Fine-tuning mitochondrial activity in Yarrowia lipolytica for citrate overproduction

Yarrowia lipolytica is a non-conventional yeast with promising industrial potentials for lipids and citrate production. It is also widely used for studying mitochondrial respiration due to a respiratory chain like those of mammalian cells. In this study we used a genome-scale model (GEM) of Y. lipolytica metabolism and performed a dynamic Flux Balance Analysis (dFBA) algorithm to analyze and identify metabolic levers associated with citrate optimization. Analysis of fluxes at stationary growth phase showed that carbon flux derived from glucose is rewired to citric acid production and lipid accumulation, whereas the oxidative phosphorylation (OxPhos) shifted to the alternative respiration mode through alternative oxidase (AOX) protein. Simulations of optimized citrate secretion flux resulted in a pronounced lipid oxidation along with reactive oxygen species (ROS) generation and AOX flux inhibition. Then, we experimentally challenged AOX inhibition by adding n-Propyl Gallate (nPG), a specific AOX inhibitor, on Y. lipolytica batch cultures at stationary phase. Our results showed a twofold overproduction of citrate (20.5 g/L) when nPG is added compared to 10.9 g/L under control condition (no nPG addition). These results suggest that ROS management, especially through AOX activity, has a pivotal role on citrate/lipid flux balance in Y. lipolytica. All taken together, we thus provide for the first time, a key for the understanding of a predominant metabolic mechanism favoring citrate overproduction in Y. lipolytica at the expense of lipids accumulation.

www.nature.com/scientificreports/ and both isocitrate and citrate accumulation. Then citrate and isocitrate are excreted to extracellular medium but citrate can also be converted to acetyl-CoA, used as precursor for lipid synthesis 25,26 . Recent studies highlighted the key role of oxygen on citrate optimized production in Y. lipolytica cultures [27][28][29] . Indeed, it has been showed that increased oxygen mass transfer rate (OTR) improves up to threefold citrate production by Y. lipolytica W29 strain 28 . Similarly, recent investigations led to citrate optimization by controlling the dissolved oxygen concentration (pO 2 ) 27,29 . Finally, in this older work, authors reported that citrate overproduction is inversely correlated to the cyanide-resistance respiratory pathway on Candida/Yarrowia lipolytica 30 . Cyanide resistance pathway is mainly present in plants and fungi and involves an alternative oxidase (AOX) protein 31,32 ; along with II NADH dehydrogenase (NDH2e), both part of the uncoupled ΔΨ-independent mitochondrial respiration 33,34 . The exact mechanism involving AOX on citrate production is not clearly determined but NDH2e-AOX pathway has been proposed as defense mechanism against Reactive Oxygen Species (ROS) accumulation in Y. lipolytica at stationary phase 35 . Although mitochondria are the main sources of ROS, other cytoplasmic proteins have recently been characterized as sources of ROS. This is the case of the long chain fatty alcohol oxidase (FAO). FAOs are located in the endoplasmic reticulum and participate in the oxidation of long chain alkanes or fatty acids, and production of long-chain aldehyde and hydrogen peroxide (H 2 O 2 ) They have been mainly studied in industrial yeasts Candida tropicalis or Candida cloacae 36 and recently identified in Yarrowia lipolytica 37 .
When glucose is used as a sole carbon source, citric acid production by Y. lipolytica at stationary phase is favored over lipid accumulation 38 . Recent studies have implemented strategies for controlling substrate feeding rate for lipid accumulation preventing the production of citric acid 21 . The authors managed to identify a range of N/C ratios favoring the production of lipid without citric acid excretion. Similarly, a GEM of Y. lipolytica has been proposed in combination with dynamic Flux Balance Analysis (dFBA), a computational algorithm, to study environmental conditions favorable for lipid accumulation over citric acid production 13 . These authors found that the cell specific oxygen consumption rate (rO 2 ) is reduced during lipid accumulation. Applying a fed-batch strategy under reduced dissolved oxygen concentration, from 50 to 1%, they improved lipid accumulation, from 200 mg/gDCW to 400 mg/gDCW and avoid citrate excretion. Computational strategies such as dFBA allow to analyze genome-scale reconstructed metabolic networks and to run knock-out or environment-based optimization strategies for biomass of metabolites production 19,39,40 .
