In the single-particle approach, a large number of randomly oriented projection images of a molecule are assembled into a consistent 3D view. Without an initial model, it is difficult to make an initial classification of the images and assign viewing angles to them.

We use a probabilistic model to assign viewing angles based on the relative similarity of the images. The model does not exploit any 3D information, but is completely based on the assumption that neighboring views have a higher similarity as views from very different angles. Performing a maximum-likelihood estimation of the viewing angles of all images simultaneously produces an algorithm similar to a self-organizing map.