If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Corresponding author at: Institute of Forensic Sciences “Luis Concheiro”, University of Santiago de Compostela, Rúa San Francisco, s/n, 15782 Santiago de Compostela, La Coruña, Spain. Tel.: +34 881812311; fax: +34 981580336.
Forensic Genetics Unit, Institute of Legal Medicine, University of Santiago de Compostela, SpainCIBERER, Genomic Medicine Group, University of Santiago de Compostela, Spain
In forensic analysis predictive tests for external visible characteristics (or EVCs), including inference of iris color, represent a potentially useful tool to guide criminal investigations. Two recent studies, both focused on forensic testing, have analyzed single nucleotide polymorphism (SNP) genotypes underlying common eye color variation (Mengel-From et al., Forensic Sci. Int. Genet. 4:323 and Walsh et al., Forensic Sci. Int. Genet. 5:170). Each study arrived at different recommendations for eye color predictive tests aiming to type the most closely associated SNPs, although both confirmed rs12913832 in HERC2 as the key predictor, widely recognized as the most strongly associated marker with blue and brown iris colors. Differences between these two studies in identification of other eye color predictors may partly arise from varying approaches to assigning phenotypes, notably those not unequivocally blue or dark brown and therefore occupying an intermediate iris color continuum. We have developed two single base extension assays typing 37 SNPs in pigmentation-associated genes to study SNP-genotype based prediction of eye, skin, and hair color variation. These assays were used to test the performance of different sets of eye color predictors in 416 subjects from six populations of north and south Europe. The presence of a complex and continuous range of intermediate phenotypes distinct from blue and brown eye colors was confirmed by establishing eye color populations compared to genetic clusters defined using Structure software. Our study explored the effect of an expanded SNP combination beyond six markers has on the ability to predict eye color in a forensic test without extending the SNP assay excessively – thus maintaining a balance between the test's predictive value and an ability to reliably type challenging DNA with a multiplex of manageable size. Our evaluation used AUC analysis (area under the receiver operating characteristic curves) and naïve Bayesian likelihood-based classification approaches. To provide flexibility in SNP-based eye color predictive tests in forensic applications we modified an online Bayesian classifier, originally developed for genetic ancestry analysis, to provide a straightforward system to assign eye color likelihoods from a SNP profile combining additional informative markers from the predictors analyzed by our study plus those of Walsh and Mengel-From. Two advantages of the online classifier is the ability to submit incomplete SNP profiles, a common occurrence when typing challenging DNA, and the ability to handle physically linked SNPs showing independent effect, by allowing the user to input frequencies from SNP pairs or larger combinations. This system was used to include the submission of frequency data for the SNP pair rs12913832 and rs1129038: indicated by our study to be the two SNPs most closely associated to eye color.
]. Recently, considerable interest has been expressed in the viability of using genetic predictors of physical characteristics when reliable eyewitness is not available to police investigations or no national DNA database entry exists. The initial goal has been the design of predictive tests for common pigmentation variation, the first of which have typed coding region single nucleotide polymorphisms (SNPs) providing inference of iris color [
]. In dark eye colors higher levels of melanin absorb more light, while an absence of this pigment in the stroma tends to disperse light and absorbs the majority of the color except blue and blue-gray. Complex intermediate tones, including green and hazel result from diverse quantities of melanin that reflect light of mixed tonalities [
]. Methods for iris phenotype classification based on quantitative analysis through digital photography have been recently proposed, estimating values of saturation and tonalities [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
], as the key human eye color regulator. A model for brown and blue determination was proposed based on regulation of OCA2 expression by promotor SNPs, principally rs12913832, embedded in HERC2 [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
]. However following identification of rs12913832, the most comprehensive study of eye color associations in the HERC2–OCA2 region and in other pigmentation related genes was the landmark study of Liu et al. [
]. The main advantage of Liu's study was the systematic analysis of associations of very closely sited HERC2 and OCA2 SNPs, and whether these were due to close linkage with rs12913832 or independent effects. Liu also proposed a viable eye color prediction system using probabilistic analysis based on multinomial logistic regression that enabled the identification and listing of the strongest SNP predictors in order of effect. Notably, a very close SNP to rs12913832:rs7183877 (115 nucleotides apart) emerged as the 10th strongest predictor while the more distant rs1129038 (8759 nucleotides distant) was not identified as a predictor after adjustment for the effect of close linkage with rs12913832. Liu also identified SNPs in genes beyond the HERC2–OCA2 complex with predictive effect in TYRP1, TYR, SLC24A4, SLC45A2, IRF4, and ASIP [
] have adapted the above findings of Liu to provide forensic eye color tests. Each study arrived at slightly different recommendations for the development of SNaPshot single base extension assays typing just the most closely associated SNPs. Walsh used the six strongest SNP predictors from Liu, providing the most reliable inference of blue and brown iris phenotypes, but these predictors have less success predicting intermediate eye colors. Recently the detection of epistatic effects has provided a slight increase in the prediction of intermediate phenotypes [
], however further studies are needed to fully explain the genetic basic of these complex ‘non-blue, non-brown’ eye colors. Mengel-From identified four SNPs, all clustered around the HERC2–OCA2 complex, that were detected to be most closely associated when using a simplified light/dark iris phenotyping regime. The difference between the two studies in identification of best predictors may partly be due to their varying approaches in the assignment of intermediate iris colors, occupying a continuum that merges into blue and brown. However, the most significant difference between the studies was the use by Mengel-From of multiple SNP loci in the HERC2–OCA2 complex plus SLC45A2, whereas Walsh chose to implement a single SNP from each of six genes (the above three plus SLC24A4, TYR, and IRF4).
