Advancing biogeographical ancestry predictions through machine learning.

Area of Science:

• Forensic genetics
• Machine learning
• Biogeographical ancestry estimation

Background:

• State-of-the-art forensic tools like Snipper and Admixture Model are used for biogeographical ancestry.
• These methods lack systematic comparison with general machine learning classifiers.
• Genetic data is tabular, suitable for general machine learning approaches.

Purpose of the Study:

• Benchmark established forensic biogeographical ancestry classifiers against TabPFN, a general machine learning classifier.
• Evaluate performance using accuracy, ROC AUC, and log loss metrics.
• Assess classification at both continental and intracontinental levels.

Main Methods:

• Benchmarking forensic classifiers (Snipper, Admixture Model) against TabPFN.
• Utilizing a published dataset for training and testing.
• Evaluating performance on continental and intracontinental classification tasks.

Main Results:

• TabPFN consistently outperformed existing forensic methods across all evaluated metrics.
• Significant performance differences observed between methods.
• TabPFN improved accuracy from 84% to 93% on a continental scale (eight populations).
• TabPFN improved accuracy from 43% to 48% for inter-European classification (ten populations).

Conclusions:

• TabPFN demonstrates superior performance for biogeographical ancestry estimation compared to current forensic tools.
• General machine learning classifiers offer a promising alternative for forensic applications.
• Further research should explore TabPFN's application in diverse forensic datasets.