Spirulina (Arthrospira platensis) protein as a sustainable alternative fining agent for wine: extraction, optimization and comparative mechanistic evaluation of proanthocyanidin removal in model wine systems

Abstract

Spirulina (Arthrospira platensis) represents a promising sustainable alternative to animal-derived proteins for wine fining, though its protein extraction and proanthocyanidin (PA) removal efficiency remain poorly characterized. This study presents the first comprehensive evaluation of a purified Spirulina protein extract, designed to avoid the sensory drawbacks of whole-biomass application. A scalable bead-milling extraction process (pH 7.0, 1 M NaCl, followed by ethanol precipitation) yielded a high-purity protein isolate (63%protein) at a satisfactory recovery yield (25%). Fining (50 g/hL) in a white wine model removed 81% of PAs, comparable to patatin, pea protein, and egg albumin, but lower than gelatin (99.6%) and casein (94.0%). In a red wine model, Spirulina protein achieved higher PA removal (30%) than the other fining agents (3-19%). A matrix-dependent removal mechanism was observed, governed by protein structural flexibility in white wine and hydrogen-bond-acceptor capacity in red wine model. Spirulina protein is an effective, sustainable fining agent under model wine conditions.

Highlights

• •Spirulina proteins obtained by bead milling are scalable, sustainable fining agents.
• •Optimal extraction at pH 7.0 and 1 M NaCl yields a high-purity protein extract.
• •∼88% proanthocyanins removal in a white wine model, comparable to plant proteins.
• •5-10× higher proanthocyanins removal in a red wine model vs commercial fining agents.
• •OPLS shows matrix-dependent interaction mechanisms, structure vs charge.

1. Introduction

Proanthocyanidins (PAs) are grape-derived flavan-3-ol polymers composed primarily of (+)-catechin, (−)-epicatechin, (−)-epicatechin-3-O-gallate, and (−)-epigallocatechin subunits. They are major contributors to wine mouthfeel, significantly influencing astringency, bitterness, and color stability, thereby shaping consumer acceptance (González-Muñoz et al., 2022). The extraction and impact of PAs on wine quality depend critically on both winemaking practices and structural characteristics, including mean degree of polymerization (mDP), degree of galloylation, B-ring hydroxylation pattern, stereochemistry, and interflavanoid linkage type (Du Preez et al., 2022).

Astringency results from interactions between PAs and salivary proteins, particularly proline-rich proteins, and is associated with a drying and puckering sensation. At low tannin concentrations, soluble protein–PA complexes are formed, whereas at higher concentrations, these complexes become insoluble, reducing salivary lubrication and increasing oral friction (Kuang et al., 2025; Soares et al., 2020). Conversely, bitterness is primarily associated with monomeric and low-molecular-weight flavan-3-ols, often galloylated, that activate the TAS2R/T2R bitter taste receptors (Dias et al., 2023). PAs structural features strongly influence sensory perception: higher mDP and a higher degree of galloylation generally enhance astringency while reducing bitterness, whereas monomers and small oligomers are more strongly associated with bitterness (Soares et al., 2020). B-ring trihydroxylation, characteristic of epigallocatechin-containing units, markedly enhances protein-tannin interactions and consequently increases astringency, but also accelerates oxidative degradation, leading to the formation of brown polymers that compromise color stability, particularly in white wines (Weilack et al., 2021). Consequently, the structural features determinants of sensory quality of PAs are intrinsically linked to their oxidative stability, highlighting the need for fining strategies that can selectively modulate specific PA fractions while preserving desirable color and sensory characteristics of wine.

Protein-based fining agents (PFAs) constitute the main technological approach used to control excessive astringency and bitterness through the formation of insoluble complexes with PAs (Marangon et al., 2019; Obreque-Slier et al., 2023). Traditional PFAs, including gelatin, potassium caseinate, egg albumin, and isinglass, exhibit fining behaviour that depends on amino acid composition, secondary and tertiary structure, molecular weight, proline and aromatic amino acid content, and hydrophobicity (Marangon et al., 2019; Obreque-Slier et al., 2023). In contrast, plant-derived proteins (pea, potato, quinoa, kiwicha) typically present lower proline content, more compact structures, and abundant aromatic amino acid residues that create defined binding pockets favoring interactions with smaller, more compact polyphenols and enhancing color preservation (Marangon et al., 2019; Obreque-Slier et al., 2023; Río Segade et al., 2019). Nevertheless, wine matrix composition, including its phenolic profile, pH, ethanol content, and polysaccharide content, also strongly modulates fining efficacy, rendering PFA performance highly wine-dependent and necessitating preliminary fining trials and tailored application to avoid sensory losses (Dias et al., 2023; Marangon et al., 2019; Obreque-Slier et al., 2023; Río Segade et al., 2019).

Regulatory, health, and sustainability concerns increasingly challenge the use of animal-derived PFAs. Egg white (ovalbumin) and milk proteins are well-documented IgE-mediated allergens whose epitopes may persist in finished wine, prompting European regulations that require analytical verification of residual levels (with typical detection and quantification limits of approximately 0.25 and 0.5 ppm, respectively) and mandatory allergen labeling (Organisation Internationale de la Vigne et du Vin (OIV), 2012). Although clinical reactions are rare and usually mild, the perceived risk of “hidden allergens” has driven demand for allergen-free and vegan-friendly wines. At the same time, gelatin production is energy- and reagent-intensive, casein is linked to resource-intensive dairy production systems, and isinglass relies on fish by-products, raising environmental and ethical concerns (Fernandes et al., 2015). In contrast, plant-derived proteins, including those recovered from grape or yeast by-products, as well as other plant-based fining agents, enable the valorization of winemaking by-products (grape seeds, lees), align with circular economy objectives, and are not classified as allergens under current European regulations, thereby avoiding costly analytical controls and mandatory allergen labeling requirements (Fernandes et al., 2015; Marangon et al., 2019). Recent evidence demonstrates that plant proteins (pea, potato) and grape seed proteins effectively remove tannins in vegan-friendly wines while providing superior color preservation, whereas lees-derived protein extracts provide sustainable fining with minimal impact on mouthfeel (Río Segade et al., 2019). Thus, there is continued interest in novel, sustainable PFAs whose extraction, composition, and matrix-dependent fining performance are better understood.

Recent studies have considerably expanded the portfolio of sustainable protein-based fining agents. Plant-derived proteins obtained from amaranth, quinoa, soybean, grape seeds, lentils, rice, and maize have emerged as promising alternatives to conventional animal proteins owing to their high protein content, low allergenic potential, sustainable production, and effective fining performance (Ficagna et al., 2019; Gordillo et al., 2021; Ma et al., 2020; Pino-Ramos et al., 2024; Tschiersch et al., 2010). However, despite their high protein content and favourable sustainability profile, microalgae remain largely unexplored as a source of protein-based wine fining agents. To date, their application has been essentially limited to the evaluation of whole Spirulina biomass rather than isolated protein fractions (Noriega-Domínguez et al., 2010). Within this context, Arthrospira platensis (Spirulina), an oxygenic cyanobacterium, has emerged as a promising protein source for alternative PFAs. Its biomass contains approximately 60-70% protein (dry weight basis), dominated by water-soluble phycobiliproteins such as phycocyanin and allophycocyanin (Amorim et al., 2021; Mohammadi, Soltanzadeh, Ebrahimi, and Hamishehkar, 2022; Khalid et al., 2025; Yaishana et al., 2025). Therefore, efficient, scalable extraction and purification strategies are needed to obtain protein fractions with high yield and purity suitable for industrial application (Mohammadi et al., 2022; Rodríguez-Couto, 2022). Accordingly, this study focused on mechanical cell disruption methods, namely ultrasound and bead milling, combined with the systematic optimization of key extraction parameters. This strategy was selected because it provides a practical balance between efficient protein recovery, operational simplicity, compatibility with food-grade processing, and industrial scalability, while enabling the production of high-yield, high-purity protein extracts suitable for wine fining. Spirulina biomass is Generally Recognized as Safe (GRAS) by the U.S. Food and Drug Administration (FDA), the unprocessed biomass is widely used as a dietary supplement and natural colorant. However, for wine fining applications, high-purity protein isolates are preferred to ensure a well-defined fining agent, thereby enabling a more reliable assessment of its allergenic potential (Bianco et al., 2023) and to minimize undesirable sensory effects associated with crude biomass, which has been reported to impart unusual taste and aroma attributes to wine (Noriega-Domínguez et al., 2010). This study aimed to develop a simple, economical, and scalable process for producing a high-purity Spirulina protein extract for use in wine fining, avoiding the sensory drawbacks previously associated with whole Spirulina biomass, and to compare its fining performance with that of commercial animal-based (gelatin, ovalbumin, casein) and plant-based (patatin, pea protein) fining agents in white and red wine model systems. More specifically, our objectives were to (i) select the simplest, scalable, and most efficient cell wall disruption method described in the literature, (ii) optimize the selected extraction process for Spirulina protein yield and purity, (iii) perform a detailed physicochemical characterization of the resulting protein fraction, (iv) compare the fining efficiency of the Spirulina protein extract with commonly used PFAs, and (v) relate the physicochemical composition of PFAs to proanthocyanidin removal in contrasting wine model matrices, in order to rationalize differences in PA removal efficiency among protein fining agents and support the rational design of targeted, sustainable fining strategies. Building upon our previous work on the optimization of Spirulina protein extraction (Costa et al., 2024), this study is, to the best of our knowledge, the first to evaluate a purified and thoroughly characterized Spirulina protein extract as a wine fining agent and to compare its PA removal efficiency with that of commercially available PFAs in both white and red wine model systems.

