Ethics statement

All research activities were conducted in accordance with James Cook University Animal Ethic Guidelines (permit approval numbers: A2775 to RAM and A2454 to PSW).

Study design and survey datasets

To assess the trophic structure of reef fish assemblages and the role of planktivore-mediated subsidies, we explored 2 datasets of underwater visual surveys. These consisted of an ocean basin scale dataset spanning most of the Indo-Pacific Ocean and a regional scale dataset focused on 3 key locations. For the ocean basin scale, we used the open access database of reef fish counts from the Reef Life Survey (RLS; [72]), which includes data on abundance and species identity. The complete description of the visual survey method used is available in Edgar and Stuart-Smith [41], but, in brief, it consists in a pair of divers simultaneously conducting fish counts on parallel “blocks.” Each belt transect is the sum of the 2 blocks, each comprising an area of 50 m × 5 m, in a total area of 500 m² per transect. The complete RLS dataset was filtered online to include only tropical locations (between the latitudes of 30° S and 30° N) in the Indo-Pacific (approximately between the longitudes of 35° E and 130° W). We then excluded nonquantitative records (i.e., variable “Block” = “0”), as well as surveys shallower than 4 m or deeper than 15 m depth. Since reef zone information is not available from the RLS dataset, this step was required to minimise the chance of surveys located in back reef zones, downstream from currents and naturally scarce in plankton and planktivorous fishes (e.g., [11,73]). From the resulting set of surveys, we only retained ecoregions (sensu Spalding and colleagues [74]) that encompassed a minimum of 4 transects. Counts were then averaged, rather than summed, between the 2 blocks to obtain a single count relative to an area of 250 m² per transect. We also averaged across transects at each site for each ecoregion. The final dataset comprised 32 ecoregions and 1,035 unique sites (Fig 2A, S1 Data).

At the region scale, we surveyed 3 key locations in the Indo-Pacific: Raja Ampat and the Banda Sea (hereafter “Raja Ampat”), in eastern Indonesia; Lizard Island, in the northern Great Barrier Reef, Australia; and Ha’apai, in the central part of Tonga (S1 Fig). Surveys in Raja Ampat and Lizard Island were conducted following a 4-phase design, with each phase targeting specific sizes, mobility categories, and common behavioural responses of fishes in progressively smaller areas. The complete method is described in [11] and involves identifying, counting, and estimating the body size (total length, in cm) of (1) large and mobile, conspicuous fishes within an area of 250 m² (50 m × 5 m); (2) smaller but also mobile, conspicuous fishes in an area of 150 m² (30 m × 5 m); (3) small, site-attached fishes within an area of 30 m² (30 m × 1 m); and, finally, (4) small cryptobenthic or larger fishes hidden in holes, also within an area of 30 m² (30 m × 1 m). All surveys were performed in slope or crest habitats, between 2 and 15 m depth, and by the same observer (RAM). In both Raja Ampat and Lizard Island, all surveys took place on fringing reefs around high islands, or on the exposed facet of lagoon-forming reefs, and thus encompassed exclusively upstream reef zones. Fish sizes were estimated with 2 cm precision for fishes under 30 cm TL and with 5 cm precision for all other fishes. All surveys were done between January and February and between October and December 2018. In Ha’apai, surveys were performed in the same types of reefs and reef habitats as Raja Ampat and Lizard Island and followed similar procedures, except that they encompassed only 2 phases. Phase 1 in Ha’apai targeted the same types of fishes as phases 1 and 2 in Raja Ampat and Lizard Island (in an area of 150 m², 30 m × 5 m), while phase 2 targeted the same types of fishes as phases 3 and 4 in Raja Ampat and Lizard Island (in an area of 60 m², 30 m × 2 m). All surveys in Ha’apai were performed by the same observer (PSW) between September 2017 and November 2018. A complete description of the procedures can be found in [75]. Overall, the regional dataset comprised 325 surveys in 3 locations.