FBA predicts flux balances based on constraints and objective functions for biochemical networks verifying steady-state assumption 39,41 . The mathematical description of the metabolic reactions relies on the stoichiometries of reactants and products and does not require enzyme kinetics parameters. Dynamic flux balance analysis (dFBA) extends the FBA framework to help model how metabolic states change over time and interact with their environment. The method was developed by Varma and Palsson to predict the behavior of E. coli during several batch runs 42 . It was then formalized by Mahadevan et al. who presented two different approaches to dynamic flux balance analysis, called the dynamic optimization approach (DOA) and the static optimization approach (SOA) 43 . The DOA method optimizes over the entire time period of interest to obtain time profiles of fluxes and metabolite levels using non-linear programming problem. The SOA method divides time in discrete intervals and a new FBA problem is solved at time t after updating the external conditions according to the FBA solution at time t − 1.
Five GEMs are known from the literature so far. The iNL895 model was the first genome-wide metabolic network of Y. lipolytica 15 . This model is based on the old consensus model of Saccharomyces cerevisiae, to which was added the lipid network. The iYL619_PCP model was released the same year and experimentally validated by studying growth on glucose and mineral medium 16 . This GEM was updated with more detailed reactions and became iYL_2.0. It has been used to optimize the accumulation of triacylglycerol (TAG) 18 . Similarly, the model iMK735 was developed to study conditions for lipid optimization in Y. lipolytica 13 . This model was validated with dFBA on mineral medium with glucose or glycerol as sole carbon source. The iYali4 model was constructed from the consensus network of S. cerevisiae 44 , iNL895 and iYL619_PCP. The model has been applied to the study of amino acid metabolism and the overflow phenomenon associated to lipid accumulation 17 . We consider that iYali4 model is the most complete regarding lipid metabolism, compared to previous models, and has been updated based on the genome sequence of the strain Y. lipolytica W29 used in this study 45 .
Here, we used the Yali4 GEM and dFBA SOA strategies with COBRA Toolbox under MATLAB to analyze flux distributions associated with citrate production and alternative oxidase (AOX) protein activity. For this end, we adjusted the iYali4 model with the mitochondrial AOX reaction identified in Y. lipolytica 46,47 and performed dFBA to identify metabolic pathways contributing to citrate optimization, using gurobi solver (Gurobi Optimizer) for linear programming (LP).

Methods
Microorganism and media composition. The Yarrowia lipolytica W29 wild-type strain was used in this study. It is from CIRM-Levures collection and was provided by Micalis, INRA (Jouy-en-Josas, France) and stored at − 80 °C on 50% (v/v) glycerol.
Two media were used in this study. The YPD medium is composed of 10 g/L yeast extract, 10 g/L peptone (BD Biosciences, France) and 20  www.nature.com/scientificreports/ Culture conditions. All pre-cultures were cultivated in 200 mL Erlenmeyer flasks on YPD medium at pH 6 and incubated on a rotary shaker at 28 °C and 100 rpm. Culture broth was taken at exponential growth phase, after 16-18 h of cultivation, centrifuged and washed in a 9 g/L NaCl prior to inoculation in mineral medium for both bioreactors and shaked flasks culture. Bioreactor culture (n = 1) were performed in the 7 L Applikon system. The temperature was maintained at 28 °C and pH at 6 by the addition of 5 M NaOH. Dissolved oxygen was maintained above 20% from air saturation by controling the stirrer speed between 300-600 rpm and aeration at 2 vvm of air and dioxygen from 0 to 1 vvm. An antifoaming agent (B200 K) was added to 0.09% v/v at inoculation.
Culture in Erlenmeyer flask (n = 1) were performed at the same condition as the pre-culture, on a rotary shaker at 28 °C. The pH of the culture was maintained between 5 and 6 by manually and periodically adding 300 µL of 5 M NaOH in sterile condition. 40 mM n-propyl Gallate (nPG) stock solution were prepared on 70% ethanol and periodically added to cultures, to a final 400 µM concentration. It is a well-known inhibitor of the alternative oxidase (AOX) enzyme and already used to study mitochondrial respiration in Yarrowia lipolytica 35 . Biomass and metabolites analysis. Cell growth was followed by a spectrophotometer with calculation of yeast optical density (OD) at 600 nm. Dried biomass was carried out and calibrated against the DO 600 nm for the calculation of extracellular metabolic fluxes per gram of dry cell weight (gDCW). We found 0.62 gDCW/L concentration of biomass with DO 600 nm = 1, for Y. lipolytica W29 grown on mineral media.
