J. R. Soc. Interface

*Author for c

Electronic sup10.1098/rsif.2

doi:10.1098/rsif.2012.0140Published online

Received 23 FAccepted 13 M

Quantitative three-dimensionalmicrotextural analyses of tooth wearas a tool for dietary discrimination

in fishesMark Purnell1,*, Ole Seehausen2,3 and Frietson Galis4,5

1Department of Geology, University of Leicester, Leicester LE1 7RH, UK2Division of Aquatic Ecology and Evolution, Institute of Ecology and Evolution, University of

Bern, Baltzerstrasse 6, CH-3012 Bern, Switzerland3Department of Fish Ecology & Evolution, Centre for Ecology, Evolution and

Biogeochemistry, EAWAG Swiss Federal Institute for Aquatic Science and Technology,Seestrasse 79, 6047 Kastanienbaum, Switzerland

4NCB Naturalis, 2333 CR Leiden, The Netherlands5VU University Medical Centre, 1007 MB Amsterdam, The Netherlands

Resource polymorphisms and competition for resources are significant factors in speciation.Many examples come from fishes, and cichlids are of particular importance because of theirrole as model organisms at the interface of ecology, development, genetics and evolution. How-ever, analysis of trophic resource use in fishes can be difficult and time-consuming, and for fossilfish species it is particularly problematic. Here, we present evidence from cichlids that analysis oftooth microwear based on high-resolution (sub-micrometre scale) three-dimensional data andnew ISO standards for quantification of surface textures provides a powerful tool for dietarydiscrimination and investigation of trophic resource exploitation. Our results suggest thatthree-dimensional approaches to analysis offer significant advantages over two-dimensionaloperator-scored methods of microwear analysis, including applicability to rough tooth surfacesthat lack distinct scratches and pits. Tooth microwear textures develop over a longer period oftime than is represented by stomach contents, and analyses based on textures are less prone tobiases introduced by opportunistic feeding. They are more sensitive to subtle dietary differencesthan isotopic analysis. Quantitative textural analysis of tooth microwear has a useful role to play,complementing existing approaches, in trophic analysis of fishes—both extant and extinct.

Keywords: cichlids; microwear analysis; trophic analysis; resource exploitation

1. INTRODUCTION

There is abundant evidence that resource polymorph-isms, and character divergence linked to competitionfor resources, are significant factors in speciation [1,2].Many examples come from fishes, but analysis oftrophic resource use in fishes is not straightforward.Stomach content analysis provides only a ‘snapshot’of diet over the few hours prior to capture, and thereforerequires relatively large samples taken over all relevantseasons to be reliable, making it time-consuming andoften logistically difficult. Stable isotopic data provideevidence of resource use integrated over a longer inter-val, but provide only an indication of relative trophicposition or placement along the littoral–profundaldepth gradient. Analysis of dietary preferences basedon functional morphology is hampered by the mismatchbetween apparent specialization in trophic morphology

orrespondence (mark.purnell@le.ac.uk).

plementary material is available at http://dx.doi.org/012.0140 or via http://rsif.royalsocietypublishing.org.

ebruary 2012arch 2012 1

and actual diet [1,3]. This creates particular problemsfor dietary interpretations that rely heavily on anatomicaldata, including hypotheses that invoke changes in trophicresource use of extinct fishes as explanations for macro-evolutionary and macroecological events [4] and as thecause of ancient adaptive radiations and diversification.

Here, we present evidence from cichlids that analysisof fish tooth microwear based on three-dimensionalmicrotextural data provides an additional, powerfultool for dietary discrimination and investigation oftrophic resource exploitation in fishes. It is particularlyuseful because the dietary signal accumulates overlonger timescales than stomach contents and thereforeavoids the ‘snapshot’ problem. In addition, analysis ofmicrowear can detect subtle dietary differences betweenindividuals and populations, even when sample sizesare small. Further, it is applicable to fossils, and tospecimens which lack stomach content or isotopic data.

Quantitative microwear analysis is a powerful toolfor the analysis of diet and feeding mechanisms,especially when applied to extinct organisms:

This journal is q 2012 The Royal Society

2 Tooth microtextures and diet in fishes M. Purnell et al.

microwear provides direct evidence of tooth use that isindependent of functional inferences derived from jawand tooth morphology, and thus breaks the circula-rity in reasoning that can cause difficulties for robustanalysis of feeding and diet in fossils [5]. Quantitativemicrowear analysis is an established method in dietaryanalysis of fossil mammals [6–8], but is starting to beapplied more broadly, and new approaches are emergingthat provide new and more refined uses. Recent work,for example, has demonstrated that quantitative toothmicrowear provides a reliable guide to trophic ecology instickleback fishes (Gasterosteus), and can be used totrack the relationshipbetweenmicroevolution and changesin trophic niche over evolutionary timescales of tens ofthousands of years [9,10]. It has also been applied to dino-saurs to test hypotheses of jaw mechanics [11]. Perhapsmore significant is the realization that a new approach toquantification of sub-micrometre scale three-dimensionalsurface texture, borrowed from engineering, offers a morerobust method for tooth microwear analysis [8,12–14].International standards for three-dimensional surfacetextural analysis are only now emerging (ISO 25178-2,currently in late stages of preparation), but the approachhas be hailed as a paradigm shift in the field of surfacemetrology [15]. Three-dimensional textural analysis hasthe potential to overcome issues of subjectivity in operatorscoring that have hampered otherapproaches toanalysis oftooth microwear [9,16].

Cichlids represent an ideal choice to further inves-tigate the usefulness of microtextural analysis ofmicrowear for dietary discrimination because of theirwell-known and well-characterized trophic diversity,and because of the well-established link between trophicecology, morphological evolution and speciation. Theyare important model organisms at the interface of ecol-ogy, development, genetics and evolution, providingtext-book examples of speciation and adaptive radiation[17]. This study has two principal objectives. First,to test the applicability of quantitative tooth micro-wear analysis to oral and pharyngeal teeth for dietarydiscrimination in teleost fishes (to date analysis of tele-osts has been restricted to oral teeth of sticklebacks).Our second objective is to test the relative power of‘standard’ two-dimensional image-based approachesand three-dimensional microtextural analysis of micro-wear for dietary discrimination and trophic analysis.Two-dimensional approaches, where microwear ismanually scored by an operator, are widely employedbecause they are relatively simple and do not requirespecialist hardware, but they are time-consumingand prone to operator error [9,12,16]. Microtexturalapproaches avoid the difficulties inherent in operatorscoring by deriving quantitative measures of surfacetexture direct from high-resolution three-dimensionaldata (point clouds). Textural parameters can bederived from scale-sensitive fractal analysis (SSFA;[8,12]) or from measures reflecting the height, ampli-tude and volumes of peaks and valleys in a surface.The latter are soon to be established as the industrystandard for textural analysis in engineering [18], buttheir applicability to tooth microwear analysis of dietremains untested. One previous analysis [19] derivedboth SSFA and ISO textural parameters from three-

J. R. Soc. Interface

dimensional data in order to investigate the functionof complex dental surfaces in ungulates, but did nottest the power of textural parameters in dietarydiscrimination. Another analysis derived from three-dimensional surface data [20] compared five morebasic measures of surface roughness (Rp, Rv, Rt, Rqand Ra, defined according to the ANSI B46.1 standard)with image-based scoring of buccal tooth microwear inextant and fossil primates. Although this study foundcorrelations between some measures of roughness andmicrowear patterns, its conclusions were generally doubt-ful regarding the usefulness of roughness measures indietary analysis, partly because roughness was found tobe highly sensitive to post-mortem abrasion.

As part of our analysis, we also test the hypothesisthat three-dimensional microtextural analysis of roughtooth surfaces can discriminate between individualswith different diets. Rough tooth surfaces, in this con-text, are not amenable to two-dimensional scoringapproaches because although worn they lack discern-able, discrete microwear features (pits and scratches).This has not previously been raised as an issue in micro-wear analysis of mammals (because of the nature oftheir microwear and the general focus on occlusalfacets), but rough tooth surfaces are likely to be encoun-tered more frequently as microwear analysis is extendedto other groups of vertebrates.

2. MATERIAL AND METHODS

For analysis of oral teeth, jaws were acquired from ninespecimens of Neochromis gigas and an unnamed speciesof Haplochromis informally referred to as Haplochromispurple-yellow (hereafter Hpy) collected from LakeVictoria (by O.S.). Neochromis gigas (NgR) is morpho-logically specialized for scraping algae off rock surfaces(epilithic algae scraper; [21]). Hpy is morphologicallyspecialized for scraping algae from macrophytes (epi-phythic algae scraper; [17]), but feeds on epilithicalgae in some sites, especially on offshore rocky islandsthat lack aquatic macrophytes. Our oral tooth samplesrepresent three populations: NgR from offshore, rockyMakobe Island, observed to feed primarily by epilithicalgae scraping [21]; Hpy, co-occurring with NgR atMakobe Island, also observed to be epilithic algae scra-pers (HpyR); Hpy from an epiphytic algae scrapingpopulation from Kissenda Bay (HpyV). Data wereacquired from the second tooth from the symphysis.

For analysis of rough tooth surfaces, lower pharyngealjaws (LPJ) were obtained from nine Astatoreochromisalluaudi, six wild and three laboratory-raised, selectedblind with respect to tooth wear. Wild specimens wereoriginally collected from the Mwanza Gulf and KissendaBay of Lake Victoria; LPJ were dissected out as part of alarger study [22]. In Lake Victoria, the dominantfood items of A. alluaudi are molluscs, particularly gas-tropods [23] and gut contents confirm this (see theelectronic supplementary material). Laboratory-raisedfish from which LPJ were obtained were fed a soft fooddiet of minced heart and liver, with vitamins and Tetra-min flakes added [24]. All specimens are housed in TheNetherlands Centre for Biodiversity Naturalis under

Tooth microtextures and diet in fishes M. Purnell et al. 3

numbers RMNH.PISC.37855 to RMNH.PISC.37872(see electronic supplementary material, table S2).

