An in vivo ratio control mechanism for phospholipid homeostasis: evidence from lipidomic studies

Abstract

While it is widely accepted that the lipid composition of eukaryotic membranes is under homeostatic control, the mechanisms through which cells sense lipid composition are still the subject of debate. It has been postulated that membrane curvature elastic energy is the membrane property that is regulated by cells, and that lipid composition is maintained by a ratio control function derived from the concentrations of type II and type 0 lipids, weighted appropriately. We assess this proposal by seeking a signature of ratio control in quantified lipid composition data obtained by electrospray ionization mass spectrometry from over 40 independent asynchronous cell populations. Our approach revealed the existence of a universal ‘pivot’ lipid, which marks the boundary between type 0 lipids and type II lipids, and which is invariant between different cell types or cells grown under different conditions. The presence of such a pivot species is a distinctive signature of the operation in vivo, in human cell lines, of a control function that is consistent with the hypothesis that membrane elastic energy is homeostatically controlled.

1. Introduction

The ability of cells across all the three domains of life to control the compositions of their lipid membranes appears to be critical for normal physiological function and for survival. A growing number of publications details composition changes in a variety of organisms, for example bacteria [1], yeast [2], plants [3], insects [4], fish [5], hibernating mammals [6] and mammalian cell culture lines [7]. Although the processes through which this control is achieved vary across domains and species, the general consensus is that lipid homeostasis is regulated both transcriptionally and biochemically [1]. This duality of control has been documented most extensively in prokaryotes, and is illustrated by the case of Escherichia coli, which regulates tightly the composition of the fatty acids, as well as the distribution of species, that occur within the lipids constituting its membranes.

Transcriptional control of lipid headgroup species in E. coli arises from a range of stress responses that lead to the mutually inverse regulation of genes plsB and dgkA that code, respectively, for the phospholipid biosynthesis enzymes glycerol-3-phosphate acyltransferase and diacylglycerol (DAG) kinase [8]. The same time control of cellular fatty acid composition in E. coli is directed at maintaining a constant ratio of saturated fatty acids (SFAs) to unsaturated fatty acids (UFAs), and is achieved through the fatty acid biosynthesis repressor (FabR), a transcriptional repressor [9–11]. FabR regulates the expression of two enzymes involved in fatty acid biosynthesis, namely β-ketoacyl-ACP synthase I and β-hydroxydecanoyl-ACP dehydratase/isomerase, by binding to specific sequences in the promoter regions of the genes, fabB and fabA, that code for these proteins and repressing transcription. UFAs bound to acyl carrier protein (ACP) or to coenzyme A (CoA) increase the affinity of FabR for its DNA target site, whereas saturated fatty acid ACP (acyl-ACP) or an unsaturated acyl-CoA decreases binding of the regulator to DNA de-repressing transcription of fabA and fabR. Because FabR is in effect a sensor for the ratio of saturated to UFAs present within a cell, then from a control engineering perspective the de novo biosynthesis of fatty acids is regulated at a transcription level through a ratio control feedback mechanism [12]. In contrast, the response of bacteria to abrupt changes in their environment is mediated biochemically through changes in the turnover of existing species. This highly dynamic response, which allows fast adaptation, entails remodelling of chains or headgroups using species from existing lipid pools. An example of this type of response in E. coli is afforded by the maintenance of the ratio between zwitterionic and anionic lipids through the direct activation of phosphatidylserine synthase (PssA) by anionic lipids [10,13]. While the molecular mechanisms involved in the binding of fatty acid CoA or fatty acid ACP to transcriptional repressors for prokaryotes are well understood, the mechanisms through which membrane composition is sensed to achieve regulation at the biochemical level are still unclear.

The lipid compositional complexity of eukaryotic cells is considerably richer than for prokaryotes and archaeotes (viz. more than 800 lipid species in eukaryotes versus less than 100 in prokaryotes). Instances of both transcriptional and biochemical control of lipid homeostasis have been reported. At the transcriptional level, there appears to be at least one transcription factor, sterol regulatory element binding protein (SREBP1), that is modulated by the ratio of saturated and UFAs in a membrane. It has been suggested that this transcription factor could be a key allostatic point of control of membrane lipid homeostasis [14]. In its unactivated state, it is attached to the nuclear envelope and endoplasmic reticulum membranes as a complex with SREBP cleavage activating protein (SCAP) and an ER-embedded protein, insulin-induced gene (InsIG). In the presence of SFAs, the SCAP–SREBP1 complex transits to the Golgi apparatus, where a two-site proteolytic cleavage of SREBP1 occurs, releasing the amino terminus of SREBP1 (H₂N-SREBP1). This fragment enters the nucleus and upregulates the SREBP1 target genes. Conversely, in the presence of UFAs, the InsIG–SCAP–SREBP1 complex is stabilized and remains in the ER [14]. Thus, while SREBP1 appears to achieve a similar regulation to FabR, it does so through a different—and more complex—mechanism. Furthermore, activation of SREBP1 appears to have pleiotropic downstream effects, and has been implicated in the regulation of several genes involved in lipid biosynthesis, for example in de novo lipid synthesis (acetylCo-A carboxylase), fatty acid re-esterification (diacylglycerolacyltransferase), glycerol-3-phosphate acyltransferase, fatty acid desaturation and elongation (stearoyl CoA desaturase), elongation of very long-chain fatty acids and also phosphatidylcholine (PC) biosynthesis, CTP: phosphocholine cytidylyltransferase (CCTα) [14].

