Evidence for Soil Phosphorus Resource Partitioning in a Diverse Tropical Tree Community

Abstract

Soil phosphorus (P) partitioning could contribute to species diversity and structure in plant communities, but field-scale evidence for P partitioning remains scarce. We hypothesized that the presence of P partitioning could be inferred from statistical associations between the spatial distributions of plants and chemical forms of bioavailable soil P. We investigated this in a diverse tropical tree community on Barro Colorado Island, Panama. We quantified potentially bioavailable forms of soil P by extraction in 2 mM citric acid followed by treatment with phosphatase enzymes. We then linked these P forms to the distribution of 189 tree species in a 50 ha forest dynamics plot by testing species–P associations against null models of random dispersal. We found that 20% of tree species were significantly ($\alpha$ = 0.05) associated with the depletion of at least one soil organic P fraction, although around half of these associations might be false rejections of the null hypothesis due to type I error. Species in the Fabaceae (legumes), which are known to express high rates of phosphatase in their roots, were most frequently associated with soil P fractions. We interpret our findings as evidence of widespread P partitioning at the community scale, affecting a relatively small proportion of tree species in this moderately fertile forest. We predict that stronger evidence of partitioning will be found at sites with lower P availability.

1. Introduction

Phosphorus (P) is an essential plant nutrient of key importance in terrestrial ecosystems [1,2]. Although dissolved orthophosphate is the most readily available P compound [3], plants are also able to use soil organic P compounds following hydrolysis by phosphatase enzymes and subsequent uptake of the released orthophosphate ion [4]. Organic P exists in soils in a multitude of different compounds [5] and is a major P pool in tropical forest soils [6]. For instance, organic P in lowland tropical forest soils in Panama accounts for about 30% of their total P, mainly in the form of phosphomonoesters such as sugar phosphates and inositol phosphates, with smaller amounts of phosphodiesters (i.e., nucleic acids and phospholipids) and, in a few soils, low concentrations of phosphonates [6]. Most soils also contain small amounts of pyrophosphate, an inorganic phosphate that is hydrolyzed by phosphomonoesterase, which is also common in dissolved form in soil solution [6].

The hydrolysis of organic P by phosphatases is governed by the nature of the bond between P and the organic moiety [7]. Thus, the availability of organic P compounds depends in part on the type of phosphatases that a plant species can synthesize. Specialization in the use of different organic P forms by co-occurring tree species might therefore promote the coexistence of species in diverse communities in low-fertility soils [8]: while some species may specialize in synthesizing phosphomonoesterase to exploit P bound in phosphomonoesters, others may use phosphomonoesterase and phosphodiesterase to hydrolyze phosphodiesters, and still other species may synthesize phytase to hydrolyze myo-inositol hexakisphosphate (phytic acid). Resource partitioning might therefore contribute to ecosystem diversity by offering additional environmental niches for plant species [8], particularly in strongly weathered soils of the tropics [9,10]. In such soils, the concentration of orthophosphate in soil solution is often undetectable [11], so plants are more likely to compete for soil organic P.

Recent tests of soil P resource partitioning among tropical trees grown in pots have revealed that coexisting species differ in their ability to use different forms of organic P [9,10,12]. Phosphorus partitioning has also been demonstrated in experiments in grasslands: a greenhouse experiment with constructed grasslands showed that P resource partitioning influenced species abundance [13], while ³³P-labeled organic phosphates were used to demonstrate P partitioning among co-occurring plants in a grassland soil [13,14]. To date, however, no attempt has been made to assess the effect of soil P resource partitioning on the community-wide coexistence of plants at field scale.

Here, we assessed niche assembly due to soil P resource partitioning in a lowland tropical rain forest tree community by linking soil P forms to the distributions of individual tree species in a large forest dynamics plot in a diverse tropical forest. We hypothesized that soil P resource partitioning would result in associations between the spatial distribution of tree species and soil organic P forms. The underlying assumption is that individuals of a species that specialize in the utilization of a particular organic P compound will be associated with low levels of that compound as they deplete it preferentially from the soil. We accounted for plant–resource associations arising from dispersal limitation rather than P partitioning by comparing species–P associations against null models of dispersal assembly.