We used a resampling procedure to scale counts from the different types of surveys and different count phases in each type of survey to the same common area (as in [11,43]). First, for each of the 4 (Raja Ampat and Lizard Island) or 2 (Ha’apai) phases of each survey, fish abundance was proportionally rescaled from the area surveyed (e.g., 250, 150, 60, or 30 m²) to a common (“resampling”) area of 100 m². Then, at each resampling iteration, a number of fish individuals equal to the rescaled abundance were randomly resampled for each phase of each survey. Thus, at each iteration, each final resampled survey (i.e., 100 m²) consisted in the final, cumulative set of fish individuals resampled from all phases of the original survey. We refer to these resampled surveys throughout the manuscript as “survey areas,” which comprise a smaller spatial scale compared to RLS sites. A total of 1,000 bootstrapping iterations were performed and all indicators with the exception of abundance (species richness, biomass, productivity, etc.) calculated as the mean across iterations. Total abundance was the criterion guiding the resampling principle, and, therefore, surveys had the same total abundance (but different species-specific abundance) across all iterations.

Defining trophic groups and plankton subsidies

Species from both datasets were classified in trophic groups following the species lists provided in [11,44]. Five broad groups were used: herbivores (including macroalgivores, microphages, grazers, and detritivores, HD), omnivores (OM), invertivores (including mobile and sessile invertebrate predators, IN), zooplanktivores (PK), and generalised macrocarnivores (including specialised piscivores and large-bodied carnivores that often consumer fishes, cephalopods, and large crustaceans, GC [44]). Here, we consider plankton subsidies to be equivalent to pelagic subsidies, and we measure plankton subsidies as the net biomass production of planktivorous reef fishes. This definition of plankton/pelagic subsidies is narrower than the one used previously (e.g., [11]) but was required because of the vast number of species with uncertain or unavailable quantitative dietary information, as well as the large spatial scale of our study. This narrow definition excludes indirect trophic links that may also result in pelagic subsidies being incorporated by reef communities (e.g., preying on planktivorous fishes or heterotrophic corals, coprophagy of planktivore waste, feeding on plankton-enriched detritus [40,50,54,55]). Conversely, this definition may not completely exclude reef-based material consumed by planktivores (e.g., resuspended algae material, fish, and crustacean eggs, emerging zooplankton [39,73,76]). While some of these links may be expected to decrease, others are expected to boost the perceived importance of pelagic material for reef fish assemblages.

There is strong evidence, however, that our estimates of plankton/pelagic subsidies are conservative. First, many of the “planktivorous” fishes that ingest substantial amounts of algae matter (e.g., Abudefduf, species of Pomacentrus, Melichthys niger, Melichthys vidua) were herein classified as “omnivores,” rather than planktivores. These, therefore, do not contribute towards our results. Some of these fishes comprised a significant proportion of the productivity, as exemplified by 3 sites in Raja Ampat with 24.7%, 35.1%, and 43.3% of the productivity comprised exclusively by Melichthys spp. If included, the overall contribution of plankton to fish productivity in these cases would almost certainly be significantly higher. Second, other sources of reef-based matter, such as fish or crustacean eggs, often represent a minor to small proportion of the diet of diurnal planktivorous reef fishes (0% to 20% in most cases) [39,76–80]. Third, even nocturnal planktivores (e.g., holocentrids of the genus Myripristis, some apogonids), which feed to a higher degree on reef-based eggs and larvae and, particularly on resident zooplankton (primarily mysid shrimps), also feed extensively on pelagic copepods and other pelagic zooplankton [76]. Finally, some of the primary prey of nocturnal planktivores are thought to feed partially on pelagic-derived phytoplankton themselves (see p. 95 of [81]), which would make their predators secondarily reliant on pelagic sources, rather than reef sources. Therefore, it is likely that we provide an underestimate of pelagic subsidies—the real extent and magnitude of their energetic contribution to fish communities will probably be substantially higher than detected here.

