GenoEye: A machine learning-based framework for the prediction of intermediate eye color phenotypes

Forensic DNA phenotyping (FDP) enables the prediction of externally visible characteristics from genetic data, providing valuable investigative leads. While current approaches achieve high accuracy for blue and brown eye color, the prediction of intermediate phenotypes such as green and hazel remains challenging, particularly in Southern European and Mediterranean populations. In this study, we developed GenoEye, an interpretable machine learning framework for three-category eye color prediction based on an expanded SNP panel, from which a final 37-SNP predictive signature was derived. The model was developed and evaluated on a cohort of 363 Italian individuals using a gradient boosting approach, with independent data used for performance assessment. GenoEye achieved high predictive performance for blue and brown eye color (AUC up to 0.97), while improving the classification of intermediate phenotypes (AUC = 0.79), a category where existing tools show limited sensitivity. Comparative analyses under multiple decision strategies demonstrated that GenoEye provides consistently higher sensitivity for intermediate eye color while maintaining high specificity. To support practical application, GenoEye is implemented as a web-based tool with integrated SHAP-based interpretability. These results indicate that GenoEye improves the resolution of intermediate eye color prediction and provides a robust and interpretable framework for forensic applications.