Matching biomolecular structures by registration of point clouds
Motivation: Assessing the match between two biomolecular structures is at the heart of structural analyses such as superposition, alignment and docking. These tasks are typically solved with specialized structure-matching techniques implemented in software for protein structural alignment, rigid-body docking, or rigid fitting into cryo-EM maps. Results: We present a unifying framework to compare biomolecular structures by applying ideas from computer vision. The structures are represented as three-dimensional point clouds and compared by quantifying their overlap. We use the kernel correlation to measure point cloud overlap, and discuss local and global optimization strategies for maximizing the kernel correlation over the space of rigid transformations. We derive a majorization-minimization procedure that can be used to register two point clouds without establishing a point-to-point correspondence. We demonstrate that the majorization-minimization algorithms outperform the commonly used Iterative Closest Point registration algorithm. Furthermore, we discuss and benchmark a randomization strategy for globally optimizing the kernel correlation. We illustrate the approach on various 3D fitting problems such as the comparison of circularly permuted structures and rigid fitting of cryo-EM maps or bead models from small-angle scattering.
PDF Abstract