Two-colour live-cell nanoscale imaging of intracellular targets

Stimulated emission depletion (STED) nanoscopy allows observations of subcellular dynamics at the nanoscale. Applications have, however, been severely limited by the lack of a versatile STED-compatible two-colour labelling strategy for intracellular targets in living cells. Here we demonstrate a universal labelling method based on the organic, membrane-permeable dyes SiR and ATTO590 as Halo and SNAP substrates. SiR and ATTO590 constitute the first suitable dye pair for two-colour STED imaging in living cells below 50 nm resolution. We show applications with mitochondria, endoplasmic reticulum, plasma membrane and Golgi-localized proteins, and demonstrate continuous acquisition for up to 3 min at 2-s time resolution.

the corresponding FWHM (b). The clusters in b represent raw (unsmoothed) image data and are plotted with a color scale normalized to the minimal and maximal pixel value within each panel. Scale bar = 100 nm. While we do not know the size distribution of transferrin receptor clusters, we anticipate that many clusters are smaller than the resolution limit of our microscope. We therefore use the size-estimates from the smallest clusters as an estimate of our resolution. A significant fraction of the size distribution is below 50 nm, suggesting that our live-cell image resolution is below 50 nm. See the Supplementary Note 1 detailing our choice of pixel size and other imaging parameters. Figure 4: Absence of crosstalk between the two channels. COS-7 cells expressing Halo-Sec61β were labeled with either (a) SiR-CA or (b) 590-CA. Cells were then fixed with 3% PFA and 0.1% glutaraldehyde. STED images of both samples were acquired using both excitation sources (594 nm and 650 nm) and detection windows (624/40 nm and 685/40 nm). When imaging with the 650 nm excitation laser and detecting in the 685/40 nm range only signal from the SiR dye is detected. Likewise, when imaging with the 594 nm excitation laser and detecting in the 624/40 nm range only signal from the ATTO590 dye is detected. The two crosstalk-free detection channels are marked by asterisks (*). Scale bars = 1 µm. The histograms of pairs of pixel values in the two detection channels on the right allow to quantify the crosstalk between the two detection channels. Linear regression of the data in the histograms (signal vs. crosstalk) yielded crosstalk estimates of 0.4% for SiR (signal in the 624/40 nm channel relative to the 685/40 nm) and 1.0% for ATTO590 (signal in the 685/40 nm channel relative to the 624/40 nm channel). While we can obtain 20-30 nm resolution with our STED nanoscope (Supplementary Figure 2), we chose a pixel size of 20 nm for live-cell STED imaging, thereby allowing a resolution no better than 40 nm according to the Nyquist limit. This decision was based on a compromise between imaging speed and photobleaching on one hand and spatial resolution on the other hand. Larger pixel sizes allow to image the same field of view significantly faster, because fewer pixels have to be recorded. This, plus the fact that the STED laser power can be reduced (since effective focus sizes below the Nyquist resolution limit are of no advantage), decreases the laser light exposure and, thereby, reduces bleaching and allows to acquire longer image sequences.

Supplementary
When considering temporal resolution in a STED nanoscope, it is important to note an inherent advantage of a point scanning geometry over widefield recording, namely that small, local structures are imaged much quicker than the time it takes to record a full frame. While it takes about 2 s to scan a ~10 x 10 µm field of view ( Fig. 1-4), structures on the 250 nm scale in the same image are fully scanned in about 50 ms. Each individual line of the image (which is accumulated from 32 line scans per color channel) is finished scanning in ~4 ms. For this reason, small fast-moving objects are not nearly as susceptible to motion blur as they would be in a widefield microscope. Hence, when observing local dynamic events limited to small sections of an image, such as the image sequences shown in the subpanels of Figure 1-4, the data should be interpreted as short, sub-second snapshots interleaved by the 2 s it takes to scan the full field of view. While the point scanning geometry inherently suppresses motion blur, it is, however, important to be aware that the top and bottom of an image are out of sync by nearly 2 s. This can lead to motion artifacts when imaging large structures which extend along the slow scanning image axis. These characteristics can lead to artifacts as shown, for example, in Supplementary Movie 1. Here, the fast moving ER tubules appear jagged only along the slow scanning axis of the image.

Confocal microscopy
Confocal images in Supplementary Figure 5 were acquired with a commercial Leica TCS SP5 microscope. 595 nm and 633 nm wavelength were used to excite ATTO590 and SiR, respectively. Fluorescence was detected using avalanche photodiodes. Imaging was performed with a 100X/1.4 NA oil immersion objective lens.

STED imaging on Leica TCS SP8 STED 3X
STED imaging in Supplementary Figure 7 was performed with a Leica TCS SP8 STED 3X. mEmerald, ATTO590 and SiR were excited with 485 nm, 594 nm and 650 nm wavelength respectively. A 592-nm laser was used for depletion of mEmerald while a 775-nm laser was used for depletion of ATTO590 and SiR. HyD detectors were used for the detection of all three channels. The detection windows were adjusted to 505-550 nm, 604-644 nm and 665-705 nm for mEmerald, ATTO590 and SiR, respectively.

Image processing and deconvolution
To reduce noise, STED movie sequences were deconvolved using the Richardson-Lucy algorithm 4, 5 implemented in python and available as part of the python-microscopy package (code.google.com/p/python-microscopy). Specifically, each 2D frame was deconvolved using a 2D Lorentzian approximation to the STED imaging point spread function (PSF) 6-8 with a full width at half maximum (FWHM) estimated from STED images of sub-diffraction sized fluorescent beads. Our principal goal of noise reduction was facilitated by initializing the algorithm with a uniform prior and terminating after a relatively low number of 10 iterations, both of which have a strong regularizing effect. Under these conditions we see a significant improvement in the signal-to-noise ratio, but little to no change in resolution. When used like this the algorithm is also relatively insensitive to the small variations in the FWHM of the PSF that may result from differing imaging conditions. Raw images were corrected for bleaching using the exponential fit method in ImageJ 9 prior to deconvolution. Images and raw movies were smoothed with a Gaussian filter with 1 pixel standard deviation using ImageJ.