[Paper Review] Non-iterative Joint and Individual Variation Explained

Qing Feng, Jan Hannig+1 more|arXiv (Cornell University)|Dec 13, 2015

Gene expression and cancer classification13 references5 citations

TL;DR

This paper introduces Non-iterative JIVE, a fast, non-iterative method for decomposing multiple data blocks into joint and individual variation components using score subspaces and perturbation theory. It achieves exact decomposition without normalization, improves robustness to data heterogeneity, and provides a 16x speedup over JIVE while ensuring identifiability and theoretical guarantees via singular value decomposition and row space analysis.

ABSTRACT

Integrative analysis of disparate data blocks measured on a common set of experimental subjects is one major challenge in modern data analysis. This data structure naturally motivates the simultaneous exploration of the joint and individual variation within each data block resulting in new insights. For instance, there is a strong desire to integrate the multiple genomic data sets in The Cancer Genome Atlas (TCGA) to characterize the common and also the unique aspects of cancer genetics and cell biology for each source. In this paper we introduce Non-iterative Joint and Individual Variation Explained (Non-iterative JIVE), capturing both joint and individual variation within each data block. This is a major improvement over earlier approaches to this challenge in terms of a new conceptual understanding, much better adaption to data heterogeneity and a fast linear algebra computation. Important mathematical contributions are the use of score subspaces as the principal descriptors of variation structure and the use of perturbation theory as the guide for variation segmentation. This leads to a method which is robust against the heterogeneity among data blocks without a need for normalization. An application to TCGA data reveals different behaviors of each type of signal in characterizing tumor subtypes. An application to a mortality data set reveals interesting historical lessons.

Motivation & Objective

• To address the limitations of iterative, normalization-dependent JIVE methods in integrating multiple data blocks.
• To provide a theoretically grounded, non-iterative algorithm that ensures identifiability of joint and individual variation components.
• To improve robustness to data heterogeneity by avoiding arbitrary normalization procedures.
• To establish a new conceptual framework for variation decomposition using score subspaces and perturbation theory.
• To enable efficient, scalable analysis of complex multi-omics and heterogeneous data sets, such as TCGA and historical mortality data.

Proposed method

• Uses row spaces (score subspaces in ℝⁿ) as the primary descriptors of variation, focusing on patterns across data objects like patients.
• Applies perturbation theory to quantify noise effects and guide segmentation of joint from individual variation.
• Defines the joint score subspace as the intersection of all individual data block row spaces: row(J) = ⋂ₖ row(Aₖ).
• Employs singular value decomposition (SVD) on a concatenated matrix M of estimated right singular vectors to extract joint components.
• Uses sequential optimization via SVD to identify joint components by maximizing Frobenius norm projections onto rank-1 subspaces orthogonal to previous components.
• Eliminates tuning parameters for joint component selection through theoretical justification via perturbation bounds.

Experimental results

Research questions

• RQ1 Can joint and individual variation be decomposed in a way that is both identifiable and computationally efficient?
• RQ2 How can perturbation theory be leveraged to distinguish true joint signals from noise in multi-block data?
• RQ3 Is it possible to eliminate the need for data normalization while maintaining robustness to heterogeneity in data scale and dimensionality?
• RQ4 What is the theoretical relationship between the estimated score subspaces and the true underlying joint variation structure?
• RQ5 How does the non-iterative approach compare in speed and accuracy to existing iterative methods like JIVE and O2-PLS?

Key findings

• The joint score subspace is uniquely defined as the intersection of all individual data block row spaces, ensuring identifiability.
• The method achieves a 16x speedup over the original JIVE algorithm when applied to real TCGA data.
• Theoretical bounds show that the first rJ singular values of the concatenated matrix M satisfy σ²ⱼ ≥ ∑ₖ₌₁ᴷ cos²θₖ, confirming joint component detection.
• The approach eliminates the need for arbitrary normalization, improving robustness to data heterogeneity across blocks.
• Applications to TCGA data reveal distinct signal behaviors in characterizing tumor subtypes, while mortality data reveal historical patterns.
• The method provides statistical guarantees for joint component selection through perturbation-theoretic justification, removing reliance on tuning parameters.