with network propagation algorithms (such as random walks or graph convolution) to mine potential associations; or using tensor decomposition techniques to map multi-dimensional similarity networks to a low-dimensional space and using the inner product operation of latent feature vectors to predict association probabilities. Existing methods generally rely on the accuracy of similarity network construction and improve prediction robustness through multimodal data fusion.

[0003] In recent years, matrix factorization techniques have demonstrated significant value in predicting biomolecular associations. Traditional nonnegative matrix factorization (NMF) methods capture potential interaction patterns between biomolecules by decomposing the original association matrix into the product of a low-dimensional basis matrix and a coefficient matrix. Advanced approaches introduce deep matrix factorization frameworks, employing multi-layer encoder-decoder structures to progressively extract high-order nonlinear features, such as reconstructing the input matrix using an autoencoder to optimize feature representation. Some studies have further extended this to multi-task learning frameworks, simultaneously handling multiple association prediction tasks such as miRNA-disease and lncRNA-disease associations, and leveraging parameter sharing mechanisms to enhance model generalization capabilities. These methods effectively uncover synergistic regulatory relationships between biomolecules through joint optimization of the low-dimensional feature space.

[0004] Existing multi-type association prediction methods suffer from the following drawbacks: reliance on high-dimensional similarity networks necessitates the integration of multimodal data such as sequence and phenotype data, leading to high collection costs and introducing noise errors; single-task modeling struggles to capture cross-molecular-level synergistic regulatory relationships, limiting cross-dataset generalization ability; and shallow feature extraction mechanisms are sensitive to false negatives and have poor adaptability to sparse data, failing to meet the demands of clinical decision-making for highly robust models. Therefore, this invention proposes a simplified input strategy for predicting multi-type associations in miRNA-lncRNA-disease networks to address the problems existing in current technologies. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose a simplified input strategy for predicting multi-type associations in miRNA-LncRNA-disease networks. This method solves the problems of existing multi-type association prediction methods, such as difficulty in data collection, significant noise, insufficient collaborative reasoning ability of multi-source data affecting cross-dataset generalization and accurate identification, and weak model robustness.

[0006] To achieve the objectives of this invention, the present invention is implemented through the following technical solution: a simplified input strategy method for predicting multi-type associations in miRNA-lncRNA-disease networks, comprising the following steps:

[0007] Step 1: Extract validated binary association data for miRNA-disease, lncRNA-disease, and lncRNA-miRNA from biomedical databases and preprocess them to generate three association matrices;

[0008] Step 2: Construct a three-layer heterogeneous network based on three types of association matrices, including a miRNA node layer, a LncRNA node layer, and a disease node layer;

[0009] Step 3: Process the heterogeneous network through a stacked deep matrix factorization module containing at least three cascaded encoder-decoder units to extract miRNAs, lncRNAs, and low-dimensional potential features of the disease layer by layer;

[0010] Step 4: Input the three types of low-dimensional latent features obtained in Step 3 into the feature fusion module to generate a unified feature representation;

[0011] Step 5: Based on the unified feature representation, generate three association prediction matrices simultaneously: miRNA-disease, LncRNA-disease, and LncRNA-miRNA. Perform multi-task joint prediction and normalize the prediction results.

[0012] Step 6: Filter high-confidence candidate associations based on the normalized confidence scores, and output them in sorted order.

[0013] Further improvements are made in the following steps: In step one, the specific preprocessing steps include data cleaning, data encoding, and cross-validation. Data cleaning removes duplicate records, conflicting experimental evidence records, and related data with confidence levels below a preset threshold. Data encoding encodes known associations as the value 1 and unknown associations as the value 0. Cross-validation is achieved by verifying the consistency of related data through at least two independent biomedical databases.