Biochemical regulation of lipid homeostasis in eukaryotes is manifested by a significant number of proteins involved in lipid biosynthesis having a strict lipid requirement for activity [15]. Examples of proteins whose functional properties have been shown to be modulated by the lipid membranes with which they associate include phospholipase A2 (PLA2), phospholipase C (PLC), phospholipase D and CCTα [15–17]. However, as in the case of prokaryotes, the molecular/biophysical mechanisms through which eukaryotic cells sense lipid composition are still the subject of debate, as is the nature of the biophysical property of biological membranes that is being regulated. What is clear is that because neither the absolute nor the relative lipid compositions are conserved in response to external stimuli, then homeostatic regulation is most unlikely to be targeted at individual molecular species. This is consistent with the high metabolic cost that would be incurred in the active control of the more than 800 chemically distinct lipid species that are present in a typical eukaryotic cell membrane. It is also consistent with eukaryotic cells having a constantly primed distributed system that can respond rapidly to changes in the local requirement for lipids, and that is decoupled from a slower centralized system that involves transcription control. Taken together these empirical observations suggest that the lipid compositional changes that are typically observed in response to rapid environmental changes reflect a more global, or supramolecular, property of eukaryotic lipid membranes that is under active regulation. Evidence for a homeostatic control mechanism comes from studies where key enzymes that synthesize phospholipids are either up- or downregulated, and the resulting effects on phospholipid composition. For example, CCT a key protein in the biosynthesis of PC [18] has been overexpressed in cells, and although the rate of PC synthesis increases, there is a concomitant increase in the turnover of PC. The net result is that the PC composition is unchanged [19–21]. Similarly, phosphatidylserine (PS) synthesis is catalysed by two different enzymes, PssA-1 and PssA-2. Knockout studies in mice show that mice deficient in the gene coding for PssA-2 develop normally [22], and that PS composition is normal in such mice with PS synthesis being accomplished by upregulation of PssA-1 [23]. Genetic transformation of bacteria such as E. coli shows similar behaviour. Amplification of PssA [24] or PS decarboxylase [25] has no significant effect on membrane phospholipid composition. Of particular interest is the fact that many of these enzymes are extrinsic proteins with significantly higher activity when associated with lipid membranes. Anionic phospholipids such as phosphatidylglycerol have an activating effect on PssA [26] and CCT [27], leading to suggestions that phospholipid homeostasis might be regulated by the balance of membrane charge.

There are currently relatively few hypotheses that address possible mechanisms that underpin the way in which lipid diversity might be regulated at a supramolecular level. The theory of homeoviscous adaptation suggests that membrane viscosity or fluidity is regulated homeostatically. Evidence of such behaviour was first reported in E. coli, where it was found that upon growth at elevated temperatures E. coli cells incorporated increasing proportions of saturated and long-chain fatty acids into their membrane lipids. This change in membrane composition resulted in identical membrane viscosities being observed [28]. Homeoviscous adaptation has since been reported in fish [29], deep sea animals [30], mammalian hibernation [31] and extended as reviewed [32] and critiqued [33]. Another well-developed theory is the intrinsic curvature hypothesis. In broad terms, the intrinsic curvature hypothesis of membrane lipid composition [34] postulates that cells maintain a constant net spontaneous curvature of each leaflet of the bilayer. Initial support for this hypothesis came from the observation that in the bacterium Acholeplasma laidlawii the membrane lipid composition was controlled so as to maintain a constant transition temperature from the fluid lamellar (L_α) to inverse hexagonal (H_II) phase [35]. Subsequently it was also found that the radii of curvature of total lipid extracts taken from A. laidlawii cells with different lipid compositions had similar values [36], and this led to the proposal that lipid regulation in this organism occurs through a ratio control system that senses the balance of bilayer-forming to non-bilayer lipids [37,38]. In the case of eukaryotic cells, the many observations of the dependence of the functional properties of membrane-associated proteins on the lipid compositions of the membranes with which they associate have been interpreted to reflect a similar type of control. We note that proteins that appear to be regulated by lipid bilayer compositions are not restricted to those involved in lipid biosynthesis, and include proteins that are associated with membrane trafficking (e.g. proteins with the Bin–Amphiphysin–Rvs domain) [16].

On the basis of these observations, it has been suggested that curvature elastic energy of a bilayer (which depends on not just the mean curvature of each monolayer, but also on the monolayer bending rigidity) could provide a more fundamental biophysical mechanism for how the functions of membrane-bound proteins might be affected by lipid composition. This type of effect was demonstrated for the enzymatic activity of CCTα [39], a translocation protein that is rate determining in the Kennedy pathway for PC biosynthesis [18]. It was postulated subsequently that several enzymes involved in lipid biosynthesis might also be regulated by membrane-stored curvature elastic energy; kinetic simulations of the eukaryotic lipid biosynthetic pathway were used to show how this elastic energy could be homeostatically controlled through a ratio control mechanism [40,41] that is analogous to that proposed for A. laidlawii [38].

Membrane-stored curvature elastic energy is related to the torque tension (τ), which arises because two monolayer leaflets with intrinsic curvature are forced flat in a bilayer such that each monolayer wishes to bend, but is prevented from doing so by being apposed to the other monolayer. τ is determined by the monolayer spontaneous curvature (c₀) and the bending rigidity (κ_M) [42]:

1.1

Experimentally, c₀ and κ_M may be estimated using small angle X-ray diffraction (SAXRD) techniques [43,44]. However, the time-consuming nature of these experiments has limited the number of systems for which accurate data are available. The broad typological behaviour of lipids can be mapped from phase diagrams constructed using polarizing optical microscopy and SAXRD experiments, and from these it has been possible to develop qualitative narratives of how relative changes in spontaneous curvature relate to changes in headgroup structures or to hydrocarbon chain length and asymmetry [45], degree of unsaturation [46], and position of unsaturation [47]. We note, however, that these qualitative relationships between spontaneous curvature and lipid structure become increasingly complex to predict with increasing ionic strength and pH, as exemplified by the case of PS [48], which is a type II lipid at low pH and type 0 at higher pH. This is a serious barrier to estimating the spontaneous curvature contribution to the curvature elastic energy for complex lipid mixtures under in vivo conditions. The situation is further complicated by the effects that organic species [49] and biopolymers [50] have on the interfacial curvatures of aggregates, as well as the effect of pairing of lipid species through specific interactions [51] or through bridging cations [52]. Because of these factors, it is very difficult to estimate a priori, even qualitatively, the relative magnitude of c₀ for a monolayer with a given lipid composition, and hence this approach is not amenable to predictions of membrane curvature elastic energy.

