A novel optimization of API release using hybrid models

Abstract

Abstract Polysaccharide-based drug delivery systems have to be closely watched in terms of drug release in the body over time if they want to be able to achieve maximum therapeutic effectiveness and maintain controlled bioavailability. The work presents a data-driven framework that employs Raman spectroscopy and uncertainty-aware machine learning models to anticipate drug release patterns in different biological media. A dataset of 155 experimental samples was dealt with, which also included more than 1,500 Raman spectral features, polysaccharide composition, release time, and environmental medium (Control, Patient, Rat, and Dog). Principal Component Analysis (PCA) was implemented for dimensionality reduction, and modeling was done with the help of Gradient Boosting Regression (GBR) and Lasso Regression (Lasso) supported by newly developed bio-inspired optimization algorithms Attack-Leave Optimizer (ALO), Self-Adaptive Bonobo Optimizer (SABO), and Black-Winged Kite Algorithm (BWKA). ‍‌‍‍‌Besides k-fold cross-validation, Friedman statistical testing, prediction interval bootstrapping, and Pareto front analysis were used to further assess model robustness and uncertainty. Of the hybrid models, GBBW (GBR-BWKA) was able to attain the highest predictive accuracy with R 2 = 0.931 and RMSE = 0.076 on the test set, thus outperforming baseline models and demonstrating the value of bio-inspired hyperparameter optimization within an uncertainty-aware computational framework. The primary contribution of this work is methodological: the integration of prediction interval estimation, multi-environment modeling, and Pareto-based model selection into a single reproducible pipeline applied to a publicly available polysaccharide Raman release dataset. No new experimental data were generated; the framework is offered as a computational tool to complement existing experimental approaches in controlled-release formulation research.