We look at the (failure) behavior of structural systems under the occurrence of seismic events. Detailed, long-running Finite Element Models (FEMs) are typically adopted for the simulation of the system structural response. However, since a very large number of simulations (e.g., several thousands) is usually required for an accurate assessment of the system failure behavior under different seismic conditions, the computational burden associated to the approach may be at times impractical. To address this issue, we resort to (fast-running) Artificial Neural Network (ANN) regression models, optimally trained to approximate the response of the original FEM. In particular, we embrace a wrapper approach based on Genetic Algorithm (GA) optimization to search for the optimal set of model inputs (features) that maximize the ANN representation accuracy. The ANN is trained, validated and tested with respect to the optimal feature subset identified; then, the results are compared with those produced by the original FEM. We illustrate the approach on a case study of seismic risk assessment involving the estimation of the fragility curves for a masonry structure.