When A. alluaudi of similar standard length are com-pared, the lower pharyngeal teeth and jaws of wildindividuals feeding on molluscs are significantly largerand more robust than those from laboratory-fed fishwith a soft diet [22,25]. In order to reduce the possi-bility that any differences in LPJ tooth microtexturereflect factors linked specifically to size of teeth,jaws and individuals—rather than differences indiet—wild-caught LPJ of two kinds were analysed.Three LPJ were of similar size to the jaws fromlaboratory-raised fish. Assessment of similarity in LPJsize was based primarily on caudal horn width anddepth of rostal keel, parameters taken in previousstudies as indicators of crushing power [26], so crushingability of these fish should be comparable to those of thelaboratory-raised fish. These wild fish have lowerstandard lengths than the laboratory fish. Three wild-caught LPJ were selected to represent fish that were ofsimilar standard length to the laboratory-raised fish(LPJ and teeth larger than the laboratory-raised fish).The standard length of all wild-caught individuals wasbetween 46 and 72.7 mm: this is above the size (40 mm)at which wild A. alluaudi start to consume molluscsand LPJ development starts to diverge in mollusc special-ists, but below the size (100 mm) at which differences indiet and trophic morphology are fully established [22,25].For lower pharyngeal jaws, three-dimensional micro-textural data were acquired from the largest teethlocated adjacent to the central suture. The two mostworn teeth of each individual were sampled.

Oral teeth were scored for two-dimensional microwearanalysis using Microware v. 4.02 [27] (see the electronicsupplementary material for details). For both oraland lower pharyngeal jaws, high-resolution three-dimensional surfaces were captured using an AliconaInfinite Focus microscope G4b (IFM; software v. 2.1.2),using �100 objective to give a field of view of 145 �110 mm. All three-dimensional data were processedusing the Alicona IFM software (v. 2.1.2) as detailed inthe electronic supplementary material. Results presentedhere are based on data that were levelled and filtered toremove long wavelength features of the tooth surface(gross tooth form). Details of statistical methods, andresults of additional analyses, are included in the elec-tronic supplementary material. Data were exploredusing ANOVA, correlations, Principal Components (oncorrelations) and linear discriminant analyses (LDA).

Sample sizes used in this study are relatively small.This is for two reasons: (i) if microwear analysis is to beapplied to historical collections, extinct taxa and fossilteeth, it must be able to discriminate dietary differencesin small samples (e.g. in fossil taxa, where teeth withsurfaces that have not been adversely affected by post-mortem processes can be rare). (ii) This project wasinitiated before any other work on fish tooth microwear,to assess the feasibility of microwear analysis in fishes,and explore alternative approaches to analysis, requir-ing multiple datasets at different sampling scales andgenerated using different techniques. This is very time-consuming, so sample sizes were small in order to preventthe project expanding beyond the available resources.

J. R. Soc. Interface

Part of our purpose is to demonstrate the most effectiveapproaches to sampling and analysis, thus allowingfuture studies to more efficiently generate data andpotentially increase sample sizes.

Testing the applicability of microwear analysis tocichlids involves a number of subsidiary hypotheses con-cerning location and scale of sampling. We were unableto reject the hypothesis that microwear of oral teethdoes not differ between dentary and pre-maxilla teeth inan individual (see electronic supplementary material,table S3; the only exception is 300 mm length data). Con-sequently, all analyses of oral teeth presented here arebased on dentary teeth. The size of the area of a tooththat is sampled for microwear analysis varies betweenstudies, but commonly includes fields of view that aretoo large to be applied to smaller fish teeth. Comparingdata captured from the same teeth, but with differentfields of view (300 and 100 mm wide), feature length, den-sity and orientation all differ significantly (see electronicsupplementary material, table S4); R (mean vectorlength; a measure of angular dispersion) does not differ.Differences linked to scale are unsurprising, given thedirect influence that sampling area will have on somemeasures (e.g. feature length), but the question remainswhether 100 mm or 300 mm data differ in their dietarydiscriminatory power. We address this below.

3. RESULTS

3.1. ANOVA of two-dimensional and three-dimensional data, oral and pharyngeal teeth

Analyses of two-dimensional variables derived fromboth 100 and 300 mm sample areas indicate that featurelength and orientation differ significantly between indi-vidual fish (R and density cannot be tested forindividuals), but for most variables, we were unable toreject the hypothesis that microwear does not differbetween the three populations (table 1). Of the eighttests, the two exceptions are 300 mm density, and100 mm length data; in both cases, pairwise com-parisons (Tukey HSD) indicate that HpyV differ fromHpyR (a difference between these populations isalso evident from pairwise analysis of orientation(300 mm)). So although there are some differences, thediscriminatory power of two-dimensional microwearanalysis in small samples of these fish is limited. Thisis somewhat puzzling, because tooth surface micro-graphs of teeth from the different populations lookquite different (figure 1). This is especially true ofthe N. gigas teeth, and it seems likely that the lack ofdifference in the two-dimensional data may reflect thedifficulty in scoring rough surfaces lacking distinctiveindividual features. If this is the case then three-dimensional data should have greater discriminatorypower, and this is borne out by our analysis in twoways. First, exploratory ANOVA reveals that seven tex-tural parameters derived from the oral teeth (Sa, Sk,Spk, Vmp, Vmc, Vvc and Sal; a mixture of height,spatial and functional parameters) differ significantlybetween populations, with pairwise comparisons(Tukey HSD) indicating significant differences betweenNgR and both Hpy populations in six cases (table 2).

(a)

(c) (g) 1 mm

(d)

(e) (i)

( f ) ( j)

(h)

(b)

Figure 1. Microwear on cichlid teeth. (a,b) Dentary teeth of Haplochromis and Neochromis. The 300 mm field of view sampled fortwo-dimensional microwear data is shown; box shows area of three-dimensional data sample. (a) Macrophyte scraping Haplochro-mis ‘purple yellow’ (specimen HpyV01). (b) Rock scraping Neochromis gigas (specimen NgR92). (c–j) Lower pharyngeal jawteeth of Astatoreochromis alluaudi. (c–f) Specimen RMNH.PISC.37865, laboratory-raised, soft food diet; (e) area sampledfor three-dimensional microtextural data (boxed in (d)), ( f ) shows three-dimensional data, 608 tilt. (g–j) SpecimenRMNH.PISC.37870, wild-caught, mollusc diet; (i) shows area sampled for three-dimensional microtextural data (boxed in(h)), ( j) three-dimensional data, 608 tilt. (a–h) Scanning electron micrographs; (f,j) Alicona-rendered three-dimensional data.(e,f,i,j) Field of view 145 mm wide; vertical colour scale from 9 to 216 mm.

Table 1. Results of statistical hypothesis testing to determine whether two-dimensional oral tooth microwear data differsbetween individual cichlids (raw feature data) and between populations. Orientation analysis is based on a Watson–Williamsmulti-sample test (values are not included for populations where the null hypothesis that the orientation has a uniform(non-preferential) distribution, cannot be rejected (Rayleigh test, p ¼ 0.05)).

test result d.f. p

dentary 300 mm images, null hypothesesfeature length does not differ between fish reject x2 ¼ 31.16 ,0.0001feature orientation does not differ between fish reject F ¼ 14.69 7,970 ,0.0001feature length does not differ between populations no F ¼ 1.92 2,6 0.23feature density does not differ between populations reject F ¼ 11.88 2,6 0.01feature R does not differ between populations no F ¼ 0.34 2,6 0.72feature orientation does not differ between populations no F ¼ 5.08 2,5 0.06

dentary 100 mm images, null hypothesesfeature length does not differ between fish reject x2 ¼ 58.12 ,0.0001feature orientation does not differ between fish reject F ¼ 15.74 6,396 ,0.0001feature length does not differ between populations reject F ¼ 8.00 2,6 0.02feature density does not differ between populations no F ¼ 2.29 2,6 0.18feature R does not differ between populations no F ¼ 0.31 2,6 0.75feature orientation does not differ between populations no F ¼ 1.44 2,4 0.34

4 Tooth microtextures and diet in fishes M. Purnell et al.

J. R. Soc. Interface

Table 2. Summary of results of exploratory ANOVA, three-dimensional datasets, testing the null hypotheses of nodifference between populations. For oral teeth, Tukey HSDindicates that where differences between parameters aresignificant, NgR differs from both HpyR and HpyV in Sa,Sk, Spk, Vmp, Vmc, Vvc; NgR differs from HpyR in Sal. ForLPJ teeth, Tukey HSD indicates that the standard-length-equivalent-wild population differs from the laboratory-raised(soft diet) population in Sa, Sq, Sk, Spk, Vmp, Vmc,Vvc, Vvv, and Sal, standard-length-equivalent-wild fishdiffer from both other populations in Svk, and from thesmaller-wild populations in Smr2.

parameteroralresult oral F

oralp

LPJresult

LPJF

LPJp

Sa reject 6.94 0.03 reject 9.48 0.01Sq no 4.30 0.07 reject 10.79 0.01Sp no 0.67 0.54 no 2.03 0.21Sv no 3.87 0.08 no 1.50 0.30Sz no 1.95 0.22 no 1.81 0.24S10z no 0.73 0.52 no 1.88 0.23Ssk no 0.57a 0.62Sku no 0.91a 0.50 no 0.22 0.81Sdq no 0.46 0.65 no 2.79 0.19Sdr no 1.43 0.31 no 3.80 0.13Sk reject 8.00 0.02 reject 7.47 0.02Spk reject 10.86 0.01 reject 6.65 0.03Svk no 2.81 0.14 reject 14.48 0.005Smr1 no 0.15 0.86 no 0.71 0.55Smr2 no 0.91 0.45 reject 5.91 0.04Vmp reject 10.84 0.01 reject 6.58 0.03Vmc reject 7.03 0.03 reject 8.61 0.02Vvc reject 9.92 0.01 reject 8.44 0.02Vvv no 2.96 0.13 reject 13.17 0.01Vvc/Vmc no 0.35 0.72 no 0.29 0.77Sal reject 6.16 0.03 reject 5.91 0.04Str no 0.35 0.72 no 0.94 0.44Stdi no 0.79 0.49 no 0.57 0.59

aIndicates Welch test result (ANOVA, unequal variances).For oral teeth, Sp, Sv, Sz, and S10z log transformed; allLPJ data log-transformed. LPJ Ssk has negative values andwas therefore excluded from the analysis.

