MOHAMMED HASHIM ALBASHIR, SAAD ALI S. ALJOHANI, ABRAR KHALID ALOUFI, ABUBAKER M. HAMAD, MOHAMMED EZZELDIEN HAMZA MUSTAFA, SAFAA ABASS YOUSIF MOHAMMED, BAYAN GHAZI AHMED ALSHARIF, HA ELTAHIR
DOI: https://doi.org/Nanotechnology offers exciting promises for medicine; yet, the potential neuropsychological effects of nanoparticle exposure remain a concern. The evidence is also converging to support that effects due to nanoparticles are likely modulated by the host genomics and microbiome composition, highlighting the need for integrative predictive frameworks.
In this study, we develop a multimodal AI model that simultaneously integrates nanoparticle physicochemical data, host microbiome profiling data, and genomic signatures to predict neuropsychological responses following exposure of nanoparticles.
Open-access datasets were used, specifically the caNanoLab and EPA nanosilver MEA data for nanoparticle characterization and neurotoxicity, Qiita/MGnify for microbiome features, and curated genomic panels specific to neural signaling pathways. Data Harmonization: Some examples of preprocessing, dimensionality reduction (PCA/autoencoders) and normalization techniques. A multimodal deep learning model was designed with 3 parallel branches: nanoparticle physicochemical features routed through gradient-boosted trees, microbial abundance vectors modeled using feedforward layers and genomic features mapped to latent embeddings. Fusion layers combined outputs for joint learning, their prediction targets were neural electrophysiological activity (i.e., spike rates, PSD shifts) and neuropsychological scores (cognitive/behavioral scale).
The proposed model achieved robust predictive performance (ROC-AUC ≈ 0.86, PR-AUC ≈ 0.81), outperforming unimodal baselines by 15–20%. Feature attribution (SHAP analysis) identified nanoparticle size and surface coating as primary physicochemical determinants, while specific microbiome taxa . and genetic variants in synaptic signaling genes contributed significantly to prediction accuracy. EEG and MEA-derived biomarkers revealed consistent alterations in alpha and beta power spectra post-exposure, aligning with behavioral outcomes reported in the literature.
The results demonstrate the feasibility of applying AI algorithms to combining nanomaterial, microbiome and genomic data for nano-exposure neuropsychological effect predictions. This framework not only facilitates the mechanistic interpretations in nanotoxicology but also enables a scalable approach to risk stratification and personalized safety evaluation in nanomedicine.
