Authors - Rebecca Hufkie, Dane Brown Abstract - Protein structure determines function, yet experimental determination methods remain costly and slow. This study presents an optimised transformer-based system for predicting protein sidechain angles and reconstructing 3D structures directly from sequence data. Trained on the SidechainNet CASP12 dataset comprising 25,044 proteins, the systematically refined model achieves 0.244 radians RMSE for angle prediction and 1.413 ̊A RMSD for structural accuracy, representing a 70% improvement over the baseline. Incorporating backbone angles, secondary structure, and evolutionary information reduces RMSD from 1.861 ̊A to 1.413 ̊A compared to sequence-only inputs. On challenging CASP12 free-modelling targets, the system scores 84 to 86 GDC. This performance is competitive with leading methods while maintaining computational efficiency through single-sequence prediction without multiple sequence alignment generation. Results indicate that specific architectural choices, including deeper networks, GELU activation, and increased embedding dimensions, combined with robust dropout and weight decay, enable highly accurate structure prediction from limited training data. This demonstrates that carefully constrained models can capture complex biological folding patterns efficiently without massive computational overhead.
Open Zoom