In the face of the potentially negative consequences of climate change on agriculture, all avenues of mitigation must be examined, and even small improvements may prove worthwhile.Bio-fuel crops have been developed as an alternative, carbonneutral energy source, among which the perennial C4 grass, Panicum virgatum L. , native to North America, can adapt to a wide range of environments, including those with marginal soils and low water input. However, in order to better manage and optimize this crop for bio-fuel production, it is important to understand the mechanisms that enable its adaptivity, and how nutrient-poor environments impact chemical composition, biomass yield and feed stock quality. A long-standing barrier to this mechanistic understanding lies in the difficulty in characterizing plant chemical composition and quantifying plant-available nutrients at the rhizosphere. Phosphorus is a critical nutrient, and poor P management poses a global risk for environmental sustainability and food security. P limitation severely restricts photosynthesis and reduces CO2 fixation, but upregulates pathways associated with organic acid/ carboxylate exudation. P limitation can also be associated with increasing biosynthesis of defense metabolites, such as increased lignification of cell walls,hydroponic vertical farming systems suggesting that changes in plant carbon allocation in response to P limitation may alter both the yield and the chemical composition of bio-fuel feed stocks, and therefore productivity.
In this light, it may be beneficial to monitor plant available P concentration and plant chemical composition during the growth season, with the goal of improving biomass production and optimizing the chemical composition for improving feed stock quality through active land management, especially when growing in marginal soils. Quantifying soil P available to plants is challenging, especially if attempting to do this dynamically during a growing season. Typical chemical extraction methods quantify only a fraction of the inorganic P pool and are typically measured in top soils prior to planting. Although the P concentration data obtained with these methods have been widely used in the literature to represent total P availability, they are not an accurate measure of P available for plant growth. Perennial grasses such as switch grass produce deep roots that explore and obtain nutrients and water from distinct locations deep into the soil, and these locations vary across the growing season and lifetime of a plant. Further, plants have developed a number of strategies to access P from different types of soil, including the adaptive secretion of compounds such as organic acids, enzymes and side rophores which either mobilize soil P directly, or indirectly through their stimulation of the rhizosphere microbiome and symbiotic fungi. Combined, dynamic growth of roots through a soil profile with distinct concentrations and chemical forms of P, an adaptive allocation of photosynthate below ground, and a micro-biome with typically unknown capacity for P mobilization, makes predicting plant available P a highly complex task.
Meanwhile, there has been renewed interest and some success in predicting plant nutrient levels using spectroscopic methods for remote sensing with the help of machine intelligence. A variety of machine learning tools have been utilized to achieve satisfactory results with independent variables obtained, for example, by visible to near-infrared spectroscopy, to predict nutrient levels in shoots in agricultural crops, total nitrogen content of soils and plant adaptive responses to stress. Most of these tools are linear models such as partial least squares regression, principal component analysis, or support vector machines often with a nonlinear kernel, likely due to their inherent robustness and reduced chance of over fitting.Thus, it is conceivable that a machine learning approach could predict nutrient availability by monitoring the biochemical signatures of plant shoots. However, to our knowledge, this aspect has not been well explored. In this paper, we use a molecular spectroscopic method to determine and quantify the organic P and inorganic P in leaf tissue. This approach also provides important information on overall plant tissue biochemistry that can be used as multi-plex signatures of a plant’s response to environmental conditions. P speciation can be quantified dynamically and feed stock quality for bio-fuel production can be inferred. We then use this tissue biochemical information to infer and evaluate plant-available P using a machine-learning model trained using a dataset from a controlled laboratory experiment. Building off this approach, we used the model to interpret plant spectral data from two field locations where contrasting available P was expected.A series of experiments in sand cultures were performed to evaluate the dose-response of plant tissue chemistry to varying N and P. The chemical signatures in plant leaves varied substantially with P and N availability in the growth media, as shown in Fig. 1. Note that the absorbance data were normalized to the maximum to show the relative concentration changes on the same scale. We observed higher cellulose, lower lignin, lower lipids, and higher organic and inorganic phosphate concentrations in the leaves of plants grown in solution with closer to optimal P concentration, and increased lipid and amide concentrations in solution with closer to optimal N concentration.
