Abstract
Cryogenic electron microscopy (cryo-EM) single-particle analysis (SPA) has become a powerful technique for macromolecular structure determination. However, its effectiveness is often constrained by limited particle numbers or missing views. To address these challenges, we present CoCoFold, a fine-tuned framework that integrates raw cryo-EM particle images into AlphaFold to directly guide atomic model prediction. CoCoFold adopts a memory-efficient tuning strategy by introducing a fused attention mechanism into AlphaFold's structure module. Moreover, a differentiable network links predicted structures with cryo-EM observations, enabling end-to-end refinement against experimental data. Benchmark experiments with the escalating quantity insufficiency and view-missing of cryo-EM observations, demonstrate that CoCoFold consistently outperforms state-of-the-art methods across multiple evaluation metrics.