In our study, we examined the effect of bringing together the additional HERC2–OCA2 SNPs identified by Mengel-From with the six Irisplex SNPs of Walsh. If an expanded SNP set provides improved predictions, adjusted for linkage, the total number of SNPs typed in a forensic test can still remain small, preserving the sensitivity and robustness of a SNaPshot-based assay. Prior to this current study we had developed two SNaPshot assays typing 37 SNPs in pigmentation-associated genes to analyze associations in skin, hair, and eye pigmentation variation in Europeans. We have used these assays to test the performance of different forensic eye color predictor combinations in 416 subjects taken from six populations of north and south Europe. Intermediate phenotypes, we defined as ‘not simple blue or simple dark brown’ iris colors, were compared to the genetic clusters obtained from Structure software. We adapted an online Bayesian classifier termed Snipper (http://mathgene.usc.es/snipper/), originally developed for ancestry analysis of genotype-based data [
] – to handle allele frequencies. This enhancement allows SNP-pair frequencies to be included in the data input by simple counting of each combination, although phase must be assumed. As Snipper handles user-defined custom SNP data, both as genotypes or frequencies, further predictor combinations could be accommodated, if identified in future. However submission of multiple-SNP haplotypes becomes complex and requires large sample counts to be made, therefore this study concentrated principally on the assessment of predictive performance when adding rs12913832–rs1129038 genotype pair data to the other SNPs of Irisplex, while identifying additional HERC2 SNPs as potential contributors to intermediate eye color variation.
Once we had identified an additional five SNPs on top of the six of Walsh and two of Mengel-From we made the 13 SNP genotype data available for use with Snipper. Beyond the rs12913832–rs1129038 HERC2 SNP pair, rs1667394 of HERC2 emerged as a potentially important interacting marker while rs7183877 appears to have a marked effect on prediction of green-hazel eye colors. Both these additional HERC2 SNPs were independently identified as significant eye color predictors by Eriksson et al. in 2010 [
Study samples were collected from 416 volunteers from the following European populations: NW Spain (Galicia) 215; NW Germany (Lower Saxony) 91; Sweden (Dalarna) 44; Austria (Innsbruck) 31; Denmark (Copenhagen) 18 and Switzerland (Zurich) 17. Informed consent was obtained in all cases, as well as information about donor's immediate ancestry. Ethical approval was granted from the ethics committee of clinical investigation in Galicia, Spain (CEIC: 2009/246). DNA was obtained from buccal swabs that were stored at room temperature. DNA extraction was made using standard phenol–chloroform isoamyl alcohol protocols. Eye colors were recorded using a 12 megapixel reflex digital camera with uniform lighting conditions (consisting of a lens-attached ring flash), photographic settings and a macro lens that allowed the pupil, iris, and sclera to fill the majority of the image field. Phenotype assignment of eye color was defined following criteria described by previous studies [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
] where “intermediates”, i.e., non blue, non brown types were classified into green-hazel, intermediate-dark, and intermediate-light with the first category tending to have most examples of a brown peri-pupillary ring within a blue outer iris. However patterns for classification of intermediate phenotypes were also described according to the presence of melanin spots, accumulation of collagen, contract furrows and Fuch's crypts as detailed by Sturm and Larsson [
]. In order to compare our studies with those of Mengel-From, we added supplementary labeling of our study individuals into “light” by grouping blue and intermediate light together plus “dark” by grouping brown with intermediate dark.