2. Materials and Methods

2.1. Study Site

We tested our hypothesis on the ForestGEO 50 ha forest dynamics plot on Barro Colorado Island (BCI) in the Republic of Panama (Figure 1). The site was chosen because our methodological approach required a biodiverse plot where individual trees were exhaustively mapped across a sufficient area to reflect the spatial distribution of tree species (e.g., clustering). The plot is located on a plateau 120–160 m above sea level [15]. Soils are Oxisols with kaolinitic mineralogy [16]. Concentrations of Mehlich-III-extractable nutrients range from 2-fold to several orders of magnitudes and exhibit a distinct spatial pattern [17,18]. The climate of BCI is characterized by a long rainy season and a pronounced dry season from mid-December until late April or early May [19]. Annual rainfall averages 2612 mm (years 1925–1986) and mean annual temperature is 26 °C, with mean monthly temperature varying by <1 °C over the annual cycle [20]. The plot supports predominantly old-growth semideciduous moist tropical forest [15,21]. The plot has been censused every 5 years since 1980, with every tree (>1 cm diameter at breast height) mapped and identified to species level [22,23]. The census in 2015 recorded 297 species among 207,719 individuals [24,25]. The effect of soil nutrients on the spatial distribution and growth rates of tree species in the plot has been extensively investigated [17,18].

2.2. Soil Sampling and Preparation

Soils were sampled during two weeks in the middle of the rainy season (25 July to 8 August 2018). We sampled on a regular 50 × 50 m grid where each alternate point was paired with an additional point at a shorter distance, yielding a total of 300 sample points [17]. At each point, we combined five 2.5 cm diameter soil cores to 10 cm depth from the corners and center of a 20 × 20 cm quadrat to yield a single composite sample. Samples were air-dried for two weeks and sieved (<2 mm), and roots were removed.

2.3. Determination of Bioavailable Soil Phosphorus

Potentially bioavailable soil P fractions were determined by a sequential procedure involving extraction with citric acid followed by fractionation of P forms by phosphatase hydrolysis [30]. The procedure is designed to mimic the mechanism by which plant roots obtain P from the soil, first by solubilizing organic P compounds through organic anion secretion, followed by release of free inorganic orthophosphate by phosphatase hydrolysis. Previous work suggests that tree species in the 50 ha plot rely on organic acids to solubilize P [18]. We used two different combinations of phosphatase enzymes to subdivide the extracted organic P into fractions relevant for P resource partitioning.

To extract P fractions, 6 g of air-dried soil was extracted with 30 mL of 2 mM citric acid (pH = 5.2) by shaking for 1 h at 180 oscillations min⁻¹. Samples were centrifuged at 8000× g for 10 min and the supernatant was used for further analysis. Reactive P (RP) in the extracts, which approximates inorganic orthophosphate, was determined colorimetrically by the molybdenum blue method [31]. Total P (TP) in the extracts was determined by acid persulfate digestion [32] and molybdenum blue colorimetry. Subtracting RP from TP gave the unreactive P (UP) in the extracts, which includes organic P and inorganic polyphosphates. The UP was further fractionated by enzyme hydrolysis [30] using the following enzymes (obtained from Sigma Aldrich, Saint Louis, MO, USA) in sodium acetate buffer: (1) 0.01 units mL⁻¹ phosphomonoesterase from Escherichia coli (EC 3.1.3.1, Sigma P5931) to determine the monoesterase-hydrolyzable phosphorus (MP); this fraction includes simple phosphomonoesters such as sugar phosphates, mononucleotides, and pyrophosphate. (2) A total of 0.01 units mL⁻¹ nuclease from Penicillium citrinum (EC 3.1.30.1, Sigma N8630) in combination with 0.01 units mL⁻¹ phosphomonoesterase from E. coli to determine the nuclease-hydrolyzable phosphorus (NP); this fraction contains phosphodiesters in the form of nucleic acids. (3) UP that was not amenable to hydrolysis by either the monoesterase or the nuclease was termed non-hydrolyzed phosphorus (NHP); this fraction is likely to include recalcitrant compounds such as higher-order inositol phosphates.

2.4. Statistical Analysis

We used the test procedure proposed by John et al. [17], which we call J-test, to quantify significant associations between P fractions (RP, MP, NP, and NHP) and tree species. Applying the J-test involved three steps: (1) based on the point soil samples, a model of the spatial distribution of each soil P fraction was used to obtain spatial predictions for the whole plot area; (2) tree census data (year 2015) were used to estimate point process models (PPM) of random dispersal for each species. (3) the models of step (1) and (2) were used to test each species–P fraction combination for associations using a Monte Carlo procedure (the actual J-test).