In this paper, we present an approach, akin in spirit to data-driven modelling, which seeks to identify the presence in lipidomic datasets of a universal characteristic that would be expected if membrane curvature elastic energy were regulated within tight bounds. We used lipidomic datasets obtained by electrospray ionization mass spectrometry (ESI-MS) to determine whether the signature of ratio control regulation [12], namely a control parameter, p_x, that is related to the ratio of weighted concentrations of type II to type 0 lipids [40], is consistent with a range of independent datasets from different cell lines (HL60 and HeLa) and from cells grown under different conditions (oleate-enriched HL60 versus normal HL60). This control parameter is a proxy for the stored curvature elastic energy, which as discussed above is difficult to determine a priori. To identify this signature, we addressed the following question: given the set of all possible lipid species {L}, is there a universal pivot species, L_p, that defines how the set can be partitioned into subsets {L₀ + L_P}, {L_II} such that the ratio control function p_x, defined as

1.2

has consistently the minimum variance across our whole range of independent lipidomic datasets?

In equation (1.2), [L_II,n] denotes the concentration of the type II lipid n, as determined experimentally and w_n is a weighting factor for lipid n, [L_0,m] is the concentration of the type 0 lipid m, and w_m is the weighting factor for lipid m. The variables a and b are the total numbers of type 0 and type II lipids, respectively. In other words, for a given lipidomic dataset equation (1.2) asks ‘what is the value of p_x if one set of lipids is nominally type II and the remaining lipids are nominally type 0?’ If p_x is homeostatically conserved then p_x values derived for other biological repeats of the same cell type ought to be identical within error.

It is unclear, in vivo, where to draw the line that demarcates type II from type 0 lipids, so we use the lipidomic datasets to identify the lipid species that marks the boundary between type II and type 0 lipids, and to determine whether this species is the same across the different lipidomic sets. To do this, we construct a coarse-grained metric of type II character that is an additive function, which takes into account the headgroup type as well as the number of unsaturated bonds in each of the acyl chains. This metric is defined by equation (2.1), as described in §2.

2. Material and methods

2.1. Cell culture protocols

HeLa S3 (H2B–GFP fusion) cells [53] were gifted by Dr D. Zink (LMU Munich) and cultured at 37°C and 5 per cent CO₂ in Dulbecco's modified Eagle's medium (D-MEM , Gibco) containing 4 mM l-glutamine and HEPES, supplemented with 10 per cent foetal calf serum (FCS; Gibco) and 5 per cent antibiotic/antimycotic (Gibco). HL-60 cells were cultured at 37°C in 5 per cent CO₂ in RPMI medium with Glutamax-1 and HEPES (25 mM) supplemented with 10 per cent FCS and 5 per cent antibiotic–antimycotic solution (10 000 units ml⁻¹ penicillin G sodium, 10 000 μg ml⁻¹ streptomycin sulphate, 25 μg ml⁻¹ amphotericin B as Fungizone in 0.85% saline). Counts of viable cells were performed using a haemocytometer; aliquots of cells were mixed with trypan blue solution, and cells that stained blue were assumed to be non-viable.

The adherent HeLa cell line was detached and split by incubating with trypsin–EDTA (5 ml, 0.25%, Gibco) at 37°C for 5 min, after gentle shaking to remove cells from the flask surfaces, trypsin was deactivated by addition of 10 ml of D-MEM, and cells were harvested by centrifugation. HeLa cells were blocked at the G1/S boundary by adding l-mimosine (Sigma–Aldrich) to the culture flasks at a final concentration of 200 µM and incubating for 16 h. Release from the block was facilitated by removing the culture medium, washing with PBS and supplementing with fresh culture medium.

2.2. Phospholipid extraction

Phospholipids were extracted from whole-cell samples using a chloroform (JT Baker)/methanol (Fisher Scientific) extraction protocol first described by Bligh & Dyer [54]. Each sample was made up to 800 µl with saline (0.9% NaCl, Sigma–Aldrich) and transferred to a glass centrifuge tube. Internal standards were added. Chloroform (1 ml) was added followed by methanol (2 ml), chloroform (1 ml) and finally water (1 ml); after each addition, the sample was mixed by vortexing. After centrifugation (10 min at 1500 r.p.m., 4°C), the lower organic layer containing the phospholipid was removed and dried under nitrogen. Samples were stored at −20°C until analysis by mass spectrometry.

2.3. Mass spectrometry

ESI-MS of phospholipids extracted from whole cells was performed on a Micromass Quatro Ultima triple quadrupole mass spectrometer (Micromass, Wythenshaw, UK) equipped with an ESI interface. Samples were dissolved in butanol : methanol : water : ammonia (conc.) (2 : 6 : 1.6 : 0.4; Fisher Scientific) and introduced into the mass spectrometer by a syringe pump at a flow rate of 5 µl min⁻¹. Nitrogen was used as the cone and desolvation gas and argon as the collision gas (3.5 × 10⁻³ mbar). PC, phosphatidylethanolamine (PE) and DAG molecular species were detected under positive ionization. PC molecular species were quantified from diagnostic precursor ion scans of the m/z = 184 fragment ion (P184) with a cone voltage of 90 V and collision energy of 32 eV. PE molecular species were quantified from neutral loss fragments (NL141), collision energy 28 eV, cone voltage 90 V. DAG molecular species were detected from a NL35 fragment with collision energy 15 eV and cone voltage 50 V. Phosphatidylinositol (PI), PS and phosphatidic acid (PA) molecular species were acquired in negative ionization. PI lipids were quantified from diagnostic precursor ion scans of the m/z = 241 fragment ion (P241) with a cone voltage of 100 V and collision energy of 39 eV. PS molecular species were quantified from neutral loss fragments (NL87) with collision energy 22 eV and cone voltage 80 V. PA molecular species were quantified from P153 fragments from 34 eV and a cone voltage of 90 V. Data were acquired and processed using MassLynx NT software and a macro developed in house. MS spectra were smoothed, baseline-subtracted, converted to centroid format and exported to individual Excel sample files. These were imported into the analyser program, and corrected for the [M+1]⁺and [M+2]⁺ ions generated by the approximately 1 per cent natural abundance of ¹³C. Detailed mass spectrometry methodology has been published previously [55].

