Optimal transport-based machine learning to match specific patterns: application to the detection of molecular regulation patterns in omics data

Thi Thanh Yen Nguyen, Warith Harchaoui, Lucile Mégret, Cloé Mendoza, Olivier Bouaziz, Christian Néri, Antoine ChambazJournal of the Royal Statistical Society Series C (Applied Statistics) 2024 · DOI · arXiv

Abstract

We present several algorithms designed to learn a pattern of correspondence between 2 data sets in situations where it is desirable to match elements that exhibit a relationship belonging to a known parametric model. In the motivating case study, the challenge is to better understand micro-RNA regulation in the striatum of Huntington’s disease model mice. The algorithms unfold in 2 stages. First, an optimal transport plan P and an optimal affine transformation are learned, using the Sinkhorn–Knopp algorithm and a mini-batch gradient descent. Second, P is exploited to derive either several co-clusters or several sets of matched elements. A simulation study illustrates how the algorithms work and perform. The real data application further illustrates their applicability and interest.

Introduction: The Challenge of Cross-Omics Matching

In large-scale biological data analysis, such as transcriptomics, researchers frequently attempt to match different molecular types to uncover regulatory mechanisms. A prominent example is the interaction between micro-RNAs (miRNAs) and messenger-RNAs (mRNAs). Typically, miRNAs act as negative regulators; when a miRNA is highly expressed, its target mRNA is degraded, leading to an inverse relationship in their expression profiles.

Figure 1: Expression profiles of Ahrr (mRNA) and Mir20b (miRNA) showing the inverse relationship across age and genetic severity (Poly Q length).

However, biological systems are rarely perfectly linear. Noise, threshold effects, and complex transformations often obscure these relationships, making traditional tools like Pearson correlation insufficient. The goal of this research is to develop a method that can learn these parametric relationships even when they are shifted or scaled.

Methodology: Weighted Transformation Optimal Transport (WTOT)

The researchers introduced WTOT (Weighted Transformation Optimal Transport), a two-stage framework designed to bridge disparate datasets:

  1. Learning the Transformation: Instead of assuming a fixed relationship, the algorithm learns an optimal transport plan (mapping mRNA to miRNA) and an optimal affine transformation simultaneously. This is achieved using the Sinkhorn-Knopp algorithm for efficient transport calculation and mini-batch gradient descent for the mapping.
  2. Pattern Extraction: The resulting transport plan is used to identify co-clusters (groups of molecules with similar behavior) or direct matches between specific miRNAs and mRNAs.

A key innovation is the "Weighted" aspect, which allows the algorithm to ignore "irrelevant" data points—biological noise that does not participate in the regulation—thereby focusing on the most significant signals.

Results and Biological Discovery

The WTOT framework was tested on synthetic data and a large-scale Huntington’s Disease (HD) mouse model dataset, comprising over 13,000 mRNA and 1,100 miRNA profiles across various ages and genetic severities.

Figure 2: Learned centroids representing the primary molecular regulation patterns identified by the WTOT algorithm.

Key Findings:

  • Peaked Profiles: WTOT excelled at identifying "peaked" expression profiles—where activity spikes or dips at specific time points—whereas traditional methods often only capture simple monotonic trends. For instance, centroid 4 in the results demonstrates a clear non-monotonic pattern.
  • Novel Pathways: In the HD dataset, WTOT identified 1,247 high-confidence matches. Crucially, it highlighted biological signatures related to extracellular matrix organization and motile cilium functions. These areas are increasingly linked to neurodegeneration but were missed by standard algorithms like WGCNA.

Conclusion and Future Work

WTOT provides a robust mathematical framework for discovering hidden regulatory patterns in complex omics data. While challenges remain regarding global convergence and handling extreme noise, the method's ability to learn the relationship between datasets offers a significant advantage over static matching techniques. Future work will expand this analysis to other brain regions and refine simulation models to better capture biological complexity.

Summary updated

Abstract

Abstract We present several algorithms designed to learn a pattern of correspondence between 2 data sets in situations where it is desirable to match elements that exhibit a relationship belonging to a known parametric model. In the motivating case study, the challenge is to better understand micro-RNA regulation in the striatum of Huntington’s disease model mice. The algorithms unfold in 2 stages. First, an optimal transport plan P and an optimal affine transformation are learned, using the Sinkhorn–Knopp algorithm and a mini-batch gradient descent. Second, P is exploited to derive either several co-clusters or several sets of matched elements. A simulation study illustrates how the algorithms work and perform. The real data application further illustrates their applicability and interest.
BibTeX
@article{nguyen2024optimal,
  author = {Thi Thanh Yen Nguyen and Warith Harchaoui and Lucile Mégret and Cloé Mendoza and Olivier Bouaziz and Christian Néri and Antoine Chambaz},
  title = {{Optimal transport-based machine learning to match specific patterns: application to the detection of molecular regulation patterns in omics data}},
  journal = {Journal of the Royal Statistical Society Series C (Applied Statistics)},
  year = {2024},
  doi = {10.1093/jrsssc/qlae005},
}
Figure