Research On Prognostic Risk Assessment Algorithm Based On Multi-omics Data Analysis And Deep Learning

Valuable information for the prediction of cancer risk can be exposed with genome-wide multi-omics data,and the correlation of multidimensional genomic measurements with cancer outcomes(survival or death)is moreover helpful for bridging the gap between genomic data and clinical practice.By integrating multiomics data and assessing medical risk based on patient-specific genes(cancer prognostic risk assessment),prognostic interventions and personalized treatment can be efficiently provided for patients,thus significantly improving the survival rate.However,there are still challenges existing in integrating,multi-omics data,with high dimensions and low sample size,into prognostic assessment algorithms.The topology relationship among patients is mostly under consideration when adopting prognostic risk assessment algorithms for multi-omics data in current neural network-based researches,in which the noise caused by the graph structure in biological networks is ignored.In order to solve the above limitations,genes with robust and biological significance were screened in combination with tumor-related knowledge by statistical test,and carried out bioinformatic analysis combined with biological knowledge.Specifically,two types of tumors were screened based on the standard of 5-year survival rate,and genes,with different expression levels in tumor tissues and normal human tissues,were screened by combining generalized likelihood ratio test and multiple method,which were then enriched into pathways by hypergeometric test.Because there is path redundancy,and the relationship among different paths can be represented by a directed acyclic graph,this paper combines the advantages of both node-based and edge-based algorithms to calculate the similarity between paths and divide function sets.Through the biological analysis and verification of the functional set,it was confirmed that the screened genes have important biological significance in the process of tumor development.Secondly,on the basis of the gene set obtained by bioinformatics analysis,a new tumor prognosis risk assessment method is proposed,which combines graph neural network,convolutional neural network and various attention mechanisms,realizes the stratification of survival risk.The model mainly consists of three frameworks:(1)A graph neural network based on a dual attention mechanism,which calculates the different contributions of neighbor nodes of a node through the attention mechanism at the node level,and performs weighted aggregation on the embedded representation of each node;The similarity relationship can be aggregated through the association layer attention mechanism among nodes.Then,the network topology among nodes can be learned,and the importance relationship of each node and the correlation between nodes and their neighboring nodes can be fully considered.Consequently,a variety of information is aggregated to successfully learn a robust semantic representation(2)The multi-omics attention mechanism is combined with the convolutional neural network.At first,the convolution operation is performed on the omics features respectively to fully learn the internal representation of the omics features.Then the omics attention mechanism is applied to weigh the contribution of different omics to the prognosis risk assessment.Finally,a global omics semantic representation is obtained by convolution fusion with the omics semantic information.(3)To better take advantage of these two prognostic models,we propose a prognostic model based on graph importance embedding and omics importance embedding.The experiment is trained on the multiomics data and survival data of pancreatic cancer through ten-fold cross-validation.The results indicate that the method takes the best effect,applicable to divide the survival subtypes of pancreatic cancer.