2.2 SNP selection, multiplex design and genotyping methods
A set of 37 SNPs considered to be associated with human pigmentation variation according to previously published studies was selected for genotyping [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
] in two SNaPshot single base extension assays detailed below. Twenty-three of these SNPs found to be most associated with eye pigmentation in a series of published studies (literature references for all 37 SNPs are listed in Table 1, Table 2) were used to test the performance of forensic iris color prediction in the six European study populations. Table 1, Table 2 also list the ranking given by Liu for the eye color predictive power of the 15 most associated SNPs. The other 8 SNPs were additional HERC2 markers analyzed by the two studies that first identified rs12913832, plus components of a three SNP haplotype in OCA2 – which had been thoroughly analyzed by Kayser et al. [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
Two novel single base extension multiplex assays were developed: SHEP 1 and SHEP 2 (i.e., skin, hair, and eye pigmentation) using Primer3 and AutoDimer [
] to design, check and optimize amplification primers creating amplicon lengths ranging from 87 to 135 base pairs (bp). Parallel extension primer sets were built in the same way and all primer sequences developed are given in Table 1, Table 2. Typical SHEP 1 and SHEP 2 primer extension profiles are shown in supplementary Fig. S1. Our aim when developing SHEP 1–2 was to examine all previously associated pigmentation SNPs and adapt these SNaPshot-based assays to smaller multiplexes later, when the most closely associated SNPs were more clearly identified. The alternative approach of designing Sequenom iPlex assays lacked the necessary flexibility to modify SNP combinations with changing knowledge of the strongest associations or interactions.
The PCR reaction was optimized using 1–10 ng of DNA in 10 μl final reaction volume, containing 1× Buffer, 1× BSA, 8 mM MgCl2, 700 μM dNTPs, 0.1–0.01 μM of each primer and 0.1 U AmpliTaq Gold polymerase (Applied Biosystems, Foster City, US: AB). Amplification conditions comprised: denaturing at 95 °C for 10 min, then 35 cycles using 95 °C for 30 s, 60 °C for 50 s, 65 °C for 40 s, then a final extension of 65 °C for 6 min. Multiplexed SNaPshot (AB) single base extension chemistry was used to type the amplified SNP combinations in two parallel extension reactions. Prior to extension with SNaPshot 2.5 μl of PCR product was treated with 1 μl of ExoSAP-IT (USB® Corporation) to remove unused dNTPs or PCR primers, run at 37 °C for 15 min followed by 85 °C for 15 min to inactivate the enzyme. Then 1.5 μl of purified PCR product was added to 2.5 μl of SNaPshot ready reaction mix (AB) plus 1.5 μl of extension primer mix (final concentration 0.2 μM). Extension conditions comprised: 30 cycles of 96 °C for 10 s, 50 °C for 5 s, 60 °C for 30 s. The extension reaction products were cleaned up with 1 μl of SAP (USB) at 37 °C for 80 min and 85 °C for 15 min. Capillary electrophoresis was performed on a Prism 3130xl Genetic Analyzer (AB) using Genemapper® Analysis Software v. 3.7 (AB).
2.3 Statistical analyses and classification models
One problem that occurs when comparing the association findings of Mengel-From with those of Walsh relates to use of light-dark and blue-intermediate-brown phenotyping regimes respectively. Because of this we decided to compare patterns of genetic clustering obtained with Structure [
] with our own assignment of five eye color phenotypes based on photography (as described in Section 2.1), to assess how our subjects grouped, i.e., whether two, three or five groups are discernible from the genetic data of 23 SNPs. Analysis using Structure v. 2.3.3 comprised: 100,000 Markov Chain steps after a burn-in of length 100,000 with four replicates for each value of K (assumed populations) from 2 to 5. The admixture and linkage model was applied using LOCPRIOR information and frequencies correlated among populations. CLUMPP and Distruct software were used to visualize and plot results [
]. We used standard approaches to calculate the optimum K value from the data based on the mean estimated probability of data stabilizing around a maximum value.
Each SNP was analyzed for Hardy Weinberg equilibrium (HWE) using 1,000,000 Markov Chain steps and assessed for pairwise linkage disequilibrium using “1000-permutation Chi-square tests” performed with Arlequin (v. 35). Significance analysis levels were adjusted for multiple tests following standard Bonferroni corrections [
Individual SNP informativeness as part of a predictive system was estimated using the Snipper classifier (http://mathgene.usc.es/snipper/), originally developed for genetic ancestry inference [
] that ranges from 0 = no divergence to 1 = maximum divergence. To evaluate the robustness of the eye color reference training sets for use with Snipper we performed two kinds of cross-validation: the classical one-out reclassification and a variant of the bootstrap analysis by randomly choosing (with replacement) a training set of 200 individuals from the reference set and classifying the remaining samples with this training set, repeating the procedure 100 times [
]. For this part of the assessment of training sets, a likelihood ratio threshold of 0.5 was used to denote a successful assignment. All calculations were made with custom programs written in R (v. 2.13.1).
