Single-particle tracking (SPT) has transformed the study of biomolecular systems by providing quantitative and qualitative insights into their structural and dynamical properties. Recent advances in fluorescence microscopy enable efficient nanoscale volumetric imaging in 3D and 4D, yet most conventional point-spread function (PSF) detection and SPT pipelines remain restricted to 2D, relying on fixed parameters and computationally heavy workflows. This underutilizes the rich spatial and temporal information available, leading to suboptimal accuracy and error-prone results that risk biasing downstream analysis. Here, we present a multi-module deep-learning toolbox that overcomes these limitations by integrating detection and linking of PSF and PSF-like objects across dimensions with limited computational cost. The first module employs a deep-learning architecture for sub-pixel localization, trained on physics-inspired simulations to generalize across wavelength, experimental conditions, and spatial context. An auxiliary prediction head simultaneously infers the underlying PSF, enabling refined feature extraction and expression deconvolution. The second module, a graph neural network pipeline, ingests N-dimensional detections and auxiliary features to construct a probabilistic linking landscape. By learning diffusional dynamics and event structures it enables accurate trajectory reconstruction and detection of complex events. While each module can be used independently, they are trained sequentially, establishing a grounded baseline that integrates detection and linking across heterogeneous anomalous diffusion behaviors and parameter-specific PSF expressions. The first module maximizes localization fidelity across imaging conditions, while the second module’s graph-based probabilistic linking encodes diffusional dynamics and event structures directly into trajectory reconstruction. Together, they ensure that detection quality is consistently propagated into robust linking performance. This modular design unifies detection, tracking, and dynamic event analysis within a single toolbox, while also supporting module-based adaptation into existing pipelines, thereby democratizing advanced quantitative analysis of high-dimensional data without complete restructuring of potentially existing workflows.
Bender et al. (Sun,) studied this question.