Sampling took place in the English Channel which is a shallow epicontinental sea, bordering with the Celtic Sea in the west and the North Sea in the east (Dauvin 2012). Environmental, planktonic and fish samples were collected during the Channel Ground Fish Survey (CGFS, Giraldo et al. 2021, https://doi.org/10.17600/18001250) in 2021 on board of the R/V Thalassa in autumn (mid-September to mid-October).

Sardine (Sardina pilchardus)

• Fish samples were collected using a GOV (Grande Ouverture Verticale) bottom trawl. Immediately after trawling, all sardines were sorted, identified, and weighed. A subsample of minimum 5 sardine individuals was randomly selected at each station (total n = 249) for subsequent lipid analysis. For those individuals, a small piece of dorsal muscle tissue was dissected. Sex, length and weight were recorded. Length and weight were measured to the nearest 0.1 cm and nearest 0 g, respectively. Spawning activity was determined from sexual maturity of gametes with premature and mature gametes recorded as spawning individuals and immature and post-spawning gametes recorded as non-spawning individuals due to the visual similarity of the latter two stages.

Salinity, depth, temperature, zooplankton, phytoplankton

• Full oceanographic profiles of the water column were conducted at each station using a conductivity, temperature, and depth (CTD) probe (total n = 122). Temperaure was measured in °C, depth in meters and salinity in PSU. Niskin bottles were used to characterize taxonomic composition of phytoplankton (n = 11). Mesozooplankton was collected from vertical hauls using a WPII net (200 µm mesh size). Samples serving for lipid analysis were pre-filtered on a nylon mesh of 500 and 1000 µm to remove small zooplankton and phytoplankton colonies and zooplankton > 1000 µm. The remaining fraction (500 – 1000 µm) was filtered through pre-combusted (450°C for 4h) GF/F filters (n = 39). Samples serving for taxonomic determination (n = 11) were stored in a formalin solution while samples used for lipid analysis were immediately frozen at -80°C.

Phytoplankton taxonomic composition

• FlowCam. Samples were digitized using an 8-bit grayscale Benchtop B2 Series FlowCam® Model VS-IV (Yokogawa Fluid Imaging Technologies, Inc., Scarborough, ME, USA). This is an image acquisition system building on optical microscopy. It can generate high-resolution images of particles in the flow, in the size range 2 μm-1000 μm (depending on the magnification/flow cell depth combination). For our study, a 4X objective (40X overall magnification) coupled with a 300 μm-depth flow-cell was used and samples were run in “AutoImage” operation mode. In this way, as described by Zarauz et al. (2007), all particles in the field-of-view of the camera (phytoplankton, zooplankton and inorganic particles) are captured and imaged at a regular user-defined interval, which allows an accurate estimation of imaged volume and consequently of particle concentration.
• A database representative of each plankton community met in the English Channel and North Sea, was built using samples taken throughout 2013 (in the frame of the IFREMER Phytoplankton and Phycotoxins monitoring network, REPHY (REPHY-French Observation And Monitoring Program For Phytoplankton And Hydrology In Coastal Waters 2023) and the SRN (Suivi Régional de Nutriments) monitoring programme (Lefebvre and Devreker 2023)). A total of 31700 images were manually classified in 20 plankton groups. Moreover, instead of manually removing inorganic particles and artefacts as is commonly done (Zarauz et al. 2007), we added 6 groups for floating dark and light dead particles, bubbles, fibers, etc. to the 20 plankton groups in order to automatically identify and then eliminate them from the statistics.
• EcoTransLearn. EcoTransLearn is an R-package that facilitates the use of Transfer Learning methods to automatically classify digital images for ecological studies (Wacquet and Lefebvre 2022). This tool includes some Convolutional Neural Networks models, widely used in Deep Learning, for many applications, and in particular for image recognition.
• All models were pre-trained on the ImageNet dataset, which is an image database, organized according to the WordNet hierarchy, and composed of approximately 1.4 M images sorted into 1000 classes. EcoTransLearn uses the Python Deep Learning toolbox Keras for both model construction step and automated classification process thanks to the R-package reticulate. For our study, the model used is the VGG16 adapted with the plankton dataset built from images acquired with the FlowCam (Wacquet and Lefebvre 2022).
• The dataset is given in cells per liter

Zooplankton taxonomic composition

Zooplankton samples scanned by a ZooScan (Gorsky et al. 2010) were subsequently treated by the machine-learning web application Eco-Taxa (http://ecotaxoserver.obs-vlfr.fr) for species determination and validated by zooplankton taxonomists. Since the ZooScan is able to detect organisms with an equivalent circular diameter of at least 300 µm, only the mesozooplankton size fraction >300 µm was retained in this study. Species composition was determined for three size classes: 300-500 µm, 500-1000 µm, >1000 µm. Zooplankton abundance is given as individuals per m³.