As a Bayesian classifier based on likelihood ratios, Snipper sorts individual likelihoods in descending order then provides a prediction based on the ratio of the two largest likelihoods. We adapted the generation of likelihoods in Snipper to work with frequency-based training sets rather than genotypes (http://mathgene.usc.es/snipper/frequencies.html), allowing the frequencies of SNP combinations consisting of closely sited loci to be included in the training sets. The rs12913832–rs1129038 SNP pairs were counted without consideration of phase, e.g., AG (rs12913832) and AG (rs1129038) was treated as the profile component AA,GG, though strand1,strand2 can comprise AG,GA. To help readers assess this system we include three frequency-based training sets as supplementary Files S1–S3 for blue:green-hazel:brown eye colors: S1 that lists 23 individual SNP frequencies; S2 combining rs12913832–rs1129038 pairs with 22; and S3 combining rs12913832–rs1129038 plus the five other Irisplex SNPs. Smaller subsets of the 23-SNP data can be made by removing SNP worksheets accordingly, but SNP combinations beyond the rs12913832–rs1129038 have not been included as these are complex and difficult to count in sufficient numbers.
In order to allow a comparison with the Irisplex classification system we also calculated positive predictive values (PPV), negative predictive value (NPV), sensitivity and specificity values for the Irisplex 6, Irisplex 5 plus the rs12913832–rs1129038 pair; and Irisplex 5 + 2 plus HERC2 SNPs rs7183877 and rs1667394. Liu provides a systematic definition of the above values and the predictive model evaluation system used in the original study and applied to Irisplex in the supplementary data of [
]. Liu defines PPV as the percentage of correctly predicted color type among the predicted positives and NPV as the percentage of correctly predicted non-color type among the predicted negatives. Sensitivity is a measure of classification success, defined as: the percentage of correctly predicted color type among the observed color type. Specificity is the percentage of correctly predicted non-color type among the observed non-color type.
IBM PASW SPSS Statistical-18 tests were used to analyze associations to phenotypes. Individual SNP associations were analyzed by logistic regression under an additive model. Adjustment for the most associated marker rs12913832 was made to detect the additional effect of other loci physically linked to this most strongly associated HERC2 SNP. We also made a SNP pair-adjusted analysis of the effect of other HERC2 SNPs adjusting for rs12913832–rs1129038 as a single variant and as two separate variants.
Following the classification approach used by Walsh et al. [
], as a complementary method to assess the informativeness of three SNP sets: the six of Irisplex; these six plus two additional SNPs of Mengel-From, and; these 8 plus an additional five SNPs we identified as contributing detectable extra eye color predictability. All AUC analyses were made on the training set (256 samples) comprising: blue, brown, and green-hazel phenotypes.
Combined-effect and interactions between SNPs were analyzed using the multifactorial dimensionality reduction system (MDR) for pairwise eye color phenotype comparisons [
A flexible computational framework for detecting, characterizing, and interpreting statistical patterns of epistasis in genetic studies of human disease susceptibility.
]. MDR is a non-parametric data mining approach that evaluates different combinations of genetic or environmental factors (SNPs in our case). A set of n SNPs is selected (where n usually comprises 1, 2, 3, and 4 SNPs) and for each n SNPs and their possible multi-factorial classes (e.g., nine genotype combinations for 2 binary loci) the ratio of cases to controls is calculated. Each cell of a SNP combination is assigned to either a low- or high-risk group depending on the ratio of cases and controls. If this ratio meets or exceeds a threshold (usually 1.0) that genotype combination is determined as high risk, otherwise it is assigned as low risk. All potential combinations of n factors are evaluated sequentially and the model that gives the lowest error in classifying cases and controls based on low- or high-risk is selected for each set of n SNPs in a training set from 9/10 of the data, evaluating a test set from the remaining 1/10 of the data, to obtain the prediction accuracy. This cross-validation procedure, consisting of a random split of data into 9/10 and 1/10 proportions is repeated 10 times using a random seed number to protect against chance divisions of the dataset.
We applied the MDR software (www.epistasis.org) to determine the best 1, 2, 3, and 4 SNPs model for our dataset, producing four different models. Of the four best models, that with the highest average testing accuracy and cross-validation consistency (number of times a model is selected as best model among the validation sets) was selected as the final, best model. We used the MDR permutation module to test the significance of the association of this final model with case status. Finally, to visualize the relationships between the SNPs from the selected model we used the interaction dendrogram. This is constructed using a hierarchical cluster analysis and implemented in the MDR software. SNPs showing a strong interaction are located together and closely in the branches and the nature of the interaction is denoted by a color scale with blue indicating redundancy, to red, when SNPs show a high degree of synergy. MDR analyses were made separately for each color comparison (blue, inter-light, GH, inter-dark and brown vs. rest).
Lastly, principal component analysis (PCA) was made to compare the geographical distribution of phenotypes among our six sampled European regions, as populations in Europe tend to show a gradient of light to dark eye color running NW to SE. We used custom programs to execute PCA written in R (v. 2.13.1).
