A model of the lumbar spine capable of taking into account realistic loads derived from human activity would be of great benefit in studying its normal biomechanical functioning as well as its in vivo behavior in injured and surgically altered states. This paper proposes a method to analyze the mechanical response of the lumbar spine subjected to loads derived from human activity, combining a non-linear finite element model (FEM) and an optimization-based force predicting algorithm. Loads borne by the lumbar spine at the T12-L1 level (joint loads) are first predicted with the optimization algorithm and then applied to the FEM, while a boundary condition prescribing the relative L1-sacrum rotation is imposed onto the FEM to account for three-dimensional physiological thorax-pelvis orientation. The prescribed rotation is achieved through the application of moments on L1. To account for the effect of these moments on lumbar joint loads, an iteration between the optimization technique and the FEM computation has been carried out. This method provides two main benefits over previous studies: first, it allows for the application of any 3D loading condition while considering the real 3D rotation measured between the thorax and the pelvis, and second, it makes it possible to estimate the moments that must be applied on L1 in order to maintain this rotation, taking them into account when predicting joint loads. As an example application of the method, results are presented for the lumbar spine mechanical response at the time of peak T12-L1 joint force during walking.