MM-PCA: Integrative Analysis of Multi-group and Multi-view Data

Jonatan Kallus, Patrik Johansson, Sven Nelander, Rebecka JörnstenarXiv 2019 · DOI · arXiv

Abstract

Data integration is the problem of combining multiple data groups (studies, cohorts) and/or multiple data views (variables, features). This task is becoming increasingly important in many disciplines due to the prevalence of large and heterogeneous data sets. Data integration commonly aims to identify structure that is consistent across multiple cohorts and feature sets. While such joint analyses can boost information from single data sets, it is also possible that a globally restrictive integration of heterogeneous data may obscure signal of interest. Here, we therefore propose a data adaptive integration method, allowing for structure in data to be shared across an a priori unknown subset of cohorts and views. The method, Multi-group Multi-view Principal Component Analysis (MM-PCA), identifies partially shared, sparse low-rank components. This also results in an integrative bi-clustering across cohorts and views. The strengths of MM-PCA are illustrated on simulated data and on 'omics data from The Cancer Genome Atlas. MM-PCA is available as an R-package.

Introduction

In modern bioinformatics and "big data" fields, researchers rarely deal with a single, clean spreadsheet. Instead, they face multi-group data (identical measurements across different cohorts) and multi-view data (different types of measurements, such as gene expression vs. DNA methylation, taken on the same subjects). The primary goal of data integration is to combine these blocks to uncover meaningful biological patterns.

However, existing methods often fall into two traps: they either assume a signal must be present in every data block to be considered "joint," or they treat anything not globally shared as noise unique to a single block. In reality, biological signals are often partially shared. For example, a specific genetic pathway might be active in two out of three patient studies, or a mutation might correlate with gene expression but not protein levels. MM-PCA (Multi-group Multi-view Principal Component Analysis) was designed to find these elusive, partially shared structures.

Schematic of data integration scenarios: (A) multi-group, (B) multi-view, (C) multi-group multi-view, and (D) the unified 3D perspective.

Methodology

MM-PCA treats data integration symmetrically using a framework called Augmented Multi-view Data (AMD). This approach views the problem as a large, sparse matrix factorization task where rows and columns are treated with equal importance.

Parsimonious Parametrization

Standard PCA requires "orthogonality" (loadings must be perpendicular), which creates a significant challenge for optimization. MM-PCA bypasses this by using Givens rotations. Instead of optimizing the matrix values directly, the algorithm optimizes the angles of rotation. This turns a constrained, non-linear problem into an unconstrained one that standard optimizers (like BFGS) can handle efficiently.

Four-Way Penalization and the "D Matrix"

The loss function includes four distinct penalties to achieve partial sharing, automatic rank selection, sparsity, and variable selection. A key innovation is the Augmented D Matrix, which provides a direct map of which biological signals belong to which cohorts and data types, serving as a "lookup table" for data sharing.

Experimental Results

The researchers validated MM-PCA using both synthetic simulations and real-world "omics" data from The Cancer Genome Atlas (TCGA).

Synthetic Simulations

In high-dimensional simulations, MM-PCA significantly outperformed existing methods like Collective Matrix Factorization (CMF) and Joint and Individual Variation Explained (JIVE) in identifying the correct number of signals and the specific directions of those signals.

TCGA Glioblastoma (GBM) Data

When applied to Glioblastoma data (3 patient cohorts and 2 data views), MM-PCA identified 24 effective components. It successfully recovered the known G-CIMP subtype (a specific methylation pattern) and correctly linked it to younger patients in an unsupervised way. Furthermore, the components found by MM-PCA were highly correlated with clinical features like survival time, patient sex, and "stemness," proving the method captures real biology rather than mathematical noise.

Conclusion

MM-PCA offers a robust framework for identifying partially shared structures in complex datasets. By treating groups and views identically and utilizing efficient rotation-based optimization, it enables more nuanced biological discovery than traditional integrative methods. While currently optimized for continuous data, future work will explore generalizations to binary and tensor data types.

Software Availability: MM-PCA is available as an R package on CRAN.

Summary updated

Abstract

Data integration is the problem of combining multiple data groups (studies, cohorts) and/or multiple data views (variables, features). This task is becoming increasingly important in many disciplines due to the prevalence of large and heterogeneous data sets. Data integration commonly aims to identify structure that is consistent across multiple cohorts and feature sets. While such joint analyses can boost information from single data sets, it is also possible that a globally restrictive integration of heterogeneous data may obscure signal of interest. Here, we therefore propose a data adaptive integration method, allowing for structure in data to be shared across an a priori unknown \emph{subset of cohorts and views}. The method, Multi-group Multi-view Principal Component Analysis (MM-PCA), identifies partially shared, sparse low-rank components. This also results in an integrative bi-clustering across cohorts and views. The strengths of MM-PCA are illustrated on simulated data and on 'omics data from The Cancer Genome Atlas. MM-PCA is available as an R-package. Key words: Data integration, Multi-view, Multi-group, Bi-clustering
BibTeX
@misc{kallus2019mm,
  author = {Jonatan Kallus and Patrik Johansson and Sven Nelander and Rebecka Jörnsten},
  title = {{MM-PCA: Integrative Analysis of Multi-group and Multi-view Data}},
  year = {2019},
  eprint = {1911.04927},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.04927},
}
Figure