Accurate prediction of human appearance from trace DNA is a central goal of forensic DNA phenotyping, yet trait prediction has remained limited to broad categorical outcomes. Here, for the first example of eye color, we present the shift from previous category-based inference to novel quantitative prediction. Specifically, we model iris pigmentation as four continuous color dimensions: BlueGray, DarkBrown, Green, and LightBrown. Each dimension measured on a scale from 0 to 1 with a resolution of 0.001. This framework captures eye color variation at a substantially higher resolution than previous multinomial classification approaches. We also developed a visually supported prediction output where several eye images meeting the prediction outcomes are generated alongside the quantitative output to aid interpretation.

For SNP selection, we analyzed 969 pigmentation-associated DNA variants that were both genotyped and imputed, including all HIrisPlex-S markers, in 3,537 individuals. Marker selection models were trained separately for each of the four color dimensions. After an initial screening step, we compared alternative variable-selection strategies using 10-fold cross-validation. Elastic net showed the best overall prediction performance, and the inclusion of interaction terms further improved prediction for several color dimensions, particularly LightBrown. The selection procedure retained 314 DNA variants including 128 interaction terms. For prediction model development, several approaches were examined to capture the phenotypic variance of all four color variables summing to 100%; with Partial Least Squares Regression (PLSR) proving the most robust. The final model consisted of 296 DNA variants including 26 interaction terms. Nested 10x10 cross-validation was performed for model building with mean R2 values of 0.664 (SD=0.046) for BlueGray, 0.656 (SD=0.028) for DarkBrown, 0.378 (SD=0.055) for LightBrown and 0.156 (SD=0.021) for Green. Classification performance, measured by AUC of the dominant category assignment, were highest for DarkBrown at 0.947(+/-0.117) and BlueGray at 0.935 (+/-0.0207), while lower for LightBrown at 0.835 (+/-0.0498) and Green at 0.772 (+/-0.0358). Independent samples (N=500) not used for model building are incorporated for external validation of overall model performance.

Together, we demonstrate for the first time, the feasibility of moving DNA-based eye color prediction beyond coarse categorical assignment toward a more detailed quantitative framework. This novel approach has strong potential to improve the resolution, interpretability, and evidential value of eye color prediction in forensic DNA phenotyping. To support future forensic application, we are developing an open-source dynamic web-based platform to make the model and its prediction output; both quantitative and visual, widely accessible for forensic DNA phenotyping.

Authors

• Susan Walsh (Indiana University Indianapolis, United States of America)
• Ziyi Xiong (Erasmus MC University Medical Center, Netherlands)
• Lauren Huntington (Indiana University Indianapolis, United States of America)
• Subhashree Ramanathan (Indiana University Indianapolis, United States of America)
• Jingxuan Zhang (Indiana University Indianapolis, United States of America)
• Shiaofen Fang (Indiana University Indianapolis, United States of America)
• Manfred Kayser (Erasmus MC University Medical Center, Netherlands)