BR-FDP-SKIN: Brazilian forensic DNA Skin phenotyping based on machine learning models

Accurate prediction of skin pigmentation from DNA remains challenging, particularly in admixed populations due to their complex genetic architecture. In this study, we evaluated a panel of 66 SNPs associated with pigmentation in an admixed South Brazilian population from Rio Grande do Sul (n=438), phenotypically classified into six Fitzpatrick skin types. Three grouping strategies (six, three, and two classes) were adopted, and feature selection was performed using VariantSpark to identify the 20 most informative SNPs. Four machine learning algorithms (SVM, KNN, MLP, and XGBoost) were tested and optimized via recursive feature addition and hyperparameter tuning. Model performance was assessed using the weighted F1-score, accuracy, precision, and recall across 30 replicates. In the six-class grouping strategy, SVM achieved the highest F1-score (0.4846±0.0067), although PCA showed substantial overlap between classes. XGBoost under the three-class grouping scheme achieved an F1-score of 0.7839±0.0047, while binary classification with SVM reached the highest performance (F1-score =0.9530±0.0031). A subset of four SNPs-rs1426654 (SLC24A5), rs11230664 (DDB1), rs16891982 (SLC45A2), and rs1448484 (OCA2)-were consistently ranked as the most informative across all grouping strategies. Furthermore, our optimized SNP panel diverged significantly from HIrisPlex-S, incorporating a set of variants absent in the established system. This study demonstrates that tailored SNP selection and phenotypic grouping strategies are essential for accurate skin color prediction in admixed populations. The resulting models were implemented into the open-access tool BR-FDP-SKIN, providing a population-specific resource for forensic and anthropological applications.