Statistical analysis was conducted using the R software environment (version 3.6.3; [33]) with the R packages geoR (version 1.8-1; [34]) for step (1) and spatstat (version 1.63.3; [35]) for step (2) and (3), as well as raster (version 3.0-12; [36]) and our own R functions that we bundled in the package ppartition (version 1.0.0; [37]).

2.4.1. Spatial Models and Predictions of Soil Phosphorus Fractions

Kriging [38,39,40] was used to obtain spatial predictions for each soil P fraction. Phosphorus concentrations below detection limit were set to half detection limit to allow them to be included in model parameter estimation and kriging. We used diagnostic plots (histograms and box plots) and the octile skew [41,42] to determine the need for transformation. Rawlins et al. [43] suggest that a transformation should be applied if the octile skew is outside the bounds of –0.2 and 0.2. Given exploratory data analysis, we Box–Cox [44] transformed (transformation parameter $\lambda$ set to 0 or 0.5) soil P concentrations (mg P kg⁻¹) and used a linear mixed spatial model (as formalized by Nussbaum et al. [45]) of transformed soil P concentration (P) at location $(x,y)$:

$$P(x,y)=\mathit{s}{(x,y)}^T\mathit{\beta }+S(x,y)+ϵ$$

where $\mathit{s}{(x,y)}^T\mathit{\beta }$ is the trend that contains covariate vector $\mathit{s}$ and coefficient vector $\mathit{\beta }$. The trend was either set to a constant coefficient (equivalent to no trend) or modeled as a 1st- or 2nd-order polynomial in the coordinates. $S(x,y)$ is a stationary autocorrelated Gaussian random field with zero mean and isotropic variogram. $ϵ$ is a normal mutually independent zero mean error with variance ${\tau }^2$ (commonly called nugget).

The model parameters were estimated using restricted maximum likelihood [46,47,48]. Kriging predictions and standard errors were eventually back-transformed to the original scale. Block predictions were received by averaging regularly spaced point predictions within 10 × 10 m blocks [49].

Model quality was evaluated by visual examination of the variogram model fit to the sample variogram as well as leaving-one-out cross-validation on the transformed scale. We calculated two measures of model accuracy [50], normalized root mean squared error (NRMSE) and mean squared deviation ratio (MSDR):

$$\mathrm{NRMSE}=\sqrt{\frac{{\sum }_{i=1}^n{\left\{p(x_i,y_i)-\widehat{P}(x_i,y_i)\right\}}^2}{{\sum }_{i=1}^n{\left\{p(x_i,y_i)-\overline{p}\right\}}^2}}$$

$$\mathrm{MSDR}=\frac{1}{n}\sum _{i=1}^n\frac{{\left\{p(x_i,y_i)-\widehat{P}(x_i,y_i)\right\}}^2}{{\widehat{\sigma }}_K^2(x_i,y_i)}$$

where $p(x_i,y_i)$ is the ith observation, $\overline{p}$ is the mean of all observations, $\widehat{P}(x_i,y_i)$ is the ith kriged point prediction, and ${\widehat{\sigma }}_K^2(x_i,y_i)$ is the kriging variance at $(x_i,y_i)$.

The NRMSE is the ratio between the errors of the spatial model and the errors of the mean model. Ideally, our spatial model should be at least better than the mean model and we should thus aim for an NRMSE considerably smaller than one. The MSDR is the ratio between observed and predicted kriging variance. Kriging only informs us reliably about the prediction uncertainty if observed and predicted kriging variance are approximately equal and thus the MSDR should be close to one [50].

2.4.2. Point Process Models of Tree Species

We used spatial point processes (also known as point process models; PPMs) to model the spatial distribution of species individuals under the null hypothesis of dispersal assembly [17,51,52,53,54,55]. The locations of the individuals of a species constitute a spatial point pattern, which can be regarded as the realization of a spatial point process [56]. Only species with at least 50 individuals were used for modeling (n = 189). Depending on the spatial arrangement of a species point pattern, we modeled the spatial distribution either with a Poisson point process or with cluster processes, which predict random or aggregated placement of species individuals in space, respectively [55]. If a species point pattern was significantly aggregated, as assessed by a Monte Carlo test using Ripley’s K [57] as a test statistic [55,56], we fitted four different types of cluster processes to a species point pattern: Thomas [58] process (modified as suggested by Diggle [59]), Matérn [60] process, Cauchy process [61], and variance gamma process [61]. We then chose the best-fitted cluster process to model the spatial distribution of a species. If a species point pattern was similarly well fitted by several cluster models, we arbitrarily preferred the Thomas process, as this is the cluster model that was chosen by John et al. [17] to represent aggregated point patterns. The goodness of fit of the PPMs was assessed by the Diggle–Cressie–Loosmore–Ford (DCLF) test [62,63,64] and simulation envelopes [65] using the empty space function and the nearest neighbor distance function as summary statistics [35].

