ABSTRACT
In an agroecosystem, plants and microbes coexist and interact with environmental factors such as climate, soil, and pests. However, agricultural practices that depend on chemical fertilizers, pesticides, and frequent tillage often disrupt the beneficial interactions in the agroecosystem. To reconcile the improvement of crop performance and reduction in environmental impacts in agriculture, we need to understand the functions of the complex interactions and develop an agricultural system that can maximize the potential benefits of the agroecosystem. Therefore, we are developing a system called the agroecosystem engineering system, which aims to optimize the interactions between crops, microbes, and environmental factors, using multi-omics analysis. This review first summarizes the progress and examples of omics approaches, including multi-omics analysis, to reveal complex interactions in the agroecosystem. The latter half of this review discusses the prospects of data analysis approaches in the agroecosystem engineering system, including causal network analysis and predictive modeling.
Integrated data analysis of multi-omics data enables agroecosystem engineering.
Plants harbor various microbes around their roots, leaves, and inside cells. While some of these microbes are pathogens that inhibit plant growth, others are symbiotic microbes that help their host plants survive, grow, and reproduce. Various symbiotic microbes are found in the rhizosphere (Lugtenberg and Kamilova 2009), endosphere (Santoyo et al. 2016), and phyllosphere (Fu et al. 2016), which increase the fitness of the host plants by increasing nutrient availability, producing phytohormones, enhancing stress tolerance, and suppressing plant pathogens (Chamkhi et al.2022; Elnahal et al.2022). However, intensive farming practices in modern agriculture exert negative effects on symbiotic microbes (French et al. 2021). For example, frequent tillage disrupts the interaction between plants and beneficial microbes such as arbuscular mycorrhizal fungi (Bowles et al.2017). Thus, switching to agricultural practices that utilize symbiotic microbes is expected to not only increase crop yield and quality, but also reduce fertilizer and pesticide use and enhance ecological resilience to climate change (Grover et al. 2011). This shift requires a detailed understanding of the functions of diverse plant-associated microbes.
The progress of plant-microbe interaction studies
The current knowledge of the functions of plant-associated microbes involves that obtained from only a fraction of the diversity of the natural microbial community. Only 1% of the total population of soil bacteria is culturable in a restricted range of media and cultivation conditions (McCaig et al. 2001), and even cultured microbes often behave differently in laboratories and natural environments (Palková 2004). Unlike laboratory experiments that use sterile media or soil, field soils contain various microbial species that compete with each other, which can alter the impact of microbes on the host plant. For example, Flavobacterium was shown to inhibit plant growth in sterile media but did not suppress it in the presence of other species (Hartman et al. 2017). Gaps exist between our knowledge of the limited number of microbes cultured in the laboratory and the behavior of natural microbial communities, including those species that cannot be cultured.
Technical advances in molecular biology are beginning to bridge these gaps. Gene sequencing technologies, such as 16S rRNA amplicon sequencing and shotgun metagenome sequencing, have enabled the determination of the abundance of hundreds of microbial species in environmental samples (Handelsman 2004). Previous studies using 16S rRNA have revealed that the composition of the rhizosphere microbial community is significantly distinct from that of bulk soils (Marilley et al. 1998; Marilley and Aragno 1999). In addition, microbial community composition is well correlated with host plant status such as the growth stage (Houlden et al. 2008; Edwards et al. 2018) and plant health (Liu et al. 2021). These results indicate toward the concept of plant holobiont (Rosenberg et al. 2007), wherein a plant and plant-associated microbial community synchronizes as a single entity. This concept could be important for considering the adaptation of host organisms, such as the human holobiont (Van de Guchte, Blottière and Doré 2018) and coral holobiont (Bourne et al. 2009). The focus of plant-microbe interaction research is shifting from a single interaction between a plant and one microbe to interconnections between a plant and a microbial community within the plant holobiont to understand the functions of microbial communities in the natural environment.
