Deep Learning Algorithm Identifies Hidden Genetic Mutations in Noncoding DNA Regions

The research team combined an experimental sequencing technique known as ATAC-seq with a deep learning-based computational tool called PRINT.

Researchers from the Children's Hospital of Philadelphia (CHOP) and the Perelman School of Medicine at the University of Pennsylvania (Penn Medicine) have successfully developed and applied an algorithm capable of identifying genetic mutations in the vast noncoding regions of the human genome that could increase the risk of common diseases. 

The findings, published in the American Journal of Human Genetics, represent a promising step toward unlocking the full potential of genetic data in diagnosing and treating complex diseases.

Shedding Light on the Genome's Dark Matter

Although only about 2% of the human genome codes for proteins that perform essential biological functions, the remaining 98% — often referred to as the "noncoding" genome — plays a crucial role in regulating gene expression. 

Variants in these noncoding regions have been increasingly linked to disease risk, but studying them has remained a challenge due to a limited understanding of the underlying regulatory code.

Past genome-wide association studies (GWAS) have helped identify broad genomic regions associated with diseases, but pinpointing which specific variant within those regions is driving the disease has proven difficult. 

Many of these disease-linked variants are found near transcription factor binding motifs — key genomic regions where proteins known as transcription factors bind to regulate gene activity. 

When these transcription factors bind to "open" regions of the DNA, they leave behind telltale signs or "footprints" that researchers can detect through sequencing techniques.

Decoding Genetic Footprints with Deep Learning

To tackle this challenge, the research team combined an experimental sequencing technique known as ATAC-seq with a deep learning-based computational tool called PRINT. 

ATAC-seq identifies regions of the genome that are "open" and accessible for protein binding, while PRINT analyzes these regions to detect specific DNA-protein interaction footprints.

"This situation is comparable to a police lineup," explained senior study author Dr. Struan F.A. Grant, Director of the Center for Spatial and Functional Genomics at CHOP. 

"You're looking at similar suspects together, so it can be challenging to know who the actual culprit is. With the approach we used in this study, we're able to pinpoint the disease-causing variant through identification of this so-called footprint."

The researchers applied these tools to genetic data from 170 human liver samples and identified 809 "footprint quantitative trait loci" (fpQTLs).

These fpQTLs represent regions in the genome where DNA-protein interactions vary depending on the specific genetic variant present, offering a precise way to locate which mutations may be influencing disease-related gene regulation.

Toward Disease Prediction & Treatment

Lead author Max Dudek, a PhD student in Grant and Almasy's labs at Penn Medicine and CHOP, highlighted the significance of this approach for future clinical applications.

"This approach helps resolve some fundamental issues we have encountered in the past when trying to determine which noncoding variants may be driving disease," Dudek said. 

"With larger sample sizes, we believe that pinpointing these causal variants could ultimately inform the design of novel treatments for common diseases."

Looking ahead, the researchers plan to expand this methodology to other tissues and organs, with the goal of identifying variant-specific disease mechanisms across different biological systems.

Stay tuned for more such updates on Digital Health News.