2. Materials and methods

2.1. Chemicals and Spirulina biomass

Spirulina powder was supplied by Allmicroalgae Company (Portugal) and was produced using a flat-panel photobioreactor followed by raceway cultivation, biomass concentration, and solar-assisted drying in a controlled-air facility. The final powder contained 63-67% protein, as determined by the Kjeldahl method (data provided by the supplier). L-amino acid standards were acquired from Sigma-Aldrich/Merck (St. Louis, MO, USA), including glutamic acid (Glu, CAS 56-86-0), aspartic acid (Asp, CAS 56-84-8), serine (Ser, CAS 56-45-1), glycine (Gly, CAS 56-40-6), threonine (Thr, CAS 72-19-5), arginine (Arg, CAS 74-79-3), β-alanine (β-Ala, CAS 107-95-9), alanine (Ala, CAS 56-41-7), tyrosine (Tyr, CAS 60-18-4), valine (Val, CAS 72-18-4), phenylalanine (Phe, CAS 63-91-2), leucine (Leu, CAS 61-90-5), lysine (Lys, CAS 56-87-1) and proline (Pro, CAS 147-85-3). Histidine (His, CAS 71-00-1) was purchased from Thermo Scientific (Waltham, MA, USA), and isoleucine (Ile, CAS 73-32-5) from Tokyo Chemical Industry (TCI, Japan). 9-Fluorenylmethylchloroformate (FMOC, CAS 28920-43-6) was obtained from Acros Organics (Geel, Belgium), o-phthalaldehyde (OPA, CAS 643-79-8) from Sigma-Aldrich, and 2-mercaptoethanol (MCE, CAS 60-24-2) from VWR (Radnor, PA, USA). Additional reagents included orthophosphoric acid (85% v/v, CAS 7664-38-2) from Pronalab (Lisbon, Portugal), sodium acetate (CAS 127-09-3) from Sigma-Aldrich, ammonium sulfate (CAS 7783-20-2) from Fluka (Buchs, Switzerland) and ethanol (96% v/v, CAS 64-17-5) from Aga (São João da Madeira, Portugal). Lysozyme (CAS 12650-88-3) was obtained from SAI Enology (Paredes, Portugal). Hydrochloric acid (37%, CAS 7647-01-0) was obtained from VWR, and Milli-Q ultrapure water was used throughout. Ethylenediaminetetraacetic acid (EDTA, CAS 60-00-4) was obtained from Chem-Lab NV (Zedelgem, Belgium). Electrophoresis mini-gels, standard protein kits, sample buffer and reducing agent were purchased from Invitrogen (Thermo Fisher Scientific, Waltham, MA, USA), and trypsin (CAS 9002-07-7) was sourced from Sigma-Aldrich.

2.2. Extraction of Spirulina proteins

Two mechanical protein extraction techniques were initially screened: bead milling and ultrasound-assisted extraction using an ultrasonic probe, with simple agitation serving as a control. Each extraction method was tested at three NaCl concentrations (0, 0.5, and 1 M) and three pH levels (4.50, 7.00, and 9.00), using a solid-to-liquid ratio of 1 g of Spirulina powder per 20 mL of extraction buffer (50 g/L; 0.1 M sodium acetate pH 4.50; 0.1 M sodium phosphate pH 7.00; or 0.1 M sodium tetraborate pH 9.00).

For bead milling, Spirulina was mixed with 50 g of 4 mm glass beads and stirred at room temperature for 20 h. The bead-milling duration (6, 12, 20, and 24 h) was subsequently optimized under the selected extraction conditions (pH 7.00 and 1 M NaCl). Ultrasound-assisted extraction was performed using a probe homogenizer (Homogenizer Optic Ivymen System, J.P. Selecta, Barcelona, Spain; > 1000 W/L) applied in three 20-s cycles. After cell disruption, the suspensions were centrifuged to recover the protein-rich supernatants. Proteins were then isolated by precipitation either by the addition of 75% (v/v) ethanol or by ammonium sulphate saturation, followed by dialysis (molecular weight cut-off of 12-14 kDa).

2.3. Optimization of bead milling extraction

Bead milling was selected as the simplest and most scalable method for Spirulina protein extraction and was therefore optimized using a replicated full factorial experimental design with two factors at three levels each: pH (4.50, 7.00, and 9.00) and NaCl concentration (0 M, 0.5 M, and 1 M) as independent variables. Protein content of the extracts and protein yield relative to Spirulina total protein content of the biomass were used as response variables and were calculated as follows:

$$\text{Protein content}\left(\%\mathrm{dw}\right)=\left(\mathrm{\Sigma a}.\mathrm{a}.{}_{\text{HPLC}}/{\mathrm{m}}_{\text{sample}}\right)\mathrm{x}100$$ (1)

$$\text{Protein yield}\left(\%\mathrm{dw}\right)=\left[\text{Protein content of extract}\left(\%\mathrm{dw}\right)\mathrm{x}{\mathrm{m}}_{\text{extract}\left(\mathrm{g}\mathrm{dry}\right)}\right]/\text{Protein content of biomass}\left(\%\mathrm{dw}\right)$$ (2)

where Σa.a._HPLC represents the sum of the amino acids determined by HPLC after acid hydrolysis (Section 2.4).

The experimental data were fitted by multiple regression to a second-order polynomial model using coded factors (Eq. 3) in STATISTICA12 software.

$$\mathrm{Y}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{\mathrm{X}}_1+{\mathrm{\beta }}_2{\mathrm{X}}_2+{\mathrm{\beta }}_3{{\mathrm{X}}_1}^2+{\mathrm{\beta }}_4{{\mathrm{X}}_2}^2+{\mathrm{\beta }}_5{\mathrm{X}}^1{\mathrm{X}}^2$$ (3)

where β_n are the regression equation coefficients and X_n the coded factors. Surface plots were generated from the fitted models.

Model adequacy was first assessed by residual diagnostics, which revealed heteroscedasticity and deviation from normality. To stabilize the variance, a Box-Cox power transformation was applied, giving λ = −1.83 for protein yield and λ = −0.17 for protein content. Analysis of variance (ANOVA), using the pure error mean square to test for lack-of-fit and factor effects, was used to assess model significance and the influence of each factor and their interaction. Because the protein content of the extracts, as determined by Kjeldahl analysis, constitutes an OIV requirement for wine protein-based fining agents, albeit with minimum thresholds that vary according to protein source (at least 5.7% total nitrogen for yeast protein, 10% for patatin and pea protein, 12% for egg albumin, 13% for casein, and 14% for gelatin), the lower acceptable limit for protein content in the present study was set 65% (Kjeldahl method), corresponding to the threshold established for plant-derived protein fining agents.

2.4. Protein, C-phycocyanin, and allophycocyanin analysis

Protein contents of the Spirulina biomass, protein extracts, and commercial PFAs were determined by the Kjeldahl method using a nitrogen-to-protein conversion factor of 6.29 for Spirulina (Safi et al., 2013), and 6.25 for commercial PFAs as a general conversion factor, recognizing that specific conversion factors vary among protein sources. Protein content was also independently estimated by summing the amino acids released after acid hydrolysis, following FAO guidelines for true protein estimation.

Amino acid composition was determined by RP-HPLC-UV (Thermo Fisher Scientific, Waltham, MA, USA) after derivatization with o-phthaldialdehyde/2-mercaptoethanol and 9-fluorenylmethyl chloroformate. To correct for amino acid degradation and incomplete release during 24 h acid hydrolysis, individual amino acid concentrations were multiplied by amino acid-specific correction factors. C-phycocyanin and allophycocyanin contents in selected extracts were quantified by UV-Vis spectroscopy (Thermo Fisher Scientific, Waltham, MA, USA) by measuring absorbance at 620 nm and 652 nm, respectively, using protein solutions at a concentration of 1 mg/mL.

2.5. FT-IR spectroscopy

Infrared spectra were recorded using a Shimadzu IRAffinity-1 FT-IR spectrometer equipped with an ATR accessory (Shimadzu, Kyoto, Japan). For each solid extract or 15 g/100 mL protein solution, 256 scans were averaged at 2 cm⁻¹ resolution, using water as background and subtracting its spectrum. Protein secondary structure was quantified from the amide I region by curve fitting assisted by second-derivative and deconvolution analysis, following established FT-IR methods. Peak assignments to secondary-structure elements were made accordingly, and fits were accepted only when the reconstructed spectra and second derivatives matched the experimental data within noise.

2.6. Thermogravimetric analysis

Thermogravimetric analysis (TGA) was conducted to assess the thermal stability of Spirulina protein extracts and commercial PFAs, and to study their decomposition behaviour as a function of temperature. Analyses were performed using a NETZSCH STA 449 F3 thermal analyser (Netzsch-Gerätebau GmbH, Selb, Germany). In a typical experiment, samples were heated from 30 to 900 °C at a heating rate of 10 °C min⁻¹ in alumina (Al₂O₃) crucibles, under a nitrogen (N₂) atmosphere (50 mL min⁻¹ purge flow; 20 mL min⁻¹ protective flow).

2.7. Protein characterization by SDS-PAGE electrophoresis

The protein extract obtained using the selected extraction method, as well as the commercial PFAs were diluted (1:22.5) in 1% (w/v) SDS and analyzed by sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE).

