To investigate the effects of patch number, patch-size heterogeneity, and dispersal, we manipulated dispersal of protists within microcosms designed to mimic different habitat patch configurations (Figs. 1, S1). The experimental design consisted of five habitat configurations crossed with nine dispersal regimes, resulting in 45 treatment combinations which were replicated four times to total 180 microcosms. The five habitat configurations consisted of one patch, or four or six patches of either homogeneous or heterogeneous patch size (SL, 4Ho, 4He, 6Ho, 6He; Fig. 1). The dispersal regimes consisted of two dispersal types (matrix, M, or local, L) which each had two ‘quality’ levels (low, L, and high, H, mortality risk or low, L, and high, H, frequency, respectively). Due to the time required for sampling, we blocked the replicates by time so that each was sampled on a different day of the week from Monday to Thursday. The experimental landscapes were kept in incubators at a constant temperature of 20 °C for the duration of the 21-day experiment. This equates to approximately 21 generations for the protists as their generation time is around one day (Clements et al. 2013). The patterns in their final abundances were therefore driven by reproduction, extinction, and colonisation within the landscapes.
The protist community consisted of eight protist species, including three apex predators (Didinium nasutum, Dileptus anser, and Stentor coeruleus) and five potential prey species (Fig. S2) (all obtained from Sciento, Manchester, UK). Didinium and Dileptus were considered to be specialists because they fed on two prey species each (Worsfold et al. 2009), whilst Stentor was considered as a generalist because it fed on four of the five available prey species (Jiang and Morin 2005; Cadotte et al. 2006; Fig. S2). To comprehensively investigate the effects of the respective predator groups, an ideal experiment would also include treatments with each predator group alone. However, due to time constraints we were unable to include any further treatments in our experimental design.
The experimental landscapes were custom-designed using FreeCAD 3D-design software (Riegel et al. 2020) then 3D-printed using a LulzBot TAZ 6 printer in black PLA filament. To ensure the landscapes were watertight they were then coated in clear epoxy resin. Experimental landscapes consisted of either a single, four, or six circular wells arranged in a ring (Fig. 1). Additionally, wells in landscapes with more than one patch (i.e., four or six patches) were either all the same size (homogeneous) or two different sizes (heterogeneous). In the heterogeneous landscapes we maximised spatial heterogeneity by alternating between small and large patches in the landscape so that each patch was adjacent to patches of a different size. The five reserve configurations (Fig. 1) each had a total volume of 48 ml and were made up of wells with a constant depth of 13 mm to standardise light penetration. Controlling for total volume enabled us to disentangle the effects of heterogeneity and patch number from the effects of landscape area by comparing landscapes with different patch-size distributions. Additionally, we ensured that the average patch size remained constant between homogeneous and heterogeneous landscapes with the same number of patches.
Each landscape was filled with experimental media containing bacteria for the bacterivorous protozoa to feed on, and a carbon source for the bacteria to consume. Experimental media was created by dissolving crushed protozoa pellets (Blades Biological LTD, UK) in spring water (Tesco Ashbeck English natural mineral water) at a concentration of 0.5 g/L and then autoclaving this mixture. On day 5 each bottle of media was inoculated with three bacteria species (Bacillus subtilis, Micrococcus luteus, and Pseudomonas fluorescens). This provided sufficient time for the bacteria to increase to levels which would sustain the protist communities which were added at the start of the experiment on day 0.
We created the experimental community by mixing the eight species in a sterilised jar at densities determined by preliminary experiments (apex predators 1 ind./ml, all prey apart from Colpidium striatum 10 ind./ml, Colpidium striatum 100 ind./ml). This mixture plus the experimental media were then distributed to fill every patch in each landscape so that each landscape contained a total of 48 ml protist-containing media.
Dispersal between wells was conducted manually rather than via printed corridors as this allowed us to investigate the effects of ‘quality’ of dispersal events by manipulating both dispersal frequency and risk of mortality during dispersal. The two types of dispersal each had two treatment levels plus a no-dispersal control which was applied to all five landscape types. “Local dispersal” occurred between adjacent habitat patches and acted as a proxy for dispersal through corridors connecting neighbouring patches. Local dispersal was either low or high frequency, representing the fact that corridor quality can influence dispersal frequency (Haddad 1999). “Matrix dispersal” occurred from each patch to any patch in the landscape, acting as a proxy for dispersal across the non-habitat matrix to any patch in the landscape. Matrix dispersal was either low or high mortality risk, representing the role an inhospitable matrix can play in causing mortality to dispersing organisms (Nowicki et al. 2014). We chose to manipulate mortality risk of matrix dispersal, but not local dispersal because, although dispersal through corridors is not risk-free, corridors are often created to provide a relatively safer route for passing wildlife, for example when over- and under-passes are added to roads (Simpson et al. 2016). To investigate if there was an interactive effect between the two dispersal manipulations, the two local dispersal manipulations plus control, and two matrix dispersal manipulations plus control were factorially crossed, leading to nine possible dispersal regimes. We chose to factorially cross the dispersal manipulations as previous work has revealed interactive effects between dispersal through corridors and dispersal across the matrix (Åström and Pärt 2013) and we wished to further disentangle this. The treatment levels of the dispersal manipulations were arbitrary as they were designed to demonstrate how general variations in dispersal regime affect diversity outcomes, rather than representing dispersal frequencies or risk of mortality during dispersal for any one taxon.