Quantification of the different lipids classes was achieved by adding known quantities of non-physiological phospholipids (internal standards) to lipid extracts during the Bligh Dyer extraction [54]. The amount of internal standard added to each sample was chosen to give a peak height similar to that of the other lipid species in the mass spectrum. Internal standards were prepared using accurately weighed quantities of solid lipids that were dissolved in chloroform. PC was quantified using 1,2-diarachidoyl-sn-glycero-3-phosphocholine, PE was quantified using 1,2-dimyristoyl-sn-glycero-3-phosphoethanolamine, PS was quantified using 1,2-dimyristoyl-sn-glycero-3-phosphoserine, PI was quantified using 1,2-palmitoyl-sn-glycero-3-phosphoinositol; all lipids were purchased from Avanti Polar Lipids. Prior to analysis by mass spectroscopy, instrument stability was assessed by running a standard mixture containing known quantities of lipids. We use this approach to ensure that any variations that are seen in our samples can be attributed to the biological origin, rather than instrumental origin.

2.4. Determination of pivot species and calculation of ratio control parameter p_x

The use of lipidomic datasets to identify which lipid species should go in the numerator of equation (1.2) and which should go in the denominator is based on a metric of the type II character of a lipid. This coarse-grained metric provides the factors (w) that weight the contribution of the concentration of each lipid species calculated by assuming a linear contribution from each of the two chains and from the headgroup to the type II characteristic of the lipid as prescribed by equation (2.1)

2.1

where c_h1 is the contribution from chain 1, c_h2 is the contribution from chain 2 and h_g is the contribution from the headgroup. Values of c_h1, c_h2 and h_g were generated iteratively from a linear distribution of random numbers between 0.01 and 1.00, a range of three orders of magnitude. The values of c_h1 and c_h2 were ranked to reflect the number of unsaturated bonds in the chain such that c_h(0) < c_h(1) < c_h(2) < c_h(3) < c_h(4) < c_h(5) < c_h(6), where the bracketed subscript denotes the number of unsaturated bonds. Headgroup contributions to the weighting factor were ranked in the order PS(h_g) < PC(h_g) < PI(h_g) < PE(h_g) < PA(h_g) < DAG(h_g).

The iterative process used to identify the occurrence of a unique pivot species is shown schematically in figure 1; a detailed scheme is shown in the electronic supplementary material, figure S1. Proxy values for the membrane-stored elastic stress, i.e. the ratio control function (p_x), were determined by using equation (1.2). One of the lipids was selected as the pivot species, L_p, such that all lipids whose weighting factors were greater than for the pivot species were assigned to the {L_II} set, with the pivot species itself and the remaining lipids with lower values of w assigned to the {L₀} set. Taking each lipid species in the lipidomic dataset as the pivot species in turn, we computed the value of p_x for 10⁶ different sets of the parameters c_h1, c_h2 and h_g, before using the next lipid species as the new pivot and regenerating 10⁶ sets of c_h1, c_h2 and h_g. This process was repeated until all of the lipids in the lipidomic dataset were trialled as a pivot species. The decision to reject parameter sets of c_h1, c_h2 and h_g, and ultimately pivot species, was determined by comparing the p_x values generated from equation (1.2) for multiple biological repeats of the same cell line, if these p_x values were similar, i.e. had a coefficient of variance that was greater than our arbitrary threshold, they were rejected. Initially this process was carried out for the HL60_oleate cell data, which comprises the ‘training’ dataset. Retained parameter sets and their respective pivot species were then put back into equation (1.2), but this time with lipid concentrations from the lipid composition data from HL60_normal cells; p_x values were recalculated across a series of biological repeats and c_v values were computed. Similarly p_x values were generated using lipid concentrations from multiple biological repeats of HeLa cells and the associated c_v values were calculated.

Figure 1.

Overview of the hierarchical process used to identify the existence of a universal pivot species that would be the signature of homeostatic control of stored curvature elastic energy in vivo. (Online version in colour.)

The concentrations of the lipids in each of the two sets {L₀} and {L_II}, as determined experimentally by ESI-MS for HL60 cells enriched in oleate (HL60_oleate), together with their weighting factors, were used to calculate a value of p_x for a given experiment. To avoid artefactual effects owing to biological variation and threshold effects, experimental data for a single flask of cells were used to calculate p_x for that flask. Typically 10–20 flasks (x) of cells were used to generate an average p_x value. The arithmetic mean (μ) and standard deviation (σ) values of p_x for all the flasks (x) were compared using the coefficient of variation (c_v), where c_v = (σ/μ) × 100%. Parameter sets that did not achieve less than a certain cut-off c_v were discarded, and the process of selection of pivot species, generation of coefficients and weighting factors, and calculation of p_x was repeated iteratively. The data from the HL60_oleate cells may be considered as analogous to a ‘training’ set. The sets of coefficients that resulted in p_x values that were below the cut-off for the HL60_oleate data were then used to compute p_x values using ESI-MS data from HL60 cells grown under normal conditions (HL60_normal), and from HeLa cells (refinement datasets). By rejecting p_x values above the cut-off conditions we obtained a set of coefficients that enabled us to identify a universal pivot species.

3. Results

3.1. Lipidomics of total cell extracts from asynchronous cell cultures

Comparing HL-60 cells grown in the presence of oleate (HL60_oleate) in the growth medium with cells grown in normal medium (HL60_normal) highlights the range of changes in lipid composition that might be expected to occur as a result of adaptation to alterations in the growth environment. Changes in the membrane lipid composition of cells, caused by adaptation to fatty acids added to culture media, have been recognized for many years [7]. In our study, we see oleate, a C18 : 1 cis fatty acid, directly incorporated into the cells, enriching the production of lipid species around the C18 : 1 chain.