The cellulose/ lignin ratio was very sensitive to P concentration, showing a 3-fold increase from <150 μM to 500 μM P, but was not consistently sensitive to N concentration. Note that the P/N concentration likely fluctuated during plant growth and between the fluid replenishment, thus the values here refer to the average concentration through the growth. Since the C/L ratio is an important metric for bio-fuel production, we suggest that higher P concentrations would produce a higher C/L ratio for potentially increased bio-fuel yield. The higher relative amide concentration in plants grown in the lowest concentration of P compared to those in intermediate P concentrations likely reflected severe P-stress in these plants, as soluble nitrogenous compounds including amino-acids and amides accumulate in other species under P-deficiency, and consistent with our other observations. P deficiency was also associated with higher lipid content, as indicated by the increased signal from carbonyl bonds, possibly related to the production of triacylglycerides as storage compounds under P limitation. Note that while lignin concentrations increased gradually with decreasing P , cellulose concentrations showed a threshold effect with a large increase between 30 and 150 μM P, and these opposing responses manifested in a cellulose/lignin ratio that is highly sensitive to P deficiency, but not to N deficiency. The increase in lignin concentration under P deficiency was possibly related to induction of defense genes and defense metabolites and the overall shift to lower cellulose and more lignin may represent a more pathogen-resistant, rigid cell wall.Because of the strong dependence of feed stock chemical composition on soil phosphorus concentration in the controlled growth experiments, we evaluated switch grass growth at two field locations contrasting in soil P availability. The soil texture of these locations differs with the RR site being a sandy loam and 3rd Stbeing a silt loam. Chemical characterization of bulk soil samples indicated a significantly higher Mehlich-III extractable P concentration in RR soils relative to 3rd St , with no significant seasonality observed . In general, plants grew taller in the RR plot than at 3rd St, reaching maximum heights at T4. Our leaf-tissue measurements showed that leaf Pi concentration increased over the course of the growing season in both field experiments. The Pi concentration of leaf tissue collected at RR was higher than that at 3rd St , consistent with the bulk soil characterization,vertical grow rack although the plant Pi as determined by FTIR was more similar than the extractable soil P data might have suggested. This may be expected due to P homeostasis and overall biomass difference. Pi concentration showed an earlier increase in plants at Red River , presumably reflecting greater uptake early on due to higher concentrations of plant-available phosphate in the soil. Further increase in the concentration of Pi, especially in the late season along with the decrease of Po at RR may reflect mobilization of Pi for translocation elsewhere in the plant, including storage tissues that support regrow in the next growing season. The trend of organic phosphates in the lower panel of Fig. 3 shows that, contrary to that of Pi, the concentration of Po plateaued in the later stages of growth at around T4, when the plant reached maximal biomass as indicated by the maximal plant heights . Po concentration decreased significantly during senescence in plants grown on the higher P soils at Red River. This explains the transient sharp increase in Pi described above and the large decrease of Po and total P . Seasonal increase of total P content in the shoots of switch grass has been observed before35, but our spectroscopic method enabled us to dissect P speciation during the growth season.
Meanwhile, we observed a similar trend in concentration of lipid signature in the late growth stage . Since the leaves we collected tended to be younger leaves to be consistent with our sand-based experiments, the maximum P concentration at T4 may reflect a combination of P uptake over the growth period, plus reallocation from old to younger leaves, resulting in higher concentrations of major P-containing molecular classes like phospholipids and/or ribosomal RNA.Because of the critical role of P in the growth of switch grass and its strong correlation with biochemical composition for this bio-fuel species, we believe the seasonal characterization of plant available P in the rhizosphere and P speciation may be beneficial for crop management and improved environmental outcomes. We demonstrate here that a machine-learning model can be used to quantify P availability using the plant leaves themselves as sensors. Since the nutrient concentration in the rhizosphere in a hydroponic substrate is relatively well controlled, this experiment allowed us to develop training data for an ML model. We achieved a principal component regression model with a high R2 of ~1 , which allows us to predict plant-available P concentrations based on the spectral data collected on the leaf tissue from field-grown plants. The predicted P concentrations are shown in Fig. 4a. Note that in a more traditional approach, the model prediction would be further validated by another independent method to evaluate the model’s accuracy. However, such a method for accurate estimation of bio-available P concentration through the soil profile over time does not yet exist in practice. We believe that our model contains an accurate statistical description of the correlation between the P concentrations in the growth media and all the spectral features in younger leaf samples, given the high accuracy achieved with large concentration range and the high affinity of P uptake; thus this model can be used for prediction of the P concentration available to each plant within the rhizosphere. The predicted P concentration available to the plants shows a gradual increase and then a sharp dip in T4 when the plant reached maximal biomass, reflecting an increase in P uptake at T4 and a quick decrease in P uptake at T5, when the shoot senesces. As a perennial plant, switch grass remobilizes and stores P in roots to support the subsequent year’s growth, consistent with previous observations with P remobilization efficiency ranging from 31% to 65% in different ecotypes. The increase in Pi in the tissue at the later stage of growth, the strong correlation of cellulose content with total P concentration, and the large reduction of plant available P in the rhizosphere provided us with a clear picture of the interaction of P availability and tissue composition during the life cycle of switch grass. This may be a consideration in the timing of harvest to achieve optimal bio-fuel yield and reserve P in the root for the optimal growth in the next year. The total P concentration in the roots followed a similar seasonal trend, showing a reduction of total P concentration near the period of maximal growth and at least a partial recovery in T5. The P concentrations measured in roots by ICP-MS at T5, a point at which plants had begun senescing, were similar across the two field locations , with a mean value of 729 ppm at 3rd St and 703 ppm at RR, respectively. The similar root P concentrations at T5 may be reflective of reduced plant P demand during senescence as well as plant P re-allocation and is described further below.