Roots of cooperation: can root graft networks benefit trees under stress?


 The occurrence of natural root grafts, the functional union of roots of the same or different trees1–3, is common and shared across tree species2. However, their significance for forest ecology remains little understood. While early research suggested negative effects of root grafting (i.e. increases the risk of pathogen transmission)4,5, recent evidence supports the hypothesis that it is an adaptive strategy that reduces stress6–8 by facilitating resource exchange9,10. Here by analysing mangrove root graft networks, we show evidence of cooperation-associated benefits of root grafting. Grafted trees were found to dominate the upper canopy of the forest, and as the probability of grafting and the frequency of grafted groups increased with a higher environmental stress, the mean group size (number of trees within groups) decreased. While root networks could form randomly (i.e. trees do not actively ‘choose’ neighbours to graft to)11,12, the increased frequency and reduced group sizes in higher-stress environments point to the existence of underlying mechanisms that regulate ‘optimal size’ group selection related to resource use within cooperating groups8,13,14. This work calls for further studies to better understand tree interactions (i.e. network hydraulic redistribution)15 and their consequences for individual tree and forest stand resilience and water-use efficiency.


MAIN
Natural root grafts, the physical connection of two roots belonging to different trees or a single individual tree, have been known about for more than 100 years 3 and are recognised in almost 200 tree species 2, 16 . However, until the last decade, we had had little information about their ecological implications for tree interactions and forest stand dynamics. Between the 1950s and the 1990s, they were mainly regarded as a phenomenon of random occurrence or a threat to forest stands due to their role as vectors of pathogen transmission in forest stands 5 and the only long-term accepted consequence of grafting as a positive trait was increased mechanical stability 2, 16,17 . Now, the common perception of forest dynamics being ruled only by competition and survival of the fittest is being challenged by the discovery of mycorrhizal networks and the re-evaluation of root-grafts as critical vectors of positive interactions amongst trees [18][19][20][21][22][23] .
Root grafts can be either non-functional or functional (cambia and vascular tissues are fused 2 ). Non-functional grafts confer higher mechanical stability through shared anchoring systems 17 , which is particularly relevant for coastal wetland forests with shallow root systems 16,24 due to anoxic sediments, and exposure to strong winds 25 . Functional grafts can additionally facilitate resource exchange 21 and promote growth 7 by mitigating the adverse effects of defoliation and budworm outbreaks 10,26 and increasing the concentration of carbohydrates in shaded trees 21 . By enabling the exchange of water, carbon, and mineral nutrients, severed trees can be kept alive through the support of grafted neighbours 27 .
Moreover, modelling approaches suggest that natural root grafts could explain short-range positive interactions that lead to large-scale fractal patterns in tree yield 28 that seem to predestine natural root grafting as a cooperative trait. However, our current knowledge is mainly based on the study of grafted pairs of trees, and spatially explicit field investigations are limited to small plots in terrestrial forests 3,6,22,29 , while the functional ecology of root grafts in wetland forests, and the effect of environmental stress on network topology remains unexplored.
Until now, the study of root grafts has required the extensive excavation of root systems 5-7,29,30 and, often, decades to gather quality information 7 . The inclusion of environmental gradients to understand positive plant interactions and their ecological implications for community dynamics 31,32 is therefore very limited. However, mangrove forests, with their distinct environmental/elevational gradients 32 and traceable shallow root system with pneumatophores (e.g., the pencil-like emerging roots of Avicennia spp.), offer an ideal model system to study the ecological role of root graft networks. Avicennia germinans L. dominates forests on hypersaline mudflats with limited tree diversity, and strong salinity gradients offer particularly satisfactory conditions to study physiological responses 33 , tree architecture, and tree interactions 25,34,35 . These mangrove specificities provide ideal conditions to investigate in the field whether root grafting can benefit trees growing under environmental stress through an analysis of individual tree attributes and spatial root graft network structures.