[0014] A further improvement is made in step three, where the input of the first layer encoder of the stacked depth matrix factorization module is the original correlation matrix, and the input of each subsequent layer encoder is the output feature of the previous layer decoder. The decoder uses a nonlinear activation function to strengthen the signal strength of known correlations and suppress noise interference from unknown correlations. The network parameters are jointly optimized by minimizing the reconstruction error and the classification loss.

[0015] A further improvement is that the nonlinear activation function is the ReLU function, which sets the negative part of the reconstructed matrix to zero in the decoder, thus preserving the known strong signal features.

[0016] A further improvement is made in step four, where the feature fusion module includes a splicing mode, a summation mode, and a dynamic weighting mode. The splicing mode splices three types of low-dimensional features sequentially along the feature dimensions into a long feature vector. The summation mode performs element-wise addition on the three types of low-dimensional features. The dynamic weighting mode generates weight coefficients for the three types of features through an attention mechanism, and generates a fused feature after weighted summation.

[0017] A further improvement is made in step five, when performing multi-task joint prediction, firstly, an initial prediction score matrix is ​​generated through matrix multiplication, then the initial score is normalized using the sigmoid function to output a confidence score of 0 to 1, and then the accuracy of the three types of association prediction is simultaneously optimized through the multi-task loss function.

[0018] Further improvements include model training and validation steps, specifically as follows: using validated associations as positive samples and unknown associations as negative samples, an equal amount of random sampling is used to construct a training set and the parameters are optimized using a weighted cross-entropy loss function. The effectiveness of the input simplification strategy is verified through ablation experiments.

[0019] Further improvements are made in that: the ablation experiment includes three comparison modes: single matrix input mode, dual matrix input mode, and full matrix input mode. The single matrix input mode uses only the miRNA-disease association matrix for prediction, the dual matrix input mode uses both miRNA-disease and LncRNA-disease association matrices, and the full matrix input mode uses all three types of association matrices simultaneously.

[0020] Computer systems applied to simplifying input strategies for predicting multi-type associations in miRNA-lncRNA-disease networks include:

[0021] Data storage unit is used to store raw biomedical data, preprocessed correlation matrix, and model parameters;

[0022] The feature processing unit is configured with a hardware accelerator for the stacked depth matrix factorization module and the feature fusion module;

[0023] The prediction processing unit is a graphics processor that performs multi-task joint prediction and normalization processing.

[0024] The results output unit is a visual interface for generating a candidate association list sorted by confidence level.

[0025] The central control unit is a processor that coordinates the execution of operational procedures by various units.

[0026] The beneficial effects of this invention are as follows: This invention achieves an innovative breakthrough in predicting associations of multiple types of biomolecules by simplifying the input strategy and using a deep feature fusion mechanism. It uses a validated binary association matrix to replace the traditional similarity network, which can effectively avoid redundant computational noise. By stacking deep matrix factorization modules and using an encoder-decoder cascade structure to progressively extract low-dimensional features, it can strengthen known association signals and suppress false negative interference. Furthermore, by introducing dynamic weighted feature fusion and a multi-task joint prediction framework, it can simultaneously optimize the collaborative reasoning ability of the three types of associations, significantly reduce the complexity of data collection, and is especially suitable for sparse data scenarios of rare diseases. It can also improve the generalization ability across datasets, accurately identify disease biomarkers and drug targets, and provide reliable decision support for personalized medicine. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a simplified input strategy for predicting multi-type associations in miRNA-LncRNA-disease networks according to an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] In current biomedical research, miRNA-lncRNA-disease association prediction primarily relies on heterogeneous network models. These models typically integrate multiple biological data sources, including molecular sequence similarity, functional similarity, and known association data, to construct multi-layered biomolecular interaction networks. Typical methods extract low-dimensional feature representations of molecular nodes from the network using matrix factorization or random walk algorithms, and then predict potential associations based on feature similarity. For example, some approaches use graph convolutional networks (GCNs) to capture the topological relationships between nodes, or utilize tensor decomposition to fuse multiple types of association matrices. The core of these techniques lies in uncovering implicit biological association patterns between molecules through complex network structures.

