In Vitro liposome release profile prediction using explainable machine learning approaches
Hamza Abu Owida, Sameer Ahmad Hasan, Areen Arabiat, Suhaila Abuowaida
Abstract
Formulation features and test settings influence liposomal in vitro release (IVR) profiles, yet it is challenging to examine these multivariable associations across varied literature data. We created an explainable computational workflow in this proof-of-concept study for classifying liposomal release phenotypes and identify formulation/assay features linked to slow and fast release.
Introduction
Liposome formulations have been shown to have varying levels of in vitro release (IVR) under different experimental conditions (formulation characteristics, e.g., lipid composition, drug concentration and amount, vesicle diameter and size distribution) and assay conditions (temperature, pH, medium composition and hydrodynamics). This dependence complicates quantitative comparison across studies and motivates predictive characterization tools that jointly
Methods
2.1. Study design and data sources
The Accelerated IVR public dataset and accompanying coding, which are housed in a GitHub repository with a CC-BY license, were utilized in all studies [26]. An SQLite relational database with linked tables for article provenance, formulation composition, lipid and excipient identities, physicochemical descriptors, IVR assay conditions, and digitized cumulative release profiles was initially assembled from 34 scholarly publications.
Results
3.1. Dataset composition and scope
169 Weibull-parameterized release profiles covering quick to sustained release behaviors over a standardized 0–168 h timeframe comprised the release-profile cohort. Three kinetic phenotypes (slow, n = 34; intermediate, n = 19; fast, n = 25) comprised the 78 labeled occurrences in the metadata-annotated cohort. The kinetic extremes were the focus of supervised modeling, which kept 59 records (slow, n = 34; fast, n = 25).
Discussion
Kinetic-model benchmarking using fâ showed that apparent agreement between a fitted model and a liposomal release profile depends strongly on the selected model form. In this dataset, Weibull provided the highest overall similarity, while first-order and Korsmeyer–Peppas generally outperformed zero-order and Hixson–Crowell. Together, these comparisons suggest that many profiles are better described by non-linear model forms rather than strictly linear release behavior, consistent with multi-mechanism transport in liposomal systems.
Conclusion
A workflow for assessing heterogeneous liposomal IVR data that is integrated, explainable, and proof-of-concept is presented in this study. The method made it possible to compare models using f2 similarity and downstream PCA-k-means phenotyping by standardizing release profiles on a common time grid and parameterizing them with Weibull descriptors. A supervised XGBoost classifier had the best cross-validated performance for differentiating between slow and fast extremes after the investigation revealed slow, intermediate, and fast release phenotypes.
Citation: Owida HA, Ahmad Hasan S, Arabiat A, Abuowaida S (2026) In Vitro liposome release profile prediction using explainable machine learning approaches. PLoS One 21(8): e0354293. https://doi.org/10.1371/journal.pone.0354293
Editor: Claudio J. Salomon, National University of Rosario, ARGENTINA
Received: March 31, 2026; Accepted: July 6, 2026; Published: August 11, 2026
Copyright: © 2026 Owida et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: danielyanes22. accelerated_IVR. GitHub repository. Available at: https://github.com/danielyanes22/accelerated_IVR. Accessed Feb 13, 2026.
Funding: The author(s) received no specific funding for this work.
Competing interests: the authors declare no conflict of interests.