Need for accurate Y. lipolytica GEM description. We used iYali4 the GEM of Y. lipolytica by Kerkhoven et al., and an updated version for Y. lipolytica W29 to study citrate optimization. Growth rate on glucose and nitrogen-limited mineral medium was predicted by the model based on glucose specific consumption rate adequately fitted to the experimental results (Table 1 and Fig. 1). It was noted that iYali4 and the other GEMs did not integrate the alternative oxidase (AOX) reaction. So we added the AOX reaction as part of the electron transfer chain (ETC), based on the hypothesis that this enzyme has a protective effect against ROS generation in Y. lipolytica and probably on citrate production 30,31,35 . The stoichiometry of AOX reaction were determined from literature and in coherence with iYali4 ETC reaction: ubiquinol + 0.25 02 = > ubiquinone + 0.5 H 2 O 32 . AOX reaction is doesn't contribute to proton pumping across the mitochondrial membrane 48 .

Results
Experimentally determined kinetic parameters. We performed a series of bioreactor batch cultivations of Yarrowia lipolytica in mineral medium for kinetic parameters determination. These cultures were running in nitrogen-limited (N-lim) conditions to favor citrate production. Glucose was used as carbon source. Initial two glucose concentrations were accessed in order to determine the N/C ratio effect on cells specific growth rate (µmax) and other kinetic physiological parameters (Table 1). Specific growth rate of 0.076 ± 0.004 h −1 is observed in the lower N/C ratio compared to 0.049 ± 0.004 h −1 in the higher N/C ratio condition. The opposite effect of the N/C ratio is observed for glucose uptake where higher glucose uptake of 1.148 ± 0.264 mmol/ gDCW/h is observed for the higher N/C ratio condition. Citrate, isocitrate, pyruvate, acetate and succinate were detected in the batch cultures but only after extracellular nitrogen source (NH4 + ) is totally consumed (between 28 and 30 h of culture). Citrate and isocitrate specific production rates are higher under low N/C ratio batch cultivation. For the remaining of our investigations, we considered the higher N/C ratio condition to avoid the glucose inhibition effect at 100 g/L. growth. The experimentally determined kinetic parameters such as specific glucose and oxygen uptake at the higher N/C (0.00204) condition were used as constrains to predict cell growth behavior. The model predicts a growth rate of 0.052 h −1 , with low 0.616% error (Table 1). Thus, the Yali4 model accurately predicts the Yarrowia lipolytica W29 strain growth in mineral medium. However, this model applied to FBA algorithm was not able to predict the glucose inhibition effect on 100 g/L glucose concentration medium. We also tried the iMK735 model, but it failed to predict accurate growth.

Predicting citrate production upon N-lim. Nitrogen limitation (N-lim) is the mostly used limitation
parameter to trigger lipid accumulation and citrate production by Y. lipolytica. Initial NH4 + and glucose concentrations were set to 102 mg/L and 50 g/L, respectively. Experimental data obtained from duplicate flasks show two growth phases (Fig. 1). In the first growth phase an exponential growth rate µmax 1 = 0.100 ± 0.008 h −1 was determined between 45 and 50 h (only residual NH4 + remain in the culture medium and were qualitatively measure by colorimetric essay). A second stable growth phase, with a µmax 2 of 0.0240 ± 0.0013 h −1 is observed thereafter until the culture reached stationary phase, at 96-100 h. Up to 6 gDCW/L of biomass concentration has been observed. Citrate production clearly started upon nitrogen starvation and cell growth limitation. It's mainly produced during the cell growth stationary phase and reached 10.9 g/L at the end of cell cultivation. Lower isocitrate concentration of 1.3 g/L was then observed. After 170 h of culture, residual 5 g/L of glucose remained in the medium (Fig. 1). The initial biomass concentration was set to 0.06 gDCW/L based on experimental data. Then we applied the dynamic Flux Balance Analysis (dFBA) algorithm to Yali4 model with the aim to simulate our experimental results. Experimental initial glucose concentration (50 g/L) and uptake rate (0.64 mmol/gDCW/h) were first set in the model according to the experimental data. Slightly adjusting the specific glucose uptake rate down to 0.60 mmol/gDCW/h allowed to better fit the experimental curve. The alternative oxidase (AOX) reaction flux was fixed to zero during the growth phase, as reported by Guerrero-Castillo et al., and free-bounded during the stationary phase 35,49,50 . Therefore, a modified dynamicFBA.m MATLAB function has been proposed to allow further dFBA simulations when growth rate reaches the stationary phase (µ = 0). Then, the initial objective function of biomass optimization during the growth phase was changed to "lipid accumulation" as it's experimentally observed upon stationary phase. Therefore, our dFBA model permitted prediction of biomass and citrate production in good agreement with the experimental results (Fig. 2a). Nevertheless, isocitrate production cannot be predicted by the model. The predicted specific growth rate of 0.052 h −1 differs from the experimental µmax 1 (0.100 ± 0.008 h −1 ) but looks closer if we consider the average of the whole culture, i.e. the mean µ of the two successive growth phases (0.062 h −1 ).