2.8. Proteomic analysis of Spirulina protein extract

Proteins present in the Spirulina protein extract were identified by proteomic analysis using an in-solution digestion protocol adapted from Albuquerque et al. (2023), as well as in-gel digestion of Coomassie-stained SDS-PAGE bands. Tryptic digests (10 μL) were analyzed by LC-MS/MS on a Thermo Fisher Scientific LTQ XL system (Waltham, MA, USA) coupled to a C18 column (Accucore Vanquish C18, 150 × 2.1 mm, 1.5 μm particle size, Thermo Fisher Scientific, Waltham, MA, USA).

MS/MS peak lists were searched with OMSSA version 2.1.9 via SearchGUI version 4.3.12, against the Arthrospira platensis UniPRotKB target/decoy database (1236 protein sequences, downloaded on March 10, 2025), using trypsin specificity (<2 missed cleavages), 20 ppm as MS¹ and 0.5 Da for MS² tolerances. Variable modifications included ammonia loss from N (−17.026549 Da), deamidation of N (+0.984016 Da), deamidation of Q (+0.984016 Da), oxidation of M (+15.994915 Da), pyroglutamate formation from E (−18.010565 Da), and pyroglutamate formation from Q (−17.026549 Da).

Peptide and protein identifications were validated using PeptideShaker version 3.0.3 (Vaudel et al., 2015), at a 1% false discovery rate (FDR), and relative protein abundances were estimated using the Normalized Spectral Abundance Factor (NSAF) approach. Raw proteomics files are available at https://doi.org/10.5281/zenodo.14999343.

2.9. Model wine solutions

The efficiency of the Spirulina protein extract and its comparison with commercial PFAs were assessed in synthetic model wine solutions. The compositions of the white wine model (WWM) and red wine model (RWM) were selected to reproduce representative physicochemical conditions and proanthocyanidin concentrations typical of the respective wine types while minimizing matrix variability, thereby enabling a controlled comparison of fining performance across different PFAs. The WWM contained 12.0% (v/v) ethanol, 3.5 g/L tartaric acid (adjusted to pH 3.20 with NaOH), 500 mg/L KCl, and approximately 19 mg/L total PAs (approximately 10 mg/L each from Vitis vinifera L. seed and skin extracts). The RWM contained 15.0% (v/v) ethanol, 3.5 g/L tartaric acid (adjusted to pH 3.50 with NaOH), 1500 mg/L KCl, and 1123 mg/L PAs from Vitis vinifera L. seed and skin extracts. Detailed composition is of both model wine solutions is given in Supplementary Material Table S.1.

2.10. Commercial protein fining agents and fining trials

For the fining trials, the selected Spirulina protein extract (AP) was compared with commonly used commercial PFAs: gelatin (GEL), casein (CAS), egg albumin (ALB), and pea protein (PEA) and patatin (PAT). All protein fining agents were tested in the red and white wine model solutions at a dose of 50 g/hL, corresponding to the maximum OIV-approved dose for plant-based protein fining agents. This standardized dosage was selected to enable a direct comparison among all fining agents under industrially relevant conditions, following OIV recommendations, although this approach does not account for differences in protein purity and solubility among the individual fining agents. An unfined control wine was included and subjected to the same handling procedures (Organisation Internationale de la Vigne et du Vin (OIV), 2022).

2.11. Quantification and structural analysis of PAs

PAs were quantified and structurally characterized after solid-phase extraction (SPE), which fractionated the sample into monomeric+oligomeric and polymeric fractions. The oligomeric and polymeric fractions were subsequently depolymerized with thioglycolic acid and analyzed by LC-MSⁿ (Thermo Fisher Scientific, Waltham, MA, USA). Sample preparation consisted of diluting 1 mL of the white wine model or 0.25 mL of the red wine model with 10 mL of phosphate buffer (pH 7.00), to which 25 mg of ascorbic acid was added.

2.12. Average physicochemical attributes, frontier orbital energies, and hydrogen-bond acceptor and donor constants

Average physicochemical attributes of the protein amino acids were described using 66 AAIndex physicochemical indices (e.g., hydrophobicity, charge, polarity), computed as amino acid composition (mol%)-weighted averages to yield potential numerical functional descriptors (Roy et al., 2009; Supplementary Material Table S.2), complemented by 13 experimentally derived variables [zeta potential (Anton Paar, Graz, Austria) and FT-IR-derived protein secondary-structure abundances; Supplementary Material Table S.2].

The hydrogen-bonding (HB) potential of protein amino acid side chains in hydroalcoholic media was captured by effective donor (α_eff) and acceptor (β_eff) constants derived from Hunter's chemical scale for 12% and 15% (v/v) ethanol-water mixtures, and by relative HB-donor/acceptor strengths derived from quantum-mechanical interaction energies and normalized into dimensionless scales. Total, polar, and nonpolar solvent accessible surface areas were also calculated to complete the descriptor set used to relate protein physicochemical properties to PA fining efficiency.

2.13. Statistical analysis

Data are reported as mean ± standard deviation. Statistical analysis of the extraction results was performed using STATISTICA 12 software (StatSoft, Tulsa, OK, USA). ζ-potential data were evaluated by one-way analysis of variance (ANOVA) followed by Tukey's HSD post hoc test. Orthogonal partial least squares (OPLS) analysis (Ferreira et al., 2024) was performed using SIMCA version 14 (Sartorius, Göttingen, Germany) to relate the physicochemical properties of PFAs (X variables) to PA removal efficiency in WWM and RWM (Y variables). Model quality was assessed using the goodness-of-fit parameters (R²X and R²Y) and predictive ability (Q²), estimated by seven-fold cross-validation. Model robustness was further evaluated by permutation testing (999 permutations) to assess the risk of overfitting. Hierarchical cluster analysis (HCA) was performed on the predictive component scores obtained from the OPLS models, using Euclidean distance as the dissimilarity metric and Ward's minimum-variance method for agglomerative clustering. Cluster robustness was assessed using the cophenetic correlation coefficient (CCC) and cluster separation metrics.

3. Results and discussion

3.1. Screening of protein extraction/isolation methods

Among the most frequently used microalgal cell-disruption techniques, the two most promising in terms of scalability, extraction yield, and sustainability were selected (Li et al., 2025), namely bead milling (Safi et al., 2014) and ultrasound-assisted extraction (Vernès et al., 2019). Protein solubility and stability are strongly influenced by pH and ionic strength, with NaCl shown to stabilize phycobiliproteins (Jaeschke et al., 2019; Silva et al., 2022; Teuling et al., 2017). Therefore, these methods were initially tested at pH 7.00 and pH 9.00 (Supplementary Material Table S.3). Simple agitation was used as a control to assess the amount of protein released by diffusion into the aqueous medium through membranes and cell walls that have already been partially disrupted during solar drying and osmotic shock.

The Spirulina biomass contained 54.6 ± 1.7% protein (dry weight, dw) as determined by the Kjeldahl method (nitrogen-to-protein conversion factor of 6.29), and 55.7 ± 4.4% protein (dw) as determined by the sum of total amino acids. These values were in good agreement and consistent with those reported in the literature, despite variations associated with methodological differences, nitrogen-to-protein conversion factors, cultivation conditions, and strain-dependent effects (Teuling et al., 2017). In addition, recent studies have shown that biomass pretreatment and protein extraction strategies can influence protein recovery and quantification, further contributing to the differences reported between studies (Chaudhary et al., 2025). For example, Safi et al. (2014) reported 53.5 ± 1.1% protein (dw; Kjeldahl, convertion factor 6.27), Milia et al. (2025) reported 62.7% (dw; Kjeldahl, convertion factor 6.25), and Oliveira et al. (1999) reported 59.4-71.5% (dw; Kjeldahl, convertion factor 6.25), decreasing with increasing growth temperature, while Becker (2007) summarized protein values ranging from 46 to 63%.

Sun-drying and osmotic shock during biomass resuspension in distilled water caused only limited cell wall disruption, resulting in low protein solubilization (Supplementary Material Table S.3). Ultrasound-assisted extraction and bead milling significantly increased Spirulina protein extraction, both yielding similar results (Supplementary Material Table S.3) and producing an extract with a protein content of 46.9% (dw) and an extraction yield of approximately 20% at pH 9.00. Both the protein yield and the extract protein content were significantly higher at pH 9.00 than at pH 7.00 (Supplementary Material Table S.3), consistent with the lower solubility of Arthrospira platensis proteins at pH 7.00 (Teuling et al., 2017).

The protein concentrations obtained at pH 7.00 by bead milling were lower than those reported by Jaeschke et al. (2019) (46.1 ± 4.4% protein without dialysis; conditions in Supplementary Material Table S.4) and also lower than those of Safi et al. (2014) (47.1 ± 0.9% by sonication without dialysis; Supplementary Material Table S.4). These lower yields can be explained by losses during dialysis (molecular-weight cut-off 12-14 kDa), which likely removed a substantial proportion of the low-molecular-weight proteins known to be extracted from Spirulina (Teuling et al., 2017). Using dialysis, Teuling et al. (2017) reported an extraction yield of 35.9 ± 9.2% and an extract protein content of 60.8 ± 1.0% with bead milling (conditions in Supplementary Material Table S.4), values closer to those obtained in the present work. Silva et al. (2022) obtained a protein yield of 40.0 ± 1.9% and an extract protein content of 43.6 ± 0.9% using bath ultrasound on previously lipid-extracted Spirulina (without dialysis; Supplementary Material Table S.4), whereas Lupatini et al. (2017) reported 50.8% protein in the initial biomass and 75.8% protein recovery after ultrasound probe treatment combined with mechanical agitation under alkaline conditions, following prior hexane lipid extraction (Supplementary Material Table S.4). Conversely, manothermosonication of Spirulina cells yielded a lower protein content of 28.4 ± 1.2% (dw) (Vernès et al., 2019; Supplementary Material Table S.4). Overall, the extraction yields and protein contents obtained during this screening fall within the range reported by previous studies using ultrasound, bead milling, and related cell disruption methods cited above, confirming that both extraction processes tested in this study are consistent with, and in some cases comparable to, the best-performing protocols described in the literature in terms of protein extraction yield.

