Study Of Non-small-cell Lung Cancer Prediction With Deep Neural Network

Non-small-cell lung cancer(NSCLC)is the most common lung cancer with high incidence and mortality.It is significant for improving the prognosis of NSCLC patients that the survival status and molecular subtypes can be predicted precisely and provide guidance for clinicians to develop personalized treatment plans.With the rapid development of highthroughput technologies,a large amount of clinical and multi-omics data of NSCLC patients have been collected.This data makes it possible to conduct extensive prediction studies on NSCLC.It turns out to be efficient to improve the predicting of survival status and molecular subtypes for NSCLC patients by integrating clinical and multi-omics data.Due to the high feature dimensionality and heterogeneity of omics data,it is difficult to extract and fuse omics features.A shortcoming of previous researches for NSCLC prediction is that only a small number of the relevant features are utilized and the high heterogeneity of omics data is ignored.The deep neural network is characterized by having the powerful ability of feature extraction and feature fusion.Therefore,this paper studies NSCLC survival classification and molecular subtype identification based on deep neural networks.The main research contents and results are as follows:The datasets for NSCLC survival classification and molecular subtype identification are constructed.The TCGA database is selected as the data source after comparing the NSCLC related databases.The clinical,gene expression,mi RNA,copy number variations,and DNA methylation profiles of 1089 NSCLC patients are collected from TCGA.In regard to the differences between various types of data and the requirements of two prediction tasks,the data are cleaned and the omics sequences are converted into the single-channel matrices by format transformation and feature normalization for the feature extraction.The survival classification dataset and molecular subtype identification dataset are constructed after data screening and grouping.A deep neural network-based survival classification model is designed.To obtain the abstract feature representation of omics data,a feature extraction module based on the convolutional neural network is proposed.By integrating the self-attentive mechanism and deep neural network,a survival classification network(ADNNC)is developed to improve the data fusion efficiency.Compared with three commonly used classification algorithms such as XGBoost,The ADNNC improves the accuracy of the NSCLC survival classification by at least 7.5% among testing dataset.A deep neural network-based molecular subtype recognition model is designed.For the problem of the high heterogeneity among omics data,a molecular subtype recognition network(ADNNT)is proposed.ADNNT is built on the self-attentive mechanism,the deep neural network and the k-mean clustering,which could enhance the utilization of omics features.Compared with commonly used clustering algorithms such as the Gaussian mixture model,the results presented by ADNNT show that the samples after clustering have larger inter-class distance and smaller intra-class distance.And the clustering results show that ADNNT can subdivide two pathological subtypes into five molecular subtypes.This work might help clinicians to develop differentiated treatment strategies and improve the prognosis for patients with specific molecular subtypes NSCLC.