Taken together, our findings of evolved isolates displaying altered biofilm formation and motility properties and the detection of mutations within genes related to biofilm formation and motility in single evolved isolates as well as across independent endpoint populations indicates that adaptation of B. subtilis to A. thaliana roots under the employed conditions is associated with alterations in these two bacterial traits. While we found that the phenotypic and genetic changes of Ev6.1 and Ev7.3 conferred a fitness advantage over the ancestor during root colonization, adaptation to one certain environment may be accompanied by a loss of fitness in other environments . This has been demonstrated for Escherichia coli which following adaptation to low temperature showed reduced fitness at high temperature . In the example of the evolution of hyper-swarmers of P. aeruginosa, the hypers warmer clones out competed the ancestor in swarming, but lost in biofilm competitions . In this study, we demonstrate that adaptation of B. subtilis to A. thaliana roots is accompanied by an evolutionary cost. When Ev6.1 and Ev7.3 each were competed against the ancestor in LB + xylan under shaking conditions, i.e. an environment where plant compounds are present but biofilm formation is not required for survival, both evolved isolates suffered a fitness disadvantage. The observation that two evolved isolates, from independent populations and with different phenotypes and genetic changes, both suffered a fitness disadvantage in a non-selective environment might suggest the generality of such an evolutionary cost accompanying adaptation to A. thaliana roots. In our EE approach, B. subtilis was adapted to plant roots in the absence of other microbes. In the rhizosphere environment under natural conditions, B. subtilis is far from being the sole microbial inhabitant.
Instead, it engages in cooperative and competitive interactions with other members of the rhizosphere microbiome. We tested whether the evolved isolate, Ev7.3, displaying increased root colonization in the selective environment relative to the ancestor, would also show improved establishment on the root under more ecologically complex conditions. We found that in the presence of a synthetic, soil-derived community, Ev7.3 displayed enhanced establishment on the root compared with the ancestor in two out of four inoculation ratios. This enhanced establishment on the root by Ev7.3 is not expected to be caused by altered antagonistic activities toward the community members. First,hydroponic bucket no major changes in the inhibition of the community members were observed in confrontation colony assays. Secondly, an increased number of Ev7.3 cells on the root did not cause a reduction in the co-colonizing community. Finally, Ev7.3 did not harbor mutations in genes directly related to secondary metabolite production. Instead, enhanced establishment on the root by Ev7.3 in the presence of the community is possibly enabled by robust biofilm formation facilitating stronger attachment to the root and enhanced utilization of plant compounds. Interestingly, a study by Molina-Santiago et al. showed that compared with a Dmatrix mutant, co-inoculation of B. subtilis WT with Pseudomonas chlororaphis on melon leaves enabled co-localization of the two species as well as the closer attachment of B. subtilis to the left surface . The robust biofilm formed on the root by Ev7.3 possibly facilitated by increased matrix production may thereby not exclude the community members on the root but could rather allow them to incorporate into the matrix. This could also explain why the enhanced establishment of B. subtilis Ev7.3 on the root did not cause a reduction in the number of community cells attached to the root. Alternatively, the community may not be majorly affected by any difference in the establishment on the root between the ancestor and Ev7.3 owing to the low abundance of B. subtilis relative to the community. Further work is needed to elucidate the interactions between B. subtilis and this synthetic community during root colonization. In summary, these findings suggest that even though B. subtilis was evolved on A. thaliana in the absence of other microbes, it became highly adapted to the plant root environment enabling better establishment on the root also when the ecological complexity increases.
How genetic adaptation to the plant root in the absence of other microbial species differs from adaptation to plant root environments with varying levels of ecological complexity is the scope of future studies.From an applied perspective, experimental evolution of B. subtilis on plant roots represents an unexplored approach for developing strains with improved root attachment abilities for agricultural use. However, a biofilm-motility trade off as observed here may be undesirable when developing biocontrol agents owing to the growing evidence of motility as an important trait for bacterial root colonization in soil systems . The phenotypes associated with the adaptation of B. subtilis to A. thaliana roots presented here as well as the accompanying evolutionary cost and the increased root colonization also in the presence of resident soil bacteria highlight the importance of considering the selective environment if evolving PGPR for biocontrol purposes.Aquaponics is an agricultural technique that is touted as a promising alternative to solve the food and environmental crisis that the world faces today. As it is a combination of aquaculture and hydroponics , this technique can have high water use efficiency, and reduce the dependency on pesticides and fertilizers which adds to the sustainability of the system. This system has been shown to consume 50–70% less water compared to traditional agricultural systems, owing to its recyclability. Moreover, this system requires less pest control and has proven to be less impacted by harsh weather conditions, leading to yield increase. As we are moving towards the era of digital agriculture, efforts have been made to design physical prototypes for smart commercial aquaponic systems. Lobanov et al. has focused on proposing a new system for reducing the carbon and nitrogen content from fish waste to produce liquid fertilizer which is added to facilitate the growth of lettuce in aquaponic set-ups .
