Journal of Colloid and Interface Science 343 (2010) 628–633

Contents lists available at ScienceDirect

Journal of Colloid and Interface Science

www.elsevier .com/locate / jc is

Phase behavior in a ternary lipid membrane estimated using a nonlinearresponse surface method and Kohonen’s self-organizing map

Yoshinori Onuki *, Kozo TakayamaDepartment of Pharmaceutics, Hoshi University, 2-4-41 Ebara, Shinagawa-ku, Tokyo 142-8501, Japan

a r t i c l e i n f o

Article history:Received 27 October 2009Accepted 2 December 2009Available online 6 December 2009

Keywords:Lipid raftLiposomeFluorescence anisotropyResponse surface methodKohonen’s self-organizing mapClustering

0021-9797/$ - see front matter � 2009 Elsevier Inc. Adoi:10.1016/j.jcis.2009.12.009

Abbreviations: CFM, confocal fluorescence micros1,2-dioleoyl-sn-glycero-3-phosphocholine; DPH, 1DPPC, 1,2-dipalmitoyl-sn-glycero-3-phosphocholine;calorimeter; FRET, Förster resonance energy transfernositol; La, liquid-disordered phase; Lb, solid-orderphase; MVS, multivariate spline interpolation; RSMincorporating multivariate spline interpolation; SM,nen’s self-organizing maps; Tm, phase transition temp

* Corresponding author. Fax: +81 3 5498 5783.E-mail address: onuki@hoshi.ac.jp (Y. Onuki).

a b s t r a c t

A novel method for investigating phase behavior in a ternary lipid membrane was developed and tested.Sixty-five model membranes composed of sphingomyelin (SM), dioleoyl phosphatidylcholine (DOPC),and cholesterol (Ch) were prepared, and fluorescence anisotropy between 25 �C and 60 �C was measured.Observed fluorescence anisotropy curves as functions of temperature were analyzed using a nonlinearresponse surface method and Kohonen’s self-organizing map. Thus, we generated a scatter plot indicatingthe distribution of membranes with similar membrane properties. The scatter plot showed that the SM/DOPC/Ch membranes resolved into six clusters with distinct membrane properties. We then conducteddifferential scanning calorimeter (DSC) measurement of membranes typical of the clusters. The resultsindicated that the membranes consisted of several phase domains (i.e., La, Lb, lo phase domains), andthe clusters were distinguished by differences in the type and content of membrane domain. This methodis accurate because the clusters were determined based on experimental values. This technique is usefulfor elucidating the phase behavior of ternary lipid membranes. These findings contribute to clarificationof domain formation.

� 2009 Elsevier Inc. All rights reserved.

1. Introduction

A lipid raft is defined as a membrane microdomain rich in cho-lesterol (Ch) and sphingolipid located in the outer leaflet of theplasma membrane. Because numerous signaling molecules suchas glycosylphosphatidylinositol (GPI)-anchored proteins [1,2] andreceptor- [3,4] or nonreceptor-type tyrosine kinases [4–6] origi-nate from these domains, the lipid raft is thought to act as a plat-form for protein segregation and signal transduction in the plasmamembrane. Lipid-driven lateral separation of immiscible liquidphases is likely to be a crucial factor in the formation of lipid raftsin cell membranes [7]. Although the concept of a lipid raft has re-cently been widely accepted, its mode of formation in the cellmembrane remains controversial.

ll rights reserved.

copy; Ch, cholesterol; DOPC,,6-diphenyl-1,3,5-hexatriene;

DSC, differential scanning; GPI, glycosylphosphatidyli-ed phase; lo, liquid-ordered-S, response surface methodsphingomyelin; SOM, Koho-erature.

Model membranes such as liposomes are effective tools for elu-cidating the formation of lipid rafts. Their lipid composition can bemanipulated to suit the purpose of the experiment, and results ob-tained in this way are likely to be consistent. In addition, severalmembrane domains are known to coexist in model membranesas well as in cell membranes [8–12].

In general, lipid bilayers are classified into three differentphases in order of increasing fluidity: a solid-ordered phase (Lb),a liquid-ordered phase (lo), and a liquid-disordered phase (La)[7,13]. The Lb and La phases are also called gel and liquid crystallinephases, respectively. Ordered and tight packing are typical of the Lb

phase membrane, whereas fast axial rotation and high lateralmobility are observed in the La phase membrane. The Ch-richmembrane exists as an lo phase membrane; this phase is interme-diate between the La and Lb phases. Its ordered packing is similar tothat of the Lb phase, but its fast axial rotation and high lateralmobility are similar to that of the La state. The lipid raft is assumedto exist as an lo phase membrane in the plasma membrane.

