The most effective method for structure determination of molecular crystals from powder data is simulated annealing. The structure solution complexity can be described by DOF parameter - the number of free parameters witch need to be determined during the calculation. A search in CSD had however revealed the current record (DOF 42, CSD code NIELSEH) is 12 year old. The other structures in CSD claimed "to be solved from powder'' with higher DOF were found to be based on known single crystal structure or to be totally non reliable based on low quality powder diffraction data. A study of possible methodical extension of simulated annealing as implemented in DASH software is given in [1].
There exist following approaches to speed up the calculation and to solve structures with high DOF:
- parallel processing on multiple CPU, multiple PC or even on rented Amazon PC cluster
- application of torsion angles statistic based on the Mogul module of CSD
- optimization of the simulated annealing algorithm parameters
We had tested this approaches on several complex structures solved recently in our laboratory namely methylergometrin maleate, metergloline phase II, ixazomib phase II and X, and selexipag phase VI. The results in the form of success rate (SR = number of correct solutions/number of required SA runs in %) is summarized in Tab. 1.
|
Compound name |
DOF |
Default DASH run SR,% |
Mogul biased improved run SR,% |
Optimized SA parameters SR, % |
Mogul bias and SD optimized parameters run SR, % |
|
methylergometrin maleate |
17 |
7 |
21 |
37 |
50 |
|
metergloline phase II |
24 |
2 |
13 |
20 |
28 |
|
ixazomib phase II |
32 |
0 |
0.2 |
0.4 |
0 |
|
ixazomib phase X |
32 |
1.4 |
0.8 |
8.2 |
4.8 |
|
selexipag phase VI |
38 |
0 |
0.075 |
0.5 |
0.1 |
The speed up of parallel processing of DASH code executed by the help of MDASH extension was tested on a PC with 2x Intel Xeon E560, 2.4 GHz processors (8 real cores). We wanted to know whatever the utilization of hyper-threading and virtual 16 processors can give any benefit. The speed tests were done for the methylergometrin maleate structure. Results are summarized in Tab. 2.
|
Execution methods |
1 CPU core from DASH GUI |
1 CPU core from MDASH |
2 CPU core from MDASH |
4 CPU core from MDASH |
8 CPU core from MDASH |
12 CPU core from MDASH |
16 CPU core from MDASH |
|
Time required to find 1 correct solution/sec |
1113 |
1144 |
470 |
240 |
108 |
116 |
126 |
The use of optimized SA parameters can clearly make a difference between finding a solution or not finding solution at all, especially for high DOF structures. The use of Mogul bias looks a bit problematic - for non usual structures with atypical geometry (ixazomib with boron atom) it can give less benefit than for common molecules. The time required for solution is reduced proportionally to number of true CPU used for calculation. The use of virtual hyper-threading CPU gives no benefit and slow-down the calculations. For methylergometrin maleate the combination of optimized SA parameters, Mogul bias used and parallel processing on 8 CPU give up to 47x speed up in comparison to standard calculation setup.
This work was supported by the Czech Science Foundation (GA ČR) project, Grant No. 16‑10035S.