From standing biomass to fish productivity: Combining growth and mortality trajectories

We obtained length–weight conversion parameters (to estimate standing biomass) and growth and mortality coefficients (to estimate productivity) for all species in the regional dataset following the procedures in [82,83]. Briefly, species-specific Bayesian length–weight regression parameters were obtained from FishBase [84,85] and used to estimate individual weight (mass) for all fishes detected in the surveys. Then, mean sea surface temperature (calculated from remote sensing data, provided in the Bio-Oracle database [113]), species maximum body length (L_max), dietary and habitat use information were used to predict a standardised Von Bertalanffy Growth Model (VBGM) parameter (K_max) for each species at each location [83]. Both L_max and K_max were then used, alongside sea surface temperature, in Pauly’s empirical regression to predict the instantaneous exponential natural mortality coefficient Z for each species, again, at each location [86]. This annual, population-level, mortality rate coefficient was then converted to a daily rate and combined with a risk function that describes the ontogenetic decline in mortality risk as individuals age and grow [82]. This resulted in an individual-level daily probability of mortality.

With growth trajectories and mortality probabilities defined for each individual fish at each location, productivity was then estimated under the individual age framework [82]. Expected somatic growth was quantified by (1) estimating the operational age of each individual fish at the time of the survey [82,87]; (2) positioning the fish in their predicted VBGM trajectory according to their operational age and obtaining their size (mass) after a 1-day growth period; and (3) subtracting their mass at the survey from their forecasted mass after growth [82]. Finally, each individual was assigned a fate at the next day after the survey, mortality, or survival, simulated randomly according to their probability of mortality. Productivity was, finally, the expected growth of all surviving individuals at each bootstrapped simulation. Given the short time interval considered (i.e., 1 day), the terms “productivity” (i.e., rate) and “production” (i.e., the expression of this rate throughout a finite time interval) are used interchangeably throughout the text. The reasoning behind and the equations underlying each step described above can be obtained in detail from [82], which also provides the R software [88] package rfishprod, used for all the above calculations.

Statistical modelling procedures

Except where otherwise stated, all modelling procedures in this study were implemented following the Bayesian framework and using the No-U-Turn Sampler (NUTS) with the Stan machinery, via the rstan interface to R and the rstanarm package [89,90]. Auxiliary priors (representing the shape parameter of gamma distributions and the phi parameter of beta distributions) were always defined following a Cauchy distribution with location = 0 and scale = 5. All other priors are defined on a case-by-case basis below. To propagate the variability and uncertainty in response variables arising from the resampling distributions described above, we ran 1 Bayesian model per resampling iteration in a total of 1,000 models per response variable/modelling procedure (below, unless otherwise stated). We used the compounded posterior distributions of the modelled parameters from these 1,000 bootstrapped models to estimate coefficients and 95% HDP intervals. By propagating variability and uncertainty in the response variables, bootstrapping our models made our approach conservative when compared to an alternative strategy, i.e., using resampling medians to run a single model per response variable. For all Bayesian bootstrapped models, we used 3 MCMC chains per model and 1,500 NUTS steps per chain, with 50% burn-in.

Following the equivariance property of MCMC sampling, quantities derived from parameters estimated via MCMC are treated as posterior distributions of parameters themselves [91]. Thus, we derive our inferences from ratios between coefficients, i.e., between intercepts or slopes of locations. All comparisons and the acronyms used to represent these ratios in figures are specified below.

To ensure the convergence of 1,000 Bayesian models per modelling procedure, we implemented 2 automated checks. These checks evaluated, for each model, whether (1) the effective sample size of all parameters was equal to at least 100 times the number of chains (as recommended in [92], in our case, 300 steps); and (2) all Gelman–Rubin R_hat statistic values were below 1.05. Only models that passed these checks by fulfilling these 2 criteria were retained in the compounded posterior distributions. Below we also list, for each model procedure, the number of models retained.

The trophic structure of Indo-Pacific reef fish assemblages and the productivity–biomass relationships of 3 key locations

We started by exploring the ocean-scale (RLS) dataset, evaluating the potential correlation between the relative abundance of each of the 5 trophic groups of reef fishes (HD, OM, IN, PK, and GC; see above) and total fish abundance. To estimate the relationship between the relative abundance of trophic groups, we fit a local polynomial regression smoother (LOESS). We used a near-unitary α parameter (>0.9) to minimise sensitivity to outliers. Noticing that planktivores were the only group positively associated with high fish abundance and that both high abundance and high planktivore proportions seemed to bear some level of spatial clustering at or near the “core” IAA ([93,94]), we then investigated the 3 selected locations in our regional scale dataset. These 3 locations spanned a steep gradient in biodiversity, from the core IAA (Raja Ampat, Indonesia), the centre of marine biodiversity in the world [42], through the periphery of the IAA (Lizard Island, Great Barrier Reef), to the central Pacific (Ha’apai, Tonga), over 6,000 km away from the core IAA.

