📊 statistics

Making Multi-Axis Models Robust to Multiplicative Noise: How, and Why?

This paper introduces MED-MAGMA, a graph-learning algorithm designed to robustly fit multi-axis models corrupted by multiplicative noise, demonstrating superior performance in learning network structures across single-cell RNA sequencing datasets compared to prior methods.

Original authors: Bailey Andrew, David R. Westhead, Luisa Cutillo

Published 2026-03-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bailey Andrew, David R. Westhead, Luisa Cutillo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, two-dimensional puzzle. On one side of the puzzle, you have cells (like different types of people in a city). On the other side, you have genes (like the specific rules or habits those people follow). Your goal is to figure out how these cells and genes influence each other to create a map of relationships—a "network."

In the world of biology, specifically with single-cell RNA sequencing, scientists try to build this map. But there's a huge problem: the data they collect is incredibly "noisy."

The Problem: The "Flickering Lightbulb" Effect

Imagine you are trying to take a photo of a room, but the lightbulbs in the room are flickering wildly. Sometimes a light is bright, sometimes dim, not because the objects in the room changed, but because the light source itself is unstable.

In biology, this "flickering" is called multiplicative noise.

The Reality: A cell might have a lot of a specific gene.
The Noise: Due to technical glitches in the machine measuring it, the machine might only catch 10% of that gene's signal, or 90% of it, randomly.
The Result: The data looks like a distorted version of reality. If you try to draw a map based on this distorted data, you'll connect the wrong things together.

Most existing computer programs (algorithms) used to draw these maps assume the data is clean, or they assume the noise is just random static (like white noise on a radio). But this "flickering lightbulb" noise is different; it multiplies the true signal by a random number. Standard tools get confused and draw messy, inaccurate maps.

The Solution: MED-MAGMA (The "Noise Filter")

The authors of this paper, Bailey Andrew and his team, built a new tool called MED-MAGMA. Think of it as a super-smart photo editor that knows exactly how to fix "flickering lightbulb" photos.

Here is how it works, using a simple analogy:

1. The "Shape vs. Size" Trick

Imagine you have a rubber sheet with a drawing on it. Someone stretches the sheet unevenly—pulling it wider in some spots and squishing it in others. The shape of the drawing (the relationships between the lines) stays the same, but the size of the distances between points changes wildly.

Old Tools: Tried to measure the exact distances between points. When the sheet was stretched, they got the wrong answer.
MED-MAGMA: Realizes that the stretching (the noise) is just a distraction. It ignores the absolute size of the numbers and focuses entirely on the relative shape and direction. It asks, "Does point A move in the same direction as point B?" rather than "How far apart are they?"

By stripping away the "stretching" (the noise) and keeping only the "shape" (the true biological relationships), MED-MAGMA can see the true network underneath the mess.

2. The "Flip-Flop" Dance

To do this mathematically, the algorithm uses a clever trick called "flip-flopping."

Imagine you are trying to tune a radio with two knobs: one for the Cell signal and one for the Gene signal.
You can't tune both perfectly at once because they are tangled together.
So, you tune the Cell knob while holding the Gene knob still. Then, you tune the Gene knob while holding the Cell knob still.
You keep flipping back and forth. With every pass, the picture gets clearer until the noise is gone and the true signal is sharp.

Why Does This Matter?

The authors tested this new tool on 30 different real-world biological datasets (like a library of different cell types). They compared it to the best existing tools.

The Result: MED-MAGMA consistently found better, more accurate maps.
The Proof: When they looked at the maps, the groups of cells that should be together (like "cancer cells" vs. "healthy cells") were actually grouped together. The old tools often mixed them up.
The "Tail" Connection: The paper also notes that biological data has "extreme" behaviors (like a gene suddenly going into overdrive). Old tools assume everything is average and smooth. MED-MAGMA is flexible enough to handle these wild, extreme spikes, which are actually very important for understanding diseases.

The Big Picture

Think of MED-MAGMA as a pair of noise-canceling headphones specifically designed for biology.

Before: Scientists were listening to a choir where everyone was singing at different, random volumes. They couldn't tell who was singing what.
Now: MED-MAGMA filters out the volume differences and lets the scientists hear the harmony. This allows them to understand how cells talk to each other and how genes control life, leading to better ways to diagnose and treat diseases.

The authors have even made this tool available as a free software package, so other scientists can use it to clean up their own noisy data and discover new biological secrets.

1. Problem Statement

The paper addresses the challenge of fitting multi-axis (Kronecker-sum-structured) graphical models to matrix-variate data corrupted by multiplicative noise.

Context: In applications like single-cell RNA sequencing (scRNA-seq), data is represented as a matrix where rows are cells and columns are genes. The goal is to simultaneously learn dependency networks (graphs) for both rows ( $\Psi_{rows}$ ) and columns ( $\Psi_{cols}$ ).
The Noise Problem: scRNA-seq data is inherently noisy due to technical biases (e.g., GC-bias, incomplete capture efficiency). This manifests as multiplicative noise, where the observed value $x_{ij}$ is the product of a latent true value $z_{ij}$ and strictly positive scaling factors $r_i^{rows}$ and $r_j^{cols}$ (i.e., $x_{ij} = r_i^{rows} r_j^{cols} z_{ij}$ ).
Limitations of Existing Methods:
- Standard multi-axis models (e.g., GmGM) assume Gaussian noise or rely on "Gaussian copula" assumptions.
- Gaussian copulas restrict dependencies, failing to capture tail dependence (extreme events occurring together), which is prevalent in biological data.
- No existing method can fit Kronecker-sum-structured models under a general multiplicative noise regime without making restrictive distributional assumptions.

