Species identification using high resolution melting (HRM) analysis with random forest classification

• Published in: Australian journal of forensic sciences Vol. 51; no. 1; pp. 57 - 72
• Main Authors: Bowman, Sorelle, McNevin, Dennis, Venables, Samantha J., Roffey, Paul, Richardson, Alice, Gahan, Michelle E.
• Format: Journal Article
• Language: English
• Published: Clovelly Taylor & Francis 02.01.2019; Copyright Agency Limited (Distributor)
• Subjects: Accuracy; Analysis; Animals; Biological materials; Classification; Computer programs; COMPUTER SOFTWARE; Criminal investigation; DNA; Fluorescence; FORENSIC SCIENCE; Forensic sciences; High resolution; high resolution melt analysis; Human subjects; Methodology; Polymerase chain reaction; predictive modelling; random forest; Random variables; rapid screening; rRNA 16S; Software; Species; Species classification; STATISTICS; AUSTRALIA
• ISSN: 0045-0618; 1834-562X
• DOI: 10.1080/00450618.2017.1315835

Species identification is an important facet of forensic investigation. In this study, human and non-human species (cow, chicken, pig, sheep, cat, dog, rabbit, fox, kangaroo and wombat) were assayed on the ViiA 7 Real-Time PCR System (Thermo Fisher Scientific) to rapidly screen for their species of origin using the high resolution melt (HRM) analysis targeting the 16S rRNA gene. Classification of HRM difference profiles using the onboard ViiA 7 software resulted in a classification accuracy of <20%. Derivative profiles (temperature versus negative first derivative of fluorescence, -dF/dT) were classified using random forest algorithms supplemented by bagging and boosting, with either a randomly partitioned test set or a variety of folds of cross-classification, in addition to a range of trees and variables. Random forest classification with bagging conditions (constructed over 500 trees) was found to considerably outperform the ViiA 7 software for species differentiation with 100% classification accuracy for biological material from humans, domestic pets (cat and dog) and consumable meats (chicken and sheep) with an average classification accuracy of 70% across all species.