3. Results
3.1 Eye color phenotypes
Observed phenotypes across the European population samples consisted of 145 blue and 64 brown eye color plus 207 non-blue, non-brown phenotypes. The latter were not placed into a single group because of the full range of tonalities observed, so this group was classified as follows: green-hazel 95; intermediate-light 40; and intermediate-dark 72. We observed a high frequency of light eye colors in northern European subjects and this decreased moving south as shown in supplementary Fig. S2, in agreement with previous observations [
Supplementary Table S1 outlines the results of formal tests for Hardy Weinberg equilibrium with most population–locus combinations not indicating significant departure from Hardy–Weinberg equilibrium and only six significant values but distributed randomly across eye color ‘populations.’
From 2530 SNP pair comparisons significant LD was found in 40 SNP pairs listed in supplementary Table S2. All cases of significant LD were randomly distributed across loci and eye color populations except SNPs rs1375164–rs4778232 in OCA2 (15:28291812–15:28281765 = 10,047 nucleotide separation).
The PCA analysis did not reveal discernable geographic stratification – i.e., the phenotypic classes that we assigned to samples repeatedly cluster independently of their European geographic distribution as indicated in Fig. 1C and D .
Fig. 1(A) Analysis of eye color phenotype population structure in European samples with K = 2 clusters using Structure software. (B) Estimation of optimum K value indicates values greater than 2 can be inferred as the optimum cluster number. (C and D) PCA comparison of phenotype assignments made across population samples (C) and the geographical distribution of the populations studied (D).
Handling the phenotypes observed as genetic populations and using blue and brown classes as pre-defined populations, Structure analysis suggested a pattern of continuity observed in intermediate phenotypes distinct from blue and brown, represented in the K:2 cluster plot of Fig. 1A. We have positioned iris photographs above each sample range to illustrate the characteristics of the phenotypes assigned in the study. The intermediate-light samples tended to cluster close to the blue end and the intermediate-dark samples mostly clustered with browns; green-hazels appear as a largely equal mixture of the two most differentiated classes of blue and brown. This suggests that the simplified light and dark eye color phenotyping regime of Mengel-From continues to have validity for the wider range of SNPs examined in our study. Nevertheless, the green-hazel phenotype does not form a distinct cluster at K:2, K:3, or K:4 (K:3 and 4 cluster plots in supplementary Fig. S3) and appears as a group where all components have mixed membership proportions. Furthermore, the optimum K estimation indicated that more than two genetic populations were detectable (probability plot in Fig. 1B). Therefore we adopted the green-hazel class as a third reference phenotype in the training sets. In order to ensure the clearest differentiation of samples among a continuum of very similar iris colors, we excluded 35 green-hazel samples whose iris images fell between the most clearly differentiated blue and green-hazel phenotypes. Similarly, a further 13 brown eye samples were excluded that fell between green-hazel and brown phenotypes. So the training sets (256 samples) comprised: 145 blue subjects; 60 green-hazel from an original 95; 51 brown from an original 64.
3.2 Association analysis
The p-values for association with light, dark and the five eye color phenotypes originally defined in our sample set, for all 37 SNPs of SHEP 1–2, are given in supplementary Table S3 and this data includes adjustment for the effect of rs12913832. Of the 14 additional skin and hair color-associated SNPs genotyped in the SHEP assays, 6 of 98 association p-values were significant (p < 0.05) after adjustment for rs12913832, but we did not pursue the analysis of these SNPs further. The levels of association found in the 23 SNPs previously identified as most closely associated with eye color are summarized in Fig. 2. Association with a significance value of p < 0.05 was recorded in twenty markers. The independent additional effect on the prediction of phenotypes was measured by adjustment with the most associated SNP rs12913832. From adjustment analysis SNPs were detected to give an additional effect differentiating light, dark and the five eye color phenotypes independently of their linkage to rs12913832, all located in the OCA2/HERC2 region – in ranked order of effect: rs1129038, rs1667394, rs916977, rs4778138, rs7495174, rs4778241, rs4778232, rs8024968, rs11636232, rs7183877. These were followed by associated SNPs in genes: SLC45A2 (rs26722, rs16891982), SLC24A4 (rs12896399), ASIP (rs1015362), TYR (rs1393350), and TYRP1 (rs1408799). Comparing eye color phenotyping approaches, the classes light and dark have the largest number of SNPs associated after adjustment for rs12913832. Additional adjusted association values for the other eleven SNPs most strongly associated in the above data were obtained, by treating the rs12913832–rs1129038 SNP pair as a single variant and as two variants, and p-values are given in supplementary Table S4. Results were consistent between both approaches indicating strongest adjusted associations were for light and dark phenotypes with rs1667394 the strongest adjusted association to dark eye color.
Fig. 2A schematic representation of the individual association analysis of 23 SNP components for each eye color class (orange cells) and their association after adjustment with rs12913832 (dark-orange cells), SNPs are listed in order of descending divergence. Rows represent each color vs. all other eye colors.