[0030] Existing deep learning frameworks primarily employ feature learning strategies in biological association prediction. Among these, autoencoders are widely used to process high-dimensional, sparse association matrices. The encoder compresses the original data into a low-dimensional latent space, and the decoder reconstructs the association matrix, thereby learning robust feature representations. Advanced methods introduce multi-task learning mechanisms to simultaneously optimize multiple association prediction targets (such as miRNA-disease, lncRNA-disease, etc.), sharing underlying features to improve generalization. Some schemes combine attention mechanisms to dynamically weight contributions from different data sources, or design hierarchical networks to progressively refine the feature extraction process. These techniques achieve the mapping from raw data to association predictions through end-to-end training.

[0031] Example 1

[0032] See Figure 1 This embodiment provides a simplified input strategy for predicting multiple types of associations in miRNA-lncRNA-disease networks, including the following steps:

[0033] Step 1: Data Acquisition and Preprocessing

[0034] Validated binary association data of miRNA-disease, lncRNA-disease, and lncRNA-miRNA were extracted from biomedical databases and preprocessed to generate three types of association matrices. Known associations were marked with a value of 1, and unknown associations were marked with a value of 0. The reference basis of technical terms was established through the data encoding stage.

[0035] This embodiment preprocesses three types of binary association data in three stages: data cleaning, data encoding, and cross-validation.

[0036] Data cleaning stage: Remove duplicate records (multiple duplicate entries of the same association), records of conflicting experimental evidence (contradictory conclusions from different sources or experimental methods), and associated data with confidence levels below a preset threshold (data with insufficient literature support or statistical significance below a preset threshold), in order to eliminate noisy data from the source and avoid interference from erroneous associations to subsequent network construction and model training.

[0037] Data encoding phase: Known associations (verified miRNA-disease / lncRNA-disease / lncRNA-miRNA relationships) are uniformly encoded as the value 1, and unknown associations (potential relationships that are not verified or lack evidence) are encoded as the value 0. By establishing a standardized binary matrix representation, structured input is provided for the construction of heterogeneous networks. Technical terms (such as "verified associations") are clearly defined through 0 / 1 encoding to ensure the compliance of the basis of reference in the claims.

[0038] Cross-validation phase: Extract similar association data from at least two independent biomedical databases (such as HMDD, LncRNADisease, DisGeNET), compare and verify the consistency of the data, and retain only the association records that co-occur in multiple databases to solve the coverage bias problem of a single database, improve the biological credibility of the data, provide highly reliable positive samples for subsequent model training, and reduce the risk of false positives.

[0039] Step 2: Construction of Heterogeneous Networks

[0040] Based on the three types of association matrices from step one, a three-layer heterogeneous network is constructed, comprising a miRNA node layer, a LncRNA node layer, and a disease node layer. The layers are connected by corresponding association matrices, wherein:

[0041] miRNA node layer: Each node represents a miRNA molecule, and node features can be initialized as an embedded representation of molecular sequence or expression profile;

[0042] LncRNA node layer: Each node represents a LncRNA molecule, and node features can be initialized based on sequence or functional annotations;

[0043] Disease node layer: Each node represents a disease, and node features can be constructed based on disease ontology (such as MeSH) or phenotypic data;

[0044] The rules for inter-layer connections are as follows:

[0045] miRNA-disease layer connections: Edges are defined using the preprocessed miRNA-disease association matrix (values ​​1 / 0), and connections are established between node pairs with known associations;

[0046] LncRNA-Disease Layer Connections: Establishing cross-layer edges based on the LncRNA-disease association matrix;

[0047] LncRNA-miRNA layer connection: Connecting two types of RNA molecule nodes through the LncRNA-miRNA association matrix;

