The application of machine learning to predict genetic relatedness using human mtDNA hypervariable region I sequences

Area of Science:

• Forensic Genetics
• Bioinformatics
• Machine Learning

Background:

• Victim identification after mass casualty events is crucial but challenging with degraded DNA samples.
• Mitochondrial DNA (mtDNA) is vital for sample origin determination and ethnic group assignment.
• Machine learning (ML) offers a novel approach to forensic genetic analysis, with limited prior application.

Purpose of the Study:

• To investigate the efficacy of ML algorithms in predicting genetic relatedness using mtDNA hypervariable region I sequences.
• To compare the performance of four ML classification algorithms (SVM, LDA, QDA, RF) hybridized with different feature extraction techniques (one-hot encoding, PCA, BoW).
• To evaluate ML model accuracy against traditional methods like Analysis of Molecular Variance (AMOVA).

Main Methods:

• Utilized hypervariable region I sequences from the GenBank database for African, Asian, and Caucasian populations.
• Implemented and compared four ML algorithms: Support Vector Machines (SVM), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Random Forest (RF).
• Combined ML algorithms with feature extraction methods including Principal Component Analysis (PCA) and Bags of Words (BoW), analyzed using WEKA and Python.

Main Results:

• PCA-SVM achieved 80-90% accuracy in WEKA, outperforming PCA-LDA, PCA-RF, and PCA-QDA.
• BoW-PCA-RF in Python reached 94.4% accuracy, surpassing other BoW-PCA combinations.
• ML models demonstrated higher accuracy than AMOVA for genetic inferences.

Conclusions:

• ML algorithms, particularly SVM and RF, are effective supplementary tools for forensic genetics casework.
• The study validates ML's potential for accurate genetic relatedness prediction and sequence classification.
• These findings support the integration of ML into forensic science for enhanced identification capabilities.