Figure 2a shows the endogenous PC composition of three asynchronous cell cultures; HL-60 cells cultured in FCS (HL60_normal) and HL60 cells cultured in FCS enriched with oleate (HL60_oleate), and HeLa cells cultured in FCS (HeLa). In the case of the two HL60 cultures, there are some differences in the dominant PC species. The three dominant PC species in the HL60_normal cultures are PC16 : 0/18 : 1 (approx. 21%), PC16 : 0/16 : 1 (approx. 13%) and PC18 : 1/18 : 1 (approx. 13%). In the case of the HL60_oleate cultures, the three dominant species are PC18 : 1/18 : 1 (approx. 28%), PC16 : 0/18 : 1 (approx. 25%) and PC16 : 1/18 : 1 (approx. 5%). Similarly, figure 2b shows that the three dominant DAG species for HL60_normal cultures are DAG 18 : 1/18 : 1 (approx. 22%), DAG 16 : 0/18 : 1 (approx. 16%) and DAG 16 : 1/18 : 1 (approx. 12%), whereas for HL60_oleate they are DAG 18 : 1/18 : 1 (approx. 42%), DAG 16 : 0/18 : 1 (approx. 22%) and DAG 16 : 1/18 : 1 (approx. 5%). These results show that, as expected, exogenous oleate causes a significant change in the lipid composition of HL60 cells. In the case of the DAG lipids, the changes are dramatic: in the presence of oleate, the DAG 18 : 1/18 : 1 rises from around 21–45% of the DAG composition detected in the samples. In addition, oleate increases the total DAG in the HL60_oleate cells such that the ratio of total PC to total DAG is approximately 3.5 : 1, compared to approximately 8.1. Analogous changes are seen in other phospholipid fractions (see the electronic supplementary material, figure S2).

Figure 2.

The endogenous PC (a) and DAG (b) composition of a single biological repeat of HL60_normal cells, HL60_oleate and HeLa S3 cells. White bars show HL60_normal cells, grey bars show HL60_oleate, whereas black bars show HeLa S3 data.

Typical cell-type-dependent variations in lipid composition are highlighted by comparing the data obtained from HeLa cells, which is an adherent cell line, with data from HL60 cells (HL60_normal). The dominant PC and DAG lipids in the HeLa cells are PC16 : 0/18 : 1 (approx. 28%), PC18 : 1/18 : 1 (approx. 15%), PC16 : 0/16 : 1 (approx. 6%), DAG 16 : 0/18 : 1 (approx. 26%), DAG 18 : 0/18 : 1 (16%) and DAG 18 : 1/18 : 1 (approx. 14%). By contrast, for HL60_normal cells, the dominant PC and DAG lipids are PC16 : 0/18 : 1(approx. 21%), PC16 : 0/16 : 1(approx. 13%), PC18 : 1/18 : 1(approx. 13%), DAG 18 : 1/18 : 1 (approx. 22%), DAG 16 : 0/18 : 1 (approx. 16%) and DAG 16 : 1/18 : 1 (12%).

The differences observed between the lipid composition of HeLa cells and HL60_normal as well as between HL60_oleate and HL60_normal are sufficiently robust and reproducible that they should reveal the signature of lipid composition being regulated by a ratio control function, directed at the regulation of stored curvature elastic energy, if this does indeed occur in vivo.

3.2. Assumptions used to constrain the search for a universal pivot species

In order to identify a signature of ratio control of lipid composition, we adopted an approach that is analogous to data-driven modelling. Having established a rigorous biophysical framework as a hypothesis to underpin the origin of this regulation, we used this to constrain the functional form of the ratio function as indicated by equation (1.2). Our approach involves two sets of explicit assumptions, namely assumptions relating to the experimental data and assumptions deriving from biophysics of regulation of stored curvature elastic energy.

Assumptions relating to the experimental data are

• — that the arbitrary threshold noise correction/cut-off of 95 per cent of the total ion count we chose for the ESI-MS data is sufficient to exclude artefacts owing to noise,

• — that the lipids we have quantified represent more than 90 per cent of the total phospholipid, by amount in the whole cells, and

• — that because we cannot quantify the relative amounts of a small number of isomeric or isobaric species, we assign and analyse for the most dominant isoforms.

The assumptions based on the biophysics of the putative control mechanism are

• — the stored elastic energy increases in the order PS < PC < PI < PE < PA < DAG,

• — stored elastic energy increases with the number of unsaturated bonds in a lipid's hydrocarbon chain,

• — DAG lipids are classified exclusively as belonging to the {L_II} set, and

• — the functional form of the control function is a ratio whose numerator and denominator are linearly dependent on the contributions of the individual lipid components.

The ranking assumed for the headgroup contribution to the stored elastic energy (h_g) is supported by spontaneous curvature and bending modulus data, as well as phase behaviour data. This is highlighted by comparing the spontaneous radii of curvature (R₀) of dioleyl containing lipids, where c₀ is 1/R₀. In water, R₀ values follow the order of 1,2-dioleyl-sn-glycerophosphoserine (DOPS; +144 Å) [44], 1,2-dioleyl-sn-glycerophosphocholine (DOPC; −140 Å) [45], 1,2-dioleyl-sn-glycerophosphate (DOPA; −130 Å) [56], 1,2-dioleyl-sn-glycerophosphoethanolamine (DOPE; −27 ± 1 Å) [56] and 1,2-dioleyl-sn-glycerol (DOG; −11 Å) [57]. In the presence of isotonic saline, electrostatic screening causes a change such that c₀ for DOPA decreases to −46 Å [56] with only a minor change observed for the zwitterionic DOPE (−23 ± 1 Å) [56]. The bending modulus (κ_M) for DOPC is approximately 13 k_BT [58]; for DOPS, it is approximately 11k_BT [44] and similar to DOPE (approx. 9 k_BT) [57], which is unchanged by the addition of up 30 mol% DOG to DOPE [57]. Therefore, because stored elastic stress (τ), equation (1.1), is dependent on κ_M and c₀, and κ_M varies little across these species, we can be confident that the contribution to membrane-stored elastics stress (for identical chain structures) follows the broad trend PS < PC < PE < DAG. κ_M for DOPA is unreported in the literature; however, Lee et al. [59] have shown that the propensity of DOPA to form inverse hexagonal phases when mixed with 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) is greater than that of DOPE, a result that indicates that DOPA is more type II than DOPE, a conclusion supported by other work [60]. 1,2-dioleyl-sn-glycerophosphoinositol (DOPI) lipids have not been extensively characterized owing to difficulties in obtaining pure lipid samples. However, a thorough study suggests that DOPI mixtures are more type II than DOPC and less type II than DOPE [61].

