The reference-free averaging of three-dimensional electron microscopy (3D-EM) reconstructions with empty regions in Fourier space represents a pressing problem in electron tomography and single-particle analysis. We present a maximum likelihood algorithm for the simultaneous alignment and classification of subtomograms or random conical tilt (RCT) reconstructions, where the Fourier components in the missing data regions are treated as hidden variables. The behavior of this algorithm was explored using tests on simulated data, while application to experimental data was shown to yield unsupervised class averages for subtomograms of groEL/groES complexes and RCT reconstructions of p53. The latter application served to obtain a reliable de novo structure for p53 that may resolve uncertainties about its quaternary structure.