Multi-omics for plant-microbe interaction
In the past decade, owing to the advances in omics technologies, significant progress has been made in elucidating the behavior of plant holobiont in an agroecosystem (Figure 1). The term omics indicates the comprehensive quantification of biomolecules. Metagenomics, the genetic sequencing of microbial populations, is an example of omics. Based on the developments in various omics techniques and computational tools, multi-omics-based studies have become a hotspot for many scientific fields, including plant and microbial sciences (Krassowski et al. 2020). Multi-omics-based studies integrate multiple omics analyses to examine processes or relationships across different biological layers that cannot be analyzed using a single omics analysis. Here, we introduce the utility of two omics technologies in plant holobiont studies and give examples of actual multi-omics studies of agroecosystems.
The concept of the agroecosystem. The agroecosystem consists of the interactions among plant-holobiont, soil, climate, insect community, and agricultural practices.
Transcriptomics, the study of the complete set of mRNAs transcribed from the genome, provides informative data to examine the genetic backgrounds of interactions in plant holobiont. A recent study has shown that the inoculation of root microbial communities derived from different soil samples significantly altered the gene expression profiles in rice and affected the expression levels of particular gene families involved in the disease resistance mechanism (Santos‐Medellín et al. 2022). Another transcriptomic analysis demonstrated that the downregulation of the iron transporter gene in plant roots under drought stress triggered a change in the microbial composition in the rhizosphere, especially leading to the enrichment of Actinobacteria (Xu et al. 2021). A study using metatranscriptomics, an analysis of the gene expression profiles of a microbial community, showed that the enrichment of Actinobacteria under drought stress was correlated with the increased expression of genes involved in carbohydrate and amino acid metabolism and transport, which may contribute to the increased adaptation of host plants (Xu et al. 2018). These transcriptomic studies have shown that plants and microbes are interconnected through the regulation of various gene expressions.
Metabolomics, the comprehensive study of a set of metabolic compounds, can be used to examine the functions of various metabolites in the plant holobiont. In particular, the interactions of secondary metabolites related to plant stress tolerance with the plant-associated microbial community have been well-studied (Pang et al. 2021). A previous study has shown that the presence of the root microbial community altered the leaf metabolite composition, which inhibited the feeding behavior of herbivore larvae (Badri et al. 2013). It has also been well-recognized that plant roots release various metabolites that recruit beneficial microbes. For example, infection by the pathogen Pseudomonas triggered a compositional change in the extracellular metabolites released from roots, such as a decrease in sugar content and an increase in amino acids, resulting in the enrichment of beneficial microbes in the rhizosphere (Yuan et al. 2018). Extracellular metabolites classified as benzoxazinoids were shown to affect the composition of the rhizosphere microbial community and suppress the feeding activity of herbivores in the next plant generation (Hu et al. 2018). Metabolomic studies have shown that plants and microbes communicate through the exchange of metabolites.
The integration of multiple omics analyses has led to the study of interaction networks in actual fields at the agroecosystem level. Ichihashi et al. (2020) conducted a multi-omics study in an arable field, using different soil management methods. They integrated plant phenotype data with the following four omics datasets: plant metabolome, rhizosphere microbiome, soil metabolome, and soil ionome. The results showed that soil management practices affected multiple omics layers. Upon reconstructing the module structure of the agroecosystem from multi-omics data, they also found a novel interaction in which soil choline promoted crop growth. Other agroecosystem-level studies using multi-omics analysis have been conducted in actual fields such as soybean fields with cover crops (Yamazaki et al. 2021), apple orchards with intercropping (Li et al. 2022), and corn-wheat rotation fields with long-term fertilization (Guiru et al. 2022). Although the application of multi-omics to agroecosystems is still in its infancy, this is a promising approach for understanding the behavior of plant holobiont in the agroecosystem.
Prospects of agroecosystem engineering system
One significant goal of the multi-omics analysis is the update of agricultural practices by agroecosystem engineering system. Conventional agricultural practices target only a limited range of factors in the agroecosystem, such as essential nutrients and pathogens, and depend on chemical fertilizer and pesticide use, neglecting the inherent functions of the plant holobiont. In contrast, the agroecosystem engineering system aims to engineer the entire agroecosystem, including diverse microbial communities and functional metabolites. A proposed example of such engineering is the establishment of a healthy microbial community that suppresses the outbreak of soil pathogens by introducing core microbes that recruit beneficial microbes (Toju et al. 2018). The vast knowledge gained from the multi-omics analysis will enable high-resolution agroecosystem engineering optimized for each field. In the data analysis leading to this point, it is necessary to clarify the relationships among the components of the agroecosystem and to simulate changes in the overall system when an intervention is made. Here, we outline the process of agroecosystem engineering in the following two steps: the reconstruction of causal networks and simulation of intervention effects.