Drivers of root grafting and implications for tree size
To understand the main drivers and consequences of natural root grafting in an A. germinans dominated forest, we focused on a seasonally hypersaline mangrove forest bordering the coast of the Gulf of Mexico (Fig. 1A). A steel rod root detection method developed to measure root length with minimal excavation 36 was modified to identify and map root graft networks in eight 900 m 2 forest stands (Fig.1B). We further related root graft frequency to biotic and abiotic variables, such as stem diameter, stand density and porewater salinity (see methods section). We also explored the height-diameter relationship of grafted and non-grafted trees with different neighbourhood asymmetries (i.e., competition pressure by neighbours; see Supplementary Information) and the relationships between the related network group attributes of group size and frequency to stand density and salinity.
Top-height trees, defined as the 20% biggest trees in a stand (as per stem diameter, see methods section), are considered to have exploited resources to their maximum ability, and thus reflect the potential productive capacity of a stand 37 . The high frequency of grafting in the most dominant trees suggests that a shared root system provides essential advantages to the forest, either by optimizing resource exploitation or by increasing mechanical stability and windthrow resistance.
In line with previous studies 6, 38,39 , the probability of grafting increased with increasing tree stem diameter (p < 0.0001; Fig. 2A). With the addition of salt stress, however, the contribution of stem diameter to grafting probability decreased for stem diameters >20 cm ( Fig. 2A; Extended Data Table 1). In upland forests, higher stand densities contribute to increased grafting probability because of reduced distances between neighbouring trees 38 . In our study, however, stand density reduced grafting probabilities. This could be due to higher resource limitations within saline environments. First, closer neighbours result in greater competition 37 , while high salt stress additionally reduces resource availability and limits growth rates 33,40 , leading to smaller stem diameters within the stand. Independently, stand density and salinity lead to smaller stem diameters, thus reducing the probability of grafting ( Fig. 2A). However, trees with smaller stem diameters have a higher probability of grafting at high salinity and stand density sites (Extended Data Table 1). The highest proportion of grafting was recorded for plots with the highest stand densities and salinities (Extended Data Fig. 1), suggesting that salt stress has direct control over root grafting.
The generalised additive mixed model that was used to explore the effect of grafting on tree size, demonstrated that grafted trees are generally taller than non-grafted trees (p < 0.01, Fig.   2B). The model (R 2 = 0.78, explaining 85% of de deviance) showed that neighbourhood asymmetry did not influence tree height of grafted trees (p = 0.46 and p = 0.25, for grafted and non-grafted trees, respectively) (Extended Data Fig. 3A grafts provide a benefit to trees, which could be due to either a potential increased growth rate 6,21 if grafts are fully functional, or due to increased mechanical stability related to an extended area for anchorage 17 . The height to stem ratio (so called slenderness) is an allometric trait 41 that determines mechanical stability. Very slender trees are more vulnerable to windthrow, while low slenderness coefficients increase wind resistance 42 . Slenderness varies throughout tree development; younger trees invest in height growth before girth as a result of competition for light, which increases their risk of mechanical failure 42 . As they reach the canopy, more resources are invested in stem girth, conferring higher mechanical stability and resistance to windthrow [41][42][43] . Further assessing changes in slenderness along the range of stem diameters ( Fig. 2C), we found a significant increase in slenderness for non-grafted trees that are subject to higher competition pressure (p=0.02, Extended Data Table 2). However, we also detected a significant, but weak negative interaction between stem diameter, grafting condition and neighbourhood asymmetry (p = 0.01, Extended Data Table 2). As non-grafted trees increase in diameter, their slenderness decreases more rapidly than it does for grafted trees (Fig. 2C).
This points towards potentially increased mechanical stability for dominant grafted trees with greater wind exposure than their neighbours 42 . This finding is consistent with the higher slenderness reported for grafted hybrid poplar clones 22 .