[0048] This embodiment directly constructs connections based on the verified binary association matrix (0 / 1), avoiding redundant noise introduced by similarity network calculations in traditional methods (such as Gaussian kernel function and cosine similarity), and clearly distinguishes known and unknown associations by discretizing edge weights (0 or 1), which can avoid the interpretive ambiguity caused by continuous similarity scores;

[0049] Step 3: Deep Feature Extraction

[0050] The heterogeneous network in step two is processed by a stacked deep matrix factorization module. The stacked deep matrix factorization module contains at least three cascaded encoder-decoder units, which extract low-dimensional latent features of miRNA, lncRNA and disease layer by layer. By strengthening known association signals, the reconstruction error and classification loss are jointly optimized to improve the quality of feature representation. The encoder performs dimensionality reduction mapping, mapping the high-dimensional association matrix to the low-dimensional latent feature space to extract deep features of miRNA, lncRNA and disease. The decoder performs feature reconstruction, attempting to restore the input association matrix. By minimizing the difference between the reconstructed matrix and the original matrix, the network parameters are optimized.

[0051] The three-level encoder-decoder structure of the stacked depth matrix factorization module is as follows:

[0052] The input to the first encoder is the original correlation matrix, and the input to each subsequent encoder is the output feature of the previous decoder.

[0053] The decoder uses the ReLU function to suppress noise (negative values ​​return to zero), suppressing noise interference from unknown correlations and amplifying known correlation signals;

[0054] Jointly optimize reconstruction error and classification loss to improve feature representation quality;

[0055] This embodiment uses layer-by-layer feature extraction and joint optimization to enable the model to learn deeper biomolecular association features and improve prediction performance.

[0056] Step 4: Feature Fusion

[0057] The three types of low-dimensional features obtained in step three are input into the feature fusion module to generate a unified feature representation. The feature fusion module supports three modes: splicing, summation, and dynamic weighting, and is used to integrate low-dimensional features of miRNA, lncRNA, and diseases to meet the needs of different data scenarios, as detailed below:

[0058] The splicing mode splices three types of low-dimensional features sequentially along the feature dimensions into a long feature vector, preserving the independent feature information of the three types of biomolecules and avoiding information loss. It is suitable for research scenarios that require explicit differentiation of molecular types (such as screening for specific biomarkers).

[0059] Additive mode: Performs element-wise addition on three types of low-dimensional features, strengthens common features across molecular categories, improves the compactness of feature representation, and is suitable for highly correlated molecular regulatory networks (such as miRNA-lncRNA co-regulating the same disease);

[0060] Dynamic weighting mode: After generating weight coefficients through an attention mechanism, the weights are summed to dynamically adapt to different data distributions. For example, in rare disease data, the weights of disease features are increased to suppress noise feature interference (such as low-quality LncRNA data) and enhance the robustness of the model.

[0061] Step 5: Multi-task joint prediction

[0062] Based on the unified feature representation in step four, three association prediction matrices—miRNA-disease, LncRNA-disease, and LncRNA-miRNA—are generated simultaneously. Multi-task joint prediction is then performed, and the prediction results are normalized to improve the model's prediction accuracy and generalization ability. The specific steps for performing multi-task joint prediction are as follows:

[0063] Using a unified feature representation, the initial prediction score matrix for the three types of associations is calculated by matrix multiplication;

[0064] The initial scores are normalized using the Sigmoid function, mapping the scores to confidence scores between 0 and 1, representing the probability of the association existing;

[0065] Construct a multi-task loss function that includes reconstruction loss (measuring the difference between the reconstructed features and the original data) and classification loss (measuring the difference between the predicted results and the true labels) to simultaneously optimize the accuracy of the three-class association predictions;

[0066] The multi-task learning in this embodiment improves the robustness and generalization ability of the model by sharing model parameters and utilizing the correlation between tasks. The normalization process makes the prediction results of different tasks comparable, which facilitates subsequent analysis and decision-making. Furthermore, the simultaneous prediction of three types of associations reduces the time cost of model training and inference.

