Passive sampling for grazer cues incorporates top-down effects in harmful algal bloom monitoring and toxin prediction

Significance

Harmful algal blooms (HABs) of Pseudo-nitzschia can introduce the neurotoxin domoic acid into marine food webs, posing a serious threat to seafood safety and wildlife conservation. Accurate HAB forecasting is challenging, and current monitoring efforts do not account for all known toxin drivers, such as top-down effects by grazers. We employed passive sampling to incorporate grazer cues from copepods, or copepodamides, into long term HAB monitoring. Copepodamides were correlated to copepod biovolume, showing utility as a proxy for grazer density, and furthermore, statistical models including copepodamides improved predictions of Pseudo-nitzschia and domoic acid. This work encourages the expansion of passive sampling in HAB monitoring, providing an ecological perspective on the complex drivers of toxin production.

Abstract

Domoic acid (DA) produced by toxigenic species of the diatom genus Pseudo-nitzschia during harmful algal blooms (HABs) bioaccumulates in filter feeders, bait fish, and commercial seafood products, posing a threat to marine organisms and public health. Copepodamides, a group of polar lipids exuded by copepods, are known to induce DA production in multiple species of Pseudo-nitzschia. However, copepodamides are not currently considered in HAB monitoring and prediction, in part because of challenges associated with measuring ambient concentrations of dilute and labile compounds. Here we employ solid phase adsorption toxin tracking (SPATT) to measure natural concentrations of copepodamides over 28 mo in Monterey Bay, California, a known hotspot for HAB events. Copepodamides were significantly correlated to copepod biovolume, confirming utility as a zooplankton proxy. In generalized linear models, copepodamides from SPATT provided greater accuracy and lead time for predicting Pseudo-nitzschia abundance and DA concentration in sentinel mussels than DA from SPATT. Field experiments in the Northeast Atlantic assessed sampling methodology, and lab experiments demonstrated DA induction in Pseudo-nitzschia multiseries isolated from Monterey Bay. Results show potential for copepodamides to improve HAB predictions and encourage development of novel approaches to include top-down information in HAB monitoring systems.

Pseudo-nitzschia is a genus of marine diatom capable of forming harmful algal blooms (HABs) that produce domoic acid (DA), a potent neurotoxin. Within the California Current System, Pseudo-nitzschia HABs have caused fatal poisonings of marine mammals and seabirds (1, 2). Extreme HAB events have also forced extended closures of the Dungeness and rock crab fisheries resulting in significant socioeconomic impacts (3). Furthermore, DA has been found across multiple trophic levels suggesting unknown vectors to seafood products that are outside traditional shellfish monitoring (3, 4). These threats highlight the need for advanced solutions to predictive models for Pseudo-nitzschia and DA to mitigate the ecological, social, and economic impacts of future HAB events.

Regions at risk of HABs monitor DA across multiple sources. For example, toxin concentration in mussel tissue (mDA), is used to regulate seafood safety, and harvesting is forbidden when samples exceed 20 µg DA g⁻¹ tissue (5). Discrete measurements from seawater can be partitioned as particulate and dissolved DA. While not readily accumulated by shellfish, the release of dissolved DA by cells through active transport or lysis is a significant component of the total DA pool (6, 7). Solid-phase adsorption toxin tracking (SPATT) is a passive sampling method that integrates dissolved toxin concentration over larger spatiotemporal scales than discrete sampling [(8), Fig. 1A]. SPATT samplers contain porous beads of adsorbent resin such as DIAION® HP20 (HP20), SEPABEADS® SP207 (SP207) and SP700 (SP700), among others (9, 10). SPATT is used globally as a semiquantitative tool, often complementing discrete methods, for monitoring a diverse suite of toxins, anthropogenic contaminants, and marine natural products (8, 11, 12).

Fig. 1.