A mixture of three different lipids, lipids with a high phasetransition temperature (Tm) (e.g., lipid with saturated acyl chains),lipids with a low Tm (e.g., lipid with unsaturated acyl chains), andCh, is required to generate membrane domains. Such membraneshave been widely used to mimic plasma membranes to elucidatethe formation or structures of lipid rafts [9,11,12,14,15]. Eventhough these membranes are much simpler than biological mem-

Y. Onuki, K. Takayama / Journal of Colloid and Interface Science 343 (2010) 628–633 629

branes, their phase behavior remains complicated. Despitenumerous research studies, consensus on membrane phasebehavior has yet to be reached. Veatch et al. generated a ternaryphase diagram by observing the surface of giant unilamellar ves-icles using confocal fluorescence microscopy (CFM) and describeda condition in which phase separation occurs [11,12]. Althoughmicroscopic observation can be used to directly detect phase sep-aration on membrane surfaces, this method has limitations. First,because giant unilamellar vesicles are not obtained from all lipidcompositions, whole-phase behavior can never be determinedwith this method. Second, because the border region betweenmembrane domains is thought to be an ambiguous structure, itis difficult to distinguish membrane domains by subjective evalu-ation. Third, because the phase behavior of ternary lipids is verycomplicated and is substantially changed by slight differences inlipid composition, a bunch of model membranes differing in lipidcomposition should be examined to elucidate phase behavior.

Fluorescence analyses such as Förster resonance energy trans-fer (FRET) and fluorescence anisotropy are also employed to iden-tify the lipid phase behavior of membranes [14,16–19]. Becausethese measurements do not require preparation of a giant unila-mellar vesicle, a wider range of lipid compositions can be exam-ined than with CFM. In addition, measurements based onfluorescence analysis enable objective and quantitative evaluationof domain formation, and the experimental procedure is not verycomplicated. However, large data sets are required to fully under-stand the relationship between lipid composition and phasebehavior and, in most cases, the collection of such large data setsis impractical.

To overcome this problem and elucidate the phase behavior ofternary lipids, we applied a response surface method incorporatingmultivariate spline interpolation (RSM-S) and a data mining tech-nique (Kohonen’s self-organizing maps; SOMs). Firstly, liposomescomposed of sphingomyelin (SM), dioleoyl phosphatidylcholine(DOPC), and Ch were prepared and their fluorescence anisotropywas measured. Using data based on fluorescence anisotropy, wedeveloped a scatter plot indicating the distribution of membraneswith similar membrane properties. Response surface methods anda data mining technique were used to compensate for the lack ofexperimental data. Using these methods, we successfully resolvedthe membranes into several clusters. This is the first technical re-port on the phase behavior of lipids conducted using response sur-face and data mining methods.

2. Experimental procedure

2.1. Materials

Ch was purchased from Wako (Osaka, Japan). Chicken egg SM(Coatsome NM-10) and DOPC (Coatsome MC-8181) were pur-chased from Nippon Oil & Fat (Tokyo, Japan). More than 70% ofthe SM is 16:0 SM. 1,6-diphenyl-1,3,5-hexatriene (DPH) was pur-chased from Aldrich (Milwaukee, WI, USA). All other chemicalswere of analytical grade and are commercially available.

2.2. Preparation of liposomes

Liposomes composed of SM, DOPC, and Ch were prepared asreported previously [19]. In brief, designated amounts of lipidsdissolved in chloroform were transferred to a flask, and the chlo-roform was removed by evaporation at room temperature undera stream of nitrogen. This procedure resulted in the formationof a thin lipid film on the wall of the flask. The film was storedovernight in a vacuum desiccator to ensure complete evaporationof the chloroform. Purified water (10 mL) was added to the flask,

and the lipids were hydrated for 30 min. The total lipid concen-tration was adjusted to 10 mM. The suspension was sonicatedfor 10 min at about 60 �C using a bath-type sonicator. After thesamples were cooled to room temperature, they were stored atroom temperature for maximum of 2 days before use in theexperiments.

