Fluolab【Nat.Methods】北京大学邓伍兰|提升300倍数据密度！SMLDM显微术实现活细胞分子动力学高精度制图
通讯作者： Wulan Deng
文章概要
引言
在活细胞成像领域，理解生物分子的空间组织与其扩散动力学之间的联系至关重要。然而，传统的单分子追踪（SPT）技术长期面临着一个两难境地：为了准确关联分子轨迹，必须保持极低的成像密度，这导致无法获取高密度的空间动力学图谱。为了克服这一限制，研究团队开发了SMLDM技术。该技术的核心突破在于利用深度学习直接从单帧曝光产生的“运动模糊”快照中提取分子的扩散系数和精确位置。这种方法彻底抛弃了复杂的轨迹连接过程，在保持单分子分辨率的同时，将扩散动力学的数据密度提升了50至300倍，为观察细胞内转瞬即逝的动态过程提供了全新的有力工具。
Fig. 1: Development of Deep-SnapTrack.
a, Schematic of the simulation pipeline. Single-molecule trajectories are synthesized into experimental-like snapshots by convolving with the microscope’s PSF, G__i,j,r, and adding noise. The ground-truth track image has a pixel size of 2 nm, whereas the snapshot is pixelated at 110 nm. b, Examples of simulated molecule trajectories, track images, probability distributions of trajectory signal intensity (Traj__i,j), pixelated Traj__i,j and snapshots with integrated noise at different values of D. c, Architecture and training workflow of Deep-SnapTrack. Simulated snapshots are upscaled tenfold via nearest-neighbor interpolation to create inputs, which are paired with ground-truth track images for training. Training is supervised by a loss function combining L1 loss and mean squared error (MSE) between the output pseudotrack and ground-truth track images. d, Representative molecule snapshots at different ground-truth D values, resolved by Deep-SnapTrack. e, SSIM between ground-truth (GT) tracks and different image sources for the molecules in d. Blue indicates snapshot images (solid) and background (dashed) versus ground-truth tracks. Orange indicates pseudotracks (solid) and processed backgrounds (dashed) versus ground-truth tracks. f, Representative molecule snapshots at different SNRs and ground-truth D values, resolved by Deep-SnapTrack. g, SSIM (mean ± s.d.) between ground-truth tracks and either snapshots (circles) or pseudotracks (squares) across various SNRs (n = 100 molecule snapshots per SNR); scale bars, 500 nm.
主要实验及结论
研究人员首先构建了名为Deep-SnapTrack的卷积神经网络，通过海量的模拟数据训练，使模型能够从带有噪声的单分子快照中重建出分子的伪轨迹（pseudotrack）。实验证明，基于扩散理论开发的TrackD算法可以根据这些伪轨迹的面积精确计算出扩散系数。在定位精度方面，团队引入了自适应的TrackL算法，它能根据分子的运动状态自动切换拟合方式，显著提升了快速扩散分子的定位精度。为了将该技术应用于复杂的细胞环境，研究者进一步开发了基于U-Net的分割网络，实现了在高密度成像条件下对单分子快照的自动提取。
Fig. 2: Estimation of molecule localization and D using a Deep-SnapTrack-resolved pseudotrack.
