Continued Optimization of a High Resolution Melt-Based Mixture Screening Tool & Evaluation as a Contributor Identification Tool

The analysis of forensic biological samples is currently performed at the end of the DNA workflow and requires an STR profile for interpretation. Sometimes samples contain minor contributors with low-copy DNA that result in profile alleles that do not pass the analytical threshold, causing convolution during profile interpretation. This research aims to optimize a screening assay and statistical model that uses high-resolution melt (HRM) curve data to make predictions regarding whether a sample is from a single contributor or contains DNA from multiple individuals. Training and validation sets containing HRM data collected from genotyped buccal swab samples were processed through a prediction algorithm using three machine learning models. The accuracy, reproducibility, and ability to predict the number of contributors in mixtures and the genotypes of single source samples using the integrated Quantifiler™ Trio-HRM assay was assessed. 75.00% of validation set samples were accurately predicted using an 8-category labeling method (7 unique genotypes + mixtures). When genotype labels were changed to “single source” for a 2-category prediction method, the prediction accuracy increased to 81.94%. After increasing the number of mixtures in the training sets so that the ratio of sample types was more even, the 2-category dataset’s overall accuracy increased to 83.30%. Additionally, the assay was 63.95% accurate in predicting the NOC in mixtures ranging from two to six contributors. The inter-analyst reproducibility of the finalized 2-category assay was 66.60% and the intra-analyst reproducibility was 62.50%, which shows the assay is consistently over 60.00% reproducible.