This repository contains R scripts to extract data and replicate Conditional Random Fields analyses used in Clark et al “Rapid winter warming could disrupt coastal marine fish community structure” (recently published in Nature Climate Change)

Overview
We apply Conditional Random Fields models to an open-source dataset of fish distributions to show that winter sea surface warming strongly predicts variation in coastal fish community compositions in the Mediterranean Sea. Specifically, we formulate an undirected network model that accounts for all combinations of other species and environmental covariates (winter and summer sea surface temperatures) on a species’ occurrence probability. By allowing species’ co-occurrence associations to vary in response to temperature changes, our model can generate future biodiversity predictions and identify putative, biologically-meaningful mechanisms that underlie our finding of changes in niche filling across seasonal temperature gradients.

Authors
This work was developed through a collaboration between Nicholas Clark, James Kerry and Ceridwen Fraser. The R code was developed and is maintained by Nicholas Clark.

Contacts
Please contact Nicholas Clark with any questions or correspondence.

Workflow
The workflow to extract primary data and complete statistical analyses is as follows:

1. Download open-source data on coastal Mediterranean fish species occurrences, functional traits and phylogenetic relationships, as well as historical and projected Intergovernmental Panel on Climate Change (IPCC) SRES A2 Sea Surface Temperature data, from the open-source FishMed database (Albouy et al. 2015). Note that these data are licensed and cannot be directly stored in this repository. However, R functions in the Functions directory are available to download the data directly into your working environment. Please follow steps in Workflow/Appendix_S1_Datacollation.pdf to extract the data.

2. Run Conditional Random Fields models using functions available in the MRFcov package (Clark et al. 2018) to analyse the binary presence-absence data downloaded in Step 1 above. While R code in Workflow/Appendix_S2_Models.pdf is available to replicate this step, note that a previous version of MRFcov was used and so some of the argument names may have changed but most steps should still work. Also please note that these functions were originally run on a high-performance computing cluster using 24 processing cores to speed up computations. The models will take several hours or more to complete even on a powerful machine. We have provided the original results to save processing time. These are available in Results/MRF.results.rda and can be loaded directly into R using the load() command.

3. Process the models to analyse their results and generate summary figures as shown in the original manuscript. The steps are outlined in Workflow/Appendix_S3_ModelProcessing.pdf. Most of these steps should be achievable even on personal laptops, apart from the fitting of the multivariate boosted regression trees (which rely on the mvtboost package, see below for instructions for downloading). This step is not partciularly slow but is very memory-heavy because each set of 500 regression trees is saved for each of the 215 fish species in order to find predictors that minimise joint prediction error. We have provided the original results of this step to save processing time. These are available in the Results/mvtb.results.rda and can be loaded directly into R using the load() command.

Key packages needed for analyses and making figures
devtools
ggplot2
cowplot
dplyr
plyr
readxl
rvest
tidyverse
stringr
sf
raster
gstat
sp
ape
viridis
MRFcov
mgcv
visreg
PhyloMeasures
pals(devtools::install_github('kwstat/pals'))
BBS.occurrences(devtools::install_github('nicholasjclark/BBS.occurrences'))
mvtboost(devtools::install_github('patr1ckm/mvtboost'))

References
Albouy, C., Lasram, F.B.R., Velez, L., Guilhaumon, F., Meynard, C.N., Boyer, S., Benestan, L., Mouquet, N., Douzery, E., Aznar, R., Troussellier, M., Somot, S., Leprieur, F., Le Loc’h, F. & Mouillot, D. (2015) FishMed: traits, phylogeny, current and projected species distribution of Mediterranean fishes, and environmental data. Ecology, 96, 2312-2313.

Clark, N.J., Wells, K. & Lindberg, O. (2018) MRFcov: Markov Random Fields with additional covariates. R package version 1.0. GitHub.

Clark, N.J., Kerry, J.T. & Fraser, C.I. (2020) Rapid winter warming could disrupt coastal marine fish community structure. Nature Climate Change DOI: https://doi.org/10.1038/s41558-020-0838-5

This project is licensed under the terms of the GNU General Public License (GNU GPLv3)