Sample sites for field detection of copepodamides by SPATT. (A) SPATT ring deployed at the Santa Cruz Municipal Wharf (SCW). (B) Global and regional maps of field sites at (C) SCW on Monterey Bay, CA, and (D) Kristineberg Center for Marine Research and Innovation (KC) in Gullmarsfjord, Sweden.

SPATT has also been adapted to passively sample copepodamides, a group of taurine conjugated lipids produced by herbivorous copepods (13, 14). Copepodamides are a kairomone for phytoplankton, signaling grazer presence and inducing defensive traits in multiple taxa (15–17). In diatoms, inducible defenses against copepod grazing include frustule strengthening, colony-size plasticity, and production of secondary metabolites, such as DA (18). A recent meta-analysis showed DA induction by copepodamides was four times stronger than by nutrient manipulations (19). Moreover, zooplankton have been observed selectively rejecting cells with elevated DA content, illustrating a mechanism for top-down control on HAB formation (20). This mechanism in situ, however, has yet to be quantified. Individual copepods produce a spectrum of copepodamides, which are divided into structure groups, dihydrocopepodamides (dhCA) and copepodamides [CA, (13)]. For clarity, total copepodamides, or the sum of both groups, will be referred to hereafter as CA-lipids. CA-lipids are short-lived in the environment, deacylating on an hourly scale to form lyso-CA and lyso-dhCA (21). Lyso-lipids do not induce defensive traits in phytoplankton (22).

Despite the tight control of toxin production by predator cues, top-down information from grazers is generally not included in HAB monitoring or forecast models. To address this knowledge gap, we used SPATT passive sampling to measure CA-lipids over 28 mo at the Santa Cruz Municipal Wharf in Monterey Bay, CA (Fig. 1 B and C). We leveraged colocated HAB monitoring to create generalized linear models comparing the capacity of CA-lipids and DA from passive samplers to predict Pseudo-nitzschia abundance and mDA. Monitoring CA-lipids with SPATT has practical significance, namely, providing an efficient, biologically relevant predictor variable for empirical HAB models. The passive sampling method was benchmarked against the current standard method for estimating discrete CA-lipid concentrations during field experiments in the Northeast Atlantic (Fig. 1 B and D), demonstrating utility for short deployments in addition to weekly monitoring. As SPATT is already integrated in many routine monitoring systems, CA-lipid analysis incorporates grazer information without the need for zooplankton enumeration.

Results

CA-lipid and Copepod Monitoring.

From February 2023 to June 2025, total CA-lipid concentration in Monterey Bay, CA ranged from approximately 1 pg g resin⁻¹ in August, 2024 to 875 pg g resin⁻¹ in March, 2025 (Fig. 2 A and B). dhCA made up more than 50% of the total CA-lipid content in 94 out of 97 samples (Fig. 2 A and B), and the mean concentration of dhCA was significantly greater than CA in every month except January (paired Wilcoxon signed-rank test, P < 0.05, SI Appendix, Table S1). The highest proportional contribution from each subtype was dhCA 18:4, which made up 35% of the total CA-lipid content and 46% of the dhCA content, and CA 20:5, which made up 7% of the total CA-lipid content and 31% of the CA content. CA-lipid concentration varied significantly across months of the year (Kruskal–Wallis test, H = 46.84, df = 11, P = 2.3 × 10⁻⁶) with March and April accounting for 44% of the mean annual CA-lipid content (Fig. 2C). Intact structures were correlated to their respective degradation products (R² > 0.20, P < 0.05, SI Appendix, Fig. S2).

Fig. 2.

Observational data of (A) dihydro-copepodamides (dhCA) and (B) copepodamides (CA) concentration over 28 mo at the Santa Cruz Wharf. One SPATT sampler containing HP20 resin was analyzed for copepodamides (CA-lipids) per date. (C) Mean monthly CA-lipid concentration from weekly observations (2023 to 2025). Bars indicate mean ± SE. Sample number (N) contributing to the mean is indicated on top of each bar. (D) Linear regressions between ln-transformed CA-lipids and copepod biovolume, where shading represents SE of the fitted line. Winter storms impeded field sampling during February 2024 and January 2025.