3.3 Comparative assessment of classification with different SNP predictors
The divergence between the three phenotypes of the training set for 23 markers, tested with Jensen and Shannon's divergence metric, provided a ranking of SNP differentiation showing rs12913832 to be by far the most divergent, in agreement with a large number of previous studies [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
]. SNP rs12913832 is followed by rs1129038, rs1667394, rs916977, rs4778138, rs749517, rs4778241, and rs4778232, all located in the region of the HERC2–OCA2 complex. This divergence data provides the order of the SNPs in the three plots of Fig. 3, indicating the relative predictive power of each of the 23 SNPs from left (strongest) to right. The cumulative divergence values were highest for the differentiation of blue and brown compared to green-hazel with blue, or green-hazel with brown, but it is also notable that divergence values differed considerably between component SNPs. HERC2 loci rs12913832 and rs1129038 gave divergences of ∼0.7 for blue-brown comparisons. Although rs1667394 gave consistently high values, over half the other SNPs in any one comparison had much lower divergence values of 0.01–0.002 and rs6058017 of ASIP was the least informative marker. The allele frequencies observed in the eye color populations are listed in supplementary Table S3.
Fig. 3Accumulated divergence values from classifications using the Bayesian system for pairwise eye color comparisons (A–C) and three-way eye color comparisons (D).
Evaluating ROC curves as an alternative phenotype assignment method to Snipper allowed a direct comparison with the system used by Walsh et al. based on the six Irisplex SNPs, all included in our set. AUC estimations of the ROC curves were made for each pairwise phenotype comparison and these are shown in Fig. 4 for varying SNP subsets. In agreement with the findings of Walsh et al., the predictions for blue and brown eye colors were mostly explained by rs12913832 producing AUC values of 0.952 and 0.965 respectively, from this single marker alone. For green-hazel predictions additional SNPs are required to improve the precision but these do not exceed AUC values above 76% with six or indeed eight SNPs, as shown in Fig. 4A and B. We reach similar AUC values to those of Walsh with values of 0.986 for blue, 0.756 for green-hazel, and 0.978 for brown (Fig. 4A). However aiming to achieve a balance between small numbers of markers and predictive value we found that the addition of other markers, shown to be independent in effect from rs12913832, raised the AUC values for all phenotypes enough to justify their inclusion. This was particularly noticeable moving from the two additional SNPs of Mengel-From to a total of 13 that included SNPs improving green-hazel prediction. Adding the two additional SNPs of Mengel-From gave AUC values of 0.996 for blue and 0.983 for brown, plus slight progression to 0.768 for green-hazel (Fig. 4B). When the additional five SNPs were incorporated to the analysis, predictability from the AUC value estimates increases significantly to: 0.999 for blue, 0.990 for brown and 0.816 for green hazel (Fig. 4C). In summary, there is evidence the incorporation of rs1129038 of HERC2 improves eye color predictability in general, while adding rs1667394 and rs7183877 (HERC2) improves hazel-green predictability.
Fig. 4AUC curves for: (A) 6 SNP predictors of Walsh et al.; (B) 6 + 2 SNP predictors of Mengel-From et al.; (C) 6 + 2 plus the five additional SNP predictors identified in this study.
MDR analysis indicated that when assessing all possible combinations of two, three, and four SNPs the best model observed was composed of two SNPs for each eye color comparison (i.e., brown vs. non-brown, etc.). In brown vs. non-brown comparisons, the best model included rs12913832 in HERC2 and rs4778138 in OCA2 (BA 0.8587, CV 10/10). The entropy analysis indicated an interaction effect between these SNPs that was redundant as shown by the blue lines of the right-hand dendrogram of Fig. 5. For green-hazel vs. non green-hazel comparisons, the SNPs selected in the best model were HERC2 rs12913832 and rs1667394 (BA 0.8322, CV 8/10) with a significance level of p < 0.001. In this case a synergist interaction was observed between these SNPs indicated by the red lines on the left-hand dendrogram of Fig. 5B. Results were confirmed when these selected SNPs and their interactions were included in a logistic regression under additive model, with a significance p-value of 2.03 × 10−15 for green-hazel and 0.00130 for brown (data not shown).
Fig. 5Analysis of SNP interactions with MDR: (A) interaction of SNP-pair combinations and (B) corresponding entropic dendrograms for green-hazel vs. non green-hazel and brown vs. non-brown comparisons. In the interaction charts left bars indicate ‘cases’ (i.e., eye color) and right bars indicate the ‘controls’ (i.e., all other eye colors), light cells represent low risk and dark cells high risk. In the left-hand dendrogram the red lines represent synergistic interactions and on the right, blue lines represent redundant interactions.
Curiously the interactive effect between rs12913832 and rs1667394 is evidently different between the two eye color comparisons shown in Fig. 5. Such complex interactive effects require detailed further analysis and we are currently pursuing these studies.
