Compressive sensing (CS) MRI aims to accurately reconstruct images from undersampled k-space data. Most CS methods employ analytical sparsifying transforms such as total-variation and wavelets to model the unknown image and constrain the solution space during reconstruction. Recently, nonparametric dictionary-based methods for CS-MRI reconstruction have shown significant improvements over the classical methods. These existing techniques focus on learning the representation basis for the unknown image for a synthesis-based reconstruction. In this paper, we present a new framework for analysis-based reconstruction, where the sparsifying transform is learnt from a reference image to capture the anatomical structure of unknown image, and is used to guide the reconstruction process. We demonstrate with experimental data the high performance of the proposed approach over traditional methods.