Screening of Respiratory Toxicity of Environmental Compounds Based on Multimodal Feature Fusion Model

The respiratory system constitutes the primary interface
between
the human body and the external environment, demonstrating particular
vulnerability to chemical toxicants through diverse exposure routes.
Current regulatory frameworks face significant limitations in respiratory
toxicity assessment, relying predominantly on expensive and low-throughput
animal testing methods while lacking systematic premarket evaluation
protocols. To address these challenges, we developed GFEnet, an innovative
multimodal deep learning (DL) framework that synergistically integrates
molecular graph features, structural fingerprints, and electron-level
properties for comprehensive cross-scale respiratory toxicity prediction.
The model was rigorously trained and evaluated across three toxicological
dimensions, including in vivo mammalian respiratory toxicity, in vitro
A549 cell cytotoxicity, and ACE2 gene regulation activity. GFEnet
demonstrated exceptional predictive capability, achieving outstanding
AUC values of 0.986, 0.965, and 0.919 on the respective test sets,
substantially outperforming conventional machine learning algorithms
and single-modality DL architectures. Systematic ablation studies
confirmed the critical contribution of each feature modality to the
model’s predictive power. When applied to screen compounds
from substances of very high concern and air pollutant databases,
GFEnet identified fluorene-9-bisphenol and Michler’s ketone
as high-priority risk candidates exhibiting consistent toxicity across
all evaluation end points. Subsequent in vivo validation using mouse
models confirmed these predictions, demonstrating that both compounds
induce significant pulmonary function impairment and histopathological
damage. This study establishes GFEnet as a robust high-throughput
screening platform for the early identification of respiratory toxicants,
effectively bridging computational toxicology with environmental health
protection and regulatory science.