Fine-grained Mining Of Disease-related MiRNAs Based On Graph Representation Learning

Increasing evidence suggests that miRNA,a key biomarker,plays a vital role in various complex human diseases.Specifically,miRNA plays a regulatory role in cell development,differentiation,growth,and metabolism.Therefore,studying miRNAs related to diseases can promote understanding of their pathogenesis,clarify the disease occurrence and development process,and help efficiently and accurately develop treatment methods.Therefore,identifying miRNAs potentially associated with diseases has become a hot research topic for better understanding the pathogenesis of diseases.Exploring the potential association of miRNAs(MDA)with diseases is mainly divided into traditional wet experiments and computational methods.Traditional wet experiments use quantitative PCR technology to determine the expression level of candidate miRNAs but need help with problems such as cumbersome experimental processes,long experimental cycles,low success rates,and the need to waste a lot of workforces and financial resources.Deep learning models based on computational science have emerged to address the costly challenges.The computational model implements an end-to-end pattern,simplifies the system design and deployment process,and is suitable for MDA mining problems.However,computational methods still face some challenges: firstly,miRNA disease relationship data has the characteristic of high-dimensional sparsity,and existing models are challenging to capture nonlinear interactions and are sensitive to noise,which affects model accuracy and stability.Secondly,existing research methods cannot embed domain knowledge,limiting MDA predictions’ accuracy and reliability.Thirdly,most current methods can only predict the relationship between miRNA and disease at a coarse-grained level and cannot accurately predict what regulatory relationship it is.The fine-grained MDA prediction task needs to be further overcome.Fourthly,single-cell sequencing technology can perform high-throughput sequencing on a single-cell basis.Studying the differences in miRNA expression among different cell types can help discover miRNAs related to disease occurrence and development.Existing methods have yet to study MDA at single-cell resolution.As a result,this paper proposes a series of methods based on graph representation learning to conduct fine-grained,in-depth research on potential disease-related miRNAs.The core innovative achievements of this paper are as follows:(1)Disease-related miRNA mining method based on similarity constraint learningGiven the noise and missing information in the dataset,this paper proposes the RSCMDA method.This method combines existing biological information to construct a new disease semantic similarity network and miRNA functional similarity network and uses similarity constraints to infer MDA.RSCMDA first constructs a miRNA functional network based on the functional similarity of miRNAs and uses DAG(Directed Acyclic Graph)theory and the semantic definition of diseases to construct a disease semantic network.Then,heterogeneous graphs are created based on the validated miRNA disease relationship.Secondly,RSCMDA utilizes these three networks data to adaptively learn new information similarity networks based on similarity information during optimization.Finally,RSCMDA uses a unified constraint framework to update miRNA and disease space information to obtain robust prediction results.The results of experimental verification and case analysis of lung tumors indicate that RSCMDA is an effective tool that provides reliable prediction and analysis for disease diagnosis and treatment.In addition,this paper constructed miRNA functional similarity networks,disease semantic similarity networks,and heterogeneous graph models of miRNA and diseases,providing a solid research foundation for subsequent MDA research.(2)Knowledge-driven fine-grained mining method for disease-related miRNAsIn order to fully utilize domain knowledge to improve the embedding representation ability of the model,this paper proposes a fine-grained MDA prediction method based on heterogeneous knowledge embedding(KDFGMDA).This method first standardizes the three regulatory relationships between miRNA and disease based on the description of the relationship between miRNA and disease in the data: upregulation(UR),downregulation(DR),or dysregulation(SR).Secondly,KDFGMDA constructed a miRNA disease knowledge graph using experimentally validated miRNA disease association information and represented it in the form of triplets <m,r,d>,where m represents head entity miRNA,d represents tail entity disease,and r represents finegrained relationships.Next,KDFGMDA uses a heterogeneous neighbor encoder with perceptual relationships to embed the representation of triplet information,embeds low dimensional data representations under each fine-grained relationship through an aggregation network,and infers the confidence score of triplet information through a matching network.Therefore,this paper provides a powerful tool for standardizing the fine-grained relationship between miRNA and diseases and constructing a knowledge graph,laying the foundation for subsequent research on the fine-grained relationship of MDA.(3)Fine-grained mining method for disease-related miRNAs based on multi-relationship graph encodingIn response to the underrepresentation of data in MDA tasks and the problems of high-order neighbor data embedding and feature fusion,multi-relation encoding,and information aggregation in knowledge graphs.This paper proposes a fine-grained association prediction method between miRNA and diseases based on a multi-graph encoder network(MRFGMDA)to address the issues above.Firstly,based on the work above,MRFGMDA meticulously reveals the relationship between miRNA and diseases,including upregulation(UR),downregulation(DR),or dysregulation(SR).Secondly,MRFGMDA uses a counter-example sampling technique to select the data most dissimilar to the positive sample as the negative sample,thereby overcoming the limitation of negative samples.Finally,MRFGMDA fully utilizes nodes’ first-order and secondorder neighbor information through weighting to aggregate richer domain information.(4)Fine-grained mining method for disease-related miRNAs based on single-cell resolutionThe development of single-cell sequencing technology has brought new perspectives to MDA research issues.The current methods cannot extract critical information from a single-cell perspective and reveal the relationship between miRNA and diseases.In response to this issue,this paper proposes a disease-related miRNA analysis method based on single-cell resolution(sc IMTA).Firstly,sc IMTA proposed a multitasking method for comprehensive analysis of single-cell sequencing data,which can simultaneously complete cell clustering and functional gene cluster mining.Secondly,to address the issue of zero values in single-cell datasets,sc IMTA proposes a "deep denoising" noise data processing method,which effectively solves the problem of false zero values in current biological datasets.Finally,sc IMTA conducted comprehensive experiments on real large-scale datasets to validate the effectiveness and efficiency of the model.The experimental performance indicators and case analysis results demonstrate that sc IMTA has superior performance and good biological interpretability.