Tensor Network based Gene Regulatory Network Inference for Single-Cell Transcriptomic Data

Olatz Sanz Larrarte, Borja Aizpurua, Reza Dastbasteh, Ruben M. Otxoa, Josu Etxezarreta Martinez · 2025

Read Tensor Network based Gene Regulatory Network Inference for Single-Cell Transcriptomic Data by Olatz Sanz Larrarte, Borja Aizpurua, Reza Dastbasteh, Ruben M. Otxoa, Josu Etxezarreta Martinez (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-06891.txt.

Introduction

Gene interactions hidden in single-cell data are inferred with tensor networks by Olatz Sanz Larrarte and four co-authors in a 2025 computational biology paper. Traditional methods for gene regulatory networks often miss higher-order and nonlinear dependencies between genes. The authors borrow tools from quantum physics to catch them.

Expression data is mapped into a lower-dimensional tensor network representation that preserves biological locality. Quantum mutual information, a dependence measure natural to tensor networks, quantifies how genes relate, and permutation testing checks that edges are not chance artifacts. Applied to single-cell RNA data from more than twenty-eight thousand lymphoblastoid cells, the method recovered a regulatory network of six pathway genes. It also uncovered several three-way regulatory mechanisms. Robust edges that reflect real biology rather than noise are the goal, with disease research and precision medicine as intended uses.

Catalog record

CREATOROlatz Sanz Larrarte, Borja Aizpurua, Reza Dastbasteh, Ruben M. Otxoa, Josu Etxezarreta Martinez
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-06891

