Primary question

To what extent do mesophotic coral ecosystems and shallow reefs share species of conservation interest?

Secondary questions consider effect modifiers:

Does the extent of community overlap vary by region or between broad taxonomic groups?

Deviations from the protocol

JSTOR (all years) was not searched as a database, despite being detailed in our protocol. This was because an information request was declined as a result of the high volume of enquiries. Journal articles were still retrieved from JSTOR through a Google Scholar search.

Articles were categorised by type as part of a supporting narrative synthesis. We generated a word cloud based on term frequencies from included titles and abstracts, in consultation with the advisory committee. The five, objectively derived, tags were: ecology, disturbance, trophic interactions, symbionts, and reproduction. Tags were assigned manually while screening on outcome at abstract level. Our full text screening for the primary review question, and quantitative synthesis, use studies tagged as ecology.

Method of data collection was not extracted, despite being identified in the protocol paper [29], as many articles used a mixture of different survey methods for shallow and deep sampling. AIC are not compared between models fitted to the full dataset and the high validity subgroup as the metric is affected by sample size.

Search strategy

We use two search strings to balance capturing relevant literature against articles captured after name changes in the field over time, as detailed in our protocol [29]. “Mesophotic” from 01/01/2010 to 31/12/2016, and “Mesophotic OR “Deep reef” OR ((Submersibl* OR Submarine* OR “Deep water” OR Trimix) AND (biolog* OR Reef* OR Cora* OR Spong* OR Alga* OR Fish*))” from the start of the source to 31/12/2009. The scoping exercise and search string development are reported in the protocol paper. The final search strings were successful in retrieving 87% of our test library.

We conducted our literature search in ISI Web of Science, Science Direct, Proquest, AGRICOLA and Scopus in English, with a full list of databases presented in our protocol [29]. We downloaded all bibliographic data from mesophotic.org at the end of 2016 and used Google Scholar for a complimentary search, capable of capturing grey literature [32]. The Google Scholar search was conducted using web-scraping software in R [33]. A detailed methodology can be found in Additional file 1, based on work by Haddaway et al. [34], retrieving up to the first 1000 entries. Following our literature search the personal libraries (Private collection of manuscripts) of authors and advisory committee members were checked for missed publications. Additional publications were also provided by the reviewers of this manuscript. Literature searches were conducted in the period from 19/04/2015 to 28/02/2017.

Article screening and study inclusion criteria

Before screening based on exclusion criteria occurred in the software EPPI Reviewer 4 [35], duplicate entries resulting from the use of multiple databases were removed. We set EPPI Reviewer to automatically mark as duplicates full bibliographic entries with 95% similarity. This level of similarity generally corresponded to the identical entries with differing capitalisation, or entries with one database field left blank in comparison to the master entry. Other possible duplicates were checked manually. When retrieving full-text articles for detailed screening not all were readily available. The lead author was contacted for any manuscripts the screening team could not find available online. A final list of full texts not retrieved after two attempts at contact with the authors was circulated to the advisory committee. If full texts were not present in personal libraries after these checks the articles were excluded. The screening was conducted by authors who had no publications in the list screened, preventing bias in the inclusion of data.

The original screening and validity appraisal criteria were reviewed to reconcile disagreements during Kappa analysis, at each level of screening, between the two reviewers. 1000 articles were coded by both reviewers at title and abstract level, 50 full texts were coded by both reviewers. Any differently assigned articles were discussed. The final criteria were as follows:

Critical appraisal

External validity considers the generalisability, and the fairness of comparison between different studies. Internal validity considers the risk of bias within a specific study. A number of methods exist to critically assess these concepts, but most are developed with randomised control trials in mind, or at least observational studies with a level of randomisation in allocation and blinding rarely found outside of the medical sciences [36].

We have no reason to believe the observation of a species depth range from one article is any more important than that from another [29]. By including geographical area and taxon as effect modifiers, we have controlled for the primary confounding sources of variability when comparing studies. We therefore consider the spatial scales of different studies are the primary remaining threat to external validity. A key based on a whole ocean region is likely to find a greater depth range for a given species than a study at a specific location, because the key is likely based on data from a broader range of abiotic conditions.

With respect to internal validity, our primary concern with mesophotic research is that sampling effort may decline with increasing depth. This can lead to an underestimate of the number of species present deeper than 30 m as rarer species are found in shallow water, but are missed at depth. Similarly, studies reporting only a small number of species may provide a less reliable estimate of community level overlap. This is because the importance of a single species depth range is elevated, when community overlap is calculated as a proportion. Our critical appraisal is therefore based on the following criteria:

Studies passing all these criteria were included in a high validity sub-group for sensitivity analysis. We choose to limit our high validity group to studies with more than ten species to avoid undue influence of single species over community level overlap values.