[0067] Step Six: Output Results

[0068] High-confidence candidate associations are selected based on the normalized confidence scores, and the results are sorted and output to achieve multi-type association prediction, providing strong support for subsequent analysis and decision-making.

[0069] In this embodiment, the verified associations marked with a value of 1 in step one are used as positive samples, and the unknown associations marked with a value of 0 are used as negative samples. An equal amount of random sampling is used to construct a training set and the parameters are optimized using a weighted cross-entropy loss function. The effectiveness of the input simplification strategy is verified through ablation experiments.

[0070] The ablation experiment in this embodiment includes the following comparison modes:

[0071] Single matrix input mode: Prediction is made using only the miRNA-disease association matrix;

[0072] Dual-matrix input mode: combining miRNA-disease and lncRNA-disease association matrices;

[0073] Full matrix input mode: uses three types of correlation matrices simultaneously;

[0074] This embodiment verifies the effectiveness of the input simplification strategy in reducing noise interference and improving model generalization by comparing single-matrix (miRNA-disease only), dual-matrix (miRNA-disease + LncRNA-disease) and full-matrix (three-class joint) input modes.

[0075] The simplified input strategy for predicting multi-type associations in miRNA-lncRNA-disease networks, as described in this embodiment, has the following applications in the biomedical field:

[0076] Disease biomarker screening: Extract miRNA / LncRNA molecules with a correlation higher than a preset threshold with the target disease from the prediction results;

[0077] Drug target prediction: Identifying modifiable regulatory targets based on high-confidence LncRNA-miRNA associations;

[0078] Personalized treatment: Integrating patient-specific expression data to generate customized medication recommendations.

[0079] Example 2

[0080] This embodiment provides a computer system for predicting multiple association types in miRNA-lncRNA-disease networks using a simplified input strategy. The system consists of a data storage unit, a feature processing unit, a prediction computation unit, a result output unit, and a central control unit. Through the coordinated operation of these five units, efficient prediction of multiple association types in miRNA-lncRNA-disease networks is achieved, supporting biomedical applications such as disease biomarker screening, drug target prediction, and personalized treatment.

[0081] The data storage unit is responsible for storing raw biomedical data, preprocessed correlation matrices, and model parameters, providing data support for the system and ensuring data security and integrity.

[0082] The feature processing unit is equipped with hardware accelerators that stack the deep matrix factorization module and the feature fusion module to perform deep feature extraction and fusion operations, thereby improving the speed and efficiency of feature processing.

[0083] The prediction computing unit utilizes a graphics processing unit (GPU) to perform multi-task joint prediction and normalization processing, generating an associated prediction matrix. The GPU has strong parallel computing capabilities, which accelerates the model training and prediction process and improves the system response speed.

[0084] The results output unit generates a candidate association list sorted by confidence level and displays the results through a visualization interface, providing an intuitive user interface to facilitate researchers in viewing and analyzing the prediction results;

[0085] The central control unit, as the core of the system, coordinates the execution of operational procedures by each unit to ensure the stable operation of the system.

[0086] This computer system, through the collaborative work of its various units, achieves efficient prediction of multiple types of associations in the miRNA-LncRNA-disease network. It is characterized by high efficiency, stability, and user-friendliness, providing a powerful tool for biomedical research and clinical applications.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for simplifying input strategy prediction of multi-type association in miRNA-LncRNA-disease network, characterized in that, The method comprises the following steps: Step 1: extracting verified miRNA-disease, LncRNA-disease and LncRNA-miRNA binary correlation data from biomedical databases and preprocessing to generate three types of correlation matrices; Step 2: constructing a three-layer heterogeneous network containing miRNA node layer, LncRNA node layer and disease node layer based on the three types of correlation matrices; Step 3: processing the heterogeneous network through a stacked deep matrix decomposition module containing at least three cascaded encoder-decoder units to extract low-dimensional latent features of miRNA, LncRNA and disease layer by layer; Step 4: inputting the three types of low-dimensional latent features obtained in step 3 into a feature fusion module to generate a unified feature representation; Step 5: based on the unified feature representation, synchronously generating miRNA-disease, LncRNA-disease and LncRNA-miRNA three types of correlation prediction matrix, performing multi-task joint prediction, and normalizing the prediction results; Step 6: filtering high-confidence candidate correlations according to the normalized confidence scores and outputting them in order of score.

