Bayesian Flexible Shape Matching with Applications to Structural Proteomics

We introduce a method for flexible shape registration using Bayesian changepoint analysis. Our approach is particularly suitable for shapes containing “hinge”-like flexibility, and is motivated by problems in structural proteomics and bioinformatics. We define a class of flexible shape spaces and an associated Procrustes-type metric, along with a highly efficient algorithm for computing flexible shape distance. We use this distance to define distributions over flexible shapes, and develop Bayesian models for several variants including both affine and rigid-body component transformations. We demonstrate the approach on several examples arising from protein structure analysis, including structure alignment and function discovery.