Tooth microtextures and diet in fishes M. Purnell et al. 5

Second, analysis of textural parameters derived fromLPJ of A. alluaudi found significant differences in 11height, functional and spatial parameters, all but twoof which differ between the laboratory-raised (soft-diet) and the standard-length-equivalent-wild fish, butnot between laboratory-raised and smaller-wild fish(Tukey HSD; table 2). Analyses including a secondtooth from each individual gave comparable results(see the electronic supplementary material). It is impor-tant to note that Sa, Sk, Spk, Vmp, Vmc, Vvc, Saldiffer significantly in both the dataset derived fromoral teeth and that derived from LPJ teeth. This is sig-nificant because obtaining comparable results from twoindependent datasets strongly suggests that our analy-sis is not being skewed by type I errors, which canarise as a consequence of multiple testing.

3.2. Correlations between two-dimensional andthree-dimensional data

For the oral teeth, analysis of two-dimensional data andthree-dimensional parameters revealed that there are no

J. R. Soc. Interface

significant correlations between any three-dimensionalparameters and feature length, feature width, featuremean orientation or R. Fifteen parameters, most ofwhich are height and functional parameters, are corre-lated with density (Sa, Sq, Sp (log), Sv (log), Sz (log),S10z (log), Sdr, Sk, Spk, Svk, Vmp, Vmc, Vvc, Vvv,Sal; see the electronic supplementary material fordetails). In all cases, the correlations are positive, soas feature density increases, so do the roughness par-ameters (see electronic supplementary material,figure S1 for visualization). This result goes some wayto explain the greater sensitivity of three-dimensionalroughness data in dietary discrimination (see discussionbelow). Our two-dimensional analysis and previouswork on fishes [9,10] shows that feature density is infor-mative, but quantification of three-dimensionalroughness clearly breaks down the signal captured cru-dely by feature density into a number of more subtleattributes of the tooth surface.

Analysis of how three-dimensional roughness par-ameters correlate with the surfaces of the LPJ teethused a qualitative scale based on visual assessment ofrelative roughness of two-dimensional images (see theelectronic supplementary material for methods of assess-ment). This qualitative ranking of surface roughness iscorrelated with 12 quantitative roughness parameters(Sa, Sq, Sp, Sz, S10z, Sdq, Sdr, Sk, Spk, Vmp, Vmc,Vvc), a mixture of functional, height and hybridparameters (see electronic supplementary material,table S1). All correlations are positive, so as qualitativeroughness increases, so do the quantitative roughness par-ameters (see electronic supplementary material, figure S2for visualization). Unlike the results of quantitative analy-sis of roughness, qualitative ranking of LPJ toothroughness is not correlated with stomach contents.

3.3. Multivariate analysis of two-dimensionaland three-dimensional data from oral teeth

The relative discriminatory power of two-dimensionaland three-dimensional microwear analyses was furthertested using principal components analysis (PCA;figure 2) and LDA. For the 100 mm two-dimensionaldata, the distribution of samples along PC axis 1 reflectsecology only moderately well: HpyV have negative values,HpyR having more positive values and NgR sit bet-ween (figure 2a). Separation between populations is notcomplete, however, as the least worn NgR specimen(NgR101; ‘2’ in figure 2) plots among HpyV, and themost-worn (NgR92; ‘1’ in figure 2) plots with HpyR.There is little ecologically informative separation of popu-lations along PC axis 2. Reducing the number of variablesanalysed does not result in ordinations with axes that cor-respond closely to ecological differences between samples(see the electronic supplementary material).

For the 300 mm two-dimensional data (figure 2b),analyses based on seven variables result in separationof the three populations on PCA axes 1 and 2: HpyVare separated from rock scraping populations alongPC axis 1; rock scraping NgR and HpyR are separatedalong PC axis 2. Results of analysis based on ‘standard’variables of mean feature length, mean feature orien-tation, angular dispersion (R) and density (n/area)

–3

–2

–1

0

1

2

3

PC2

8

4

6

1

3

5

7

2

9

–4 –3 –2 –1 0 1 2 3 4

(a)

HpyV

NgR

HpyR

–3

–2

–1

0

1

2

3

23

4

56

1

8

7

9

–4 –3 –2 –1 0 1 2 3 4

(b)

HpyVNgR

HpyR

–3

–2

–1

0

1

2

3

PC2

3

1

2

8

496

57

–4 –3 –2 –1 0 1 2 3 4

(c)

HpyVNgR

HpyR

–6

–4

–2

0

2

4

6

3

9

12

84

6

57

–8 –6 –4 –2 0 2 4 6 8

(d)

HpyV

NgR

HpyR

–3

–2

–1

0

1

2

3

PC2

16

1814

17

13

15

11

10

12

–5 –4 –3 –2 –1 0 1 2 3 4 5PC1

(e)

laboratory-raised

small-wild

size-equivalent-wild–4

–2

0

2

4

16

18

14

13

17

11

15

10

12

–8 –6 –4 –2 0 2 4 6 8PC1

( f )

laboratory-raisedsmall-wild

size-equivalent-wild

Figure 2. Results of principal components analysis, first two axes. (a) Oral teeth, two-dimensional data, 100 mm sampling area, alltwo-dimensional variables (feature mean length, s.d. of length, feature mean width, s.d. of width, preferred orientation, R and featuredensity.) PC1 accounts for 49% of the variance, with all the variables derived from scratch dimensions loading approximately equallyand positively (eigenvectors of 0.47–0.48). PC2, accounts for 22% of the variance, and reflects the two variables derived from scratchorientation (eigenvectors: preferred orientation ¼ 0.72; R ¼ 20.54). (b) Oral teeth, two-dimensional data, 300 mm sampling area; alltwo-dimensional variables. PC1 accounts for 58% of the variance, and all the variables derived from scratch dimensions load approxi-mately equally and positively (eigenvectors of 0.40–0.45), while preferred orientation and feature density load negatively (20.37,20.32, respectively). PC2, accounting for 18% of the variance, reflects strong positive loading of R (0.83) and negative loading offeature mean length and s.d. of length. (c) Oral teeth, three-dimensional data, parameters that differ (ANOVA). All seven par-ameters (Sa, Sk, Spk, Vmp, Vmc, Vvc and Sal) load approximately equally on PC1 (eigenvectors of 0.33–0.39) which accountsfor 93% of variance. (d) Oral teeth, three-dimensional data, all parameters. PC1 (60% of variance) reflects heaviest loading ofthe following parameters (.2.4, Sq, Sdr, Svk, Vvv, Sv(log), Sz(log). PC2 (21% of variance) reflects positive loadings (2.5–3.0) ofSmr2, Sal and Std, and negative loadings (22.5 to 23.7) of Ssk, Sku and Sdq. (e) LPJ teeth, three-dimensional data (log), par-ameters that differ (ANOVA). Most parameters (Sa, Sq, Sk, Spk, Vmp, Vmc, Vvc, Vvv and Svk) load equally and positively onPC1, with eigenvectors of 0.3 or above; the loading of Sal is less, and the loading on Smr2 is negative. ( f ) LPJ teeth, three-dimen-sional data (log), all parameters. The first two axes account for 82% of variance; 15 parameters load approximately equally on PC1,with eigenvectors of between 0.23 and 0.26. Key to specimen numbers: 1 ¼ NgR92, 2 ¼ NgR101, 3 ¼ NgR104, 4 ¼ HpyR118, 5 ¼HpyR163, 6 ¼ HpyR205, 7 ¼ HpyV01, 8 ¼ HpyV02, 9 ¼ HpyV03, 10 ¼ 37864, 11 ¼ 37865, 12 ¼ 37866, 13 ¼ 37867, 14 ¼ 37868,15 ¼ 37869, 16 ¼ 37870, 17 ¼ 37871, 18 ¼ 37872.

6 Tooth microtextures and diet in fishes M. Purnell et al.

are similar. PCA of the three-dimensional roughnessparameters demonstrated by ANOVA to differ signifi-cantly between populations (figure 2c) reveals thatNgR specimens are clearly separated from Hpy

J. R. Soc. Interface

populations along PC axis 1. There is no clear separ-ation of populations along PC axes 2 or 3. Analysisof all three-dimensional roughness parameters (figure2d)reveals that except for one specimen (HpyV03; ‘9’ in

–54–52–50 3

1264

578 9

1600 1610 1620 1630 1640

canonical axis 1

HpyV NgRHpyR(a)

–9–10

–8–7–6

10

12

111513

141618

17

–40 –35 –30 –25 –20

laboratory-raisedsmall-wildsize-equivalent-wild(b)

Figure 3. Results of stepwise linear discriminant analysis ofthree-dimensional roughness data. (a) Oral teeth: three micro-textural parameters—Spk, Svk and Smr2—are enough toassign specimens to their correct trophic groups (canonicalaxis 1 explains 99.94% of variance; scores for axis 1 are corre-lated with Spk; scores for axis 2 are correlated with Svk andSmr2; see the electronic supplementary material). (b) Lowerpharyngeal teeth: three microtextural parameters—Svk,Smr1 and Str—are enough to assign specimens to their correcttrophic groups (canonical axis 1 explains 99.77% of variance,axis 2 the remaining 0.23%; scores for axis 1 are correlatedwith Svk; scores for canonical axis 2 are correlated with Str.Circles are 95% confidence limits for the means. Numbersrefer to specimens (figure 2).

Tooth microtextures and diet in fishes M. Purnell et al. 7

figure 2) NgR and Hpy populations are separated alongPC axis 1. PC axis 2 also provides ecological separationof populations: HpyV ranging from negative to near-zero values, HpyR around zero, and NgR with positivevalues (the only exception is specimen HpyR205 (‘6’ infigure 2), the least worn HpyR tooth that plots withHpyV). There is no separation of populations alongPC axis 3 (10% of variance). Analyses based on thedataset where recently erupted teeth (with little wear)were substituted with data from the third tooth fromthe symphysis produced very similar results to thenon-substituted dataset, but with slightly betterseparation of populations.

