Background/Objectives: Poly(lactic-co-glycolic acid) (PLGA) microparticles are widely used for sustained drug delivery, yet the release behavior reported in the literature remains difficult to predict across studies. It was hypothesized that this limitation reflects insufficient information content in commonly reported formulation variables rather than model inadequacy. Methods: A curated dataset of 321 PLGA microparticle formulations from 113 publications comprising 89 drugs and 4913 release observations was analyzed. Early time release was parameterized using Korsmeyer–Peppas descriptors (n, K), and burst release was quantified as the 24 h cumulative release. Machine learning models were evaluated using formulation-grouped cross-validation, applicability-domain analysis, and leave-one-study-out validation to assess cross-laboratory transportability. Results: Under formulation-grouped validation, predictability was limited (stacked ensemble: R2=0.156 for n, R2=0.169 for K, burst R2=0.100). Leave-one-study-out validation yielded negative pooled R2 values for all targets (−0.061, −0.040, and −0.180, respectively), indicating failure to generalize across laboratories. Applicability-domain filtering did not materially improve performance, supporting the interpretation that prediction is limited by missing or inconsistently reported variables rather than covariate extrapolation alone. Conclusions: These results reveal an information-limited regime in PLGA release prediction in which the literature covariates enable only weak formulation-level prediction under grouped validation and cannot support transferable models. Minimum reporting priorities are therefore proposed, including standardized characterization of polymer molecular weight, end-group chemistry, quantitative emulsification and solvent-removal parameters, and microstructural or porosity measurements, to enable reproducible formulation screening.
Sharma et al. (Mon,) studied this question.