Zooplankton net tows were paired with SPATT deployments from January 2024 to June 2025. Copepod biovolume ranged from 0.25 mm³ L⁻¹ on April 19, 2025, to 0.0004 mm³ L⁻¹ on July 26, 2024, and abundance ranged from 15,531 individuals m⁻³ on March 14, 2025, to 40 individuals m⁻³ on July 26, 2024 (SI Appendix, Fig. S3A). All intact CA-lipid structures were positively correlated with copepod biovolume (Fig. 2D, all R² > 0.21, P < 0.01; SI Appendix, Table S2). The strongest correlations to copepods occurred in dhCA 20:5, CA 14:0, and dhCA 18:4, followed by total CA-lipids, or the sum of all monitored structures (SI Appendix, Table S2). Copepod biovolume was more strongly correlated with abundance (Spearman ρ = 0.90, P ≪ 0.001, N = 55) than with average body size (Spearman ρ = 0.54, P = 2.0 × 10⁻⁵). The difference was statistically significant, revealing abundance as a stronger driver of copepod biovolume (percentile bootstrap, 20,000 resamples, 95% CI [0.178, 0.569], P = 2 × 10 ⁻⁴).

Harmful Algal Bloom Monitoring.

Pseudo-nitzschia abundance, mDA, and particulate DA peaked multiple times during the study period (Fig. 3 A and B and SI Appendix, Fig. S4). Particulate DA was correlated to mDA, indicating that both methods reflected the event magnitude (Spearman’s ρ = 0.66, P = 5.0 × 10⁻¹¹). Not all Pseudo-nitzschia blooms produce DA, and elevated mDA coincided with events in May 2023 and April to May 2025, but not May 2024 (Fig. 3B). DA from three different SPATT resins (HP20, SP207, and SP700) had statistically different distributions during the study period, highlighting their different adsorbance characteristics (Kruskal–Wallis, H = 7.57, df = 2, P = 0.02). DA in SPATT increased during periods of elevated cell counts, mDA, and particulate DA in spring HAB events from 2023, 2024, and 2025 (Fig. 3C and SI Appendix, Fig. S4). SPATT also showed high sensitivity and captured relatively small increases in mDA during January 2024, August 2024, and November 2024 (Fig. 3C).

Fig. 3.

Observational HAB monitoring at the Santa Cruz Municipal Wharf from February 2023 to June 2025. (A) Toxigenic Pseudo-nitzschia, (B) DA in mDA, and (C) DA from SPATT. Colors indicate different SPATT resins (SP700, HP20, and SP207). Scaled measurements of copepodamides (CA-lipids) from SPATT are shown as points on the secondary axis in the top two panels.

Pseudo-nitzschia abundance and mDA exhibited a strong correlative relationship in real time (Week Lag = 0, Spearman’s ρ = 0.60, ^{P = 9.1 x 10–8}, SI Appendix, Figs. S4 A and S5 B). In addition, Pseudo-nitzschia abundance from 1 wk prior also had a relatively strong correlation to mDA (SI Appendix, Fig. S5B). Of the three resins used, DA from SP207 (DA-SP207) had the strongest correlation to Pseudo-nitzschia abundance and mDA. The correlation to Pseudo-nitzschia abundance was strongest with a 3-wk lag in DA-SP207 (Spearman’s ρ = 0.31, P = 0.02), and the correlation to mDA was similar across 1-3 wk lags. CA-lipids from HP20 (CA-lipid-HP20) were also significantly correlated to both response variables, though with longer lag times. CA-lipid-HP20 from 6 wk prior had an especially strong relationship with Pseudo-nitzschia abundance (Spearman’s ρ = 0.58, P = 2.8 × 10⁻⁷, SI Appendix, Fig. S5A) and mDA (Spearman’s ρ = 0.35, P = 4.3 × 10^–3, SI Appendix, Figs. S4 B and S5 B).

