SMDNet: A Self-Training-Aware and Multi-Modal-Adaptive Deep Learning Network for Low-Data Aβ₄₂ Probe Design and...

Amyloid-β (Aβ) plaque accumulation is a crucial hallmark of Alzheimer's disease, and fluorescence imaging can support disease diagnosis and monitoring. However, Aβ₄₂ probe development is often hindered by trial-and-error experiments due to subtle structure-property effects. Here, we developed SMDNet, a self-training-aware and multimodal-adaptive deep learning (DL) framework for low-data Aβ₄₂ probe design and optimization. SMDNet combines iterative self-training with confidence-aware and distribution-aware sampling to improve data quality, while integrating molecular graphs, fingerprints and protein descriptors through cross-attention and protein-conditional adaptive layer normalization. Ablation studies, external validation and generalization analyses confirmed the strong predictive ability of SMDNet. Interpretability analyses further highlighted chemically meaningful substructures associated with model predictions. As a proof of concept, SMDNet guided the rational design of five ThT-derived probe candidates, with TA3 showing favorable binding affinity and high-contrast imaging performance. Additional validation on coumarin- and naphthalimide-based candidates further supported useful predictive discrimination across distinct fluorescent scaffold classes.