We conducted dispersal manipulations following sampling and nutrient replenishment. Dispersal occurred weekly for all regimes apart from high frequency local dispersal treatments. Matrix dispersal was conducted by gently mixing a patch through pipetting then removing 0.6 ml from each patch onto a sterile Petri dish. This mixture was then gently mixed and either 0.48 ml (equivalent to 20% mortality, the “low mortality” treatment) or 0.12 ml (equivalent to 80% mortality, the “high mortality” treatment) was pipetted back into every patch (Fig. 1). The remaining solution was discarded, representing the mortality risk associated with dispersing across the matrix. Every patch was topped up with 0.12 ml (low mortality) or 0.48 ml (high mortality) of sterile nutrient medium to account for the volume discarded. This dispersal mode meant that a protist may move into any patch including returning to their initial patch (Fig. 1).
We conducted local dispersal by gently mixing and then pipetting 1.2 ml from every patch onto a sterile petri dish labelled with the initial patch location. 0.6 ml from the 1.2 ml droplet was pipetted into both patches adjacent to the original, transferring protists to their neighbouring patches (Fig. 1). This occurred weekly for the low frequency local dispersal, and twice weekly for high frequency local dispersal, on the sampling day and then again three days later. When a dispersal regime consisted of both local and matrix dispersal, all the media to be dispersed was removed from the landscape prior to pipetting any back, ensuring that no protists were dispersed more than once. This approach to dispersal meant that the dispersal rate was slightly lower for matrix dispersal than local dispersal, as a small proportion of the 0.6 mL taken from a patch in the matrix dispersal treatment was returned to the original patch.
We conducted sampling once a week for each treatment block. Sampling involved searching a well with a microscope for a maximum of five minutes and recording presence or absence of each species. A species was marked as present when at least one individual of the species was spotted. On day 21, the final sampling day, we calculated the abundance of each species by counting individuals in a 0.5 ml subsample taken from each well and multiplying this value according to the well’s total volume. We pipetted up and down within a well to ensure a roughly homogenous distribution of individuals, thus minimising potential bias caused by aggregation of individuals, then pipetted the 0.5 ml subsample onto a sterile petri dish for counting (as in Clements et al. 2013). If a species which was not known to be extinct appeared absent from the subsample, we then checked the whole well and counted all individuals of that species within the well. Although 0.5 ml is a small volume relative to the total volume of the larger wells, this was the maximum amount we could sample in the available time, in particular due to high densities of certain species (>100 individuals in a single subsample).
Following weekly sampling, we conducted nutrient replenishment by mixing each well, removing 10% of each well’s total volume and replacing this with fresh sterile nutrient medium to fill the well to its original volume. This prevented build-up of waste materials and compensated for any evaporation which may have occurred.
The following parameters were used as descriptive variables in our analyses: number of patches (one, four, or six), patch-size heterogeneity (homogeneous or heterogeneous), matrix dispersal (none, low mortality, or high mortality), and local dispersal (none, low frequency, or high frequency). Using data from presence / absence sampling, we calculated the overall number of extinctions in a landscape and probability of specialist (Didinium nasutum and Dileptus anser) and generalist (Stentor coeruleus) apex predator presence. Using the final population count data, we calculated Shannon diversity of each landscape to measure γ diversity. This was calculated using the ‘entropart’ package in R (Marcon and Hérault 2015).
To investigate the relationship between the predictor variables and the response variables, we conducted analyses using Generalised Linear Models (GLMs). The γ diversity GLM had a Gaussian error distribution, specialist and generalist predator presence probabilities had a binomial error distribution, and extinctions had a Quasi-Poisson error distribution due to underdispersion. The saturated model included all descriptive variables and their two-way interactions. To select the best model for each response variable we used Akaike’s Information Criterion corrected for small sample size (AICc, Burnham and Anderson 2002) which ranks models with different parameter combinations by ∆AICc, selecting the ‘best’ model with the lowest AICc. Model simplification was conducted using the ‘dredge’ function from the MuMIn package in R (Bartón 2020) which ranks models according to their AICc values. All models within ∆AICc ≤ 2 of the top model were considered equivalent in their descriptive ability. These top models were subsequently averaged using the ‘model.avg’ function in MuMIn, producing coefficients which were extracted and used for plotting and parameter reporting. In addition to reporting P values and effect sizes, we used hierarchical partitioning (Chevan and Sutherland 1991) to estimate the proportion of total variation in each response variable explained by each descriptive variable, thus demonstrating their importance. This analysis was conducted using the ‘hier.part’ package in R (Walsh and MacNally 2020). Furthermore, to investigate the effect of probability of generalist predator presence on probability of specialist predator presence at the patch level we performed a logistic regression (GLM with binomial error distribution). All analyses were conducted in R (version 3.6.3, R Development Core Team 2020).