Few-Shot Early-Release Adaptation Enables Generalizable Prediction of PLGA Drug Release
8 affiliations
- 1Department of Biomedical Engineering
- 2Pennsylvania State University
- 3Harold & Inge Marcus Department of Industrial & Manufacturing Engineering
- 4Department of Neurosurgery
- 5Penn State Health Milton S. Hershey Medical Center
- 6Penn State College of Medicine, Pennsylvania State University
- 7Materials Research Institute (MRI
- 8State University
Abstract
Reliable prediction of long-term drug release remains challenging for formulations developed under experimental conditions that differ from those represented in literature-derived training data. Standard formulation descriptors capture important physicochemical and formulation characteristics, but they do not fully represent study-specific factors that shape the realized release trajectory. Here, we develop a machine-learning-based few-shot adaptation strategy that combines formulation descriptors with sparse early release measurements to forecast the remaining release trajectory under previously unseen experimental settings. Using a public dataset comprising 321 formulations, 4,913 release measurements, and 113 source studies, we evaluate the framework using strict study-grouped validation. Zero-shot prediction shows limited generalization to new experimental domains, whereas incorporating only a few early observations produces a marked and reproducible improvement in predictive performance. Across repeated held-out study splits, two early measurements increase mean pooled from −0.084 to 0.333 and reduce formulation-level MAE by approximately 30%, with further improvements as additional observations are incorporated. Although absolute predictive accuracy remains moderate, the improvement is consistent across repeated splits and demonstrates that limited target-specific information can substantially reduce the generalization gap. These findings show that sparse early release measurements can serve as an effective calibration signal for machine-learning models, providing a practical strategy for improving long-term drug-release forecasting under previously unrepresented experimental conditions.
Related
- Publication-block calibration reveals wide uncertainty in PLGA release endpoint prediction
- Peripheral Charge Gating Enables Redox-Controlled Capture and Complete Release of Phosphate by a Europium(III) Complex in Water
- Machine Learning Prediction of Subsequent Drug Release Trajectories using Early Release Information in PLGA Microparticles
- β-Sitosterol–Naringenin Co-Loaded Nanophytosomes: Encapsulation, Compound-Specific Release Kinetics, Machine Learning-Assisted Release Prediction, and Cell-Dependent Anticancer Activity
- Wireless, spatially controlled in-situ methotrexate drug encapsulation and release from PEDOT films
BibTeX
@misc{liu2026fewshot,
title = {{Few-Shot Early-Release Adaptation Enables Generalizable Prediction of PLGA Drug Release}},
author = {Tengxiao Liu and Juchen Zhang and Alireza Mansouri and Hongtao Sun},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010043/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010043/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Few-Shot Early-Release Adaptation Enables Generalizable Prediction of PLGA Drug Release
AU - Tengxiao Liu
AU - Juchen Zhang
AU - Alireza Mansouri
AU - Hongtao Sun
PY - 2026
DA - 2026/10/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010043/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010043/v1
AB - Reliable prediction of long-term drug release remains challenging for formulations developed under experimental conditions that differ from those represented in literature-derived training data. Standard formulation descriptors capture important physicochemical and formulation characteristics, but they do not fully represent study-specific factors that shape the realized release trajectory. Here, we develop a machine-learning-based few-shot adaptation strategy that combines formulation descriptors with sparse early release measurements to forecast the remaining release trajectory under previously unseen experimental settings. Using a public dataset comprising 321 formulations, 4,913 release measurements, and 113 source studies, we evaluate the framework using strict study-grouped validation. Zero-shot prediction shows limited generalization to new experimental domains, whereas incorporating only a few early observations produces a marked and reproducible improvement in predictive performance. Across repeated held-out study splits, two early measurements increase mean pooled from −0.084 to 0.333 and reduce formulation-level MAE by approximately 30%, with further improvements as additional observations are incorporated. Although absolute predictive accuracy remains moderate, the improvement is consistent across repeated splits and demonstrates that limited target-specific information can substantially reduce the generalization gap. These findings show that sparse early release measurements can serve as an effective calibration signal for machine-learning models, providing a practical strategy for improving long-term drug-release forecasting under previously unrepresented experimental conditions.
ER -