a, Schematic of the SMLDM concept. A molecule snapshot is processed by the Deep-SnapTrack network to generate a pseudotrack. The TrackD algorithm calculates D from the pseudotrack area using equation (8). Depending on the D value from TrackD, the TrackL algorithm calculates the molecule’s centroid using either Gaussian fitting or the weighted centroid of the pseudotrack. b, Plot of pseudotrack area (_PT_area; mean ± s.d.) versus ground-truth diffusion coefficients (_D_ground-truth), fitted with equation (8). Data points represent 100 molecules simulated at each of 27 logarithmically spaced D values (exposure time: 30 ms, pixel size: 110 nm, SNR: 56 (I__peak / σ__noise)). Their pseudotrack images were generated using Deep-SnapTrack, and _PT_area was subsequently calculated for each molecule to plot against its _D_ground-truth. c–e, Comparison of diffusion coefficients: D from pseudotrack (_D_pseudotrack; mean ± s.d.) versus _D_ground-truth (c), D from MSD analysis (_D_MSD; mean ± s.d.) versus _D_ground-truth (d) and _D_pseudotrack versus _D_MSD (e). Data are from 100 molecules at each of 27 D values. Pearson correlation coefficients were calculated for log-transformed D values. f,g, Scatter plots of _D_pseudotrack versus _D_MSD for datasets simulated with varying pixel sizes (at a constant 30-ms exposure; f) and varying exposure times (with a constant 110-nm pixel size; g). h, From left to right, a ground-truth 30-ms track overlaid with its geometric center, a molecule snapshot overlaid with the localization from elliptical Gaussian fitting, the corresponding pseudotrack overlaid with the localization from the TrackL algorithm and a comparison of all three localizations. i, Average localization precision from elliptical Gaussian fitting of snapshots or from the weighted centroid of pseudotrack images across different values of D. Deviation from the geometric center of the molecule trajectory was used to calculate localization precision. Precision was calculated as the deviation from the trajectory’s geometric center. For each value of D, 100 snapshots at an SNR of 17 (_I_peak / _σ_noise) were analyzed.
Fig. 3: Development of U-Net-based molecule segmentation and MPALM pipeline.
a, U-Net training. The network was trained on paired simulated snapshots (four molecules per 32 × 32-pixel image) and ground-truth mask images. A loss weight map was applied to enforce learning of intermolecule borders. b, Dice coefficient (mean ± s.d.) of U-Net segmentation performance on simulated snapshots of identical trajectories across varying SNRs (n = 100 snapshots per SNR level). c, Overview of the MPALM pipeline with exemplary data; scale bars, 500 nm. d, Probability distribution of log10 (D) measured by MPALM for benchmark Halo-tagged molecules exogenously expressed in U2OS cells and 0.1-μm TetraSpeck microspheres immobilized on coverslips. Dashed lines represent the two-component GMM fit. e,f, Quantification of the bound fraction, free fraction and D of the free component (_D_free; all shown as mean ± s.d.) derived from the GMM fits for MPALM (e) and saSPT (f). For d and e, the following were the numbers of biologically independent cells: 1×NLS (n = 9), 2×NLS (n = 9), 3×NLS (n = 10), 6×NLS (n = 10), FOXA2 (n = 10) and H2B (n = 12); fields of view of beads sample (n = 8). For f, the following were the numbers of biologically independent cells: 1×NLS (n = 8), 2×NLS (n = 12), 3×NLS (n = 8), 6×NLS (n = 12), FOXA2 (n = 20) and H2B (n = 16); fields of view of beads sample (n = 8). Significance was determined by one-way analysis of variance (ANOVA) with a Tukey’s post hoc test; *P < 0.1, **P < 0.01 and ***P < 0.001; NS, not significant.
在活细胞应用中，这项被称为Mobility-PALM (MPALM) 的系统展示了惊人的解析能力。在染色质组织研究中，MPALM揭示了核小体在活细胞内聚集形成的微米级染色质域，并发现高密度区域的分子运动明显受限。在药物研发相关的GPCR研究中，研究人员观察到μ-阿片受体在激动剂刺激下会形成特定的信号微域，而偏向性激动剂则表现出不同的动力学特征。此外，该技术还成功记录了局灶粘附（FA）在移动过程中的分子组装与拆解，以及生物分子凝聚体在相分离早期的非均匀扩散特性。这些实验结果一致表明，SMLDM能够在纳米尺度上同时绘制分子的空间分布与动力学异质性图谱。
Fig. 4: Visualizing the diffusivity distribution of nuclear proteins in living cells using high-density MPALM.
