Exploring genetic influences on adverse outcome pathways using heuristic simulation and graph data science

Highlights

• AI methods can be leveraged to uncover patterns of genetic regulation in AOPs using observational data.

• The AOP framework and AI methods can be implemented to gain novel insights into toxicity-mediated mechanisms and outcomes.

• This is the first study into how GP can be used to understand the genetic mechanisms underlying toxicity-mediated disease.

• The simultaneous association observed for variants in the AHR and ABCB11 genes indicates increased risk for liver cancer.

• Our socioeconomic deprivation approach provides a tool for improving social justice in environmental health studies.

Abstract

Adverse outcome pathways provide a powerful tool for understanding the biological signaling cascades that lead to disease outcomes following toxicity. The framework outlines downstream responses known as key events, culminating in a clinically significant adverse outcome as a final result of the toxic exposure. Here we use the AOP framework combined with artificial intelligence methods to gain novel insights into genetic mechanisms that underlie toxicity-mediated adverse health outcomes. Specifically, we focus on liver cancer as a case study with diverse underlying mechanisms that are clinically significant. Our approach uses two complementary AI techniques: Generative modeling via automated machine learning and genetic algorithms, and graph machine learning. We used data from the US Environmental Protection Agency’s Adverse Outcome Pathway Database (AOP-DB; aopdb.epa.gov) and the UK Biobank’s genetic data repository. We use the AOP-DB to extract disease-specific AOPs and build graph neural networks used in our final analyses. We use the UK Biobank to retrieve real-world genotype and phenotype data, where genotypes are based on single nucleotide polymorphism data extracted from the AOP-DB, and phenotypes are case/control cohorts for the disease of interest (liver cancer) corresponding to those adverse outcome pathways. We also use propensity score matching to appropriately sample based on important covariates (demographics, comorbidities, and social deprivation indices) and to balance the case and control populations in our machine language training/testing datasets. Finally, we describe a novel putative risk factor for LC that depends on genetic variation in both the aryl-hydrocarbon receptor (AHR) and ATP binding cassette subfamily B member 11 (ABCB11) genes.

Introduction

Informatics and computational methods have revolutionized biomedical research and enabled scientists to explore questions that are either infeasible or impossible through traditional experimentation alone [32]. In environmental health and toxicology, common computational tasks include building and training models that predict various chemical properties, conducting statistical analysis of observational and epidemiological data to better understand exposure-related health outcomes, and performing network analyses to discover key processes in biochemical pathways, among others [17], [38]. Despite the successes made using these methods, some key deficiencies have become apparent in toxicological research, such as a lack of richly structured, multimodal biomedical data describing chemicals and the biological systems that respond to chemical exposure [31] and a paucity of novel methods for discovering new knowledge from these complex data resources [36]. In this paper, we employ both to gain new insights into a phenomenon of growing interest: the influence of genetics on susceptibility to an adverse outcome following specific chemical exposures.

Adverse Outcome Pathways (AOPs) are pathway-like descriptions that outline the mechanistic associations between molecular exposure events and higher-order clinical and population-level outcomes that may arise from the exposure [2], [26]. AOPs consist of molecular initiating events (MIEs), key events (KEs), and adverse outcomes (AOs). By definition, a KE is any internal step within an AOP at some level of biological organization, and an MIE is a particular kind of KE that both initiates an AOP and is comprised of a molecular interaction between a toxicant and a body component. AOPs are classified according to their respective health outcomes, and AOPs associated with similar outcomes often overlap to create an ‘AOP network.’ An AOP’s set of KEs can include genetic polymorphisms that are associated with higher risk to the adverse outcome. For example, colon cancer AOPs include 53 unique SNP associations originally derived from GWAS [37]. This study will attempt to look at the influence genetic phenomena have on susceptibility to adverse outcomes after specific chemical exposures using AOPs as a framework for reference.

Methodologically, one area in particular that has experienced rapid growth, and holds great promise in all areas of biomedicine, is artificial intelligence (AI). AI broadly aims to construct computational systems that make intelligent decisions based on available data, knowledge, and/or human input. The scope of what comprises AI is broad, and usually nebulously defined. In this paper, we explore two areas within AI: Evolutionary algorithms and graph data science. Evolutionary algorithms are a family of algorithms that imitate processes found in biological evolution to optimize a system (e.g., a predictive model, a symbolic mathematical equation, or even another algorithm). Unsurprisingly, evolutionary computation is often used in computational biology, for example, in the context of simulating natural systems or processes [10], [11], [22] and building machine learning classifiers that perform well on a specific task [16], [23], [28]. Graph data science refers to the quantitative analysis of graphs – sometimes known as networks (e.g., biological networks), and comprised of a set of nodes connected by a set of edges that define relationships between those nodes [7], [27]. Some tasks within graph data science involve community detection [9], identification of the shortest paths linking two nodes in a graph [12], determining ‘hub nodes’ that play critical roles in the global structure of a graph [7], [41], and using computational algorithms that yield quantitative understandings of the behavior and characteristics of a given graph [1], [15]. Since AOPs can be represented as graphs, graph data science provides a powerful set of tools for discovering properties of AOPs that are not obvious through manual inspection.

Here, we propose a novel approach to gain understanding of the mechanisms underlying genetic influences on toxic adverse outcomes, without the inclusion of associated case-control information, that leverages these two areas of AI, and subsequently evaluates the approach in the context of toxicity-mediated adverse outcome pathways involved in liver cancer (LC). Briefly, we train interpretable generative models to construct synthetic datasets resembling real-world LC AOP genotype data via the HIBACHI software, and introspect the best models produced by HIBACHI (Heuristic Identification of Biological Architectures for simulating Complex Hierarchical genetic Interactions) for the most prominent AOP SNPs that influence LC outcomes. HIBACHI is a command line utility based on genetic programming (GP) that generates (synthetic) datasets with interactions between input features [24], [25]. It uses the (μ + λ) evolutionary algorithm [6] to construct trees of primitive mathematical operations that can represent interactions between independent variables. For example, when applied to genetic data, these feature interactions may represent epistasis or mechanisms underlying polygenic traits. HIBACHI can take an existing dataset – referred to in the context of GP as a model – as input, which is then used to evaluate the fitness of candidate output datasets. Our hypothesis is that HIBACHI can create synthetic datasets of SNPs involved in AOPs that behave the same as real data for the same AOPs. This will allow us to explore the interpretable generative models used to create the synthetic data, which gives insights into interactions between specific features in the real data used to train HIBACHI. Conceptually, this process can be likened to a brute-force version of symbolic regression [18] that avoids pitfalls arising from statistical analyses on genetic data with complex interactions between features [40]. Importantly, this approach utilizes genomic and phenotypic data from real-world populations, combined with information and knowledge sourced from publicly available, open access databases describing mechanisms of toxicity. Our methods are generalizable to other diseases of interest and provide a new framework for toxicologists to explore genetic mechanisms that underlie toxic adverse outcome susceptibility.