Independent Component Analysis for Unraveling the Complexity of Cancer Omics Datasets

Nicolas Sompairac, Petr V. Nazarov, Urszula Czerwińska, Laura Cantini, Anne Biton, Askhat Molkenov, Zhaxybay Zhumadilov, Emmanuel Barillot, François Radvanyi, Alexander N. Gorban, Ulykbek Kairov, Andreï ZinovyevInternational Journal of Molecular Sciences 2019 · DOI

TL;DR

This review highlights Independent Component Analysis (ICA) as a powerful matrix factorization technique for dissecting the complexity of cancer omics data. By treating biological signals as independent sources, ICA offers unique advantages over traditional methods like PCA for tasks such as data preprocessing, dimensionality reduction, and identifying functional biological subsystems, ultimately aiding in a deeper understanding of cancer.

What Problem Does This Paper Address?

Cancer research generates vast amounts of high-throughput molecular data (omics data) from various sources like transcriptomes, methylomes, and proteomes. Understanding the intricate biological processes underlying cancer from these complex, multi-layered datasets is a significant challenge. Raw omics data often contains noise, technical biases, and signals from different biological sources (like various cell types within a tumor) that are intertwined. Traditional methods like Principal Component Analysis (PCA) can reduce dimensionality but may not effectively disentangle these underlying biological signals. The paper addresses the need for robust analytical tools that can effectively "unmix" these complex signals, identify meaningful biological patterns, and provide interpretable insights into cancer biology.

Methods

The paper focuses on Independent Component Analysis (ICA), a matrix factorization technique. Unlike PCA, which aims to find orthogonal components that maximize variance, ICA seeks to find components that are statistically independent. This is based on the assumption that the observed omics data is a linear mixture of underlying independent sources.

Key aspects of ICA discussed include:

  • Mathematical Foundation: ICA models data $X$ as a sum of products of component vectors ($a_k$ and $s_k$), aiming to find these components by maximizing the statistical independence of the $s_k$ vectors.
  • Algorithms: The review mentions popular algorithms like fastICA and JADE, which often rely on maximizing the non-Gaussianity of the component distributions.
  • Preprocessing: A crucial step is "whitening" or "sphering" the data, which involves transforming the data to have unit variance and zero correlation, preparing it for ICA.
  • Comparison with other methods: ICA is compared to PCA and Non-Negative Matrix Factorization (NMF), highlighting ICA's strengths in interpretability and reproducibility, especially when components are stabilized.
  • Technical Considerations: The paper delves into practical aspects like data log-transformation (essential for RNA-Seq data), choosing whether to maximize the independence of metagenes (gene-related components) or metasamples (sample-related components), and determining the optimal number of components.

Datasets

The review discusses applications of ICA across various omics datasets, primarily focusing on cancer. While specific datasets used for training or benchmarking are not explicitly listed as new contributions in this review paper, it references numerous studies that have applied ICA to:

  • Transcriptomic data: From bulk tumors and single-cell RNA-Seq.
  • Methylome data: DNA methylation profiles.
  • Proteome data: Protein and phosphoform abundances.
  • Multi-omics datasets: Combinations of the above.

The paper also references several large public datasets and consortia, such as The Cancer Genome Atlas (TCGA) and the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC), which have been utilized in the studies reviewed.

Key Insights and Ideas

  • ICA as a "blind source separation" tool for biology: The core idea is that complex biological signals (like gene expression patterns, cell type contributions) are like independent "sources" that get mixed in observed data. ICA can "unmix" them, revealing these underlying biological signals.
  • Superior interpretability over PCA: While PCA finds components that explain the most variance, these components can be mixtures of many biological factors. ICA's independent components often correspond more directly to interpretable biological processes or cell types.
  • Reproducibility is key: Stabilized ICA (using methods like icasso or sub-sampling) leads to more reproducible results across different datasets, which is crucial for building reliable biological knowledge.
  • Deconvoluting cellular heterogeneity: ICA is powerful for unmixing bulk omics data to estimate the proportions and characteristics of different cell types within a tumor microenvironment, a critical step for understanding tumor complexity.
  • Functional subsystems as building blocks: ICA can identify "functional subsystems" (modules of genes/proteins) that act together in response to stimuli or disease states, offering a systems-level view of cellular processes.
  • Multi-omics integration potential: ICA can be extended to tensorial ICA, allowing for the simultaneous analysis of multiple omics data types, potentially revealing cross-level biological interactions.

