B0512
Title: Unsupervised learning approaches for multi-OMICS data
Authors: Marina Evangelou - Imperial College London (United Kingdom) [presenting]
Abstract: It is increasingly common these days for biomedical studies to generate multiple OMICS datasets for the same individuals. The conventional approaches for understanding the relationships between the OMICS datasets and the complex traits of interest (e.g. diseases) would be through the analysis of each dataset separately from the rest. Similarly, if researchers are interested in understanding the relationships between the OMICS datasets, they will perform pairwise tests with the features of the two OMICS datasets. It is illustrated that integrating multiple OMICS datasets improves understanding of their in-between relationships and improves their predictive performance. Two alternative data integration approaches will be presented: an extension of sparse canonical correlation analysis (sCCA) for the integration of multiple (more than 2) OMICS datasets. Although sCCA is an unsupervised learning approach, it is illustrated that by including the response variable as one of the datasets the predictive performance is increased. The second approach presented, named multi-SNE, is an extension of the well-known t-SNE approach for dimensionality reduction and visualisation of multi-view data. By incorporating the obtained low-dimensional embeddings of multi-SNE into the K-means clustering algorithm, it is shown that sample clusters are accurately identified.