Causal network analysis using multi-omics data
To realize agroecosystem engineering, we first need to reconstruct the causal network underlying multi-omics data. Network analysis can explain the effects of, eg the introduction of a certain soil metabolites on microbial communities and plant phenotypes. Thus, the causal network can be used to propose prescriptions that improve target parameters such as crop yield and quality. Many studies have used correlations to reconstruct the interaction network from the omics data. However, the correlation network involves spurious correlations, ie correlations that are not causally related, but originate from coincidence or unobserved common causes among the two elements. In addition, the direction of the causal relationship cannot be determined from the correlations. Hence, further experimental tests are required to identify causal relationships within the correlation network. However, it is difficult to verify multiple effects in a complex network generated from multi-omics data. Therefore, there is a need for computational methods to distinguish causal relationships from spurious correlations and infer the directions of causal effects (Figure 2).
Two different networks from multi-omics data. The constructed networks are different between correlation and causal networks. Blue and red lines (arrows) represent positive and negative connections, respectively. Dashed lines represent spurious correlations.
This need has led to the development of statistical methods to infer causal relationships from observational data. This area is known as causal discovery and causal structure learning. Although methods of causal discovery for plant-microbe interactions using time-series data have also been compared (Suzuki et al. 2022), we focus here on causal discovery using non-time-series data, which are more common in multi-omics data than time-series data. Since the 1990s, two types of algorithms have been developed: the constraint-based approach represented by the PC algorithm (Spirtes et al. 2000) and the score-based approach represented by the Greedy Equivalence Search algorithm (Chickering 2002). These are non-parametric models that make no assumptions about data distribution and functional forms in causal relationships. However, they fail to uniquely identify the correct causal structure, and instead output multiple candidates with the same conditional independence. Researchers have found that a semi-parametric model, called the Linear Non-Gaussian Acyclic Model (LiNGAM), can completely identify the correct causal network under certain assumptions, including linear functions and non-Gaussian data distribution (Shimizu et al. 2006). Several algorithms have been developed for LiNGAM, such as the ICA-LiNGAM algorithm that uses independent component analysis (Shimizu et al. 2006) and DirectLiNGAM algorithm that repeats regression and independence tests between an explanatory variable and a residual (Shimizu et al. 2011). These models have already been applied to several studies in social science (Helajärvi et al. 2014; Louvigné et al. 2018; Chen et al. 2018; Yamada, Ohkubo and Shimizu 2020) and natural sciences (Banks and Banks 2019). LiNGAM has demonstrated its utility in these studies, and is at the frontier of causality discovery research.
The application of these causal discovery models to agroecosystem multi-omics data presents several challenges. The first challenge is the high-dimension and low-sample-size dilemma. Causal discovery algorithms, such as LiNGAM, require a large sample size compared to the number of variables to reconstruct an accurate causal graph (Shimizu, Hyvarinen and Kawahara 2014). However, multi-omics data generated from a field experiment often have a limited sample size of ˂ 100 with numeric variables being more than tens of thousands, owing to the integration of multiple omics. Therefore, effective feature selection and reduction methods must be considered to establish a robust pipeline. The second challenge is the violation of model assumptions by multi-omics data. LiNGAM makes several assumptions, including a linear relationship and an acyclic graph structure. However, biological systems often have nonlinear relationships with feedback loops. To accommodate such relationships, algorithms that loosen these assumptions are under development (Hoyer et al. 2009; Lacerda et al. 2012). The third challenge is the heterogeneity of the multi-omics data. Multi-omics data are composed of datasets from different sampling locations such as leaves, roots, and soils. In addition, each dataset undergoes different measurements and computational processes. All these data backgrounds produce different signal-to-noise ratios across omics layers, which may lead to false results in the causal discovery. It is necessary to discuss the methods that cope with such heterogeneous noises, such as weighting of each omics dataset or data transformation into a latent space with common signal features (Miao et al. 2021). Overcoming these challenges will allow us to replicate the causal network of agroecosystems from multi-omics data and suggest where manipulations can improve crop performance.