Network formation
Trees can benefit from functional root grafts through the increase of foraging area via communal root systems 1,2 .Root networks could also mitigate salinity-induced physiological drought through water redistribution between stems 26,27 . Further, shared carbohydrate pools could improve tree responses to both abiotic and biotic stress 19,20 . These factors likely contribute to the dominance of the forest canopy by grafted individuals. We challenged this hypothesis through the analyses of network topologies along the salt-stress gradient, where the lack of resource exchange would result in random network formation patterns, and similar network topologies along the stress gradient.
In forests, the location of individual trees is fixed after their establishment, and network formation is determined by physical, genetic and size proximity 38 , limiting any preferential attachment processes. Although the root networks in our study might fit a scale-free powerlaw distribution (p = 0.21; Fig. 3A), that is, they might possess patterns of continuous growth and preferential attachment 11 , we found no significant power of determination to reject random network formation when comparing the power-law distribution to log-normal (p = 0.99), Poisson (p = 0.93) or exponential distributions (p = 0.99; Fig. 3A). In this context, grafting could be a random process 2,6 ; however, the distribution of the node degree (number of trees connected to a given tree via root grafts), the group frequency distribution and group size along the stand-density gradient (Fig. 3B, C) point to underlying mechanisms that select for optimal group size in cooperative groups 11 .
Most of the grafted trees were connected to one (61%) or two (29%) individuals, while connections to four partners were rare (1 %). Hence, the node degree's relative frequency is smaller with increasing node degree (Fig. 3A). Additionally, the average node degree is negatively correlated to stand density (R² = 0.40; p = 0.05; Fig. 3B) and the frequency of grafted trees (R² = 0.93; p < 0.001; Fig. 3B). Consequently, the average node degree was highest in the plot with the lowest stand density and highest grafting frequency (Fig. 3B).
Likewise, as A. germinans stand density increased, the average number of trees forming groups became smaller (R² = 0.62; p = 0.01; Fig. 3C), whereas the frequency of groups increased (R² = 0.73; p < 0.01; Fig. 3C). This is in line with network theory findings of cooperative interactions increasing with environmental stress 14, 44 . Such interactions, however, do not come without costs. It costs each cooperating individual to provide a benefit to its neighbours, and to be selected as an adaptive trait within a population, the net gain of the cooperative trait should be greater than its cost 44 .
There is evidence that in unweighted networks, selection favours cooperation when the benefit-cost ratio " # $ % exceeds the average number of neighbours ( ) (i.e., node degree): " # $ % > . Thus, most cooperative groups tend to have few members 14,44,45 and higher probabilities of direct reciprocity (i.e., pairwise ties) 45 . The average node degree in the mangrove root networks we studied was smaller at sites with high salinity and high stand density compared with low salinity and medium stand density (Fig. 3B). However, stand densities were similar in low-and high-stress environments. As the cost of cooperation increases under stressful conditions 14 , assuming that a tree receives constant benefit from its cooperating neighbours, the critical " # $ % ratio decreases with increasing stress. Thus, larger tree groups are not selected under situations of limited resource availability. Most of the grafted mangrove groups (73%) consisted of only two or three members. However, of the groups that included more than two trees, 72% had no more than the minimum required number of connections for a cooperative system (each individual had at least one connection for cooperation within its group members). This supports the hypotheses that functional root grafts can only be maintained if there is a long-term payoff for all group members and underlying mechanisms selected for optimal group size in root networks.

Study site
The study site is located on the central coast of the Gulf of Mexico (GoM) at the La Mancha lagoon (Fig. 1A)  For the pre-established plots, existing tree parameters were recovered from a publicly available database 52 , including a unique ID, species, x-and y-axis positions in the plot, stem diameter at 130 cm from the soil surface ( +,-) and height ( ). For the newly established plot, the same tree parameters were measured using a laser rangefinder (Laser Rangefinder Forestry Pro 550; Nikon Vision Co., Ltd, Tokyo, Japan), and tree positions were determined using a compass and the rangefinder following standard forestry procedures 53 .
For each plot, during April and September 2017, two pseudo-replicate pore water samples were collected from each corner and the middle of the plots from 20 cm below the ground surface using a custom-made pore water extractor 54 and immediately analysed for pH, salinity, temperature and redox potential (Ultrameter II; Myron L Company) 55 .

