Overview of the artificial neural network training, application, and validation.

a. The data used for training the network were generated by simulating 336 different head types (with an example simulation for one head type provided here), where a head type is defined by mean bone thickness, mean BM intensity and mean body fat thickness. There are 5000 datapoints per head each with 50 intensities per datapoint. OT: outer table, IT: inner table, BM: bone marrow.

b. We evaluated the performance of the neural network using two metrics: 1. the overlap between the predicted and the true BM location (defined by the simulation), and 2. the ratio between the predicted signal intensity of the BM and the true intensity of the BM.

c. Illustration of the process of extracting the intensity arrays from the MRI scan, locating the BM, and computing its intensity.

d. The algorithm was validated on a real dataset consisting of 30 individuals that were scanned 10 times each at 3-day intervals. For each scan, we computed the intensity of the outer bone (blue) and of the BM (yellow) which displayed high concordance across scans of the same individuals. The quality control measures correctly identified scans not producing reliable measures (13 of 300, 11 of which fail due to too few datapoints).

e. Application of the algorithm to a cohort of monozygotic and same-sex dizygotic twins. Correlation in BMA between twin pairs in coloured boxes.

The calvarial BMA phenotype

a. BM intensity distribution in males (blue) and females (yellow)

b. BMA age trajectories for males and females. Males in blue and females in yellow (never HRT) and orange (ever HRT). The effects of age and HRT on female BMA levels were both statistically significant (p-value < 2.0E-16)

c. Sex-dimorphic relationship between BMA and calvarial bone mineral density.

d. Relationship between BMA and BMI. Fitted lines in B, C, and D were computed using a general additive model and shading represents 95% confidence intervals.

BMA genome-wide significant loci

a. QQ-plot of p-values for the male, female and joint discovery GWAS

b. Comparison of male (outer), female (middle), and joint (inner) discovery GWAS significant loci.

c. Manhattan plot of the 41 genome-wide significant loci in the joint discovery GWAS named after gene closest to the top lead SNP (bold). Other genes in each locus for which genome-wide significant SNPs were eQTLs, are listed below the closest gene. Three loci were closely adjacent to another more significant locus and are marked “+”. SNPs with p < 0.05 not plotted.

Gene expression in mesenchymal lineage cells of mouse BM.

a. The differentiation of BM mesenchymal cells. Early (EMP), intermediate (IMP), and late mesenchymal progenitors (LMP). Lineage committed progenitors (LCP). Marrow adipogenic lineage precursor (MALP). Lipid-laden Adipocyte (LiLA). Osteoblast (OB). Osteocyte (Ocy).

b. A clustering of the gene expression profiles across cell types. Gene names are for mouse genes (these broadly match human gene names). Of the 40 genes which are closest to lead SNP for each locus, 28 had expression in the scRNAseq data (bold). LiLA expression data was not available in this dataset.

Genetic correlation and overlap of BMA with other traits

rg: genetic correlation. Q-values are FDR corrected P-values. Genetic overlap: number of causal variants that overlap (i.e. are shared) between BMA and the second trait. Total number of estimated causal variants for a trait (i.e. explaining 90% of SNP heritability) is the sum of those unique to trait and overlap (large total is indicative of high polygenicity). NA*: The publicly available summary statistics of the BMI GWAS contained only 3.3M SNPs, it was thus not possible to apply the MiXeR method to this dataset. Sex-specific genetic correlations, standard errors for causal variant partitioning, and Akaike Information criterion values (Table S9). Blue highlighting for: genetic correlation p < 0.01, overlap fraction > 0.5, correlation of effect sizes within overlapping > 0.5 or < -0.5.