Multistep Machine Learning Pipeline For Polymeric Nanoparticle Design

Integrating machine learning (ML) into nanotechnology represents a promising strategy for rational design and accelerated development of drug delivery systems. However, studies in this field are scarce and face methodological and interpretative problems. This study presents a modular ML pipeline for the predictive modeling of nanoparticles produced via nanoprecipitation using isoniazid as a model drug. The workflow was structured into three sequential steps: (1) binary classification to predict nanoparticle formation, (2) multiclass classification to estimate size ranges, and (3) regression to refine size prediction. Several algorithms were used, including Extreme Gradient Boosting, Random Forest, Artificial Neural Networks (ANN), Generalized Linear Models, and Naive Bayes. A total of 90 formulations were evaluated over three iterative experimental rounds. In each cycle, models were retrained with new data and used to simulate virtual formulations, thereby guiding the selection of experiments to reduce data imbalance and improve prediction accuracy. The ANN algorithm consistently outperformed other models in all steps, achieving an R² > 0.9 in both classification and regression tasks. Classification outputs were used as constraints in the regression phase to improve robustness. The final pipeline demonstrated high predictive performance across a broad size range (75-768 nm), with maximum absolute errors below 40 nm. Validation with new formulations confirmed the model's reliability and generalization capacity. This approach may significantly reduce experimental workload while providing a scalable framework. Overall, the proposed ML-guided strategy supports data-driven decision-making in nanopharmaceutical research, enabling systematic formulation development aligned with Quality-by-Design principles.