Testing the phylogenetic gambit: how much functional diversity can we reliably conserve if we prioritize phylogenetic diversity?

In the face of the biodiversity crisis, it is argued that we should prioritize species in order to capture high functional diversity (FD). Because species traits often reflect shared evolutionary history, many researchers have advocated for a “phylogenetic gambit”: maximizing phylogenetic diversity (PD) should indirectly capture FD. For the first time, we empirically test this gambit using data from >15,000 vertebrate species and ecologically-relevant traits. Maximizing PD results in an average gain of 18% of FD relative to random choice. However, this average gain hides the fact that in over 1/3 of the comparisons, maximum PD sets contain less FD than randomly chosen sets of species. These results suggest that, while maximizing PD protection can help to protect FD, it represents a risky strategy. Statement of authorship FM, MP, MC, SD, GVDR, RG, AOM, CT and WP conceived the design of the study. FM and GVDR conducted the analysis. FM, RG, MP and WP interpreted the results and wrote the first draft of the manuscript. All authors edited the final version. Data accessibility statement Most of the data is publicly available (see methods). The Fish data is available upon request. Code accessibility statement R functions developed in this paper are available at https://github.com/FloMazel/FD_PD_Max


Introduction
We are in the midst of a period of heightened biological extinction, with rates several orders of 63 magnitude higher than background rates estimated from the fossil record 1-3 . In addition to Following this logic, phylogenetic diversity has formed the basis of global conservation 82 schemes, notably the EDGE program 18 has been used by restoration biologists 19 and has been 83 widely embraced by researchers across the biodiversity sciences 20-23 . Despite this enthusiasm, 84 the critical question of whether maximizing PD will actually capture more FD than prioritization 85 schemes that ignore phylogeny has, to our knowledge, never been empirically tested 16 . Some 86 studies have discussed 24,25 and documented the relationship between FD and PD, both at 87 regional 26 and global scales 20,22 , and many of these studies have shown that maximizing PD 88 does not maximize FD. However, such studies do not test the fundamental phylogenetic gambit 89 at the heart of all PD-based conservation strategies: maximizing PD captures more FD than 90 5 randomly choosing species. No one would dispute that the best way to maximize FD is to 91 prioritize FD, but phylogenetic diversity has emerged as prioritization tool because we rarely 92 have sufficient trait data to calculate FD. Here we test whether PD-based conservation passes a 93 much less stringent, but ultimately more fundamental, test: is conserving on the basis of PD 94 better than conserving at random? Worryingly, a recent theoretical study has indeed 95 demonstrated that PD could be a poor surrogate for FD and, in some scenarios, prioritizing 96 species on the basis of PD could actually lead to capture less FD than if species were simply 97 selected at random 16 .

98
This points to the need for empirical tests of whether -within a given species pool-sets of 99 species selected to maximize PD actually contain more FD than sets of species selected without 100 regard to evolutionary relatedness. We clarify what our goals are in testing the utility of PD to 101 capture FD. First, we take as given that maximizing PD is not the overarching goal per se of PD-102 maximization schemes, but rather that a PD maximization strategy is valued for its ability to 103 capture more FD compared to a strategy that ignores phylogeny. Second, it is important to note 104 that we are selecting species sets to maximize PD or FD within a region. While this is a 105 simplification, as conservation actions often aim to select sets of areas (e.g. in reserve design), 106 the only global phylogenetically-informed conservation initiative is species-centered 18 (EDGE).

107
Critically, the question we raise has been shown to be distinct from asking whether traits have 108 phylogenetic signal (whether closely related species tend to share similar sets of traits), since strategy that maximize PD captures FD relative to an optimal and a random strategy. To do so, 128 we compare FD accumulation curves (i.e. FD computed for increasing proportion of the species 129 pool considered) across these three different sampling strategies: the random sampling (i.e. 130 rarefaction curve, averaged over 1000 sets), the maxPD (surrogacy, averaged over 1000 sets) 131 sampling (i.e. the sets that maximize PD) and the maxFD (optimal) sampling (i.e. sets that 132 maximize FD, see legends). Then, we measure the surrogacy of PD for FD (SDPD-FD) as the area 133 between the random and the maxPD curve ('A', see legend) divided by the area between the 134 random and the maxFD curve ('A+B', see legend). If SDPD-FD is positive, PD is a good surrogate 135 for FD (the maximum value being 1 where PD is an optimal surrogate) while when SDPD-FD is 136 negative preserving species based on PD is worse than preserving them at random.