Data extraction

Maximum and minimum depths for each species reported in an article were extracted manually. A subset of 12 articles were checked for consistency of data extraction by a second reviewer. Discussion of these extractions produced the following clarifications. Depth ranges for adults and juveniles of a species were combined into a single depth range. For the purposes of defining multiple studies within an article, locations were considered different when across multiple countries, when provided. Australia and the USA were exceptions because of their size, here we took into account the seas around Australia, and the islands/coastlines sampled of the USA. The review was conducted by authors who had no publications in the list screened at full text, preventing bias in the inclusion of data. Data were restricted to species occurring on tropical and sub-tropical reefs when multiple habitats were studied in a single article, such as mangroves and reefs. If these data were not available, the corresponding author was contacted. Contact was attempted twice before asking members of the advisory committee to try a third time with any researchers they knew personally. Failure to respond to these contact attempts led to exclusion of the data set from the analysis (Additional file 2). A list of all articles contributing data to the final meta-analysis can be found in Additional file 2 [37–68], a list of the studies resulting from these articles can be found in Additional file 3.

Potential effect modifiers/reasons for heterogeneity

The location and broad taxonomic group were extracted from articles as effect modifiers, following consultation with the advisory committee [29], as different regions may have different transition depths and taxa may respond differently to depth.

Data synthesis

After reviewing the retrieved articles, a meta-analysis for the primary review question was deemed appropriate. The list of observed depth ranges was limited to species observed shallower than 30 m. Of these species, the proportion also present deeper than 30 m was calculated, grouped by article, site, and taxon. Log odds ratios were calculated in R [69] in the package metafoR [70] and passed to a mixed-effects model. A maximum-likelihood estimator with Knapp and Hartung adjustments [71] was used with location and taxon as effect modifiers. Likelihood-ratio tests assessed the statistical significance of model elements and determined whether to retain an interaction term.

Our sensitivity analysis and tests for bias follow our protocol [29], and were as follows. Funnel plots were used to identify potential outliers [72], and a regression test for asymmetry assessed the chance of publication bias affecting the dataset [72]. As a sensitivity analysis, Rosenthal’s fail-safe number [73] was calculated for the model to determine how many studies, averaging to null results, are required to lose statistical significance. To check sensitivity further, the analysis was repeated on a subset of high validity studies and the effects compared to a full model.

The broader research field was summarised using the bibliographic data from papers retained at full text level. Articles by year, categorisation, geographical area, and the mesophotic depth limit used were counted and plotted.

Review descriptive statistics

Our methods retrieved almost 170,000 articles, with 3011 screened at full text (Fig. 1). Kappa analysis was used for consistency checking and refining of the inclusion criteria by two reviewers for 1000/122,305 at title and abstract level and 50/311 at full text level. A list of all ecology articles screened at full text level, with decisions, can be found in Additional file 4. The following summary and figures of the systematic review, as opposed to the meta-analysis, are created from the data extracted from all articles; not excluded on outcome at the abstract level and verified as mesophotic at full text level. This brief discussion provides context of the wider field of mesophotic research, before we address the primary review question with a meta-analysis. The bibliographic data for these 461 articles were deposited with mesophotic.org.

There was a notable acceleration in the number of articles per year as we entered the twenty first century (Fig. 2a). Publication levels were relatively stable during the twentieth century with only 1 year yielding more than 10 papers. The earliest article we identified was from 1960.

After more than half a century of research, the mesophotic field is still largely descriptive and focussed on pattern. This is supported by the majority of studies being assigned an ecology tag during our screening process (Fig. 2b). The next most common tag was assigned six times less frequently and indicates articles considering disturbance or damage on mesophotic reefs. Other common MCE study types refer to trophic interactions, symbioses, and reproduction.

This exercise was used as an opportunity to see how researchers define mesophotic reefs. 27% of articles include a depth limit in their introductions. The overwhelming majority of papers providing a definition state MCEs start at 30 m depth (89%), however, some papers exist claiming the upper bound is 20 m or even as deep as 60 m.

Mesophotic reefs in the western Atlantic, Caribbean, and Gulf of Mexico, are the focus of most published studies, followed by those on archipelagos in the Pacific (Fig. 2d). Australian reefs represent another hot spot of activity, but are nonetheless not as well studied. The Red Sea has provided a relatively large number of articles for its area, while the Indian Ocean has been comparatively understudied.

