Physics-informed topological representations for interpretable nanotoxicity modelling in metal-oxide nanoparticles
Аннотация
The rapid growth of engineered nanomaterials has outpaced conventional toxicological screening, motivating predictive frameworks that are interpretable as well as accurate under data-limited conditions. Most machine-learning approaches to nanotoxicity operate as black boxes and give little insight into the physical structure of nano–bio interactions. Here we develop the physics-informed Hodge network (PIHN), an interpretable new approach methodology (NAM) that integrates Derjaguin–Landau–Verwey–Overbeek (DLVO) interaction physics with algebraic topology to construct interaction-aware representations for nanotoxicity prediction. PIHN builds DLVO-weighted simplicial complexes from physicochemical descriptors and applies a Hodge decomposition to extract gradient-, curl- and harmonic-associated features, with the relative contribution of these components inferred from validation performance rather than imposed a priori. We evaluate the framework on a curated NanoWiki-derived set of 61 experimental records covering two metal-oxide families—amorphous silica ($${\text{SiO}}_{2}$$) and hafnium dioxide ($${\text{HfO}}_{2}$$)—expanded to 1037 samples by constrained perturbation, and we keep three evaluation protocols distinct throughout because they answer different questions. On the conventional random 80/20 benchmark, PIHN attains $${R}^{2}$$ = 0.715, within the interval spanned by tuned tree ensembles (0.704 for extra trees, 0.695 for gradient boosting); bootstrap confidence intervals overlap, and no paired test separates them, so no accuracy advantage over established ensembles is claimed. That benchmark is itself optimistic: because augmented replicates of one record fall on both sides of a random split, it rewards recall of a near-duplicate. Under record-grouped cross-validation, which confines every replicate of a record to onefold, the tree ensembles and PIHN each lose roughly 0.14 in $${R}^{2}$$ and PIHN reaches 0.588 ± 0.138 against 0.585 ± 0.128 for extra trees, a difference of + 0.003 ( p = 0.97). The record-grouped value, not the benchmark, is the one to quote for accuracy. In leave-one-study-out transfer between the two material families, no model transfers reliably: PIHN and k -nearest neighbours each exceed the held-out family mean in one direction and fall below it in the other, and every other model fails in both (PIHN $${R}^{2}$$ = + 0.027 and − 0.027; k -nearest neighbours − 0.107 and + 0.200). These small positive values are not robust to the choice of Hamaker constant either, so no claim of cross-chemistry generalisation is made. A matched feature-map control shows a generic degree − 3 polynomial expansion of the same four descriptors performing comparably (0.609 against 0.588), so no predictive advantage over elementary feature engineering is claimed either. The learned gradient–curl mixing parameter proves split-dependent rather than stable, and we therefore report it as an exposed, inspectable quantity rather than as evidence for a biological mechanism. The contribution we claim is transparency of the representation and an explicit account of what each evaluation protocol can and cannot support.