139
We find that selecting the most phylogenetically diverse sets of species within a given 140 taxonomic family or within a given geographical location (large grid-cells across the globe) 141 captures, on average, 18% more FD than sets of randomly chosen species (i.e. SPD_FD = 18%, SD 142 +/-6.5% across pools, see Figure 1 and S1-2). Although the surrogacy is generally positive, there 143 was substantial variation across species pools. For example, the surrogacy of PD varies widely 144 from a minimum of -85% to a maximum of 92%, meaning that selecting the most 145 phylogenetically diverse sets of taxa can capture either 85% less (or 92% more) FD than sets of   However, even if in the majority cases maximizing PD does, on average, better than an 171 averaged random selection, this does not capture the reliability of its performance. The PD-172 maximization and the random selection strategies exhibit variation: simply by chance, random 173 selection of species can capture very high (or, conversely, very low) FD, and the same may be 174 true (to a previously unstudied degree) for PD. The extent of this variation is important: if it is 175 less than the average difference, PD-maximization is a reliable strategy as it will always yield 176 more FD, but if it does not, then PD-maximization could be unreliable for individual 177 conservation interventions. To contrast these two situations, we measured the fraction of times 178 that, within each species pool, the PD-maximization strategy yielded more FD than random 179 selection (see methods). PD-based selection was the best choice in 64% of cases (SD across 180 species pool=9%, see Supplementary Table 1 and Fig. S5), making it the better strategy but not 181 a perfectly reliable one. Thus, while the PD-maximization strategy has a consistent positive 182 effect (i.e. the average PD-maximization strategy yields more FD than the average random 183 strategy), its effect is weak (i.e. the PD-maximization strategy still yields less FD than the 184 random strategy in 36% of the trials within a species pool). 185 We next explored the drivers of surrogacies values across species pools. Surrogacy of PD 186 appears to weaken as the species pool richness increases (on average, Spearman Rho between 187 absolute surrogacies and species richness = -.15), most clearly seen in the tropics and in 188 species-rich families such as the Muridae (rats, mice and allies) and Columbidae (pigeons and 189 allies) ( Fig. 2-3). This is likely because our measure of FD (see Methods) rapidly saturates as the 190 number of selected species increases and species from these large pools harbor high functional 191 redundancy, such that a random prioritization scheme performs relatively well, or at least no 192 worse than other strategies (Fig. S6). In contrast, FD can be greatly increased by prioritization of 193 species using PD from species poor assemblages or clades. This is particularly the case in spatial 194 assemblages containing multiple taxonomic orders, which are both phylogenetically and 195 ecologically divergent from one another. Interestingly, the PD-FD relationship was not 196 consistent across taxonomic scale: we found that, in contrast to patterns at the family level, for 197 certain mammalian and avian orders (which are older than the families described above), using 198 PD to select species is much worse for capturing FD than choosing species at random (see, for 199 example, the Afrosoricidae, Chiroptera, and Charadriiformes in Fig. S7). 200 We then explored whether we can explain this variability within-and between-datasets, 201 and in particular, why for some assemblages/clades, a PD-prioritization strategy fails to capture 202 more FD than random choice. It is often implicitly assumed that phylogenetic signal (i.e. the 203 degree to which closely related species tend to harbor similar sets of traits) can be used to

209
For mammals, regions where PD did worse than random were located in the Sahara, 210 south western Patagonia, southern Africa including parts of Madagascar, and New Guinea 211 ( Figure 2). These latter two in particular are of concern since they are global conservation 212 priorities on the basis of species endemism and habitat loss. We suggest two historical reasons 213 for such idiosyncratic poor performance of PD. First, there is a tendency for a large carnivore 214 species, either a top predator (e.g., cheetahs in the Sahara or foxes in Patagonia) or a large 215 scavenger (e.g., the hyena in South Africa) to co-occur with a close relative with distinct traits in 216 these areas (e.g., a desert cat with the cheetah or the aardwolf with the hyena, see Fig. S10).

217
Only one of these closely-related species will tend to be selected under prioritization schemes 218 that maximize PD, thus reducing the volume of the convex hull on average when the 219 functionally distinct one is not selected (the large predator or scavenger). This seems also to 220 drive the low surrogacy of PD in Charadriiformes (especially Larus and Sterna; see Figure S10).

