tensorOmics: Data integration for longitudinal omics data using tensor factorisation

The paper introduces **tensorOmics**, an R package that utilizes tensor factorization to provide a unified framework for both supervised and unsupervised integration of longitudinal multi-omics data, preserving temporal structures and multi-layer correlations that traditional matrix-based methods often fail to capture.

The Gist

The Problem: The "Flattened Movie" Mistake

Imagine you are trying to understand a complex, high-stakes drama—like a long-running TV series. To understand the story, you need to know three things:

1. Who the characters are (the biological molecules).
2. What they are doing (their activity levels).
3. When things happen (the timeline of the story).

Currently, most scientists trying to study biology over time make a mistake: they take the entire TV series, chop it up into individual still photos, and throw them all into one giant pile. This is called "flattening" the data.

When you do this, you lose the "plot." You might see that a character is crying in one photo and laughing in another, but you’ve lost the connection that shows they were crying because of something that happened in the previous scene. In biology, if you flatten your data, you lose the "temporal trajectory"—the beautiful, flowing story of how a body responds to a drug or a disease over time.

The Solution: tensorOmics (The Ultimate Movie Director)

The researchers created a new tool called tensorOmics. Instead of turning the movie into a pile of photos, tensorOmics treats the data like a high-definition film. It keeps the "three-way structure" intact: Who (the molecules), What (the measurements), and When (the time).

Think of tensorOmics as a master film editor who can do two incredible things:

1. The Detective (Supervised Analysis): It can look at the footage and say, "Aha! This group of characters is acting differently because they are in the 'Action Movie' genre (the treatment group), while these others are in a 'Romance' (the control group)." It can specifically hunt for the scenes that distinguish one group from another.
2. The Orchestrator (Multi-Omics Integration): In a real movie, you have actors, lighting, music, and costumes all working together. In biology, you have different "layers" of data (like DNA, proteins, and bacteria). tensorOmics doesn't just look at the actors; it looks at how the music and the lighting change at the exact same moment as the actors' emotions. It finds the "coordinated response"—the moment where the music swells because the actor started crying.

Why Does This Matter?

The researchers tested their "film editor" on three very different "movies":

• Antibiotic recovery: Watching how the human body's internal ecosystem recovers after being hit by drugs.
• Anaerobic digestion: Watching how complex biological systems process waste.
• Fecal transplants: Watching how a change in gut bacteria affects a whole system.

In all these cases, tensorOmics succeeded where older methods failed. It didn't just see that things changed; it saw how they changed over time and how different biological layers "danced" together to create a response.

The Bottom Line

tensorOmics is a new digital toolkit for scientists. It allows them to stop looking at biology as a collection of snapshots and start seeing it as a continuous, multi-layered movie. By preserving the "plot" of time, scientists can better understand how diseases progress and how treatments actually work.

Technical Summary

Technical Summary: tensorOmics

1. Problem Statement

The integration of multi-omics data (e.g., transcriptomics, proteomics, metabolomics) is essential for understanding complex biological systems. However, current computational approaches face two critical limitations when dealing with longitudinal data (measurements taken from the same subjects over multiple time points):

• Loss of Temporal Structure: Traditional integration methods rely on matrix-based techniques (e.g., PCA, PLS). To apply these to longitudinal data, researchers often "flatten" the data into a 2D matrix (concatenating time points into features). This process destroys the inherent multi-way structure, obscures temporal trajectories, and violates the statistical assumption of independence between observations.
• Lack of Integrated Supervised Learning: Existing tensor-based methods are largely unsupervised, meaning they can find patterns but cannot effectively use phenotypic information (e.g., treatment vs. control) to drive the discovery of discriminant molecular signatures. Furthermore, there is a lack of unified frameworks that can simultaneously handle multiple omics layers and temporal dimensions.

2. Methodology

The authors propose tensorOmics, a computational framework that utilizes tensor factorization to preserve the natural three-way structure of longitudinal data: $\text{Samples} \times \text{Features} \times \text{Time}$.

The framework is built on tensor decomposition, which decomposes high-dimensional arrays into low-rank components that represent latent biological processes. The methodology is divided into two main settings:

A. Single-Omic Setting:

• Unsupervised: Uses Tensor PCA to identify the primary axes of variation across samples, features, and time.
• Supervised: Uses Tensor PLS Discriminant Analysis (Tensor PLS-DA) to find molecular signatures that maximize the separation between predefined phenotypic groups (e.g., treated vs. untreated).

B. Multi-Omic Setting (Multi-block Analysis):

• Unsupervised: Uses Tensor PLS and Block Tensor PLS to integrate multiple omics layers, identifying coordinated patterns that exist across different biological layers (e.g., how a change in the microbiome correlates with a change in the metabolome over time).
• Supervised: Uses Block Tensor PLS Discriminant Analysis to identify multi-layer molecular signatures that are most predictive of a specific biological outcome or treatment response.

3. Key Contributions

• Preservation of Multi-way Structure: Unlike matrix methods, tensorOmics treats time as a distinct dimension, allowing for the modeling of temporal dynamics and repeated measurements.
• Unified Framework: It provides a single ecosystem that bridges the gap between single-omic and multi-omic analysis, and between unsupervised exploration and supervised discrimination.
• Dimensionality Reduction & Compression: The tensor-based approach inherently manages the "curse of dimensionality" by compressing high-dimensional omics data into interpretable, low-rank latent factors.
• Software Implementation: The framework is implemented as an R package, making it accessible for the broader biological research community.

4. Results and Validation

The framework was validated through three diverse biological case studies:

1. Antibiotic Perturbation: Demonstrated the ability to distinguish between treatment groups and track the recovery of molecular profiles over time.
2. Anaerobic Digestion Systems: Showed how the method captures complex, time-dependent shifts in microbial and chemical environments.
3. Fecal Microbiota Transplantation (FMT): Revealed coordinated molecular responses across different omic layers that were missed by traditional cross-sectional (non-temporal) methods.

In all cases, tensorOmics successfully identified biologically meaningful temporal signatures and demonstrated superior performance in capturing the "co-movement" of different molecular layers over time.

5. Significance

The significance of tensorOmics lies in its ability to provide a mathematically rigorous way to study biological trajectories. By moving beyond "snapshots" of biological states and instead modeling the "flow" of molecular changes, researchers can better understand the causal and temporal relationships within complex systems. This is particularly vital for precision medicine, where understanding how a patient's molecular profile evolves in response to a drug is critical for determining treatment efficacy and timing.