TorchRef: an open-source PyTorch framework for crystallographic refinement

Seidel, H.-P.; Standfuss, J.; Weinert, T.

doi:10.1107/S1600576726009258

Figure 2
AlphaFold benchmark; 767 Phaser-solved candidates. Panels (a)–(c) use the conserved subset of 725 structures on which all three engines refined and Phenix-validated successfully. Panel (d) additionally excludes trajectories with less than 0.02 total $[R_{{\rm free}}]$ improvement, leaving 676 REFMAC5, 663 phenix.refine and 666 TorchRef trajectories. R factors are Phenix-validated under constant scoring conditions; geometry RMSZ is REFMAC5-validated against a uniform restraint library. Engines: unrefined AlphaFold prediction (green), REFMAC5 (purple), phenix.refine (blue), TorchRef (red). (a) $[R_{{\rm free}}]$ versus $[R_{{\rm work}}]$ scatter. Dashed lines mark per-engine medians. (b) Geometry RMSZ strip (bond, angle, chiral, main-chain B value); point = median, band = 25–75th percentile. (c) Wall-clock runtime (log scale). (d) Convergence speed: fraction of total $[R_{{\rm free}}]$ improvement versus macrocycle, shown as the median across structures with a 25–75th percentile band.

JOURNAL OF
APPLIED
CRYSTALLOGRAPHY

ISSN: 1600-5767

Volume 59| Part 6| December 2026|

https://doi.org/10.1107/S1600576726009258

Open

access