Forensic DNA workflows are built around the STR profile. It’s also where analysts first learn a sample has a problem — a mixture they didn’t anticipate, a template too low to produce usable results. By then, the options for adjusting the approach have narrowed considerably.
Kiersten Fultz, a master’s student in Forensic Science at Virginia Commonwealth University, has been working on an assay designed to surface that information earlier. Using high-resolution melt (HRM) curve data collected immediately after the qPCR quantification step, her lab’s screening tool runs predictions through a machine learning algorithm to assess whether a sample is single source or a mixture — before the STR profile is ever built.
Fultz presented her poster at ISHI 36 as a Student Ambassador. We spoke with her about the research, what her optimization work changed, and what the field might eventually do with a tool like this.
What problem is this research trying to solve, and how does the assay approach it?
The core issue is timing. Current workflows require a full STR profile before an analyst can determine sample composition — and that’s late in the process to discover a complication. As Fultz put it: “Evidentiary biological samples are not able to be assessed until the end of the DNA analysis workflow where a STR profile is developed. And this process can cause problems where if you have a low template DNA sample or even a mixture sample, this is not seen until the STR profile is produced.”
The Quantifiler Trio HRM assay her lab developed addresses this by using HRM curve data — already collected during quantification — as an input to a prediction modeling algorithm built in RStudio. The algorithm uses machine learning models trained on genotyped buccal swab samples to generate a prediction: single source, or mixture.
The composition of the training set matters. The more samples, and the more varied the sample types, the sharper the predictions. Fultz drew a deliberate analogy to how AI systems learn from new input — while being careful to note the assay doesn’t use what would typically be called AI software.
What did your optimization work change, and what were the results?
Fultz’s role wasn’t to redesign the assay but to optimize how the algorithm interprets its output, and to build a new function for estimating the number of contributors (NOC) in mixture samples.
Adjusting the labeling method made a measurable difference. Collapsing from eight categories down to a two-category system — single source versus mixture — improved accuracy, and after rebalancing the training set to more evenly represent mixture samples, the two-category prediction accuracy reached 83%.
The NOC function performed at approximately 64% accuracy across mixtures of two to six contributors. One model she tested accurately predicted six-person mixtures. “We were happy with those results as well,” she said.
The genotyping function remains unfinished. Current accuracy is around 25%, and Fultz is direct about it: “Future research is looking at different machine learning models within the code, such as neural networks or a ten fold validation that could hopefully increase the genotyping accuracy of this assay as well.”
Who is this tool built for, and what’s the practical benefit?
The intended user is the analyst running samples. “We want this to be a tool that they could use during their traditional workflow so that it saves them time and overall money so that they don’t have to reanalyze samples and waste time on interpretation and subjectivity,” Fultz said.
Earlier detection of a mixture gives analysts more room to respond — adjusting sample volume, changing the analytical approach, flagging the sample for a different interpretation strategy — before resources have been committed to a profile that may need to be redone.
What was it like presenting this research at ISHI as a Student Ambassador?
Fultz described the Ambassador experience as one that opened doors she wouldn’t have found on her own. “I love being an ambassador because I feel like I already knew people even before I got here,” she said.
Her advice to students considering the program was characteristically direct: “Even if you’re an introvert like me, it gets you out there and into opportunities that you wouldn’t have any other way.”
Fultz was finishing her degree at the time of the conference. She has since graduated from VCU’s accelerated bachelor’s-to-master’s program in Forensic Science and begun working as a DNA analyst — the exact role this tool is built for.