Although isoelectric precipitation of Spirulina proteins has frequently been reported (Kamaruding et al., 2022; Safi et al., 2014), ethanol and ammonium sulfate precipitation are often preferred, as they offer greater selectivity and process control and are more robust, reproducible, and scalable, particularly for complex biological matrices (Amorim et al., 2021; Irvani et al., 2024). Therefore, after extraction, proteins were isolated either by ethanol precipitation (75% v/v) or by ammonium sulfate saturation (Supplementary Material Table S.5), both yielding good protein recoveries with simple and cost-effective procedures. For extracts obtained by ultrasound, no significant differences in protein content were observed between the two isolation methods, whereas for bead-milled extracts, a significantly higher protein yield was obtained using ethanol precipitation.

3.2. Optimization of protein extraction conditions by bead milling

Ultrasound-assisted extraction (UAE) is widely reported as an efficient method for Spirulina protein recovery, providing high yields with limited protein denaturation under optimized conditions (Burdějová et al., 2024; Khalid et al., 2026). In the present study, preliminary screening showed that ultrasound and bead milling produced comparable protein yields and extract protein contents. As no major differences were observed between the two methods, bead milling coupled with ethanol precipitation was selected and further optimized, owing to its similar extraction performance combined with greater operational simplicity, lower equipment and operating costs, and easier industrial scale-up (Li et al., 2025). A 3² full factorial design was applied, with pH and salt concentration evaluated at three levels each (Table 1), enabling response surface modeling of nonlinear effects and interactions related to net charge and ionic shielding on protein solubility (Bezerra et al., 2008). Protein extraction data (Supplementary Material Table S.6) supported a quadratic model describing the combined influence of pH and salt concentration (Fig. 1), which was statistically significant (ANOVA, p < 0.0001) and showed high explanatory power (R² = 0.914, R²adj = 0.894). Both factors, including their linear and quadratic terms, and their interaction significantly affected protein recovery (Supplementary Material Table S.6), although a significant lack-of-fit (p < 0.05) suggests that additional variables, such as biomass batch variability, may further improve the predictive accuracy of the model.

Table 1.

Protein content and protein yield of Spirulina extracts obtained by bead milling at different pH values and NaCl concentrations.

pH	NaCl (M)	Protein content (%)	Protein yield (%)
4.50	0	44.8 ± 1.2DE	9.2 ± 0.2A
4.50	0.5	36.1 ± 5.6BC	11.2 ± 1.7AB
4.50	1	32.2 ± 2.6B	10.4 ± 0.8AB
7.00	0	21.5 ± 0.4A	14.5 ± 0.2BC
7.00	0.5	40.8 ± 3.9CD	32.0 ± 3.1E
7.00	1	63.2 ± 2.1F	25.2 ± 0.8D
9.00	0	49.1 ± 0.5E	28.2 ± 0.3DE
9.00	0.5	21.8 ± 1.2A	17.4 ± 1.0C
9.00	1	23.8 ± 2.8 A	17.1 ± 2.0C

Protein content was determined by HPLC. Protein yield represents the percentage of protein recovered from the initial biomass. Values are mean ± standard deviation. Different letters within the same column indicate significant differences among treatments (Tukey HSD test, p < 0.05).

Fig. 1.

Response surfaces for Spirulina protein extraction yield as a function of pH and NaCl concentration. The fitted quadratic model is: z = − 0.0449 + 0.0868 × pH - 0.0927 × pH ² + 0.0130 × C_NaCl - 0.0274 × C_NaCl²-0.0357 × pH × C_NaCl.

Experimentally, the highest protein yield was obtained at pH 7.00 with 0.5 M NaCl (32.0 ± 3.1%), which was not significantly different from that obtained at pH 9.00 without salt (28.2 ± 0.3%). Salt addition increased protein recovery at pH 7.00 up to 0.5 M but reduced yields at higher ionic strength, as well as at all salt concentrations tested at pH 9.00. The model predicted a maximum extraction yield of approximately 35% at pH 8.00 and 0.46 M NaCl (95% confidence interval, 26% to 66%).

Previous studies have shown that Spirulina protein extractability is strongly controlled by pH and ionic strength, consistent with the trends observed in the present work. Increased NaCl concentration can enhance protein recovery through osmotic shock and improved cell disruption at high ionic strength (Hadiyanto & Adetya, 2018), but the present results indicate a strong pH dependence. At pH 7.00, moderate salt concentrations significantly improved protein recovery, in line with the higher protein solubility reported near neutral pH (Kamaruding et al., 2022) and with previous reports of synergistic pH-salt effects during bead milling (Fabre et al., 2022). In contrast, at pH 9.00, salt addition reduced extraction yield, and at pH 4.50 it had no significant effect, suggesting the occurrence of salting-out phenomena or limited protein solubility under suboptimal charge conditions. Although higher recoveries can be achieved using harsh alkaline extraction conditions (Bashir et al., 2016), the moderate yields obtained under milder conditions underline the need to optimize bead-milling parameters to balance yield, protein functionality, and process sustainability.

The quadratic model for extract protein content as a function of pH and salt concentration was statistically significant but showed limited predictive power (R² = 0.18; R²adj ≈ 0; Supplementary Material Table S.7 and Fig. S.1), in contrast to the strong model obtained for protein extraction yield. This reflects the fact that extract purity depends not only on protein recovery but also on the co-extraction of non-protein components, such as polysaccharides. Nonetheless, both pH and salt concentration significantly affected extract purity: at pH 7.00, increasing salt concentration up to 1 M improved the protein content of the extract (i.e., higher purity without a proportional increase in mass), whereas at pH 4.50 salt strongly decreased protein content, and at pH 9.00 it consistently lowered extract purity relative to salt-free condition. These results show that conditions maximizing protein yield do not necessarily optimize extract purity, reinforcing the need to jointly consider recovery and selectivity when designing Spirulina protein extraction processes, since the protein content of PFAs is regulated by OIV, and was set in this work equal to the minimum nitrogen content established for plant proteins (>10% total nitrogen, on a dry basis). For wine fining applications, the best compromise reflects a balance between sufficient protein recovery and adequate extract purity to ensure effective fining performance.

Although the model predicted a maximum protein extraction (approximately 35%) occurred at pH 8.00 and 0.46 M NaCl, this yield did not differ significantly from those obtained experimentally at pH 7.00 with 0.5 M NaCl or at pH 9.00 without salt. Because the predicted optimum did not differ significantly from the best-performing experimental conditions, no additional independent validation experiments were performed. Accordingly, pH 7.00 and 1 M NaCl were selected because this condition represents the best practical compromise between protein yield and extract purity. Higher salt levels favored protein-enriched extracts, reducing the need for further purification. Accordingly, extracts obtained at pH 7.00 with 1 M NaCl and at pH 9.00 without salt showed significantly higher protein purity than the extract obtained at pH 7.00 with 0.5 M NaCl extract (Table 1; detailed amino acid composition, % mol, of the extracts is shown in Supplementary Material Table S.8) and were therefore selected for detailed characterization by FT-IR and TGA, in line with OIV requirements for concentrated, contaminant-free protein extracts (Organisation Internationale de la Vigne et du Vin (OIV), 2022). The results were then compared to commercial animal- and vegetable-derived PFAs.

The FT-IR spectra of both Spirulina protein extracts were highly similar (Supplementary Material Fig. S.2), showing a protein profile consistent with literature phycocyanin spectra, with characteristic Amide A (3310-3270 cm⁻¹), Amide I (1700-1600 cm⁻¹), Amide II (1550-1400 cm⁻¹), and Amide III (1240-670 cm⁻¹) bands, indicating predominantly proteinaceous material. Bands at 1448 and 1418 cm⁻¹ were assigned to C—H wagging, while the broad 1080-1030 cm⁻¹ feature, typical of biliproteins, arose from overlapping bilin chromophore modes, combining pyrrole-ring C—C/C—N stretching with methine-bridge and pyrrolic C—H out-of-plane deformations.

C-phycocyanin (C-PC) and allophycocyanin (A-PC) contents were 7.6 ± 0.5% and 10.4 ± 0.9%, respectively, for the pH 7.00/1 M NaCl extract and 3.7 ± 0.1% and 4.7 ± 0.9%, respectively, for the pH 9.00/no-salt extract, meaning that C-PC + A-PC accounted for 31% and 16% of the respective Spirulina protein isolates, consistent with A₆₂₀/A₂₈₀ purities ratios of 0.29 ± 0.02 and 0.16 ± 0.02, respectively (Supplementary Material Fig. S.3).