We first examined abundance distributions for all species detected (i.e., not only planktivores) in the 3 locations. This unveiled striking distinctions in the relative status of abundant species between the locations. These patterns were then further explored by contrasting total fish abundance, α, β, and γ diversities, size and trophic structure, and productivity–biomass (PB) relationships between locations [43]. The median number of species per resampled survey (scaled as spp. 100 m⁻²) was our metric of α diversity. We used the multivariate dispersion between resampled surveys (betadisper [95]) as our metric of β diversity, with a nonmetric multidimensional scaling (nMDS) on a Bray–Curtis species dissimilarity matrix used to help visualise patterns. Finally, γ diversity was estimated as the cumulative number of species in resampled surveys using sample-based rarefaction curves with the Coleman method [96,97].

Comparisons between localities (Raja Ampat, Lizard Island, and Ha’apai) were done via permutation tests (in betadisper), overlapping of 95% confidence intervals of estimates (rarefaction curves) or via Bayesian generalised linear models (GLMs). Bayesian GLMs had negative binomial (for total abundance and α diversity) or gamma (for standing biomass and productivity) error structures and used log link functions in both cases, with the following model structure: where y_s were the observed values of each indicator for each survey area s; and represent the mean of the negative binomial and gamma distributions, respectively; β₀ is an overall intercept; and β₁ is an effect (intercept) for each locality, both of which with weakly informative, normally distributed priors. Differences between localities were summarised by the ratio of the posterior distributions of locality intercepts (i.e., β₀+β₁) between Raja Ampat and Lizard Island (hereafter β_RL) or between Raja Ampat and Ha’apai (hereafter β_RH). As total fish abundance was the criterion used to inform the resampling procedure described above (see Study design and survey datasets), abundance was the same for each survey across all resampling iterations. Therefore, only 1 model (instead of 1,000 bootstrapped models) with 3 MCMC chains, 5,000 steps per chain, a thinning of 1 every 3 steps, and a 50% burn-in was used to compare total fish abundance between localities. For the other 3 comparisons among localities (i.e., species density, biomass, and productivity), all 1,000 models used parameters as above and passed the convergence checks, being thus retained for the compounded posterior distributions.

PB relationships describe variability in relationships between standing biomass and productivity and have been shown to be linked with reef fish turnover and exploitation rates [43]. We defined a model for PB relationship as: where y_s represents the observed values of productivity in each survey area s; β₀ is an overall intercept; β₁ is an effect (intercept) for each locality; and β₂ is an effect (slope) for the standing biomass values of each survey area. We also compared PB relationships between localities using ratios. However, in this case, β_RL and β_RH are the ratios between the posterior distributions of the slopes of the curves in Raja Ampat and Lizard Island, and Raja Ampat and Ha’apai, respectively. For PB relationships, 994 out of 1,000 bootstrapped models passed the convergence checks and were retained for the compounded posterior distributions.

Size structure was investigated using biomass and productivity size spectra. These consisted of the cumulative biomass or productivity per fish size class (body mass in g) in log₁₀ bins. We visually examined the overlap between size-spectra resampling quantiles for each size bin. Trophic structure was first also visually evaluated with an nMDS using a square root and Wisconsin-transformed, Bray–Curtis dissimilarity matrix calculated from the mean productivity of the 5 trophic groups at each resampled survey. This revealed a clear separation in the dots, particularly between those of Raja Ampat and the other 2 localities, which appeared to be driven by associations with distinct trophic groups. We then quantified these associations using Pearson’s correlations. We calculated Pearson’s r values as the slopes of ordinary least squares linear regressions between the standardised productivity of each trophic group and standardised total productivity [98] at each locality, with intercepts fixed at 0: where ∑x_s is the total productivity of each survey area s; x_t,s is the productivity of each trophic group t at each s; and r_t is Pearson’s r for each t. Prior to standardisation, all variables were log₁₀ transformed to decrease dispersion among data points. All correlations were calculated for each resampling iteration assemblage, resulting in bootstrapped distributions of Pearson’s r that were analogous to the compounded posterior distributions of the bootstrapped Bayesian models above. We summarise these distributions with their 95% quantile intervals.

