Somewhere in a massive database in the UK, there are X-ray images of nearly 60,000 people's spines. And until recently, those images were just sitting there, full of secrets about why some people's backbones decide to go rogue and curve sideways. A team of researchers just taught an AI to measure all of them - and discovered genetic clues that decades of traditional research had missed.
The Scoliosis Problem Nobody Talks About
Scoliosis - that sideways curvature of the spine you probably got screened for in middle school - is the most common spinal deformity in humans. About 3% of children develop it, and we still don't fully understand why. The condition can range from barely noticeable to requiring surgery, and while we know genetics plays a role (if your parent has it, you're significantly more likely to develop it), the specific genes involved have remained frustratingly elusive.
The problem? Studying scoliosis genetics traditionally requires diagnosing patients one by one, which introduces all sorts of inconsistencies. One doctor's "mild scoliosis" might be another's "watch and wait." Medical records are messy, incomplete, and vary wildly between healthcare systems.
Enter the Machines
Researchers from the University of Texas at Austin and their collaborators decided to take a different approach. Instead of relying on clinical diagnoses, they grabbed DXA scans (the same type of scan used to check bone density) from the UK Biobank - a treasure trove of health data from half a million Brits - and trained a deep learning model to measure spine curvature automatically.
The AI learned to identify individual vertebrae in the images, mark their positions, and calculate how much the spine deviates from where it should be. On a test set of 150 people who also had traditional Cobb angle measurements (the gold standard for scoliosis assessment since 1948), the automated measurements correlated at 0.83 with the clinical readings. Not perfect, but good enough to capture meaningful variation across tens of thousands of people.
Here's where it gets clever: instead of just classifying people as "has scoliosis" or "doesn't have scoliosis," the researchers used their measurements as a continuous variable. Because human spines exist on a spectrum - there's no magical threshold where a normal spine suddenly becomes a scoliotic one.
Two New Suspects in the Genetic Lineup
Armed with precise curvature measurements for 57,588 people, the team ran a genome-wide association study (GWAS) - essentially scanning millions of genetic variants to see which ones correlate with curvier spines.
They found two genetic regions that previous studies had missed entirely. One is near a gene called SEM1 (also known as SHFM1), which is part of the cellular machinery that breaks down proteins and has been linked to limb development. The other is a long non-coding RNA on chromosome 3, nestled between genes called EDEM1 and GRM7 - a stretch of DNA that doesn't code for proteins but might still influence how nearby genes behave.
The really wild part? Their quantitative approach outperformed a case-control study with ten times the sample size. Measuring things precisely, it turns out, beats having more data points measured poorly.
What Your Curved Spine Might Be Telling You
Beyond genetics, the researchers found some intriguing associations. People with more spine curvature tended to have differences in leg length, reduced muscle strength, lower bone density, and higher rates of knee osteoarthritis (but not hip osteoarthritis, oddly). Whether these are causes, effects, or just fellow travelers on the same genetic pathway remains to be figured out.
The findings also held up across populations. When the researchers compared their results with data from the Biobank Japan, they found significant genetic overlap despite the different ancestries - suggesting these genetic factors are broadly relevant across human populations.
The Bigger Picture
This study is part of a growing trend: using AI to extract quantitative measurements from medical images at scale, then connecting those measurements to genetics. It's a powerful approach because it sidesteps all the messiness of clinical diagnoses - the varying thresholds, the missed cases, the documentation inconsistencies.
The limitation? The scans come from adults aged 40-69, so the researchers can't distinguish between people whose curves started in adolescence and those who developed them later due to age-related degeneration. Different origin stories might mean different genetic underpinnings.
Still, finding two new genetic loci for a condition this common is no small feat. Each discovery is a thread that researchers can pull on, potentially leading to better understanding of why spines curve in the first place - and maybe, eventually, how to prevent or treat it more effectively.
References
- Zeosky, M., et al. (2026). Deep learning-based precision phenotyping of spine curvature identifies novel genetic risk loci for scoliosis in the UK Biobank. npj Digital Medicine. DOI: 10.1038/s41746-026-02540-6
- MedlinePlus. Adolescent idiopathic scoliosis. https://medlineplus.gov/genetics/condition/adolescent-idiopathic-scoliosis/
- NCBI Gene Database. SEM1 26S proteasome subunit. https://www.ncbi.nlm.nih.gov/gene/7979
- Bycroft, C., et al. (2018). The UK Biobank resource with deep phenotyping and genomic data. Nature, 562(7726), 203-209. DOI: 10.1038/s41586-018-0579-z
- Physiopedia. Cobb Angle. https://www.physio-pedia.com/Cobb_Angle
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.