TGA showed a higher residual mass at 900 °C for the pH 9.00/no-salt extract (24.1% vs. 19.9%), together with a lower onset degradation temperature and a smaller second-stage mass loss (44.3% vs. 52.5%) compared with the pH 7.00/1 M NaCl extract (Supplementary Material Fig. S.4 and Table S.9), indicating more non-protein residue and lower thermal stability in the former. Overall, these results identify the pH 7.00/1 M NaCl extract as the optimal compromise between high protein content and minimal non-protein contaminants. Finally, from an industrial application perspective, increasing the bead-milling time at pH 7.00/1 M NaCl from 6 to 24 h (Supplementary Material Table S.10) did not significantly affect protein yield or extract protein content, indicating that the extraction process can be substantially shortened without loss of performance.

3.3. Spirulina protein extract characterization

The Spirulina protein extract obtained through bead milling at pH 7.00/1 M NaCl for 20 h was characterized by SDS-PAGE under reducing conditions (Fig. 2A), showing well-resolved protein bands mainly in the 14-30 kDa range, together with additional protein bands observed between 30 and 100 kDa. Selected protein bands (Fig. 2A, numbered 1-6) were excised for in-gel digestion followed by LC-MS/MS analysis for protein identification. Table 2 summarizes the proteins identified in the excised protein bands, with NSAF quantification showing that approximately 70% of the major excised bands (4 out of 6) corresponded to phycobiliproteins, mainly the allophycocyanin α chain (P72504) and the phycocyanin β subunit (A9UKJ0), consistent with the prominent 17-28 kDa bands observed in the SDS-PAGE profile (Chaiklahan et al., 2011). These proteins appeared across multiple electrophoretic bands, likely due to protein-polysaccharide complexation, which can alter electrophoretic mobility and cause smearing, or to partial protein degradation, generating fragments with different molecular weights (Li et al., 2021).

Fig. 2.

Characterization of Spirulina protein extracts and commercial fining agents. (A) SDS-PAGE profiles of the protein fining agents used in this study. Lanes: molecular weight marker (kDa) are shown on the left (Marker), the numbers (1 to 6) on the Spirulina protein extract (AP) correspond to the excised and identified protein bands (see Table 2, for protein identification); Spirulina protein extract (AP); thermally denatured Spirulina protein extract (DAP, 100 °C/30 min); patatin (PAT); pea protein isolate (PEA); egg albumin (ALB); casein (CAS); and gelatine (GEL); (B) Variation of the zeta potential of protein fining agents as a function of pH. Different lowercase letters correspond to significantly different ζ-potential values (ANOVA, Tukey's HSD test, p < 0.05); (C) Second derivative analysis and curve-fitted amide I region (1710-1580 cm⁻¹) from FT-IR spectra of Spirulina protein extract (AP) and denatured AP (DAP).

Table 2.

Proteins identified in Spirulina extracts by LC-MS/MS after in-gel and in-solution tryptic digestion. For in-gel digestion, proteins are listed by NSAF within each excised band. For in-solution digestion, proteins are listed by total NSAF. pI, theoretical isoelectric point; MW, molecular weight; NSAF, normalized spectral abundance factor.

Method	NSAF	Accession	Description	pI*	Coverage [%]	Peptides	MW (kDa)
In-gel, band (1) 198 kDa	3.49E+04	P34831	Small ribosomal subunit protein uS2	6.24	7.57	1	27.8
	3.07E+04	Q1KMR6	Photosystem II CP47 reaction center protein	5.96	5.11	1	56.1
	2.78E+04	P48293	Protein RecA	5.06	5.48	1	39.2
	1.80E+04	Q307D5	Type I site-specific deoxyribonuclease	6.53	2.37	2	116.9
	1.54E+04	J9PHT9	Nitrite reductase	5.53	4.63	1	62.9
	9.11E+03	O32393	Histidine kinase	4.99	1.66	1	133.8
	8.79E+03	Q307C0	Type I restriction enzyme endonuclease subunit	5.43	2.73	1	117.4
In-gel, band (2) 49	6.30E+04	B2BN85	CpcH	6.94	8.12	1	30.8
	3.05E+04	A0A3P3ZKU1	Succinate dehydrogenase	5.72	3.13	1	63.5
	9.13E+03	B6DU62	Histidine kinase	5.67	0.82	1	219.2
In-gel, band (3) 28	1.65E+05	Q307B9	YgjP-like metallopeptidase domain-containing protein	9.88	11.57	2	28.9
	1.28E+05	Q306Z1	HNH endonuclease	11.12	11.69	1	17.1
	1.15E+05	P72504	Allophycocyanin alpha chain	4.89	14.29	1	17.4
	9.52E+04	Q9S431	Rod linker polypeptide CpcI	8.87	7.64	1	32.6
	9.34E+04	Q307B5	Cytosine-specific methyltransferase	7.68	12.87	2	50.1
	8.45E+04	B2BN67	CpcH	7.82	6.64	1	30.8
	7.53E+04	A0A385HVZ3	Ribulose-1.5-bisphosphate carboxylase/oxygenase large subunit (Fragment) (A0A385HVZ3_ARTPT)	8.95	6.3	1	28.9
	6.35E+04	B1A631	DNA-binding transcriptional activator HetR	5.72	8.67	1	34.9
	6.25E+04	Q307C2	Uncharacterized protein (Fragment)	9.44	2.91	1	35.4
	5.10E+04	Q307D7	ATP-binding protein	5.65	5.72	1	46.1
	4.64E+04	Q306Z4	HNH endonuclease	10.33	4.94	1	48.2
	3.00E+04	B6DU61	Serine/threonine protein kinase	6.23	2.05	1	82.0
	2.74E+04	Q307A1	Site-specific DNA-methyltransferase (adenine-specific)	5.04	3.44	1	90.7
	1.97E+04	O32393	Histidine kinase	4.99	1.16	1	133.8
	1.69E+04	Q19PI2	Magnesium chelatase	5.12	1.88	1	148.1

Method	NSAF	Accession	Description	pI*	Coverage [%]	Peptides	MW (kDa)
In-gel, band (4) 17	3,08E+04	A9UKJ0	Phycocyanin beta subunit	4.96	19,19	3	18,0
	2,17E+04	P72504	Allophycocyanin alpha chain	4.89	17,39	2	17,4
	1,61E+04	A0A0C4VYX4	ApcB protein	6.25	17,39	2	17,3
	1,15E+04	A0A3P3ZKQ8	Cytochrome c class II	8.44	9,8	1	23,0
	1,15E+04	Q307A7	Res	6.17	5,71	1	28,1
	1,02E+04	U3M0X3	Fatty aldehyde decarbonylase (Fragment)	4.78	11,65	1	23,4
	7,78E+03	D4ZX35	Type II methyltransferase M.AplI	8.32	8,41	1	38,9
	7,11E+03	A0A385HVZ3	Ribulose-1,5-bisphosphate carboxylase/oxygenase large subunit (Fragment)	8.95	7,48	1	28,9
	5,39E+03	P72505	Allophycocyanin beta chain	7.67	10,56	1	17,4
	5,30E+03	Q307C4	Methyltransferase	8.28	4	1	42,8
	4,42E+03	Q307E0	Putative serine protease	5.06	5,86	1	51,3
	2,14E+03	A0A3P3ZKT0	Chaperone protein ClpB	5.30	1,72	1	98,5
	1,86E+03	O32393	Histidine kinase	4.99	1,91	1	133,8
	1,15E+03	B6DU62	Histidine kinase	5.67	1,34	1	219,2
In-gel, band (5) 14-17	7.24E+04	P72504	Allophycocyanin alpha chain	4.89	21.74	2	17.4
	4.11E+04	A9UKJ0	Phycocyanin beta subunit	4.96	8.14	1	18.0
	2.73E+04	P34831	Small ribosomal subunit protein uS2	6.24	7.57	1	27.8
	2.52E+04	A0A3P3ZKW1	Succinylglutamate desuccinylase	8.55	6.19	1	35.8
	1.99E+04	B6DU59	Serine/threonine protein kinase	7.16	3.49	1	56.0
	1.70E+04	E0Z0D5	ChpX	5.63	3.96	1	50.4
	1.57E+04	P27868	Acetolactate synthase (Fragment)	6.16	1.73	1	63.4
	1.52E+04	Q307C0	Type I restriction enzyme endonuclease subunit	5.43	2.73	1	117.4
	1.33E+04	B9A0P6	60 kDa chaperonin (Fragment)	5.07	3.23	1	53.2
	1.08E+04	Q307D0	Type I restriction enzyme endonuclease subunit	9.19	2.86	1	65.2
	6.20E+03	O32393	Histidine kinase	4.99	1.66	1	133.8
	6.11E+03	Q307D5	Yype I site-specific deoxyribonuclease	6.53	2.17	1	116.9
	3.83E+03	B6DU62	Histidine kinase	5.67	0.77	1	219.2
In-gel, band 6) 14	3.49E+04	P72504	Allophycocyanin alpha chain	4.89	17.39	3	17.4
	2.63E+04	A9UKJ0	Phycocyanin beta subunit	4.96	11.05	1	18.0
	2.31E+04	P72505	Allophycocyanin beta chain	7.67	19.25	2	17.4
	1.25E+04	E1AW19	SQDG synthase (Fragment)	5.55	17.57	1	16.4
	1.15E+04	P72509	C-phycocyanin alpha subunit	5.83	8.02	1	17.6
	9.73E+03	A0A0C4VZ00	CpcI protein	8.33	2.43	1	32.5

