Optimizing Pharmacokinetic Property Prediction Based on Integrated Datasets and a Deep Learning Approach.

Area of Science:

• Computational chemistry
• Pharmacokinetics
• Drug discovery

Background:

• Oral bioavailability (OBA)-related properties (solubility, lipophilicity, permeability) are vital in drug discovery.
• Experimental measurement of these properties is expensive and time-consuming.
• Computational prediction models exist but face limitations in dataset capacity and algorithm adaptability.

Purpose of the Study:

• To address limitations in predicting OBA-related molecular properties by optimizing both datasets and algorithms.
• To develop and validate a novel deep learning model for accurate prediction of solubility, lipophilicity, and membrane permeability.

Main Methods:

• Construction of benchmark datasets for aqueous solubility (log S), lipophilicity (log D), and Caco-2 cell membrane permeability (log Papp).
• Development of a multi-embedding-based synthetic network (MESN) using a deep learning algorithm.
• Synthesis of multiple molecular embedding types within the MESN model.
• Validation against state-of-the-art methods and independent datasets.

Main Results:

• MESN demonstrated superior performance in predicting aqueous solubility, lipophilicity, and membrane permeability compared to existing methods.
• Dimension reduction and atomic feature similarity analyses confirmed the clustering and diversity of MESN-extracted molecular embeddings.
• Control studies using other algorithms and independent datasets corroborated MESN's predictive power.

Conclusions:

• The developed MESN model offers significant improvements for predicting key OBA-related molecular properties.
• The model's effectiveness on benchmark datasets highlights its potential utility in accelerating drug discovery pipelines.
• Optimization of both data and algorithms is crucial for enhancing the applicability of molecular property prediction models.