In Vitro liposome release profile prediction using explainable machine learning approaches

Area of Science:

• Pharmaceutical Sciences
• Computational Chemistry
• Biotechnology

Background:

• Liposomal in vitro release (IVR) profiles are influenced by formulation and testing conditions, making analysis complex.
• Understanding these multivariable associations is crucial for effective liposomal drug development.

Purpose of the Study:

• To develop an explainable computational workflow for classifying liposomal release phenotypes.
• To identify formulation and assay features associated with slow and fast liposomal release.

Main Methods:

• Benchmarking kinetic models and simulating Weibull-parameterized release curves.
• Clustering release profiles using Principal Component Analysis (PCA) and k-means clustering.
• Training supervised classifiers (XGBoost) on formulation and IVR descriptors using a public dataset.

Main Results:

• Three distinct kinetic phenotypes (slow, moderate, fast) were identified, with PCA explaining 98.4% of variance.
• XGBoost showed promising performance in classifying extreme slow vs. fast release subgroups.
• Key influential features identified include media pH, drug loading, lipid transition temperature, and media temperature.

Conclusions:

• Explainable machine learning can support risk-based IVR technique development and hypothesis generation for liposomal formulations.
• Prospective validation on independent datasets is necessary before clinical or regulatory application.