Generalization of the default sparsification neighborhood size

Determine whether the default sparsification neighborhood size $n_{\text{nbr}}$ used in the weighted similarity aggregation framework generalizes across diverse real cryo-EM datasets.

Background

The aggregation framework constructs a particle-similarity matrix from weighted co-assignment matrices and sparsifies each row by retaining only its nnbrn_{\text{nbr}} strongest similarities before clustering. Hyperparameter ablations show that this parameter has a much larger effect on classification accuracy than the exponential weight-heuristic parameters.

Although the experiments use a fixed default value and demonstrate successful results on the Tomotwin and EMPIAR-10076 datasets, the authors explicitly leave unresolved whether that value is transferable to other real datasets. Establishing such generalization would be important for routine deployment of the framework without dataset-specific tuning.

References

Whether our default value for $n_{\text{nbr}}$ generalizes across diverse real datasets remains to be tested.

— A meta-algorithm for ab initio reconstruction of complex mixtures in cryo-EM  (2608.25388 - Kaz et al., 26 Aug 2026) in Section 6, “Discussion,” paragraph beginning “Several limitations warrant discussion”