Neural network fills gaps in protein maps from raw diffraction data

The work

Technical · from the record

The authors trained CrysFormer, a hybrid 3D vision transformer and CNN, to output protein electron-density maps from two inputs: a Patterson map computed directly from simulated diffraction data, and an incomplete template map derived from an AlphaFold Database prediction with 3 to 7 of its 15 residues removed. Training used a synthetic data set of 15-residue fragments in P21 unit cells, built from 589 546 Patterson map and ground-truth density pairs expanded to 1 634 839 examples by generating up to three different residue-omission patterns per pair. On a held-out test set of 64 070 initial examples, the paper reports higher in-model-region Pearson correlation with the ground truth and lower unweighted and FOM-weighted phase errors for the model predictions than for the same partial-structure templates processed with the SIGMAA PARTIAL baseline, both over the full test set and over a 19 436-example subset with the poorest alignment of the AlphaFold-derived template to the ground truth. Pearson correlations of the predictions showed no apparent dependence on the solvent content of the ground-truth unit cell.

How AI was used

A single learned model does the central work. 15-residue fragments were extracted from a basis of nearly 38 000 curated PDB structures, placed in P21 unit cells, and used to compute structure factors with gemmi sfcalc and then Patterson and ground-truth electron-density maps with the CCP4 FFT program at resolution limits binned across 1.75-2.3 A. Matching AlphaFold Database v4 predictions were located through UniProt identifiers from PDBe SIFTS, aligned to each fragment by Needleman-Wunsch sequence alignment and then by structural alignment with the PyMOL align command as a stand-in for molecular-replacement placement, and had a randomly chosen contiguous block of 3 to 7 residues removed from one or both ends to form partial-structure template maps. CrysFormer, a 3D vision transformer with Nystrom approximate attention combined with scale-equivariant 3D convolution and batch-normalisation layers, was then trained from scratch to regress the ground-truth density from the Patterson map and the single same-sized partial-structure template, using mean-squared error plus a negative Pearson correlation term, a Schedule-Free AdamW optimiser under a OneCycle schedule, and data-parallel training over two GPUs. The trained model was run over the held-out test examples, and its predicted maps and derived structure factors were scored with Phenix get_cc_mtz_pdb and CCP4 CPHASEMATCH against the same templates improved by the non-learned SIGMAA PARTIAL procedure.