CR-I-TASSER is a hybrid method to integrate I-TASSER and cryo-EM density map for high-quality protein structure determination. Starting from density map, it first uses deep convolutional neural networks (CNNs) to predict C-alpha positions, which are used to improve threading templates by the sequence-independent template and C-alpha position superpositions. Next, the deep-learning boosted threading templates are reassembled by the I-TASSER based structure assembly simulations to generate full-length atomic models under the guidance of density map and template restraints. Large-scale benchmark tests showed a significant advantage of CR-I-TASSER over other de novo and refinement-based approaches in cryo-EM structure determination.