SC-AE: An Improved Spectral Clustering Unsupervised Feature Selection Algorithm Guided by Autoencoders Based on Pan-cancer Multi-view Omics Data

Omics data is characterized by high-dimensional small samples and contains a great number of gene features unrelated to diseases. Pan-cancer exhibits heterogeneity and can be subdivided into different subtypes. Identifying disease subtypes is pivotal for advancing precision medicine, and disease subtype analysis through omics data has become a popular method. Without knowing the sample labels, how to select the most relevant subset of features for pan-cancer subtypes becomes an immediate challenge. This paper proposes an improved spectral clustering unsupervised feature selection algorithm guided by autoencoders (SC-AE) based on pan-cancer multi-view omics data to solve it. In the feature selection stage of spectral clustering, this paper proposes the evaluation metrics of feature importance to measure whether the feature is important. It takes into account both feature discernibility and feature independence. Through the autoencoder, the low-dimensional and denoised latent space of the original features is learned as the target representation to further guide the selected feature subset. We propose new evaluation metrics. It evaluates whether a feature subset is an important feature based on both clustering quality and clinical significance. The algorithm has been used on multi-view data from five pan-cancer datasets. Compared with the current six mainstream unsupervised feature selection algorithms, the optimal feature subset obtained by SC-AE exhibits better clustering effects and has more realistic clinical significance.