SPTnet

Single-particle tracking turns fluorescence microscopy video into quantitative statements about how individual biomolecules move. The conventional route to those statements is a chain of independent steps — detect spots, localize them, link them across frames, then fit a motion model to the resulting trajectory — and each step discards information the next one needs. Localization errors propagate into linking, linking errors propagate into motion inference, and mean-square-displacement fitting requires long tracks that live-cell experiments rarely provide.

SPTnet, from Cheng Bi, Fang Huang and colleagues at Purdue University, replaces the entire chain with one network trained against a single global loss. The video goes in; trajectories, generalized diffusion coefficients and Hurst exponents come out together. Framing detection, tracking and parameter inference as one set-prediction problem means the network can use motion behaviour to disambiguate linking and use linking to sharpen motion estimates, which sequential pipelines cannot do.

The Hurst exponent is the payoff. Under a fractional Brownian motion model it distinguishes subdiffusion, ordinary Brownian motion and superdiffusion on a continuous scale, so heterogeneous or actively transported populations can be separated without assuming a motion type in advance. The preprint was posted in February 2025.

#Key Features

• Single global loss over the whole pipeline: Hurst exponent error, diffusion coefficient error, per-frame detection probability and coordinate distance are optimized jointly, so trajectory reconstruction is informed by motion behaviour rather than geometry alone.
• Two-stream transformer architecture: A spatial transformer resolves positions within each frame while a temporal transformer processes all frames together, and a fusion module lets global dynamics disambiguate emission patterns that are ambiguous frame by frame.
• Anchor-free set prediction: Track queries are assigned to ground-truth molecules by the Hungarian algorithm rather than a fixed grid or anchor boxes, so videos with varying molecule counts, start times and durations are handled without reconfiguration.
• Fisher-information loss weighting: Per-track loss terms are normalized by their statistical precision limits, so short fast tracks are not penalized against a standard they cannot physically meet.
• Physically realistic training data: Simulated videos are rendered from phase-retrieved pupil functions, Perlin-noise heterogeneous backgrounds and sub-exposure intermediate frames, so system aberrations, autofluorescence and motion blur all appear during training.

#Technical Details

The architecture comprises a customized 3D-ResNet backbone that extracts spatio-temporal feature maps with 3D convolutions, two encoder-decoder transformer streams (Spatial-T and Temporal-T), a feature fusion module and a four-headed multilayer perceptron emitting detection probability, coordinates, Hurst exponent and generalized diffusion coefficient. Training used more than 200,000 simulated videos of single molecules following fractional Brownian traces under Poisson photon noise. On 30-frame simulated data, SPTnet's Hurst exponent error sat 12.5% ± 7.4% above the Cramér-Rao lower bound, against 35.6% ± 27.4% for mean-square-displacement analysis, 142.2% ± 68.6% for rescaled-range analysis and 338.2% ± 86.7% for the wavelet second-order derivative method — roughly 2.8-fold better than MSD overall, and 11-fold and 3.6-fold better at the strongly anomalous extremes H = 0.1 and H = 0.9. For generalized diffusion coefficients under Brownian conditions the error was 21.5% ± 8.2% above the bound versus 55.9% ± 5.1% for MSD. Experimental validation spanned three imaging systems: fluorescent beads in glycerol-water mixtures matched theoretical predictions, low-SNR supported lipid bilayer data resolved immobile and Brownian populations, and sptPALM imaging distinguished the motion behaviours of Reticulon 4 and Sec61β along endoplasmic reticulum tubules in live COS-7 cells.

#Applications

SPTnet targets the live-cell regime where conventional tracking is least reliable: low signal-to-noise, short trajectories, structured autofluorescent backgrounds and fast diffusers that blur within a single exposure. Membrane receptor dynamics, endoplasmic reticulum protein mobility, chromatin locus motion and intracellular cargo transport all produce data of this kind, and all benefit from a Hurst exponent estimate that does not require a long track. Because the training data generator is driven by a phase-retrieved pupil function, groups can retrain for their own microscope's point spread function rather than accepting a mismatch. The repository ships training and inference scripts, a MATLAB training-data generator with a GUI, visualization tools, trained models and a user manual under an MIT license.

#Impact

SPTnet shows that the end-to-end framing that reshaped object detection transfers cleanly to single-molecule imaging, and that doing so recovers information the classical pipeline throws away — its precision approaches the Cramér-Rao bound on parameters where mean-square-displacement analysis falls well short. The gains are largest exactly where biological interpretation is hardest, on anomalous motion far from Brownian. The scope is deliberately narrow: the model outputs trajectories and fractional Brownian motion parameters rather than serving as a general imaging backbone, alternative anomalous-diffusion models such as continuous-time random walks are not among its outputs, and applying it to a new optical system is best done after regenerating training data for that system's point spread function.