An extinction event in planktonic Foraminifera preceded by stabilizing selection

Unless they adapt, populations facing persistent stress are threatened by extinction. Theoretically, populations facing stress can react by either disruption (increasing trait variation and potentially generating new traits) or stabilization (decreasing trait variation). In the short term, stabilization is more economical, because it quickly transfers a large part of the population closer to a new ecological optimum. However, stabilization is deleterious in the face of persistently increasing stress, because it reduces variability and thus decreases the ability to react to further changes. Understanding how natural populations react to intensifying stress reaching terminal levels is key to assessing their resilience to environmental change such as that caused by global warming. Because extinctions are hard to predict, observational data on the adaptation of populations facing extinction are rare. Here, we make use of the glacial salinity rise in the Red Sea as a natural experiment allowing us to analyse the reaction of planktonic Foraminifera to stress escalation in the geological past. We analyse morphological trait state and variation in two species across a salinity rise leading to their local extinction. Trilobatus sacculifer reacted by stabilization in shape and size, detectable several thousand years prior to extinction. Orbulina universa reacted by trait divergence, but each of the two divergent populations remained stable or reacted by further stabilization. These observations indicate that the default reaction of the studied Foraminifera is stabilization, and that stress escalation did not lead to the emergence of adapted forms. An inherent inability to breach the global adaptive threshold would explain why communities of Foraminifera and other marine protists reacted to Quaternary climate change by tracking their zonally shifting environments. It also means that populations of marine plankton species adapted to response by migration will be at risk of extinction when exposed to stress outside of the adaptive range.

).   (a) Shell size decreases over time but shell size variation remains rather constant. (b) Shell roundness variation increases due to the increase in abundance of the more variable small population.        Figure S . Results of a randomization approach to test trends in morphological variation during marine isotope stage in the Red Sea. All variation parameters were calculated for randomly resampled samples (with replacement, 1000 replications), where the sample size was kept as in the original samples.
(a-b) Shell size and roundness variation of Orbulina universa (large population only) as coe cient of variation. (c-d) Shell size (coe cient of variation) and shell shape variation (variance of the Riemannian shape distance from the grand mean) of Trilobatus sacculifer. In all cases, the random sampling (null-model) shows no trend at all, and the observed trend is always above the null-model earlier during the section and below the null-model when approaching the local extinction. This shows that the observed stabilization trend is stronger than can be expected by chance and is not biased by the decreasing sample sizes toward the local extinction.

Error discussion
Geometric morphometric analyses are error prone due to the large amount of manual steps involved, which can bias the reproducibility of the results. Amongst others, this includes the error of the measurement device and the measurer. We limited the device error by using the same microscope and camera for all photographs and using a constant magnification per species. The error by the measurer was limited by ensuring that all tasks were applied by the same researcher per specimen. Two other sources of error are only relevant for the geometric morphometric analyses of Trilobatus sacculifer. They were quantified using analysis of variances-based (ANOVA) approaches proposed by Yezerinac, Lougheed et al. [ ], and are discussed in the following.

. Error due to manual orientation of specimens
While all specimens per species were homogeneously oriented by the same researcher, eliminating a personal error term, a mis-orientation of specimens could still occur. To estimate that error, we repeatedly (eight times) reoriented three randomly selected specimens covering the whole observed size range, and estimated the landmark position error due to orientation. The analyses shows that manually orienting the specimens did not introduce a large error. The mean squares of the replicates (2219) is much smaller than the residual mean squares (23 739), with the ANOVA being highly insignificant (p = 0.999). This indicates a high reproducibility of landmarks regardless of small orientation errors, and the relative measurement error associated with specimen orientation sums up to only 1.07 .

. Error due to manual landmark placement
A second source of error is the problem of misplacing landmarks in the specimen images. Due to the large amount of samples it was not feasible to replicate landmark extraction for all samples, instead we replicated this step for two samples, one with on average very small specimens (1439.5-1440 cm) and one with on average very large specimens (1488-1488.5 cm). The reasoning behind this is that larger specimens have a higher e ective resolution, because they contain more pixels under constant magnification, and morphological details are better visible in larger specimens when the magnification is kept constant. It is thus likely that the landmark extraction error is not independent of specimen size. For both samples we replicated the landmark extraction and calculated the session error (i.e. the average mismatch between replicates) and the individual error (i.e. the error per specimen) as well as the relative measurement error (Table S ). In both samples we observe that the residual mean squares (error variance) of the session is much larger than the session factor mean squares, and that the ANOVA results for the session error are insignificant, indicating a high replicability of landmark positions in di erent sessions. The individual error ANOVA is highly significant in both cases, with the individual factor mean squares (between-specimen variance) being much larger than the residual mean squares (within-specimen variance). This indicates that the observed di erences between specimens are much larger than what could be explained by errors in landmark extraction. Accordingly, the relative measurement errors associated with landmark misplacement are very small, with 0.264 for small specimens and 0.148 for large specimens.

. Allometry analyses
One last potential problem in morphometric analyses, especially in geometric morphometrics, can be the influence of ontogeny on shape. This can be tested in two ways [ ]. ( ) One can adapt the univariate allometric equation and calculate the Riemannian shape distances between the smallest individual and all other individuals, and then calculate the regression between size and shape distance to the smallest individual. When doing this for our data for T. sacculifer ( Fig. S a) we find that the regression is insignificant (p = 0.919) and shell size explains less of the observed shape change than the null-model (R 2 < −0.001). ( ) Alternatively, one can regress the partial warps of the landmark configuration on centroid size data in a multivariate regression approach to investigate the shape change as a whole dependent on size [ , eq. . ].
Doing so reveals a significant allometric component in T. sacculifer shape (p < 0.001), but it explains only 4.10 of the observed shape changes and is thus practically negligible, because it cannot be responsible for the signals we see in the data. It is furthermore mainly limited to a widening of the aperture (Fig. S b), which is no trend that we observe in the T. sacculifer population in association with increasing stress levels. We thus conclude that allometry is no problem in our analysis of T. sacculifer shape changes with environmental stress.

(b)
Figure S . Allometry of Trilobatus sacculifer from marine isotope stage in the Red Sea. (a) The univariate allometry approach shows no significant correlation between shape (expressed as Riemannian shape distance from the smallest individual) and size. (b) The shape change with ontogeny is nearly exclusively limited to a widening of the aperture.