Response Prediction For Cancer Treatment Based On Deep Learning

With the continuous development of medical technology in recent years,more and more patients are detected and diagnosed at the early stage of cancer,and the mortality rate of some common cancers is decreasing year by year.Since cancer treatment does not apply to universal treatment,the same drug may have different therapeutic effects for different patients.The need for personalized cancer drug response prediction has emerged.The development of new models around machine learning techniques to address drug response prediction has become a new hot topic of interest.This paper revolves around the problem of data shift between gene library data and clinical data in personalized drug response prediction tasks.Shapley Value is used to perform feature selection on differential gene features to enhance the generalization ability of the model trained.Parameter adaptive optimization mechanisms using evolutionary algorithms for optimization are also integrated to MR.DRP,a parameter adaptive machine learning framework for personalized drug response prediction.Details are as follows:This paper focuses for the first time on the key problem that drug response prediction tasks need to face when going to the clinic: data shift between gene library data and clinical data.A Shapley Value based feature selection is proposed to moderate the data shift phenomenon.And a series of analytical experiments were conducted on the Cisplatin drug dataset and the Paclitaxel drug dataset.The experimental results showed that the Shapley Value based feature selection method successfully focused on genes with distribution differences on the gene library data and clinical data.The covariance matrix adaptive evolution algorithm was then used to search the space of possible parameter solutions.The model was trained based on the optimal solutions obtained from the search and validated on the clinical data.The results show that the proposed parametric adaptive machine learning framework MR.DRP achieves the best prediction results compared with other traditional machine learning models.Finally,this paper presents an enrichment analysis of the most important part of the features of the network model and points out some of the important pathways that influence the judgment of drug response.