Machine-Learning-Guided Exploration of a Small Orthogonal PBAE Polymer Library for <i>In Vitro</i> siRNA Delivery Optimization

Pulmonary siRNA delivery is a promising therapeutic approach for future pandemics and many noninfectious lung diseases. Polymeric nanocarriers, especially poly(β-amino ester)s, are easily tunable and versatile delivery systems to protect RNA from degradation. To navigate this broad chemical space and improve control over synthesis parameters, an orthogonal polymer library of amphiphilic-spermine-based poly(β-aminoester)s was investigated for gene knockdown and toxicity. Subsequently, a nested leave-one-out cross-validation approach was chosen to screen different machine-learning (ML) models and capture useful information within the limited data set. Experimental characterization features enabled descriptive model interpretation using SHAP and feature ablation, identifying tentative feature–performance associations. For prospective testing, we trained a synthesis-parameter-only model to prioritize candidate synthesis conditions within the investigated design space. One ML-prioritized candidate, ML-2, was synthesized and evaluated experimentally. ML-2 showed improved in vitro knockdown and a more favorable toxicity profile compared with the previous lead polymer, providing an initial proof-of-concept for model-guided candidate prioritization. In vivo experiments, however, highlighted the lack of in vitro – in vivo correlation after model optimization for in vitro performance.