metals.json updated

[GaribMurshudov]

New metal-ligand distance file. New file contains distance statistics for all metal-non metal "bonds" depending on coordination number. If the distance distribution have multiple modes then all they are listed (small modes may be absent)

[GaribMurshudov]

Do I need to do anything else on this branch?

[keitaroyam]

Giving some description about how the file was updated may help.

I need to do some tests.

[GaribMurshudov]

I added some description.

[keitaroyam]

I just wondered why some MAD values were so different from the old metals.json. For example, K-O distance:

# Old
   coord    median       mad      mean       std  count
0      3  2.753206  0.155887  2.808922  0.191985     36
1      4  2.602137  0.046944  2.618844  0.059042     34
2      6  2.761015  0.103126  2.754573  0.139016    722
3      8  2.818235  0.040012  2.818390  0.060503   4680
4     10  2.881260  0.057027  2.879357  0.073333     76

# New
   coord    median       mad      mean       std  count
0      2  3.251116  0.136209  2.905904  0.574963      8
1      3  2.753206  0.096156  2.808922  0.189300     36
2      4  2.602137  0.030799  2.618844  0.058167     34
3      6  2.761015  0.077418  2.754573  0.138920    722
4      8  2.818235  0.027879  2.818390  0.060496   4680
5      9  2.852053  0.031486  2.843358  0.081653     18
6     10  2.879385  0.052707  2.883545  0.075945     87
7     11  2.675899  0.008922  2.698282  0.046524      4
8     12  2.799075  0.074610  2.822165  0.110755     16
9     16  2.782353  0.024644  2.772753  0.042812     16

For coord = 8, it was 0.04 and now is 0.028, despite the same counts. Cs-Cl is much more different:

# Old
   coord   median       mad      mean       std  count
0      3  3.48288  0.000000  3.482880  0.000000      1
1      6  2.91600  0.240004  3.095503  0.270005      9

# New
   coord   median    mad      mean       std  count
0      3  3.48288  0.000  3.482880  0.000000      1
1      6  2.91600  0.001  3.095503  0.254563      9

Has the calculation method been changed?

[Lekaveh]

I checked and recognized that in the previous version deprecated "mad" function from pandas was used which is actually Mean Absolute Error. In the last calculation mediad absolute deviation (np.median(np.abs(distances - np.median(distances)))) was used.

[keitaroyam]

Thank you! Could you provide a link to the calculation code, if available on Github?

[Lekaveh]

It is not available on Github. I can post it here

[keitaroyam]

This metals.json has some inappropriate values. For example, the longest Mg-O distance has too small std/mad, which causes very strong force in (servalcat) refinement if Mg-O distance is long in the model.

 {'coord': 12,
  'median': 2.081025467416805,
  'mad': 0.00028628237234551435,
  'mean': 2.0810127465104333,
  'std': 0.00033989375892045,
  'count': 12,
  'modes': [{'mode': 2.0810127465104333,
    'std': 0.00033989375892045,
    'weight': 1.0}]}

Servalcat was fixed to cap the sigma in 0.4.94, but not released. Probably it is better to fix metals.json as well?

[GaribMurshudov]

It looks like that there is one example in the COD: Mg with coordination 12. And all these values came from that example. Capping is a good practice.
How should we change metals.json? Should it be done by @Lekaveh during generation or should there be a postprocessing?