LogPpred: An AI-Based Predictive Model for Accurate Estimation of Molecular LogP

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Pharmacokinetics

Background:

• Lipophilicity, quantified by the n-octanol/water partition coefficient (LogP), is crucial for small molecule pharmacokinetics.
• Accurate LogP estimation is vital in early drug discovery for molecular design and ADMET property optimization.

Purpose of the Study:

• To develop LogPpred, an artificial intelligence (AI)-based predictor for molecular lipophilicity.
• To provide a reliable tool for early assessment of lipophilicity in drug discovery.

Main Methods:

• Systematic evaluation of machine learning algorithms and molecular representations using a dataset of 13,536 experimentally determined LogP values.
• Development of a Gaussian Process Regression model utilizing RDKit molecular descriptors.
• Validation using independent internal, external (OECD-derived), and experimental datasets.

Main Results:

• The LogPpred model achieved a mean absolute error (MAE) of 0.34 on an internal test set and 0.84 on an external validation set.
• Experimental validation yielded an MAE of 0.66, confirming predictive reliability.
• LogPpred demonstrated superior performance compared to existing LogP prediction tools, with improved MAE for in-domain predictions (0.28 internal, 0.62 external).

Conclusions:

• LogPpred is a robust and accurate AI tool for predicting molecular lipophilicity.
• The model offers reliable early assessment of LogP, supporting medicinal chemistry and drug discovery efforts.
• Applicability domain analysis confirms the reliability and transparency of LogPpred predictions.