Generalized Linear Models.

Prior observations of DA and CA-lipids measured from SPATT samplers were used as predictors for generalized linear models of binary response variables, Pseudo-nitzschia abundance (PN`) and toxin in mussels (mDA`). For PN`, DA-SP207 was a significant predictor with lags of 1 and 3 wk, though the best model based on Akaike information criterion (AIC) included a 1-wk lag (Fig. 4A). The area under the curve (AUC) for this model was 0.61 ± 0.14 (95% CI). In contrast, models using CA-lipid-HP20 to predict PN` were significant across all the tested lags (SI Appendix, Table S3). Consistent with results from the correlation analysis (SI Appendix, Fig S5) a lag of 6 wk produced the lowest AIC and the highest model quality (Fig. 4B). The AUC for this CA-lipid-HP20 model was 0.85 ± 0.09, showing strong performance and increasing prediction by 5 wk over the best DA-SP207 model.

Fig. 4.

Generalized linear models for predicting Pseudo-nitzschia abundance (PN`) by (A) DA from SPATT (DA-SP207) with 1-wk lag and (B) copepodamides from SPATT (CA-lipid-HP20) with a 6-wk lag. Generalized linear models for predicting DA in mDA` by (C) DA-SP207 with 1-wk lag and (D) CA-lipid-HP20 with a 7-wk lag.

For mDA`, we tested two response thresholds, the median toxin concentration (0.025 μg DA g mussel<sup>−1</sup>) and a more stringent, 3rd quartile threshold (0.10 μg DA g mussel<sup>−1</sup>). Results for the former are contained in SI Appendix, Table S4. Increasing the threshold produced stronger model fits for both predictors (SI Appendix, Table S5). The best DA-SP207 model had a 1-wk lag (Fig. 4C), and the AUC was 0.70 ± 0.12 (SI Appendix, Table S5). In contrast to PN`, CA-lipid-HP20 with a 7-wk lag was the best predictor of mDA` (Fig. 4D). The AUC for CA-lipid-HP20 was 0.74 ± 0.15, higher than the best DA-SP207 model.

The F₁ score for predicting PN` with CA-lipid-HP20 ranged from 0.27 to 0.54 (SI Appendix, Table S3). For predicting mDA` with CA-lipid-HP20, F₁ = 0.39 for the 7-wk lag model, slightly below the best CA-lipid-HP20 model (SI Appendix, Table S5). These scores suggest the models would benefit from improving recall or reducing the false negative rate. However, CA-lipid-HP20 models were more successful at predicting true positives and had lower false negative rates than DA-SP207 models for both response variables (SI Appendix, Table S6).

CA-lipid Experiments.

A Pacific strain of Pseudo-nitzschia multiseries induced by CA-lipids resulted in 0.9 ± 0.1 ng DA mL⁻¹ (mean ± SE) in the highest CA-lipid treatment (Fig. 5A). The experiment showed a significant difference in toxin production (ANOVA, F_2,9 = 58.21, P = 7.10 × 10⁻⁶) and chain length (Fig. 5B, ANOVA, F_2,9 = 22.22, P = 3.30 × 10⁻⁴) between treatments. For total DA, the highest CA-lipid treatment was significantly greater than both the 0 and 1 nM treatments (Tukey HSD, P = 1.47 × 10⁻⁵ and P = 1.92 × 10⁻⁵, respectively), while the 0 and 1 nM treatments were not different (P = 0.95). Similarly, pairwise analysis of chain length showed the highest treatment (10 nM CA-lipid) was reduced relative to 0 nM and 1 nM treatments (Tukey HSD, P = 8.93 × 10⁻³ and P = 2.52 × 10⁻⁴, respectively). We could not distinguish pair-wise statistical differences in cell DA quota due to high variance in cell counts (Fig. 5C). There was also no significant difference among treatments for Chlorophyll a (Chl-a, Fig. 5D, ANOVA, F_2,9 = 0.66, P = 0.55) or cell counts (Fig. 5E, Welch’s ANOVA, F = 0.41, P = 0.68), which suggests that CA-lipids did not have a detectable effect on growth rate in the experiment.

