quick and dirty LogS approximation with xgboost+rdkit.
Aqueous solubility is usually expressed as LogS, the logarithm of aqueous solubility expressed in mol/l. Correctly estimating this value is not that easy, but a regression model on known molecules can give a somewhat decent first guess. The approach here uses XGB regressor with molecule features represented by rdkit 2D descriptors.
Putting this up here for those in a hurry who dont have time to drop this data in a regressor themselves. I did a grid search and provide a decently optimized model as is.
This is based on the data from Estimation of Aqueous Solubility of Chemical Compounds Using E-State Indices by Tetko and co-workers
Their numbers: r2 0.91, RMSE 0.62
Numbers of this model: r2 0.93, RMSE: 0.57 (using the provided splits)
So this regressor's performance is about the same as the published method. A previous version of this model used ECFP fingerprints which has r2 of about 0.86
sklearn pandas rdkit>2023.03 (needed for Descriptors.CalcMolDescriptors()) and xgboost
Obtain the input data using
wget https://vcclab.org/lab/alogps/logs.txt
(make sure to check license/downloading conditions)
to retrain just run
python train.py
import function mols2logS and put an array of mols in there
example:
from XGBlogS import mols2logS
from rdkit import Chem
mols = [Chem.MolFromSmiles(smi) for smi in ["C(CCCC)CC(C)CC","OCCc1ccccc1"]]
print(mols2logS(mols))outputs [-4.9830585 -1.0931162]
and that's all. enjoy!