Cerebra: a computationally efficient framework for accurate protein structure prediction

Abstract

Remarkable progress has been made in the field of protein structure prediction in the past years. State-of-the-art methods like AlphaFold2 and RoseTTAFold2 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 protein structure prediction framework, Cerebra, for improving the computational efficiency of protein structure prediction. In this innovative network architecture, multiple sets of atomic coordinates are predicted parallelly and their mutual complementary is leveraged to rapidly improve the quality of predicted structures through a novel attention mechanism. Consequently, Cerebra markedly reduces the model training consumption, achieving a training acceleration of at least 7 folds, in comparison to OpenFold, the academic version of AlphaFold2. When evaluated on the CAMEO and CASP15 sets, the Cerebra model insufficiently trained on a single GPU only shows slight performance inferiority to the published OpenFold model.

GDFold2: a fast and parallelizable protein folding environment with freely defined objective functions

“…Modeling the protein structures compatible with such prediction results is feasible in Rosetta only by converting the information-abundant C α coordinates into distributions of inter-residue distances, a procedure that clearly leads to severe information loss. In another work, we developed a novel protein structure prediction framework Cerebra 22 , which significantly improves the computational efficiency by predicting multiple sets of atomic coordinates in parallel and leveraging their complementarity for prediction error suppression. In each set of the prediction results, Cerebra outputs the relative translations and rotations (in quaternions) of the local coordinate systems of all residues in the reference frame of one anchor residue.…”