2.3. Fluorescence anisotropy measurement

The liposome was labeled with DPH by adding 10 lL of 10 mMfreshly prepared DPH stock solution in tetrahydrofuran to 1000 lLof liposome suspension and then incubating the mixture at 37 �Cfor 2 h in the dark to complete the labeling. The samples were di-luted 50 times with purified water. The fluorescence anisotropy ofDPH in the liposomes was measured using a fluorescence spectro-photometer (F-450; Hitachi, Tokyo, Japan) at an excitation wave-length of 351 nm and an emission wavelength of 430 nm. Thetemperature range during measurement was 25–60 �C. Steady-state fluorescent anisotropy was calculated using the followingequation:

r ¼ IVV � GIVH

IVV þ 2GIVHð1Þ

where r is anisotropy and IVV and IVH are the intensities measuredparallel and perpendicular to the polarized exciting light, respec-tively. The G factor was defined as IVV/IVH, which is equal to the ratioof the sensitivities of the detection system for vertically and hori-zontally polarized light. The G factor of our detection system was1.202.

2.4. Differential scanning calorimeter measurement

The liposome suspensions were lyophilized using a freeze dryer(FD-1; Tokyo Rikakikai, Tokyo, Japan) under reduced pressure. Thefreeze-dried liposomes (5 mg) were placed in aluminum pans forthe DSC measurement. DSC measurements were performed usinga Thermo Plus DSC 8230 instrument (Rigaku, Tokyo, Japan)equipped with a refrigerated circulator (Rigaku, Tokyo, Japan).The scan rate was set to 1 �C/min.

2.5. Clustering of SM/DOPC/Ch membranes into membranes withsimilar properties

The procedure for clustering membranes is shown in Fig. 1.Sixty-five model liposomes with different lipid compositions wereprepared (Fig. 2). Fluorescence anisotropy was measured at tem-peratures ranging from 25 �C to 60 �C. dataNESIA software, version3.0 (Yamatake Corp., Tokyo, Japan) was used for RSM-S. The ob-served fluorescence anisotropy values at 25, 30, 35, 40, 45, 50,55, and 60 �C and the differences in values from 25 �C to 60 �C wereused as tutorial data for generating response surfaces using RSM-S.Fluorescence anisotropy values of untested lipid compositionswere predicted by reading points on the response surfaces. Thenumber of untested lipid compositions for the prediction was5041. Lipid composition and corresponding fluorescence anisot-ropy values for every 5 �C from 25 �C to 60 �C were regarded asan input data set. Namely, 5106 data sets, including experimentaland predicted data, were used for SOM clustering. SOM clusteringwas performed using Viscovery software, SOMine version 4.0(Eudaptics Software, Vienna, Austria). The number of nodes inthe output was set at 2000. Thus, a scatter plot indicating the dis-tribution of distinct membranes was developed using referencevectors for each cluster.

Generating a ternary phase diagram

Collecting experimental data(Fluorescence anisotropy measurement)

Generating response surface by RSM-S (Correlation model)

Predicting characteristics of untested lipid compositions

Classification of lipid composition into several clusters by SOM

Estimating lipid composition belonging to each cluster

Fig. 1. Flow chart for estimating a ternary phase diagram for SM/DOPC/Ch byintegrating RSM-S and SOM.

10 10

0202

0303

0440

0505

0606

0707

0808

0909

001001DOPC SM

Ch

mol

% o

f DO

PC

mol%

of SM

Fig. 2. Lipid compositions of model membranes. The maximum contents of SM,DOPC, and Ch were 100%, 100%, and 90%, respectively. Each point represents thelipid composition of a model membrane. Sixty-five membranes with different lipidcompositions were prepared.

630 Y. Onuki, K. Takayama / Journal of Colloid and Interface Science 343 (2010) 628–633

3. Results

3.1. Estimation of the distribution of membranes with similarmembrane properties

Fluorescence anisotropy curves as functions of temperature ob-tained from 65 model membranes differed markedly in shape andvalue, indicating the diversity of the model membranes (data not

shown). The experimental values were processed using RSM-S,and then response surfaces indicating the relationships between li-pid composition and fluorescence anisotropy were generated. Onbehalf of all response surfaces generated using RSM-S, those at25, 40, and 60 �C are shown in Fig. 3a–c. We also generated a re-sponse surface representing differences in values from 25 �C to60 �C (Fig. 3d). Low values were observed for DOPC-rich mem-branes (neighborhood of the left-bottom vertex), whereas high val-ues were observed for membranes containing a large amount of Ch.The values of SM-rich membranes (neighborhood of the right-bot-tom vertex) were markedly lower at higher temperatures, implyingstructural alteration of membranes. Prediction accuracy was eval-uated by leave-one-out cross validation (Fig. 4). The correlationcoefficient was very high (r = 0.992), indicating that the RSM-Sconstructed a reliable model of the correlation between lipid com-position and fluorescence anisotropy. The response surfaces en-abled prediction of a large number of fluorescence anisotropyvalues. The number of untested lipid compositions for predictionwas set at 5041.