Reduced simulated OxPhos during citrate production and lipid accumulation. Glucose consumption reached the upper-bound rate value of 0.60 mmol/gDCW/h during exponential growth phase (expphase), and it was constrained to 0.39 mmol/gDCW/h at stationary growth phase (stat-phase). Predicted NH4 + and oxygen specific consumption rates are respectively 0.468 mmol/gDCW/h and 1.54 mmol/gDCW/h at expphase, and both drop to 0 and 0.217 mmol/gDCW/h respectively at stat-phase. Carbon dioxide (CO 2 ) is produced only on exp-phase (Fig. 2b). These results are hallmarks of reduced cell energy demand. To specifically illustrate that, we investigated the simulated mitochondrial activity, emphasing the fluxes along the complexes of the electron transfer chain (ETC). The model reports reduced ATP production rate by complex V (CV) at stat-phase (Fig. 2c). Complex I (CI) and complex II (CII) are inhibited at stat-phase. The model predicts active complexes III (CIII) and complex IV (IV) on stat-phase with a drastic decrease. Further, the added alternative oxidase (AOX) reaction is active on stat-phase and reached 0.022 mmol/gDCW/h oxygen consumption rate. Surprisingly, the model also predicts NADH dehydrogenase (NADH2e) activity at stat-phase (0.08 mmol/ www.nature.com/scientificreports/ gDCW/h) whereas this enzyme shows being inactive during exp-phase. This is a hallmark of the uncoupled mitochondrial respiratory pathway as reported by Guerrero-Castillo et al. NADH2e, instead of CI, supplies electrons to AOX through ubiquinone/ubiquinol cycling without contributing to the mitochondria intermembrane proton gradient. This result clearly suggests that this uncoupled alternative pathway is contributing to citrate production, as mainly observed on the stationary phase. We have then assessed this hypothesis proceeding by stimulating Krebs enzymes activity.

Krebs enzymes activity sustain mitochondrial citrate accumulation at stat-phase. When
the simulated growth reaches stat-phase, Krebs enzymes activity seem to sustain mitochondrial citrate accumulation. Pyruvate dehydrogenase (PDH) activity is decreased from 0.672 mmol/gDCW/h at exp-phase to 0.120 mmol/gDCW/h at stat-phase. It remains the dominant mitochondrial Acetyl-CoA supplier, joined with mitochondrial malate dehydrogenase which furnishes oxaloacetate for citrate synthesis (Fig. 2d). Citrate synthase (CS) activity also decreases from 0.612 to 0.120 mmol/gDCW/h. The model predicts inactive aconitase during stat-phase but active, on the reverse way, from isocitrate to citrate, during the stationnary growth phase. NAD-dependent mitochondrial isocitrate dehydrogenase (IDHm) activity is also reversely operating converting oxoglutarate to isocitrate. The cytosolic NADP-dependent counterpart (IDH) is inversely active during growth phases. Both isozymes show increased fluxes during stat-phase associated with citrate production. All the other Krebs enzymes, from oxoglutarate dehydrogenase (OXDH) to fumarase, are active during growth phase and totally inactive on stat-phase (Fig. 2d). Amino acids seem also involved in citrate accumulation since aspartate transamination is a source of oxaloacetate used by citrate synthase for mitochondrial citrate synthesis.
