Early machine learning methods provided continuous advances in feature-based classification, especially for nanopore sensing to overcome the inherent limitations such as the intrinsic noise, non-stationarity, and high dimensionality of ionic current signals. However, their reliance on handcrafted descriptors constrained sensitivity and scalability. Deep learning has since transformed the field by automating feature extraction, enhancing signal-to-noise ratio, and enabling robust classification across diverse biomolecules. Architectures, such as convolutional neural networks, recurrent neural networks, and transformers, have advanced tasks from signal denoising to multimodal fusion, driving applications in nucleic acid variant calling, protein modification profiling, and glycan analysis with femtomolar-level sensitivity. Remaining challenges, including data scarcity, interpretability, and real-time deployment, point toward emerging directions, including self-supervised and few-shot learning, physics-informed and explainable artificial intelligence (AI), and lightweight models for point-of-care use. The evolution from traditional machine learning to deep learning, and ultimately toward foundation models and cross-modal frameworks, positions AI-enhanced nanopore sensing as a cornerstone for next-generation molecular analytics and precision medicine.
Gao et al. (Fri,) studied this question.