Subsequently, we analyzed experimental and predicted fluores-cence anisotropy values using SOM clustering and classified SM/DOPC/Ch membranes into clusters with similar membrane proper-ties. Viscovery software includes several clustering techniquessuch as SOM-Ward, Ward, and SOM-Single-Linkage. The SOM-Ward technique was employed for clustering because it is consid-ered the most efficient. SM/DOPC/Ch membranes were divided intosix clusters using SOM clustering. Consequently, based on the ref-erence vectors of clusters, we estimated a scatter plot indicatingthe distribution of lipid compositions with similar membraneproperties (Fig. 5). Cluster 1 was the most abundant in Ch, whereascluster 6 was the most abundant in DOPC. Cluster 3 was an SM-richmembrane.

3.2. Characterization of distinct membranes

We characterized distinct membranes according to SOM. Forthe following experiments, the centroid lipid composition of thecluster was assumed to be typical of the cluster (Table 1).

Fig. 6 shows the fluorescence anisotropy curves of the clusters.The symbols in the figure represent values predicted by RSM-S andSOM. The predicted values were highly consistent with experimen-tal values (r = 0.994, data not shown). Except for cluster 3, the val-ues decreased as the amount of DOPC in the membrane increased.The highest values were observed for cluster 1, whereas the lowestvalues were observed for cluster 6 (Fig. 6). The values of cluster 3decreased markedly from 0.25 to 0.15.

Freeze-dried liposomes were used for DSC measurement toavoid a large peak caused by water. The DSC curve of cluster 3showed an endothermic peak corresponding to the Tm of SM(Fig. 7). Similarly, endothermic peaks corresponding to the Tm ofDOPC were observed in clusters 5 and 6, which are DOPC-richmembranes (Fig. 7).

4. Discussion

To clarify the phase behavior of ternary lipid membranes, SM,DOPC, and Ch were selected as components of the model mem-brane. SM, a high Tm lipid, is a component of ordered phase mem-branes, whereas DOPC, a low Tm lipid, is a component of disorderedphase membranes, respectively. SM also forms a lo phase mem-brane with Ch. Fluorescence anisotropy was used to investigatethe phase behavior of the ternary lipid mixture. The La, Lb, and lophases of lipid membranes can be distinguished according to fluid-ity. Fluorescence anisotropy is commonly used to determine mem-brane fluidity. Each phase membrane can be identified by analysis

0.30

0.25

0.20

0.15

0.10

0.05

10

20

30

40

50

60

70

80

90

100

10

20

30

40

50

60

70

80

90

100

Ch

DOPC SM

mol

%of

DO

PC m

ol%of SM

(a)0.30

0.25

0.20

0.15

0.10

0.05

10

20

30

40

50

60

70

80

90

100

10

20

30

40

50

60

70

80

90

100

Ch

SMCPOD

mol

%of

DO

PC m

ol%of SM

(b)

0.30

0.25

0.20

0.15

0.10

0.05

10

20

30

40

50

60

70

80

90

100

10

20

30

40

50

60

70

80

90

100

Ch

DOPC SM

mol

%of

DO

PC m

ol%of SM

(c)

0.18

0.16

0.14

0.12

0.10

0.08

0.06

10

20

30

40

50

60

70

80

90

100

10

20

30

40

50

60

70

80

90

100

Ch

mol

%of

DO

PC m

ol%of SM

(d)

DOPC SM

Fig. 3. Response surfaces generated by RSM-S for fluorescence anisotropy values at 25 �C (a), 40 �C (b), and 60 �C (c) and differences in values from 25 �C to 60 �C (d). Pointsrepresent lipid compositions of the model membrane. Experimental values at each temperature were used as tutorial data and response surfaces were generated using RSM-S.