Root graft data collection
A non-destructive method was used to detect the potential location of root grafts using a portable Doppler ultrasound probe (DU; SonoTrax Basic; Edan Instruments GmbH, Hessen, Germany) and a set of steel rods. The mangrove roots were gently located with steel rods with the DU probe placed on the tree stem. Following an adapted method originally developed to measure the woody root extensions of A. germinans 37 , the probe was then gradually moved from the stem to the consecutive rods in contact with the target root. Each tree was examined following the consecutive order of the tree tag numbers within the plots by assessing their grafting to all immediate neighbours.
Placing the DU on a tree stem collar ring, a steel rod was used to probe the soil to shallow depths, and an amplitude monitor indicated when a root belonging to the stem was touched.
Leaving this first steel rod in contact with the root, a second rod was used to further probe close to the first rod in the assumed direction of the course of the root until another positive signal was obtained. The interchangeable waterproof probe of the DU was then attached to the second steel rod, having been proofed to be in contact with the initial root, and the process repeated until either the root was too deep or too thin to be followed or led to another tree. In the latter case, the probe was held on the second tree stem and the last verified steel rod was used to again probe until another positive signal was returned by the DU from the second stem. The DU-located root graft was then verified by localised excavation of each target tree's neighbour. Although we were unable to verify false negatives, we calculated a 6% probability of finding false-positive connections (i.e. 12 false positives out of 200 connections detected), all identified false-positives were treated as non-grafted trees. We did not have any means to evaluate false-negative rates.
Using this method, a total of 376 A. germinans tree-root systems were followed during April and May 2017. These were mapped and used to determine the grafted network topology: node degree (number of direct connections for each tree), number of groups of grafted trees and mean group size (number of individuals within a group).
To estimate the pressure each tree receives from its neighbours, an index of neighbourhood asymmetry was calculated as a function of the size and distance of all neighbouring trees ( 3 ) within a 5 m radius of the target tree ( 4 ; see Supplementary Information). A large index of neighbourhood asymmetry implies that the neighbours are large and in close proximity, potentially exerting higher competition pressure on a target tree than a small neighbourhood asymmetry would. The 5 m radius was chosen because it had been previously identified as the optimal radius for detecting the responses of trees to its neighbours at the same study site 25 . Neighbourhood asymmetry was only calculated for trees where their complete neighbourhood was within the limits of the sampling plots (187 trees) to avoid biased neighbourhood asymmetry sizes related to incomplete information for neighbouring trees located outside a plot.

Data analysis
Both the density of the target species A. germinans and the total stand density (including selected as per stem diameter because it was measured in the field and is considered more accurate than tree height, which is estimated through stem diameter measurements 53 . Logistic regression was implemented using a generalised mixed effects model to assess the probability of grafting as a function of stem diameter, total stand density and salinity. The model included site identity as a random effect and stem diameter, site salinity and total stand density as fixed effects after assessing the autocorrelation between response variables (Extended Data Fig. 2) and all intra-and cross-level interactions between stand density and salinity. All the variables were z-transformed using the mean and standard deviation of each variable across all sites.
To explore the effect of root grafting and neighbourhood pressure on tree allometry, in the generalised additive mixed effects model (GAMM), salinity and condition were included as fixed effects (cyclic cubic regression spline), neighbourhood asymmetry and stem diameters were included as smooth terms with smooth functions (Duchon spline) and the sampling plot was included as a random effect. The best model explaining tree height was selected using a minimal Akaike information criterion value following a stepwise removal of non-significant response variables (N = 138 single-stem A. germinans trees with a computed asymmetric neighbourhood).
In the existing database of tree parameters 52 , multiple-stem trees are recorded following the traditional convention of summing the diameters of each stem but by measuring only the height of the tallest stem 35  Network parameters (node degree, number of groups per hectare and group size) were used to assess random network formation by comparing the probability of networks having a scalefree power-law distribution with random process distributions (i.e. log-normal, exponential and Poisson). Scale-free networks do not occur randomly because a relative change in one node results in a proportional, relative change in another node. Scale-free power-law distributions indicate the continuous expansion of networks and preferential attachment, where new nodes are constantly added and previously well-connected nodes are more likely to acquire new connections 11, 56 . We then related network node degree, group size and frequency (number of trees grafted within groups and frequency of groups per hectare, respectively) to stand density and site salinity using linear regressions. All 376 A. germinans trees, including multi-stemmed trees, were used for this analysis.
All the statistical analyses were conducted using R programming language 57 . Specifically, we used the lme4 58 , DHARMa 59 and gamm4 60 packages for the logistic regression and the GAMM construction and diagnosis. For the network analyses, we used igraph 61 to estimate the node degrees and PoweRlaw 62 to explore the distribution. All figures presented were developed using ggplot2 63 .

Data availability
Data on allometric attributes of the trees (i.e. height, stem diameter and position in stand) are publicly available at https://doi. org/10.5525/gla.researchdata.657. Root Network formation and sediment salinity databases will be made publicly available from the research data repository of the University of Glasgow upon acceptance for publication. All other data supporting findings, including coding resources, are available from the corresponding author upon request.