Abstract
Precise computational prediction of the molecular lipophilicity, expressed as
the logarithm of partition coefficient (log P), is still one of the most important
but also difficult missions in the early stages of the drug discovery process.
This is because the property substantially determines a pharmacokinetic
profile (Absorption, Distribution, Metabolism, Excretion, and Toxicity)
(ADMET) for a potential drug candidate. Generating well-generalizing
Quantitative Structure Property Relationship (QSPR) tools for log P (and
other related properties) becomes very challenging as the fine but non-local
electronic and steric effects that determine log P can be difficult to model
within classical approaches, while fragmentary datasets combined with
relatively small data literature on these systems do not have high-quality data
suitable for deep learning. In this work, we tackle these issues by providing an
in-depth analysis of the AttentiveFP architecture (a specialized Graph Neural
Network for learning from molecular-graph representations) tailored at
maximizing learning effectiveness through a dual-stage attention mechanism.
We assumed that allowing the AttentiveFP model to learn atomic and
bond-specific variations of importance weights would outperform other QSPR
models with fixed Atom and Bond-importance. The AttentiveFP model was
trained and validated on the largest public benchmark log P dataset
(consisting of information from PubChem/ChEMBL), which, to our
knowledge, is also the largest dataset that has been studied. Our results
showed that the AttentiveFP model successfully reached a state-of-the-art
prediction performance: we obtained a Root Mean Square Error (RMSE) of
0.6447 on the test set, which was much better than the one achieved by a
well-performing descriptor-based method (RMSE = 0.655) and our model got
even better performance after parameter tuning with RMSE of 0.6225.
Importantly, examination of the learned attention weights yielded high-fidelity
visualizations validating that the model can recognize chemically relevant
substructures related to lipophilicity, viz. aromatic ring systems and large
halogenated chains to be crucial determinants. Our results validate
AttentiveFP as a strong, interpretable, and chemically aware method that can
achieve the same high prediction accuracy needed for successful virtual
screening and/or rational drug design to optimize the all-important log P
value.