0.00

0.10

0.20

0.30

0.40

0.00 0.10 0.20 0.30 0.40

Experimental values

Pre

dict

ed v

alue

s

y = 0.981x - 0.004 r = 0.992

Fig. 4. Relationship between experimental values and values predicted by fluores-cence anisotropy. The prediction accuracy of RSM-S was evaluated by leave-one-outcross validation. One data pair was omitted from the data set, and then theprediction model was fabricated using the RSM-S. The missing data were predictedby the RSM-S model, and then the process was repeated for all data.

0

10

20

30

0101

20 20

0303

0404

0505

0606

0707

0808

90 90

100 100

DOPC SM

Ch

mol %

of SM

mol

% o

f DO

PC

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Cluster 5

Cluster 6

Fig. 5. Scatter plot indicating the distribution of SM/DOPC/Ch membranes withsimilar membrane properties. The phase diagram was based on the referencevectors of SOM. Large circles represent the centroids of clusters.

Table 1Typical lipid compositions of clusters in membranes.

SM (mol%) Ch (mol%) DOPC (mol%)

Cluster 1 25.0 69.4 5.6Cluster 2 36.7 40.8 22.5Cluster 3 79.0 8.5 12.6Cluster 4 23.4 33.6 43.0Cluster 5 24.7 18.0 57.3Cluster 6 12.1 14.5 73.3

Y. Onuki, K. Takayama / Journal of Colloid and Interface Science 343 (2010) 628–633 631

of the curves because their responses differ, i.e., the fluorescenceanisotropy values of Lb and lo phase membranes are much higherthan those of La phase membranes. The measurements were per-formed between 25 �C and 60 �C. This range includes the transitionof SM from the Lb phase to the La phase. Therefore, if a membranecontains a Lb phase membrane rich in SM, a marked decrease influorescence anisotropy should be observed at its Tm. In contrast,La and lo phase membranes are presumed to be stable within this

0.00

0.05

0.10

0.15

0.20

0.25

0.30

25 30 35 40 45 50 55 60

Temperature ( oC)

Flu

ores

cenc

e an

isot

ropy

Cluster 1

Cluster 2

Cluster 4

Cluster 3

Cluster 5

Cluster 6

Fig. 6. Fluorescence anisotropy of DPH for clusters in the typical membrane. Thetemperature was scanned at 1 �C/min. Lines and symbols represent experimentaland predicted values estimated using RSM-S and SOM, respectively. ( , s)cluster 1; ( , h) cluster 2; ( , e) cluster 3; ( , d) cluster 4; ( , N)cluster 5; ( , j) cluster 6.

-20 0 20 40 60 80 100 120 140 160

Temperature (oC)

SM

DOPC

Ch

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Cluster 5

Cluster 6

Fig. 7. Differential scanning calorimetry thermograms of freeze-dried liposomesuspensions. The temperature was scanned at 1 �C/min.

632 Y. Onuki, K. Takayama / Journal of Colloid and Interface Science 343 (2010) 628–633

temperature range. Previously, we described distinct fluorescenceanisotropy curves derived from Lb, lo, and a mixture of lo and Lamembranes composed of dipalmitoyl phosphatidylcholine (DPPC),DOPC, and Ch [19]. An advantage of fluorescence anisotropy is thatmeasurements can be conducted under identical conditionsregardless of the components and composition of the membranes.To make the most of this attribute, the lipid composition of themodel membranes were designed to be as diverse as possible.

The maximum amounts of SM, DOPC, and Ch were 100%, 100%,and 90% of total lipids, respectively (Fig. 2).

Based on the fluorescence anisotropy values of 65 model mem-branes, RSM-S was used to generate response surfaces of the rela-tionships between lipid composition and fluorescence anisotropy(Fig. 3). RSM-S is a nonlinear response surface method developedin our laboratory. Multivariate spline interpolation (MVS) is inte-grated into RSM-S as a method of generating the response surface.MVS involves a boundary element method [20]. It can estimatenonlinear relationships between factors and characteristics withhigh accuracy. To date, we have applied this method to formulationoptimization of pharmaceutics, and various findings have sug-gested that RSM-S is a promising tool for the development of phar-maceutics [21–26]. The response surfaces showed that thefluorescence anisotropy values changed substantially and in acomplicated manner as lipid composition and temperature chan-ged (Fig. 3). Thus, to fully understand the lipid phase behavior ofthe whole range of lipid compositions tested, a vast number of datasets would have to be compiled.

