Genetic control of the HDL proteome

Nathalie Pamir; Calvin Pan; Deanna L. Plubell; Patrick M. Hutchins; Chongren Tang; Jake Wimberger; Angela Irwin; Thomas Q. de Aguiar Vallim; Jay W. Heinecke; Aldons J. Lusis

Normalization of shotgun proteomic data is a continuous struggle in the field (Välikangas et al., 2018) . HDL is a rather uncomplicated mixture containing only ~100 proteins. However, its proteome is driven by the 10 most abundant proteins, with 65% being made up of APOA1 and 15% of APOA2 (Toth et al., 2013) . The normalization strategy should conserve the compositional bias of the HDL. The normalization that directly adjusts scale such as Total Count (TC) and Upper Quartile (UQ) fails to accommodate compositional bias. The normalization strategies that adjust scales using landmarks in the distribution Median, (Med), Differentially expressed (DESeq), Trimmed mean (TMM) are promising approaches for HDL proteome however, detailed analyses need to be performed for their validation to be used on smaller libraries of stochastic count data from mass spectrometry. The quartile (Q) and reads per kilobase per million mapped (RPKM) (equivalent of normalizing spectral counts to protein length) have adverse effects on intra-sample variance and on distribution bias (Dillies et al., 2013) . The TC and UQ normalizations favor the most abundant proteins and are unfriendly for mixtures with a distribution bias. That said, TC normalization is often the preferred method for shotgun HDL proteomics as it controls for differences in instrument response, digestion efficiency, and amounts of loaded protein digest but fails at conserving the distribution bias as it tends to accommodate the changes in the abundant proteins (Vaisar et al., 2007) . That is partly why the HDL protein quantification by shotgun proteomics is not optimal and the correlation with immunobased assays are moderate (Hoofnagle et al., 2012) . Therefore, we opted to include a second normalization approach by spiking yeast carboxy peptidase at levels ~8 fold lower than APOA1 and to correspond to the median/mean abundance of the typical HDL proteome (Carvalho et al., 2008; Liu et al., 2004) . The QTLs identified using both proteomic information are mostly overlapping with TC normalization resulting in ~30% more significant QTLs. The relationships among inbred lines of mice was inferred from the high-density SNP map where strains cluster according to their genealogy (Cervino et al., 2005) . We employed the same approach: The 155 proteins that were present in at least 20% of the strains loosely predicted the relatedness of the strains according to their genealogy for inbred strains and according to the breeding scheme for recombinant strains . Almost half of these proteins (81) were present in greater than 80% of the strains and only 34 were shared by all the strains. The strain dependent distribution of HDL proteome across 93 strains validates our previous studies with only 5 strains (Pamir et al., 2016) . However, the comparison of the clustering patterns between microarray data or the SNPs did not reach full agreement as genetic variation explains only a fraction of the variation and a very small part of the genome is involved in regulating HDL (data not shown). The 93 strains are represented by N=1-5 with a distribution of ~4, 9, 75, 9, and 1 % for N=1,2,3,4, and 5 respectively. Even though, these N are not optimal to calculateintra and -inter strain variation, the broad sense heritability calculations captured 65 proteins that have greater than 10% heritability. Among which APOA2 has a score of 0.62 which is consistent with its strong association with HDL-C loci -a highly heritable trait.