RocketX

RocketX predicts de novo three-dimensional protein structures from amino acid sequences using deep learning to infer inter-residue geometric constraints and iteratively refine models.

Key Features:

• Closed-Loop Feedback Mechanism: Integrates GeomNet, a structural simulation module, and EmaNet to iteratively refine protein structure predictions.
• GeomNet (Geometric Constraint Prediction Network): Extracts co-evolutionary features from multiple sequence alignments (MSA) sourced from sequence databases and uses an improved residual neural network to predict inter-residue geometric constraints.
• Structural Simulation Module: Simulates folding of structure models based on the geometric constraints predicted by GeomNet.
• EmaNet (Model Quality Evaluation Network): Extracts 1D and 2D features from folded models and uses a deep residual neural network to estimate inter-residue distance deviations and per-residue local Distance Difference Test (lDDT) scores, providing dynamic feedback to GeomNet.
• Benchmarking and Performance: Evaluated on 483 benchmark proteins and 20 CASP14 free-modeling (FM) targets, outperforming trRosetta (without templates) and RaptorX and showing comparable accuracy to template-integrating methods in CAMEO blind tests on medium and hard targets.

Scientific Applications:

• Protein Structure Prediction: Predicts three-dimensional protein structures from amino acid sequences to support analysis of protein function and interactions.
• Folding Mechanism Insights: Provides template-free predictions that contribute to investigating intrinsic protein folding mechanisms.

Methodology:

GeomNet extracts co-evolutionary features from MSAs from sequence databases and uses an improved residual neural network to predict inter-residue geometric constraints; a structural simulation module folds models based on those constraints; EmaNet extracts 1D and 2D features from folded models and uses a deep residual neural network to estimate inter-residue distance deviations and per-residue lDDT scores and returns dynamic feedback to GeomNet in a closed-loop iterative refinement.