Prediction of Influence Transmission by Water Temperature of Fish Intramuscular Metabolites and Intestinal Microbiota Factor Cascade Using Bayesian Networks

Abstract

Aquaculture is receiving attention as one of the solutions to the global food problem. Therefore, it is essential to clarify the impact of fish and their environment on the stable supply and uniformity of the quality of fish provided as meat. Nuclear magnetic resonance can comprehensively acquire metabolite information in foods nondestructively and is suitable for measuring physical properties for quality control. Moreover, recent advances in machine learning methods and artificial neural network (ANN) analysis have contributed to the analysis of comprehensive information. In this study, we sampled a wide variety of fish from the natural sea and analyzed them using a scheme incorporating ANN. As a result, it was found that anserine, an antioxidant, was found to be reduced in fish muscles, and this destabilized the homeostasis of other metabolites at low water temperature. We also concluded that the fish muscle metabolic state was stabilized in warm water. Furthermore, a relationship between water temperature and the intestinal microbiota of fish was established. In this study, we evaluated the relationship between the metabolic profile changes in fish muscle and external environmental factors and predicted connection strength and order using machine learning and ANN. We conclude that our proposed scheme for estimating the degree and direction of the influence of environmental factors on organisms by using ANN will work.

1. Introduction

Technological progress is expected to promote a range of human activities, such as ensuring food security, including with regard to marine resources. Fish are becoming an increasingly popular component of the human diet, with progress in aquaculture potentially solving various food security-related issues [1,2,3]. However, several fish types still cannot be successfully farmed, whereas other problems, such as drainage, feed, taste of the produce, and stability of the supply of aquaculture products, remain to be resolved. The failure to resolve these problems is related to the difficulty in analyzing the effects of particular environmental factors on fish [4]. As such, there is an urgent need to promote the analysis of the effects on metabolite variations in meat for uniform fish growth and food quality.

Nuclear magnetic resonance (NMR) enables the non-destructive identification and measurement of metabolites, such as amino acid, organic acids, lipids, and proteins, and has played an important role in omics technologies [5]. Omics technologies involve the application of various measurement techniques enabling a comprehensive analysis of a particular sample [6,7]. Particularly in microbiology, the advances using next-generation sequencing (NGS) technology for DNA encoding 16S ribosomes have made it possible to analyze diverse microbiomes which are difficult to culture, for example, the intestinal microbiome [8]. Advances in computational technology have enabled the study of comprehensive data from NMR and NGS multivariate analyses [9,10]. Furthermore, it has been shown that machine learning and artificial neural network (ANN) methods are effective in analyzing such complex data and are able to make predictions because ANN identifies arbitrary non-liner multiparametric discriminant function from big data [11].

Metabolites in living organisms can be greatly influenced by host conditions, such as genetic background, age, disease, behavior, and external environmental factors [12,13,14].

The environmental factors to which living organisms are exposed interact with each other and exert various influences on them [15,16,17]. Symbiotic microbes influence their hosts, for example, by affecting the development and optimization of immunity and working cooperatively in digestion and maintaining host health; thus, symbiotic microbes and their interactions can potentially be incorporated into the concept of the exposome [18,19,20].

Furthermore, “temperature” as energy in environmental factors influences the skeletal muscular metabolic fluctuations of organisms and the growth of microorganisms [12,21,22]. In fact, the fish intestinal microbiota are affected by various environmental factors, including temperature [23].

The concept is particularly important as it extends across various realms, such as the physical, scientific, and biological spheres. The concept of the exposome has been widely adopted; then, the government and other local governments have continued to investigate and amass basic environmental information to look back freely on its impact. This environmental information is also used in Japan by various local governments that have released surveys and databases on measurements of air, water, soil temperature, chemical substances, and other variables.

In this study, we aimed to establish a scheme to estimate the impact of environmental factors on organisms by characterizing the relationship between fish habitats based on the data previously collected [24,25,26], with water temperature as the environmental factor index. Various variables can be considered as environmental factors, but convertible energy is expected to affect several environmental factors and is the most basic variable that can be measured and for which data can be accumulated by local governments. In this study, it was decided that temperature would be used as a proxy for available energy. Based on the accumulated data and the water temperature at the time of sampling, we evaluated the effects of water temperature on the metabolic profiling of fish muscle using NMR spectroscopy. We also hypothesized that changes in the intestinal microbiota are an intermediate factor affecting fish phenotypes and visualized the relationship between temperature changes and the intestinal microbiota in fish. Moreover, we attempted to predict the important factors that altered the cascade of metabolites using an artificial neural network, Bayesian networks, which are probabilistic neural net analyses. Bayesian networks are graphical models that use arrows to indicate probabilistic dependencies between variable nodes [27]. Recently, they have been applied to many fields, such as breast cancer detection and the estimation of gene expression networks [28,29,30]. Therefore, it is possible to speculate on the effects of environmental factors on fish phenotypes, including intestinal microbiota.