LDA provides further evidence that difference in dietand feeding are reflected in microwear and micro-textural differences between populations. For stepwiseLDA of the 100 mm two-dimensional data (excludingorientation; see the electronic supplementary materialfor details), six variables are required for correct assign-ment of all specimens to their populations (canonicalaxis 1 accounts for 93% of variance, axis 2 the remain-ing 7%), but the significance of this LDA is doubtful(Wilks’ Lambda ¼ 0.007, p ¼ 0.42). Probabilities ofcorrect assignment to populations are all either 99 or100 per cent except for the least worn NgR specimen(101) which has a 49.9 per cent probability of beingassigned to the HpyV population. For the 300 mmtwo-dimensional data, five variables are required forcorrect assignment of all specimens to their populations(canonical axis 1 accounts for 99.7% of variance, axis 2the remaining 0.3%; Wilks’ Lambda ¼ 0.00004, p ¼0.0006). Probability of correct assignment to populationsis 100 per cent for all specimens (see the electronicsupplementary material for further details).

Stepwise LDA of the three-dimensional roughnessdata suggests that these data have greater potentialfor discrimination between populations and for dietarydiscrimination than the two-dimensional data becausefewer variables are required for correct and significantassignment. NgR specimens can be distinguished fromthe other populations on the basis of a single parameter(any of Sa, Sk, Spk, Vmp, Vmc, Vvc). Stepwise LDA ofall three-dimensional roughness parameters reveals thatthree (Spk, Svk, Smr2) are enough to assign all speci-mens to their correct groups, with clear separationalong the first canonical axis, representing an ecologicalspectrum from specialized epiphytic algae scrapers(HpyV), through facultative epilithic scrapers (HpyR),to specialized epilithic scrapers (NgR) (figure 3; Wilks’Lambda ¼ 0.002, p ¼ 0.0001; probability of correctassignment to populations ¼ 100% for all specimens; seethe electronic supplementary material). Cross-validationfurther supports this LDA as robust.

Clearly, both two-dimensional scored data andthree-dimensional roughness data from cichlid oralteeth are informative with regard to trophic ecology:no single PC axis provides perfect ecological separationof cichlid populations, but two axes combined producenon-overlapping distributions in PCA space; LDA cor-rectly discriminates between groups. Analyses basedon microwear compares favourably with analysesbased on stomach (and gut) contents, which vary con-siderably from individual to individual (see the

J. R. Soc. Interface

electronic supplementary material), with no clear-cutpatterns in the data. One ‘rogue’ individual from eachof the NgR and HpyV populations, for example, con-tained large volumes of insect larvae, presumably as aresult of opportunistic feeding just prior to capture.PCA of these data shows separation of HpyR fromthe other two populations along PC axis 1 (which cap-tures 43% of the variance), but no separation of NgRand HpyV. Results of stepwise LDA are comparable(see the electronic supplementary material). On bal-ance, for these samples, microwear and microtexturaldata provide a more reliable guide to dietary groupingsthan gut contents. This is supported by correlations ofPC scores from the analysis of stomachs with those fromthe analysis of three-dimensional variables (axes 1–3):there are no significant correlations between scores fromthe two analyses when all individuals are included, butexclusion of the two ‘rogue’ specimens yields significantcorrelations between PC 1 scores (rs 2 0.79, p ¼ 0.036;rogue specimens excluded from PCA of stomachs). N issmall, so caution is required, but this correlation supportsthe hypothesis that three-dimensional microwear on oralteeth records a dietary signal and that it is less likely tobe distorted by opportunistic feeding than analysisbased on stomach contents.

3.3. Multivariate analysis of three-dimensionaldata from pharyngeal teeth

Multivariate analysis of three-dimensional data fromthe LPJs is similarly informative with respect to diet.PCA of parameters that differed significantly betweengroups (ANOVA) provided separation of the threepopulations into non-overlapping clusters in a spacedefined by PC axes 1 and 2 (figure 2e). Together,these two axes account for 95% of the variance. PCaxis 1, in particular, provides good dietary separation,

8 Tooth microtextures and diet in fishes M. Purnell et al.

with laboratory-raised soft-diet fish having negativevalues, smaller-wild fish having values around zero,and size-equivalent-wild fish having more positivevalues. There is just a little overlap between the twowild groups along this axis; axis 1 is strongly correlatedwith stomach contents (rs ¼ 0.96, p ¼ 0.0005). Resultsof PCA of all three-dimensional parameters (figure 2f )are similar, but with more overlap between groups onPC axis 1. This axis generally corresponds to the samedietary axis as that of figure 2e (from negativelaboratory-raised, through smaller-wild, to positive size-equivalent-wild fish) and although there is overlapbetween each of the groups, the correlation with stomachcontents is significant (rs ¼ 0.85, p ¼ 0.015).

Analyses based on the second most-worn tooth, ontwo teeth from each fish, and on mean values for par-ameters derived from two teeth per fish, are reportedin the electronic supplementary material. Analysesbased on mean values gave results that were similar toanalyses based on most-worn teeth in terms of separ-ation of dietary groups, but no better. Other datasetswere less useful in distinguishing between groups.

Results of LDA of LPJ teeth are comparable to thePCA results. When all parameters are included inthe stepwise LDA, three (Svk, Smr1, Str) are enough tocorrectly discriminate between populations (figure 3;Wilks’ Lambda ¼ 0.01, p ¼ 0.0016; probability of correctassignment to populations is 99.9 or 100%, except for twospecimens with probabilities of 98 and 91%, so the prob-ability of any of the specimens being mis-assigned is low;see the electronic supplementary material). The firstcanonical axis is correlated with stomach contents (rs ¼

0.852, p ¼ 0.015). Cross-validation further supports thisLDA as robust.

Analysis based on parameters that differ (ANOVA;most-worn tooth, filtered data, log transformed) indi-cates that only three are required to assign all samplesto their correct trophic group (Sq, Svk and Vvv), andthat the first axis is correlated with stomach contents(rs ¼ 0.964, p ¼ 0.0005). However, the significance ofthis LDA is doubtful (Wilks’ Lambda ¼ 0.09, p ¼0.07; see the electronic supplementary material).Inclusion of two additional parameters (Spk andVmc) produces a more informative LDA, also with cor-rect assignment of all specimens (Wilks’ Lambda ¼0.0002, p ¼ 0.0025; probability of correct assignmentto populations ¼ 100% for all specimens; see theelectronic supplementary material).

4. DISCUSSION

For the pharyngeal teeth, the results of microtexturalanalysis are correlated with diet, and for the oralteeth microwear and microtextural data performbetter than stomach contents in discriminating betweenthe populations. This may seem surprising at first sight,but in small sample sizes, opportunistic feeding in thehours before capture clearly has the potential to intro-duce significant noise into trophic analysis based onthe ‘snapshot’ provided by stomach contents. Toothmicrowear and microtextural data, on the other hand,integrate the effects of feeding on the teeth over a

J. R. Soc. Interface

longer period of time. This is especially true of non-occlusal teeth, where unlike the wear facets onmammal teeth microwear and microtextural signal isnot constantly reset through tooth-on-tooth abrasion.Furthermore, A. alluaudi and other pharyngeal crush-ing cichlids do not simply ingest crushed molluscs:ejection of crushed shells is a major component offood processing in these fishes [23], and this has obviousconsequences for dietary analysis based in gut contents.

Analysis of microwear patterns and textures in fishesand other aquatic vertebrates clearly has great potential.The small sample sizes of this exploratory analysis reducethe statistical power of some of our tests, but both two-dimensional and three-dimensional approaches have thepower to discriminate between fishes with differentdiets and different ways of obtaining similar diets, evenwhen a single tooth per individual is analysed, andwith the constraints on sampling area imposed by thesmall size of many fish teeth. This holds both for differ-ences between oral teeth and for differences betweenpharyngeal teeth. Inevitably sampling of only onestandardized tooth position will lead to greater noise ina dataset, but even with small sample sizes, a trophicsignal can be detected. In comparing our microwearresults with stomach contents, we are fully aware thatanalysis of stomach contents based on small samplesizes would not be considered reliable. However, thatmicrowear analysis of small samples can correctly assignindividuals to subtly different trophic sub-groups givesa good indication of its power.

The power of two-dimensional microwear analysis isreduced by the well-known difficulties of operator biasand the lack of comparability between studies carriedout by different groups and using different methods.Three-dimensional analysis avoids these pitfalls, andoffers the potential for methodological standardizationthat would allow comparability of results and dataacross studies and between laboratories (not possiblewith two-dimensional approaches). Our results demon-strate that analysis of three-dimensional data hasgreater power than two-dimensional analysis in dietarydiscrimination, and this clearly argues that three-dimensional approaches are the better way forward.Currently analysis of orientation of microwear features,useful for analysis of feeding mechanics and tooth move-ments [11], provides an exception to this: extraction oforientation data is simpler and easier using two-dimensional approaches, and orientation data probablysuffers from fewer operator errors (when scoring micro-wear features, aligning a cursor with feature orientationis less error prone than precise and accurate recogni-tion of feature end points). However, methods for theextraction of orientations from three-dimensionalmicrotextural data are under development [28].

In terms of approaches to data analysis, ANOVA,PCA and LDA all have potential, but PCA, in particu-lar, has the power to ordinate samples on the basis ofmicrotextural roughness data alone, with few assump-tions and without a priori knowledge of diets andtrophic niche.

In conclusion, multivariate analysis of microtexturalroughness data represents a powerful tool for testinghypotheses of trophic similarity and difference in

Tooth microtextures and diet in fishes M. Purnell et al. 9

fishes and other aquatic vertebrates which are extinct/fossil or otherwise lack dietary data. It has the potentialto provide robust tests of macroevolutionary scenariosthat invoke dietary and trophic change in extinctaquatic vertebrates. In addition, because tooth micro-wear/microtexture develops over a period of time(several days at least), it is less prone to the snapshotbias of stomach content analyses. It is more sensitiveto subtle dietary differences than isotopic analysis.Thus, quantitative microtextural analysis of microwearhas a useful role to play, complementing existingapproaches, in trophic analysis of extant fish andother aquatic vertebrates.

The Novartis Foundation funded M.A.P.’s visit to Leiden towork with F.G.; Frans Witte and Wessel de Priester arethanked for their assistance during this visit. Additionalsupport for M.A.P. was provided by NERC grant NE/G018189/1. Paul Hart, Laurent Darras and Dave Baines arethanked for discussion and analytical assistance.

