基于多组学融合和对抗自编码器的生存分析模型

A Survival Analysis Model Based on Multi-Omics Integration and Adversarial Autoencoder

摘要: 多组学整合分析可以利用不同组学之间的互补信息，有利于系统全面地理解癌症疾病的分子生物学机制。多组学数据的高维小样本属性，导致传统的生存分析模型存在严重的过拟合问题。深度学习模型可以从高维数据中进行自动特征提取，在处理复杂的多组学数据方面具有显著优势。为了有效地整合多组学数据，本文提出了基于对抗自编码器的多组学特征提取网络。结合1D-CNNCox生存分析模型，构建了基于多组学融合和生成对抗网络的GAN-1DCCox模型。在8种不同癌症类型的TCGA数据集上进行了消融和对比实验，相比流行的生存分析基准模型，GAN-1DCCox模型取得了更高的C指数值。结果表明GAN-1DCCox模型能够有效地融合多组学数据，筛选出重要的预后特征基因，提升了模型的生存预测性能和稳健性。

Abstract: Multi-omics integration analysis can utilize complementary information from different omics, beneficial for a more systematic and comprehensive understanding of the molecular biology mechanisms of cancer diseases. The high-dimension small-sample size of multi-omics data leads to serious overfitting issues in traditional survival analysis models. Deep learning models can automatically extract features from high-dimensional data and have significant advantages in processing complex multi-omics data. In this study, we proposed a survival analysis model based on multi-omics integration and adversarial autoencoder, called GAN-1DCCox model, which consists of a multi-omics feature extraction module based on generative adversarial networks and a 1D-CNNCox survival analysis module. GAN-1DCCox model achieved the highest C-index values in both ablation and comparative experiments on TCGA datasets of 8 different cancer types. It indicates that GAN-1DCCox model can effectively integrate multi-omics data and screen out important prognostic signature genes, and thereby improving the prediction performance and robustness of survival analysis model.