3.5 Predictive performance
The informativeness of the reference samples used as a 23-SNP training set for the Snipper classifier was tested by cross-validation and gave classification success rates of: blue 97.93%; green-hazel 98.93%; and brown 92.16%. The modified bootstrap analysis revealed comparable success rates of: blue 98.27%, green-hazel 97.81%, and brown 96.67%. These measurements of classification success/error support the approach used to create three distinct reference classes for the classification system, despite excluding a proportion of complex intermediate phenotypes from the reference set.
The results based on the AUC curves, outlined in Section 3.3 above, indicated rs1129038 improves the predictive performance of the six Irisplex SNPs. There was also a marked jump in AUC value for green-hazel with the inclusion of rs7183877. Additionally, the MDR results in Section 3.4 suggest a synergistic interaction between rs12913832 and rs1664394. Therefore we examined the effect on sensitivity, specificity, NPV and PPV, for three SNP sets: Irisplex 6; Irisplex 5 with the rs12913832–rs1129038 pair and; Irisplex 5 + 2 with HERC2 SNPs rs7183877 and rs1664394.
Table 3 details these measures of predictive performance obtained by making Snipper classifications with the three SNP sets. The table indicates the effect of adding rs1129038 or three additional HERC2 SNPs is to improve the sensitivity of green-hazel predictions, that jump from 3.3% (Table 3A) to 75.3% (Table 3B) with the addition of rs1129038. Specificity is also improved for light, blue, intermediate-light, intermediate-dark, brown, and dark (Table 3B). Overall, each specificity is raised in value when adding rs1129038 except green-hazel, however this phenotype gains the most improved balance between sensitivity and specificity. Table 3C indicates the addition of further HERC2 SNPs does not necessarily improve the overall classification performance of Snipper and does not gain better balance between sensitivity and specificity. Since predictive models do not take account of interactive effects this may explain the lack of improvement in the sensitivity/specificity values when including two additional HERC2 SNPs (the four in total of Table 3C).
Table 3AUC % predictive value estimates from Snipper classifications for eye colors from pairwise comparisons of the full range of eye colors with three sets of SNP predictors: (A) Irisplex 6; (B) adding rs12913832-rs1129038 as a SNP pair PPV: positive predictive value, NPV: negative predicted value, ND: not determined. Bold values indicate the perceived best balance between sensitivity and specificity.
In the present study of six European populations, we confirmed the effect of most of the SNP predictors previously published for blue/brown or light/dark iris colors [
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
]. Specifically rs12913832, rs1129038, and rs116363232 in HERC2 and rs1289399 in SLC24A4 showed strong associations when making appropriate adjustment for the effect of rs12913832 or rs12913832–rs1129038. The markers rs1800407 in OCA2, rs16891982 in SLC45A2, and rs1393350 in TYR are also good predictors for forensic tests. The SNP rs12203592 in IRF4, reported to be among the six best predictors by Walsh et al., was not found to be a good predictor in our study. Interestingly, the frequency of the minor allele of rs12203592 had a high frequency in the Northern Ireland sample analyzed recently in an extended study of Irisplex in European populations [
] but has a much lower frequency in the samples of our study. This may explain the weaker predictability of rs12203592 in our analyses and raises the issue of population differences within Europe that will require extended examination to properly gauge the value of SNPs outside of the OCA2–HERC2 complex. We also confirmed the previously reported AUC values for the six Irisplex SNPs of Walsh proposed for forensic blue/brown eye color prediction. When considering the extension of these six predictors our study findings lead us to recommend the inclusion of HERC2 SNP rs1129038. This SNP improves the predictability for all eye colors as indicated by the AUC analyses and we have shown it has an additional effect to rs12913832 after adjustment analysis. In addition to rs1129038 the predictability of complex intermediate phenotypes increased when incorporating markers, in order of decreasing divergence: rs916977, rs7183877 (HERC2), rs4778138, rs7495174, rs4778241, rs8024968, rs1375164 (OCA2), rs1408799 (TYRP1), and rs26722 (SLC45A2). In particular, rs7183877 makes a detectable contribution to the prediction of green-hazel eye color. We have also begun epistasis studies with initial MDR results reported here and these already point to a synergistic interactive effect within HERC2 for rs1667394 with rs12913832 for green-hazel prediction. Therefore the three SNPs: rs1129038, rs1667394, and rs7183877, despite their close proximity in HERC2, provide additional information for eye color prediction after adjustment with rs12913832. This is consistent with the idea of complex regulatory function interactions within the OCA2–HERC2 complex [
]. We note that including S European population samples in our study provided 50% non-blue, non-brown subjects, but these comprised just 10% of Liu's study. Incorporating rs1129038, rs7183877, and rs1667394 with the six SNPs of Irisplex had the most marked effect on green-hazel predictability, but we also achieved a slight improvement in the prediction of blue and brown eye colors (AUC levels of 99% for blue and 98% for brown). However, further studies analyzing the independence of the additional OCA2–HERC2 markers closely located in the same chromosomal region are recommended, in particular utilizing the extra power provided by family studies.
