Machine Learning Prediction of Subsequent Drug Release Trajectories using Early Release Information in PLGA Microparticles
Abstract
Accurate prediction of drug-release behavior from PLGA microparticles may reduce the experimental effort required for formulation characterization, yet conventional prediction approaches rely primarily on static molecular and formulation descriptors and do not incorporate experimentally observed release behavior as it becomes available during an ongoing experiment. Here, we develop a machine-learning framework that integrates molecular and formulation descriptors with early-release measurements to predict subsequent drug-release trajectories from PLGA microparticles. A dataset comprising 4,913 release observations from 321 formulations was analyzed using random forest regression with formulation-grouped cross-validation. Using a common set of 89 formulations, we systematically evaluated the predictive value of increasing amounts of early-release information. Prediction accuracy improved with additional early measurements, but the incremental benefit progressively diminished, with two and three measurements capturing 64.7% and 91.9% of the total predictive gain achieved with four measurements, respectively. The value of early information also varied substantially across formulations, with greater benefits generally observed for formulations that were more difficult to predict from static descriptors alone. Trajectory-level analysis further revealed that early observations constrained subsequent release trajectories to different extents across formulations. These findings demonstrate that experimentally observed early-release behavior provides formulation-specific predictive information beyond static descriptors and establish a data-driven framework for quantifying information sufficiency in drug-release characterization.
Related
- The Limited Contribution of Structural Information Relative to Language-Model Signal in Data-Limited Cryptic Pocket Prediction
- Biocomposite Microparticles Incorporating Ibuprofen and ZnO Nanoparticles for Controlled Drug Delivery and Antimicrobial Therapy
- Pickering‐Engineered Microparticles for Magnetically Guided Motion and Light‐Triggered Catalysis
- Foldability Is an Intrinsic Property of Protein Sequences: A Lattice Model Study with Information-Theoretic Approach and Unsupervised Machine Learning
- Folding and Function as Two Sides of One Coin: Statistical Mechanics, Information Theory, and Deep Learning of Binding, Catalysis, and Allostery in an Exactly Enumerable Lattice Protein
BibTeX
@misc{liu2026machine,
title = {{Machine Learning Prediction of Subsequent Drug Release Trajectories using Early Release Information in PLGA Microparticles}},
author = {Tengxiao Liu and Juchen Zhang and Hongtao Sun},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010440/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010440/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Machine Learning Prediction of Subsequent Drug Release Trajectories using Early Release Information in PLGA Microparticles
AU - Tengxiao Liu
AU - Juchen Zhang
AU - Hongtao Sun
PY - 2026
DA - 2026/10/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010440/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010440/v1
AB - Accurate prediction of drug-release behavior from PLGA microparticles may reduce the experimental effort required for formulation characterization, yet conventional prediction approaches rely primarily on static molecular and formulation descriptors and do not incorporate experimentally observed release behavior as it becomes available during an ongoing experiment. Here, we develop a machine-learning framework that integrates molecular and formulation descriptors with early-release measurements to predict subsequent drug-release trajectories from PLGA microparticles. A dataset comprising 4,913 release observations from 321 formulations was analyzed using random forest regression with formulation-grouped cross-validation. Using a common set of 89 formulations, we systematically evaluated the predictive value of increasing amounts of early-release information. Prediction accuracy improved with additional early measurements, but the incremental benefit progressively diminished, with two and three measurements capturing 64.7% and 91.9% of the total predictive gain achieved with four measurements, respectively. The value of early information also varied substantially across formulations, with greater benefits generally observed for formulations that were more difficult to predict from static descriptors alone. Trajectory-level analysis further revealed that early observations constrained subsequent release trajectories to different extents across formulations. These findings demonstrate that experimentally observed early-release behavior provides formulation-specific predictive information beyond static descriptors and establish a data-driven framework for quantifying information sufficiency in drug-release characterization.
ER -