Future predictions using multi-omics modeling
The causal network inferred from multi-omics data leads to a better understanding of predictive models and simulations utilizing multi-omics data for agroecosystem engineering (Figure 3). Agricultural models have long been studied and utilized for crop production (Jones et al. 2017). For example, integrated crop models, such as the Agricultural Production Systems Simulator (APSIM) and Decision Support System for Agrotechnology Transfer (DSSAT), use weather, soil, and management method data to predict and simulate crop growth (Jones et al. 2003; Keating et al. 2003). However, more holistic models, which include the processes related to plant-associated microbes, have not yet been developed. Given that plant-associated microbes have a significant impact on crop phenotypes, the integration of the knowledge from multi-omics data into models will improve the prediction accuracy. Furthermore, models constructed using multi-omics data can predict not only crop yield but also new targets such as crop qualities and environmental impacts.
The workflow of agroecosystem engineering by predictive modeling. Modeling and simulation using multi-omics data support agroecosystem engineering.
Successful crop models have been built on process-based models; however, machine learning models have recently been used to overcome the limitation of process-based models. The actual processes in crop growth involving many microbes, transcripts, and metabolites are often too complex to fit simplified process-based models. Machine learning models, such as random forest and support vector machine, are effective to predict the output from complex processes using multivariate data with sufficient samples. These models have successfully predicted the ecological functions of root microbes, including carbon decomposition, soil health, and crop phenotypes (Chang et al. 2017; Thompson et al. 2019; Wilhelm, van Es and Buckley 2022). One of the issues in using machine learning models is their low interpretability compared with process-based models, which prevents an understanding of the rationale behind the predictions (Carvalho, Pereira and Cardoso 2019). Using a causal network to select features and describe causal relationships among explanatory variables can help in the interpretation of the models. Furthermore, to assemble models that are more interpretable and accurate, previous studies proposed incorporating the outputs of a process-based model into the variables of machine learning model (Feng et al.2019) or incorporating the outputs of machine learning model into the variables of the process-based model (Droutsas et al. 2022). Integration of machine learning models using multi-omics data into the process-based crop models will improve the resolution of crop prediction.
Digital transformation is now creating a new trend called the “digital twin” in modeling methodology, and its application in agriculture has already been discussed (Pylianidis, Osinga and Athanasiadis 2021; Nasirahmadi and Hensel 2022). Unlike conventional models, digital twins automatically perform simulations using real-time data, such as weather sensors, soil sensors, drone cameras, and satellite images, which enables real-time monitoring and high-resolution simulations. The sensing data can be assimilated into crop models by machine learning techniques called data assimilation (Jin et al. 2018). As digital twins are used for anomaly detection in health care (Gupta et al. 2021), they can also be used for the early detection of crop disease outbreaks by combining an initial set of multi-omics data with real-time weather data and drone images. Although the “digital twin” trend is still in the early stages of research, the development of agroecosystem engineering system will be accelerated by these new concepts.
Conclusion
Developments in molecular biology have demonstrated close relationships between plants and microbes, and multi-omics analysis has begun to reveal the behavior of this plant holobiont in a complex agroecosystem. We are visualizing the causal network of the agroecosystem from multi-omics data and building predictive models that include the functions of diverse interactions in the agroecosystem. This agroecosystem engineering system will contribute to the establishment of agricultural practices with less environmental impact.
Acknowledgments
We appreciate the support and advice of our lab members and collaborators. We would like to thank Editage (www.editage.com) for English language editing.
Funding
This work was supported by Cabinet Office, Government of Japan, Moonshot Research and Development Program for Agriculture, Forestry and Fisheries (JPJ009237, funding agency: Bio-oriented Technology Research Advancement Institution), and by the RIKEN Junior Research Associate Program.
Disclosure statement
No potential conflict of interest was reported by the authors.
