Nested Machine Learning Facilitates Increased Sequence Content for Large-Scale Automated High Resolution Melt Genotyping.

Stephanie I Fraley,Tza-Huei Wang,Samuel Yang,Justin Hardick,Yu-Hsiang Hsieh,Pornpat Athamanolap,Karen C Carroll,Charlotte A Gaydos,Richard E Rothman,Billie J Masek

doi:10.1038/srep19218

Abstract

High Resolution Melt (HRM) is a versatile and rapid post-PCR DNA analysis technique primarily used to differentiate sequence variants among only a few short amplicons. We recently developed a one-vs-one support vector machine algorithm (OVO SVM) that enables the use of HRM for identifying numerous short amplicon sequences automatically and reliably. Herein, we set out to maximize the discriminating power of HRM + SVM for a single genetic locus by testing longer amplicons harboring significantly more sequence information. Using universal primers that amplify the hypervariable bacterial 16 S rRNA gene as a model system, we found that long amplicons yield more complex HRM curve shapes. We developed a novel nested OVO SVM approach to take advantage of this feature and achieved 100% accuracy in the identification of 37 clinically relevant bacteria in Leave-One-Out-Cross-Validation. A subset of organisms were independently tested. Those from pure culture were identified with high accuracy, while those tested directly from clinical blood bottles displayed more technical variability and reduced accuracy. Our findings demonstrate that long sequences can be accurately and automatically profiled by HRM with a novel nested SVM approach and suggest that clinical sample testing is feasible with further optimization.

Highlights

Acid melting interactions become more complex for longer stretches of sequence[1,2,11,12], we hypothesized that long amplicons could produce additional distinguishing melt features beyond Tm alone
To increase the profiling breadth and sequence information generated from a single primer set and to test the ability of High Resolution Melt (HRM) to resolve variations in long amplicons, we here explore the use of a ~1000 bp amplicon covering six hypervariable regions in the 16 S rRNA genetic loci (16 S) genetic loci for HRM curve-based identification
We hypothesized that long amplicon melting of the 16 S gene could identify as many as or more organisms as short amplicon melting of any single region alone, despite the amplicon being approximately six times longer than that typically used for HRM

Summary

Introduction

Acid melting interactions become more complex for longer stretches of sequence[1,2,11,12], we hypothesized that long amplicons could produce additional distinguishing melt features beyond Tm alone. We and others have previously designed HRM assays for 16 S-based identification using multiple separate amplification reactions and primer sets to interrogate short sequence stretches within a genetic loci[4,5,6,7,9,22,23,24,25] This was successful for identifying a pure sample, but fails in non-pure samples where multiple genotypes present a mixture of distinct sequences contributing to a single melt curve[10]. To increase the profiling breadth and sequence information generated from a single primer set and to test the ability of HRM to resolve variations in long amplicons, we here explore the use of a ~1000 bp amplicon covering six hypervariable regions in the 16 S genetic loci for HRM curve-based identification. We explore a new nested OVO SVM approach to maximize the benefit of long amplicon curve shape complexity and advance HRM’s ability to accomplish large-scale profiling

Methods

Results

Conclusion