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

In this paper we develop a graph-learning algorithm, MED-MAGMA, to fit multi-axis (Kronecker-sum-structured) models corrupted by multiplicative noise. This type of noise is natural in many application domains, such as that of single-cell RNA sequencing, in which it naturally captures technical biases of RNA sequencing platforms. Our work is evaluated against prior work on each and every public dataset in the Single Cell Expression Atlas under a certain size, demonstrating that our methodology learns networks with better local and global structure. MED-MAGMA is made available as a Python package (MED-MAGMA).

💡 Research Summary

This paper addresses a significant gap in multi-axis graphical modeling by developing a method robust to multiplicative noise, a common feature in real-world data like single-cell RNA sequencing (scRNA-seq). The authors introduce MED-MAGMA (Matrix Elliptical Distribution’s Multi-Axis Graphical Modelling Algorithm), a novel algorithm designed to fit Kronecker-sum-structured models (where the precision matrix is Ψ_cols ⊕ Ψ_rows) when the observed data is subject to element-wise multiplicative scaling.

The core problem stems from technical biases in applications like scRNA-seq, where gene-specific capture efficiency and cell-specific sequencing depth introduce noise that multiplies the underlying latent signal. Existing multi-axis models (e.g., GmGM) assume additive Gaussian noise and fail to account for this, leading to biased network estimates. The proposed model formalizes this as: the observed data matrix X is a Hadamard product between a latent Gaussian matrix Z (with Kronecker-sum precision structure) and an outer product of unknown, strictly positive row and column scaling vectors R_rows and R_cols.

The methodological innovation lies in a three-pronged approach: 1) Design of a noise-invariant function (f): A carefully constructed function f is applied to the data, which removes all information corrupted by the multiplicative scaling, leaving only the “direction” of the latent signal Z. This function uses geometric means across the entire matrix, rows, and columns for normalization. 2) Expectation-Maximization (EM) framework: With y=f(X) as the observed variable and Z as latent, an EM algorithm is derived. The M-step reduces to solving a standard GmGM problem, but requires the conditional expectations of the sufficient statistics ZZ^T and Z^TZ given y. 3) Efficient approximation via Laplace’s method: Computing these conditional expectations involves intractable integrals over the fiber space f^{-1}(y). The authors employ Laplace’s method to approximate these integrals efficiently. This requires finding the point z* in the fiber that minimizes z^T Ω z, which is achieved by a flip-flop algorithm that breaks the problem into alternating quadratic programming sub-problems over R_rows and R_cols.

The authors provide comprehensive validation. On synthetic data, MED-MAGMA accurately recovers the true row and column network structures under multiplicative noise, outperforming the noise-agnostic GmGM. Most compellingly, they conduct a systematic evaluation on every public dataset in the Single Cell Expression Atlas with fewer than 1000 cells. Using metrics that assess both local network structure (e.g., hub gene identification) and global structure (e.g., clustering quality after dimensionality reduction), MED-MAGMA demonstrates superior performance across the majority of datasets. A deep dive into the E-MTAB-7249 dataset shows that networks learned by MED-MAGMA lead to clearer separation of biologically meaningful cell clusters and more coherent gene regulatory modules compared to baselines.

In summary, this work makes both a practical and theoretical contribution. Practically, it delivers a robust algorithm for a critical noise model in modern bioinformatics. Theoretically, it generalizes multi-axis models to the richer class of matrix-variate elliptical distributions, enabling the capture of tail dependencies. The algorithm is implemented and made available as an open-source Python package, facilitating adoption and further research.