Fig. 5.

Induction of a Pacific strain of P. multiseries with copepodamides (CA-lipids). (A) Total DA, (B) chain length, (C) cell DA quota, (D) chlorophyll-a (Chl-a), and (E) cell count. Start samples were taken before CA-lipid induction and were not included in statistical analyses. All samples depicted as the mean ± SE (N = 4). Letter groups indicate statistical significance by ANOVA + Tukey HSD (α = 0.05).

Field experiments from the Northeast Atlantic were used to assess reproducibility and interoperability of the passive sampling method. Short-term SPATT deployments were successful, and CA-lipids were recovered from all samples (SI Appendix, Fig S1). CA-lipid recovery exhibited an additive pattern with no significant differences between the sums of sequential 6- and 12-h deployments compared to the full 24-h deployment (t test, P > 0.05). SPATT was also compared to solid phase extraction (SPE), the main sampling method for CA-lipids (SI Appendix, Fig. S1 A and B). The relative abundance of CA-lipids in SPE and SPATT samples differed only in 20:5 dhCA and 22:6 CA (PERMANOVA + Bonferroni-corrected Wilcoxon test, P = 0.04 and P = 0.02, respectively; SI Appendix, Fig. S1F). Additional details are contained in SI Appendix.

Discussion

Grazers structure phytoplankton communities through induced defenses (18, 23) and selective grazing (24, 25). However, mechanisms of top-down control are not generally accounted for in HAB monitoring, and their relative importance in natural HAB assemblages is poorly understood. Here we integrated CA-lipids with an existing observing system, facilitating the inclusion of nonconsumptive grazing effects in empirical forecast models for Pseudo-nitzschia and DA. CA-lipids improved accuracy and lead-time relative to traditional toxin monitoring with SPATT, and CA-lipids induced toxin production in locally isolated P. multiseries, demonstrating grazer induced toxin production in a Northeastern Pacific strain of Pseudo-nitzschia. CA-lipids from passive sampling were correlated with copepod biovolume in weekly sampling, and CA-lipids were successfully recovered from short-term deployments, expanding applications for future study.

Improving early warning for toxins in commercial seafood is a main objective of HAB monitoring and seafood safety. Routine, in situ monitoring of phytoplankton and toxin plays a vital role in HAB bulletins that provide information about near real-time risks (26, 27). Public health departments and coastal agencies use this information to make decisions about staffing and sampling priorities, often with only days to a week of advanced notice. In weekly observations from Monterey Bay, CA, Pseudo-nitzschia abundance and DA in mDA were strongly correlated in real time, illustrating the difficulty of using cell counts alone to predict toxins in shellfish or vice-versa. CA-lipids from SPATT provided 6 to 7 wk of prediction time for HABs, extending the lead-time of DA-SP207 and other modeling attempts by 1 to 1.5 mo (28–32). CA-lipid-HP20 models also exhibited a lower false negative rate for predicted Pseudo-nitzschia and the more stringent response variable for mDA, than DA-SP207 models. False negatives in shellfish samples are the most harmful outcome for seafood safety because true occurrences of contamination can be missed. CA-lipids from HP20 SPATT deployed 6 wk prior accurately alerted 22% of mussel contaminations above 0.1 µg DA g mussel⁻¹, compared to 14% of contaminations using DA from SP207 deployed 1 wk prior. While neither predictor convincingly modeled the response variable alone, results suggest routine CA-lipid monitoring may benefit forecast models, especially for systems that already conduct routine shore-based or ship-board passive sampling.