Method	NSAF	Accession	Description	pI*	Coverage [%]	Peptides	MW (kDa)
In-solution	5.36E+04	A9UKJ0	Phycocyanin beta subunit	4.96	16.3	11	18.0
	1.47E+04	P72504	Allophycocyanin alpha chain	4.89	19.9	3	17.4
	1.47E+04	B2BN66	CpcD	5.36	11.5	2	9.7
	8.73E+03	E0Z0D1	CcmN	5.72	13.2	1	18.4
	8.60E+03	A0A3G2BB91	Gas vesicle protein A	4.65	12.5	1	7.6
	7.78E+03	P72509	C-phycocyanin alpha subunit	5.83	17.9	3	17.6
	6.81E+03	Q1KMR6	Photosystem II CP47 reaction center protein	5.96	9.4	3	56.1
	6.68E+03	C8CMN9	Multi-sensor hybrid histidine kinase (Fragment)	5.39	19.4	1	11.4
	6.62E+03	Q307C7	Type I RM system S subunit	8.98	13.7	3	44.1
	6.19E+03	P13552	Elongation factor Tu	5.16	15.9	3	44.7
	5.85E+03	P13576	Small ribosomal subunit protein uS12	11.32	7.3	1	13.8
	5.36E+03	Q307D6	TIGR02646 family protein	5.06	20.0	2	28.9
	5.21E+03	Q306Z0	HNH endonuclease (Fragment)	11.58	15.0	1	15.3
	4.43E+03	B2BN68	CpcI	8.65	9.7	2	32.7
	4.40E+03	A0A0C4VZ00	CpcI protein	8.33	7.6	2	32.5
	4.08E+03	Q0ZUC6	Histidine kinase (Fragment)	4.75	11.6	2	36.3
	4.00E+03	Q307A0	Type I RM system S subunit	9.23	8.2	2	47.1
	3.92E+03	P72505	Allophycocyanin beta chain	7.67	10.6	1	17.4
	3.85E+03	Q307C4	Methyltransferase	8.28	10.1	2	42.8
	3.75E+03	Q306Z3	HNH endonuclease	10.19	9.4	2	40.0
	3.75E+03	Q307B8	DNA (cytosine-5-)-methyltransferase	8.77	10.3	2	48.7
	3.57E+03	Q6WJF7	Superoxide dismutase (Fragment)	5.23	17.1	1	18.6
	3.49E+03	Q307D7	ATP-binding protei	5.65	10.5	2	46.1
	3.42E+03	A0A3G2BBM2	Protein RecA	6.33	8.3	1	25.5
	3.21E+03	D0PSV9	Nitrate/nitrite transporter (Fragment)	9.31	5.9	1	44.0
	3.11E+03	A0A3P3ZKT0	Chaperone protein ClpB	5.30	8.8	4	98.5
	2.90E+03	B9A0P6	60 kDa chaperonin (Fragment)	5.07	4.6	2	53.2
	2.90E+03	B2BN67	CpcH	7.82	8.1	1	30.8
	2.88E+03	A0A3P3ZKW4	Transposase IS200/IS605 family	10.31	5.2	1	28.9
	2.79E+03	A0A3P3ZKU1	Succinate dehydrogenase	5.72	10.1	2	63.5
	2.62E+03	A0A4D8US14	Probable transcriptional regulator ycf27	6.36	5.5	1	28.6
	2.58E+03	A0A385HVZ3	Ribulose-1.5-bisphosphate carboxylase/oxygenase large subunit (Fragment)	8.95	2.8	1	28.9
	2.50E+03	O07333	Phytoene synthase	5.44	4.5	1	35.9

Method	NSAF	Accession	Description	pI⁎	Coverage [%]	Peptides	MW (kDa)
In-solution	2.47E+03	Q307A4	Methyltransferase	8.80	3.4	1	34.3
	2.43E+03	G0YSA0	Rod shape-determining protein (Fragment)	5.54	9.0	1	32.9
	2.33E+03	E0Z0D2	CcmO	5.12	5.3	1	28.7
	2.25E+03	Q306Z9	Site-specific DNA-methyltransferase (cytosine-N(4)-specific)	9.02	3.3	1	44.7
	2.00E+03	Q307A8	Cytosine-specific methyltransferase	9.05	8.3	1	35.3
	1.88E+03	Q307A1	Site-specific DNA-methyltransferase (adenine-specific)	5.04	1.8	2	90.7
	1.86E+03	E0Z0D5	ChpX	5.63	2.6	1	50.4
	1.79E+03	A0A1S6GXD8	1-deoxy-D-xylulose 5-phosphate reductoisomerase	5.03	5.1	1	42.9
	1.73E+03	A7U8S1	Putative glutathione reductase (Fragment)	5.16	3.8	1	39.7
	1.71E+03	P27868	Acetolactate synthase (Fragment)	6.16	1.6	1	63.4
	1.70E+03	Q307A5	Cytosine-specific methyltransferase	9.24	4.1	1	47.4
	1.59E+03	P0DXD4	Argininosuccinate lyase	5.32	3.7	1	51.6
	1.57E+03	A0A3P3ZKS4	Histidine kinase	6.05	6.0	1	54.7
	1.54E+03	A0A3P3ZKX0	Glycogen synthase	5.57	6.1	1	54.5
	1.54E+03	Q2PPX5	Hydroxylamine reductase	5.28	4.4	1	60.6
	1.53E+03	A0A3P3ZKS3	Amine oxidase.flavin-containing	8.78	4.7	1	69.9
	1.50E+03	A0A3P3ZKS5	Protein translocase subunit SecA	5.14	1.5	1	105.7
	1.49E+03	Q2PPX6	HoxF	5.58	4.3	1	57.2
	1.46E+03	E3VS06	Alpha amylase	5.54	3.1	1	67.6
	1.40E+03	E5LCI2	Alpha-amylase	5.03	3.9	1	57.3
	1.32E+03	Q59950	Acetolactate synthase (Fragment)	5.64	3.4	1	66.5
	1.30E+03	Q307C0	Type I restriction enzyme endonuclease subunit	5.43	3.2	2	117.4
	1.28E+03	O32392	Adenylate cyclase	5.72	1.4	1	56.4
	1.23E+03	A0A4D8URJ0	Transketolase	5.73	3.9	1	72.4
	1.18E+03	Q307D0	Type I restriction enzyme endonuclease subunit	9.19	3.4	1	65.2
	1.17E+03	A0A3P3ZKR1	Periplasmic oligopeptide-binding protein of oligopeptide ABC transporter	5.11	3.5	1	67.9
	1.14E+03	Q307D9	Site-specific DNA-methyltransferase (adenine-specific)	5.43	3.5	1	80.4
	1.14E+03	J9PHT9	Nitrite reductase	5.53	4.6	1	62.9
	1.04E+03	Q0ZUC5	Histidine kinase	6.05	4.9	1	68.9
	9.70E+02	P13550	Elongation factor G	5.02	1.6	1	76.7
	8.60E+02	C0KJQ0	Histidine kinase (Fragment)	5.15	3.6	1	87.0
	7.56E+02	D3KZ43	Methionine synthase	4.99	1.6	1	130.9
	4.17E+02	B6DU62	Histidine kinase	5.67	0.8	1	219.2

^⁎Theorethical isoelectric point (pI) computed using compute pI/Mw tool from Expasy portal at https://www.expasy.org/

Phycobiliproteins are the major components of the light-harvesting phycobilisome complexes (Safi et al., 2014), aligning with previous studies showing that Spirulina soluble protein fractions are dominated by phycobiliproteins, alongside Rubisco (∼50-60 kDa) and photosynthetic enzymes (∼38/98 kDa) (Teuling et al., 2017).

In-solution digestion complemented this approach by broadening proteome coverage and avoiding gel-related losses (Albuquerque et al., 2023). A total of 66 proteins were identified by in-solution digestion, with the same top NSAF hits (P72504, A9UKJ0; MWs 17.4-18.0 kDa) as those obtained by in-gel digestion (Table 2). Representative MS/MS spectra are shown in Supplementary Material Fig. S.5. Together, these complementary approaches validate the dominance of phycobiliproteins and photosynthetic enzymes in the soluble fraction.

From an allergenicity perspective, phycocyanin has been implicated as a potential allergen, as evidenced by a reported case of Spirulina-induced anaphylaxis in a 14-year-old patient, in which IgE reactivity was associated with the β chain of C-phycocyanin (Le et al., 2014). More broadly, a recent short report on edible algae suggests that several algal components can elicit allergic responses in humans, although the prevalence of such reactions in the general population remains poorly characterized, highlighting the need for further studies in this area (James et al., 2023). In addition, protein extraction methods can influence the proteomic profile and, consequently, the potential allergen load of Spirulina extracts, indicating that tailored extraction strategies may help modulate exposure to specific allergenic proteins (Irvani et al., 2024). Compared with previous studies that characterized Spirulina protein concentrates mainly for food or nutraceutical applications, the present work provides complementary information directly relevant to wine fining, by linking the identified protein profile and potential allergenic components to extraction conditions and to the requirements for oenological protein preparations.

