Advancement of high-throughput technologies in omics studies had produced large amount of information that enables integrated analysis of complex diseases. Complex diseases such as cancer are often caused by a series of interactions that involve multiple biological mechanisms. Integration of multi-omics data allows more advanced analysis using features from various aspects of biology. However, analysing cancer multi-omics data on a large scale could be challenging due to the high dimensionality of the data. The recent development of advanced computational algorithms, especially deep learning, had sparkednumerous efforts in applying these algorithms in multi-omics studies. This study aims to investigate how deep learning algorithms, namely stacked denoising autoencoder (SDAE) and variational autoencoder (VAE) can be used in cancer classification using multi-omics data. Moreover, this study also investigates the impact of feature selection in multi-omics analysis through the implementation of an embedded feature selection. The multi-omics data used in this study includes genomics, methylomics, transcriptomics and clinical data for a case study of lung squamous cell carcinoma. The classification performance has beencompared and discussed in terms of the effectiveness of different models and the impact of feature selection. Results showed that VAE outperforms SDAE with 91.86% accuracy, 22.73% specificity and 0.21% Matthews Correlation Coefficient (MCC).