2. The method of claim 1, wherein the method is used to predict the multi-type associations in the miRNA-LncRNA-disease network based on the simplified input strategy. In step 1, the specific steps of preprocessing include data cleaning, data encoding and cross-validation. Through data cleaning, duplicate records, experimental evidence conflict records and correlation data with confidence lower than the preset threshold are removed. Through data encoding, known correlations are encoded as numerical 1 and unknown correlations are encoded as numerical 0. Through consistency verification of at least two independent biomedical databases, cross-validation is achieved. 3.The method of claim 1, wherein: In step 3, the input of the first layer encoder of the stacked deep matrix decomposition module is the original correlation matrix, and the input of each subsequent layer encoder is the output feature of the previous layer decoder. The decoder uses a nonlinear activation function to strengthen the signal strength of known correlations and suppress the noise interference of unknown correlations. Network parameters are optimized by minimizing the reconstruction error and classification loss.

4. The method of claim 3, wherein the method is used to predict the multi-type associations in the miRNA-LncRNA-disease network. The nonlinear activation function is ReLU function, which sets the negative part of the reconstructed matrix to zero and retains the strong signal features of known correlations.

5. The method of claim 1, wherein the method is used to predict the multi-type associations in the miRNA-LncRNA-disease network. In step 4, the feature fusion module includes concatenation mode, summation mode and dynamic weighting mode. The concatenation mode concatenates the three types of low-dimensional features along the feature dimension to form a long feature vector. The summation mode performs element-wise addition operation on the three types of low-dimensional features. The dynamic weighting mode generates weight coefficients of the three types of features through attention mechanism, and generates fusion features after weighted summation. 6.The method of claim 1, wherein: In step 5, when performing multi-task joint prediction, the initial prediction score matrix is generated through matrix multiplication operation, then the initial score is normalized by sigmoid function to output confidence scores between 0 and 1, and then the accuracy of three types of correlation prediction is optimized simultaneously through multi-task loss function.

7. The method of claim 1, wherein the method is used to predict the multi-type associations in the miRNA-LncRNA-disease network. It also includes a model training and verification step, which is as follows: taking verified correlations as positive samples and a certain amount of random samples from unknown correlations as negative samples to construct a training set and optimize parameters using weighted cross-entropy loss function, and verifying the effectiveness of input simplification strategy through ablation experiment.

8. The method of claim 7, wherein the method is used to predict the multi-type associations in the miRNA-LncRNA-disease network. The ablation experiment includes three comparative modes of single matrix input mode, double matrix input mode and full matrix input mode, the single matrix input mode only uses the miRNA-disease association matrix for prediction, the double matrix input mode jointly uses the miRNA-disease and LncRNA-disease association matrix, and the full matrix input mode simultaneously uses the three types of association matrix.

9. A computer system for applying the multi-type relationship method for predicting the miRNA-LncRNA-disease network according to any one of claims 1-8, characterized in that, Comprise: a data storage unit for storing original biomedical data, preprocessed association matrix and model parameters; a feature processing unit configured with a hardware accelerator of stacked depth matrix decomposition module and feature fusion module; a prediction operation unit of a graphics processor for performing multi-task joint prediction and normalization processing; a result output unit of a visualization interface for generating a candidate association list in confidence order; a central control unit of a processor for coordinating the operation of each unit.

Images