This is the algorithmic ECFP inversion approach developed in our lab (https://github.com/lich-uct). It is fully refactored from the previous version and has superior performance now. If you would like to use the old version anyway, you can find it here: https://github.com/dehaenw/ECFPinvert/tree/da6301b511d076b341da67d50bb9edd7691307dc
ECFP is a type of circular fingerprint used to encode the presence of substructures in molecules. In this prototype implementation we use RDKit's MorganFingerprint as our ECFP. By a fragment per fragment build up of the molecule and checking it at every step, a good degree of reconstruction is possible. An alternative NN based method for inversion was developed earlier and can be found here: https://github.com/bayer-science-for-a-better-life/neuraldecipher Their paper also contains some good background on why it is interesting to show some ECFP can be inverted.
make a conda env or venv with rdkit=2022.09 and numpy=1.23.4. yes, it is important to have these exact versions (for now).
here is a minimally working code block to invert the ECFP4(2048) of strychnine, a rather complex molecule:
from rdkit import Chem
from search import ECFPInvert
import utils
utils.initialize_atomtypes("CHEMBL")
utils.set_fp_settings(2,2048)
strychnine = Chem.MolFromSmiles("C1CN2CC3=CCO[C@H]4CC(=O)N5[C@H]6[C@H]4[C@H]3C[C@H]2[C@@]61C7=CC=CC=C75")
strychnine_fp = utils.get_fp(strychnine)
inv = ECFPInvert()
s, info = inv.run_search(strychnine_fp)
if s:
print(f'inverted strychnine in {info["time"]} seconds')
else:
print("failure")if that worked it should print something like inverted strychnine in 9.421 seconds
I will add example notebooks in a few days. Stay tuned!
Exact reconstruct only. Can't be used to reconstruct non existent fingerprints that do not correspond to a real structure. This version had ca 97% success inverting ChEMBL like ECFP6(4096).