How does it work?

[catastropiyush]

This works very nicely!
But I wanted to know how this works?

using a privileged set of chembl-mined reduced graphs and fragments, this script generates random molecules using string operations.
 the output is (highly non-canonical) rdkit parsable SMILES strings

How did you train model1000.pkl.gz ?

Thank you for your time,
Piyush

[dehaenw]

Thank you for the interest. The way this works is quite simple: A smarts pattern is applied on every chembl molecule dividing it into "Rings", "Subsitituents" and "Linkers". Basically, it will take the molecule and break every non-ring bond that has at least one participating ring atom (somewhat similarly to what one would do when decomposing molecules into scaffolds). In this case, the top1000 occurring fragments in chembl get retained, and these are used as the likelihood for fragment occurrences, and used to craft randomly recombined molecules. I think when that file gets unpickled, you'll see it will have something like a dictionary with fragments and occurrences in there.

I'll have a look if i can find the code I used to generate these and drop it here. Feel free to ask any further questions if anything is unclear.

Also, in my other repo, a conceptually very similar approach is taken to breaking down ChEMBL molecules. This is then also used for generating random molecules (a little bit more slowly, but with nicer structures). The decomposition code is here https://github.com/dehaenw/lacan/blob/main/lacan/decompose.py and a generation example is here: https://github.com/dehaenw/lacan/blob/main/lacan/example_notebooks/generate_molecules.ipynb

[catastropiyush]

Thank you for answering my question. I'll try to folllow the steps and do it but would appreciate if you can find the code.
Will have a look at lacan too.

[dehaenw]

dear @catastropiyush i have looked into my archives and found the exact notebook I used to generate this pkl gz file. I added it to the repo unchanged, you can find it here. It's a little bit messy but I hope the main logic is clear. I will try to reconstruct the thought process on the bottom of this message.

https://github.com/dehaenw/tinygenerator/blob/main/notebooks/fragment_dataset.ipynb

the fragmentation is performed on the guacamol list of a million or so SMILES. Those files can be found here https://figshare.com/projects/GuacaMol/56639

Indeed, the whole thing is based on fragmenting molecules using the SMARTS pattern [*]-!@[R]. Going through the notebook, I see i used the following approach: when i deconstruct a molecule by cutting the bonds that match that SMARTS pattern, the molecule gets decomposed into a certain amount of fragments. At the same time, I also retain a reduced graph showing how the fragments are connected (in this graph every vertex is fragment and every edge is a direct connection between fragments), retaining only the degree. For example, imagine 1,2,3 Bromo Chloro Fluoro benzene. In this case the molecule would get degraded into four fragments: Br*, Cl*, F* and c1ccc(*)c(*)c1*. The reduced graph would be *(*)(*)*, so one vertex with degree 3 corresponding to the benzene ring and three vertices with degree one corresponding to the halides. The most popular fragments and reduced graphs get retained and stored in the pkl gz file. When generating a molecule, a random reduced graph (also mined from the above SMILES dataset) gets selected and populated with random degree matched fragments. Due to some clever SMILES trick, this can be done without have to instantiate an RDKit Mol but using string operations on the fragments SMILES. This is why this method is very fast.

[catastropiyush]

Thank you for the detailed reply and sending the code as well. I'll go through it.
Learning cheminformatics now so this helps a lot.

Thank you

[steto123]

Hello,
many thanks for your nice work. But can you short explain what CCONN,NSSD,NSSDW,CGC are?

And it is possible to generate moleculs only to a given molecular formula like C8H7BrO?

(For a short explanation: I work on the field of NMR Spectroscopy and one of my interests is to use my own calculation system to find the best structure using GA like algorythm, similar to J. Meiler (J. Chem. Inf. Comput. Sci. 2001, 41, 1535-1546) )

[dehaenw]

@steto123 because this tool is generating molecules unconditionally, it is not optimal for the molecular formula constrained problem. For this problem I recommend https://jcheminf.biomedcentral.com/articles/10.1186/s13321-022-00604-9 which directly generates molecular graphs adhering to some given molecular formula.

Regarding the field names, sorry for that, this was for internal use so they are not so sensibly named. If I recall correctly, these are the meanings:

nssd: N-substructure dictionary: a dictionary where N, the degree of the fragment is the key and the item is the smiles of the fragment

nssdw: N-substructure dictionary weights: occurrence counts of these fragments in the reference database (some version of ChEMBL)

cg: common graphs (these are the reduced graph referred to above in this issue), sorted by occurrence

cgc: common graph counts: again occurrence counts of the common graphs

cconn: common graph connectivity: this is an array with the degree the individual fragments filling up the reduced graph should have. For example, for *(*)(*)* that would be [3,1,1,1].

[steto123]

Many Thanks for your fast answer. I think this is verry helpful for me.

sincerly

St. Thomas