Previous studies have proposed diverse mechanistic roles for DA including trace metal acquisition and detoxification (33), microbe-host signaling (34), and grazer deterrence (19). From a top-down perspective, CA-lipids induce DA in multiple species of Pseudo-nitzschia isolated from different environments (15, 20, 22, 35). Recent experimental work has also demonstrated that toxin content is positively correlated to copepod rejection rate, and moreover, Pseudo-nitzschia spp. induced by CA-lipids had significantly higher rejection rates than the noninduced control (20). We measured approximately 0.03 pg DA cell⁻¹ in P. multiseries after induction by CA-lipids. This was ten times greater than the maximum DA content of P. multiseries isolated prior to an extreme bloom event in Monterey Bay (36). Particularly when grazer abundance is high during February and March, grazer-induced toxin may be a nonnegligible fraction of total DA in the California Current system.

Natural populations of Pseudo-nitzschia respond to multiple, interacting toxin drivers including pH, temperature, and nutrient limitation (37, 38). Species-specific microbiome associations may also play a role in regulating DA (34), and grazer-induced DA is variable across species (20) and copepodamide structure (39). Even with this variability, large-scale processes including ocean warming, coastal upwelling, and the El Niño Southern Oscillation have been linked to toxin events caused by mixed Pseudo-nitzschia assemblages on the United States west coast (3, 40, 41), and environmental drivers have been integral to building forecast models (26, 27). The best CA-lipid-HP20 models performed in similar ranges (AUC 0.7 to 0.8) to published bottom-up models from the California Current (28, 32), which supports consideration of grazer cues in future modeling efforts.

CA-lipids play an important role in structuring plankton communities, yet ambient concentrations are challenging to observe because they are dilute and highly labile in seawater. The short half-life of CA-lipids contrasts with DA, which can persist for months in the dissolved form (6, 21). Thus, CA-lipid measurements from passive samplers likely reflect local production, while DA measurements include information about circulation dynamics (42). In field experiments from the Northeast Atlantic, CA-lipid profiles were largely consistent across SPE and SPATT methods, suggesting shortened (6 to 24 h) deployments provide comparable information about CA-lipid composition to discrete measurements regardless of the different sorbents used. This is advantageous because monitoring discrete CA-lipids requires seawater to be pumped across the SPE cartridge in the field to prevent degradation, whereas SPATT resins are simply deployed and retrieved at the field site. In addition, passive sampling over a 24-h period can integrate local changes in CA-lipid concentration due to zooplankton vertical migration and patchiness. Field experiments from the Northeast Atlantic provided a foundation for future studies to consider CA-lipid analysis from passive samplers from biologging (43), autonomous vehicles, and research cruises.

Chemical signaling is key to structuring marine microbial communities (23), but the wealth of information contained in dissolved organic matter is rarely considered in ecological monitoring. In addition, improving prediction capacity for HABs would be an enormous advantage to shellfish aquaculture, regulatory agencies, and marine conservation efforts. Passive sampling of CA-lipids shows promise for practical applications in HAB monitoring. Results from generalized linear models gave management-relevant lead times amenable to use in early warning systems, and the method can be readily applied to future studies in plankton ecology and food web dynamics.

Materials and Methods

CA-lipid Induction and Field Experiments.

In culture experiments conducted at the University of California Santa Cruz, we tested the effect of three CA-lipid concentrations (0, 1, and 10 nM) on toxin production, growth, and chain length of toxigenic P. multiseries (Strain EBL 27, Monterey, CA). The induction method was adapted from ref. 22. To test the suitability of measuring CA-lipids with SPATT over shorter deployment times and assess SPATT in relation to discrete water samples by SPE, a 24-h field experiment was performed at the Kristineberg Center for Marine Research and Innovation (58.25°N, 11.45°E, Fig. 1 B and D). Experimental values are reported as mean ± SE unless otherwise noted. Additional details and statistical analyses for both experiments are contained in SI Appendix.