221
Second, lineages in which traits evolve very slowly will contribute little to FD, even over long 222 periods of time (branch lengths) that contribute greatly to PD. For example, in New Guinea 223 many co-occurring bats with similar traits diverged long ago, such that they are always selected 224 in the PD maximizing set, but do not add much to the convex hull, resulting in a poor surrogacy 225 of PD for FD. Such strong ecological niche conservatism is common in mammals 29 , e.g. in the 226 Geomyidae: two basal branches of the Geomyidae tree harbor very similar traits (species 227 descending from these branches are actually grouped in the same genus Thomomys) while 228 being distantly related in the phylogenies we used (Fig. S10). As such, they will be selected in 229 all PD maximizing sets, but will not contribute greatly to FD. test the phylogenetic gambit of whether PD prioritization captures more FD than random 251 selection (which has not, to our knowledge, been tested) 16 . Here we have shown that the 252 phylogenetic gambit holds: that PD is an effective conservation metric to capture FD. Yet we 253 13 also show that it remains something of a gambit: PD is good 'on average', but there is still some 254 risk associated with taking it. 255 We found that prioritizing the most phylogenetically diverse set of taxa in a region or 256 clade will result in an average gain of 18% functional diversity relative to applying the same 257 conservation effort without considering phylogeny, but this gain will decrease as species commonly focus on 'unusual' clades with rapid divergences (e.g., cichlids); we show here that 268 divergence does not have to be that spectacular (e.g., African carnivores) to alter the PD-FD 269 relationship. Third, we found that while this strategy, on average, captures FD well, it is also 270 somewhat unreliable, and 36% of the time will not capture more FD than random choice. This 271 means that while the PD gambit can be a bet worth taking, it is still a bet with associated risk, 272 not a sure thing.

273
Our objective in this paper is to test the phylogenetic gambit using empirical datasets.

274
This means that we do not aim to provide a coherent prioritization strategy 35 , or ready-to use 275 conservation guidelines. Indeed, we simplistically and implicitly assume that chosen species will 276 either be saved or go extinct, and we have not linked our various scenarios to any particular (and therefore of their associated FD or PD). While our study is thus not directly applicable, the 281 14 test we are conducting is actually critical to validate (or invalidate) the use of PD in conservation 282 as a whole. While it is not clear whether our results would generalize to other taxa (although 283 we hope that others will extend our work and test the phylogenetic gambit in other systems), 284 we do feel it is important to consider the uncertainty that has been introduced into our analysis 285 as a result of uncertainty associated with the spatial scale of our analysis, our phylogenetic 286 data, and our choice of trait and measurement of FD.

287
The scale of conservation activities can vary, from the global scale of the hotspots 288 approach to local protected areas within a single country, but, unfortunately, the connection 289 between these scales remains unclear. For example, if the motivation for protecting FD is to 290 maintain community-driven ecosystem functions and services 6,36,37 , the value of a regional or 291 global focus may be questionable 38 , and studies are increasingly focusing on local scales 6 .

292
Ecologists are refining and improving our understanding of how local assemblages assemble 293 within a regional context 39 , and while the concept of the 'regional pool' of species is 294 increasingly being viewed as a simplification, it is unlikely that regional-and local-scale patterns 295 are totally disconnected. We emphasize that our results are relatively robust to variation in 296 spatial scale (see Fig. S3), but we acknowledge that future studies should test the phylogenetic 297 gambit at more local scale as well.

298
The set of species that maximize PD obviously rely on the phylogenetic hypothesis used.

299
No hypothesis is perfect or without uncertainty, and these phylogenetic uncertainties could in 300 turn impact the composition of the set of species that maximize PD and hence the surrogacy 301 values we compute. In this study, we explicitly took into account these uncertainties by using  The motivator of our test of the surrogacy value of PD for FD is the fact that ecologically-307 relevant trait data is in short supply, especially for rare and data-deficient species. Indeed, if it 308 were not for this relative paucity of data, we could simply prioritize species based on their 309 unique contribution to FD directly. Although there have been massive and well-funded efforts 310 to collect and curate trait data from across the Tree of Life 42-44 , we are still far from having 311 comprehensive coverage. Furthermore, despite recent progress 45 , it is still not fully understood 312 which traits are most relevant for responses to environmental change, or that contribute most 313 to certain ecosystem functions and services, and how these vary among systems. Our analysis 314 suffers from a similar data limitation. We chose these traits because they are frequently 315 collected in ecological studies, not because we know they are ecologically important. Our 316 assumption is that their phylogenetic distribution is typical of those traits that are most 317 desirable for the purpose of conservation and that our primary results are therefore widely 318 applicable. While we did test the robustness of our results to the variation of trait information 319 retained to compute FD ( Figure S1), it is true that, overall, we used a rather limited set of traits. 320 We acknowledge that it is possible that many other potential valuable traits are not captured 321 by our measure of FD. One of the ideas behind the use of PD is that phylogeny might account 322 for these for unmeasured and unmeasurable traits 9,14,15 , however, as this hypothesis is not 323 testable (we do not have these traits), it seems risky to assume it is true. Our objective here is 324 to test the phylogenetic gambit given the limited set of traits that we have: we consider that 325 carrying out our imperfect test is more informative than not carrying any test at all.

