Abstract
High-aspect-ratio nanostructures enable nuclear delivery through transient nuclear envelope (NE) disruption, but sporadic nucleoporation limits efficiency. This study presents a label-free, machine-learning approach detecting nucleoporation from morphological changes. U2OS cells cultured on silicon nanopillars undergo NE disruption detected via Ku-80 mislocalization. A custom longest-line algorithm quantifies Ku-80 versus DAPI intensity profiles outside the nucleus, establishing ground truth for 714 cells (451 intact, 263 porated). An orientation-invariant variational autoencoder, pre-trained on publicly available mouse embryonic fibroblast images, generates 32-dimensional embeddings for cell and nuclear shapes from binary masks. These embeddings, combined with four morphological descriptors (cell area, nucleus-to-cell area ratio, centroid distance, axis alignment) to reconstruct actual cell-nucleus geometry, train a support vector machine (SVM). Training comprised 462 cells from five experimental chips; testing used 258 cells from two separate chips. The model achieves 87.0% area under receiver operating characteristic curve and 82.9% test accuracy. SHAP analysis reveals nucleus-to-cell area ratio as the strongest predictor, with specific nuclear features (surface smoothness, localized protrusions, radial bulging) and cell features (boundary complexity, elongation patterns, polarization) critically influencing nucleoporation probability. This demonstrates a simple AI workflow for investigating whether cell/nuclear shapes correlate with or predict phenotypes/events, enabling high-throughput, non-invasive monitoring with relatively small datasets.