3.4. Potential of Spirulina protein extract as a wine fining agent: comparison with animal and plant-derived protein fining agents in model wine solutions

The fining performance of the Spirulina protein extract (AP) toward PAs was evaluated against the commercial PFAs gelatin (GEL), egg albumin (ALB), and casein (CAS) and the plant proteins patatin (PAT) and pea protein (PEA), in WWM and RWM systems (Supplementary Material Tables S.11 and S.12.). These model systems were designed to reflect the typical PA composition of grape skins, seeds, and wine (Bambina et al., 2024). To assess the impact of protein structural modifications, as previously reported for plant-derived PFAs (Marangon et al., 2019), a thermally denatured Spirulina protein extract (DAP), was also prepared and tested. Thermal denaturation was confirmed by spectroscopic analysis, which revealed a loss of the C-PC and A-PC absorption maxima and an increase in the A₆₂₀/A₂₈₀ ratio. These spectral changes result from chromophore reorientation following protein unfolding, which alters the conformation of the tetrapyrrole chromophore and consequently modifies its light absorption characteristics (Scheer & Kufer, 1977, Supplementary Material Fig. S.3.). All fining trials were conducted at 50 g/hL, corresponding to the maximum OIV approved dosage for plant protein fining agents (Organisation Internationale de la Vigne et du Vin (OIV), 2022). Although the AP extract presented a light green color in the solid state, no odor was detected after freeze-drying. After fining, no color or unusual aroma was detected in either wine model.

In the WWM system, all PFAs significantly reduced the total PAs content (Fig. 3; Supplementary Material Table S.11). Animal-derived PFAs were the most effective, with reductions of 99.6% (GEL), 94% (CAS), and 84% (ALB). Plant-derived PFAs reduced PAs by 86% (PAT) and 82% (PEA). AP-based treatments reduced total PAs by 81% (AP) and 84% (DAP), with no significant differences between the native and denatured extracts. Their fining efficiencies were comparable to those of ALB, PAT, and PEA, but significantly lower than those of GEL and CAS (approximately12.5% lower on average). AP and DAP effectively removed monomeric flavan-3-ols, decreased PA galloylation, and lowered mDP, indicating an overall fining performance broadly similar to that of commercial PFAs in WWM. The results obtained for ALB, GEL, and CAS concerning their fining efficiency and impact on structural features of PAs are in line with those obtained by Cosme et al. (2008) for white wine.

Fig. 3.

Effect of protein fining agents on the mean degree of polymerization (mDP) and content (mg/L) of (a) monomeric flavan-3-ols, (b) oligomeric PAs, (c) polymeric PAs, and (d) total PAs in white and red wine model system. Control: untreated; AP: Spirulina protein extract; DAP: thermally denatured AP; PAT: Patatin; PEA: pea protein; GEL: Gelatine; ALB: egg Albumin; CAS: Casein. Different uppercase letters correspond to significantly different contents (ANOVA, Tukey's HSD test, p < 0.05). Different lowercase letters correspond to significantly different mean degree of polymerization (ANOVA, Tukey's HSD test, p < 0.05). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

In the RWM system (Fig. 3; Supplementary Material Table S.12), all PFAs again significantly reduced total PAs, although to a much lesser percentage than in the WWM. Animal PFAs reduced total PAs by 12-19%, preferentially removing polymeric PAs (PPAs) over oligomeric PAs (OPAs), while PAT and PEA reduced PAs by only 3% and 11%, respectively. In contrast, AP showed the highest reduction in total PAs (30%), followed closely by DAP (29%), with a particularly strong effect on OPAs. The results obtained for GEL, PAT, and PEA are in accordance with those of Río-Segade et al. (2019), where GEL was more efficient than PEA and PAT. Only minor decreases in mDP were observed in the RWM, mainly for DAP, GEL, and CAS. Significant reductions in PA galloylation were observed only for AP, while DAP, GEL, and CAS decreased the (E)GC/(E)C ratio. Animal and plant-derived fining agents were unable to reduce either the monomeric flavan-3-ols or oligomeric PAs in the red wine model, in accordance with the results of Dias et al. (2023).

Although percentage reductions were lower in the RWM, the absolute amount of PAs removed was higher than in the WWM, due to the substantially (approximately 60%) higher initial PA content. The greatest increase in absolute PA removal between the RWM and the WWM was observed for AP and DAP (approximately 21%), followed by CAS (12%), GEL (10%), ALB and PEA (8%), and PAT (2%). This contrasting behaviour between the WWM and the RWM mainly reflects differences in the PA:Protein ratio, which strongly influences protein-polyphenol interactions and aggregation pathways. At low ratios, small soluble complexes are formed; at intermediate ratios, multidentate cross-linking yields insoluble oligomers; and at high ratios, macroscopic precipitation occurs (Pascal et al., 2007). These interactions are further modulated by solution conditions and tannin structural features, particularly average polymer size (Ropiak et al., 2017).

To better understand the contrasting behaviour of PFAs in both model systems and identify the physicochemical drivers of fining efficiency, all PFAs were characterized in terms of amino acid composition (Supplementary Material Table S.13), molecular-weight distribution (SDS-PAGE; Fig. 2A), zeta potential (Fig. 2B), and secondary structure by FT-IR (Fig. 2C and Supplementary Material Table S.14). Animal PFAs showed the highest protein contents, whereas AP showed protein levels similar to those of PAT and PEA (Supplementary Material Table S.13), and both SDS-PAGE profiles (Fig. 2A) and protein secondary structure profiles (Fig. 2C and Supplementary Material Table S.14) of the commercial PFAs agreed with literature data. Zeta potential (Fig. 2B), which reflects surface electrostatics and thus colloidal stability, was negative for AP between pH 2.80 and 3.80 and became more negative with increasing pH, consistent with the reported pI values of Spirulina proteins. In contrast, most other PFAs (except PEA) were positively charged over this range, and their charge decreased with increasing pH, in line with their known pI value. PEA shifted from slightly positive to zero near pH 3.50 and became negative at pH 3.80, matching its reported pI. However, all PFAs displayed absolute zeta potential values below 30 mV, indicating poor intrinsic colloidal stability and a tendency to aggregate (Li et al., 2018).

Simple amino acid composition is a powerful and broadly applicable descriptor of protein behaviour, capable of predicting interaction propensity with an accuracy comparable to that of more complex domain- or structure-based features (Roy et al., 2009). Building on this evidence, 66 amino acid-based physicochemical descriptors from the AAindex database, complemented with amino acid side chain hydrogen-bond donor and acceptor constants, were calculated and used as input variables (Supplementary Material Tables S.2. and S.15) for exploratory orthogonal partial least squares (OPLS) models developed to predict total PA removal in the WWM and the RWM. The OPLS models (Fig. 4A and B) showed excellent fit and predictive power without evidence of overfitting (WWM: 1 predictive +3 orthogonal components, R²Y = 0.941, Q² = 0.861; RWM: 1 predictive +3 orthogonal components, R²Y = 0.967, Q² = 0.906; Supplementary Material Tables S.16 and S.17, respectively), with model validity supported by permutation tests and cross-validation (WWM CV p < 0.00977; RWM CV p < 0.00273; Supplementary Material Figs. S.6 and S.7, and Tables S.18 and S.19), in line with recommended OPLS criteria. Relevant descriptors were selected using VIP > 1 (Fig. 4C and D), significant regression coefficients via modified jack-knifing (Supplementary Material Figs. S8 and S9), and S-plot thresholds (|p(cov)| ≥ 0.10, |p(corr)| ≥ 0.5) to retain robust, influential predictors (Figs. 4E and F).

Fig. 4.

OPLS-DA analysis of protein fining efficiency (% PA removal) in WWM and RWM. (A) WWM score plot; (B) RWM score plot; (C) WWM VIP plots; (D) RWM VIP plots; (E) WWM S-plot; (F) RWM S-plot; (G) SUS-plot comparing predictive components (WWM vs RWM; V7 opposite effect); (H) WWM HCA dendrogram; (I) RWM HCA dendrogram.

In the WWM, higher PA removal by PFAs correlated positively with increased protein content of PFAs, random-coil content (V82, V89, V84, V80), a larger hydrodynamic radius R_h (V65), and a higher proline-rich index (V4), followed by conformational flexibility (V64), relative distance to the natively unfolded protein frontier of the charge-hydrophobicity diagram of Uversky (V66), partition energy (V26), percentage abundance of hydrophobic amino acids (V2), ratio of total β-turn/(total helix+ total β-sheet) (V83), intrinsic disorder parameter (%) (V63), value of θ(i) (V44), ratio of HB acceptor constant β /HB donor constant α (V58), side-chain torsion angle φ (V41), and percentage area β-turn (band at 1674 ± 2 cm⁻¹) (V74), and negatively with a higher abundance of negatively charged residues (V7), localized electrical effects (V53), Ser, the number of full nonbonding orbitals (V25), Asp, JJT2 hydrophobicity (V19), Phe, Leu, the Kyte-Doolittle hydrophobicity scale (V10), percentage abundance of polar amino acids (V1), number of hydrogen-bond donors (V24), the Zimmerman polarity index (V16), hydration number (V27), side-chain angle θ (V40), Glu, ratio of polar/hydrophobic amino acids (V3), Asn, total hydrogen-bond acceptor constant β (V55), STERIMOL minimum width of the side chain (V37), and the smoothed upsilon steric parameter (V15).

