This command-line script is designed to compare the packing of a set of input structures of a molecule (polymorphs, co-crystals, solvates, and hydrates). An all-to-all comparison of the structures is performed, considering only the heaviest components in each structure, and a packing-similarity dendrogram or tree is constructed using hierarchical clustering. This dendrogram shows the similarity between groups of structures and how these groups relate to one another.

CSD Python API (v. 1.0 or later), matplotlib and standard python packages.

Instructions on running example scripts through the CSD Python API miniconda installation can be found in the repository ReadMe.md

• CSD-Materials or CSD-Enterprise

python Packing_Similarity_Dendrogram.py -h will show the help text:

usage: Packing_Similarity_Dendrogram.py [-h] [-m similarity_matrix.txt]
[-ns 25] [-nm 15] [-o]
[--allow_molecular_differences]
[--clustering_type {complete,single,average}]
[-s] [-ct 0.5] [-at 25] [-dt 0.25]
input_file

Packing_Similarity_Dendrogram.py - Construct a dendrogram for an input set of structures based
on packing-similarity analysis

positional arguments:
input_file Set of structures to perform analysis on
[.mol2/cif/res/ind].

optional arguments:

-h, --help show this help message and exit
-m similarity_matrix.txt, --matrix similarity_matrix.txt
NumPy matrix containing existing packing similarity
results.
-ns 25, --n_structures 25
Number of structures to take from input set.
-nm 15, --n_molecules 15
Size of molecular packing shell to use for analysis
(must be consistent with input matrix, if used).
-o Flag for whether to save packing similarity results
(text file and mol2 overlays).
--allow_molecular_differences
Flag for whether to allow for molecular differences
between structures (e.g. for salts).
--clustering_type {complete,single,average}
Type of clustering to employ
-s, --strip Strip all terminal atoms and alkyl chains, up to any
hetero atom (O, N, S) or cyclic atom. This cuts the


molecule down to core structural features and may
reveal more general structural similarities, including
those between molecules with different conformations.

-ct 0.5, --conf_tol 0.
RMSD threshold for considering two conformations to be
the same (when merging at level 1).
-at 25, --angle_tol 25
Tolerance for angles (in degrees) used by packing
similarity.
-dt 0.25, --dist_tol 0.
Fractional tolerance for distances (0.0 - 1.0) used by
packing similarity.
-pd 5, --pad_length 5
pad right hand side of plot to make room for structure names

python Packing_Similarity_Dendrogram.py input_file

where input_file should be a format recognised by the CSD Python API (e.g. cif, res, mol2, gcd or ind). The default output of the tool consists of figures (as .png files) of the packing-similarity dendrogram and a heat map of the packing similarity between the structures. The file names are prefixed with the stem of the input file ( i.e. if roy.gcd is the input, then the similarity dendrogram will be called roy_packing_similarity_tree.png).

If a large set of structures are inputted, the top N structures can be selected using the -nm options.

Using the -o option will result in overlays being saved for each comparison (as .mol2 files). The matrix of similarities is also saved as a raw numpy matrix ( i.e. as roy_similarity_matrix.txt), and can be read back in (with -m) to skip the packing-similarity analysis, if, for example, a different clustering algorithm is desired or part of the script has been changed.

The --allow_molecular_differences option can be used when comparing crystal structures of closely related moleculese.g. salts and free forms.

The -s option will strip all terminal atoms and carbon atom chains up to hetero atoms (e.g. the methyl of a methoxy will be removed), which may be useful for identifying more coarse- grained similarity that ignores small changes in the periphery of the molecule. A new cif is created containing the stripped molecule and analysis is performed using this file. Note that disordered experimental structures present in the input database will cause problems when this option is selected and should be removed from the file.

The remaining options control the packing-similarity settings, such as number of molecules, thresholds etc. Matches of only one molecule between structures may not necessarily

correspond to good agreement between two conformations (as the distance and angle thresholds can be more coarse-grained than an RMSD tolerance). The -ct option can therefore be used to specific an RMSD tolerance for whether two conformations are matched and in, turn, whether two clusters that have only one molecule in common should be merged at that level.

An example dendrogram (produced using CBZdataset.ind and related files; see supporting material of Cryst. Growth Des., 2009, 9, 1869 – 1888 ) is shown below in Figure 1. To produce this dendrogram, the script compares all 50 structures in the input to each other. For a default cluster size of 15, the match can range from 0, which indicates a completely different conformation and no similarity in packing, to 15, where the two structures are isostructural.

In comparing the structures different linkage criteria can be used. The default behaviour is to link structures based on the smallest difference between them (referred to as single- linkage clustering). Structures are initially grouped together based on their best matches, such that if A matches B with 15/15 and C with 14/15 and B matches C with 15/15, then A, B and C are all grouped together at level 15. Groups are then merged with the group which they share the best overlay, based on a single structure in each group. This means that if D has an 8/15 molecule match with B, then the group A, B, C will join with D at level 8, even if A and C had smaller similarities with D.

At the highest level ( e.g. 15 in the example below) isostructural crystal structures are listed together ( e.g. “CBZDBF18, CBZDBF19”). At lower levels, where groups or structures share a similarity then the parent groups are draw one level higher (with blue dots) and then join at the level they match. For example, structures 37 and 4 at the bottom of the plot have 5 / molecules in common, so both are drawn at level 6 and then merge at level 5. This group of 37 and 4 then shares a similarity with 32, 09 and 30, as well as all other structures apart from 31 at level 3, so all of them merge.

Alternative clustering modes can be invoked by using --clustering_type, which takes one of three arguments: single, complete and average. Single linkage clustering is the default discussed above. Complete clustering will instead only join two clusters at the level where all the structures in the two clusters match, which is equivalent to joining clusters at the worst or lowest level connecting them. Figure 2 shows the same database as Figure 1 using this approach, which gives more definite clustering but will also hide some similarities due to the hierarchical nature of the clustering.

If there are dangling nodes without an associated identifier, it is likely that there multiple structures given the same name in the supplied file (searched through multiple databases say)

Figure 1 : Example dendrogram based on a database of carbamazepine solid forms using the default single-linkage clustering.

Figure 2 : Example dendrogram based on a database of carbamazepine solid forms using complete- linkage clustering.

The final option for clustering type is to merge clusters based on their average packing- similarity agreement. The result for carbamazepine is shown in Figure 3. With average linkage clustering, two cluster can join at non-integer values and therefore the plot may become more complex visually.

There are several alternative schemes for linking clusters ( e.g. weighting the average based on the cluster sizes) that could also be implemented.

Figure 3 : Example dendrogram based on a database of carbamazepine solid forms linking clusters based on their average packing similarity.

The tool considers only the heaviest component in each crystal structure to enable comparison of multi-component forms with pure forms. Results for Z ’ > 1 systems may not fully reflect differences between structures as the best match is retained by default in the packing-similarity analysis.

Anthony Reilly(2016)

For feedback or to report any issues please contact support@ccdc.cam.ac.uk