Abstract
Background: Dissolution is commonly modeled forward from formulation inputs. We propose an inverse-learning framework where machine learning (ML) infers mechanistically interpretable release-kinetic parameters from dissolution profiles and relates them to formulation and pharmacokinetic (PK) descriptors within Quality by Design (QbD). Methods: An in silico dataset of release profiles was generated using six kinetic models (zero-order, first-order, Higuchi, Hixson–Crowell, Korsmeyer–Peppas and Weibull). Each profile was fitted to all candidate models to select the best mechanism and estimate its parameters. ML regressors (tree-based ensembles and artificial neural networks) were trained using full time series and engineered summaries (e.g., lag time, fractional release at 2/6/12 h and AUC0–24); PK features (e.g., Tmax, Cmax, CL, Vc) were incorporated when applicable. Performance was evaluated by cross-validation and external tests using R² and error metrics, and interpretability was assessed with SHAP. Results: Gradient boosting performed best for simpler kinetics (zero/first/Higuchi/Hixson–Crowell; R²≈0.99; AFE/AAFE≈1.00–1.01). For complex kinetics, neural networks were best for Korsmeyer–Peppas (R²=0.89), and boosting remained strong for Weibull. SHAP highlighted AUC0–24, Tmax and Cmax as dominant predictors. External experimental profiles showed good agreement by visual comparison. Conclusion: Inverse learning can recover mechanistically meaningful release parameters from dissolution data and connect them to formulation and PK descriptors, supporting faster and more transparent modified-release design under QbD.