Our proposed framework clusters the data using the i-means method, which consists of two well-known machine learning tools: K-means and random forest [31]. Alternatively, it has been reported that energetically stable substances show less variability, and the same is true for the microbiome [32]. Herein, this approach evaluates the stability of the data from the centrality of the data using Mahalanobis’ distance and evaluates the environmental factors associated with unstable data which was defined as a state away from the group centrality. The evaluation method was an a priori algorithm to extract the association rules that determine how temperature influences the phenotype. Finally, a Bayesian network was used to predict the sequence of each factor. Computations were performed in R.

2. Materials and Methods

2.1. Data Preparation

As a phenotype, we used data from the NMR analysis of meat from fish collected from 2011 to 2015 in the seas around Japan (Figure 1) [24,25,26]. NGS data were collected from 2012 to 2016 in the same areas as the NMR data. Overall, 59 samples from 7 orders of taxonomy of NMR data and 123 samples from 16 orders of taxonomy of NGS data were used. The NMR and NGS data were processed to obtain the composition ratio to the sum of the measurement in fish for each sample. As an environmental factor, data on water temperature from 2010 to 2016 were extracted from the Japan Meteorological Agency website. Daily data were averaged on a monthly basis. We added the samples information (body size, orders of taxonomy, sampling area)and water temperature of the month to which the sampling date belongs to NMR data (

Table 1. Summary of fish sample characters used for NMR data.

Order Name	Number	Length Ave. (mm)	SD	Weight Ave. (g)	SD
Aulopiformes	3	101.7	±88.4	28.6	±2.2
Gadiformes	4	371.3	±251.5	391.5	±126.2
Myliobatiformes	3	167.3	±30.4	78.4	±37.7
Perciformes	37	153.1	±83.4	92.6	±137.1
Scorpaeniformes	5	185.6	±83.0	178.7	±238.7
Tetraodontiformes	5	238.0	±78.5	266.9	±209.0
Zeiformes	2	295.0	±162.6	581.5	±671.1

). Length and weight were ranked within the order of taxonomy and “Long” was defined as higher than the interquartile range; “Tiny” was defined as lower than the interquartile range.

2.2. i-Means Analysis

NMR data were clustered using the i-means method, which consisted of two parts. First, data were clustered by k-means and given a class name. Then, random forest was performed using the tentative class information from k-means, and the results for which the rate of accuracy was particularly high were clustered. In this study, the simplest division into two groups was performed.

2.3. Mahalanobis’ Distance

We used the code from https://github.com/kazwd2008/MSD (accessed on 18 January 2023) to find the centroid in the group decided by i-means and to decide the weight of the variables. Mahalanobis’ distance was computed using R. We also defined the data at a distance of one standard deviation of Mahalanobis’ distance or more from the center as representing an unstable state.

2.4. Association Analysis (a Priori)

Association analysis was performed using the R package Arule. Data for a priori were converted using data to zero-one data by ranking, and Top or bottom interquartile of whole variables belonging to “High” or “Low” were “one”; others were zero. This calculation used an a priori algorithm. We set the extraction conditions as follows: support = 0.063, confidence = 0.25, and lift > 1. The association networks were visualized using the free software Gephi.

2.5. Bayesian Network Analysis

Bayesian network analysis was performed using the R package bnlearn [33]. We used the water temperature-related metabolites and microbial phyla obtained by the association analysis for Bayesian network analysis. Additionally, water temperature information was added to each sample. Considering the month in which the sample was taken, information about the experienced water temperature was classified: Cold1 when the temperature was low in the current month, Cold2 when the temperature was low one month ago, Cold3 when the temperature was low two months ago. Similarly, “Hot” water information was added. This non-numerical information was handled as a “factor” in R, and the values of NMR and NGS were handled as “numeric”. A directed acyclic graph was computed using the original data structure (not zero-one) by the hill climb method and bootstrapped 200 times to predict the strength of the edge. The edge strength information obtained by bootstrapping was applied as line thickness to the initial directed acyclic graphs. The results were depicted by the R package Rgraphviz. Additionally, temperature status information was added in blacklist, so that it is not linked to other temperature status information.

3. Results

First, we extracted data on water temperature published by the Japan Meteorological Agency. Since water temperature was reported as daily values, these were transformed into monthly data (Figure 2). This scatter plot shows that the temperature of all sea areas in this study tended to peak around August and reached their nadir from January to March. Since the temperature range was 15 °C–30 °C, we decided to regard 25 °C or more as “hot” water, 20 °C or less as “cold” water, and the range between these as “lukewarm” water.

The shape represents the year. The color shows the measurement location in Figure 1. The point represents the monthly average. The horizontal line in the figure indicates the boundary of the temperature zone between hot and cold water.

Next, we clustered data of fish collected in each sea area with water temperature using the i-means. We attempted to build multiple clusters; however, some types of fish only formed a single cluster; hence, we classified them into two groups. Using the cluster center and each variable weighted by the Mahalanobis’ distance, the distance between each sample and the cluster to which it belonged was computed (Figure 3). The boxplot showed that the clusters were generally well-formed, indicating the clear differentiation of some samples.

