| 01 Jul 2026
Reconstructing pico- and nanophytoplankton assemblages from long-term coastal thermohaline observations
Abstract. Over the past 30 years, the Gulf of Lions in the northwestern Mediterranean Sea has undergone various environmental changes that have impacted marine ecosystems. At the beginning of the 21st century, shifts were observed in small pelagic fish and zooplankton communities. However, little is known about the impact of these changes on phytoplankton communities, especially pico- and nanophytoplankton (hereafter referred as PicoNano), which play a crucial role in the northwestern Mediterranean Sea. One major limitation is the lack of data on these size fractions during the observed shifts. In this study, we show that vertical profiles of temperature and salinity can serve to infer PicoNano assemblages from the thermohaline properties measured in a coastal area, the Bay of Marseille (BoM). We use functional data analysis and clustering methods to identify recurring vertical thermohaline structures over 20 years of temperature and salinity data (1994–2024), measured at low and high frequencies. To address the issue of truncated vertical profiles, we employ a reconstruction method based on functional principal component analysis. Using flow cytometry data collected over the past 11 years (2014–2024), we characterize PicoNano assemblages and the nutrient concentrations associated with each recurring thermohaline structure. We identify three thermohaline structures associated with the most oligotrophic conditions observed in the BoM, all involving a thermally stratified water column. The occurrence of these stratified conditions does not show any significant trend over the past 20 years. However, vertical structures indicative of Rhône River intrusions have been less frequently observed since 2004, while mixed water columns are more frequently highly homogeneous over the same period. We show that highly homogeneous water columns favour different PicoNano assemblages compared to mixed but less homogeneous water columns. Overall, these results suggest that conditions favouring larger nanoeukaryotes have become less frequent since 2004. This work highlights that easily measurable variables, such as vertical profiles of temperature and salinity, can provide valuable insights into more complex variables—such as nutrient concentrations and PicoNano assemblages—when direct observations are unavailable, a situation commonly encountered in long-term and high-frequency monitoring.
This manuscript links two decades of thermohaline profiles to phytoplankton assemblage shifts at a single Mediterranean coastal station. The dataset is impressive, the analytical approach is interesting, and the writing is clear. However, I have a few concerns, mainly that the authors often use a causal language that doesn’t match what the data show, and several analytical decisions lack justifications, which I explain below:
My first concern is that main conclusion covers a decade with no biological data. The abstract states that conditions favoring larger nanoeukaryotes have grown less frequent since 2004. The PicoNano time series begins in 2014. Ten years of the claimed shift were never observed. The paper shows that thermohaline structures changed after 2004. It shows that different structures associate with different assemblages after 2014. From these two facts it infers a third: that assemblages probably changed between 2004 and 2014. My issue is the language in the abstract and discussion. It reads as observed fact rather than inference. The causal chain: wind shift, then mixing, then smaller nanoeukaryotes, then reduced copepod lipids, then sardine decline. Each link rests on correlation across separate time series. None are modeled together. The chain's foundation is a single Wilcoxon test at p = 0.031, uncorrected, out of four tests run.
Second, the clustering pipeline rests on undocumented choices. The classification of 1,619 TS profiles into nine vertical structures is the analytical backbone of the paper, yet several decisions that directly determine the result are presented without validation. First, the authors report that BIC selected 9 clusters using a VVV model, but BIC has a well-known tendency to favor a larger number of clusters as sample size grows, and it is not guaranteed to identify ecologically meaningful groupings. No sensitivity analysis is provided showing how the biological or temporal results change under 7 or 11 clusters. Second, the clustering is performed on the first 4 MFPCA scores, justified by a “plateau” in explained variance, but the proportion of total variance actually retained by 4 components is never reported. This information is needed to assess whether the low-dimensional representation is adequate. Third, the uncertainty threshold of 0.25 (posterior probability > 0.75) is cited to Scrucca et al. (2023) as a general guideline, not validated for this specific dataset or application. With ~25% of profiles unclassified, the “NC” category is large enough to affect temporal trend estimates; the sensitivity of GAM results to this threshold needs to be tested.
Finally, the Prochlorococcus signal may be an instrument artifact rather than ecology. The paper acknowledges, briefly, that surface Prochlorococcus counts likely miss high-light-adapted ecotypes. These cells lose pigment under strong light and drop below detection. This matters because the paper's central summer-autumn pattern is exactly what that artifact would produce. Stratified oligotrophic water brings high light and photo-inhibition, pushing cells below detection. Mixed water brings low-light-adapted cells to the surface, where the instrument sees them. The paper treats the resulting abundance pattern as ecology, but no evidence is given to rule out instrument bias as the source. The authors should run a sensitivity analysis that removes Prochlorococcus from the analysis and report whether the assemblage separation holds.
Minor comments: