This paper introduces a novel approach to deformation modeling based on partial-least-squares-regression and its application to the registration of first-pass magnetic resonance perfusion image sequences. The method relies on the extraction of intrinsic correlations between the latent factors of both the input and output signals for deriving a simple yet accurate deformation model. The strength of the technique has been demonstrated with both numerically simulated data sets and a myocardial perfusion study which involves nine patients with known coronary artery disease. The method represents a step forward for the commonly used principal component analysis for salient motion feature extraction. The proposed technique should be applicable to studies that involve deformation prediction such as motion adaptive radiotherapy and imaging, where the deformation of the target organ can be predicted from externally measurable signals.