We also investigated mitochondrial-to-cytosolic citrate transport to decipher which pump is solicitated during growth and citrate production phases. Four mitochondrial membrane citrate antiports are identified in the Yali4 model: citrate-phosphoenolpyruvate (CIT-PEP), citrate-oxoglutarate (CIT-OXO), citrate-isocitrate (CIT-ICIT) and citrate-malate (CIT-MAL). The model predicts inactive CIT-PEP on both phases. During growth phase, the model predicts mitochondrial to cytosolic citrate fluxes through CIT-MAL transporter and a lower inverse CIT-ICIT flux contributing to cytosolic isocitrate. During stat-phase, the net cytosolic citrate flux is ensured by CIT-OXO antiport since the positive CIT-MAL flux is totally reversed to mitochondria through CIT-ICIT antiport (Fig. 2e).
Optimization of citrate synthesis. We also intended to investigate metabolic triggers promoting citrate overproduction. For that we simulated higher citrate production rate at stationary phase and identified major effectors (inhibiting or increasing reaction fluxes) of carbon overflow resulting in increased extracellular citrate production. We first restricted our study to Krebs enzymes, ETC complexes and reactions involved in the citrate pathway. Our results show that PDH-PC-CS route is the main contributor of carbon overflow to mitochondrial citrate accumulation. These simulated enzymes activities are proportional to citrate production rate (Fig. 3). Likewise, the mitochondrial aconitase works on the reversed direction and triggers isocitrate to citrate flux. A third phenotype associated with citrate overproduction is an increased extracellular oxygen consumption rate (Fig. 3). This is counterintuitive since less ATP is required during the stationary growth phase and OxPhos complexes activities are decreased, except for the complex I (CI), which increases concomitantly with citrate production rate. Interestingly, the simulated increased CI flux is not accompanied with overexpressed complexes III and IV. Looking more closely at the intracellular reactions, it appears that cytosolic reaction of longchain alcohol oxidation, ensured by fatty alcohol oxidase (FAO), is the main route of ROS production (Figs. 4,5). Intracellular oxygen consumption is balanced by catalase reduction of hydrogen peroxidase (H 2 O 2 ) with a specific 0.013 mmol/gDCW/h flux rate. This result also emphasizes that the NADH2e-AOX alternative route for NADH oxidation and O 2 reduction is inactive when citrate is overproduced during the stationary phase (Figs. 4, 6). Based on these observations, we have then experimentally assessed the effect of AOX inhibition on citrate production in batch cultures of Y. lipolytica in shake flasks in mineral medium at 50 g/L of glucose. When the culture reached the stationary phase, we added 100 µM n-propyl gallate (nPG), a well-known inhibitor of AOX, at 72 h then repeated every 24 h. Experimental results show optimized citrate production in the nPGimplemented medium (Fig. 6). Up to 20.5 g/L of citrate is produced, compared to 10.9 g/L found on control media (Fig. 1), biomass reaches 5.8 gDCW/L and only traces of glucose remain at the end of the culture (Fig. 6). Experimental data are well predicted by the model, except for glucose consumption during stat-phase. To our knowledge this is the first report of Y. lipolytica citrate overproduction on mineral medium based on a culture management strategy designed using mathematical predictions.
Another way to confirm these predictions would have been to add salicylhydroxamic acid (SHAM), another well-known and used inhibitor of AOX 51 . The dFBA model of Y. lipolytica also was also used to predict the effect of adding various Krebs intermediates such as α-ketoglutarate, malate and oxaloacetate for citrate optimization. Whereas simulations reported an improved citrate production (data not showed), these results weren't experimentally confirmed, compared to the control ( Supplementary Fig. SI II).

