Most current methods using genetic markers to predict externally visible characteristics (EVCs) rely on Multinomial Logistic Regression models, where the independent variable is the number of minor alleles in the k-th SNP, as an integer in the [0,2] domain. Such an approach has major drawbacks, since the prediction probabilities based on the minor allele count assume a linear effect. As a consequence, the distance between the heterozygote and both alternative homozygous genotypes is considered identical, a condition rarely reflected in actual biological effects. Such assumptions ignore events such as dominance, penetrance, and other epistatic effects that are eventually present. As a result, prediction performance can be impacted, especially when evaluating intermediary phenotypical categories.

In this work, we propose an alternative method of evaluating genotypic data to predict EVC traits using Item Response Theory methods. In this approach, the EVC in question is considered a latent, unobservable trait, which is continuously distributed in a population. Genotypes are then taken to be observable manifestations of these hypothesized traits, which may not be directly observed but must instead be inferred from the manifest responses (genotypes).

To verify the feasibility of this approach, we have evaluated 67 distinct SNPs located in 9 different genes involved in melanin expression. A generic trait called "pigmentation level" was proposed, and the genotypes of around 600 highly admixed Brazilians were coded as a polytomous IRT model. Individuals were classified according to their hair, eyes, and skin colors. Individual item curves were generated for each individual locus, adopting 1 (Rasch model), 2, or 3-parameter models, and overall performance was assessed. Cronbach-alpha measures were also accessorily employed to evaluate each marker performance. Parameter estimation (including difficulty, discrimination, and pseudo-guessing for each item, as well as the latent trait score theta for each subject) from the genetic data was performed through Bayesian optimization, using R scripts designed for this task. The resulting discrimination curves for each item show different effects for genotypical combinations, and the overall latent parameter theta obtained for each subject reflects individual pigmentation levels as observed in hair, eyes, and skin colors. Therefore, IRT can be a powerful statistical tool to evaluate EVC models, assisting in both classification tasks and SNP selection in the panel development phase.

Authors

• Eduardo Avila (Federal University of Rio Grande do Sul, Brazil)
• Cassio Ritzel (Federal University of Rio Grande do Sul, Brazil)
• Alessandro Kahmann (Federal University of Rio Grande do Sul, Brazil)
• Marcio Dorn (Federal University of Rio Grande do Sul, Brazil)
• Clarice Alho (Federal University of Health Sciences of Porto Alegre, Brazil)
• Stela Castro (Federal University of Rio Grande do Sul, Brazil)