Subsequently, individual fish were ranked in the terms of NMR and NGS data and organized into quartiles. An association analysis was conducted with water temperature data (Figure 4). Figure 4A shows that unstable samples, with large deviations in Mahalanobis’ distance, were more common in small Perciformes in a cold environment. A strong relationship between them was also shown via NMR. Low concentrations of anserine were strongly associated with unstable metabolism. However, the steady state, close to the vector center, was observed in hot water environments (Figure 4B). Gut microbe population occurrence, particularly proteobacteria, seemed to be associated with higher water temperature changes.

Finally, we performed the Bayesian network analysis, an artificial neural network, using factors that were linked to unstable or temperature information in association analysis (Figure 5). We decided on a directed acyclic graph from origin data and bootstrapped to predict connection strength. In the NMR data, the fish metabolites that experienced the environmental effects of cold water showed a direct link with serine, valine, inosine, and glycerol (Figure 5A). There was also a strong connection between anserine and beta-alanine. However, the hot water conditions exerted only the direct effect of alanine and valine, but it appears to eventually affect organic acids (Figure 5A). In the result of NGS data, the cold water environment was strongly related to Gram-positive bacteria, such as actinobacteria and firmicutes, and was also weakly related to cyanobacteria (Figure 5B). Alternatively, the experience of the hot water environment appeared to be related to planctomycetes and proteobacteria. Moreover, the current hot water affected various other bacterial groups (Figure 5B).

4. Discussion

We first extracted data from a public database to determine the changes in water temperature (Figure 2). Data showed that even in adjacent sea areas, water temperature could have monthly average differences of 1 °C–2 °C, along with clearer annual variations. The causes of this were mainly the weather and differences in the number of hours of sunshine. The annual range of variation of the water temperature in the seas near Japan was ~15 °C. It was speculated that cold water would limit the efficiency of enzymatic reactions as well as the replication of microorganisms to some extent [34].

The distribution of Mahalanobis’ distances for groups classified was confirmed using i-means. Both studied groups had a data structure close to the center (Figure 3). Similar to the physiochemical state, the stable state of biological phenomena was considered to be the state of less energy dispersion [32]. Therefore, samples with Mahalanobis’ distances close to the center of the group were considered to be in a stable state, whereas those with long Mahalanobis’ distances were in an unstable one. Specifically, we defined samples as being in an unstable state when they were more than one standard deviation away from the center of the group to which they belonged.

To clarify the association between stable/unstable state and water temperature estimated from Mahalanobis’ distances, we analyzed the association rules using an a priori algorithm. We previously reported that this method was able to detect nonnumerical factor association rules from biological states, for example the relationship between creatine content in muscle and animal habitat [31]. Therefore, it is considered possible to analyze data on the relationship between data stability and nonnumerical information, such as water temperature change. Indeed, correlations were found between state stability and ranking of numerical or nonnumerical information. Figure 4A indicates that if the water temperature remained below 20 °C for a while, the growth of fish would be difficult, and fish meat would transition to an unstable state. This might be attributable mainly to the inhibition of enzymatic reactions. It is known that low-temperature conditions may reduce oxidative stress [35]. Our findings indicated that if the water was maintained at a high temperature (Figure 4B), the muscular metabolite state would likely be stable. This supports the hypothesis that enzymatic reactions are important for stabilizing the state. In addition, the intestinal microbiota of fish appears to be affected during the transition between cold and hot conditions, possibly because it acts as an intermediate factor in environment changes.

Finally, we attempted to predict the cascade with temperature at the most upstream by the Bayesian network analysis. In the cold water state, the a priori algorithm showed that anserine could be at low levels, but just before that, it was related to beta-alanine concentrations, which was reasonable because anserine is a compound of beta-alanine and methylhistidine. Therefore, it could be stated that the Bayesian network was working. In other words, fluctuations in inosine and IMP signals, which connected temperature and beta-alanine, are predicted to be involved in the final anserine concentration. However, the a priori algorithm revealed that the microbiota showed a decrease in the abundance of firmicutes and actinobacteria and an increase in the abundance of proteobacteria due to the cold water environment, but the Bayesian network analysis suggested that the decrease in firmicutes and actinobacteria abundance occurred first, followed by the increase in proteobacteria abundance. After that, it was speculated that the abundance of other bacterial groups increased as the water temperature increased from that of the cold water.

5. Conclusions

In this study, we used a machine learning and network analysis approach to investigate the relationship between metabolites in fish muscle, water temperature, and intestinal microbiota by using an NMR-based bioinformatics methodology. Overall we studied the relationship between the environmental factor index from published data by local governments and our laboratory data.

By using our analysis scheme with ANN, it was possible to extract important factors from multiple variables and use environmental factors or their indices to estimate the degree and direction of their impact. Therefore, when it comes to aquaculture, it is possible to estimate the optimum environment for aquaculture.

Our proposal scheme can reveal any cascade that consists of environmental factors and internal fluctuations. In this study, we used water temperature as a representative value of environmental factors, but it is also possible to handle other data, considering that environmental factors influence each other. In addition, various organization collect and publish data including environmental factors, and our approach will help solve the problem not only for the experimental data accumulated so far, but also for the experimental data that will be obtained in the future. Therefore, multivariate ANN analysis including the result of our study will be able to predict the results of many biological effects from the measurement results of various environmental factors.