In this work we propose a penalised regression model in which the covariates are known to be clustered into groups, and the clusters are arranged as nodes in a graph. We are motivated by an application to genome-wide association studies in which the objective is to identify important predictors, single nucleotide polymorphisms (SNPs), that account for the variability of a quantitative trait. In this applications, SNPs naturally cluster into SNP sets representing genes, and genes are treated as nodes of a biological network encoding the functional relatedness of genes. Our proposed graph-guided group lasso (GGGL) takes into account such prior knowledge available on the covariates at two different levels, and allows to select important SNPs sets while also favouring the selection of functionally related genes. We describe a computationally efficient algorithm for parameter estimation, provide experimental results and present a GWA study on lipids levels in two Asian populations.