Abstract
In this paper, we present two novel approaches, PICNIC2-LOCAL and PICNIC2-GLOBAL, for refining AlphaFold protein tertiary structure (TS) predictions using deep 3D residual neural networks. These approaches aim to improve the accuracy and consistency of TS predictions, addressing a key challenge in protein structure prediction. PICNIC2-LOCAL and PICNIC2-GLOBAL effectively learn by putting each atom of the AlphaFold model in a voxel grid, with feature extractions of each atom, to learn the structural conformations and deviations to the native structure. PICNIC2-LOCAL learns the features locally, in which we used a smaller voxel grid with the target atom put in the center. Whereas for PICNIC2-GLOBAL, the voxel grid covers the whole protein. We trained on 30,077 AlphaFold TS predictions from the UniProt database and performed a blind-test evaluation on AlphaFold TS predictions from CASP15. The results indicate consistent improvement in atom-specific (reduction of per-atom deviation to native) and whole-protein (measured by TM-score) structural accuracy. Both models excel at refining the FM/TBM and FM target proteins in CASP15. PICNIC2-LOCAL and PICNIC2-GLOBAL can be freely accessed at https://github.com/zwang-bioinformatics/PICNIC/ and http://dna.cs.miami.edu/PICNIC/.