2.4.3. J-Test

The test of species–P association, the J-test [17], is based on an index of association:

$$I{A}_i=n_i{\widehat{\Omega }}_i$$

where $n_i$ is the number of individuals of a species in block i and ${\widehat{\Omega }}_i$ is the kriged block prediction in block i for a P fraction.

The J-test actually consists of two tests. The first test, called the $IA$ mean test, is based on a test statistic that is given by the mean of the $IA$ ($I{A}_{mean}$) over the whole plot area:

$$I{A}_{mean}=\frac{1}{N}\sum _{i=1}^k{n}_i{\widehat{\Omega }}_i$$

where N is the total number of individuals of a species in the plot and k is the total number of kriged blocks.

The second test, called the $IA$ standard deviation test, is based on a test statistic that is given by the standard deviation of the $IA$ ($I{A}_{sd}$) over the whole plot area:

$$I{A}_{sd}=\frac{1}{N}\sqrt{N\sum _{i=1}^k{n}_i{\widehat{\Omega }}_i^2-\sum _{i=1}^k{\left(n_i{\widehat{\Omega }}_i\right)}^2}$$

We used the fitted PPM of a species to simulate the spatial point patterns (1000 simulations for each species) and calculate the distribution of both test statistics under the null hypothesis of dispersal assembly. We then calculated the test statistics for the observed spatial pattern and used a two-sided test for $I{A}_{mean}$ and a one sided-test for $I{A}_{sd}$ (observed value smaller than the distribution of values under the null) to assess the significance of the species–P association. Species–P association was only denoted as significant if both test results were significant.

2.4.4. Comparison with Species Phosphorus Affinities

In accordance with the prediction that P resource partitioning is of particular importance in environments where the concentration of plant-available P is low [8], we hypothesized that species of the 50 ha forest plot with significant associations to organic P fractions would be low-P specialists. Hence, we compared our set of significant associated species with species P affinity values reported by Condit et al. [66], who found that plant-available P (extracted by anion exchange membranes) is a strong predictor of species occurrence in 1 ha plots across a soil P gradient in Panama. We classified low-P and high-P specialists as species with P affinities <−0.8 and >0.8, respectively [66].

3. Results

3.1. Geostatistics of Phosphorus Fractions

Concentrations of citric-acid-extractable P fractions varied up to 300-fold (Figure 2). This is remarkable given that a previous study of enzyme-labile organic P of five Panamanian soils spanning a strong P gradient reported a much lower 44-fold variation [30]. However, the magnitude of variability is mainly due to a few outliers. The dataset of citric-acid-extractable P concentrations is available online [67].

Linear correlations among P fractions and with topographical features were weak (<0.4; see Supplementary Materials File: Table S1), with the exception of a moderate correlation between NHP and elevation (0.53). Experimental semivariances of Box–Cox transformed concentrations revealed spatial autocorrelation in each P fraction up to a distance of >200 m (Figure 3). Variogram maps indicated isotropy for all P fractions (see Supplementary Materials File: Figure S1). Parameter values of the fitted spatial models are given in Supplementary Materials File: Table S2. For the NHP fraction only, we included a second-order trend in the spatial model of transformed P concentration.

The nugget-to-sill ratio of all four spatial models was high (0.46–0.53, Supplementary Materials File: Table S2), causing smooth prediction surfaces and high kriging standard errors (Figure 4). Indeed, the level of the kriging standard error was similar to the predictions, except for NHP, where the inclusion of trend in the model increased the confidence in the predictions strongly. The cross-validation statistic NRMSE (Supplementary Materials File: Table S2) reflects this pattern of prediction uncertainty, with a high value for all P fractions (0.88–0.95) apart from NHP (0.64). Given cross-validation, the spatial models accurately estimated the kriging standard error, as indicated by a MSDR of 1.0 for each model (Supplementary Materials File: Table S2).