REFERENCES

1 Robinson, B. W. & Wilson, D. S. 1998 Optimal foraging,specialization, and a solution to Liem’s paradox. Am.Nat. 151, 223–235. (doi:10.1086/286113)

2 Schluter, D. 1994 Experimental evidence that competitionpromotes divergence in adaptive radiation. Science 266,798–801. (doi:10.1126/science.266.5186.798)

3 Bellwood, D. R., Wainwright, P. C., Fulton, C. J. & Hoey,A. S. 2006 Functional versatility supports coral reef biodi-versity. Proc. R. Soc. B 273, 101–107. (doi:10.1098/rspb.2005.3276)

4 Vermeij, G. J. 1977 The Mesozoic Marine Revolution evi-dence from snails predators and grazers. Paleobiology 3,245–258.

5 Lauder, G. V. 1995 On the inference of structure fromfunction. In Functional morphology in vertebrate palaeon-tology (ed. J. J. Thomason), pp. 1–18. Cambridge, UK:Cambridge University press.

6 Walker, A. C., Hoeck, H. N. & Perez, L. 1978 Microwearof mammal teeth as an indicator of diet. Science 201,808–810. (doi:10.1126/science.684415)

7 Teaford, M. F. 1988 A review of dental microwear and dietin modern mammals. Scan Microsc. 2, 1149–1166.

8 Scott, R. S., Ungar, P. S., Bergstrom, T. S., Brown, C. A.,Grine, F. E., Teaford, M. F. & Walker, A. 2005 Dentalmicrowear texture analysis shows within-species dietvariability in fossil hominins. Nature 436, 693–695.(doi:10.1038/nature03822)

9 Purnell, M. A., Hart, P. J. B., Baines, D. C. & Bell, M. A.2006 Quantitative analysis of dental microwear in three-spine stickleback: a new approach to analysis of trophicecology in aquatic vertebrates. J. Anim. Ecol. 75,967–977. (doi:10.1111/j.1365-2656.2006.01116.x)

10 Purnell, M. A., Bell, M. A., Baines, D. C., Hart, P. J. B. &Travis, M. P. 2007 Correlated evolution and dietarychange in fossil stickleback. Science 317, 1887. (doi:10.1126/science.1147337)

11 Williams, V. S., Barrett, P. M. & Purnell, M. A. 2009Quantitative analysis of dental microwear in hadrosauriddinosaurs, and the implications for hypotheses of jawmechanics and feeding. Proc. Natl Acad. Sci. USA 106,11 194–11 199. (doi:10.1073/pnas.0812631106)

12 Scott, R. S., Ungar, P. S., Bergstrom, T. S., Brown, C. A.,Childs, B. E., Teaford, M. F. & Walker, A. 2006 Dental

J. R. Soc. Interface

microwear texture analysis: technical considerations. J.Hum. Evol. 51, 339–349. (doi:10.1016/j.jhevol.2006.04.006)

13 Ungar, P. S., Merceron, G. & Scott, R. S. 2007 Dentalmicrowear texture analysis of Varswater Bovids andearly pliocene paleoenvironments of Langebaanweg, Wes-tern Cape Province, South Africa. J. Mamm. Evol 14,163–181. (doi:10.1007/s10914-007-9050-x)

14 Ungar, P. S., Grine, F. E. & Teaford, M. F. 2008 Dentalmicrowear and diet of the Plio-Pleistocene homininParanthropus boisei. PLoS ONE 3, e2044. (doi:10.1371/journal.pone.0002044)

15 Jiang, X., Scott, P. J., Whitehouse, D. J. & Blunt, L. 2007Paradigm shifts in surface metrology. II. The currentshift. Proc. R. Soc. A 463, 2071–2099. (doi:10.1098/rspa.2007.1873)

16 Grine, F. E., Ungar, P. S. & Teaford, M. F. 2002 Errorrates in dental microwear quantification using scanningelectron microscopy. Scanning 24, 144–153. (doi:10.1002/sca.4950240307)

17 Seehausen, O. 2006 African cichlid fish: a model systemin adaptive radiation research. Proc. R. Soc. B 273,1987–1998. (doi:10.1098/rspb.2006.3539)

18 International Organization for Standardization. 2007Geometrical product specifications (GPS)—Surface tex-ture: Areal—Part 2: Terms, definitions and surfacetexture parameters, p. 1–42.

19 Schulz, E., Calandra, I. & Kaiser, T. M. 2010 Applyingtribology to teeth of hoofed mammals. Scanning 32,162–182. (doi:10.1002/sca.20181).

20 Estebaranz, F., Galbany, J., Martınez, L. M. &Perez-Perez, A. 2007 3-D interferometric microscopyapplied to the study of buccal enamel microwear. InDental perspectives on human evolution: state of the artresearch in dental paleoanthropology (eds S. E. Bailey &J.-J. Hublin), pp. 391–403. Amsterdam, The Netherlands:Springer.

21 Seehausen, O., Lippitsch, E., Bouton, N. & Zwennes, H.1998 Mbipi, the rock-dwelling cichlids of Lake Victoria:description of three new genera and fifteen new species(Teleostei). Ichthyol. Explor. Freshw. 9, 129–228.

22 Hoogerhoud, R. J. C. 1986 Ecological morphology of somecichlid fishes. DPhil, University of Leiden, Leiden,Netherlands.

23 Hoogerhoud, R. J. C. 1987 The adverse-effects of shellingestion for molluscivorous cichlids, a constructional mor-phological approach. Netherlands J. Zool. 37, 277–300.(doi:10.1163/002829687X00099)

24 Smits, J. D. 1996 Trophic flexibility through spatial reallo-cation of anatomical structures in the cichlid fishAstatoreochromis alluaudi. DPhil, University of Leiden,Leiden, The Netherlands.

25 Witte, F., Barel, C. D. N. & Hoogerhoud, R. J. C. 1990Phenotypic plasticity of anatomical structures and itsecomorphological significance. Netherlands J. Zool. 40,278–298. (doi:10.1163/156854289X00309)

26 Smits, J. D., Anker, G. C., Witte, F. & Barel, K. D. N.1997 Comparative functional anatomy of the pharyngealjaw apparatus in two morphs of Astatoreochromis alluaudi(Pisces, Cichlidae). Netherlands J. Zool. 47, 313–347.(doi:10.1163/156854297X00049)

27 Ungar, P. S. 2001 Microware 4.0. A semiautomated imageanalysis procedure for the quantification of dental micro-wear, 4th edn. Fayetteville, Arizona, USA: Privatelypublished.

28 Jiang, X. Q., Zeng, W. H., Scott, P., Ma, J. W. & Blunt, L.2008 Linear feature extraction based on complex ridgelettransform. Wear 264, 428–433. (doi:10.1016/j.wear.2006.08.040)

1

ELECTRONIC SUPPLEMENTARY INFORMATION ADDITIONAL DETAILS OF METHODS Studied specimens of Haplochromis purple-yellow (Hpy) and Neochromis gigas (Ng) were all males of similar standard length, and were selected blind with respect to tooth wear. Dentaries and premaxillae were dissected from preserved individuals, skin and muscle were carefully peeled away using tweezers, and jaws were then boiled in individual vials of distilled water for 4 hours to further clean tooth surfaces. Care was taken to avoid contact between instruments and teeth, but in order to further minimise the possibility of inadvertently recording preparation artefacts as microwear, the tooth closest to the symphysis, where right dentary and premaxilla were physically separated from their neighbours, was not analysed.

For scanning electron micrography and data acquisition jaws were sputter coated with gold (2 minutes). Both 2D and 3D data were collected from original gold-coated tooth surfaces. SEM images were acquired using a Jeol 6400 Scanning Electron microscope (Secondary electron images). Some teeth were subsequently reimaged using a Hitachi S-3600N. In 2D microwear analysis, most statistical testing was based on the summary outputs from Microware for each tooth (feature mean length, feature density, preferred orientation, and R (providing a measure of angular dispersion [1])). For a few tests, how microwear varies within individuals for example, raw data were extracted from Microware 4.02 output as x/y coordinates and processed by using simple trigonometric functions in MS Excel to derive the length, width, and long-axis orientation for every feature in a sample site.

The Alicona Infinite Focus microscope G4b has a CCD of 1624 x 1232 pixels. In theory, for a field of view of 145 µm, this equates to a lateral sampling distance of 0.09 µm, but the limits imposed by the wavelength of white light mean that lateral optical resolution is actually about 0.35 - 0.4 µm. For all samples, vertical resolution was set at 20 nm, and the lateral resolution factor for the IFM was set at 0.3. Exposure and contrast (gamma) settings were adjusted to maximise data quality in terms of measurement repeatability (this is estimated automatically by the IFM software during data capture) for each sample. Adjusting exposure and contrast do not affect the values for 3D measurements. All data were acquired using polarized co-axial incident illumination.

All 3D data were processed using the Alicona IFM software (version 2.1.2) as follows: 1, data editing (to delete dirt and dust particles from the surface); 2, data levelling and filtering (all points levelling; application of Gaussian wavelength filter to remove long wavelength features of the tooth surface (gross tooth form)); 3, generation of standard parameters from analysis of the resulting roughness surface. Most roughness parameters conform to the current draft ISO 25178-2 standards [2], and include: Height parameters (quantifying the distribution of height values along the z-axis), Spatial parameters (quantifying direction and spatial periodicity of the surface), Hybrid parameters (combining the information present on the x, y and z axes of the surface, quantifying aspects of the spatial shape of the data), and Functional parameters related to measures of volumes, such as peak material, calculated from the areal bearing ratio curve. All parameters analysed and brief definitions are provided in Table S1.