After finding a steep and tight relationship between planktivore productivity and total productivity (see main text), we also looked at the overall relationship between the proportional productivity of planktivores (predictor) and total productivity (response) using Bayesian generalised models. Because there was no reason to expect a specific shape for this relationship, we fit a series of models with different functions of proportional planktivore productivity: (1) a thin plate spline function (within a generalised additive model); (2) a log-linear function; (3) a second-degree polynomial function; and (4) a third-degree polynomial function. We compared these models using leave-one-out cross validation (using the package loo [99]), which showed that the best model (i.e., with the highest expected log predictive density (ELPD)) was model 3 (ELPD_Model3 = −475.9; ELPD_Model1 = −477.6; ELPD_Model2 = −532.7; ELPD_Model4 = −476.8). This model had the form: where y_s represents the observed values of total productivity for each survey area s; β₀ is an overall intercept; and β₁ and β₂ are the linear and quadratic effects, respectively, of x_pPK, the proportional planktivore productivity of each s. Here, the emphasis was on the shape of the relationship between predictor and response, rather than on the exact values (or boundaries) of the coefficients, and, therefore, we fitted only 1 model (i.e., instead of bootstrapping 1,000 models) using the median values of total productivity and relative planktivore productivity across the 1,000 resampling iterations. We used 3 MCMC chains, 5,000 steps per chain, a thinning of 1 every 3 steps, and a 50% burn-in. We also calculated Bayesian R² values for each step following [100], which we summarise with their median and 95% HPD.

Integrating datasets and predicting Indo-Pacific patterns of planktivore contribution to fish productivity

Finally, we integrated the ocean-scale, Indo-Pacific wide RLS dataset and our region scale, high-definition surveys from Raja Ampat, Lizard Island, and Ha’apai. To do that, we first modelled the effects of a series of potential predictors of the proportional contribution of planktivorous fishes to total fish productivity in the region-scale dataset. We then performed model selection (described below) to narrow down a final set of variables, which were used to predict the likely contribution of planktivores to the fish productivity of 1,031 sites that included planktivores, out of the total of 1,035 sites spread across most of the Indian and Pacific Oceans. Three sites from 1 ecoregion had unusually high values of one of the predictors and were subsequently excluded (see below). Thus, the final ocean-scale dataset used to predict planktivore contribution to total productivity consisted of 1,028 sites in 31 ecoregions.

Our initial set of potential predictors of planktivore relative productivity included 4 distinct classes: species-level predictors (average maximum species size, average species growth potential, and average species body shape), community-level predictors (planktivore abundance, species richness, and species composition), environmental predictors (bathymetric slope, mean and maximum surface current velocity, mean pelagic net primary productivity, and survey depth), and a proxy for human impacts (gravity of human impacts). These variables combine our knowledge of the drivers of fish biomass production (e.g., size, growth, overexploitation [43,101]), zooplankton availability flow (i.e., current speeds, phytoplankton productivity [14,39,73]), and physical constraints imposed by the energetic expenditure of swimming against the flow (e.g., [102,103]). Some of these predictors have also been found to drive the abundance and biomass of zooplanktivorous fishes across large spatial scales (e.g., [21,22]).

Species-level predictors were sourced from the same datasets used to estimate biomass and productivity (see From standing biomass to fish productivity: Combining growth and mortality trajectories above). Maximum species size and growth potential consisted of maximum length and K_max parameter values for each individual. Species body shape was the standardised length–weight parameter a, which should increase in value as species change from slender and elongated to rotund, deep-bodied [83,104]. Average values of these variables were obtained by aggregating across all individuals for each survey. Community-level predictors consisted of the following: (1) the abundance of all planktivorous species in each survey; (2) number of species of planktivore species in each survey; and (3) taxonomic composition. Taxonomic composition was quantified by the 2 main axes (2 orthogonal variables) of a principal coordinate analysis (PCoA) on a Bray–Curtis dissimilarity matrix of family-level abundances at all sites. Abundance data were fourth root and range transformed prior to the analysis. Following molecular phylogenetic evidence, we considered Caesionidae as part of Lutjanidae [105] and Microdesmidae as part of Gobiidae [106]. Analyses used the R package vegan [107].

