Evaluation of supervised machine-learning methods for predicting appearance traits from DNA.

Area of Science:

• Forensic genetics
• Anthropological genetics
• Computational biology

Background:

• DNA-based prediction of externally visible characteristics (EVCs) is an established forensic and anthropological genetics technique.
• Multinomial logistic regression (MLR) is commonly used, but the performance of other classification methods for EVC prediction remains under-investigated.

Purpose of the Study:

• To systematically compare the performance of MLR against popular machine learning (ML) classifiers: support vector machines (SVM), random forest (RF), and artificial neural networks (ANN).
• To determine if ML classifiers offer advantages over existing models for predicting categorical pigmentation traits using established DNA markers.

Main Methods:

• Phenotypes (eye, hair, skin color) were predicted using genotypes from IrisPlex, HIrisPlex, and HIrisPlex-S DNA markers.
• Performance comparison of MLR, SVM, RF, and ANN classifiers, including hyperparameter tuning to optimize performance.

Main Results:

• All four classification methods demonstrated similar predictive performance for the tested EVCs.
• No single method was substantially superior across all traits; performance varied slightly between traits and more significantly across trait categories.
• Machine learning methods did not provide a significant advantage for predicting categorical pigmentation traits with the selected DNA markers.

Conclusions:

• Current machine learning classifiers do not offer a performance advantage over traditional methods for predicting categorical pigmentation traits from DNA.
• Further research may be needed to explore different ML algorithms or marker sets for improved EVC prediction.