Data-synthesis

Across the whole dataset, the median proportion of shallow species also present below 30 m is 0.67. The meta-analysis was performed on articles assigned as Eco at abstract level screening, and subsequently passed full-text level screening, as well as two further publications provided as titles at the review stage. Of these 81 + 2 articles (Fig. 1), data were successfully extracted from 35, yielding 52 studies (combinations of broad taxon and locations within articles) after contacting corresponding authors (Additional files 2). The consistency of data extraction was checked by a second reviewer for 12/35 articles included. The data can be found in Additional file 3. Ecology papers we could not source a full text for are listed in (Additional file 5).

Both location and taxon were included as effect modifiers to aid biological interpretation and reduce the variance assigned to between study heterogeneity (I² of 10.1% reallocated). An interaction term was not included as intended in our protocol. This decision was made as studies were not evenly distributed across all factor-level combinations (Additional file 6) and a likelihood-ratio test failed to distinguish the two models (χ² = 10.8, P = 0.14). A Q-test detected significant between-study heterogeneity (QE = 182.2, df = 27, P < 0.0001). An I² of 72.3% can be interpreted as “substantial” under Cochrane review guidelines [74], but does increase with study number [75] and may be less important for an ecological synthesis. The R² of our model is 75.9%. A forest plot of effect size contributions from individual studies can be found in Additional file 7.

Figure 4 visualises the proportion of shallow species present deeper than 30 m, predicted from the model, for each modifier level. The central Pacific has the highest community overlap between shallow and mesophotic reefs at 72% (95% CI 53–85%). Studies off the cost of Brazil report the lowest at 30% (7–72%) when ignoring areas represented by a single study. For taxa, models of algal groups all returned predictions of over 75% overlap. 86% (37–98%) of shallow-water Antipatharia were found on mesophotic reefs, while octocorals were predicted to have the lowest overlap of all factor levels at 38% (15–68%). The amount of studies retrieved for a given factor level does influence the width of the confidence interval. Scleractinia with 18 studies has a prediction ranging by 29%, whereas Octocorallia with three studies ranges by 53%. Data underlying the figure can be found in Additional file 8.

As part of the sensitivity analysis 22 studies were excluded from the high validity group, following our validity appraisal criteria. The analysis was then re-run to quantify the influence of these studies over our synthesis (Fig. 4). These studies are indicated in the comments column of Additional file 2. The dataset-wide estimate of community overlap dropped by 10 percentage points to 0.57. No change in broad pattern is noted, only a slight increase in the range of predictions, and a slight decrease in estimated overlap. We would generally expect a high validity sub-group analysis to narrow the prediction intervals. Our results suggest that the overriding influence may be the reduction in sample size, and that studies outside of the high validity group are largely similar to those retained. This is despite low validity studies operating at larger spatial scales and with unbalanced sampling effort with depth. A regression test provides no evidence for publication bias in the dataset (t = 0.5382, df = 33, P = 0.594), also visualised with a funnel plot (Additional file 9). This is further supported by a Rosenthal’s fail safe number [73] of 1612 studies, it is highly unlikely this many mesophotic surveys remain unpublished given the number we retrieved.

Quantitative results

Two-thirds of shallow-water species can be found on MCEs (Fig. 3). This suggests MCEs should potentially be considered as a significant conservation target when protecting shallow reefs. This observation is in broad agreement with expert opinion regarding the Caribbean; suggesting over 40% similarity for shallow and mesophotic benthic taxa and approximately 60% for fish [8]. Semmler et al. [8] use Jaccard similarity and will therefore estimate lower similarity than this study. Jaccard similarity returns a lower score for reefs with mesophotic-specialist taxa while in this study community overlap was calculated excluding species not observed on shallow reefs.

As individual studies would lead us to expect, community overlap varies by region (Fig. 4, Additional file 8). The highest proportion of shallow species present on mesophotic reefs was predicted for the central Pacific at 72% (95% CI 53–85%). Brazilian reefs and the Coral Sea with the Great Barrier Reef both harbour less than half of shallow reef taxa on MCEs at 31% (7–72%) and 41% (20–65%) respectively. A number of drivers may explain these patterns but there are two likely candidates.

Reefs located in biodiversity hot spots, such as the Coral Triangle and Red Sea [76], may experience greater competition and subsequent specialisation [77]. This may lead to narrower depth ranges and so less community overlap in comparison to species-poor regions such as the Caribbean and Hawai’i. Additionally, the boundaries between mesophotic communities and shallow reef communities may vary between sites [22]. This is likely underpinned by abiotic factors interacting with species physiology. The current use of an arbitrary 30 m depth limit, derived from recreational SCUBA limitations, may be limiting our ability to accurately report on ecological pattern. If the transition depth occurs deeper, a larger portion of shallow water species will be present below 30 m. The deviation in community transition depth from 30 m between sites may explain some of the variability in our estimates of community overlap. The maximum depth limits of photosynthetic corals are known to vary by region [78] and with light [2] and loosely correlate with our findings.

