CryoConvNeXt enables robust identification of low-abundance molecular species in experimental cryo-EM data.

Rare but biologically important molecular species are easily lost when reconstruction-based cryo-EM classification is applied to heterogeneous samples dominated by more abundant particles. We developed CryoConvNeXt, a deep-learning classifier that combines cyclic equivariance with adaptive frequency filtering. It is trained on simulated projections and adapted to experimental data by self-training, an unsupervised domain adaptation technique where the model acts as its own annotator. We tested Cryo-ConvNeXt on cryo-EM data collected for this study from controlled binary and ternary mixtures of Catalase, Apoferritin, and HSP60. Manual particle curation provided reference labels for benchmarks with class ratios from 1:1 to 1:16. CryoConvNeXt retained minority-species recall across this range. In contrast, cryoSPARC's recall collapsed in four of six pairwise conditions, as early as 1:4 when Catalase was the minority species. By recovering low-abundance species that conventional classification fails to recover reliably, CryoConvNeXt takes an important step towards solving a central problem in quantitative visual proteomics: measuring the molecular composition of complex biological samples.