Paper List

期刊: ArXiv Preprint
发布日期: 2026-03-11
BioinformaticsComputational Biology

Single Molecule Localization Microscopy Challenge: A Biologically Inspired Benchmark for Long-Sequence Modeling

Technische Universität Wien

Fatemeh Valeh, Monika Farsang, Radu Grosu, Gerhard Schütz
Figure
Figure
Figure
Figure

30秒速读

IN SHORT: This paper addresses the core challenge of evaluating state-space models on biologically realistic, sparse, and stochastic temporal processes, which are not captured by existing benchmarks focused on dense, regularly sampled data.

核心创新

  • Methodology Introduces SMLM-C, the first benchmark dataset specifically designed to evaluate long-sequence models on sparse spatiotemporal localization data with known ground truth, spanning dSTORM and DNA-PAINT modalities.
  • Methodology Formulates SMLM reconstruction as a sequence-to-set prediction task, requiring models to disentangle overlapping localization clouds by jointly exploiting spatial and temporal context over up to 10,000 frames.
  • Biology Reveals that state-space model performance degrades substantially as temporal discontinuity increases (e.g., detection accuracy drops from ~73% to ~62% when average off-time increases from 100 to 1000 frames), highlighting fundamental challenges in modeling heavy-tailed blinking dynamics.

主要结论

  • State-space models show limited absolute performance on SMLM reconstruction, with the highest detection accuracy reaching only 73.4% ± 1.23% (S5-L on μ_off=100 frames) and dropping to 69.6% ± 0.21% (Mamba-2-L on μ_off=1000 frames) under a 20 nm matching threshold.
  • Model performance is strongly influenced by temporal sparsity, with all evaluated architectures (S5 and Mamba-2) showing degraded performance as average off-time increases from 100 to 1000 frames, indicating fundamental challenges in handling long-range temporal dependencies.
  • Mamba-2 demonstrates better robustness to long temporal gaps, outperforming S5 in the long off-time regime (μ_off=1000 frames), while S5 performs better under shorter dark states (μ_off=100 frames), suggesting architectural differences in handling temporal discontinuity.
研究空白： Existing benchmarks for state-space models focus on synthetic tasks or domains with dense, regularly sampled signals, leaving a critical gap in evaluating their performance on biologically realistic data characterized by extreme sparsity, irregularity, and heavy-tailed temporal dynamics.

摘要: State space models (SSMs) have recently achieved strong performance on long-sequence modeling tasks while offering improved memory and computational efficiency compared to transformer-based architectures. However, their evaluation has been largely limited to synthetic benchmarks and application domains such as language and audio, leaving their behavior on sparse and stochastic temporal processes in biological imaging unexplored. In this work, we introduce the Single Molecule Localization Microscopy Challenge (SMLM-C), a benchmark dataset consisting of ten SMLM simulations—spanning dSTORM and DNA-PAINT modalities with varying hyperparameter—designed to evaluate state-space models on biologically realistic spatiotemporal point-process data with known ground truth. Using a controlled subset of these simulations, we evaluate state space models and find that performance degrades substantially as temporal discontinuity increases, revealing fundamental challenges in modeling heavy-tailed blinking dynamics. These results highlight the need for sequence models better suited to sparse, irregular temporal processes encountered in real-world scientific imaging data.