Cryo-EM heterogeneity analysis stops treating particle poses as fixed: SOLVAR refines poses alongside principal components and leads ab-initio benchmarks
Synopsis
Continuous-heterogeneity analysis in cryo-EM has long assumed particle poses are known, yet those poses come from homogeneity-assuming refinement, so their errors leak into the recovered structural variability. SOLVAR assumes the volume covariance matrix is low-rank, rewrites principal-component estimation as an optimization solvable by stochastic gradient descent, and folds particle poses and contrast into the same optimization. On CryoBench synthetic data it scores highest on most datasets, and in an ab-initio pipeline seeded by RELION poses its per-image FSC area under the curve exceeds that of the other ab-initio methods evaluated by the benchmark.
Interpretation
Continuous-heterogeneity analysis in cryo-EM can now correct particle poses while estimating structural variability, instead of taking poses as a fixed input. Earlier covariance-based methods, and most other heterogeneity methods, require the user to supply particle poses and cannot themselves refine those initial poses; in realistic pipelines the poses are computed by software that ignores heterogeneity and are distorted by it. On the CryoBench datasets, initial poses from RELION homogeneous refinement followed by SOLVAR with pose optimization reduce pose error on most datasets; the same ab-initio pipeline's per-image FSC area under the curve exceeds that of the other ab-initio methods evaluated by CryoBench.
The method constrains the volume covariance matrix to be low-rank, avoiding direct estimation of an N³×N³ matrix. Early covariance methods estimated the entire covariance matrix, whose size is prohibitively large for practical computation and limited them to very low resolution; more recent methods estimate the principal components directly instead. After the low-rank substitution the objective and gradient cost O(nr²N² + r²N³) operations, and with mini-batch stochastic gradient descent total complexity scales linearly with the principal-component volume size N³; on the IgG-1D dataset with 100,000 particle images of 128 pixels and r = 10, SOLVAR takes 25 minutes without pose optimization and 90 minutes with it, against 60 minutes for CryoDRGN and 20 minutes for RECOVAR on the same A100 node.
Perspective
The evaluation covers CryoBench synthetic datasets plus two experimental datasets, EMPIAR-10076 of 50S ribosomal-subunit assembly intermediates and EMPIAR-10180 of the pre-catalytic spliceosome; for the experimental data, principal components are estimated on images downsampled to 128×128 pixels and volumes are then reconstructed at the original 256×256 pixels. The intended readers are researchers and pipeline developers working with cryo-EM single-particle data who need to correct poses during heterogeneity analysis.
The number of components must be supplied by the user and is typically not known in advance; the Gaussian heterogeneity assumption is described by the authors as not physically grounded and somewhat arbitrary, though close to models used in common cryo-EM software and empirically effective. The authors also note that the per-image FSC area-under-the-curve metric used by CryoBench is itself susceptible to bias, and list extension to a fully ab-initio setting and frequency marching for faster runtime as future directions.
