In some cases more complicated approaches to refinement of macromolecular structures are not necessary

In the manuscript published in IUCrJ under the title `Modeling a unit cell: crystallographic refinement procedure using the biomolecular MD simulation platform Amber' (Mikhailovskii et al., 2022), the authors described the results of the refinement of 84 protein structures using a novel approach that utilizes molecular dynamics and the maximum-likelihood potential that encodes the structure-factor based restraints. They claimed the procedure was superior to traditional approaches used for refinement of macromolecular structures, exemplified by programs such as Refmac5 (Murshudov et al., 2011) or Phenix (Adams et al., 2010). In our opinion, the only example for which the re-refined structure was deposited in the Protein Data Bank (PDB) does not provide convincing proof for the superiority of this refinement method.

The structure chosen as an example represented Type III antifreeze protein isoform HPLC12 (PDB entry 2msi; DeLuca et al., 1998). The structure was originally refined at 1.9 Å resolution, but it appears that the refinement was not finalized. In particular, the deposited coordinates lacked any solvent that should have been visible at this resolution, thus despite good validation statistics (Table 1), that structure has to be considered of uncertain quality.

Table 1
Selected statistics of the three models of Type III antifreeze protein isoform HPLC12

	Original structure (PDB entry 2msi)	Amber-refined (PDB entry 7q3v)	Refmac5-refined (PDB entry 9cbe)
Protein atoms (non-hydrogen)	485	1952	495
Solvent atoms	0	155	39
R/Rfree	0.193/0.261	0.160/0.194	0.152/0.189
Clashscore	3.99 (98th percentile)	4.45 (97th percentile)	1.94 (100th percentile)
MolProbity score	1.61 (92nd percentile)	1.64 (90th percentile)	1.37 (98th percentile)
Bad bonds	0	0	0
Bad angles	8	16	1

In the procedure used by Mikhailovskii et al. (2022), the original unit cell that contained a single protein molecule in the asymmetric unit in the space group P2₁2₁2₁ was expanded to P1 with identical unit-cell parameters, but containing four independent molecules. Thus, the number of refinable parameters was significantly increased, though the amount of available diffraction data remained the same. Their re-refined structure consisted of 1952 protein atoms and 155 water molecules (Table 1). The resulting model (PDB entry 7q3v) corrected obvious mistracing at the N-terminus and added missing solvent. As described by the authors, this was accomplished after very extensive calculations that could only be done on a computer equipped with GPUs.

To put this effort into perspective, we re-refined the structure 2msi with Refmac5, after manually correcting the erroneous tracing of the N-terminus and several improper side-chain rotamers with the program Coot (Emsley et al., 2010). Waters were added automatically in Coot on the basis of 3.5σ peaks in the F_o − F_c electron density map. Statistics of the resulting structure (PDB entry 9cbe), some obtained with MolProbity (Chen et al., 2010), are also shown in Table 1. The process took less than an hour of human time and negligible computing time on a standard PC.

In principle, the more adjusted parameters, the better the agreement between the observed and calculated functions. However, if more parameters do not lead to better agreement, the strict Ockham razor should be mercilessly applied to avoid unnecessary complication of the problem. It is clear that, at least in this case, the complicated procedure involving molecular dynamics did not work any better than manual refitting and standard crystallographic refinement. Although the procedure utilized by Mikhailovskii et al. (2022) might be potentially very useful in some specific cases, the sole example out of the 84 that were deposited in the PDB was maybe not an optimal selection. A more difficult or complicated case might have provided a clearer indication of the usefulness of this novel structure refinement procedure.