- This event has passed.
Heidelberg Kolloquium – Britta Velten, EMBL
11. December 2017 @ 16:15 - 17:00
Multi-Omics Factor Analysis for an unsupervised integrative analysis of heterogeneous molecular data
Thanks to technological advances it has become feasible to jointly prole multiple molecular layers of large cohorts of samples (e.g., patient tissues), with the aim of gaining a more comprehensive understanding of biological processes in health and disease. However, principled methods for integrative analysis of such multi-omic data remain lacking. Of particular interest are unsupervised methods that map out the major molecular \dimensions”, which could then be associated with phenotypes and clinically relevant traits.
We present Multi-Omics Factor Analysis (MOFA), a method to uncover the principal axes of variation across heterogeneous datasets. MOFA infers a low-dimensional data representation in terms of (latent) factors that capture biological as well as technical sources of variability across data modalities. Sparsity assumptions on the factor loadings enable decomposing variation into axes present in all, some, or single data types and promote interpretable factors that can directly be linked to molecular drivers. For a broad applicability to multi-omic studies, the statistical model underlying MOFA can accommodate dierent data types and handle missing data. Its inference algorithm scales to large sample sizes. Once learnt, the factors enable a variety of downstream analyses, including identication of sample subgroups, data imputation, and the detection of outlier samples.
We applied MOFA to data derived from 200 patients with chronic lymphocytic leukaemia, where somatic mutations, RNA expression, DNA methylation and ex-vivo drug response proles were measured. Here, MOFA captured the key drivers of interpatient variability, thereby enhancing data interpretation as well as prediction accuracy of clinical outcomes.