Discussion
Few studies already reported optimization of citrate in Aspergillus niger and Yarrowia lipolytica by modulating mitochondrial metabolism with specific effector compounds or by cell engineering strategies 52,53 . Our predicted results show that the oxidative phosphorylation (OxPhos) pathway is globally reduced when the cell culture reaches the stationary phase (Fig. 2c). These results were confirmed by a series of oxygraphic experiments (cell respiration) carried out at different growth states, therefore with cells exhibiting different mitochondrial membrane potential (Supplementary Fig. SI. I). We observed that mitochondrial respiration is functional though the alternative NADH2e-AOX pathway whereas complex I (CI) is inactive (Fig. 2c). This is in good agreement www.nature.com/scientificreports/ with the proposed association/dissociation mechanism of NADH2e/CIV channeling electron from the external NADH to cytochrome pathway or towards the alternative oxidase 35 . In agreement with the literature we propose AOX activity as a protective mechanism against mitochondrial ROS production in Y. lipolytica and some other yeasts and plants species 51,54 . Indeed, OxPhos is considered as the main ROS producer, specially through CI and CIII 54 . Our results of simulated intracellular oxygen consumption upon citrate optimization show that fatty   www.nature.com/scientificreports/ acid oxidation may also be a source of ROS production. The long-chain alcohol oxidase (C12) reaction provided by the fatty alcohol oxidase gene (FAO) consumes O 2 at stationary phase and triggers the oxidation of cytosolic alcohols. Theses reaction is accompanied by the production of hydrogen peroxide (H 2 O 2 ) as major ROS species. FAO gene has been recently identified in Y. lipolytica as key in ω-oxidation of long-chain fatty acids 37 . One may consider a global mechanism in Y. lipolytica where citrate is mainly used for lipids accumulation at the entry in stationary phase, followed by ω and β-oxidation of fatty acids upon lipid saturation and exposure to free oxygen species. Fatty acids oxidation is thus thought to trigger cell defense mechanisms such as catalase (CAT) and alternative NADH2e-AOX oxidase to circumvent ROS accumulation (Fig. 4). Further oxygenation of the culture by controlling pO 2 or OTR or inhibition of AOX activity is then expected as a key to trigger citrate overproduction. The proposed mechanism is in good agreement with the work of Kavšček et al., and Sabra et al., showing that the fine modulation of oxygen consumption enzymes in Y. lipolytica favors citrate overflow metabolism. A recent study of citrate production in Aspergillus niger by disruption or overexpression of aox1 gene, in addition to antimycin A during cell growth, showed an improved citrate production in overexpressed aox1 strains compared to the parental strain 55 . Up to 169.1 g/L citrate is reported for one of the overexpressed strains against 158.9 g/L for the parental stain. According to our understanding of the results presented, the highest amount of citrate obtained for the overexpressed aox1 strain should only be attributed to a better germination rate. Moreover, antimycin A compound used in the study is stated as an inhibitor of AOX protein, which is in contradiction with the effect sought by the overexpression of aox1. In the absence of contradicting our study, Hou et al. (2018) confirm the hypothesis that AOX may contribute to clearing ROS species and improving cell survival. Figure 5. Illustrated balance of citrate overflow. Oxoglutarate and aspartate, malate and pyruvate mainly contribute to mitochondrial citrate overflow. ATP citrate lyase (ACL) inactivation allows cytosolic citrate accumulation and citrate production. Fluxes to lipogenesis is globally limited. Fatty acid accumulation from lipolysis is predicted to trigger ROS accumulation. Figure 6. Inhibition of AOX activities by nPG triggers citrate production experimentally (n = 1) and in silico. Up to 20 g/L of citrate is produced during the stat-phase.

Conclusion
We applied dFBA algorithm to the existing Yali4 model of Yarrowia lipolytica W29 strain metabolism to predict favorable metabolic conditions for citrate optimization. The model first predicted carbon overflow to lipid accumulation and citrate production during stationary phase. Y. lipolytica respiration is drastically decreased during citrate production phase and shifts from OxPhos to the alternative ΔΨ-independent pathway ensured by NADH2e/AOX route. Isocitrate dehydrogenase and aconitase are predicted to operate in the backward direction and favor mitochondrial citrate accumulation. Simulated overproduction of citrate is accompanied by inhibition of the NADH2e/AOX pathway as well as increased lipid oxidation and ROS production. The prediction of citrate overproduction has been confirmed experimentally by inhibition of AOX protein. We also challenged the importance of ROS in balancing lipid accumulation and citrate production fluxes by adding glutathione to Y. lipolytica culture in the presence of nPG. Our results showed a global inhibition of citrate production. We propose AOX gene as a target for metabolic engineering and industrial production of citrate in food and pharmaceutical industries. www.nature.com/scientificreports/ Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/.