Résumé
Background : Principal Components Analysis is a standard and computationally efficient method to explore large SNP data sets. We propose in this study additional interpretations of PCA results about the characterization of population genetic structure when dealing with SNP data. In particular, we evaluate how SNP typological values obtained from PCA are related to F-statistics and may help to identify footprints of selection. Results : We show that a normed PCA on biallelic SNP haplotypes is equivalent to a Multiple Correspondence Analysis and to a PCA on the r correlation matrix, where r represent the signed square root of the r2 linkage disequilibrium measure. Each resulting principal component describes a typology and provides a measure of the underlying SNP contributions which may further be interpreted in terms of correlation ratio and variance reduction. In addition, PCA can be partitioned into sub-analyses (between-group, within-group). Betweengroup PCA maximises the variance between groups and delivers principal components with maximum FST. Only per-group allele frequencies and relative frequencies are needed to compute between-group PCA. Finally, chromosomal regions containing SNPs with high contributions may be interpreted as footprints of selection. As an illustration of the approach we analyzed human chromosome 2 haplotypes sampled from three HapMap populations (from African, Asian and European origin). We showed that SNPs within or close to EDAR and LCT genes exhibit the highest typological values, in agreement with previous studies. Conclusions : When applied to biallelic SNP data, our PCA based proposed approach enables to describe the genetic structuring of populations and to quantify for each typology the contributions of SNPs by FST statistics. Taking into account spatial dependences of SNPs allows in turn to identify genomic regions contributing to the structuring of populations which might be interpreted as footprints of selection. Finally, this approach was proven computationally efficient since it can handle data including several hundreds of thousands SNPs within less than one hour on a standard computer.