2

Table S1. Short definitions and categorization of 3D areal surface texture parameters

parameter definition Sa Average height of surface height Sq Root-Mean-Square height of surface height Sp Maximum peak height of surface height Sv Maximum valley depth of surface height Sz Maximum height of surface height S10z Ten point height of surface segmentation Ssk Skewness of height distribution of surface height Sku Kurtosis of height distribution of surface height Sdq Root mean square gradient of the surface hybrid Sdr Developed interfacial area ratio hybrid Sk Core roughness depth, Height of the core material functional Spk Mean height of the peaks above the core material functional Svk Mean depth of the valleys below the core material functional Smr1 Surface bearing area ratio (the proportion of the surface which consists of peaks above the

core material) functional

Smr2 Surface bearing area ratio (the proportion of the surface which would carry the load) functional Vmp Material volume of the peaks of the surface (ml/m²) functional Vmc Material volume of the core of the surface (ml/m²) functional Vvc Void volume of the core of the surface (ml/m²) functional Vvv Void volume of the valleys of the surface (ml/m²) functional Vvc/Vmc Ratio of Vvc parameter to Vmc parameter functional Sal Auto correlation length. Horizontal distance of the auto correlation function (ACF) which

has the fastest decay to the value 0.2. Large value: surface dominated by low frequencies. Small value: surface dominated by high frequencies.

spatial

Str Texture aspect ratio (values range 0-1). Ratio from the distance with the fastest to the distance with the slowest decay of the ACF to the value. 0.2-0.3: surface has a strong directional structure. > 0.5: surface has rather uniform texture.

spatial

Std Texture direction (°). Derived from the maximum of the angular power spectrum. Std = 90° means a dominant lay parallel to the y-axis.

spatial

Stdi Texture direction index (values range 0-1). Average value of the angular power spectrum divided by the value of the dominating direction. Smaller values correspond to stronger directional structures.

spatial

In order to test comparability and relative discriminatory power of the different approaches,

3D data from the cichlid oral tooth surfaces were acquired from the same dentary teeth as had been subjected to 2D analysis. The only exception to this was specimen HpyR118, the dentary tooth of which was accidentally destroyed subsequent to SEM analysis. 3D data from the premaxilla of this specimen were substituted (at 100µm scale sampling 2D analysis results indicate that scratch length on dentary and premaxilla teeth in this specimen do not differ). It became apparent during acquisition of 3D data, as a result of our increasing experience, that two of the teeth in the original data set were less worn that others because they were only recently erupted at the time the fish were caught. In order to test whether a modified sampling strategy was more effective in dietary discrimination, we conducted a second analysis in which data from the 3rd tooth from the symphysis was substituted in cases where the second tooth was only recently erupted.

PCA and LDA of oral teeth were based on data acquired from dentaries only. For LPJ Spearman Rank correlation was used to test relationships between fish size, diet and microtexture (in the form of results from PCA and LDA). For 2D microwear data, where the results of Shapiro-Wilks tests indicated that variables were non-normally distributed (p > 0.05), non parametric (Wilcoxon/Kruskal Wallis tests, 2-sample Z tests), and Chi2 tests were employed. Tests for differences within individual fish (e.g. dentary-premaxilla comparisons) and within individual teeth (300 µm – 100 µm sample area comparisons) employed matched pairs t-tests. For 2D orientation

3

data, Watson-Williams multisample tests were used: samples where the null hypothesis that the orientation has a uniform (non-preferential) distribution could not be rejected (Rayleigh test, p = 0.05) were excluded from such analyses (because where distributions are uniform mean orientation data and tests based on non-uniform distributions are meaningless). These tests of orientation data were carried out using Oriana 2.02e [3]. All other statistical analysis was carried out using JMP 8.

For 3D microtextural datasets, a few variables were found to deviate from normality, but for most of these we were unable to reject the null hypothesis of normality (Shapiro-Wilks) for log-transformed data, and the latter were used for analysis. Where homogeneity of variance tests revealed evidence of unequal variances, Welch ANOVA was used.

Analysis of correlations between 2D and 3D data for oral teeth was based on the 100 µm data set (the 300 µm dataset is unsuitable for analysis because of the differences in scale between the 3D and 2D data). Analysis of correlations of the LPJ teeth was more difficult because they are not amenable to collection of 2D data. In this case we tested for rank correlations between 3D parameters and a qualitative scale of relative roughness generated by seven people independently ranking 2D images of the surfaces from which 3D data were acquired, from smoothest to roughest. Although no two individuals produced the same ranking, pairwise correlations between rankings were all significant (rs range from 0.67 to 0.98, p < 0.05). The mode of rank position for each sample was taken as a proxy for qualitative roughness and used for testing for correlation with quantitative roughness parameters. This ranking must be used with caution: although there was a high degree of agreement in recognizing the 2 or 3 specimens corresponding to the qualitatively most and least worn, there is less agreement about the specimens between the extremes. Indeed, if the two least and the three most worn specimens are excluded from analysis only 3 of 21 pairwise correlations are significant (specimens selected as most/least worn based on mode of rank position; 3 most worn specimens had equal rank).

For the LPJ data, where longer wavelength features of the surface might reflect tooth surface wear rather than aspects of original tooth form and might thus be informative with respect to diet, analysis was carried out on both levelled non-filtered 3D data, and levelled and Gaussian-filtered data. ANOVA, PCA and LDA was applied to four different datasets: one comprising data from only the most worn tooth from each individual (based on visual inspection; non-filtered and filtered), one comprising data from only the second most worn tooth of each (non-filtered and filtered), one comprising the average of the two teeth from each individual (non-filtered and filtered), and one comprising all the teeth sampled (non-filtered and filtered).

The significance of linear discriminant analyses reported in the main text was assessed using Wilks’ Lamda, and robustness was tested through cross validation using a jack knife approach: for both the oral and LPJ 3D datasets one specimen was randomly deleted from the dataset, the LDA was rerun, and the resulting discriminant function was then applied to the complete dataset with the deleted specimen reinstated. Three iterations of this process for each dataset failed to produce a LDA that resulted in specimens being misassigned, indicating that the LDAs are robust.

Stomach and gut content data were also collected for most of the fish (see ESM Table S2).

4

Table S2: Stomach content data for Haplochromis p-y, Neochromis gigas and Astatoreochromis alluaudi. Stomach contents data for Hpy and NgR was collected by OS. All stomachs except NgR104 were full to very full; for the empty stomach, gut contents were used for analysis. For Haplochromis and Neochromis gut contents were quantified in terms of estimated volume percentages. Dietary data for lab raised A. alluaudi are based on what they were fed. For five of the six wild fish gut contents data were recorded in the field (by Hoogerhoud). Data for the sixth fish was unavailable to us. Melanoides and Bellamya are gastropods; Spaerium and Caeltura are bivalves. Taxon specimen museum

number ecology Stomach contents

Neochromis gigas NgR92 RMNH.PISC 37855

Rock scraper

5% cladophora, 95% insect larvae

Neochromis gigas NgR101 RMNH.PISC 37856

Rock scraper

90% cladophora, 10% insect larvae

Neochromis gigas NgR104 RMNH.PISC 37857

Rock scraper

100% cladophora

Haplochromis purple-yellow

HpyR118 RMNH.PISC 37858

Rock scraper

5% cladophora, 30% macrophyte chunks, 32.5% filamentous blue greens, 32.5% diatoms

Haplochromis purple-yellow

HpyR163 RMNH.PISC 37859

Rock scraper

14.3% macrophyte chunks, 31% filamentous blue greens, 38.1% brown detritus, 9.5% zooplankton

Haplochromis purple-yellow

HpyR205 RMNH.PISC 37860

Rock scraper

31.6% cladophora, 10.5% insect larvae, 21.1% macrophyte chunks, 6.3% filamentous blue greens, 12.6% brown detritus, 12.6% diatoms, 5.3% sand

Haplochromis purple-yellow

HpyV01 RMNH.PISC 37861

Veg. scraper 100% insect larvae

Haplochromis purple-yellow

HpyV02 RMNH.PISC 37862

Veg. scraper 100% macrophyte fragments

Haplochromis purple-yellow

HpyV03 RMNH.PISC 37863

Veg. scraper 5% insect larvae, 75% macrophyte fragments, 10% filamentous blue greens, 10% diatoms

Astatoreochromis alluaudi

37864 RMNH.PISC 37864

Lab-raised minced heart and liver, with vitamins and Tetramin flakes

Astatoreochromis alluaudi

37865 RMNH.PISC 37865

Lab-raised minced heart and liver, with vitamins and Tetramin flakes

Astatoreochromis alluaudi

37866 RMNH.PISC 37866

Lab-raised minced heart and liver, with vitamins and Tetramin flakes

Astatoreochromis alluaudi

37867 RMNH.PISC 37867

Smaller-wild

9.4 mg Bellamya; 4.46 mg Melanoides; 1 mg ostracods

Astatoreochromis alluaudi

37868 RMNH.PISC 37868

Smaller-wild

Bellamya, Melanoides, ostracods, cladocera, chironomids present

Astatoreochromis alluaudi

37869 RMNH.PISC 37869

Smaller-wild

7.1 mg Bellamya; 3.4 mg Melanoides; ostracods, cladocera, chironomids present

Astatoreochromis alluaudi

37870 RMNH.PISC 37870

Size-equiv. wild

14.5 mg Bellamya; 49 mg Melanoides; 2.6 mg Spaerium; 5 mg Caeltura; 2.2 mg ostracods

Astatoreochromis alluaudi

37871 RMNH.PISC 37871

Size-equiv. wild

3.18 Bellamya; 21.4 mg Melanoides; 12.86 mg Spaerium; 6.5 mg ostracods

Astatoreochromis alluaudi

37872 RMNH.PISC 37872

Size-equiv. wild

5

ADDITIONAL DETAILS OF ANALYSIS OF ORAL TEETH Table S3: Results of statistical hypothesis testing to determine whether 2D microwear sampled on dentary and premaxilla differs (300µm and 100µm fields of view; matched pairs t-test, dentary and premaxilla pair within fish matched. Orientation analysis based on Watson-Williams multisample test). 100µm, pairwise dentary - premaxilla comparisons Reject? t p Feature length does not differ, dent-pmax no 1.95 0.09 Feature R does not differ, dent-pmax No 0.34 0.74 Feature density does not differ, dent-pmax No -0.47 0.65 F df p Feature orientation does not differ, dent-pmax No 2.00 1,13 0.18 300µm, pairwise dentary - premaxilla comparisons Reject? t p Feature length does not differ, dent-pmax Yes -3.23 8 0.012 Feature R does not differ, dent-pmax No 1.62 8 0.14 Feature density does not differ, dent-pmax No -0.11 8 0.91 F df p Feature orientation does not differ, dent-pmax No 4.31 1,13 0.06 Table S4. Results of statistical hypothesis testing to determine whether microwear sampled at different scales differs within fish (300 µm and 100 µm fields of view; matched pairs t-test. Orientation analysis based on Watson-Williams multisample test). Dentary, pairwise 100-300 comparisons Reject? t p Feature length does not differ, 100-300 Yes 8.83 <0.0001 Feature R does not differ, 100-300 No 0.97 0.36 Feature density does not differ, 100-300 Yes -5.57 0.0005 F df p Feature orientation does not differ, 100-300 Yes 5.68 1,13 0.03

