| Cancer diagnosis and treatment has always been a timeless topic in the medical field.According to a report from the World Health Organization,cancer has become the leading cause of premature death in 112 countries.Patients with the same cancer,due to their different genes,have vastly different responses to the same anti-cancer drugs.With the development of high-throughput sequencing technology,doctors can more quickly and comprehensively consider factors such as the patient’s genetic composition,thus quickly and effectively customizing medical plans.Due to the ultra-high dimensional cell line gene expression data and relatively small sample size,traditional machine learning algorithms do not perform well in solving this problem.In the field of deep learning,neural networks can fully grasp the intrinsic features of the data,model the original data,and to some extent make up for the shortcomings of traditional machine learning requiring high data features,so it is of great practical significance to study deep learning in the sensitivity of genes to anti-cancer drugs.Based on the above background,this paper takes breast cancer cell lines as an example.extracts 51 gene data of breast cancer cell lines and 295 drug structure data from GDSC and PubChem databases,and proposes corresponding improved algorithms based on the original model.In order to better improve the accuracy of the model,this paper introduces a spatial transformer network to improve the encoding module and decoding module in the deep autoencoder network,and replaces the original activation function.Experimental results demonstrate that the proposed model outperforms both the principal component feature extraction network and the network without feature extraction in terms of both accuracy and stability.The feature extraction is performed using a channel-wise concatenation method.Experimental results indicate that the stability of the model is significantly improved while maintaining the accuracy of the model.To address the one-dimensional nature of both cell line gene data and drug structure data,this paper employ s one-dimensional convolution kernels instead of two-dimensional ones.This approach not only effectively addresses the problem of mainstream networks being unable to handle one-dimensional biological data,but also reduces the model parameters by over 10 times.The proposed model is a multi-input AE-ResNet model,with the main part of the model being a one-dimensional residual network,which replaces the original identity modules with pre-activation modules.During the training process,various learning rate schedulers are introduced to select the best training model.Finally,this paper compares the proposed model with the latest two deep learning networks,and shows superior performance in multiple metrics. |
