22 May 2025 in Journal of Physics: Photonics
Diffusion is a fundamental process in many scientific disciplines, and accurately characterizing diffusion at the single-molecule level is crucial for understanding complex dynamic systems. To advance and benchmark anomalous diffusion (AnDi) analysis methods in the presence of motion changes, an international team of scientists launched the second AnDi Challenge in March 2024, assessing and comparing new and existing methods for diffusion estimation in dynamic systems. In response to the challenge, we introduce U-LFormer, a deep learning framework designed for precise motion analysis in single-particle tracking (SPT) experiments. By integrating advanced network architecture, training objective, and motion change analysis, U-LFormer excels in change points detection and diffusion parameters recognition across trajectories of various lengths. In the second AnDi Challenge, U-LFormer achieved first place in the Video Track Ensemble task and second place in the trajectory track single-trajectory task, being the only method to consistently rank among the top performers across all tasks. Further evaluations in this study demonstrate the robustness and versatility of U-LFormer, pushing the frontiers of diffusion analysis in SPT experiments.
28 October 2026 in Advanced Science
Single‐particle tracking (SPT) serves as a fundamental technique for deciphering complex microscopic dynamics in biological and physical systems by resolving trajectories of individual particles. However, linking localizations into trajectories remains challenging in realistic experiments, where crowding, blinking, and localization errors lead to linking ambiguity. Both statistical and deep‐learning approaches have been explored in these conditions, yet reliable link selection that generalizes across varying data distributions remains elusive. Here, we propose ReLink, a general deep‐learning framework for reliable trajectory linking in SPT. By integrating physical priors into geometric representations and leveraging confidence‐based adaptive filtering at inference, ReLink enables automated selection of trustworthy linkages, thereby preserving the fidelity of downstream trajectory‐based analyses. Across challenging simulated scenarios spanning broad experimental conditions and motion regimes, ReLink consistently outperforms representative statistical and deep learning‐based trajectory linking methods in both edge‐level and trajectory‐level evaluation metrics, leading to markedly improved downstream inference on public benchmarks. We further validate ReLink on diverse real‐world experiments, from free diffusion systems to crowded intracellular environments, demonstrating its practical significance for quantitative interpretation of microscopic dynamics and its promise for next‐generation trustworthy and automated SPT.