I have a ~4.3 angstrom data set of a trigonal crystal of a seven subunit protein complex which I can scale in P3, P31, P32, P321, P3121 and P3221 with similar statistics:
P3
Shell Lower Upper Average Average Norm. Linear Square
limit Angstrom I error stat. Chi**2 R-fac R-fac
50.00 9.25 1296.8 89.2 23.5 1.233 0.064 0.077
9.25 7.35 356.3 18.5 9.7 1.512 0.065 0.066
7.35 6.42 97.1 8.2 7.5 1.584 0.143 0.140
6.42 5.83 55.2 8.3 8.1 1.503 0.247 0.241
5.83 5.42 51.4 9.4 9.3 1.438 0.297 0.284
5.42 5.10 47.0 10.5 10.5 1.469 0.374 0.345
5.10 4.84 48.3 11.8 11.9 1.421 0.398 0.383
4.84 4.63 43.6 12.9 13.1 1.474 0.488 0.449
4.63 4.45 40.3 14.1 14.2 1.530 0.546 0.477
4.45 4.30 30.8 14.7 15.0 1.601 0.732 0.631
All reflections 203.8 19.6 12.3 1.477 0.125 0.085
P3121:
Shell Lower Upper Average Average Norm. Linear Square
limit Angstrom I error stat. Chi**2 R-fac R-fac
50.00 9.14 1242.9 51.8 18.3 1.200 0.057 0.068
9.14 7.26 314.0 11.2 6.5 1.454 0.070 0.069
7.26 6.35 86.9 5.3 5.0 1.499 0.158 0.152
6.35 5.77 51.9 5.5 5.3 1.248 0.264 0.252
5.77 5.35 46.9 6.1 6.0 1.213 0.330 0.305
5.35 5.04 44.3 6.9 6.7 1.137 0.393 0.363
5.04 4.79 43.4 7.7 7.4 1.109 0.434 0.407
4.79 4.58 39.2 8.5 8.1 1.128 0.533 0.478
4.58 4.40 34.2 9.1 8.6 1.115 0.634 0.549
4.40 4.25 24.9 9.9 9.3 1.064 0.872 0.766
All reflections 199.0 12.4 8.1 1.216 0.127 0.080
Unit cell parameters: 129.653 129.653 358.280 90.000 90.000 120.000
The systematic absences are consistent with either P31, P32, P3121, or P3221. Analyzing the cell contents in P3121 suggests either 1 (Matthews coefficient of 3.86, 68.2% solvent) or 2 mol/ASU (Matthews coefficient of 1.93, 36.38% solvent)
I built a molecular replacement model (a polyala model containing about 2/3 of the protein complex) and ran phaser in multiple space groups with one (for P3121 or P3221) or two (P31, P32) copies of the model. Runs in P32 or P3221 gave no solutions or solutions with TFZ around 4-5. When run in P31 or P3121, phaser output solutions with TFZ> 11.0 and what appeared to be good packing.
Rigid body refinement on the P3121 solution failed to improve the Rfactor (it hovered around 55.3%). Adding the missing subunits (as polyala chains) based on the phaser solution and refining with rigid body refinement resulted in a model with an Rfree to 48.5. Refining with torsion angle dynamics and restrained group B-factor refinement made the Rfree worse – it jumped up to about 55.6%. The Rwork values were similar to the Rfree values for each attempt. I also tried DEN refinement with similar results.
Rigid body refinement of the P31 phaser solution gave an Rfree of about 54.4%. Adding the missing subunits and running rigid body refinement again improved the Rfree to 53.0. Refining with torsion angle dynamics and restrained group B-factor refinement again made the Rfree worse (increased to 54.5%).
I analyzed the reflection file processed in P31 using detect_twinning.inp in cns. The data did not appear to be perfectly merohedrally twinned, but in the test for partial merohedral twinning, the twin fraction calculated for “2 along a,b” was 0.475. I repeated rigid body refinement, then torsion angle dynamics with restrained group b-factor using the calculated twinning parameters. This brought the free R down to 46.3%, but caused significant divergence between Rwork and Rfree (Rwork =21.4%(!)). The Rfree is fairly constant across resolution shells, but Rwork drops dramatically with low resolution reflections (In the 50 – 9.14 ang shell, Rwork = 12.3%!).
I’m guessing that because the twinning fraction is near 0.5, detwinning is not working. Does anyone have any suggestions about how to successfully refine this structure (assuming it is possible)? Should we average the twin related reflections to generate perfectly twinned data, and if so, how do we do that? Is the twinning likely responsible for our difficulty refining the structure or could there be a problem with the space group assignment? Why does including the partial twinning in our refinement cause Rfree and Rwork to diverge so dramatically? Given the trouble I’ve had so far and the poor quality of the data, I’m about ready to give up on this structure, but if anyone has any ideas please let me know.