For instance, Buboltz et al. investigated phase behavior in mix-tures of DPPC, DOPC, and Ch using FRET measurements [17] andalso encountered the problem of the number of data sets requiredfor interpretation of their results. To address this issue, they exam-ined 1294 kinds of lipid compositions and prepared a ternaryphase diagram. Although their experimental procedure was muchsimpler than others, the collection of such a large amount of exper-imental data must have been very arduous. However, this strategywas not feasible with our experimental approach.

We relied on the high prediction accuracy of RSM-S to overcomethe problem of a relatively small number of data sets. Leave-one-outcross validation showed that the response surfaces generated in thisstudy have a good correlation between predicted and experimentalvalues (r = 0.992) (Fig. 4). We also evaluated the model correlationsusing data sets consisting of untested lipid compositions. In thecourse of the study, we identified clusters in six typical membranes(Table 1) and measured their fluorescence anisotropy values (Fig. 6).Because these lipid compositions were not used in generating re-sponse surfaces, we regarded them as untested. There was a highcorrelation coefficient (r = 0.994, data not shown) between the pre-dicted and experimental values. Thus, a large number of fluores-cence anisotropy values were predicted with high accuracy. As thismethod is applicable to any experimental data, we consider it an effi-cient method of compensating for a lack of experimental data.

Experimental and predicted fluorescence anisotropy valueswere analyzed using SOM clustering. SOM clustering is now re-garded as a powerful tool for data mining. SOM is a feedforward-type neural network model that consists of one input layer andone output layer [27]. The SOM algorithm is based on unsuper-vised, competitive learning. The network ultimately associatesthe output nodes with groups or patterns of input vectors byrepeating the learning. This method was described in detail inour previous articles [22,25]. SOM can accommodate several vari-ables as input vectors and takes them into account in determiningclustering. We used a series of fluorescence anisotropy values fordifferent temperatures as input vectors to represent the curve.Using this method, we were able to classify membranes with sim-ilar properties based on fluorescence anisotropy curves obtained.

As a result of the SOM clustering analysis, SM/DOPC/Ch mem-branes were divided into six clusters (Fig. 5). Apart from cluster3, the main difference between clusters was the DOPC content ofthe total lipid fraction. The fluorescence anisotropy values of typi-cal membranes decreased progressively as the amount of DOPC inthe membrane increased (Fig. 6). Endothermic peaks correspond-ing to the Tm of DOPC were observed in DOPC-rich membranes(clusters 5 and 6) (Fig. 7). This indicates that a considerableamount of La phase membrane domains rich in DOPC are present

Y. Onuki, K. Takayama / Journal of Colloid and Interface Science 343 (2010) 628–633 633

on their surfaces. Based on these findings, the clusters were distin-guished according to the ratio between the disordered phase do-main and the ordered phase domain.

We intended to use differences in the fluorescence anisotropyvalues between 25 �C and 60 �C to estimate the distribution of theLb phase domain. As anticipated, substantial changes in the valueswere observed for SM-rich membranes (Fig. 3d). As the distributionwas quite similar to that of the Lb phase domain reported by de Al-meida et al. [28], we considered the pattern shown in Fig. 3droughly indicative of the distribution of the Lb phase domain. Clus-ter 3 was abundant in membranes that exhibited a substantialchange in anisotropy. The typical membrane showed a marked de-crease in fluorescence anisotropy (Fig. 6) and an endothermic peakcorresponding to the Tm of SM (Fig. 7). Thus, cluster 3 representsmembranes rich in Lb phase membrane domains. In addition, a partof cluster 5 (Ch < 10 mol%, 40 mol% < SM < 80 mol%) showed aslight but obvious difference in anisotropy value (Fig. 3d and 5).These membranes were probably composed of a mixture of Lb andLa phase domain membranes. In contrast, the fluorescence anisot-ropy values of clusters 1, 2, and 4 hardly changed, even thoughthe temperature increased to 60 �C; therefore, the most orderedphase domains were assumed to exist as an lo phase membrane.This is supported by the phase diagram of de Almida et al. [28].

The results of this study show that ternary lipid membranescomposed of SM, DOPC, and Ch are formed by distinct membranedomains such as the La, Lb, and lo phase domains, and several clus-ters with similar membrane properties can be distinguishedaccording to the type and content of the domain. This techniqueis useful for elucidating the phase behavior of lipid mixtures.