326
In conclusion, we found that maximizing PD results in an average gain of 18% of FD 327 relative to random choice. However, this average gain hides the fact that in over 1/3 of the 328 comparisons, maximum PD sets contain less FD than randomly chosen sets of species. These 329 results suggest that, while maximizing PD can help capture FD, it represents a risky strategy. If 330 maximizing PD is a risky strategy, then, should we abandon the use of PD in conservation? We 331 believe that before such dramatic decision, our test should be repeated across space, traits and 332 taxa, in order to narrow the uncertainties of our results. This is why we now urge others to 333 expand our simple phylogenetic gambit test to other clades and other traits in order to test the 334 generality of our findings. We hope that our study will stimulate the production of numerous 335 tests to finally rigorously assess the usefulness of PD in conservation.

365
The total number of grid cells was 3,646. Domestic and aquatic mammals were excluded 366 from the analysis. In order to make sure our results were not driven by the important trait 367 difference between volant and non volant mammals, we repeated our results excluding bats.

368
For birds, we repeated our analysis using the full ranges. Finally, we evaluated the 369 robustness of our result to the spatial resolution considered by repeating our analysis at a    (the brute force approach). However, this is not feasible in practice as the numbers of 458 combinations of selected species was too high (e.g., 10 71 possible sets for all mammal 459 assemblages). To rapidly and efficiently find the set of species that aim to maximize FD, we 460 developed a novel (at least in ecology) greedy algorithm. In brief, our approach iteratively 461 (starting with two species) select the species that is the furthest from the centroid of the 462 already selected set. To avoid selecting two species that are far from the centroid but close to 463 each other, we penalized the distance to the centroid by the distance to the closest neighbour 464 in the already selected set. Here we present in details the greedy algorithm we used to find the 465 set of species that maximize FD: 466 Step 1. Select the two species with the highest trait distance 467 Step 2. Compute the centroid of these two selected species 468 Step 3. Compute distances between species not in the set and this 'set centroid'.

469
Step 4. Penalize these distances by adding the following factor f (Eq. 1) 470 f = K x e L x minD (eq. 1) 471 with K and L being penalizing factors and minD the distance between a given candidate 472 species and the nearest species already in the selected set.

473
Step 5. Select the species that maximized the penalized distance 474 Step 6. Go back to step one with this new set of species until the desired number of To avoid arbitrarily setting the penalizing parameters, we tested 1000 pairs of parameters 478 drawn from a truncated normal distribution (mean=1, sd=.5) and retained the parameter pairs 479 that yielded the maximal FD.

480
In tests of subsets of the data for which finding the true maxFD was feasible, we found our 481 approach to adequately approximate the true maxFD and to produce a very good 482 approximation of the true degree of PD's surrogacy for FD (Fig. S2).

484
Measuring performance and surrogacy of prioritization schemes. 485 We use a common approach 27,28 to quantify the extent to which a given surrogate (here, the 486 maxPD choice) reaches a certain objective (here, maximize FD). Species from a given pool (i.e., This surrogacy metric is at 100% when the surrogate perfectly meets the objective (i.e., the 497 maxFD and maxPD curves are identical and the max PD set is the maxFD set), 0% when the 498 surrogate is not better than randomly chosen sets of species (i.e., the random and maxPD 499 curves are identical) and is negative if the surrogate choice is worse than random (i.e., the 500 maxPD curve is below the random curve). Correlates of SPD-FD were evaluated using Spearman 501 correlations.

502
Apart from focusing on average tendencies, we quantified the variability of the FD yielded by independently, the number of cases where FDrandom>FDmaxPD across the 1000 random *1000 506 maxPD sets combinations (i.e. 10 6 comparisons). We then averaged theses number across % of 507 selected species and report statistics across datasets (Supp. Table 1). 508