We deal with the problem of gene selection when genes must be selected group-wise, where the groups, defined a priori and representing functional families, may overlap. We propose a new optimization procedure for solving the regularization problem proposed in Jacob et al. (2009), where the group lasso penalty is generalized to overlapping groups. While in Jacob et al. (2009) the proposed implementation requires replication of genes belonging to more than one group, our iterative procedure, provides a scalable alternative with no need for data duplication. This scalability property allows avoiding the otherwise necessary pre-processing for dimensionality reduction, which is at risk of discarding relevant biological information, and leads to improved prediction performances and higher selection stability.