Nanoparticle safety stops being a material-only question: a host-response axis lifts biosafety AUC from 0.66 to 0.90

Predicting nanoparticle biosafety no longer rests only on the particle's physicochemical descriptors; three host-response programmes enter the model as an interpretable information layer. Material-centred frameworks such as nano-QSAR and computational nanotoxicology have treated biological responses as individual toxicity endpoints or mechanistic signatures, which transfers poorly across heterogeneous experimental systems. Across 893 experimental contexts spanning six nanoparticle classes, more than 40 cell types or models, and 405 in vitro and 488 in vivo conditions, three feature configurations were compared after study-level splitting into training and independent test sets.

Predictive performance rises as each information layer is added, with exposure and biological context contributing the largest gain and the cellular-response layer adding a further increment. Material properties alone gave limited discrimination, adding exposure conditions and biological context improved it substantially, and adding the three response programmes improved it again. With Random Forest, biosafety ROC-AUC went from 0.66 to 0.88 and then 0.90, and recovery from 0.64 to 0.77 and then 0.78; in separate analyses biosafety exceeded 0.90 in vitro and in vivo, while recovery reached 0.83 in vitro and 0.70 in vivo.

On cellulose–cobalt nanocomposites absent from model development, independent molecular and cellular measurements landed where the response axis predicted. These nanocomposites were not part of the 893-context discovery dataset and were not used during model development, making them a test on unseen materials. In BV2 microglia, IFNγ raised NF-κB p65 fluorescence from about 189.4 to 605.2 arbitrary units, and C3Co1 and C3Co5 lowered it to about 209.3 and 141.8; the viable population fell from about 86% in controls to 38.2% and recovered to about 80.8% and 92.1%.

The same three programmes also organize transcriptomic data outside the nervous system, though each dataset occupies its own configuration rather than a single fixed pattern. In lung and liver datasets, in vitro and in vivo, injury, exposure and therapeutically improved conditions occupied distinct positions in the response space instead of shifting uniformly along one programme. Four independent transcriptomic datasets covered carbon black nanoparticles, multi-walled carbon nanotubes, silver nanoparticles and lipid nanoparticles; outcome labels came from the original studies' conditions and phenotypes, independent of the expression values used to compute programme activity.

The demonstrated scope is modelling on 893 neurological experimental contexts, validation of cellulose–cobalt nanocomposites absent from model development in 2D BV2 microglia and 3D spheroids, and independent brain, lung and liver transcriptomic datasets; the authors propose that direct cross-organ prediction of biosafety and recovery will require prospective outcome-labelled studies.

Response programmes are a manual qualitative abstraction of heterogeneous readouts by functional direction, which the authors say does not imply quantitative equivalence across assays or models; under in vivo recovery conditions the configuration without cellular responses marginally outperformed the one with them, so the response layer's contribution depends on experimental context; in the cross-organ analysis the axis describes the accompanying transcriptional configuration rather than independently establishing recovery.