As reported in the main text, exploratory ANOVA of the first 3D data set from the oral teeth (same teeth sampled as 2D analysis) revealed significant differences between Sa, Sk, Spk, Vmp, Vmc, Vvc, and Sal. A Tukey HSD procedure revealed significant pairwise differences for most of these variables between NgR and both Hpy populations. ANOVA of the data set with two substitutions of 3rd teeth from the symphysis revealed similar results: significant differences between Sa, Sk, Spk, Vmp, Vmc, Vvc, and Sal (Welch ANOVA). Tukey pairwise comparisons however, were a little different: most variables (Sa, Sk, Vmc, Vvc) differed between HpyV and NgR, indicating that these should be assigned to different groups, with HpyR not differing significantly from either of the other populations. Spk and Vmp differentiated NgR from both Hpy populations. Analysis of correlations between 2D data and 3D parameters revealed that in oral teeth there are no significant correlations between any 3D parameters and feature length, feature width, feature mean orientation, or R. Fifteen parameters are correlated with density, most of which are height and functional parameters (Sa, r = 0.79, p = 0.01; Sq, r = 0.87, p = 0.002; Sp (log), r = 0.75, p = 0.02; Sv (log), r = 0.90, p = 0.001; Sz (log), r = 0.88, p = 0.002; S10z (log), r = 0.73, p = 0.02; Sdr, r = 0.81, p = 0.01; Sk, r = 0.73, p = 0.02; Spk, r = 0.68, p = 0.04; Svk, r = 0.92, p = 0.00; Vmp, r = 0.67, p = 0.05; Vmc, r = 0.75, p = 0.02; Vvc, r = 0.71, p = 0.03; Vvv, r = 0.92, p = <0.001; Sal, r =, 0.71, p = 0.03). In all cases the correlations are positive: as feature density increases, so do the roughness parameters. For the LPJ teeth, qualitative ranking of surface roughness was correlated with twelve quantitative roughness parameters (Sa, rs = 0.7227, p = 0.0278; Sq, rs = 0.6975, p = 0.0367; Sp , rs = 0.9412, p = 0.0002; Sz, rs = 0.8404, p = 0.0046; S10z, rs = 0.874, p = 0.0021; Sdq, rs = 0.832, p = 0.0054; Sdr, rs = 0.832, p = 0.0054; Sk, rs = 0.8152, p = 0.0074; Spk, rs = 0.7899, p = 0.0113; Vmp , rs = 0.7899, p = 0.0113; Vmc, rs = 0.7227, p = 0.0278; Vvc, rs = 0.8152, p = 0.0074). These are a mixture of functional, height and hybrid parameters (see Table S1). All correlations are positive: as qualitative roughness increases, so do the quantitative roughness parameters. Figure S2 provides a

6

degree of visualization, showing renderings of unfiltered and filtered 3D surface data for the qualitatively most worn (specimen 37872) and least worn of the LPJ teeth (specimen 37864. It is worth noting that unlike the results of quantitative analysis, the qualitative ranking of roughness is not correlated with diet (rs = 0.49, p = 0.27).

Principal components analysis of all 2D variables is discussed in the main text. For the 100µm data, reducing the number of variables analysed does not result in ordinations with axes that correspond closely to ecological differences between samples, but they are informative nonetheless. PCA based on ‘standard’ variables of feature mean length, preferred orientation, R and feature density, for example, gives the following results. PC1 accounts for 39% of the variance, and reflects positive loadings of feature mean length and R, with negative loadings from preferred orientation and feature density. There is no separation of populations long this axis. PC2 accounts for 31% of the variance, feature mean length and preferrred orientation loading positively, R and feature density loading negatively. Along this axis, HpyR are separated from the other populations. PC3 also accounts for a significant amount of variance (20%) and reflects positive loading of feature mean length and feature density, and negative loading of preferred orientation and R. N. gigas and HpyV are separated along this axis. Plotting PC2 against PC3 thus results in non-overlapping distribution of the three populations. For the 300µm 2D data, analyses based on ‘standard’ variables of feature mean length, preferred orientation, R and feature density give similar results to analyses based on all 2D variables. Preferred orientation and feature density load positively on PC1, with feature mean length loading negatively. This axis accounts for 54% of the variance. PC2 (27% of variance) reflects strong positive loading of R (0.92) and negative loading of feature mean length. Compared to the seven variable analysis, separation of populations in PCA space is reduced, and separation of HpyV from rock scrapers along axis two is incomplete, with specimen NgR92 plotting among HpyV specimens. This highlights a problem with 2D scoring approaches to microwear: where surfaces are very rough, individual features are more difficult to recognise and score, and the resulting microwear data may be misleading.

Stepwise LDA of the 2D 100µm data reveals that 7 variables are required for correct assignment of all specimens to their populations, but that correct assignment of all but one specimen (NgR101, misclassified as HpyV) is achieved with only 4 variables (feature mean length, SD of length, preferred orientation, SD of orientation). Omitting orientation data, values for which, as noted above, are not meaningful for some populations, 6 variables are required to assign all specimens to their correct groups (feature mean length, SD of length, feature mean width, SD of width, R and feature density). Canonical axis 1 explains 93% of variance, axis 2 the remaining 7%. The significance of the LDA is doubtful (Wilks' Lambda, 0.007, p = 0.42), and although the probabilities of correct assignment to groups for all teeth except one are are either 99% or 100%, the least worn NgR specimen (NgR101) has a 49.9% probability of being assigned to the HpyV population. Scores for canonical axis 1 are correlated with feature length (r = -0.80, p = 0.01) and the SD of length (r = -0.74, p = 0.02). No other correlations between variables and canonical scores are significant. Analysis including 5 of these variables (all but feature mean width) correctly assigns all specimens except NgR101, but the significance of the LDA is again doubtful (Wilks’ Lamda 0.015, p = 0.16).

Analysis of the 300µm data produces a more informative LDA. Five variables are required to assign all specimens to their correct groups (SD of length, feature mean width, SD of width, preferred orientation, R). Canonical axis 1 explains 99.7% of variance, axis 2 the remaining 0.3%. Wilks' Lambda is significant (0.00004, p = 0.0006) and the probability of correct assignment to populations is 100% for all specimens. Scores for canonical axis 1 are correlated with feature width SD (r = -0.7051, p = 0.034); scores for canonical axis 2 are correlated with feature orientation (r = 0.83, p = 0.005). No other correlations between variables and canonical scores are significant. Three variables (feature mean width, preferred orientation, R) correctly assign all but one specimen (HpyV01 misclassified as HpyR; Wilks’ Lamda = 0.04, p = 0.02. If orientation is excluded, correct assignment of all specimens cannot be achieved.

7

Results of PCA analysis of the 3D dataset where recently erupted teeth were substituted with data from the 3rd tooth from the symphysis (specimens HpyR163 and HpyR205) were as follows: Results from analysis of the six parameters that differ significantly produces similar results to the non-substituted dataset, but with better separation of populations. All six parameters load positively and approximately equally (eigenvectors 0.38 – 0.42) on PC1, which accounts for 93% of the variance. NgR specimens are clearly separated from Hpy populations on this axis. PC2 (4% variance) reflects heavy positive loading of Vvv (0.82), and negative loading of Spk and Vmp (0.36, 0.4). With the exception of HpyV01, HpyV and HpyR are separated along this axis (HpyV positive values, HpyR negative), and the three populations have non-overlapping distributions in the space defined by PC1 and PC2. Analysis of all 3D roughness parameters produces very similar results to the non-substituted dataset. The same parameters load on PC1 (59% of variance), with similar eigenvectors; NgR and Hpy populations are separated along this axis, with the exception of one specimen (HpyV mp03). PC2 (23% of variance) reflects positive loadings (2.3 – 3.2) of Smr2, Sal and Stdi, and Vvc (functional and spatial parameters), and negative loadings (-2.4 - -3.5) of Ssk, Sku, Sp, and Sdq (height and hybrid parameters). This axis provides good ecological separation of populations, slightly better than the non-substituted dataset, with non-overlapping distribution in the space defined by PC1 and PC2. HpyV range from negative to near zero values, HpyR are around zero, and NgR have positive values.

Results of stepwise LDA of the oral tooth 3D roughness data are reported in the main text and in Fig. 3. Three parameters (Spk, Svk, Smr2) are enough to assign all specimens to their correct populations (canonical axis 1 explains 99.94% of variance, axis 2 the remaining 0.06%; Wilks' Lambda, = 0.002, p = 0.0001; probability of correct assignment to groups is 100% for all specimens; scores for canonical axis 1 are correlated with Spk (r = 0.8729, p = 0.002); scores for canonical axis 2 are correlated with Svk (r = 0.80, p = 0.01) and Smr2 (r = -0.97, p < 0.0001). No other correlations between parameters and canonical scores are significant.

LDA of the 3D dataset with substitution of less worn teeth has similar discriminatory power to the unsubstituted dataset, but some of the parameters that discriminate differ. When parameters that differ significantly are analysed, 4 variables are required to assign specimens to their correct groups (Sk, Spk,Vmp, Vmc), but when all variables are analysed only three are required (Vmp, Std and Stdi). In both cases, the three populations are separated along the first canonical axis, and this axis corresponds to the HpyV, HpyR, NgR ecological spectrum discussed in the main text.

