Nanoparticle Protein Corona Prediction

What this research found

Can a nanoparticle's physicochemical properties alone predict which blood proteins will coat it — the so-called protein corona? Models were trained on 144 observations covering 25 serum proteins across six formulations, using a dataset assembled synthetically from published corona compositions because raw proteomics data could not be obtained. The answer came back negative: tested on held-out nanoparticle types, random forest scored an R² of -0.158 and gradient boosting -0.679, both worse than a dummy model that simply predicts the mean (-0.013).

• Prediction across formulations failed outright. Under leave-one-group-out validation over the six nanoparticle types, random forest reached R² = -0.158 and gradient boosting -0.679, against -0.0134 for a mean-predicting baseline, meaning both models were worse than useless for unseen particle types.
• Protein properties outrank particle properties in the random forest's feature importances: protein length (0.235), protein molecular weight (0.170), ligand molecular weight (0.129) and protein hydropathy (0.115) all beat zeta potential (0.079) and every surface-chemistry indicator.
• No individual physicochemical variable tracked protein abundance. The strongest correlation was ligand hydrophobicity at r = 0.0906 (p = 0.28), and a Kruskal-Wallis test across five surface chemistries found no effect on abundance (H = 4.7458, p = 0.31).
• The intuitive charge-matching mechanism did not show up: nanoparticle zeta potential and protein isoelectric point correlated at r = 0.0043 (p = 0.96) across all pairs, and r = -0.0087 among proteins of above-median abundance.
• Adding 3D structure helped only marginally. For the nine most abundant proteins whose predicted structures were retrieved, relative exposed hydrophobic surface area correlated with abundance at r = 0.267 (p = 0.49) — better than the sequence-based hydropathy score at r = -0.1015, but neither is significant.
• The work still delivered a tool: a calculator that takes size, zeta potential, surface chemistry and core material and returns a predicted corona fingerprint plus a dysopsonin-to-opsonin stealth ratio. Its own output flags the weakness — one test case rated a cationic amine-coated particle as low immune risk, which the report notes contradicts established literature.

How it was done

Because raw corona proteomics could not be pulled from public repositories, a 144-row dataset was constructed synthetically from published compositions, covering 25 serum proteins across six formulations with gold, PLGA, polystyrene and silica cores, sizes from 20 to 100 nm, and zeta potentials from -42.3 to +32.7 mV. Thirteen predictors combined protein descriptors — length, molecular weight, isoelectric point and hydropathy — with particle size and charge plus RDKit-computed descriptors for each surface ligand. Dummy, random forest and gradient boosting regressors were then compared under leave-one-group-out cross-validation grouped by nanoparticle type, which deliberately tests generalisation to formulations the model has never seen. AlphaFold structures were retrieved for the nine most abundant proteins and solvent-accessible surface area computed with the Shrake-Rupley algorithm to test a structural hydrophobicity hypothesis, and the fitted model was wrapped in a command-line design calculator with rule-based safety warnings.

Data sources

• Synthetic dataset of 144 observations — 25 serum proteins across 6 nanoparticle formulations, built from literature-reported corona compositions
• Tenzer et al., ACS Nano 7(4):3253-3263 (2013) — size-dependent corona composition
• Walkey et al., ACS Nano 8(3):2439-2455 (2014) — corona fingerprinting and cellular interaction
• Monopoli et al., Journal of the American Chemical Society 133(8):2525-2534 (2011)
• Ke et al., ACS Nano 11(11):11773-11784 (2017)
• AlphaFold Protein Structure Database — 9 predicted structures retrieved and analysed

Figures from this analysis