Karimanzira et al. introduced the concept of building an intelligent aquaponics system incorporating predictive analysis, system optimization and anomaly detection for maximizing productivity in commercial aquaponics through early fault detection . Mahanta et al. formulated a laboratory set-up as a prototype to grow soybeans in hydroponic solution using plasma activated water to decrease the amount of heavy metal uptake and optimize yield . Rau et al. designed a smart IoT based sensing and actuation system for growing rice by controlling the concentration of magnesium and nitrogen in hydroponic solution along with monitoring the environmental parameters of the greenhouse . A similar design was proposed by Dhal et al. using a smart IoT system for real time sensing and regulation of nutrients in commercial aquaponic set-ups depending on the season in which the crops were grown . Timsina et al. proposed the use of Machine Learning models to regulate nutrients for growing cereals in farmlands, but efforts are yet to be made for growing them in aquaponic set-ups .There have been recent advancements in the field of Smart Aquaponics which involve monitoring environmental parameters as well as plant growth through different Machine Vision-Based approaches in an IoT environment . Arvind et al. implemented an AutoML model trained with an XGradient boost algorithm, with 10-fold cross-validation taking into account the different sensor values recorded in the greenhouse and the fish count in the aquaponic tank which was extracted using the mask R-CNN image segmentation. This algorithm was used to control the triggering of the actuators in the system that in turn control the environmental parameters in the greenhouse; ensuring significant improvements in yield and water conservation, when compared with the conventional methods. Languico et al. did a comparative study of three ML estimators: K-Nearest Neighbours , Logistic Regression , and Linear Support Vector Machine , on the visionfeature extracted images of lettuce in a smart aquaponics set-up to monitor diseases that the crop may incur in its lifetime. A similar kind of study was done by Maleki-Kakelar et al. where multiple ML algorithms like linear and quadratic regression models, fuzzy systems and genetic programming was used to conduct regression analysis for improving urease activity aimed at strengthening the behaviour of soils . A study on images of lettuce leaves was conducted by Concepcion II et al. to detect diseases, wherein different feature selection processes were used to select the top four attributes for training the ML models .
A study on smart nutrient regularization for replenishing the Nitrogen, Phosphorus and Potassium content of the soil has been done by Ahmed et al. using genetic algorithms to provide the optimal level of nutrients needed for high production level of crops . Hiram Ponce et al. used a combination of Convolutional Neural Network for extracting features from tomato leaves along with a combination of Artificial Hydrocarbon Network as the dense layer to predict deficiency of nutrients in tomato plants . A similar kind of Deep Neural Network was implemented by Yadav et al. for apple foliar disease classification using Plant Pathology image data-sets . Nevertheless, limited research has been conducted on monitoring and regulating nutrient concentrations in the aquaponic solution using ML-based approaches. The current work focuses on building a ML algorithm that monitors the nutrient status of hydroponic irrigation water and outputs a recommendation system for regulation of these parameters. The main motivation behind this entire approach is to select the most important nutrients that need to be regulated in aquaponic environments depending on the output of Machine Learning classifiers trained on small data-sets. The main issues which one may face while designing such a data-driven approach has been discussed in the next paragraph. One of the major challenges with automation in aquaponics is the lack of sufficient data and the vast number of predictors that have to be used for making any inferences. This could result in what is referred to as the “Curse of Dimensionality” leading to the available data becoming sparse . Thus, it becomes extremely important to reduce the dimensionality of the data-set without losing valuable information. For this, stackable planters feature selection techniques become more relevant. To state a few, Recursive Feature Elimination and ensemble techniques such as the Extra Trees Classifier have proven to be highly effective. It is also equally important to check for the separability of the classes in the data-set. Many data visualization techniques like Principal Component Analysis and Multi-Dimensional Scaling plots can be used to understand how linearly separable the data is and what classifiers would be best suited for the purpose. Another major drawback especially in the case of small data-sets is the problem of “over-fitting” the data . Traditional ML and Deep Learning algorithms have a high probability of performing poorly on small data-sets.
A solution to this problem was suggested by Shao et al. who proposed deep Reinforcement Learning algorithms named MONEADD and did a comparative study with Knapsack and Traveling Salesman problem to design neural networks for different combinatorial optimization problems without much feature engineering which showed better scalability when distributed on multiple GPUs . On a similar note, to address this problem of over-fitting, Braga-Neto et al. proposed Bolstered Error Estimation method, which uses the same data for both classifier design and error estimation . In this form of error estimation, the variance setting for the Bolstering kernel is determined in a non-parametric manner from the data. For all the linear classification rules, the integrals in the Bolstered error estimation are computed in the same way. For the non-linear classification rules, a small number of Monte-Carlo samples are generated. In this study, three types of Bolstered error estimators, namely the Gaussian Bolstered re-substitution error estimator, semi-Bolstered re-substitution error estimator, and the Gaussian Bolstered Leave-One-Out error estimator, have been used with linear classifiers like LDA and SVC, along with non-linear classifiers like CART and KNN, and their results have been compared to identify the best performing classifier in this case. Based on the performance of the classifier with utmost optimal performance, a set of recommendation rules were prescribed and a comparative study was done on how this proposed Machine Learning based approach resulted in more optimal yield as compared to the baseline model.The data-set which was used in this case was recorded from three commercial aquaponic facilities located in Caldwell, Bryan, and Grimes counties in Texas, USA which are large producers of lettuce and other greens.