Using simulated microhaplotype genotyping data to evaluate the value of machine learning algorithms for inferring DNA...

Area of Science:

• Forensic genetics
• Population genetics
• Computational biology

Background:

• Accurate inference of the number of contributors (NoC) is critical for DNA mixture interpretation.
• Challenges in NoC inference arise from allele sharing and dropout in complex mixtures.
• Microhaplotypes (MH) offer high polymorphism, potentially aiding NoC determination.

Purpose of the Study:

• To analyze the impact of allele sharing and dropout on NoC inference using microhaplotypes (MH).
• To evaluate the effectiveness of MH and compare three NoC inference methods: Maximum Allele Count (MAC), Maximum Likelihood Estimation (MLE), and Random Forest Classification (RFC).

Main Methods:

• Simulated over 40 million complex DNA mixture profiles (NoC 2-8) using 100 polymorphic MH from the Southern Han Chinese (CHS) population.
• Included unrelated (RM) and related individuals (parent-offspring - PO, full-sibling - FS, second-degree kinship - SE).
• Compared MAC, MLE, and RFC algorithms for NoC inference under varying dropout and kinship conditions.

Main Results:

• Number of detected alleles varied with marker polymorphism, kinship, NoC, and dropout.
• MAC and MLE performed best for RM, followed by SE, FS, and PO types.
• RFC excelled in PO types, followed by RM, SE, and FS types.
• Recall decreased with increased NoC and dropout for all methods.
• MLE was better at low NoC; RFC performed better at high NoC/dropout.
• RFC models trained with prior kinship information outperformed a general RFC model.

Conclusions:

• Microhaplotypes show promise for NoC inference in complex DNA mixtures.
• RFC models demonstrate superior performance in challenging scenarios (high NoC, high dropout), especially when trained with specific kinship information.
• Recommendations are provided for building machine learning models for NoC inference in forensic genetics.