Community overlap can be broken down further by predicting the effect of different taxon levels (Fig. 4, Additional file 8). The two taxonomic groups with the largest number of studies contributing to the synthesis, fish and Scleractinia, are estimated as having 64% (46–79%) and 57% (42–71%) of shallow taxa on MCEs respectively. These values are relatively close to our synthesis-wide estimate. Other taxonomic groups have relatively large confidence intervals, but all three algal taxa return high community overlap of over 75%. This may be surprising for autotrophic taxa, but it appears a 30 m boundary is within the physiological envelope of most algal species identified. The high estimate of 86% (37–98%) for Antipatharia, though unintuitive for a taxon traditionally considered ‘deep’ [79], likely results from limiting the analysis to species observed shallower than 30 m. For the most part, if a black coral is observed shallower than 30 m it is likely a deep species encroaching on shallow reefs. Alternatively, a small number of species may have evolved towards shallow-water emergence. The smaller number of species contributing to estimates for taxa other than fish and Scleractinia highlights the need for taxonomic diversity in future mesophotic studies. It is also important to maintain shallow sampling for contemporary comparisons to allow for site to site variation.

Our finding of over half of shallow taxa being recorded on MCEs is not in direct contention with research suggesting mesophotic-specific communities on MCEs [8, 22, 32, 80–82]. Firstly, we deliberately exclude taxa only observed on mesophotic reefs from our analysis. This is as MCEs may extend below 150 m [2] while the majority of the studies synthesised here do not sample as deep. Any estimate of the proportion of mesophotic taxa not on shallow reefs would likely be misleading because of an absence of evidence (Additional file 10). Additionally, the lower limit of MCEs is acknowledged to be variable [21, 24]. Further analyses considering overlap at the lowest limit of MCEs would therefore be inappropriate because of a similar lack of sampling. Our primary research question considers the potential for shallow water species to exploit deep reef refuges.

Secondly, we consider MCEs in a narrow sense as a 30 m depth limit. As depth increases the ratio of shallow taxa to mesophotic-specialists declines (Additional file 10) [8]. When sampling is evenly distributed across the whole depth range of an MCE it is possible to detect depth-specific community structure [22] and high levels of geographically endemic biodiversity [20]; though this likely varies between locations [22]. It is important to note consensus in the field is to recognise an upper mesophotic zone, in which shallow species are gradually lost with depth, and a lower mesophotic zone containing specialist communities [21, 25], rather than rigidly following the 30 m boundary. Any conclusions regarding shallow taxon occurrence deeper than 30 m is therefore likely driven by pattern in the upper mesophotic. We are reticent to perpetuate the definition of these zones by arbitrary depth limits because of the potential for between site variation [22].

Qualitative analysis

At the outset of this project www.mesophotic.org represented the largest publicly available compilation of mesophotic literature, with the metadata of 293 articles available. This study succeeded in boosting this number to at least 461, considerably increasing the visibility of research. Absent from this list of papers are any published in 2017 or later. The new database offered the opportunity to categorise these studies, describing the field as a whole.

Past reviews have noted the geographical bias of mesophotic research [2, 12, 78]. This is to some extent unavoidable because of the locations of research groups in developed countries near accessible reefs (i.e., Hawai’i, US Virgin Islands, and Israel among others). We confirm the bias towards studies in the central Pacific and western Atlantic region. A better global understanding of MCEs requires increased research attention in the Coral Triangle and the wider Indian Ocean to the west. Currently MCE research attention focuses on reefs with lower coral species richness [76]. Though this may be convenient for a number of reasons, it may mean ecological differences between shallow and mesophotic reefs resulting from greater competition and niche diversification are missed.

In addition to geographic biases, the mesophotic field lacks diversity in terms of study types. As expected for a field which has only recently taken off (Fig. 2a), a large proportion of the studies focus on ecological description. This has already been noted across existing research and narrative reviews [12] but we lend support to the observation with a larger literature base. This pattern has been shifting over time [12], but there is certainly room for an increase in experimental, modelling, and longitudinal studies considering processes instead of pattern.

A final interesting observation is that the term “mesophotic” has been defined differently across published work. Though the overwhelming majority of papers define mesophotic reefs as starting at 30 m, the mesophotic research strategy more loosely states they can start from 30 to 40 m [1]. This 30 m limit is rooted in recreational dive limits, rather than biological boundaries which can vary [22]. Not surprisingly the next most common definition is for mesophotic communities starting at 40 m. More interestingly, six papers within the last 10 years define MCEs from 50 m with another one at 20 m and one at 60 m. As the number of papers on mesophotic reefs increases it is important to remain consistent in the application of terminology, or specify explicitly how these boundaries may be recognised to vary.

Conclusion