Code accompanying the paper "On the value of using 3D-shape and electrostatic similarities in deep generative methods". The paper can be found at https://doi.org/10.33774/chemrxiv-2021-sqvv9

To run the code on the default dataset use the command:

python deepFMPO.py -f ./Data/molecules.smi -l ./Data/lead.smi -o results.sdf

To use other sets of molecules (one of molecules that will be optimized and one of molecules that will be fragmented), use the command:

python deepFMPO.py -f your_fragment.smi -l your_lead.smi -o your_results.sdf

In both cases several parameters can be set in the global_parameters.py file. The most important parameter is the type of charges to be used: gasteiger, mmff or psi4. If psi4 charges are used, methodPsi4 and basisPsi4 must also be set. Please note that using psi4 charges the calculation requires a lot of time, so this option should be used only with small datasets. In The same file the target values can also be setted.

To save the generated fragments in an sdf file use the script decoding_to_sdf.py, with the command:

python decoding_to_sdf.py

The fragments are reported in the same way in which they are present in the similarity tree.

The program is written in Python 3.7 using the following Python libraries:

• rdkit
• scipy
• resp
• psi4
• dask
• ESP-Sim
• numpy
• sklearn
• keras
• pandas
• bisect
• Levenshtein
• A backend to keras, such as theano, tensorflow or CNTK

Please note that the ESP-sim can be obtained from the "Comparison of electrostatic potential and shape" GitHub repo: https://github.com/hesther/espsim