Machine learning tool improves accuracy of genome mapping studies

University of Virginia School of Medicine scientists have identified a widespread source of error in a popular method for studying the genome – our genetic instructions – and created a machine-learning tool to correct it. 

The free tool could improve the reliability of both conventional and single-cell data generated using this method, giving researchers a clearer view of how gene activity is controlled in health and disease and providing stronger foundations for future diagnostic and drug-development efforts.

Although nearly every cell in the body contains the same DNA sequence, different cells use different sets of genes to maintain their identities and functions. Much of that control comes from the "epigenome" – chemical modifications and structural features of chromosomes that influence whether genes are turned on or off without changing the DNA sequence itself. The CUT&Tag (Cleavage Under Targets & Tagmentation) method can map these epigenomic features efficiently from very small samples and even from individual cells. However, the UVA researchers, led by Chongzhi Zang, PhD, found a "hidden bias" in the method that creates artifacts resembling genuine biological signals.

The DNA sequence is like the sheet music. The epigenome determines which notes to play, when they are played and by what instruments. When a technical artifact looks like a real signal, researchers can be led toward the wrong biological mechanism. That risk is especially serious in single-cell data, where the true signals are already sparse."

Chongzhi Zang, UVA's Department of Genome Sciences and director of computational genomics, UVA Comprehensive Cancer Center

Finding and fixing hidden bias

CUT&Tag has many advantages over prior methods and has quickly become popular among scientists. But it also has weaknesses that have gone unnoticed until recently. The assay relies on an enzyme called Tn5 transposase, which naturally favors open, accessible regions of the genome. Zang and colleagues found that this preference can create misleading and confusing research results that the scientists described as "severe" after examining nearly 300 published datasets.

Having discovered the unexpected scope of the problem, Zang and his team developed PATTY (Propensity Analyzer for Tn5 Transposase Yielded bias), a tool that uses machine learning to correct the CUT&Tag bias. It is effective both in conventional data from many cells and in sparser single-cell data, the scientists report.

"Detecting true signals in noisy data is like finding a needle in a haystack, and it is even more difficult when many pieces of hay look like real needles," Zang said. "PATTY does not detect signals simply by subtracting a background. It learns how a real needle differs from hay and uses the learned model to reduce the artifact while preserving real signals."

Zang and his colleagues have made PATTY available as a free, open-source software package at Github (https://github.com/zang-lab/PATTY) and Zenodo (https://doi.org/10.5281/zenodo.20078737). They hope their tool will improve genomic research and eventually help get new diagnostics and medicines to patients faster, a major mission of UVA's new Paul and Diane Manning Institute of Biotechnology.

"We believe that bias correction should become a routine part of CUT&Tag analysis," Zang said. "Cleaner data can keep scientists from wasting time and resources pursuing technical artifacts and can make real biological differences easier to see. More broadly, PATTY provides a conceptual framework for correcting similar biases in other genomic technologies."

Findings published

The researchers have described the development of PATTY in an article in the scientific journal Nature Communications. The article is open access, meaning it is free to read. The UVA research team consisted of Shengen Shawn Hu, Qingying Chen, Megan C. Grieco, Mengxue Tian and Zang. Collaborators Zhangli Su and Anindya Dutta of the University of Alabama at Birmingham and Lin Liu of Shanghai Jiao Tong University also contributed. The scientists have no financial interest in the work. 

The research was supported by the National Institutes of Health, grants R35GM133712, R21HG012981, R00CA259526 and R01CA060499.