The assumption that the contribution to the stored elastic stress from the hydrocarbon chain components (c_h1 and c_h2) increases with the number of unsaturated bonds in each fatty acid chain is also based on experimental observations. An illustrative example is the suppression of the bilayer to inverse hexagonal phase transition occurs with increasing unsaturation from PE 16 : 0/18 : 1 (L_α to H_II = 75°C) with one unsaturated bond, to PE 18 : 3/18 : 3 (L_α to H_II =−25°C) with six unsaturated bonds, to PE 20 : 4/20 : 4 (L_α to H_II =−30°C) with eight unsaturated bonds [46]. We note that we have not included overall chain length, or the position of the unsaturation, in our calculations. This leads to a coarse-graining effect such that increasing chain length as well as unsaturation position have no effect on the elastic energy of a membrane.

3.3. Identifying a universal pivot species

As indicated earlier, the signature feature of ratio control arising from stored elastic energy would be the existence of a pivot species (L_p), i.e. a lipid species that divides the set of lipids {L} into the subsets {L₀} and {L_II}, that is invariant for different cell types and cells grown under different conditions. To determine a set of candidate pivot species, equation (1.2) was used to calculate values of p_x using the lipid composition data for asynchronous HL60_oleate cells. Only parameter sets for which the resulting p_x had a coefficient of variation (c_v) of below 15 per cent (denoted by ) were retained. The parameter sets corresponding to were then used with lipid composition data from asynchronous HL60 cells (HL60_normal) and with lipid composition data from HeLa cells to obtain two further sets of p_x, denoted by and that satisfied the 15 per cent c_v cut-off. The mean c_v, i.e. the arithmetic mean of c_v(oleate), c_v(normal) and c_v(HeLa) of each of these sets was calculated to find parameter sets that consistently returned p_x values with low c_v across all three lipidomic datasets.

Each of the ca 150 lipids present in the lipidomic datasets was trialled as a pivot species, but only 26 led to p_x values that satisfied the criteria of an initial c_v(oleate) of less than 15 per cent and mean c_v with a variance of less than 10 per cent. Figure 3a shows pivot species that gave the lowest mean c_v values with a variance of less than 10 per cent. These results suggest that the pivot species that gives the lowest c_v, with the least variance, across each of the cell types/conditions is the PE species with one unsaturation (PE 0 : 1). Figure 3b shows the p_x values for the HL60_oleate and HL60_normal cells. For comparison, p_x values are shown for two parameter sets, one leading to c_v approximately 10 per cent and the other to c_v 30–40%. Figure 3b highlights that the parameter set with c_v approximately 10 per cent gives similar values of p_x from the two very different lipid composition data, whereas the set with the higher c_v does not.

Figure 3.

(a) Initial parameter evaluation of ratio control functions. 1 × 10⁴ iterations of equation (1.2) for each pivot species were explored. Each of the ca 150 lipids in the total lipid extract was trialled as a potential pivot species. The coefficient of variation c_v (c_v = (σ/μ) × 100%) was used to restrict solutions to within a cut-off c_v value. Initial data were obtained for HL60_oleate cells with a cut-off of 15 c_v(oleate) of 15%, where the number of biological repeats (S) = 16, parameter sets were then tested against the HL60_normal (S = 16) and HeLa lipid composition data (S = 8) to generate the c_v(normal) and c_v(HeLa). (a) Shows the lowest mean c_v (i.e. the arithmetic average of c_v(oleate), c_v(normal) and c_v(HeLa)) for those pivot species where the standard deviation of the mean c_v was less than 10%. Error bas show the standard deviation of the mean c_v. (b) Shows the individual p_x value of each of S flasks of cells (biological repeats). Black circles and crosses depict a successful parameter set for HL60_normal and HL60_oleate cells respectively. Black triangles and pluses depict unsuccessful ratio functions for HL60_oleate and HL60_normal cells respectively, which were discarded because c_v values were in excess of the cut-off value.

Although species PE 0 : 1 emerges as the most likely candidate for the universal pivot species, it is clear from figure 3a that several other lipid species might also be good candidates. In order to differentiate between these species, we explored 10⁶ new parameter sets in conjunction with the three main lipidomic datasets (HL60_oleate, HL60_normal, HeLa), and assessed six species as possible pivot species. These were PE 0 : 0, PE 0 : 1, PE 1 : 1 and PA 0 : 0, PA 0 : 1, PA 1 : 1. The results of these calculations are shown in figure 4, where the mean c_v for the three different cell types is shown plotted against the standard deviation of the mean c_v.

Figure 4.

Evaluation of different candidate pivot species. Six different pivot species that from figure 3 emerged as possible candidates for universal pivot species. These were PA (a) 0 : 0, (b) PA 0 : 1, (c) PA 1 : 1 and (d) PE 0 : 0, (e) PE 0 : 1, (f) PE 1 : 1. Plots show the mean c_v (i.e. the arithmetic average of c_v(oleate), c_v(normal) and c_v(HeLa)) versus the standard deviation of the mean c_v.

