An approach to examining model dependence in EM reconstructions using cross-validation.

Reference bias refers to a common problem in fitting experimental data to an initial model. Given enough free parameters, a good fit of any experimental data to the model can be obtained, even if the experimental data contain only noise. Reference-based alignment methods used in electron microscopy (EM) are subject to this type of bias, in that images containing pure noise can regenerate the reference. Cross-validation is based on the idea that the experimental data used to assess the validity of the fitting should not be the same data as were used to do the fitting. Here we present the application of cross-validation to one form of reference-based alignment: 3D-projection matching in single-particle reconstructions. Our results show that reference bias is indeed present in reconstructions, but that the effect is small for real data compared to that for random noise, and that this difference in behavior is magnified, rather than diminished, during iterative refinement.