A1269
Title: AI-assisted STR genotyping from high-resolution melt curves for rapid forensic screening
Authors: Parastoo Bybordi - Virginia Commonwealth University (United States)
Indranil Sahoo - Virginia Commonwealth University (United States) [presenting]
Rahul Ghosal - University of South Carolina (United States)
Kiersten Fultz - Virginia Commonwealth University (United States)
Tracey Dawson Green - Virginia Commonwealth University (United States)
Chenlu Ke - Virginia Commonwealth University (United States)
Abstract: Rapid and reliable DNA profiling is essential for modern forensic investigations, yet conventional short tandem repeat (STR) genotyping requires time-intensive amplification and electrophoresis. High-resolution melt (HRM) analysis provides a faster alternative by generating fluorescence-based melt curves whose shapes reflect underlying genotypes; however, manual interpretation is subjective and intensity-based computational approaches are sensitive to DNA concentration and lack forensic relevance. A shape-based, AI-driven framework is proposed for early STR genotype prediction from single-source HRM curves. After interpolation to a common temperature grid, curves were modeled as functional data and functional principal component analysis (FPCA) scores were used as inputs to a random forest classifier within a nested cross-validation framework with class weighting and SMOTE to address imbalance. Grouping genotypes into three groups with similar derivative profiles further improved separability. Across five outer validation folds, the model achieved a mean accuracy of 0.623 (SD 0.071) and balanced accuracy of 0.632 (SD 0.054), with high specificity (0.760). These results show that functional representations capture discriminative melt-curve shape information and provide a rapid, concentration-invariant screening tool for prioritizing samples and generating early exclusionary information in forensic workflows.
