Performance of four models for eye color prediction in an Italian population sample

• Published in: Forensic science international : genetics Vol. 40; pp. 192 - 200
• Main Authors: Salvoro, Cecilia, Faccinetto, Christian, Zucchelli, Luca, Porto, Marika, Marino, Alberto, Occhi, Gianluca, de los Campos, Gustavo, Vazza, Giovanni
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.05.2019
• Subjects: Algorithms; Decision Trees; Eye Color - genetics; Eye color prediction; Female; Forensic DNA phenotyping; Genotype; Genotyping Techniques - instrumentation; Humans; Italy; Male; Models comparison; Models, Genetic; NGS SNP genotyping; Phenotype; Polymorphism, Single Nucleotide; Sequence Analysis, DNA - methods; Italy; NGS SNP genotyping; Eye color prediction; Models comparison; Forensic DNA phenotyping
• ISSN: 1872-4973; 1878-0326
• DOI: 10.1016/j.fsigen.2019.03.008

•Comparison of four eye color prediction models in an Italian sample showed strengths and weaknesses for each of them.•The IrisPlex, Ruiz and Allwood models provided 60–69% of correct predictions –using the recommended thresholds.•The IrisPlex system showed the lowest frequency of errors, followed by the Allwood and the Ruiz models.•The major issues were related to the high prevalence of intermediate eye colors in the Italian sample.•Model adjusting suggests the need for a further genetic and phenotypic characterization of intermediate colors. Forensic DNA phenotyping (FDP) has recently provided important advancements in forensic investigations, by predicting the physical appearance of a subject from a biological sample, using SNP markers. The majority of operable prediction models have been developed for iris color; however, replication studies to understand their applicability on a worldwide scale are still limited for many of them. In this work, 4 models for eye color prediction (IrisPlex, Ruiz, Allwood and Hart models) were systematically evaluated in a sample of 296 subjects of Italian origin. Genotypes were determined by a custom NGS-based panel targeting all the predictive SNPs included in the 4 tested models. Overall, 60–69% of the Italian sample could be correctly predicted with the IrisPlex, Ruiz and Allwood models, applying the recommended threshold. The IrisPlex model showed the lowest frequency of errors (17%), but also the highest number of inconclusive results (18%). In the absence of the threshold, the highest proportion of correct predictions was again obtained with the IrisPlex model (76%), followed by the Allwood (73%) and the Ruiz (65%) models. Lastly, the Hart predictive algorithm had the lowest error rate (2%), but the majority of predictions (87%) were restricted to the less informative categories of “not-blue” and “not-brown”, and correct color predictions were obtained only for 11% of the sample. As observed in previous studies, the majority of incorrect and undefined predictions were ascribable to the intermediate category, which represented 25% of the Italian sample. An adjustment of the IrisPlex (multinomial logistic regression) and Ruiz models (Snipper Bayesian classifier) with Italian allele frequencies gave only minor improvements in predicting intermediate eye color and no remarkable overall changes in performance. This suggests an incomplete knowledge underlying the intermediate colors. Considering the impact of this phenotype in the Italian sample as well as in other admixed populations, future improvements of eye color prediction methods should include a better genetic and phenotypic characterization of this category.