3.2. Point Process Models of Tree Species

The spatial patterns of 186 species were tested as significantly aggregated. We therefore fitted the aggregated 186 species by a cluster PPM, whereas the remaining 3 species were fitted by a Poisson PPM. However, the goodness-of-fit assessment (i.e., DCLF test and simulation envelopes [65]; see Supplementary Materials File: Figure S2) revealed that only 182 PPMs were fitted well to the point patterns. We therefore excluded seven species from the J-test. Model type, parameter estimates, and goodness-of-fit statistics of PPMs are given in Supplementary Materials File: Table S3.

3.3. Species–Phosphorus Associations

We tested 182 species with well-fitted PPMs for soil P associations. Of these, 36 species (20%) at the 0.05 level and 58 species (32%) at the 0.10 level were significantly associated with at least one of the four P fractions (

Table 1. p values of significantly associated species ($\alpha$ = 0.05) of the 50 ha forest dynamics plot on BCI. Only species with p value of the $I{A}_{mean}$ test ($p_{mean}$) and p value of the $I{A}_{sd}$ test ($p_{sd}$) below the significance level (indicated by bold print) for a phosphorus fraction are listed. High- or low-P specialists according to P affinity values by Condit et al. [66] (reported in [68]) are indicated. Taxonomic information is based on [69] and subfamily names on [24].

Family	Species	Authority	RP		MP		NP		NHP		P Specialist
			$p_{\mathit{mean}}$	$p_{\mathit{sd}}$	$p_{\mathit{mean}}$	$p_{\mathit{sd}}$	$p_{\mathit{mean}}$	$p_{\mathit{sd}}$	$p_{\mathit{mean}}$	$p_{\mathit{sd}}$
Anacardiaceae	Spondias radlkoferi	Donn. Sm.	0.933	0.633	0.018	0.008	0.422	0.200	0.060	0.036	high P
Annonaceae	Guatteria lucens	Standl.	0.144	0.059	0.002	0.001	0.649	0.317	0.002	0.001
Annonaceae	Mosannona garwoodii	Chatrou & Welzenis	0.505	0.185	0.018	0.007	0.104	0.049	0.050	0.014
Annonaceae	Unonopsis pittieri	Saff.	0.507	0.251	0.144	0.084	0.480	0.736	0.044	0.022
Apocynaceae	Tabernaemontana arborea	Rose ex J. D. Sm.	0.404	0.249	0.392	0.209	0.018	0.020	0.951	0.399
Burseraceae	Protium costaricense	(Rose) Engl.	0.366	0.164	0.068	0.034	0.991	0.496	0.048	0.017
Burseraceae	Protium panamense	(Rose) I. M. Johnst.	0.160	0.075	0.052	0.027	0.018	0.007	0.072	0.033	low P
Burseraceae	Protium stevensonii	(Standl.) Daly	0.865	0.371	0.174	0.074	0.637	0.282	0.034	0.017
Burseraceae	Trattinnickia aspera	(Standl.) Swart	0.400	0.816	0.304	0.146	0.236	0.135	0.036	0.018	low P