In RWM, greater PA removal was associated with higher net relative strength of HB acceptors (V61), Arg, Ala, ratio negative charge index/positive charge index (V9), ratio of polar accessible molecular surface area/nonpolar accessible molecular surface area (V52), ratio of the relative strength of HB acceptors/ relative strength of HB donors (V62), ratio of polar accessible molecular surface area/total accessible molecular surface area (V50), Sweet-Eisenberg hydropathy (V12), relative strength of HB acceptors (V59, and negative charge index (V7), whereas more positive zeta potential (V67), lower free energy of transfer to water (V28), Pro, higher average accessible surface area (V46), ratio of nonpolar accessible molecular surface area/total accessible molecular surface area (V51), and Asn were negatively related to PA removal. In contrast to the WWM, where lower protein content resulted in lower fining efficiency of PFAs (Fig. 4C and G), protein content had no effect on fining efficiency in the RWM (Fig. 4D and G). In fact, protein features enhancing fining efficiency differ markedly between the WWM and the RWM. As can be observed in the SUS-plot comparing the predictive components (Fig. 4G) of both models, largely distinct predictor sets were obtained, with no shared positive drivers. Asn abundance was consistently associated with reduced fining efficiency in both models, and the negative-charge index (V7) was positively correlated with PA removal in the RWM but negatively correlated in the WWM. Concordance among the score plots, VIPs, coefficients, S-plots, and the SUS-plot supports the robustness and internal consistency of the OPLS analysis and indicates that different protein-polyphenol interaction mechanisms dominate in the WWM versus the RWM. However, given the limited number of fining agents tested relative to the number of descriptors considered, these mechanistic interpretations should be considered exploratory. These differences are likely due to variations in PA:protein ratios and matrix composition.

Although the OPLS models identify the protein descriptors most strongly associated with fining performance, they do not establish causal relationships. Therefore, the mechanistic interpretations presented below are based on the integration of these statistical associations with the current understanding of protein-tannin interactions reported in the literature. Tannin-protein interactions involve a combination of hydrogen bonding, hydrophobic interactions, van der Waals forces, π-π stacking, and pH-dependent electrostatic interactions (Quan et al., 2019), while recent work suggests that water-mediated hydrogen-bond networks also contribute to complex stabilization (Yoshimura et al., 2025). No single mechanism dominates; instead, tannin structure, protein conformation, solvent, and thermodynamic conditions jointly govern binding mode and aggregation kinetics, with more apolar tannins such as pentagalloylglucose tending to form hydrophobic “coatings” and more polar procyanidins favoring hydrogen-bonding-driven cross-linking (Frazier et al., 2006). Therefore, the superior performance of AP in the RWM is attributed to its higher content of negatively charged residues (V7) and higher ratio of negatively to positively charged amino acids (V9), together with its more negative zeta potential (V67), which reflects a greater abundance of aspartate and glutamate residues capable of acting as hydrogen-bond acceptors. Together with the strong positive loadings of HB-acceptor strength descriptors (V61, V62, V59), these results suggest that hydrogen bonding, rather than electrostatic attraction, plays a key role in promoting protein-proanthocyanidin interactions under the RWM conditions. In contrast, in the WWM, the most efficient PFAs (gelatin and casein) are rich in proline, and in the case of gelatin also hydroxyproline (V4) and adopt more random-coil (V80, V82, V84, V89), flexible conformations (V63, V64, V65, V66), which increases the accessible binding surface area and enable cooperative, multistage aggregation, in agreement with prior observations that tannins interact non-specifically with compact globular proteins (e.g., BSA) but follow a two-stage mechanism with random-coil gelatin (Le Bourvellec & Renard, 2012).

OPLS models for polymeric proanthocyanidin (PPA) removal in both the WWM and the RWM showed high fit and predictive power (WWM: 1 predictive +3 orthogonal components, R²Y = 0.944, Q²Y = 0.865, CV p < 0.00908; RWM: 1 predictive +4 orthogonal components, R²Y = 0.982, Q²Y = 0.938, CV p < 0.00308), showing approximately the same distribution and weighting of variables as that found for total PAs (y = 1.0053 + 0.0015, R² = 0.9998 for the WWM and y = 1.1644-0.0114, R² = 0,9286 for the RWM). For OPAs, a significant model was obtained only for the RWM (1 predictive +4 orthogonal components, R²Y = 0.962, Q²Y = 0.854, CV p < 0.0336), whereas the WWM OPA model lacked predictive validity (1 predictive +2 orthogonal components, R²Y = 0.683, Q²Y = 0.364, CV p < 0.516), most probably due to their low concentration. In the RWM, total galloylated PAs were also well described by the OPLS model (1 predictive +3 orthogonal components, R²Y = 0.977, Q²Y = 0.946, CV p < 0.000425), and the galloylation percentage of total PAs (1 predictive +2 orthogonal components, R²Y = 0.948, Q²Y = 0.901, CV p < 0.000752) was driven largely the same variables as total PAs, PPAs, and OPAs. No significant galloylation model was obtained for WWM, and the (E)GC/(E)C ratio model was likewise not significant (1 predictive +6 orthogonal components, R²Y = 0.932, Q²Y = 0.700, CV p < 0.528).

Hierarchical clustering of the OPLS predictive components further distinguished PFAs according to fining performance in each model. In the WWM, three clusters were identified (mean silhouette = 0.86, cophenetic correlation = 0.94, Fig. 4H): GEL (Cluster 1) and CAS (Cluster 2), characterized by high proline content and random-coil structure, showed the highest efficiency (GEL > CAS); and a lower-efficiency group comprising PAT, ALB, DAP, AP, and PEA (Cluster 3), was associated with lower random-coil content and a higher negative-charge index. In the RWM, three clusters were also identified (mean silhouette = 0.75, cophenetic correlation = 0.83, Fig. 4I): AP and DAP formed the most effective cluster, characterized by higher hydrogen-bond acceptor capacity, and negative zeta potential; GEL, CAS, ALB, and PEA formed an intermediate-performance group; and PAT (Cluster 3) showed the lowest efficiency.

These results indicate that native and denatured Spirulina protein extracts are at least comparable to commercial PFAs and, under red-wine-like conditions, can surpass them in proanthocyanidin removal. From a practical perspective, these findings highlight the potential of Spirulina-derived proteins as a sustainable alternative to conventional animal- and plant-derived fining agents. Although further optimization of protein extraction and production costs will be important to enhance commercial competitiveness, the fining performance demonstrated in this study supports their further evaluation under industrial winemaking conditions.

4. Conclusions

To the best of our knowledge, this is the first study to comprehensively characterize and evaluate the proanthocyanidin fining efficiency of a purified Spirulina protein extract in white and red wine model solutions. Firstly, a high-purity protein isolate with 63% protein content was obtained, with a 25% yield, by bead milling at pH 7.00 with 1 M NaCl followed by ethanol precipitation. These processing conditions offer a good compromise between protein recovery from whole biomass and minimal co-extraction of non-protein components. Secondly, under model wine conditions, the Spirulina protein extract showed PA fining efficiency (81% PA removal) that was similar to that of plant proteins (82-86% PA removal) and egg albumin (84% PA removal) and lower than that of gelatin (99.6%) and casein (94.0%%) in the white wine model, but superior (30% PA removal) to that of all commercial fining agents in the red wine model (3-19% PA removal). This superior performance in the red wine model is consistent with the higher negative charge, more negative zeta potential, and stronger hydrogen-bond-acceptor capacity of the Spirulina protein extract, suggesting that these properties contribute to enhanced protein-proanthocyanidin interactions under red wine model conditions. Conversely, the lower performance in the white wine model compared with that of gelatin and casein is consistent with the lower random-coil content, proline abundance, and reduced conformational flexibility of the Spirulina protein extract, suggesting that these structural characteristics contribute to the observed differences in fining efficiency. Thermal denaturation of Spirulina proteins did not improve their fining efficiency. Overall, these results position Spirulina proteins as promising next-generation fining agents, capable of matching or surpassing current animal and plant-based PFAs in PA management, while simultaneously addressing sustainability challenges; however, validation in real white and red wines is required before practical application can be confirmed. These mechanistic interpretations are supported by the OPLS modeling and physicochemical characterization of the fining agents but remain inferential and require further validation through direct studies of protein-proanthocyanidin interactions in complex wine matrices. In addition, future studies should address pilot-scale production, economic feasibility, batch-to-batch reproducibility, shelf stability, storage conditions, and regulatory approval to facilitate the industrial implementation of Spirulina-derived protein fining agents and their acceptance by the wine industry.

CRediT authorship contribution statement

Elisa Costa: Writing – review & editing, Writing – original draft, Visualization, Investigation, Formal analysis, Data curation. Miguel Ribeiro: Writing – review & editing, Writing – original draft, Methodology, Data curation, Conceptualization. Luís Filipe-Ribeiro: Writing – review & editing, Writing – original draft, Visualization, Validation, Resources, Project administration, Methodology, Formal analysis. Fernanda Cosme: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Funding acquisition, Formal analysis, Data curation, Conceptualization. Fernando M. Nunes: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Funding acquisition, Formal analysis, Data curation, Conceptualization.

Funding sources

The CQ VR is funded by FCT through the projects UID/00616/2025 (https://doi.org/10.54499/UID/00616/2025) and UID/PRR/00616/2025 (https://doi.org/10.54499/UID/PRR/00616/2025), and Ph.D. grant to E. Costa (SFRH/BD/144107/2019, https://doi.org/10.54499/SFRH/BD/144107/2019). This work was mainly funded by the European Regional Development Fund (ERDF) through NORTE 2020 (North Regional Operational Program 2014/2020) to the project “BluEnology - Desenvolvimento de um adjuvante proteico de microalgas para uso enológico”, no. NORTE-01-0247-FEDER 047005.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material

Data availability

Data will be made available on request.