Cerebra: a computationally efficient framework for accurate protein structure prediction

Remarkable progress has been made in the field of protein structure prediction. Representative methods like AlphaFold and RoseTTAFold achieve prediction accuracy close to experimental structural determination, but at the cost of heavy computational consumption for model training. In this work, we propose a new framework, Cerebra, for improving the computational efficiency of protein structure prediction. In this network architecture, multiple sets of atomic coordinates are predicted parallelly and their mutual complementarity is leveraged to rapidly improve the quality of predicted structures through a new attention mechanism named Path Synthesis Attention (PSA). Consequently, Cerebra markedly reduces the computational consumption, achieving evident acceleration on both model training and inference, in comparison to OpenFold, the academic version of AlphaFold2. When evaluated on the CAMEO 2025 dataset, the Cerebra model trained on a limited number of GPUs shows a comparable performance to the published OpenFold model that has been thoroughly optimized using a plethora of computational resources. Moreover, ablation studies confirm that the PSA mechanism efficiently accelerates the network convergence, suggesting its potential in handling the many-body effects of protein geometries in graph-based models. Finally, the fast inference of Cerebra effectively assists hallucination-based de novo protein backbone design, allowing the production of abundant backbone structures with sufficient designability and novelty.