Research On Anti-Tumor Drug Response Prediction Method Based On Knowledge Graph Embedding

Cancer is a complex disease,and due to the heterogeneity of tumours and individual differences in cancer patients,the same anti-cancer drug has a wide range of therapeutic effects on different patients with the same type of cancer.Providing individualised treatment to each patient can effectively increase the likelihood of recovery.Studying the response of tumour cells to anti-tumour drugs is key to the treatment of cancer and the development of novel anti-cancer drug candidates,which play a crucial role in personalised cancer treatment.Currently,data on gene expression profiles,gene regulation,copy number variation and other data on the response of drugs to tumour cells in a growing number of human cancer cell lines provide the basis for the development of drug response prediction models.In this paper,we propose a knowledge graph-based embedding method for anti-tumour drug response prediction.First,a tumour cell-gene fusion network is constructed by fusing tumour cells with gene regulatory networks based on tumour cell gene expression profiles.Then,based on the resulting tumour cell-gene fusion network,a knowledge graph embedding(KGE)technique is used to learn a low-dimensional feature vector of tumour cells.Finally,the combined low-dimensional feature vector of tumour cells and molecular fingerprints of drugs were used as input features,and the tumour cell-drug response data were used as labels to train machine learning-based drug response prediction models for predicting tumour cell-drug responses.In this paper,two types of models,binary classification and regression,are trained separately.For binary classification,logistic regression models,random forest binary classification models,support vector machine binary classification models and deep neural network binary classification models are trained;for regression,linear regression models,random forest regression models,support vector regression models and deep neural network regression models are trained.In terms of experimental validation of the method in this paper,the good drug response prediction capability of the proposed method was firstly illustrated through several experiments such as five-fold cross-validation,comparison with original gene expression prediction performance,comparison with hotspot gene prediction performance and comparison with prediction performance of existing methods.And then validated through power-law distribution analysis of tumour cell-gene fusion network,verification of the learning capability of knowledge graph embedding,low-dimensional The effectiveness of this paper’s method in extracting gene regulatory information was then validated by experiments such as power-law distribution analysis of tumor cell-gene fusion networks,verification of the learning ability of knowledge graph embedding,visualization of low-dimensional feature vectors,and literature validation of new drug-tumor cell sensitivity relationships.Finally,the effect of different training sample sizes on the prediction accuracy of the model and the effect of changing the gene regulation network on the prediction accuracy of the model are analysed.The knowledge graph embedding-based drug response prediction method can effectively extract the gene interaction information in the gene regulatory network to improve the accuracy of drug response prediction.The method in this paper reduces the dimensionality of tumour cell genomic features from 17419 to 128 dimensions,and also provides a new solution to the dimensional catastrophe of genomics data.