Methodology Development - Sparse MFA-FABIA: A Hybrid Method for Detecting Weak Signals in Multi-Omic

Background and Motivation: Multi-omics integration holds immense potential for uncovering complex biological patterns. However, weak signals, subtle yet biologically essential variations, often remain hidden due to overwhelming high-signal features and noisy data structures. Conventional factorization methods like Multiple Factor Analysis, MFA [Abdi et al., 2007, Abdi et al., 2013] excel at balancing contributions across omics layers, while Factor Analysis for Bicluster Acquisition, FABIA [Hochreiter et al., 2010; Kasim et al., 2016] provides sensitivity to sparse, hidden biclusters. Each has unique strengths, but neither fully addresses the challenge of low-signal detection in high-dimensional, heterogeneous datasets. A hybrid approach, Sparse MFA-FABIA, offers a promising solution: leveraging MFA’s balanced global integration with FABIA’s sparse, factor-driven biclustering to detect signals that may otherwise be drowned out. This novel method could significantly improve the discovery of weak but biologically critical patterns, such as regulatory modules or rare disease markers.

Problem Statement: High-signal features (dominant clusters, strong differential patterns) often mask weak but meaningful signals. In multi-omics, this imbalance makes it difficult to: identify rare but essential biomarkers, capture subtle interactions across data layers and/or retain interpretability while controlling for noise. The lack of methods explicitly designed for weak signal recovery represents a critical methodological gap in integrative omics analysis.

Objectives

1. Develop Sparse MFA-FABIA: Create a hybrid model combining MFA’s cross-omics balancing with FABIA’s sparse bicluster extraction.
2. Enhance weak-signal detection: Implement sparsity-inducing penalties and Bayesian priors to isolate subtle signals obscured by dominant patterns.
3. Validate through simulations and case studies: Benchmark performance against existing methods on both simulated low-signal datasets and real biological data.
4. Extend method flexibility: Incorporate nonlinear transformations and weighted omics contributions to broaden applicability.

Proposed Methodological Approach: The proposed methodology combines MFA to align and normalize multi-omics contributions with FABIA-inspired sparse factorization to extract rare but structured biclusters while controlling for dominant signals. Sparsity penalties and Bayesian priors will enhance weak-signal detection and stabilize inference. Possibly, nonlinear extensions through kernelization and weighted contributions will capture complex interactions and balance omics influence. Validation will include simulations embedding weak signals, benchmarking against existing methods, and applications to real datasets such as cancer, rare disease, and single-cell omics.

References

1. Abdi, H., & Valentin, D. (2007). Multiple factor analysis (MFA). Encyclopedia of measurement and statistics, 657-663.
2. Abdi, H., Williams, L. J., & Valentin, D. (2013). Multiple factor analysis: principal component analysis for multitable and multiblock data sets. Wiley Interdisciplinary reviews: computational statistics, 5(2), 149-179.
3. Hochreiter, S., Bodenhofer, U., Heusel, M., Mayr, A., Mitterecker, A., Kasim, A., ... & Clevert, D. A. (2010). FABIA: factor analysis for bicluster acquisition. Bioinformatics, 26(12), 1520-1527.
4. Kasim, A., Shkedy, Z., Kaiser, S., Hochreiter, S., & Talloen, W. (2016). Biclustering for Cloud Computing. In Applied Biclustering Methods for Big and High-Dimensional Data Using R (pp. 389-398). Chapman and Hall/CRC.