Validation Design Governs the Apparent Predictive Performance of Machine Learning for PLGA Drug Release
Abstract
Reliable generalization is essential for machine-learning models intended to predict drug-release behavior across heterogeneous experimental settings. However, multi-study experimental data often contain closely related formulations generated under shared materials, processing conditions, release protocols, and laboratory practices, creating hidden dependencies between training and test data. Such dependencies can make model performance appear stronger than its ability to generalize to genuinely new experimental domains. Here, we examine how validation topology governs the apparent predictability of poly(lactic-co-glycolic acid) (PLGA) microparticle drug release using a public dataset containing 321 formulations and 4,913 release observations collected across 113 studies. Using a common LightGBM prediction framework, we compare validation schemes ranging from random observation splitting to strict study-grouped holdout. Predictive performance decreases markedly as train–test independence becomes more stringent, with 𝑅! declining from 0.939 under random point splitting to 0.616–0.625 under formulation-level holdout and 0.372 when entire studies are withheld. Controlled overlap experiments further show that retaining samestudy sibling formulations in the training set systematically increases apparent performance, while nearest-neighbor analysis reveals lower prediction error for formulations lying closer to the training distribution, particularly when their nearest training counterpart originates from the same study. These results show that formulation-level independence alone does not ensure experimental-domain independence and demonstrate that apparent drug-release predictability is strongly conditioned by validation design. Study-aware validation therefore provides a more rigorous basis for assessing generalization of machine-learning models trained on heterogeneous multi-study experimental data.
Related
- Steric‐Pressure‐Driven Topological Reconstruction for High‐Performance Zintl Phase Thermoelectrics with Machine‐Learning Validation
- Group-aware validation sharply reduces the apparent accuracy of extraction-category classification in public cannabis laboratory data
- Washing order rather than ethanol concentration governs protein restructuring, porosity and emulsifying performance of defatted soybean meal particles
- Nitrogen content governs thermoelectric performance in TPU/SWCNT composites.
- Systematic evaluation of LLM-based zero-shot model construction in chemistry and materials science: predictive performance and response reliability
BibTeX
@misc{liu2026validation,
title = {{Validation Design Governs the Apparent Predictive Performance of Machine Learning for PLGA Drug Release}},
author = {Tengxiao Liu and Juchen Zhang and Hongtao Sun},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010439/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010439/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Validation Design Governs the Apparent Predictive Performance of Machine Learning for PLGA Drug Release
AU - Tengxiao Liu
AU - Juchen Zhang
AU - Hongtao Sun
PY - 2026
DA - 2026/10/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010439/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010439/v1
AB - Reliable generalization is essential for machine-learning models intended to predict drug-release behavior across heterogeneous experimental settings. However, multi-study experimental data often contain closely related formulations generated under shared materials, processing conditions, release protocols, and laboratory practices, creating hidden dependencies between training and test data. Such dependencies can make model performance appear stronger than its ability to generalize to genuinely new experimental domains. Here, we examine how validation topology governs the apparent predictability of poly(lactic-co-glycolic acid) (PLGA) microparticle drug release using a public dataset containing 321 formulations and 4,913 release observations collected across 113 studies. Using a common LightGBM prediction framework, we compare validation schemes ranging from random observation splitting to strict study-grouped holdout. Predictive performance decreases markedly as train–test independence becomes more stringent, with 𝑅! declining from 0.939 under random point splitting to 0.616–0.625 under formulation-level holdout and 0.372 when entire studies are withheld. Controlled overlap experiments further show that retaining samestudy sibling formulations in the training set systematically increases apparent performance, while nearest-neighbor analysis reveals lower prediction error for formulations lying closer to the training distribution, particularly when their nearest training counterpart originates from the same study. These results show that formulation-level independence alone does not ensure experimental-domain independence and demonstrate that apparent drug-release predictability is strongly conditioned by validation design. Study-aware validation therefore provides a more rigorous basis for assessing generalization of machine-learning models trained on heterogeneous multi-study experimental data.
ER -