DualNeuro: A Dual-Model Framework for Selectivity Between Neurotoxic Insecticides and CNS Drugs

The insect nervous system represents a primary target for insecticide development. However, the evolutionary conservation of neural targets between insects and mammals, the off-target safety of neurotoxic insecticides has raised growing concerns. Existing studies have primarily focused on single-task applications, such as pesticide toxicity prediction or central nervous system (CNS) drug screening, while integrated frameworks that jointly assess neurotoxic insecticidal potential and human CNS risk remain scarce. In this study, a neurotoxic insecticide classification model (Pest-LGBM) and a CNS drug classification model (CNS-XGB) were developed using neurotoxic/non-neurotoxic insecticide datasets and CNS/non-CNS drug datasets, respectively. Both models exhibited robust classification performance and generalization ability. Pest-LGBM achieved a 100% positive prediction rate on External Test Set 1, whereas CNS-XGB achieved negative prediction rates of 80.0%, 85.0%, and 92.2% on the training, validation, and External Test Set 1, respectively. Furthermore, we established a dual-model prediction framework, termed DualNeuro, which effectively discriminates compounds with high insecticidal potential but low predicted CNS risk from those exhibiting dual-high responses. Systematic structural characterization at both the global molecular level and core scaffold level was performed via chemical space analysis and Murcko scaffold enrichment. Neurotoxic insecticides were primarily enriched in phosphorus/sulfur-containing ester motifs and hydrophobic aromatic structural units, whereas CNS drugs preferentially contained complex fused-ring scaffolds and nitrogen-containing moieties. Overall, this study provides a computational framework to support safety-oriented design of neurotoxic insecticides, non-target risk assessment, and early-stage CNS drug candidate screening.