Cannabaceae	Celtis schippii	Standl.	0.917	0.477	0.106	0.053	0.284	0.809	0.010	0.003
Celastraceae	Monteverdia sieberiana	(Krug & Urb.) Biral	0.613	0.680	0.036	0.019	0.777	0.547	0.044	0.029
Chrysobalanaceae	Hirtella triandra	Sw.	0.066	0.032	0.034	0.018	0.607	0.645	0.020	0.008
Clusiaceae	Chrysochlamys eclipes	L. O. Williams	0.120	0.071	0.010	0.005	0.607	0.254	0.002	0.001	low P
Clusiaceae	Symphonia globulifera	L. f.	0.585	0.306	0.040	0.028	0.839	0.386	0.044	0.022	low P
Elaeocarpaceae	Sloanea terniflora	(Moc. & Sessé ex DC.) Standl.	0.731	0.295	0.262	0.111	0.314	0.138	0.018	0.010
Fabaceae (Mimosoideae)	Inga cocleensis	Pittier	0.913	0.583	0.010	0.009	0.603	0.271	0.162	0.106
Fabaceae (Papilionoideae)	Erythrina costaricensis	Micheli	0.501	0.329	0.032	0.012	0.144	0.062	0.092	0.028
Fabaceae (Papilionoideae)	Lonchocarpus heptaphyllus	(Poir.) DC.	0.963	0.765	0.218	0.089	0.016	0.011	0.048	0.023
Fabaceae (Papilionoideae)	Platypodium elegans	Vogel	0.372	0.168	0.372	0.176	0.022	0.019	0.827	0.534
Fabaceae (Papilionoideae)	Swartzia simplex var. continentalis	(Sw.) Spreng.	0.312	0.160	0.044	0.020	0.176	0.085	0.008	0.004
Malvaceae	Herrania purpurea	(Pittier) R.E. Schult.	0.372	0.156	0.202	0.105	0.482	0.165	0.014	0.006
Moraceae	Maquira guianensis	Aubl.	0.573	0.252	0.563	0.282	0.484	0.204	0.004	0.002
Moraceae	Poulsenia armata	(Miq.) Standl.	0.250	0.124	0.028	0.015	0.833	0.329	0.008	0.004
Moraceae	Sorocea affinis	Hemsl.	0.559	0.204	0.004	0.002	0.088	0.034	0.002	0.002
Moraceae	Trophis caucana	(Pittier) C. C. Berg	0.855	0.433	0.056	0.037	0.803	0.539	0.028	0.015	high P
Myristicaceae	Virola nobilis	A. C. Sm.	0.657	0.317	0.056	0.028	0.589	0.591	0.024	0.013
Myristicaceae	Virola sebifera	Aubl.	0.136	0.066	0.400	0.166	0.200	0.111	0.046	0.017
Myrtaceae	Chamguava schippii	(Standl.) Landrum	0.358	0.217	0.028	0.033	0.084	0.059	0.597	0.310
Rubiaceae	Pentagonia macrophylla	Benth.	0.146	0.094	0.060	0.028	0.517	0.184	0.002	0.001
Rubiaceae	Psychotria limonensis	K. Krause	0.567	0.094	0.266	0.056	0.741	0.459	0.046	0.008
Rutaceae	Zanthoxylum acuminatum	(Sw.) Sw.	0.156	0.089	0.004	0.001	0.521	0.333	0.104	0.032
Rutaceae	Zanthoxylum ekmanii	(Urb.) Alain	0.821	0.399	0.302	0.166	0.118	0.057	0.024	0.016
Salicaceae	Laetia thamnia	L.	0.917	0.462	0.068	0.034	0.006	0.004	0.188	0.079
Sapotaceae	Chrysophyllum argenteum	Jacq.	0.218	0.096	0.320	0.120	0.494	0.192	0.002	0.001
Simaroubaceae	Simarouba amara	Aubl.	0.983	0.457	0.238	0.083	0.022	0.013	0.430	0.144
Solanaceae	Cestrum schlechtendalii	G. Don.	0.450	0.203	0.008	0.004	0.404	0.140	0.020	0.011	high P