As noted in the main text, stomach contents vary considerably from individual to individual: two of the rock scraping NgR contained large volumes of green algae (Cladophora; 90% and 100%) and two of the HpyV had large volumes of macrophyte chunks (100% and 75%). The third specimen from each of these populations contained large volumes of insect larvae. HpyR stomach contents are more varied. PCA results show separation of HpyR from the other two populations along PC1, but HpyV and NgR overlap one another. This axis accounts for only 43% of the variance, and reflects equal positive loadings (4.0 - 4.4) of filamentous blue-green algae, brown detritus, zooplankton, and sand. PC2 accounts for 23% of the variance, and reflects heavy loading of diatoms and thus separates members of the HpyR population from one another, rather than separating any of the populations. PC3 accounts for 15% of the variance, and reflects heavy positive loading of small macrophyte fragments, and heavy negative loading of cladophora. This axis thus goes some way towards separation of HpyV and NgR populations, but the two specimens with high volumes of insect larvae are, unsurprisingly not separated.

The results of stepwise LDA are comparable. Five variables are required for all specimens to be assigned to their correct groups, but the confidence with which two specimens (NgR92 and HpyV01 – the two with large volumes of insect larvae) are assigned to their groups is low, and does not increase with addition of more variables. Results of stepwise LDA of stomach contents are less clear than analyses of microwear data: HpyR are separated from other populations along canonical axis 1 (strongly loaded by macrophyte chunks), while HpyV and NgR fishes are distributed along axis 2, reflecting weighting of cladophora and insect larvae.

8

ADDITIONAL DETAILS OF ANALYSIS OF LPJ TEETH The main text reports the results of ANOVA of the filtered 3D parameters derived from the

most worn tooth in each individual. ANOVA of the filtered 3D parameters derived from the second teeth sampled in each individual found significant differences only in Svk, Sal and Vvv (Welch ANOVA). Analysis based on mean values derived from filtered parameter values for both teeth sampled in each individual yielded similar results to the analysis of the most worn teeth. Significant Differences were found between a similar but not identical list of height, functional and spatial parameters (Sa, Sq, Sk, Spk, Svk, Vmc, Vvc, Vvv, Vvv/Vmc, Sal). A Tukey HSD procedure revealed that for all but one of these, the significant differences are between the lab-raised (soft-diet) and the standard-length-equivalent-wild fish, not between lab-raised and smaller-wild fish. For Vvc/Vmc, lab-raised fish differ from the smaller-wild fish.

The results of analysis of filtered 3D parameters from all teeth (2 from each individual) is again similar. Significant differences were found between Sa, Sq, Sk, Spk, Svk, Vmp, Vmc, Vvc, Vvv, and Sal. However, a Tukey HSD procedure revealed more complex pairwise differences: for several parameters (Sa, Sq, Spk, Vmp, Vvc), lab raised fish differ from both wild groups; for Sk and Vmc, the significant differences are between the lab-raised and the size-equivalent wild fish; for Sal differences are between the size-equivalent-wild fish and the two other groups; while for Svk and Vvv all three groups differ from one another.

Analysis of the levelled but non-filtered 3D surface datasets, whether derived from most worn teeth, second teeth, means of the two, or all teeth, found no significant differences between the three groups of fish.

Results of stepwise LDA of the LPJ tooth 3D roughness data are reported in the main text and in Fig. 3. When all parameters are included, three (Svk, Smr1, Str) are enough to correctly discriminate between populations (Wilks' Lambda = 0.01, p = 0.0016). Canonical axis 1 explains 99.77% of variance, axis 2 the remaining 0.23%. The probability of correct assignment to populations is 99.9% or 100%, except for two specimens with probabilities of 98% and 91%, so the probability of any of the specimens being mis-assigned is low). Scores for canonical axis 1 are correlated with Svk (r = 0.90, p = 0.0009); scores for canonical axis 2 are correlated with Str (r = 0.89, p = 0.001). No other correlations between variables and conical scores are significant.

If stepwise LDA is limited to those parameters that differ (ANOVA) three are required to assign all samples to their correct trophic group (Sq, Svk, Vvv), but the significance of this LDA is doubtful (Wilks' Lambda, 0.09, p = 0.07). Canonical axis 1 accounts for 87.97% of variance, axis 2 the remaining 12.03%. The probability of correct assignment to groups is between 91.1% and 99.8% for most specimens, but two have lower probabilities: specimen 37865 has a 33% probability of misassignment to the smaller-wild group, and specimen 37869 has a 46% probability of being misassigned to the lab-raised soft diet group. Stepwise inclusion of two additional parameters (Spk and Vmc) produces a more informative LDA, also with correct assignment of all specimens (Wilks' Lambda = 0.0002, p = 0.0025; probability of correct assignment to populations = 100% for all specimens). Canononical axis 1 accounts for 99.96% of variance, axis 2 the remaining 0.04%. All 5 parameters are correlated with scores for canonical axis 1 (Sq, r = -0.8198, p = 0.007; Spk, r = 0.7092, p = 0.03; Svk, r = -0.8899, p = 0.001; Vmc, r = -0.7844, p = 0.01; Vvc, r = -0.7716P = 0.015). No parameters are correlated with scores for canonical axis 2.

9

Table S5: Summary of results of stepwise LDA of LPJ teeth. For each dataset the parameters required to assign all samples to their correct trophic group are listed. For mean filtered data, only 2 variables are required, but addition of a third variable (in brackets) increases the separation between groups and the probability that samples are correctly assigned. Analysis results Parameters required, 100% discrimination Most worn tooth, raw data, all parameters Sk, Sa, Sp, Sdr, Str 2 teeth, raw data, all parameters Spk, Smr1, Vmp, Sq, Sp, Sv, S10z, Sdq Mean of 2 teeth, raw data, all parameters Vvc/Vmc, Ssk, Sdr Most worn tooth, filtered data (log transformed), all parameters

Svk, Smr1, Str

Most worn tooth, filtered data (log transformed), parameters that differ (ANOVA)

Sq, Svk, Vvv

2 teeth, filtered data (log transformed), all parameters Sz, Sdq, Svk, Smr2 2 teeth, filtered data (log transformed), parameters that differ (ANOVA)

Sk, Spk, Svk, Vmp, Vmc, Vvv

Mean filtered data from 2 teeth (log transformed), all parameters

Svk, Vvc/Vmc, (Sal)

Mean filtered data from 2 teeth (log transformed), parameters that differ (ANOVA)

Svk, Vvc/Vmc, (S10z)

Table S6. Spearman’s rank correlations between results of multivariate analyses of LPJ 3D data and fish size and diet. PCA axis 1 and first canonical axis are derived from analysis of LPJ 3D data; fish size = standard length, diet = mg of hard shelled prey material in guts. N is low, so p values should be treated with caution. Fish size and diet are not correlated: rs = 0.2594, p = 0.5742). Standard length (n = 9) Hard shelled prey in gut (mg; n =7) Analysis result (PCA Axis 1) rs p rs p Most worn tooth, unfiltered data, all parameters

0.267 0.488 0.037 0.937

Most worn tooth, filtered data, all parameters 0.250 0.516 0.852 0.015 Most worn tooth, filtered data, parameters that differ (ANOVA)

0.233 0.546 0.927 0.003

Most worn tooth, filtered data (log transformed), all parameters

0.250 0.516 0.852 0.015

Most worn tooth, filtered data (log transformed), parameters that differ (ANOVA)

0.250 0.516 0.964 0.0005

Second tooth, filtered data (log transformed), all parameters

-0.083 0.831 0.704 0.077

Mean filtered data from 2 teeth (log transformed), all parameters

0.117 0.765 0.964 0.0005

Mean filtered data from 2 teeth (log transformed), parameters that differ (ANOVA)

0.100 0.798 0.927 0.003*

Analysis results (first canonical axis) Most worn tooth, unfiltered data, all parameters

-0.867 0.002 -0.371 0.413

Mean of 2 teeth, unfiltered data, all parameters -0.571 -0.717

0.180 0.03

-0.148 0.751

Most worn tooth, filtered data (log transformed), all parameters

-0.317 0.406 -0.927 0.003

Most worn tooth, filtered data (log transformed), parameters that differ (ANOVA)

0.317 0.406 0.927 0.003

Mean filtered data from 2 teeth (log transformed), stepwise, 2 parameters from all

-0.283 0.460 -0.889 0.007

Mean filtered data from 2 teeth (log transformed), stepwise, 3 parameters from those that differ

0.283 0.460 0.889 0.007

Mean filtered data from 2 teeth (log transformed), stepwise, 3 parameters from all

0.067 0.865 0.889 0.007

10

Supplementary figures

Figure S1. Images showing the 100 µm areas sampled for 2D scoring, with the unfiltered and the filtered 3D surface data collected from the two samples of oral teeth exhibiting the highest (a-c, specimen NgR104) and lowest density of 2D microwear features (d-f, specimen HpyVMP01) (respectively epilithic and epiphytic scraping specialists). As noted in the text, several roughness parameters are positively correlated with the density of features (Sa, Sq, Sp (log), Sv (log), Sz (log), S10z (log), Sdr, Sk, Spk, Svk, Vmp, Vmc, Vvc, Vvv, Sal. (a, d) scanning electron micrographs, field of view 100 x 100 µm. (b, c, e, f) Alicona rendered 3D data, 60° tilt, field of view 145 µm wide; (b, e) raw, levelled 3D data, vertical colour scale from 10 to -24 µm, (c, f) filtered 3D data (roughness surface), vertical colour scale from 5 to -5 µm.

Figure S2. 3D surface data from the qualitatively most worn (a, b, specimen 37872 a size-equivalent-wild fish) and least worn of the LPJ teeth (c, d, specimen 37864; a lab-raised soft-food fish) As noted in the text, several roughness parameters are positively correlated with qualitative assessment of roughness. Alicona rendered 3D data, 60° tilt, field of view 145 µm wide; (a, c) raw, levelled 3D data, vertical colour scale from 6 to -13 µm, (b, d) filtered 3D data (roughness surface), vertical colour scale from 5 to -5 µm.

11

REFERENCES 1. Zar J.H. 1999 Biostatistical analysis. fourth ed. New Jersey, Prentice Hall; pp. 663. 2. International Organization for Standardization. 2007 Geometrical product specifications (GPS) - Surface texture: Areal - Part 2: Terms, definitions and surface texture parameters (ISO 25178-2); pp. 42. 3. Kovach W. 2006 Oriana Version 2.02e. Anglesey, Wales, Kovach Computing Services.