Example single molecule tracking data from: A high-throughput platform for single-molecule tracking identifies drug interaction and cellular mechanisms

The regulation of cell physiology depends largely upon interactions of functionally distinct proteins and cellular components. These interactions may be transient or long-lived, but often affect protein motion. Measurement of protein dynamics within a cellular environment, particularly while perturbing protein function with small molecules, may enable dissection of key interactions and facilitate drug discovery; however, current approaches are limited by throughput with respect to data acquisition and analysis. As a result, studies using super-resolution imaging are typically drawing conclusions from tens of cells and a few experimental conditions tested. We addressed these limitations by developing a high-throughput single-molecule tracking (htSMT) platform for pharmacologic dissection of protein dynamics in living cells at an unprecedented scale (capable of imaging > 106 cells/day and screening > 104 compounds). We applied htSMT to measure the cellular dynamics of fluorescently tagged estrogen receptors (ER) and screened a diverse library to identify small molecules that perturbed ER function in real-time. With this one experimental modality, we determined the potency, pathway selectivity, target engagement, and mechanism of action for identified hits. Kinetic htSMT experiments were capable of distinguishing between on-target and on-pathway modulators of ER signaling. Integrated pathway analysis recapitulated the network of known ER interaction partners and suggested potentially novel, kinase-mediated regulatory mechanisms. The sensitivity of htSMT revealed a new correlation between ER dynamics and the ability of ER antagonists to suppress cancer cell growth. Therefore, measuring protein motion at scale is a powerful method to investigate dynamic interactions among proteins and may facilitate the identification and characterization of novel therapeutics.

Methods
All image acquisition using SMT was performed on a custom-built HILO microscope based on a Nikon Ti2, motorized stage, stage top environmental chamber (OKO labs), quad-band filter cube (Chroma), custom laser launch with 405 nm, and 561 nm wavelengths, coupled to a Nikon TIRF illumination module by fiber optic element (KineFlex HPV-P-3-S-405..640-0.7-APC-P2) delivering >10 mW and >600 mW of power in a gaussian beam with a FWHM of approximately 250 µm to the back focal plane of the objective. Angle of inclination and beam direction were adjusted by micrometer on the TIRF illuminator and empirically set to maximize and flatten the signal across the camera field of view. Fluorescence emission was passed through a high-speed filter wheel (Finger Lakes Instruments) and collected with a backlit CMOS camera (Prime 95b, Teledyne). Images were acquired with a 60X 1.27 NA water immersion objective (Nikon). Environmental chamber was set to 37° Celsius, 95% humidity, and 5% CO2. For each field of view, 200 SMT frames were collected at a frame rate of 100 Hz, with a 2 msec stroboscopic laser pulse. 10 frames of the Hoechst channel were collected at the same frame rate for downstream registration of trajectories to nuclei. Sample focus was maintained using the reflection-based Perfect Focus System to determine the position of the coverglass and apply an empirically-determined offset to focus into the sample.

Image acquisition produced one JF549 movie and one Hoechst per field of view. The JF549 movie was used to track the motion of individual JF549 molecules, while the Hoechst movie was used for nuclear segmentation. Tracking was accomplished in three sequential steps – detection, subpixel localization, and linking – using a combination of existing methods. Briefly, spots were detected using a generalized log likelihood ratio detector to test every 11×11 pixel window using a gaussian kernel with a radius of 1.25 pixels and with a log likelihood detection threshold 1468. After detection, the estimated position of each emitter was refined to subpixel resolution using Levenberg-Marquardt fitting69–71 with an integrated 2D Gaussian spot model72 starting from an initial guess afforded by the radial symmetry method73. Detected spots were linked into trajectories using a custom modification of a hill-climbing algorithm with a maximum linking radius of 1.25 µm and allowing a maximum of 2 gap frames where a spot may go undetected but still be linked within the same trajectory10,74. The same detection, subpixel localization, and linking settings were used for all movies used in this manuscript.

For nuclear segmentation, all frames of the Hoechst movie were averaged to generate a mean projection. This mean projection was then segmented with a neural network trained on human-labeled nuclei, the output of which is a mask assigning a semantic category to each pixel in the image75. Each spot was assigned to at most one nucleus using its subpixel coordinates.

To recover dynamical information from trajectories, we used state arrays35, a Bayesian inference approach, with the “RBME” likelihood function and a grid of 100 diffusion coefficients from 0.01 to 100.0 µm2 s-1 and 31 localization error magnitudes from 0.02 to 0.08 µm. For each assay well, 10,000 trajectories were randomly sampled from the aggregated pool of nuclear trajectories, and this set of trajectories was used for state inference. After inference, localization error was marginalized out to yield a one-dimensional distribution over the diffusion coefficient for each field of view. For single-cell analysis, we performed SMT and nuclear segmentation on a mixture of U2OS cells bearing H2B-HaloTag, HaloTag-CaaX, or free HaloTag. We then evaluated the marginal likelihood of each of a set of 100 diffusion coefficients on the set of trajectories within each segmented nucleus30. These marginal likelihood functions were clustered with k-means (3 clusters), and the marginal likelihood functions for each cell were ordered by their cluster index to produce the heat map. To estimate the fraction bound (fbound), we integrated the state array posterior distribution below 0.1 µm2 s-1. To estimate the free diffusion coefficient (Dfree), we computed the mean of the posterior distribution above 0.1 µm2 s-1 by renormalizing the distribution density of the bins >0.1 µm2 s-1 and computing the bin-weighted mean.