Key points are not available for this paper at this time.
Single-particle tracking (SPT) provides a powerful approach for probing dynamic molecular processes in living cells with high spatial and temporal resolution. Yet traditional analysis pipelines, which often rely on manual tuning or simplified models, are limited by the complexity, noise, and heterogeneity inherent to biological systems. Recent advances in machine learning (ML), especially deep learning (DL), have reshaped the SPT workflow, including particle detection, trajectory linking, motion classification, denoising, and biophysical inference. In this review, we systematically assess how ML/DL methods, including convolutional neural networks (CNNs), recurrent architectures, and Bayesian deep learning, improve the accuracy, robustness, and interpretability of SPT analyses. We survey techniques ranging from CNN-based detection and linking to statistically principled frameworks for uncertainty quantification, highlighting the versatility and effectiveness of ML/DL in overcoming persistent challenges and revealing new biological insights. We also discuss practical considerations for deployment, including selection of suitable problem domains and construction of large, high-quality training data sets. This review aims to provide a comprehensive and accessible guide to the current landscape of ML in SPT, offering both a critical evaluation of existing state-of-the-art methods and a reference for future development.
Zhang et al. (Mon,) studied this question.