5. Conclusions

We successfully classified SM/DOPC/Ch membranes into sixclusters with similar membrane properties. Our study, whichencompassed a wide range of lipid compositions, contributes tothe elucidation of the behavior of ternary lipid mixtures. Althoughan immense amount of experimental data would have been re-quired to interpret the data using conventional means, we over-came this problem by using RSM-S and Kohonen’s self-organizingmap. Furthermore, because the clusters were determined basedon experimental values, the accuracy of our method is assured.Our method should be useful for related studies and contributeto elucidating the formation of lipid rafts.

Acknowledgments

The authors are grateful to Yamatake Corporation for providingus with dataNESIA version 3.0. We are also grateful to Ms. Eri Imajoat Hoshi University for her kind assistance with the experimentalwork.

References

[1] D.A. Brown, J.K. Rose, Cell 68 (1992) 533–544.[2] K. Simons, E. Ikonen, Nature 387 (1997) 569–572.[3] J. Matkó, J. Szöllõsi, Immunol. Lett. 82 (2002) 3–15.[4] T.M. Stulnig, J. Huber, N. Leitinger, E.M. Imre, P. Angelisova, P. Nowotny, W.

Waldhausl, J. Biol. Chem. 276 (2001) 37335–37340.[5] A.M. Shenoy-Scaria, J. Kwong, T. Fujita, M.W. Olszowy, A.S. Shaw, D.M. Lublin, J.

Immunol. 149 (1992) 3535–3541.[6] I. Stefanova, V. Horejsi, I.J. Ansotegui, W. Knapp, H. Stockinger, Science 254

(1991) 1016–1019.[7] E. London, Biochim. Biophys. Acta 1746 (2005) 203–220.[8] C. Dietrich, L.A. Bagatolli, Z.N. Volovyk, N.L. Thompson, M. Levi, K. Jacobson, E.

Gratton, Biophys. J. 80 (2001) 1417–1428.[9] N. Kahya, D. Scherfeld, K. Bacia, B. Poolman, P. Schwille, J. Biol. Chem. 278

(2003) 28109–28115.[10] B.D. Ladbrooke, R.M. Williams, D. Chapman, Biochim. Biophys. Acta 150 (1968)

333–340.[11] S.L. Veatch, S.L. Keller, Biophys. J. 85 (2003) 3074–3083.[12] S.L. Veatch, S.L. Keller, Biochim. Biophys. Acta 1746 (2005) 172–185.[13] D.A. Brown, E. London, Annu. Rev. Cell Dev. Biol. 14 (1998) 111–136.[14] Megha, P. Sawatzki, T. Kolter, R. Bittman, E. London, Biochim. Biophys. Acta

1768 (2007) 2205–2212.[15] J. Wang, Megha, E. London, Biochemistry 43 (2004) 1010–1018.[16] O. Bakht, P. Pathak, E. London, Biophys. J. 93 (2007) 4307–4318.[17] J.T. Buboltz, C. Bwalya, K. Williams, M. Schutzer, Langmuir 23 (2007) 11968–

11971.[18] Megha, O. Bakht, E. London, J. Biol. Chem. 281 (2006) 21903–21913.[19] Y. Onuki, C. Hagiwara, K. Sugibayashi, K. Takayama, Chem. Pharm. Bull. 56

(2008) 1103–1109.[20] D.T. Sabdwell, Geophys. Res. Lett. 14 (1987) 139–142.[21] H. Arai, T. Suzuki, C. Kaseda, K. Ohyama, K. Takayama, Chem. Pharm. Bull. 55

(2007) 586–593.[22] M. Nishikawa, Y. Onuki, K. Isowa, K. Takayama, AAPS Pharm. Sci. Technol. 9

(2008) 1038–1045.[23] Y. Onuki, M. Morishita, K. Takayama, J. Control, Release 97 (2004) 91–99.[24] Y. Onuki, M. Hoshi, H. Okabe, M. Fujikawa, M. Morishita, K. Takayama, J.

Control, Release 108 (2005) 331–340.[25] Y. Onuki, K. Ohyama, C. Kaseda, H. Arai, T. Suzuki, K. Takayama, J. Pharm. Sci.

97 (2008) 331–339.[26] K. Takayama, Y. Obata, M. Morishita, T. Nagai, Pharmazie 59 (2004) 392–

395.[27] K. Kohonen, Self-organizing Maps, Springer Series in Information Sciences,

Berlin, 1995.[28] R.F. de Almeida, A. Fedorov, M. Prieto, Biophys. J. 85 (2003) 2406–2416.