4. Discussion

The data in figure 4 confirm the existence of a universal pivot species, namely PE 0 : 1, because this gives the lowest mean c_v with the lowest standard deviation. The effectiveness of the different PE species as pivot species follows the order PE 0 : 1 (mean c_v ∼10%) is more effective than PE 0 : 0 (mean c_v ∼13%), whereas PE 1 : 1 does not produce a mean c_v below 15 per cent. For the PA species, PA 0 : 1 leads to a mean c_v of just under 10 per cent and consequently this species PA 0 : 0 might be considered to be a better candidate for a universal pivot species than PE 0 : 1. However, we note that the lower c_v values for PA 0 : 1 correspond to higher standard deviations than for PE 0 : 1. What this means in practice is that while PA 0 : 1 leads to an overall lower c_v, it does not do so consistently across all datasets. By contrast, PE 0 : 1 combines low c_v and low standard deviation, thereby satisfying the criteria required of a unique pivot species. Tables 1 and 2 shows parameter sets, p_x and c_v values that have given the lowest five mean c_v values from figure 4, restricted to values where the standard deviation of the mean c_v is less than 10 per cent.

Table 1.

The five parameter sets giving the lowest mean c_v values from figure 3, restricted to c_v values, where the standard deviation of the mean c_v is less than 10%.

no	ch(0)	ch(1)	ch(2)	ch(3)	ch(4)	ch(5)	ch(6)	PS (hg)	PC (hg)	PI (hg)	PE (hg)	PA (hg)	DAG (hg)	pivot species	Lp	cv(mean)
1	0.30	0.43	0.57	0.72	0.74	0.85	0.91	0.67	0.75	0.76	0.77	0.89	0.97	PE 0 : 1	0.56	10.00
2	0.29	0.39	0.58	0.72	0.73	0.9	0.97	0.56	0.56	0.56	0.58	0.65	0.71	PE 0 : 1	0.39	10.04
3	0.26	0.37	0.5	0.51	0.76	0.79	0.9	0.71	0.72	0.73	0.75	0.86	0.91	PE 0 : 1	0.47	10.05
4	0.26	0.37	0.57	0.75	0.78	0.85	0.91	0.66	0.7	0.72	0.74	0.78	0.96	PE 0 : 1	0.47	10.15
5	0.19	0.3	0.41	0.43	0.58	0.78	0.97	0.58	0.63	0.64	0.64	0.71	0.84	PE 0 : 1	0.31	10.18

Table 2.

c_v and p_x values derived from parameter sets displayed in table 1.

no	cv(oleate)	cv(normal)	cv(HeLa)	px(oleate)	px(normal)	px(HeLa)
1	10.33	10.70	8.97	1.02 ± 0.11	0.84 ± 0.10	0.46 ± 0.04
2	10.04	10.84	9.25	0.50 ± 0.05	0.43 ± 0.05	0.23 ± 0.02
3	9.86	10.85	9.42	0.63 ± 0.07	0.73 ± 0.07	0.33 ± 0.03
4	10.45	10.97	9.04	0.73 ± 0.08	0.62 ± 0.07	0.32 ± 0.03
5	10.15	10.89	9.49	0.34 ± 0.04	0.30 ± 0.03	0.15 ± 0.02

Tables 1 and 2 show some interesting and reproducible features. First, when considering the different headgroups a clear trend appears in the data. From table 2, it is evident that in all five examples the relative headgroup contributions to the ratio control function, which is a proxy for the stored elastic energy, follow the same pattern, i.e. the PC, PI and PE headgroups make almost identical contributions. The contribution from PS is slightly less than is the case for PC, PI and PE. However, PA and DAG make significant contributions to the stored elastic energy. Second, changes in contributions to stored elastic stress owing to unsaturation in the hydrocarbon chains are also reproducible. Also of interest is the observation that the magnitude of the ratio control function, p_x, is within error the same for HL60_oleate and HL60_normal, but this value is consistently around a factor of two greater than that observed in HeLa.

Figure 5 shows the relative contributions, w, of selected lipids to the ratio control function calculated using parameter set 5. It is interesting to note that for the same degree of unsaturation, w_DAG is only a factor of approximately 3.3 higher than w_PC. From literature values of c₀ and κ_M, presented earlier, we would expect that DAG should increase the stored elastic energy of a membrane by a factor of 8 relative to PC. The discrepancy between these numbers highlights the fundamental difficulty of applying physical parameters from in vitro single-component systems to in vivo data. In particular, we note that, in vivo, the presence of large numbers of proteins, particularly transmembrane proteins, and cytoskeletal elements with a range of compressibilities means that κ_M is likely to be dominated by the protein component of the membrane.

Figure 5.

(a) Relative in vivo contributions to stored elastic stress (w) for the most common lipids found in mammalian cells derived from parameter set 5, table 1 and equation (2.1). The pivot species (PE 0 : 1) marks the boundary between the sets of type II and type 0 lipids. (b) Relative in vivo contributions (w) to stored elastic stress bases for different headgroups with the 1 : 1 chain.

Figure 5 provides a further insight into the position of PE 0 : 1 as a universal pivot species. The dotted line represents the magnitude of w_{PE 0 : 1}; it is clear that w_{PI 0 : 1}, w_{PC 0 : 1} and w_{PS 0 : 3} are similar in magnitude. Although the species leading to the lowest mean c_v is still PE 0 : 1, the proximity of these w values suggests that in situations where a cell is unable to produce PE 0 : 1, then one of these other lipid species could define the {L₀} and {L_II} lipid sets used in the ratio control function.

We have presented a new methodology for interrogating quantitative lipid composition datasets in order to uncover the occurrence of lipid regulation in vivo through a ratio control function. The biophysical framework underlying the hypothesis that cells maintain the stored curvature elastic energy within tight bounds leads to the expectation that lipid composition might be regulated through a ratio control function in which type II lipids appear in the numerator, whereas type 0 lipids appear in the denominator of the function. Instead of attempting the currently impossible task of ‘synthesizing’, the overall membrane-stored elastic energy by summing the contribution of each of the 140 or so lipids, we have opted to use lipid composition data from over 40 different asynchronous cell populations, with different lipid compositions, to generate coefficients that reflect these contributions as manifested in vivo. Central to this approach is the existence of a universal pivot lipid species that marks the boundary between type II and type 0 lipids. Using lipid composition data from HL60 cells grown under two different conditions and from HeLa cells we have found that PE 0 : 1, corresponding in vivo to POPE, emerges as a universal pivot species.

