Bigger molecules

[kirtov]

Hello!

Thank you for the implementation of MolGAN

How should I update your code to use with bigger molecules (~20-30 heavy atoms)?

It seems, that only passing another dataset (with bigger molecules) after 100 epochs model generated very bad SMILES like these:

*.*.*.*.*CC.C.C.C.C.C.C.C.C.C.C.C.CC.CCN.CCNC.O.O.O
*.*.*.*.*.*.*.*.*C(*)C.C.C.C.C.C.C.CC.CCN.CCO.CO.N.N
*.*.*.*.*.*.*.*.*.*=C(C)CC.C.C.C.C.C.C.CC.CCC.CCCCO
*.*.*.*.*.*C.*C*.C.C.C.C.C.C.C.C=C.CC.CC(C)F.CNC.N.O
*.*.*.*.*.*.*.*.*.*C.*CC.C.C.C.C.C.C.C.C.CCNCO.CN.O
*.*.*.*.*.*.*.*.*.*OCC.C.C.C.CCC.CCCC.CSC.N.N.O.O
*.*.*.*.*.*.*.*.*.*.*.*.*CCOC(C)C*.*CN=*.C.C.C.CC.O
*.*.*.*.*.*.*CC.C.C.C.C.C.C.C.C.C.C.C=CC.CC.CC(C)S.N.N
*.*.*.*.*.*.*.*.*C.*C.C.C.C.C.C.C*CC.CC.CC.COO.N.O
*.*.*.*.*.*.*.*C(*)(*)O.*CC(C)CC.C.C.C.CCC.CCCN.N.O
*.*.*.*.*.*.*.*.*.*.*.*.*CN.*O.C.C.C.C.CCC.CCCC.O.S
*.*.*.*.*.*.*.*.*.*C.*C.*CSCC(C)C(CC)C(C)*C(C)C.C.C
*.*.*.*.*.*.*CC(C)CCC.*N.C.C.C.C.C.C.C.CN.CO.F.O.O.O
*.*.*.*.*.*.*.*.*C.*C.*C.*C(C)C.C.C.C.C.C.C.CC.CCN.O
*.*.*.*.*.*.*CC.*NC.C.C.C.C.C.C.C.C.C.CC.N.N.NCC(O)O
*.*C.C.C.C.C.C.C.C.C.C.C.C.C.C.C.CC.CNCO.N.N*C(=O)O.O
*.*.*.*.*.**.*C.*CC=[SH]*.C.C.C.C.C.C.C.CC.CC.CCC.N.O
*.*.*.*.*.*.*.*C(C)(F)CC(C)CC.C.C.C.C.C.CC.CCC.N.N.N.O

[nicola-decao]

Hi,
unfortunately, it is a very hard question to answer. When using big molecules there are other challenges to face as the adjacency matrix that the model has to learn is much more complex (then difficult to learn). I would say you could try to play with the architecture/ layers but I do not have a final answer to that.