Exploiting Instrumental Variables and Regularized Least Squares for Reverse-Engineering Large Scale Biological Networks from Noisy Data