Many cryo-EM reconstruction runs produced a much stronger classification result than the best single run in a 45-class synthetic test, according to an arXiv preprint. The ensemble reported 97.4% accuracy, compared with 34.6% for the best single-shot run. Its median angular pose error, the gap between estimated and reference viewing angles, was 1.71 degrees, versus 74.1 degrees for the single run.
The study asks whether aggregating weak multiclass ab initio, or from-scratch, reconstruction jobs can resolve complex mixtures beyond what a single run can handle. The proposed algorithm is organized around ensemble generation, weighted similarity scoring and clustering.
From many runs to one assignment
The experiments repeatedly used cryoSPARC's “Ab-Initio Reconstruction” job through the cryosparc-tools API. Parameters were kept at their defaults except for the class count and random seed.
Final assignments were based on weighted co-assignment similarities, a score of how often particle images were placed together across jobs. The similarity matrix was then sparsified, meaning each row kept only its strongest nearest-neighbor links, before average-linkage agglomerative clustering grouped the particles.
The one-run ceiling
The synthetic tests used Tomotwin-100, which contains 100 structures and 1,000 images per structure, plus Tomotwin-45-asym, a 45-class asymmetric subset.
The single-shot baseline showed a sharp change as the mixture grew. On diagonal cases through 32 classes, single-shot runs were near-perfect and had median pose errors of 2 to 3 degrees. When complexity reached 45 classes, runs using 32 or more classes failed to converge, while those using 16 or fewer converged.
On Tomotwin-45-asym, the ensemble reported 97.4% for each of three classification measures: accuracy, precision and macro F1. Its mean pose error was 6.11 degrees, and its median pose error was 1.71 degrees.
On Tomotwin-100, ensemble accuracy was 75.0%, versus 15.2% for the one-shot baseline. The authors described that difference as a five-fold improvement.
From controlled mixtures to ribosomal particles
The harder test used the full, unfiltered EMPIAR-10076 stack of E. coli large-ribosomal-subunit assembly intermediates. It included four major states, two rare species and junk.
Without manual filtering, the ensemble separated all four major assembly states from junk. Homogeneous refinement produced reconstructions at resolutions of 4 Å or better for all four states.
In the annotation comparison, a five-class result reached 75% accuracy and comprised the four major classes plus junk. A seven-cluster analysis did not capture the rare A and F states, while it split D and E into plus and minus subtypes.
The choice that mattered most
Across both datasets, accuracy typically improved with early hierarchical depth and additional replicates. Parallel execution meant wall-clock time depended mainly on tree depth rather than on the total number of jobs.
Parameter sensitivity was uneven. Weight parameters changed accuracy by fewer than 10 percentage points, but nnbr, the nearest-neighbor setting used in sparsification, produced swings of more than 45 percentage points.
The reported figures are point estimates: no confidence intervals or variance estimates were given for the benchmark results, and the real-data reconstruction results had no uncertainty intervals.
In these tests, the major states were recovered, but rare states A and F were not, and accuracy changed sharply with the sparsification neighborhood.
Still a preprint
The document is an arXiv version 1 preprint dated 26 August 2026. It acknowledges support from Princeton Research Computing, public and foundation funders, Princeton initiatives and industry partners, and states that funders had no role in the research or publication process.
Paper data and sources
Original title: A meta-algorithm for ab initio reconstruction of complex mixtures in cryo-EM
Authors: Alkin Kaz, Arda Kaz, Ellen D. Zhong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text