Comparative Analysis of Chemical Descriptors by Machine Learning Reveals Atomistic Insights into Solute-Lipid...

Area of Science:

• Drug formulation and development
• Computational chemistry and cheminformatics
• Physical chemistry of lipids

Background:

• Drug solubility in lipid excipients remains a complex challenge in pharmaceutical development.
• Accurate prediction of solubility is crucial for rational drug design and formulation.
• Advancements in understanding molecular structure-property relationships are needed.

Purpose of the Study:

• To investigate novel molecular descriptor sets for predicting drug solubility in medium-chain triglycerides (MCTs).
• To develop and compare quantitative structure-property relationship (QSPR) models using machine learning.
• To identify key molecular determinants influencing drug solubility in lipidic environments.

Main Methods:

• Collected and analyzed an extended dataset of 182 experimental drug solubility values in MCTs.
• Evaluated four classes of molecular descriptors: 2D/3D descriptors, Abraham solvation parameters, ECFPs, and SOAP descriptors.
• Employed regularized regression algorithms and preprocessing schemes to build and validate QSPR models.

Main Results:

• The smooth overlap of atomic position (SOAP) descriptor yielded the most accurate and interpretable QSPR model.
• Atom-centered SOAP descriptors allowed for atomic-level contribution analysis of molecular motifs influencing solubility.
• High predictive accuracy (RMSE = 0.50) was achieved for models using 2D/3D, SOAP, and Abraham solvation descriptors on a test set.
• Models based on extended connectivity fingerprints (ECFPs) showed inferior predictive performance.
• Uncertainty estimations were incorporated to define model applicability domains.

Conclusions:

• The SOAP descriptor provides a powerful tool for predicting drug solubility in MCTs with high accuracy.
• This in silico approach enhances computational formulation development and rational drug design.
• The findings facilitate the prediction of drug loading in lipidic excipients, optimizing pharmaceutical formulations.