Computational Intelligence Modeling of the Macromolecules Release from PLGA Microspheres-Focus on Feature Selection

Area of Science:

• Biomaterials Science
• Computational Chemistry
• Machine Learning

Background:

• Poly-lactide-co-glycolide (PLGA) microspheres are crucial for controlled drug delivery.
• Predicting drug release profiles from PLGA is complex due to numerous influencing factors.
• Selecting relevant PLGA properties is a significant machine learning challenge.

Purpose of the Study:

• To formulate critical attribute selection for PLGA as a multiobjective optimization problem.
• To minimize prediction error for dissolution profiles while reducing feature set size.
• To compare bio-inspired optimization algorithms with LASSO for feature selection.

Main Methods:

• Utilized multiobjective optimization to balance prediction accuracy and model simplicity.
• Applied Antlion Optimization (ALO), binary ALO, Grey Wolf Optimization (GWO), and Social Spider Optimization (SSO).
• Evaluated feature selection using diverse predictive models including random forests, ANNs, and fuzzy logic systems.

Main Results:

• Achieved a Normalized Root Mean Square Error (NRMSE) of 15.97%, comparable to existing methods (15.4%).
• Successfully reduced the number of selected input features to nine, compared to eleven in previous studies.
• Demonstrated the effectiveness of bio-inspired algorithms in identifying critical PLGA attributes.

Conclusions:

• Bio-inspired optimization algorithms provide an effective approach for PLGA feature selection.
• The proposed method enhances model simplicity without compromising predictive accuracy for drug release profiles.
• This work contributes to more efficient and accurate modeling of PLGA-based drug delivery systems.