Global gene network exploration based on explainable artificial intelligence approach.

Heewon Park,Seiya Imoto,Koji Maruhashi,Satoru Miyano,Rui Yamaguchi,Gary Stein

doi:10.1371/journal.pone.0241508

Abstract

In recent years, personalized gene regulatory networks have received significant attention, and interpretation of the multilayer networks has been a critical issue for a comprehensive understanding of gene regulatory systems. Although several statistical and machine learning approaches have been developed and applied to reveal sample-specific regulatory pathways, integrative understanding of the massive multilayer networks remains a challenge. To resolve this problem, we propose a novel artificial intelligence (AI) strategy for comprehensive gene regulatory network analysis. In our strategy, personalized gene networks corresponding specific clinical characteristic are constructed and the constructed network is considered as a second-order tensor. Then, an explainable AI method based on deep learning is applied to decompose the multilayer networks, thus we can reveal all-encompassing gene regulatory systems characterized by clinical features of patients. To evaluate the proposed methodology, we apply our method to the multilayer gene networks under varying conditions of an epithelial–mesenchymal transition (EMT) process. From the comprehensive analysis of multilayer networks, we identified novel markers, and the biological mechanisms of the identified genes and their reciprocal mechanisms are verified through the literature. Although any biological knowledge about the identified genes was not incorporated in our analysis, our data-driven approach based on AI approach provides biologically reliable results. Furthermore, the results provide crucial evidences to reveal biological mechanism related to various diseases, e.g., keratinocyte proliferation. The use of explainable AI method based on the tensor decomposition enables us to reveal global and novel mechanisms of gene regulatory system from the massive multiple networks, which cannot be demonstrated by existing methods. We expect that the proposed method provides a new insight into network biology and it will be a useful tool to integrative gene network analysis related complex architectures of diseases.

Highlights

Gene regulatory networks are crucial for understanding complex mechanisms of diseases
We focus on the results of the low epithelial–mesenchymal transition (EMT) region (i.e., Low.E&High.C2 versus Low.E&Low.C2) and group the genes as follows, ZEB1 and OVOL2: OVOL2 is one of the well-known EMT markers, and the interaction in the EMT process of OVOL2 and ZEB1 has been demonstrated in many studies: OVOL-TFs control mesenchymal–epithelial transition (MET) through a regulatory feedback loop with EMT-inducing TF, ZEB1 [41]; OVOL2 restricts EMT by directly inhibiting EMT-inducing factors including the ZEB1 system; A regulatory network containing OVOL2–ZEB1 mutual repression results in a fourstate EMT, i.e., epithelial, intermediate, intermediate, and mesenchymal states [42]; OVOL2 suppresses ZEB1 expression by binding to the ZEB1 promoter [39]
We found that gene TP63, which is known as a keratinocyte TF, plays a key role in the EMT process through interaction with the well-known EMT markers, i.e., TGF-β, GRHL2, and miR-200n family: Ectopic 4Np63a expression in normal human epidermal keratinocytes yields the EMT phenotype in a TGF-β-dependent manner; Knockdown of all isoforms of p63 leads to the EMT phenotype through loss of GRHL2 and miR-200 family genes [61]

Summary

Introduction

Gene regulatory networks are crucial for understanding complex mechanisms of diseases. The NetworkProfiler was applied to construct gene networks for 762 cancer cell lines characterized by EMT process, where EMT-related modulators for each cell line were measured based on 50 EMT-related genes labeled in the Molecular Signatures Database They focused on E-cadherin, which connects epithelial cells at adherens junctions, and identified 24 candidate regulators. Park et al [5] suggested that cancer characteristics are not uniformly distributed, and the Gaussian kernel function used to control the effect of samples in the NetworkProfiler leads to extremely small amount of weight for modeling a target sample having rare cancer characteristic, because the Gaussian kernel function is based on a constant bandwidth To address this problem, Park et al [5] proposed a robust version of NetworkProfier based on an adaptive bandwidth via the k-nearest neighbor rule, and constructed a drug sensitivity-specific gene network based on the Sanger dataset from the Cancer Genome Project

Methods

Results

Conclusion