The human skeletal form underlies our ability to walk on two legs, but unlike standing height, the genetic basis of limb lengths and skeletal proportions is less well understood. Here we applied a deep learning model to 31,221 whole body dual-energy X-ray absorptiometry (DXA) images from the UK Biobank (UKB) to extract 23 different image-derived phenotypes (IDPs) that include all long bone lengths as well as hip and shoulder width, which we analyzed while controlling for height. All skeletal proportions are highly heritable (~40-50%), and genome-wide association studies (GWAS) of these traits identified 179 independent loci, of which 102 loci were not associated with height. These loci are enriched in genes regulating skeletal development as well as associated with rare human skeletal diseases and abnormal mouse skeletal phenotypes. Genetic correlation and genomic structural equation modeling indicated that limb proportions exhibited strong genetic sharing but were genetically independent of width and torso proportions. Phenotypic and polygenic risk score analyses identified specific associations between osteoarthritis (OA) of the hip and knee, the leading causes of adult disability in the United States, and skeletal proportions of the corresponding regions. We also found genomic evidence of evolutionary change in arm-to-leg and hip-width proportions in humans consistent with striking anatomical changes in these skeletal proportions in the hominin fossil record. In contrast to cardiovascular, auto-immune, metabolic, and other categories of traits, loci associated with these skeletal proportions are significantly enriched in human accelerated regions (HARs), and regulatory elements of genes differentially expressed through development between humans and the great apes. Taken together, our work validates the use of deep learning models on DXA images to identify novel and specific genetic variants affecting the human skeletal form and ties a major evolutionary facet of human anatomical change to pathogenesis.

The human skeletal form underlies bipedalism, but the genetic basis of skeletal proportions (SPs) is not well characterized. We applied deep-learning models to 31,221 x-rays from the UK Biobank to extract a comprehensive set of SPs, which were associated with 145 independent loci genome-wide. Structural equation modeling suggested that limb proportions exhibited strong genetic sharing but were independent of width and torso proportions. Polygenic score analysis identified specific associations between osteoarthritis and hip and knee SPs. In contrast to other traits, SP loci were enriched in human accelerated regions and in regulatory elements of genes that are differentially expressed between humans and great apes. Combined, our work identifies specific genetic variants that affect the skeletal form and ties a major evolutionary facet of human anatomical change to pathogenesis.

Electronic health records (EHRs) are often incomplete and inaccurate, reducing the power of genome-wide association studies (GWAS). Moreover, the variables within these records are often represented in binary codes, masking variation in disease severity among individuals. For some diseases, such as knee osteoarthritis (OA), radiographic assessment is the primary means of diagnosis and can be performed directly from medical images. In this work, we trained a deep learning model (DL-binary) to ascertain knee OA cases from anteroposterior (AP) dual-energy absorptiometry (DXA) scans and achieved clinician level performance. Applying this model across 29,257 individuals from the UK Biobank (UKB), we identified 2,603 (240%) more cases than currently diagnosed in the ICD-10 record. Individuals diagnosed as cases by DL-binary had higher rates of self-reported knee pain, knee pain for longer durations and with increased severity compared to control individuals. We trained another deep learning model to measure the minimum knee joint space width (mJSW), a quantitative phenotype linked to knee OA severity. Despite the DL-binary phenotype and mJSW being highly genetically correlated (92%), the heritability of mJSW was an order of magnitude greater than the ICD-10 code M17 or DL-binary phenotypes. In a GWAS run on mJSW, we identified 18 genome-wide significant loci, as opposed to 1 and 6 at the same sample size using either case-control (DL-binary and ICD-10 code M17) phenotype. This improved power also translated to better polygenic risk score (PRS) prediction for knee OA diagnosis in a holdout dataset of 371,686 individuals. We also show that reduced mJSW, but neither case-control phenotype is associated with increased risk of adult fractures, a leading cause of injury-related death in older individuals. For diseases with radiographic diagnosis, our results demonstrate the enormous potential for using deep learning to phenotype at biobank scale, both for improving power for gene discovery and for epidemiological analysis.