2. Methodology: MED-MAGMA

The authors propose MED-MAGMA (Matrix Elliptical Distribution's Multi-Axis Graphical Modelling Algorithm). The core innovation is modeling the data as a generalization of elliptical distributions to the matrix-variate case, allowing for arbitrary strictly positive multiplicative noise.

Model Formulation

The model assumes:

Latent Data: $z \sim \mathcal{N}(0, \Omega^{-1})$ , where $\Omega = \Psi_{cols} \oplus \Psi_{rows}$ (Kronecker sum).
Observation: $x = (r_{cols} \otimes r_{rows}) \circ z$ , where $\circ$ denotes the Hadamard (element-wise) product, and $r$ represents the noise scaling factors.
Identifiability: The model is non-identifiable regarding the scale of $\Omega$ and $r$ . The authors impose geometric mean constraints ( $\prod r_i = 1$ ) to resolve this, focusing on the sparsity structure of $\Omega$ rather than its scale.

Algorithmic Steps

The algorithm uses Expectation-Maximization (EM) combined with specific transformations to handle the noise:

Denoising Transformation ( $f$ ):
Since the likelihood of the mixture distribution is intractable, the authors define a function $f(x)$ that maps the data to a "quotient space," effectively removing the multiplicative noise while preserving the orientation of the latent vector.
$f(X)_{ij} = \frac{x_{ij}}{(\prod_{i'j'} x_{i'j'})^{\frac{1}{d_{rows}d_{cols}}} (\prod_{j'} x_{ij'})^{\frac{1}{d_{cols}}} (\prod_{i'} x_{i'j})^{\frac{1}{d_{rows}}}}$
This function is proven to be invariant to the multiplicative noise factors.
E-Step (Expectation):
The algorithm computes the expected sufficient statistics ( $S_{rows}$ and $S_{cols}$ ) of the latent variable $z$ given the observed transformed data $y = f(x)$ .
- Because the integral over the "fiber space" (the set of all $z$ mapping to $y$ ) is intractable, the authors use Laplace's approximation.
- They find the point $z^*$ in the fiber space that minimizes the quadratic form $z^T \Omega^{(t)} z$ . This is solved via a flip-flop quadratic programming approach over the noise parameters $r_{rows}$ and $r_{cols}$ .
- The expectation is approximated as $g(z^*) + \text{correction term}$ , where $g$ represents the sufficient statistics.
M-Step (Maximization):
The maximization step reduces to the same optimization problem as the GmGM algorithm (a fast Kronecker-sum Gaussian graphical model estimator), but using the approximated sufficient statistics $S_{rows}$ and $S_{cols}$ instead of raw data moments.

3. Key Contributions

Theoretical Extension: Generalizes multi-axis graphical models from Gaussian assumptions to elliptical distributions, enabling the modeling of multiplicative noise and tail dependence.
Algorithm Design: Developed MED-MAGMA, the first algorithm to fit Kronecker-sum-structured models under multiplicative noise using an EM framework with Laplace approximation.
Noise Handling: Introduced a rigorous mathematical transformation ( $f$ ) that isolates and removes multiplicative noise directions without discarding structural information.
Comprehensive Validation: Validated the method on every public scRNA-seq dataset in the Single Cell Expression Atlas under 1,000 cells (30 datasets), a level of systematic evaluation rarely seen in the field.

4. Results

The paper evaluates MED-MAGMA against GmGM (standard Gaussian) and a single-axis baseline.

Synthetic Data:
- MED-MAGMA remained robust across varying noise strengths, maintaining high Area Under Precision-Recall (AUPR), Assortativity, and Adjusted Mutual Information (AMI).
- GmGM and the baseline performance degraded significantly as noise strength increased.
Real-World Data (scRNA-seq):
- Across 30 datasets, MED-MAGMA consistently achieved higher Assortativity (tendency of similar cells to connect) and AMI (agreement with ground-truth cell types) compared to GmGM, even when GmGM was equipped with the "nonparanormal skeptic" (a method to handle non-Gaussian marginals).
- Case Study (E-MTAB-7249): Applied to ovarian cancer data, the learned gene network identified modules that strongly correlated with stromal and neoplastic cell subtypes. Robustness tests (using the Robin package) showed Bayes factors of $10^8$ (cells) and $59$ (genes), indicating the learned networks are highly stable against edge perturbations.
Tail Dependence: The paper demonstrates that real scRNA-seq data exhibits tail dependence (extreme co-expression), which Gaussian copula models fail to capture, whereas MED-MAGMA naturally accommodates it.

5. Significance

Biological Relevance: By accounting for the specific nature of multiplicative noise in sequencing data (e.g., capture efficiency), MED-MAGMA provides more accurate gene regulatory and cell-cell interaction networks.
Model Flexibility: The shift from Gaussian copulas to elliptical distributions allows for the modeling of "extreme" biological events that are correlated, a feature critical for understanding rare cell states or stress responses.
Practical Utility: The authors released the method as an open-source Python package (MED-MAGMA), making it accessible for the broader computational biology community.
Limitations: The current implementation scales cubically with matrix dimensions, making it suitable for datasets up to a few thousand cells/genes but challenging for massive-scale datasets (tens of thousands). Future work aims to optimize the runtime of the sufficient statistic approximation.

In summary, this paper provides a robust statistical framework for learning multi-axis networks in noisy, high-dimensional biological data, significantly outperforming existing Gaussian-based methods by explicitly modeling the multiplicative noise inherent in modern sequencing technologies.