Lipid analysis

• Lipid extraction. Dorsal muscles of sardine stored at -80°C were ground with a Retsch homogenizer into liquid nitrogen. Approximately 250-300 mg of the homogenized powder was placed in pre-combusted glass vials to which we added 6 ml of solvent mixture (CHCl3:MeOH, 2:1, v:v). As described in Mathieu-Resuge et al. (2023), lipid extracts were flushed under nitrogen gas, sonicated, vortexed and rested few hours to ensure complete lipid extraction.

         Approximately 250-300 mg of zooplankton were scrapped/recovered from filters stored at -80°C, suspended in 6 ml of solvent mixture (CHCl3:MeOH, 2:1, v:v) and processed as above.

• HPTLC analysis of reserve lipids. Neutral lipids (mostly reserve lipids) were analysed by high-performance thin-layer chromatography (HPTLC) on HPTLC glass plates (20×10mm) pre-coated with silica gel 60 from Merck (Darmstadt, Germany) and quantified by using a scanning densitometer (Automatic TLC Sampler 4 and TLC Scanner 3 respectively, CAMAG, Switzerland) as previously described in (Haberkorn et al. 2010). Briefly, a preliminary run was carried out to remove possible impurities using hexane:diethyl ether (1:1), and the plate was activated for 30 min at 120 °C. Lipid samples (4 µl) were spotted on the plates by the CAMAG automatic sampler. The neutral lipids were separated using a double development with hexane:diethyl ether:acetic acid (20:5:0.5) as first solvent system followed with hexane:diethyl ether (93:3) as a second solvent system. Lipid classes appeared as black bands after dipping plates in a cupric sulfate (3%), phosphoric acid solution (8%) and heating for 20 min at 180 °C (charring). Seven neutral lipid classes (categorized as storage lipids: free fatty acids, sterol esters, glyceride ethers, monoacylglycerol, diacylglycerol, wax ester, and triacylglycerol; considered as structural lipids: sterols) were identified based upon standards (Sigma–Aldrich, France) and coloring techniques. The charred plates were read by scanning at 370 nm, and black bands were quantified by Visiocats software. Results were expressed as mg of each identified neutral lipid class per g of muscle/zooplankton wet weight.
• Preparation of Fatty acid methyl esters (FAME). A 500µL aliquot of sardine total lipid extract was, after adding C23:0 (2.3µg) as internal standard (free fatty acid form), evaporated and hydrolysed in 1 mL of KOH-MeOH (0.5 M) for 30 min at 80 °C, and then transesterified with 1.6 mL of MeOH:H2SO4 (3.4 %; v/v) for 10 min at 100 °C to form fatty acid methyl ester (FAME).  We extracted FAME by adding 800 μL of hexane and 1.5 mL of hexane-saturated distilled water, and by shaking and centrifuging 10 min at 1,000 rpm. The denser aqueous phase was discarded and this step was repeated twice by adding only 1.5 mL of hexane-saturated distilled water. Finally, the samples were placed in the freezer at -20°C without removing the aqueous phase. After several hours, we quickly transferred the unfrozen upper organic phase into 2 mL vials, which are flushed with N2 and stored in a refrigerator until GC-FID analysis.
• As zooplankton lipids may contain high amount of wax ester, FAME preparation was performed differently to eliminate unsaponifiable compounds such as alcohol released from wax esters. One mL of zooplankton lipid extract was evaporated under N2 flux. Fatty acids were saponified with 1 mL of KOH-MeOH (0.5 M) for 3 min at 90°C. After cooling, 0.5 mL of water and 0.4 mL of ethanol were added and sample was vortexed. To eliminate unsaponifiable compounds 2 mL of hexane were added. Sample was vortexed again, centrifuged 10 min at 1000 rpm and upper phase was discarded. This step was repeated twice with 1mL hexane. To release saponified fatty acids, 0.5 mL of 6N HCl was added. After vortexing, 2 mL of hexane was added and sample centrifuged 10 min at 1000 rpm and upper phase was transferred in a 7mL vial. This step was repeated twice with 1mL hexane. Combined hexane upper phases with C23:0 (2.3µg) as internal standard was evaporated and then transesterified with 1.6 mL of MeOH:H2SO4 (3.4 %; v/v) for 10 min at 100 °C to form fatty acid methyl ester (FAME). Extraction of FAME was performed as above and stored in a refrigerator until analysis.
• GC-FID analysis of FAME. FAME composition of sardine muscle and zooplankton was determined using a gas chromatograph system (HP - Agilent 6890, Agilent Technologies, USA) equipped with a JW DB wax (30 m length x 0.25 mm i.d. x 0.25 μm film thickness), with an on-column injector at 60ºC and a FID detector at 300ºC. Peaks were identified by comparison with retention times of external known standards (Supelco 37 Component FAME Mix, PUFA No.1 and No.3, and Bacterial Acid Methyl Ester Mix from Sigma) using Chemstation software (Agilent). FA are reported using a shorthand notation of A:Bn-x, where A indicates the number of carbon atoms, B is the number of double bonds and x indicates the position of the first double bond relative to the terminal methyl group. FAME content was converted into fatty acid content based on 23:0 internal standard. Total FA content was calculated as the sum of all identified FA. Data are expressed as proportion of the total FA composition (%).