I rebuilt a set of models (LDA, recursive prtitioning, random forest).
Individually none are prticularly good, so the final results are
derived from an ensemble model.

I considered experiements that had a "reactants insoluble"
classification as NO. The ensemble prediction on the 65 compound
training set is

     ens.pred
depv  No Yes
  No  43   2
  Yes  5  15

When I predict the new set of molecules, I get a number of YES's but
I'll only report those that were predicted to be soluble by 2 or more
models (out of 3).

Molecules: 6,14,19,72,83,87

smiles id
CC(C)
(C)NC(=O)C(c1ccc2ccccc2c1)N(C)C(=O)CCCC3=CC=C4C=Cc5cccc6c5C4C3C=C6  6
Cc1ccc(cc1)C(C(=O)NC(C)(C)C)N(C)C(=O)CCCC2=CC=C3C=Cc4cccc5c4C3C2C=C5
14
CCCCNC(=O)C(c1cc(cc(c1)OC)OC)N(Cc2ccccc2)C(=O)C#Cc3ccccc3 19
Cc1ccc(cc1)S(=O)(=O)CNC(=O)C(c2ccc(cc2)C(F)(F)F)N(C)C(=O)c3c(cccc3O)O
72
CCCCNC(=O)C(c1ccccc1)N(c2ccccc2)C(=O)CCCC3=CC=C4C=Cc5cccc6c5C4C3C=C6
83
Cc1ccc(c(c1)C)C(C(=O)NC(C)
(C)C)N(C)C(=O)CCCC2=CC=C3C=Cc4cccc5c4C3C2C=C5 87