and Supplementary Materials File: Table S4). A total of 0 or 2 species were significantly associated with RP, 15 or 27 with MP, 6 or 21 species with NP, and 26 or 36 species with NHP at the 0.05 or 0.10 level, respectively. All species that tested as significant were associated with low levels of soil P. Most species were representatives of the Fabaceae family (n = 5 at the 0.05 level and n = 8 at the 0.10 level), while representatives of the Moraceae family (n = 4 at the 0.05 level and n = 5 at the 0.10 level) and the Burseraceae family (n = 4 at the 0.05 level and n = 5 at the 0.10 level) were also commonly significant. The remaining tree species were from a variety of different families.

Out of 33 (at the 0.05 level) and 52 (at the 0.10 level) significantly associated species with P affinity values [66], we found that 4 and 7 species, respectively, were low-P specialists (with P affinity < −0.8).

John et al. [17] warned that testing hundreds of species with different soil variables results in an inflated type I error. They quantified type I error by testing tree species distributions of a plot against soil nutrient maps from another site. They found that, on average, 11% of tree species were significantly associated ($\alpha$ = 0.05) with the swapped soil maps. Therefore, the percentage of tree species significantly associated with at least one P fraction at the 0.05 level decreases to 9% if type I error rate is applied.

4. Discussion

4.1. Quality of Spatial Models and PPMs

The reliability of the J-test result is presumably reduced due to the quality of the spatial models and the PPMs. For the spatial models of soil P, the presence of outlying values in the dataset biased the parameter estimates because the distribution of the response is assumed to be multivariate Gaussian. Skewness could not be eliminated by transformation, and we therefore stress that parameter estimates as well as cross-validation statistics may be influenced disproportionately by extremes [70]. Nonetheless, we preferred the use of restricted maximum likelihood fitting over the conventional method in the geostatistics of fitting a variogram model to robust experimental variogram estimates because the latter method includes arbitrary decisions about lag interval, bin width, and maximum lag distance, which can also strongly influence parameter estimation [71]. Given the cross-validation statistic MSDR, the models are “correct” in a sense that they reliably estimate the uncertainty in the predictions. Note, however, that the prediction uncertainty is relatively high, except for NHP. Thus, we have limited confidence in the predictions of the spatial models.

For PPMs, our quality assessment by the DCLF test and simulation envelopes verifies goodness-of-fit but not overfitting. The latter could be checked by validation techniques, which, to our knowledge, do not exist for cluster point processes [35]. In addition, the DCLF test is conservative [65], which results in a lower true significance level at which deviations from a good fit are detected. Thus, our selection of 182 goodness-of-fit-tested PPMs may include badly fitted and overfitted models.

4.2. Implications of Species–Phosphorus Associations for Phosphorus Resource Partitioning

Our results support our main hypothesis that the preferential depletion of different soil P compounds by plants due to P resource partitioning may cause spatial associations between tree species and low concentrations of P fractions. Given that species were significantly associated with mainly organic P fractions, we conclude that the soil organic matter is a relevant P resource for several tree species on BCI.

In a previous study [17], tree species in the 50 ha plot were tested for associations with a broad spectrum of soil nutrients, including P, which were extracted by Mehlich-III (M3) solution and quantified by inductively coupled plasma (ICP) spectrometry. About 22% of species were found to be significantly associated ($\alpha$ = 0.05) with low levels of M3-P [72], which is remarkably similar to the 20% of species in our study associated with citrate-extractable P pools. The M3-P fraction can contain a considerable proportion of organic P [73], which for Panamanian soils can account for >50% of total M3-P [74]. The findings of our study, where species were almost exclusively associated with organic P fractions, suggest that the significant associations of species with M3-P of the John et al. [17] study were mainly caused by tree species exploiting soil organic P.

The association of species with NHP was even stronger than associations with MP and NP, which suggests that species specialize in using compounds of the NHP fraction. The NHP fraction potentially contains organic P compounds that are amenable to phytase hydrolysis [30]. Phytases are known to be used by some plant species to acquire P from organic matter (e.g., [10]) and, consequently, this enzyme is included in models of organic P resource partitioning [8]. Phytase-hydrolyzable compounds might play a prominent role in the organic P pool: in a previous study of Panamanian forest soils, phytase-hydrolyzable P constituted on average 40% of NHP [30]. Furthermore, a recent study of topsoil samples from the 50 ha forest plot revealed that P extracted by a solution of 0.25 M NaOH and 50 mM ethylenediaminetetraacetate and determined by ³¹P nuclear magnetic resonance (NMR) spectroscopy contained considerable amounts of potentially phytase-hydrolyzable P, including myo- and scyllo-inositol hexakisphosphates [74]. We therefore interpret species associations with the NHP fraction as being linked to the acquisition of P from inositol phosphates.

4.3. Association of Soil Phosphorus Forms with Legumes

We found that several legumes had significant associations with organic P fractions, suggesting that tree species in the Fabaceae are among the primary users of soil organic P forms in the plot. Legumes have been previously reported to have relatively high root phosphatase activity [75,76], including in several Panamanian species [77], which they can up-regulate in response to soil P deficiency [76,78,79]. It has been hypothesized that the ability of legumes to acquire P from soil organic matter by phosphatase hydrolysis may explain the abundant occurrence of this family in lowland tropical forest communities [75], where tree species are assumed to be limited by P availability (e.g., [68]). As the synthesis of phosphatase enzymes requires nitrogen (N) [80], it has been suggested that legumes may benefit from their ability to fix dinitrogen as opposed to non-fixers [75]. This nutrient trading hypothesis proposes that legumes, by fixing N and investing it in phosphatase synthesis for P acquirement, “trade” N for poorly available P. However, evidence for this hypothesis remains scarce [76,77,81] and the greater phosphatase activity of legumes appears rather to be a phylogenetically conserved trait [76]. Whatever the mechanistic link, and notwithstanding the limitations of our study in terms of the risk of type I error, our results add to the growing body of evidence that legumes are among the most prominent users of soil organic P in lowland tropical forest communities.