Key Results

The review doesn't present new experimental results but synthesizes findings from numerous studies. Key qualitative results highlighted include:

  • ICA has been successfully applied for data cleaning and preprocessing, removing technical biases from datasets.
  • It's effective for dimensionality reduction, creating lower-dimensional representations that retain biological meaning.
  • ICA aids in unsupervised deconvolution of cell types, revealing immune cell infiltration patterns and other cellular contributions in tumors.
  • It facilitates data integration and meta-analysis, allowing for comparisons and consensus building across multiple studies.
  • ICA components have been shown to correlate with clinical outcomes and molecular subtypes of cancer.
  • Studies comparing ICA to PCA and NMF often show ICA (especially stabilized versions) providing more reproducible and interpretable metagenes.

Limitations and Future Work

  • Interpretation remains a challenge: While ICA components are often more interpretable than PCA, assigning definitive biological meaning to each component still requires significant effort, often involving functional enrichment analyses and biological network mapping.
  • Choice of parameters: Determining the optimal number of components and the appropriate non-linearity function for ICA can be non-trivial and dataset-dependent.
  • Assumptions: ICA relies on the assumption of statistical independence and non-Gaussianity of the underlying sources, which may not always hold perfectly in biological systems.
  • Future Directions:
    • Exploring the roots of ICA in artificial neural networks for novel autoencoding-based blind source separation techniques.
  • Developing hybrid learning approaches that combine ICA-extracted features with deep learning or other machine learning models.
  • Injecting formalized biological knowledge into ICA to better capture complex hierarchical organization of biological systems.
  • Further development and standardization of ICA applications for emerging omics data types (e.g., epigenomics, proteomics) and single-cell data.
  • Expanding ICA applications beyond transcriptomics to multi-omics integration and non-omics data (e.g., imaging, clinical records).

Summary updated

Abstract

Independent component analysis (ICA) is a matrix factorization approach where the signals captured by each individual matrix factors are optimized to become as mutually independent as possible. Initially suggested for solving source blind separation problems in various fields, ICA was shown to be successful in analyzing functional magnetic resonance imaging (fMRI) and other types of biomedical data. In the last twenty years, ICA became a part of the standard machine learning toolbox, together with other matrix factorization methods such as principal component analysis (PCA) and non-negative matrix factorization (NMF). Here, we review a number of recent works where ICA was shown to be a useful tool for unraveling the complexity of cancer biology from the analysis of different types of omics data, mainly collected for tumoral samples. Such works highlight the use of ICA in dimensionality reduction, deconvolution, data pre-processing, meta-analysis, and others applied to different data types (transcriptome, methylome, proteome, single-cell data). We particularly focus on the technical aspects of ICA application in omics studies such as using different protocols, determining the optimal number of components, assessing and improving reproducibility of the ICA results, and comparison with other popular matrix factorization techniques. We discuss the emerging ICA applications to the integrative analysis of multi-level omics datasets and introduce a conceptual view on ICA as a tool for defining functional subsystems of a complex biological system and their interactions under various conditions. Our review is accompanied by a Jupyter notebook which illustrates the discussed concepts and provides a practical tool for applying ICA to the analysis of cancer omics datasets.
BibTeX
@article{sompairac2019independent,
  author = {Nicolas Sompairac and Petr V. Nazarov and Urszula Czerwińska and Laura Cantini and Anne Biton and Askhat Molkenov and Zhaxybay Zhumadilov and Emmanuel Barillot and François Radvanyi and Alexander N. Gorban and Ulykbek Kairov and Andreï Zinovyev},
  title = {{Independent Component Analysis for Unraveling the Complexity of Cancer Omics Datasets}},
  journal = {International Journal of Molecular Sciences},
  year = {2019},
  doi = {10.3390/ijms20184414},
}
Figure