CA-lipid and Copepod Field Observations.

SPATT were deployed to measure dissolved CA-lipids over 28 mo from February 2023 to June 2025 at the Santa Cruz Municipal Wharf (36.96°N, 122.02°W, Fig. 1 B and C). Methodology for CA-lipid analysis from SPATT was previously published in ref. 21, and additional details are contained in SI Appendix. Briefly, eight intact CA-lipid structures were quantified by reversed phase liquid chromatography tandem mass spectrometry (LC–MS/MS, SI Appendix, Table S1). Lyso-lipids were reported as relative concentrations with max-min scaling. Copepods were also collected in weekly vertical net tows from January 2024 to January 2025 at the SCW. Additional sampling details and statistical analyses are contained in SI Appendix.

Harmful Algal Bloom Field Observations.

Additional field data from the Santa Cruz Municipal Wharf including mDA, DA from three different SPATT resins (HP20, SP207, SP700), Pseudo-nitzschia abundance, and particulate DA were retrieved over the same period as CA-lipids from https://erddap.sccoos.org/erddap/tabledap/HABs-SantaCruzWharf.subset. DA measurements were obtained using LCMS as described in ref. 10. SPATT analyzed for DA contained 3 g of resin and were deployed for 7 d in the same conditions as the HP20 SPATT analyzed for CA-lipids. Pseudo-nitzschia australis and P. multiseries were enumerated with species specific large subunit rRNA probes using methods from ref. 44 and combined to form a “seriata size class”, which accounts for the two main DA-producing species in the region (29, 36). The enumeration method is comparable to traditional microscopy and image classification algorithms (45). Differences in data distributions of three SPATT resins were assessed with the Kruskal–Wallis test (α = 0.05).

Generalized Linear Models.

mDA and Pseudo-nitzschia abundance were chosen as response variables. Predictor variables were DA concentration from SP207 SPATT (DA-SP207) and CA-lipid concentration from HP20 SPATT (CA-lipid-HP20). Resin SP207 was selected because the adsorbance pattern for DA is historically the most correlated to mDA at this site. The datasets were aligned by sample week, and lags of 1 to 8 wk were created for the predictor variables. Gaps of length 1 or 2 wk were filled by linear interpolation. Gaps larger than 2 wk were removed from all variables. The resulting datasets spanned 2 y from April 5, 2023, to May 28, 2025, and included 84 and 76 datapoints for Pseudo-nitzschia abundance and mDA, respectively. Spearman correlations were used to compare lagged predictors and response variables. Response variables were binned into binary variables prior to model calculations.

Binary response variables for Pseudo-nitzschia abundance and mDA are referred to as PN` and mDA`, respectively. PN` used a threshold of presence/absence where PN` = 1 for cell counts greater than zero. We tested two thresholds for mDA`; 1) the median concentration (0.025 μg DA g mussel^-1) and 2) the 3rd quartile concentration (0.100 μg DA g mussel⁻¹). Generalized linear models (family = binomial, ln(predictor + 1)) were created using the lme4 package in R (46). Top predictor variables were selected by AIC, where lower AIC signifies better model fit. The AUC ± 95% CI for the receiver operator characteristic (ROC) was used to score model performance, where AUC = 0.5 signifies chance and AUC = 1.0 signifies perfect classification. The F₁ score was used to evaluate trade-offs in precision and recall. F₁ scores approaching 1.0 signify the model is correctly predicting positive cases and catching a high percentage of true positives. Model statistics were calculated with the pROC and yardstick packages in R (47–49).

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (CSV)

Dataset S02 (CSV)

Dataset S03 (CSV)

Dataset S04 (CSV)

Dataset S05 (CSV)

Data, Materials, and Software Availability

Software have been deposited in Analysis Code S1 (https://doi.org/10.5281/zenodo.21142450) (50). All other data are included in the manuscript and/or supporting information.