The proper design and construction of a forensic eye color predictive test requires the consideration of certain factors examined in this study, in addition to exploring the effect of expanding the number of SNP predictors. Firstly, it is important to adequately define the phenotypic classes. We collected samples from six populations from north and south Europe with a range of phenotypic variability likely to be wider than would be found within single geographic locations. Reducing the complexity of the intermediate eye color range in the reference samples provided an informative training set and helped to identify key additional predictors. Nevertheless we found the AUC values for intermediate eye colors reached a plateau at ∼82% with 13 SNPs. This indicates additional factors have yet to be identified for intermediate eye color expression, including: undetected SNPs and associations; epistatic interactions; additional pigmentation genes; and environmental effects (e.g., changing iris color with age, as suggested by a recent study which also explored several complex SNP–SNP interactions [
]). It is worth noting that OCA2 is among the larger human genes (344 kb) and is likely to harbor many low frequency coding SNPs within the total 24 exons. Therefore, at the moment, a forensic eye color test maintaining a reasonably manageable multiplex level has probably reached the limit of predictive power for intermediate eye colors. The problem of a uniform, objective recognition of intermediate iris colors distinct from brown/blue among a group of observers (such as eye witnesses) also remains a source of variation in predictive performance of forensic tests, over and above complex SNP associations. Secondly, of equal importance to the definition of phenotypes is the classification method used. For routine inference of eye color from a forensic test the classifier must have a degree of flexibility that allows the end-user to re-configure SNP profiles according to their operational needs, as well as the scope to include extra markers from newly discovered or closely linked SNP associations. Luckily, both the classifiers of Walsh et al. and the online Bayesian system proposed here allow the user to make their own decisions about a probability threshold: a value limit that balances the ‘benefit’ of sufficiently successful predictions against the ‘cost’ of too many erroneous predictions or too many that are undefined because of an unduly high probability cut-off. These limits correspond to suggested values of 0.7 using the Walsh classifier and a likelihood ratio of 3:1 we applied using Snipper. However it may be the case that users wish to explore the effect of extra SNPs using their own test sets and Snipper allows likelihoods to be collected for such test sets analyzing a range of marker combinations. Though we used AUC analysis in order to compare the performance of extra OCA2–HERC2 SNPs with the established six markers of Walsh, we actually found Snipper easier to use to assess each new marker one at a time, but more importantly to upload linked SNP data by applying the SNP rs12913832–rs1129038 pair frequencies observed in each training set phenotype. Another advantage of Snipper is the ability to deal with incomplete profiles that have missing SNP data, common when analyzing highly degraded DNA. Therefore profiles where weakly predictive SNPs are missing are valid and likely to obtain high probabilities for blue or brown eye colors.
Acknowledgments
YR was supported by the Fundation Gran Mariscal de Ayacucho (FUNDAYACUCHO). MVL was supported by funding from Xunta de Galicia INCITE 09 208163PR and this work was in part supported by additional funding from Xunta de Galicia: PGIDTIT06P-XIB228195PR. JS was supported by the German Academic Exchange Service (DAAD). ÁC and CP acknowledge the support of the Areces Foundation and the EuroForGen NoE. The authors would like to thank all the anonymous donors who participated in the study.
Appendix A. Supplementary data
The following are the supplementary data to this article:
Supplementary Fig. S1Electropherograms of SHEP 1 and SHEP 2. Legend: Typical SHEP 1 and 2 electropherograms.
Supplementary Fig. S2Distribution of eye color phenotype frequencies in study. Legend: Distribution of eye color phenotype frequencies observed across the European populations studied. Pie charts are proportional to sample sizes.
Supplementary Fig. S3Structure analysis of eye color phenotypes at K = 3 and K = 4. Legend: Cluster plots based on the same analysis as Fig. 1 for K:3 and K:4 clusters.
Blue eye color in humans may be caused by a perfectly associated founder mutation in a regulatory element located within the HERC2 gene inhibiting OCA2 expression.
A flexible computational framework for detecting, characterizing, and interpreting statistical patterns of epistasis in genetic studies of human disease susceptibility.
00130-5&id=gr1.jpg){kind=link}
00130-5&id=gr2.jpg){kind=link}
00130-5&id=gr3.jpg){kind=link}
00130-5&id=gr4.jpg){kind=link}
00130-5&id=gr5.jpg){kind=link}
00130-5&id=mmc1.jpg){kind=link}
00130-5&id=mmc2.jpg){kind=link}
00130-5&id=mmc3.jpg){kind=link}
00130-5&id=mmc4.jpg){kind=link}