4.4. The Influence of Phosphorus Availability on Phosphorus Resource Partitioning

The overall small number of tree species associated with P fractions might be explained by the relatively high availability of soil P in the 50 ha forest plot. The total P concentration of the dominant soil type in the 50 ha plot is high, sometimes exceeding 1000 mg P kg⁻¹ soil [82], which is approximately equivalent to the 90^th percentile of a global dataset of 802 soil samples [83]. However, despite the high total P, plant-available P concentrations can be relatively low due to the strong P fixation capacity of the soil, which contains high concentrations of clays and metal oxides. In Panama, available P is related strongly with species distributions [66] and growth rates [18], with strong P limitation where the concentration of resin P (soil phosphate extracted by anion exchange membranes) falls below 2 mg P kg⁻¹ soil [68]. Beneath this threshold, the activities of phosphomonoesterase and phosphodiesterase rise markedly, suggesting that organic P becomes increasingly important for P supply to soil organisms. In the BCI plot, the resin P concentration of 66% of topsoil samples (sampling protocol identical to our study) was >2 mg P kg⁻¹ soil [84], suggesting moderate (and not strong) P limitation.

Sufficient soil P availability might impede plant investment in phosphatase synthesis, which, in turn, may result in the observed low amount of species in the 50 ha forest plot that are associated with organic P fractions. In contrast, the plant demand of soil P from organic sources may increase where P availability is low [8]. Hence, in strongly P-limited environments, we would expect a relatively higher proportion of plant species to be associated with organic P fractions than reported here. Accordingly, we predicted that species in the 50 ha plot that are associated with organic P fractions would also show distributional affinities to low-P soils across a soil P gradient in Panama [66]. However, only a minor proportion of the associated species in the 50 ha forest plot were low-P specialists. Nonetheless, we recommend that future work on P resource partitioning should focus on plant communities in particularly low-P forest plots, such as the 52 ha plot at Lambir Hills National Park in Malaysia [85].

4.5. Limitations of the Methodology

We interpret the associations between tree species distribution and P fractions as evidence of the utilization of P forms by tree species. However, the association of tree species with nutrients does not imply a direction of influence or prove causation [17]. Several other mechanisms may explain the observed pattern of association. For instance, plant residues (litter and roots) are a source of soil organic P, so species–P associations might reflect interspecific variation in organic P forms. Indeed, foliar P concentrations vary substantially across species on BCI [72]. For instance, foliar P concentrations tend to be higher in legume tree species versus non-legumes [86]. However, if litter input drives the spatial distribution of soil organic P, we would expect some tree species (and legumes in particular) to be associated with high levels of organic P fractions, which was not the case in our study. Moreover, most organic P of tropical soils presumably originates from microbial residues because living microorganisms account for up to two-thirds of the total soil organic P [87], which can be released in response to seasonal changes, such as wetting and drying [88].

Furthermore, observed spatial patterns of tree species might be related to correlation of P forms with other soil nutrients. However, covariation of P fractions in the BCI forest dynamics plot was weak (Pearson correlation ranged from 0.16 to 0.38, Supplementary Materials File: Table S1) and, more importantly, John et al. [17] found that the correlation of M3-P with other soil nutrients in the BCI plot was low (Pearson correlation ranged from 0.045 to 0.350). Given that M3-P includes a considerable proportion of organic P (see above), we assume that the concentration of this fraction is similar to the total citric-acid-extractable P fraction of our study.

Our method assumes implicitly that the spatial distribution of soil P is stable over years to decades, as this distribution reflects the preferential depletion of organic P fractions by stationary individual trees. Although it has been shown that organic P pools vary seasonally in tropical forests with a strong dry season [87], to our knowledge, the spatiotemporal variation in organic P in tropical forests over a period of several decades has not been investigated. However, the Pearson correlation between M3-P of John et al. [17] and our citrate-extractable TP fraction is 0.67, indicating that the spatial pattern of organic P was relatively stable over a period of more than two decades.

5. Conclusions

The mechanisms driving species coexistence in hyper-diverse tropical forests remains a controversial topic in community ecology. Here, we provide evidence that P partitioning contributes to niche differentiation in these ecosystems, especially for species in the Fabaceae. Future tests of the hypothesis should focus on strongly P-limited sites, where P partitioning is predicted to be particularly widespread. For neotropical forests, the list of significantly related species that we provide here can serve as a starting point for a more focused selection of appropriate sites for observational studies or tree species for pot experiments.