记录一下今天下午EMBL组的Britta Velten下山来给我们讲多组学因子分析的工具（MOFA）。此人是EMBL Huber组的博后，他们实验室主要进行多组学分析和统计计算。今天她讲的这个工具是他们实验室开发的，综合多组学数据进行分析，致力于找到因子的最优组合，以表征疾病。
Integration of data from different omics is challenging
Challenges:
- Heterogeneous data sets have to be modeled under different statistical assumptions
- Appropriate regularization strategies are required to avoid over-fitting
- Approaches based on marginal correlations suffer from spurious associations and a high multiple testing burden
- Missing data is commonly encountered, both in terms of single values as well as individual samples that miss a certain assay.
Exploratory analysis: From single- to multi-omics
Principle component analysis (PCA) decomposes high-dimensional data Y into a low-dimensional representation by uncorrelated factors.
An extension to multi-omic data requires:
- Disentangling variation between different data sets:
- Which axes of variation are unique to a single data set?
- Which capture co-variation between different biological layers?
- Accounting for heterogeneity of different data sets
- Characterization and annotation of the major axes of variation
Multi-Omics Factor Analysis (MOFA)
MOFA is a computational method for discovering the principal sources of variation in a multi-omics dataset. MOFA infers a set of (hidden) factors that capture biological and technical sources of variability across a joint low-dimensional representation of multiple data sets, thereby enabling a variety of downstream analyses, including factor annotation, data imputation and the detection of outlier samples.
Important properties of MOFA for multi-omic applications:
- Interpretability by two-level sparsity
- factor- and view-wise sparsity by an ARD prior
- factor- and feature-wise sparsity by a spike and slab prior
- Scalable inference using a variational framework
- Handling of missing values
- Modeling non-Gaussian views
References
Multi-Omics factor analysis disentangles heterogeneity in blood cancer