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Tensor Network based Gene Regulatory Network Inference for Single-Cell Transcriptomic Data Olatz Sanz Larrarte, Borja Aizpurua, Reza Dastbasteh, Ruben M. Otxoa, Josu Etxezarreta Martinez Abstract Deciphering complex gene-gene interactions remains challenging in transcriptomics as traditional methods often miss higher-order and nonlinear dependencies. This study introduces a quantum-inspired framework leveraging tensor networks (TNs) to optimally map expression data into a lower dimensional representation preserving biological locality. Using Quantum Mutual Information (QMI), a nonparametric measure natural for tensor networks, we quantify gene dependencies and establish statistical significance via permutation testing. This constructs robust interaction networks where the edges reflect biologically meaningful relationships that are resilient to random chance. The approach effectively distinguishes true regulatory patterns from experimental noise and biological stochasticity. To test the proposed method, we recover a gene regulatory network consisted of six pathway genes from single-cell RNA sequencing data comprising over 28.000 28.000 lymphoblastoid cells. Furthermore, we unveil several triadic regulatory mechanisms. By merging quantum physics inspired techniques with computational biology, our method provides novel insights into gene regulation, with applications in disease mechanisms and precision medicine. I Introduction Computational analysis, and more precisely computational biology, has revolutionized the ability to understand complex biological processes Markowetz2017 . This branch of science studies biological systems, focusing on data analysis, mathematical modeling, and computational simulations. Furthermore, from the beginning of the XXIst century, advances in high-performance technologies and digitization have led to an exponential growth in biological data Altaf2014 . A key milestone in this process was the Human Genome Project (HGP) , completed in 2003, which marked the beginning of mass genomics by generating the first partial sequencing of the human genome Collins2003 . In 2005, the emergence of next-generation sequencing (NGS) technologies drastically reduced the costs and time associated with genomic analysis, in turn increasing the amount of data generated Mardis2008 . Simultaneously, the concept of "omics" emerged to describe the systematic study of sets of molecules, such as genes (genomics), messenger RNA transcripts (transcriptomics), proteins (proteomics) and metabolites (metabolomics). In transcriptomics, technologies such as RNA-Seq make it possible to measure the amount and sequence of messenger RNA in an organism, tissue, or cell at a specific time Morin2008 . These tools not only extract meaningful information, but also enable the creation of predictive models to explore complex biological phenomena beyond traditional experimental capabilities. In this context, inference of functional relationships between genes has emerged as a key area in molecular biology wang2009 . Under this framework, inference of gene regulatory networks (GRN) has become an important area in molecular biology, providing a formal representation of gene-gene interactions using graph representations, \mathrm{G(V,E)} , where nodes, \mathrm{V} , correspond to genes while edges, \mathrm{E} , represent regulatory interactions measured by transcription factors. These networks are fundamental to understand molecular mechanisms, including gene functions and their regulation in cellular processes Friedman2000 . However, GRN reconstruction remains challenging because experimental data (e.g., gene expression) only provides indirect evidence of regulatory interactions—such as correlations or perturbation responses—rather than direct observations of causality. Computational methods must therefore contend with these ambiguities, alongside technical noise and biological complexity Margolin2006 ; faith2007 ; haury2012 . In order to infer GRNs from gene expression data, different methods have been developed. These methods use statistical models to detect dependencies between gene expression profiles and establish possible regulatory relationships between them. Common approaches include correlation networks zhang2005 , information-theoretic scores butte1999mutual ; mousavian2016 , regression-based models Salleh2017 ; Kim2009 , Gaussian graphical models schafer2005 , Bayesian and Boolean networks Lahdesmaki2006 ; Ilya2002 ; Friedman2000 or dynamic models oates2012 ; morrissey2010 . It should be noted that each method has its own assumptions and therefore specific limitations that restrict its applicability to more complex biological systems Chen2018 ; Pratapa2020 ; Diaz2022 . For example, many traditional methodologies assume linear or simple relationships between variables, which is not always true, and more importantly, none of these conventional methods fully exploit the simultaneous and inter-regulatory connections between all genes. As a result, new ways for GRN inference are being proposed, including using quantum computers to do so roman2023 . Tensor Networks (TNs) are compact mathematical representations designed to model high-dimensional systems with complex correlation structures cirac2009 ; schollwock2011 . Originally, at the turn of the century, those were developed to address problems in quantum many-body systems orus2019 and fluids gourianov2025 , allowing efficient representation of complex quantum states in high-dimensional systems. Recently, these techniques have found applications in machine learning and data science sengupta2022 ; wang2023 , due to their striking ability to handle large and highly correlated data with reduced computational complexity. Mathematically, TNs take advantage of decompositions such as singular value decomposition (SVD) to capture correlations between system components, allowing less relevant dimensions to be efficiently truncated. For example, Matrix Product State (MPS) and Tree Tensor Network (TTN) models are two common TN structures, with applications in quantum many-body systems. In this work, we propose a Tensor Network (TN)-based framework to infer Gene Regulatory Networks (GRNs) from single-cell transcriptomic data, overcoming the limitations of pairwise interaction models by capturing higher-order dependencies among genes. Unlike traditional approaches restricted to binary correlations, our TN formulation naturally encodes multi-way regulatory relationships through its multilinear structure, preserving biological context while enabling efficient computation. By representing gene interactions as interconnected low-rank tensors, our method maintains the intrinsic high-dimensional structure of expression data, revealing complex regulatory motifs. The following sections discuss the theoretical foundations of this approach, detail the methodology, and present experimental results that highlight the ability of TNs to reveal intricate regulatory mechanisms that may be overlooked by classical techniques. As a result, our TN-based GRN inference method is a flexible tool that can capture gene regulation relationships in an accurate and computationally tractable manner. II Quantum Computation Theory Quantum computation leverages quantum mechanics to represent and manipulate information. Unlike classical bits, which can be either 0 or 1 , quantum bits ( qubits ) can exist in a superposition of both states within a two-dimensional Hilbert space. The state of a single qubit is written as: |\psi\rangle=\alpha|0\rangle+\beta|1\rangle,\quad\text{with }|\alpha|^{2}+|\beta|^{2}=1, where \alpha,\beta\in\mathbb{C} are complex probability amplitudes. In our gene analogy, |0\rangle represents an inactive gene and |1\rangle an active gene; the gene’s quantum state reflects a probabilistic combination of both. Upon measurement in the computational basis \{|0\rangle,|1\rangle\} , the system collapses to the state |0\rangle with probability |\alpha|^{2} or |1\rangle with probability |\beta|^{2} . When considering two qubits, their joint state is formed via the tensor product: \begin{split}|\psi\rangle&=|\psi_{1}\rangle\otimes|\psi_{2}\rangle\\ &=\alpha_{1}\alpha_{2}|00\rangle+\alpha_{1}\beta_{2}|01\rangle+\beta_{1}\alpha_{2}|10\rangle+\beta_{1}\beta_{2}|11\rangle.\end{split} This product assumes that the qubits are independent. If they interact, the system may exhibit entanglement , a uniquely quantum phenomenon where states cannot be factorized into individual components. For a general system of N qubits, the global state is: \begin{split}|\psi\rangle&=\sum_{i_{1},i_{2},\dots,i_{N}}c_{i_{1}i_{2}\dots i_{N}}|i_{1}\rangle\otimes|i_{2}\rangle\otimes\dots\otimes|i_{N}\rangle\\ &=\sum_{i_{1},i_{2},\dots,i_{N}}c_{i_{1}i_{2}\dots i_{N}}|i_{1}i_{2}...i_{N}\rangle,\end{split} where c_{i_{1}\dots i_{N}}\in\mathbb{C} and \displaystyle\sum_{i_{1},i_{2},\dots,i_{N}}|c_{i_{1}\dots i_{N}}|^{2}=1 ensures normalization. This formalism allows us to represent multiple genes within a unified computational framework, where their joint behavior can be described using concepts inspired by quantum systems. Such representations capture dependent gene activity—including synergistic regulation maeder2013 , epistasis, or other higher-order interactions—without requiring physical quantum effects. Instead, the approach uses mathematical tools from quantum theory to efficiently model complex dependencies that classical methods might overlook. III Matrix Product State: an architecture for GRN inference MPSs

Read the complete text