An estimate of the “gravity of human impacts” for each site was obtained from the global dataset provided by [101]. We first converted the shapefile to a raster and then extracted values using the bilinear method, which interpolates the value of each point coordinate from its 4 nearest neighbours [108].

Finally, bathymetric slope, mean and maximum current velocity, and pelagic primary productivity were obtained from global rasters using the package sdmpredictors [109], which provides an interface between R and the Bio-Oracle and MARSPEC databases. Bathymetric slope is derived from the SRTM30_PLUS V6.0 [110,111], a 30 arc-second resolution, global digital elevation, and seafloor topography model). Surface current speeds are derived from the ORAP5-ECMWF [112,113], a 1° resolution global ocean model driven by atmospheric forcing fluxes. Surface pelagic NPP is obtained from PISCES-SEXTANT-IFREMER [113,114], a global 0.25° resolution biogeochemical model. Mean and maximum current velocities and pelagic NPP are monthly averages for the period 2000 to 2014 and were interpolated and downscaled by Bio-Oracle [113], using a kriging model to achieve a final resolution of 5 arc-minutes. Bathymetric slope, mean, and maximum current velocity values for each site were extracted from a buffering area with a radius of 10 km around each site, while surface pelagic NPP was extracted from a buffering area with a radius of 30 km around each site. All extractions were accomplished using the raster package [108] in R.

Before model selection, variables observed to be overdispersed were log₁₀ transformed, and Pearson’s correlations were used to assess the potential for multicollinearity. This exposed high correlations between planktivore abundance, species richness, and the first PCoA axis (r > 0.6 for all combinations, r = 0.80 between abundance and richness); species average maximum size and K_max (r > −0.76); and maximum current velocity and pelagic primary productivity (r = 0.91). Thus, we immediately excluded planktivore species richness, the first PCoA axis, species average K_max, and maximum current velocity. All remaining variable comparisons had r < 0.5. Model selection was performed on frequentist beta regression candidate model structures using the package betareg [115], with tools from MuMIn [116] in R. After narrowing down the model structure, the final model was refitted as a Bayesian beta regression following: where y_s represents the proportional planktivore productivity in each survey area s; is the mean of the beta distribution; β₀ is an overall intercept; and β₁ to β₄ are effects (slopes) for log₁₀-transformed planktivore abundance (β₁), average species maximum size (β₂), mean current speed (β₃), and pelagic primary productivity (β₄). For this predictive model of relative planktivore productivity, all 1,000 models passed the convergence checks and were retained in the compounded posterior distributions. We also calculated the Bayesian R² for this modelling procedure by aggregating R² values for all models in a compounded posterior distribution.

Despite the magnitude of differences in the geographic coverage and sample size between the ocean and regional datasets, heterogeneity in the range of the predictors and taxonomic composition were minor (S1 Text, S7 Fig), except for 3 sites that had unusually high pelagic primary productivity (i.e., >0.08 g m⁻³ day⁻¹, or more than 4 times over that of any other site). Given that these sites were located in a major urbanised area (Hong Kong), and hence likely subject to eutrophic conditions, these 3 sites were excluded before the predictive model above. Furthermore, a pairwise comparison between predicted proportional planktivore productivity in the ocean-scale dataset and empirically obtained values (region-scale dataset) revealed (1) substantial agreement for Raja Ampat and Lizard Island; and (2) similar mean values, despite a narrower range with lower maximum estimates in the ocean dataset for Ha’apai, likely to make our model conservative (S1 Text). Therefore, we used posterior linear prediction with this model’s coefficients to predict the likely contribution of planktivores to the productivity of each site in the large-scale dataset. These site-level posterior predictions of planktivore relative productivity were summarised by their median values.