It is worth considering the extent to which this mechanism is compatible with the other theories of phospholipid homeostasis. Homeoviscous adaptation, as we have previously noted, is the process where membrane viscosity is preserved. Because the viscosity of the hydrocarbon region of a membrane is related to the stored elastic energy, it is not currently possible to determine experimentally which of these two physical parameters is in fact sensed and regulated. The difficulty in basing an analysis on viscosity is that there is no simple way of relating how different molecular structures contribute to the overall viscosity and hence, it is not possible to construct a metric for use in equation (1.2). We have also considered that membrane surface charge might be controlled homeostatically. In principle, a metric that produces weighting factors, based on the contribution of different headgroups to the surface potential could be proposed, allowing the lipidomic data to be interrogated for evidence of a control function that is based on electrostatic interactions. Clearly, there is no reason why multiple mechanisms of phospholipid homeostasis should not occur simultaneously within cells, and in future work we intend to extend our data-driven modelling to look for the signature of other membrane properties that might be regulated in vivo.

The signature of regulation of lipid composition through a ratio control function is the observation that the lipidomic data in our studies all point to the same lipid species, PE 0 : 1, as marking the division between the lipids that contribute to the numerator and denominator of equation (1.2). It is interesting to note that, in general, it may be possible for this particular species to be absent from the lipidome of some organism without invalidating the existence of a ratio control function. As can be seen from figure 2, a number of species (e.g. PA 0 : 0, PA 0 : 1, PA 1 : 1, PE 0 : 0 or PE 1 : 1) have values of w that are comparable with that of PE 0 : 1. Consequently, in the absence of PE 0 : 1, any of these species could become the pivot species. The implication of this is that many cellular processes involving enzymes that interact with membranes, and whose activity is related to the membrane-stored elastic energy, would not be dependent on the presence of a specific lipid species for function, but would be driven by a supramolecular property of the membrane. As noted previously, many of the enzymes of phospholipid biosynthesis such as CCT, PLA₂, PLC, etc., are regulated by membrane composition and in principle by stored elastic stress, and our findings suggest the possibility that membrane-stored elastic stress might provide a stable baseline for the activation of membrane interacting proteins. The extent to which a given protein is modulated would depend on changes in membrane-stored elastic energy that arise within some critical distance from the protein, and this affords a cell distributed local control over the maintenance of curvature elastic energy. The possibility, and indeed likelihood, that other distributed control systems are also operative (e.g. based on electrostatics) affords additional robustness to the system. Clearly, such distributed control will also intersect with transcriptional control; though the two have significantly different response times.

We have restricted our studies to asynchronous cells to avoid complications arising from the time-dependent spatially heterogeneous distributions of lipids that may occur during the cell cycle. This means that the values of p_x that we have derived stem from an amalgam of different ratio control functions, possibly one for each subcellular membrane fraction such that each of these functions would potentially lead to a different ratio control value p_x. As methods for subcellular fractionation develop, it will become possible to conduct organelle/compartment specific lipidomic studies. These studies will reveal the extent to which lipid regulation through ratio control applies to the different subcellular compartments and whether there are multiple control mechanisms in operation.

One interesting feature of the proxy for the stored elastic energy that we obtain is that the magnitude of individual lipid contributions to stored elastic stress, in vivo, is less than that predicted from the small number of physical measurements on single-component systems. We attribute this to the term κ_M, which we expect to be significantly different in protein-rich lipid membranes of the cell, than in simple lipid mixtures.

The approach we have developed lays the foundation for a new way of evaluating the way cells change their lipid compositions in response to changes in their external environment, but also has an application where curvature elastic stress is implicated in drug mechanism of action [62,63]. This approach involves the use of extensive datasets to validate a postulated homeostatic control mechanism, namely that eukaryotic cells control the stored curvature elastic energy of their membranes. This postulated mechanism afforded a biophysical framework that enabled us to specify the nature of the control function, namely a ratio involving type II to type 0 lipids, as well as the way the lipid concentrations affect the control function. A method for calculating a set of coefficients, one for each lipid type, which quantify the contribution each lipid makes to the control function/stored elastic energy was proposed based on biophysical considerations. A recursive method was then implemented to generate coefficients, combine these with lipid species concentration data, and subject to the choice of a pivot species, to calculate the magnitude of the resulting control function. By selecting only sets of coefficients that gave variance below a certain cut-off, we were able to find that there is a universal lipid species that is consistent with the diverse datasets we used. This finding of a signature of ratio control from in vivo data is a significant development in understanding how eukaryotic cells may achieve lipid regulation at a distributed, local level. While our method has been demonstrated for eukaryotic cells, it will be of considerable interest to apply it to the investigation of prokaryotic lipid regulation that does not involve transcription control.

Our approach to capturing the contribution of individual lipids to the control function (or the stored elastic energy) is, of necessity, relatively coarse grained. However, improvements in the developing field of quantitative lipidomics will reveal greater structural information, e.g. through the disentangling of isomeric and isobaric lipids, will provide more information about the positions of unsaturated bonds, and will provide increasing access to quantification of the less abundant species, such as lyso phospholipids [56,64] and PI phosphate lipids [65]. These developments will allow the construction of more fine-grained control functions.

Our studies have been performed on total lipid extracts of whole cells. Subcellular fractionation, in combination with quantitative lipidomic studies of organelle membranes as described recently [66], will provide information into the way ratio control of lipid composition varies spatially within a cell and may lead to new insights into the critical interplay between localized and distributed biochemical/biophysical control of lipid composition and centralized transcription level control.