A Multimodal Generative AI Framework for Predicting the Toxicity of Nanoparticles

Predicting the cytotoxicity of engineered nanoparticles remains a significant challenge due to the vast combinatorial diversity of their physicochemical properties. In this study, we developed a multimodal generative framework to synthesize high-fidelity nanoparticle candidates with predefined toxicity indices. We used a large language model to extract heterogeneous data from scientific articles and utilized SciBERT-based embeddings to encode unstructured textual toxicity summaries. Four generative architectures-CTGAN, TVAE, WGAN-GP, and TabDDPM-were benchmarked using the Synthetic Data Vault quality score. The TabDDPM demonstrated superior performance in capturing complex structure-activity relationships, achieving an SDV quality score of 0.78. The case study validation and feature evolution analysis prove the practical efficacy of the TabDDPM. To validate the physical plausibility of the best generated model, we conducted coarse-grained molecular dynamics simulations in the GROMACS 2026.0 engine using the Martini 3.0.0 force field. Comparative analysis of safe and toxic nanoparticles candidates revealed that the toxic variant induced 2.4 times higher electrostatic stress (88.76 kJ/mol) and significantly prolonged membrane equilibration times. The safe candidates had a lower center-of-mass distance between the nanoparticle and the hydrophobic core of the lipid bilayer compared to the toxic counterpart. These results confirm that the proposed generative approach not only replicates statistical distributions but also captures the underlying biophysical mechanisms of membrane disruption, providing a potentially robust tool for the in silico design of biocompatible nanomaterials.