a, Representative images of the indicated Halo-tagged molecules in living U2OS cells: PALM (molecule density), diffusivity map and MPALM. The MPALM image uses the HSV color model (hue: diffusivity; saturation: density; value: 1 for signal and 0 for background). The color bar indicates D values; scale bar, 5 μm. b, H2B MPALM analysis: full-cell H2B MPALM image (i); scale bar, 5 μm. Zoomed-in views of the yellow square in i (ii–vi) show PALM density (ii), diffusivity map (iii), MPALM (iv), single-molecule localizations color coded by diffusivity (v) and seven chromatin density classes identified by the HMRF model (vi; class 1 = lowest density, class 7 = highest density). White arrows indicate a dense CD, and yellow arrows indicate the peripheral region outside of the dense CD; scale bars, 500 nm. Also shown are zoomed-in views of the white dashed square in vii (vii–ix), showing chromatin classes (vii); a diffusivity map overlaid with borders of the IC, PC and CD (viii) and an MPALM image overlaid with the same borders (ix); scale bars, 200 nm. For a and b, we imaged six biologically independent cells for each sample and obtained similar results. c, Heat map showing the probability of H2B–Halo localization densities across different D ranges within the seven chromatin classes (grouped as IC, PC or CD).
Fig. 5: MPALM of the μOR in response to agonist and antagonist treatment.
Halo-tagged μOR was imaged in living U2OS cells treated with DMSO, 10 μM DAMGO, 1 μM PZM21 or 10 μM naloxone using TIRF illumination and a 100-ms exposure per frame. a, MPALM image of membrane μOR; scale bar, 5 μm. For a, we imaged eight biologically independent cells for each treatment and obtained similar results. b–d, Zoomed-in views of the yellow square in a, showing PALM (molecule density; b), a diffusivity map (c) and MPALM (d); scale bar, 1 μm. e, Model of μOR dynamics under the indicated treatments; HT, HaloTag. f, Probability distribution of log10 (D) for Halo–μOR under each treatment. Dashed lines represent the two-component GMM fit; PDF, probability distribution function. g–j, Quantification of μOR populations (mean ± s.d.) from the GMM fits in f: fraction of molecules in the ‘immobile’ (g) and ‘mobile’ (h) states and their corresponding mean diffusion coefficients (i and j). Error bars represent s.d. (n = 6 biologically independent cells per sample). Significance was determined by one-way ANOVA with a Tukey’s post hoc test; *P < 0.1 and ***P < 0.001.
Fig. 6: MPALM reveals the dynamic biomolecular organization of FAs and optoDroplets.
a–d, Representative images of Halo-tagged paxillin in a living U2OS cell: widefield (a), PALM (density; b), diffusivity map (c) and MPALM (d); scale bar, 5 μm. For a to d, we imaged 16 biologically independent cells and obtained similar results. e,f, A five-frame time-lapse sequence showing the downward movement of an FA, displayed as PALM (e) and MPALM (f) images. Zoomed-in views from the yellow boxes highlight FA disassembly over time. Time labels indicate the midpoint of the merged frames; main scale bar, 500 nm; zoom scale bar, 200 nm. g, Widefield images of TMR-stained HaloTag-modified optoDroplets captured in each imaging cycle; scale bar, 5 μm. h–l, Zoomed-in views of the condensate boxed in g, showing bulk TMR signal (h), PALM density (i), MPALM (j), single-molecule localizations color coded by diffusivity (k) and diffusivity map (l). Time labels indicate the midpoint of merged frames; scale bars, 500 nm.
总结及展望
SMLDM框架的成功开发，标志着超分辨成像从单纯的“结构制图”向“动态制图”的重大跨越。它不仅解决了传统追踪技术在数据密度上的瓶颈，还通过深度学习与物理模型的结合，实现了对细胞内复杂动力学过程的高灵敏度监测。该方法具有极强的通用性，可以轻松集成到现有的多种单分子成像系统中。未来，随着荧光探针性能的进一步提升和算法的优化，SMLDM有望在多靶点交互作用、药物机理筛选以及疾病病理诊断等领域发挥核心作用，为生命科学研究提供更深层次的分子景观。