True colors: A literature review on the spatial distribution of eye and hair pigmentation

Abstract

DNA-based prediction of externally visible characteristics has become an established approach in forensic genetics, with the aim of tracing individuals who are potentially unknown to the investigating authorities but without using this prediction as evidence in court. While a number of prediction models have been proposed, use of prior probabilities in those models has largely been absent. Here, we aim at compiling information on the spatial distribution of eye and hair coloration in order to use this as prior knowledge to improve prediction accuracy. To this end, we conducted a detailed literature review and created maps showing the eye and hair pigmentation prevalence both by countries with available information and by interpolation in order to obtain prior estimates for populations without available data. Furthermore, we assessed the association between these two traits in a very large data set. A strong limitation was the quite low amount of available data, especially outside Europe. We hope that our results will facilitate the improvement of already existing and of novel prediction methods for pigmentation traits and induce further studies on the spatial distribution of these traits.

Testing the impact of trait prevalence priors in Bayesian-based genetic prediction modeling of human appearance traits

“…Although due to data availability issues the number of predictors was lower in our freckles prediction modeling than previously [13] this discrepancy shall not affect our main outcomes for freckles significantly, since we applied the same reduced marker set to both the model with and without priors. Regarding the prior information, we surprisingly noticed that there is a limited spatial and population-specific trait prevalence information available for hair, skin and eye color, hair structure [24] and even non-existent for other traits such as freckles. We therefore exhaustively investigated the impact of the choice of prior values for the different trait categories on a fine-grained grid of all possible sets, or tupels, of values to obtain a general picture of the impact of priors on prediction performance.…”

“…All three prediction models are publicly available via https://hirisplex.erasmusmc.nl/. An alternative statistical tool for the prediction of eye, hair and skin color from genotype data is offered by Snipper [8,22,23], which uses pairwise likelihood ratios to present prediction outcomes, while other pigmentation prediction tool models were also developed (see [24] for a review). While some of these models show high prediction accuracies for some pigmentation categories, more research is currently under way in order to improve existing tools, either by including more SNP predictors after they have been identified in large-scale gene mapping studies, or by using alternative prediction methods.…”

“…To this end, we consider EVCs such as eye, hair and skin color for which prior-free genetic prediction models have previously been established [9,11,12], but also traits such as hair structure and freckles for which the first prediction models were recently proposed without considering priors [13,15,16]. Given the sparsity or even lack of spatial or population-specific prevalence information available for each of these EVCs [24], we investigated the impact of prevalence-informed priors across a grid in the complete space of all possible values for each trait category, thereby emulating the (mis-) specification of the informative prior values. Prediction modelling was performed by applying previously proposed DNA predictors in datasets from different populations inside and outside of Europe.…”

Evaluation of supervised machine-learning methods for predicting appearance traits from DNA

“…Some exceptions include fuzzy logic, artificial neural networks and classification trees used by Liu et al [13] for eye colour prediction modelling and Snipper [14], which is a Bayesian classifier that provides the prediction results as likelihood ratios. Further exceptions include the iterative naïve Bayesian approach from Maroñas and Söchtig [22,23] for skin and hair color respectively, and classification trees and partition modeling applied by Allwood et al [24] (see [25] for a further review).…”

Forensic determination of hair deposition time in crime scenes using electron paramagnetic resonance

“…This gap in knowledge is unexpected as hair fibers are one of the most common types of trace evidence found at crime scenes because healthy humans lose (on average) 50-150 telogen phase hairs a day [6], pulling hair is relatively easy especially during a struggle [7], and hair is generally very persistent, especially on rough fabrics and beneath fingernails [8]. Hair strands found at a crime scene can provide confirmatory proof to link to a person or animal, provide environmental exposures or drug history of the donor with their extended window of detection compared to blood and urine, and via DNA profiling, identify the source of the evidence (subject and/or species; [9][10][11][12][13][14][15][16][17]). Other advantages of using hair samples include the ability to store samples at room temperature, they are difficult to adulterate, and the non-invasive nature of collecting hair samples compared to blood or urine samples facilitates the process [17,18].…”