Integrated Platform for Quantification of Nanoparticle Transport Across Biological Barriers Using AI/ML Analysis

The capacity of engineered nanoparticles (NPs) to traverse biological barriers is a central determinant of efficacy in nanomedicine, targeted drug delivery and nanotoxicology, yet it remains one of the least reproducibly measured properties in the field. Established characterisation methods — Transwell® permeability assays, inductively coupled plasma atomic emission spectroscopy (ICP-AES), fluorescence microscopy and flow cytometry — each suffer from low throughput, marked inter-laboratory variability, or a dependence on particle labelling that perturbs the physicochemical properties governing transport. We describe and validate a fully integrated platform that couples a commercially available cross-flow microfluidic chamber bearing sequential porous-membrane cell barriers with label-free brightfield microscopy and a five-stage artificial-intelligence / machine-learning (AI/ML) pipeline. Intracellular NP accumulation, principally within lysosomes, produces characteristic organelle darkening that is detected without labelling, segmented by a residual-attention U-Net (ResAt-UNet; IoU = 0.85, precision = 93.2%, recall = 86.9%) and converted into quantitative transport-efficiency (TE) and barrier-integrity metrics. Across a factorial matrix of two surface chemistries, five core sizes (15–150 nm), four concentrations (10–500 µg mL⁻¹), two human barriers (HUVEC and hCMEC/D3) and two magnetic-field states (0 T, 1 T), PLGA-coated 15 nm SPIONs at 100 µg mL⁻¹ achieved the highest TE — 10.8 ± 1.5% (HUVEC) and 3.4 ± 0.6% (hCMEC/D3) at 0 T, rising to 13.2 ± 1.6% and 4.1 ± 0.7% under 1 T magnetic guidance (a realistic +22% relative gain). A dynamic three-zone quality-control (QC) architecture, gating every transport datum on matched transendothelial electrical resistance (TEER) and viability. Multi-factor ANOVA identified barrier identity, core size and coating as the dominant determinants of TE (all p < 0.001; partial η² = 0.43, 0.41 and 0.32, respectively). Gradient-boosted modelling with SHapley Additive exPlanations (SHAP) reproduced this ranking and, critically, showed that QC filtering raised cross-validated R² from 0.74 to 0.85 — establishing that the QC architecture materially improves predictive performance rather than cosmetically tidying the data.